content
stringlengths 1
15.9M
|
|---|
\section{\uppercase{Introduction}}
\label{sec:introduction}
\noindent The use of UAVs (Unmanned Aerial Vehicles) has seen extensive growth in recent years for industrial, military, and consumer use. One of the most common types of UAVs, although a naturally unstable system, is the quadrotor. Quadrotors are typically operated by remote control using a joystick \cite{Fernando2013}. There has also been an increasing interest in autonomous control of a quadrotor as seen in research in recent years.
Autonomous PID control of a quadrotor has been tested by \cite{Jun2013} and \cite{Harandi2010} where each orientation angle (pitch, roll, yaw) had an individual PID controller that was experimentally tuned. These two control approaches, as well as another controller by \cite{Zhan2012}, dealt with PID control only for attitude stabilization. Another PID controller by \cite{Li2011} added position stabilization as well as attitude stabilization. In the work by \cite{Bai2012} robust PID control was developed to deal with disturbances, although it only dealt with attitude stabilization. In a different approach to PID tuning, pole selection was performed using transfer function analysis to produce an attitude and position stabilized controller \cite{Sa2013}.
Using state space-methods for quadrotor control, the work by \cite{Reyes-Valeria2013} produced two gain matrices for a gain scheduled controller using LQR gain selection. This work used one gain matrix when the quadrotor was far away from the trajectory, and a second matrix when the quadrotor was on the desired trajectory; this work did not deal with disturbances \cite{Reyes-Valeria2013}. In a comparison of control methods \cite{Al-Younes2010} tested a PID controller, LQR controller, and non-linear Adaptive Integral Backstepping Controller for attitude stabilization; this work did not deal with trajectory tracking.
This paper presents a linear state-space method of control system design for the purpose of attitude and position stabilization as well as path tracking. The gain matrix that is designed within this work is independent of the vehicle's properties, which mean it is valid for any quadrotor configuration (provided it holds to the standard general design of a quadrotor, four fixed rotors with a mass in center).
\section{\uppercase{Modeling}}
\noindent To model the system the quadcopter configuration shown in Figure \ref{fig:QuadrotorDiagram} was used. In Figure \ref{fig:QuadrotorDiagram} it can be seen that roll is counterclockwise about the \textit{x-axis}, pitch is counterclockwise about the \textit{y-axis}, and yaw is counterclockwise about the \textit{z-axis}. Additionally Rotors 1 and 3 rotate counterclockwise, while Rotors 2 and 4 rotate clockwise.
\begin{figure}[h]
\centering
\includegraphics[width=2.5in]{figs/QuadrotorDiagram}
\caption{Quadrotor Diagram.}
\label{fig:QuadrotorDiagram}
\end{figure}
\subsection{Non-Linear Dynamics}
\noindent To derive the dynamic equations for this quadrotor first the rotation matrix from Inertial (North East Down at initial position) to Body frame (Shown on Figure \ref{fig:QuadrotorDiagram}) is defined (Equation (\ref{eq:RotationMatrix})). The roll \(\phi\) is the rotation counterclockwise about the x-axis, the pitch \(\theta\) is the rotation counterclockwise about the y-axis, and the yaw \(\psi\) is the rotation counterclockwise about the z-axis.
\begin{equation}
\footnotesize
\label{eq:RotationMatrix}
\centering
R_{IB}
=
\left[
\begin{array}{ccc}
c\theta c\psi & c\theta s\psi & -s\theta \\
s\phi s\theta c\psi - c\phi s\psi & s\phi s\theta s\psi + c\phi c\psi & s\phi c\theta \\
c\phi s\theta c\psi + s\phi s\psi & c\phi s\theta s\psi-s\phi c\psi & c\phi c\theta
\end{array}
\right]
\end{equation}
Next the set of generalized coordinates is set to be \((x,y,z,\omega_x,\omega_y,\omega_z)\), where \(x,y,z\) are the position component from initial position measured in the body axis, and \(\omega_x,\omega_y,\omega_z\) are the angular velocity components in the body axis. Lagrangian modeling techniques were then used to produce the non-linear dynamic equations of the system shown in Equation (\ref{eq:non-linearModel}) where the values for \((w_1, w_2, w_3, w_4)\) are defined in Equation (\ref{eq:wDefintion}). These equations use the quadrotor parameter \(k\) for rotor speed torque constant, \(b\) for rotor speed to drag constant, \(m\) for quadrotor mass, \(L\) for the half distance between the rotors, \(g\) is the gravitational constant, and \((I_x,I_y,I_z)\) are the components of the inertia matrix (principle axes as shown in Figure \ref{fig:QuadrotorDiagram}). The values \(\omega_{M_1},\omega_{M_2},\omega_{M_3},\omega_{M_4}\) (in Equation (\ref{eq:wDefintion})) are the four rotors angular speeds respectively. The values \(F_{D_x},F_{D_y},F_{D_z}\) represent the disturbance forces to the systems that also include aerodynamic effects lumped in with the disturbances. The values \(\tau_{D_x}, \tau_{D_y}, \tau_{D_z}\) are the disturbance torques to the system which include the aerodynamics torques and gyroscopic torques lumped in with the disturbances. This was done to simplify the modeling while still considering all internal and external disturbances. The linearized state space model does not consider the disturbance terms. In this study the rotor inertia's are neglected (assumed to be small).
\begin{equation}
\label{eq:non-linearModel}
\centering
\left[
\begin{array}{c}
\ddot{x} \\
\ddot{y} \\
\ddot{z} \\
\dot{\omega_x} \\
\dot{\omega_y} \\
\dot{\omega_z}
\end{array}
\right]
=
\left[
\begin{array}{c}
-gs{\theta} + F_{D_x}/m \\
gs{\phi}c{\theta} + F_{D_y}/m \\
gc{\phi}c{\theta} + w_1 \left(k/m\right) + F_{D_z}/m \\
-w_2 \left(kL/I_x\right) + \tau_{D_x}/I_x \\
-w_3 \left(kL/I_y\right) + \tau_{D_y}/I_y \\
w_4 \left(bL/I_z\right) + \tau_{D_z}/I_z
\end{array}
\right]
\end{equation}
\begin{equation}
\label{eq:wDefintion}
\left[
\begin{array}{c}
w_1 \\
w_2 \\
w_3 \\
w_4
\end{array}
\right]
=
\left[
\begin{array}{c}
\omega_{M_1}^2 + \omega_{M_2}^2 + \omega_{M_3}^2 + \omega_{M_4}^2 \\
\omega_{M_2}^2 - \omega_{M_4}^2 \\
\omega_{M_1}^2-\omega_{M_3}^2 \\
\omega_{M_1}^2 - \omega_{M_2}^2 + \omega_{M_3}^2 - \omega_{M_4}^2
\end{array}
\right]
\end{equation}
While these non-linear equations represent the system well, it is desirable to linearize them to simplify the control system design.
\subsection{Linearized Dynamics}
\noindent To linearize Equation (\ref{eq:non-linearModel}) first the traditional small angle approximations are made (\(\sin{x} \approx x, \, \cos{x} \approx 1\)) followed by substitutions for \(u_1,u_2,u_3\) and \(u_4\) from Equation (\ref{eq:uDefintion}).
\begin{equation}
\label{eq:uDefintion}
\left[
\begin{array}{c}
u_1 \\
u_2 \\
u_3 \\
u_4
\end{array}
\right]
=
\left[
\begin{array}{c}
g+w_1\left(k/m\right) \\
-w_2\left(kL/I_x\right) \\
-w_3\left(kL/I_y\right) \\
w_4 \left(bL/I_z\right)
\end{array}
\right]
\end{equation}
These substitutions satisfy linearization of Equation (\ref{eq:non-linearModel}); however, it is desirable to include the states for linear speed \((\dot{x},\dot{y},\dot{z})\) as well as Euler angular rates \((\dot{\phi},\dot{\theta},\dot{\psi})\) since we would like to control these as well. Considering the Euler angular rate relationships in Equation (\ref{eq:eulerRates}), it is clear these need to be linearized as well.
\begin{equation}
\label{eq:eulerRates}
\left[
\begin{array}{c}
\dot{\phi} \\
\dot{\theta} \\
\dot{\psi}
\end{array}
\right]
=
\left[
\begin{array}{c}
\omega_x+\left(\omega_y\sin{\phi}+\omega_z\cos{\phi}\right)\\
\omega_y\cos{\phi}-\omega_z\sin{\phi}\\
\left(\omega_y\sin{\phi}+\omega_z\cos{\phi}\right) / \cos{\theta})
\end{array}
\right]
\end{equation}
Once again using small angle approximation and using i) the product of two small angles is near zero, ii) the product of a small angle and an angular rate is small, a rough linearization for the Euler angular rates can be defined as \(\dot{\phi} \approx \omega_x\), \(\dot{\theta}\approx\omega_y\), and \(\dot{\psi}\approx\omega_z\). This final linearization leads to the linearized state-space model for the quadrotor in the form \(\dot{X}=AX+BU\) defined in Equation (\ref{eq:linearStateSpaceModel}). Note that to save space the \(AX\) and \(BU\) terms have been multiplied together to form vectors (Equation (\ref{eq:linearStateSpaceModel})). Furthermore, the \(A\) and \(B\) matrices can be recovered by knowing that the state vector \(X\) is \((\dot{x},\dot{y},\dot{z},x,y,z,\omega_x,\omega_y,\omega_z,\phi,\theta,\psi)^T\), and the input vector \(U\) is \((u_1,u_2,u_3,u_4)^T\). Note that this input vector \(U\) is used to solve for the rotor speeds (\(\omega_{M_1},\omega_{M_2},\omega_{M_3},\omega_{M_4}\)) when controlling the quadrotor.
\begin{equation}
\label{eq:linearStateSpaceModel}
\left[
\begin{array}{c}
\ddot{x} \\
\ddot{y} \\
\ddot{z} \\
\dot{x} \\
\dot{y} \\
\dot{z} \\
\dot{\omega}_x \\
\dot{\omega}_y \\
\dot{\omega}_z \\
\dot{\phi} \\
\dot{\theta} \\
\dot{\psi}
\end{array}
\right]
=
\left[
\begin{array}{c}
-g\theta \\
g\phi \\
0\\
\dot{x} \\
\dot{y} \\
\dot{z} \\
0\\
0\\
0\\
\omega_x \\
\omega_y \\
\omega_z \\
\end{array}
\right]
+
\left[
\begin{array}{c}
0 \\
0 \\
u_1 \\
0 \\
0 \\
0 \\
u_2 \\
u_3 \\
u_4 \\
0 \\
0 \\
0 \\
\end{array}
\right]
\end{equation}
\subsection{Controllability}
From Equation (\ref{eq:linearStateSpaceModel}), \(\dot{X}=AX+BU\), the \(A\) and \(B\) matrices can be determined as described the in previous section. Using these \(A\) and \(B\) matrices the controllability test matrix from Equation (\ref{eq:controlTestMatrix}) can be calculated \cite{Friedland1986}.
\begin{equation}
\label{eq:controlTestMatrix}
Q =
\left[
\begin{array}{ccccc}
B & AB & ... & A^{10}B & A^{11}B
\end{array}
\right]
\end{equation}
It is found that \(\text{rank}\left(Q\right)=12\) meaning that the system is controllable and that designing of the state-space controller can continue unhindered.
\section{\uppercase{System Stabilization - Pole Placement}}
\noindent To stabilize the system with a control law, first the stability of the open loop system must be assessed. It is determined that \(\text{det}\left(sI-A\right)=s^{12}\), which means that there are twelve poles at zero and the system is marginally stable. This is not a desirable behavior, therefore a gain matrix for a closed loop control law must be determined to produce desirable system dynamics and stability.
\subsection{Closed Loop Stabilization}
\noindent The feedback law that was selected to produce desirable system behavior is
\begin{equation}
\label{eq:controlLaw}
U=-GX
\end{equation}
where \(G\) is the \(4\times12\) gain matrix. The new closed loop system is thus \(\dot{X}=A_cX\) where \(A_c=\left(A-BG\right)\). To simplify the selection of the control gains it is assumed that: i) vertical control is mainly a function of \(\dot{z}\) and \(z\), ii) roll control is mainly a function of \(\dot{y}\), \(y\), \(\omega_x\). and \(\phi\) (roll), iii) pitch control is mainly a function of \(\dot{x}\), \(x\), \(\omega_y\), and \(\theta\) (pitch), and iv) yaw control is mainly a function of \(\omega_z\) and \(\psi\) (yaw). This leads to a gain matrix of the form seen in Equation (\ref{eq:gainMatrixStructure}).
\begin{equation}
\label{eq:gainMatrixStructure}
G^T=\left[
\begin{array}{cccc}
0 & 0 & g_7 & 0 \\
0 & g_3 & 0 & 0 \\
g_1 & 0 & 0 & 0 \\
0 & 0 & g_8 & 0 \\
0 & g_4 & 0 & 0 \\
g_2 & 0 & 0 & 0 \\
0 & g_5 & 0 & 0 \\
0 & 0 & g_9 & 0 \\
0 & 0 & 0 & g_{11} \\
0 & g_6 & 0 & 0 \\
0 & 0 & g_{10} & 0 \\
0 & 0 & 0 & g_{12} \\
\end{array}
\right]
\end{equation}
The dynamic behavior and the stability of this system is defined by the poles of this system, therefore one would typically think the next step is to select a set of desirable poles for this system. The problem that arises is that while ideally a direct solution for the gains in this matrix can be found from the desired closed loop poles, the characteristic equation found from \(\text{det}\left(sI-A_c\right)\) produces a set of non-linear equations that are difficult to analytically solve. Since all that is needed is a set of poles with good system behavior, an analytical solution is not necessary if a set of poles can be found numerically that produces good system behavior. Therefore, \textit{Simulated Annealing} was used to solve for a set of gains that would match a set of pole criteria.
To create the pole selection criteria it is assumed that a pair of dominant poles will exist that define the majority of the dynamic behavior of the system. This gives a good starting point for selecting a set of poles. In this paper the \(2\%\) setting time of this dominant pole system is defined by \(t_s \approx 4/\zeta\omega_n\) (where \(\zeta\) and \(\omega_n\) are the damping ratio and natural frequency of the second order system, respectively) \cite{Ogata2010}, the \(10\%\) to \(90\%\) rise time is \(t_r \approx 1.8/\omega_n\) \cite{Franklin2010}, and the overshoot is \(OS\% \approx 100\%\times e^{-(\zeta\pi/\sqrt{1-\zeta^2})}\) \cite{Ogata2010}.
After careful inspection it was observed that a pair of dominant poles with real components located at \(-6\) and low damping ratios of (\(\zeta=0.1\)) have \(t_s \approx 0.67\text{s}\), \(t_r \approx 0.03 \text{s}\), and \(\%OS \approx 73\%\). Similarly, a pair of dominant poles with real components located at \(-6\) and high damping ratios of (\(\zeta = 1.0\)) have \(t_s \approx 0.67\text{s}\), \(t_r \approx 0.3 \text{s}\), and \(\%OS \approx 0\%\). Therefore it was concluded that if the poles of the system were all less than \(-6\) in the real component, that it was likely the overall system would have \(t_s < 1.0\text{s}\), \(t_r < 0.5\text{s}\), and \(\%OS < 100\%\). This was deemed to be sufficient performance. To find the poles it was also determined that the poles should not be too far to the left to prevent excessive control effort, so poles were searched for with real components in the range of \(\ge -6\) and \(\le -30\). Using the Simulated Annealing search algorithm a set of gains for the gain matrix \(G\) were found that satisfied these criteria. Equation (\ref{eq:gainValues}) shows the gain values found, and Equation (\ref{eq:clPoles}) shows the values of the corresponding closed loop poles.
\begin{equation}
\label{eq:gainValues}
\left[
\begin{array}{c}
g_1 \\
g_2 \\
g_3 \\
g_4 \\
g_5 \\
g_6 \\
g_7 \\
g_8 \\
g_9 \\
g_{10} \\
g_{11} \\
g_{12}
\end{array}
\right]
=
\left[
\begin{array}{c}
32.8 \\
608.0 \\
394.5 \\
862.9 \\
47.1 \\
657.9 \\
-397.8 \\
-1124.2 \\
39.4 \\
552.7 \\
30.1 \\
623.4
\end{array}
\right]
\end{equation}
\begin{equation}
\label{eq:clPoles}
\text{Closed Loop Poles}
=
\left[
\begin{array}{c}
-28.32 \\
-20.36 \\
-6.47 \\
-6.28 \\
-16.40 + 18.41i \\
-16.40 - 18.41i \\
-15.05 + 19.92i \\
-15.05 - 19.92i \\
-6.29 + 6.65i \\
-6.29 - 6.65i \\
-6.25 + 2.92i \\
-6.25 - 2.92i
\end{array}
\right]
\end{equation}
At this point it is interesting to notice that this gain matrix has been derived independent of any parameters of the quadrotor. This means that this gain matrix is valid for a wide variety of quadrotor configurations which agree with the assumptions and control structure.
\section{\uppercase{Error Tracking Controller Design}}
\label{sec:errorTrackingDesign}
\noindent With the closed loop system stabilized, it is now desirable to modify the controller to allow the quadrotor to track a reference position. To do this a new closed loop system is designed such that \(\dot{e}=A_c e\) is the new system, where \(e\) is the error state vector for the control system. This results in the new control law as defined in Equation (\ref{eq:errorControlLaw}).
\begin{equation}
\label{eq:errorControlLaw}
U=-Ge
\end{equation}
The next step is to define the error state vector \(e\). For this error tracking system it is desirable that both a set position in inertial space \((x,y,z)\) and the UAVs yaw \((\psi)\) can be tracked. All other quadrotor states are ideally zero (angular and linear speeds are zero). This would seem to lead to an easy conclusion that the error vector should be the same as \(X\), except that \((x,y,z,\psi)\) are subtracted by their desired reference values. The problem with this error vector design is that with the linear model the control thinks that it can just pitch or roll a large amount to increase the linear speeds to large values. This pushes the pitch and roll outside of the linear range of the controller, causing instability and an inevitable crash of the quadrotor. To account for this a better error vector is used where the linear speeds and angular yaw rate are given errors proportional to the position and yaw errors (Equation (\ref{eq:errorVector}). When setting a position it is more likely that a user will be specifying an inertial position coordinate; however, since the quadrotor works in body frame for \((x,y,z)\) the rotation to body frame must be applied. This is performed using Equation (\ref{eq:posRotation}). In Equation (\ref{eq:errorVector}) the parameters \(k_1\) and \(k_2\) are proportionality constants for the position and yaw errors. While in this paper both \(k_1\) and \(k_2\) were set to \(1.0\), which produced good results, further tuning of these parameters can be performed.
\begin{equation}
\label{eq:errorVector}
e
=
\left[
\begin{array}{c}
\dot{x}+k_1(x-x_{B_{des}}) \\
\dot{y}+k_1(y-y_{B_{des}}) \\
\dot{z}+k_1(z-z_{B_{des}}) \\
0 \\
0 \\
0 \\
\omega_x \\
\omega_y \\
\omega_z+k_2(\psi-\psi_{des}) \\
\phi \\
\theta \\
0
\end{array}
\right]
\end{equation}
\begin{equation}
\label{eq:posRotation}
\left[
\begin{array}{c}
x_{B_{des}} \\
y_{B_{des}} \\
z_{B_{des}}
\end{array}
\right]
=
R_{IB}
\left[
\begin{array}{c}
x_{I_{des}} \\
y_{I_{des}} \\
z_{I_{des}}
\end{array}
\right]
\end{equation}
With Equation (\ref{eq:errorVector}) the input vector \(U\) can be determined using the gain matrix \(G\), and consequently the desired rotors speeds (\(\omega_{M_1},\omega_{M_2},\omega_{M_3},\omega_{M_4}\)) can be solved from the values of \(U\).
In addition to this error vector (Equation (\ref{eq:errorVector})), to further improve stability of the system, the maximum linear speed error is saturated at \(1.0\text{m/s}\) and the maximum angular rate error is saturated at \(3.14\text{rad/s}\). These values are subject to further adjustments (depending on the quadrotor properties, especially speed); however, for the purpose of this paper these values produced good results as shown in Section \ref{sec:simulationErrorTracking}.
\section{\uppercase{Simulation}}
\noindent With the stabilized controller and the error tracking controller both defined, the results of simulating the system are now presented. The quadrotor parameters used during simulations are those obtained by \cite{Harandi2010}, and are shown in Table (\ref{tab:quadParams}) with inertial matrix (\(I\)) shown in Equation(\ref{eq:inertialmatrix}).
\begin{table}[h]
\caption{Quadrotor Parameters.}\label{tab:quadParams} \centering
\begin{tabular}{|m{4cm}|c|}
\hline
Half Distance between opposite Rotors (L) & \(0.27\text{m}\) \\\hline
Mass (m) & \(1.4\text{kg}\) \\ \hline
Rotor Speed to Torque Constant (k) & \(11 \times 10^{-6} \text{Ns}^2\) \\ \hline
Rotor Speed to Drag Constant (b) & \(1.1 \times 10^{-6} \text{Nms}^2\) \\ \hline
Max Rotation Speed of Motors (\(\omega_{M_{max}}\)) & \(637.75\text{rad}/\text{s}\)\\ \hline
\end{tabular}
\end{table}
\begin{equation}
\label{eq:inertialmatrix}
\footnotesize
\text{I}
=
\left[
\begin{array}{ccc}
8.1 & 0 & 0 \\
0 & 8.1 & 0 \\
0 & 0 & 14.2
\end{array}
\right] \times 10^{-3} kg \cdot m^2
\end{equation}
The model used for the error tracking simulation is shown in Figure \ref{fig:controlDiagram}. The stability analysis diagram is the same except that the "Control Gain" block has a direct input of \(X\). In the simulation the gain matrix, which was designed using linear control design techniques, is used to control the system. The quadrotor is simulated by the non-linear model (Equation (\ref{eq:non-linearModel})). Disturbances are provided to the non-linear model in the stability simulation.
\begin{figure}[h]
\centering
\includegraphics[width=2.5in]{figs/controlDiagram}
\caption{Control Structure.}
\label{fig:controlDiagram}
\end{figure}
\subsection{Perturbation Response: Stabilized System}
\noindent To verify the stability of the control system with the selected gains, the closed loop stabilized system (ie. Hover Controller) is subjected to perturbation. The disturbances of \(1\text{m/s}\) for all linear speeds, and \(0.1\text{rad/s}\) for all angular speeds, can be seen in Figure \ref{fig:perturbDisturb}. The system response for the Euler angles can be seen in Figure \ref{fig:perturbEuler}, and for the linear inertial positions in Figure \ref{fig:perturbInertialPosition}.
\begin{figure}[h]
\centering
\includegraphics[width=3.25in]{figs/perturbDisturb}
\caption{Perturbation Simulation: Disturbances.}
\label{fig:perturbDisturb}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=3.25in]{figs/perturbEuler}
\caption{Perturbation Simulation: Euler Angles.}
\label{fig:perturbEuler}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=3.25in]{figs/perturbInertialPos}
\caption{Perturbation Simulation: Inertial Positions.}
\label{fig:perturbInertialPosition}
\end{figure}
From Figures \ref{fig:perturbEuler} and \ref{fig:perturbInertialPosition} it can be seen that the system settles quickly after the perturbation ends, reaching steady state in about \(1\text{s}\), which is considered good performance.
\subsection{Random Disturbance Response: Stabilized System}
\noindent In a second simulation (using Stabilized System, ie. Hover Controller) an extreme case of random disturbances to the linear and angular speeds of the quadrotor (to evaluate the stability in these conditions) was performed. The applied disturbances were normally distributed with a variance of \(10\text{m/s}\) for the linear speed, and \(1\text{rad/s}\) for the angular speed as shown in Figure \ref{fig:randDisturb}. The system response for the Euler angles can be seen in Figure \ref{fig:randEuler}, and for the linear inertial positions in Figure \ref{fig:randInertialPosition}.
\begin{figure}[h]
\centering
\includegraphics[width=3.25in]{figs/randDisturb}
\caption{Random Simulation: Disturbances.}
\label{fig:randDisturb}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=3.25in]{figs/randEuler}
\caption{Random Simulation: Euler Angles.}
\label{fig:randEuler}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=3.25in]{figs/randInertialPos}
\caption{Random Simulation: Inertial Positions.}
\label{fig:randInertialPosition}
\end{figure}
It can be seen from Figures \ref{fig:randEuler} and \ref{fig:randInertialPosition} that while under significant disturbances the system stays stable, and hovers without moving more than \(\pm 0.25\text{m}\) in any direction from its starting point, according to Figure \ref{fig:randInertialPosition}.
\subsection{Error Tracking}
\label{sec:simulationErrorTracking}
\noindent For the final simulation the error tracking system as designed in Section \ref{sec:errorTrackingDesign} was tested under moderate disturbances to observe the systems ability to track a position and yaw (i.e. desired \((x,y,z)\) position and a desired yaw (\(\psi\))). The disturbances were normally distributed with the linear speeds having a variance of \(1\text{m/s}\), and the angular speeds having a variance of \(0.1\text{rad/s}\) (Figure (\ref{fig:trackingDisturb})). The desired inertial position was set at \((10\text{m},\,5\text{m},\,-2\text{m})\), and the desired yaw was set at \(3\,\text{rad}\). The initial state of the system was at inertial position \((0\text{m},\,0\text{m},\,0\text{m})\) and yaw \(0\,\text{rad}\). The system response for the Euler angles can be seen in Figure \ref{fig:trackingEuler}, and for the linear inertial positions in Figure \ref{fig:trackingInertialPosition}.
\begin{figure}[h]
\centering
\includegraphics[width=3.25in]{figs/trackingDisturb}
\caption{Tracking Simulation: Disturbances.}
\label{fig:trackingDisturb}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=3.25in]{figs/trackingEuler}
\caption{Tracking Simulation: Euler Angles.}
\label{fig:trackingEuler}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=3.25in]{figs/trackingInertialPos}
\caption{Tracking Simulation: Inertial Positions.}
\label{fig:trackingInertialPosition}
\end{figure}
The results of the tracking simulation in Figures \ref{fig:trackingEuler} and \ref{fig:trackingInertialPosition} show that the quadrotor was able track well to the desired linear positions and desired yaw angle. The system stayed stable during this movement and performed well while tracking under moderate disturbances.
\section{\uppercase{Conclusion}}
This paper presented a linear state-space approach at designing a stable hover controller and a stable tracking controller (for inertial position and yaw angle) for a quadrotor Unmanned Aerial Vehicle (UAV). In designing these controllers the gain matrix for the control system was selected using a linearized model of a quadrotor using Simulated Annealing to find gains which would produce a desirable set of closed loop poles. This gain matrix was derived independent of any quadrotor properties (e.g. inertia, dimensions, mass), meaning that it is valid for a wide range of quadrotor configurations provided they agree with the assumptions and control structure.
The tracking controller designed in this paper works by providing linear velocity and angular velocity error references that are proportional to the linear position and angular position errors as feedback to control system. This type of error feedback prevents problems that would occur with a direct error feedback (i.e. subtracting actual state from reference state) where the system would be pushed out of its operating range and become unstable due to the way the system was linearized for control design.
In simulations the designed gain matrix was used to test system stability under a perturbation disturbance and under random normally distributed disturbances to the linear and angular speeds. Under a perturbation the simulated quadrotor performed well, converging back to zero error in about \(1\text{s}\). The simulated quadrotor also performed well under random normally distributed disturbances (mean of 0, variance of \(10\text{m/s}\) for linear speeds, variance of \(1\text{rad/s}\) for angular speeds) with only \(\pm 0.25\text{m}\) of movement from its zero position in any direction, according to Figure \ref{fig:randInertialPosition}.
The tracking controller was also simulated for a desired position of \((10\text{m},\,5\text{m},\,-2\text{m})\), and the desired yaw of \(3\,\text{rad}\). The simulated quadrotor was subject to moderate normally distributed disturbances with varience of \(1\text{m/s}\) for linear speeds, and a variance of \(0.1\text{rad/s}\) for angular speeds. The tracking controller performed well as the simulated quadrotor achieved the desired position and yaw while staying stable.
The results of this work show that this method of controlling a quadrotor for position tracking and hover stability performs well for this simulation and set of model parameters. This provides support for the hypothesis that since the gain matrix that was derived is independent of the quadrotor properties, it can be applied to any quadrotor system with a similar configuration (given that it agrees with assumptions and control structure). Future work would involve investigation into the effects of state estimation (measurement noise and error) on the performance of this controller, simulation on a wide variety of model parameters, and experimental testing of this control system on a real quadrotor to verify the simulation results.
\input{ICINCO14_refs.bbl}
\end{document}
|
\section{Introduction}
\label{sec:intro}
Beside the prophecy made by \citet{weaver1991science}, identifying in the study of ``organized complexity'' (i.e., systems made by many interconnected elements) as the new theoretical frontier of Sciences, the Dutch electrical engineer Bernard Tellegen, already in 1952 demonstrated the existence of regularities and laws dependent only on the wiring architecture of the studied systems \cite{mikulecky2001network,tellegen1952general}.
These regularities cause systems made by elements obeying different physical laws (e.g. electrical circuits, molecules, metabolic networks) to display a very similar mesoscopic behavior. These ``network laws'' are still largely unknown and there is a huge room of investigation for catching these principles.
This is the main focus of the emerging research field of complex networks analysis \cite{wei2014new,havlin2010,dorogovtsev2008critical,xiao2008symmetry,lacasa2013correlation,lee2006statistical,costa2007characterization}.
A well-known approach to characterize complex networks consists in analysing the fractal properties of their topology (i.e., scaling properties), which is usually implemented by means of the box counting method \cite{song2005self}.
Recently, new methods \cite{dan2012multifractal,li2014fractal,furuya2011multifractality} have been proposed to generalize this analysis to the multifractal setting \cite{harte2010multifractals}.
A network (or equivalently, a graph) is a very general mathematical construct having the power of unifying apparently different objects like correlation matrices, recurrence plots, adjacency matrices. The only necessary property to call something a graph is the existence of a wiring scheme that, for any pair of constituting elements, returns a binary information (or in certain cases, the strength) about the existence of a link between the elements.
The definition of this wiring scheme allows to transfer virtually any problem to the graph domain, granting the possibility to take advantage of all the tools coming from classical graph theory and the more recently developed complex networks analysis.
In particular, when studying time series it is possible to generate a network (in the form of recurrence plots, correlation matrices, or other mapping methods) that preserves their temporal correlation structure in the form of topological invariants \cite{donner2011,marwan2007recurrence,campanharo2011duality}.
Nonetheless, the time series / network ``isomorphism'' works in both directions. The exploration of large networks by means of a random walker, taking at each vertex different directions along the graph topology according to some probabilistic criteria \cite{bonaventura2014characteristic}, has the potential to offer a dynamical perspective of the studied network.
However, while the former approach has been deeply investigated \cite{donner2011}, less attention has been devoted to the latter in the case of (multi)fractal characterization, apart from few recent exceptions \cite{nicosia2013characteristic,weng2014time,zhou2014fractal}.
In this paper we exploit the Multifractal Detrended Fluctuation Analysis (MFDFA) \cite{kantelhardt2002multifractal,PhysRevE.74.016103,bashan2008comparison}, a generalization of the Detrended Fluctuation Analysis (DFA) \cite{peng1995quantification}, to study time series obtained from complex networks via stationary unbiased random walks (RW).
The MFDFA builds upon a generalization of the so-called Hurst exponent as a detector of long-range correlations \cite{serinaldi2010use,barunik2010hurst}.
At the basis of Hurst exponent is the idea of characterizing time series in terms of their degree of \textit{persistence}: roughly speaking, a series is long-range correlated (persistent) if the underlying process has memory of the past states, a property that is firstly noticeable as a heavy-tail in the corresponding autocorrelation function.
Brownian motion corresponds to Hurst exponent equal to 0.5 and it is considered as the baseline uncorrelated process.
Series with Hurst exponent greater than 0.5 are considered as persistent; series with Hurst exponent smaller than 0.5 are anti-persistent (consecutive values tend to be very different).
Additionally, if the value of this exponent does not vary significantly with the magnitude of fluctuations, then the time series is considered monofractal and it can be consistently analysed via DFA; in the opposite case, it is multifractal and the MFDFA is a more suitable choice.
If the studied time series corresponds to a sequence of discrete observables attached to the vertices of a network and the ordering is determined by the subsequent encounters of a random walker exploring the graph, then its persistence / antipersistence property can be translated into the assortative / disassortative character of the graph with respect to said observables.
An assortative graph \cite{newman2002assortative} is a graph in which vertices with similar properties (typically the degree is used, but in theory any property of the vertex can be taken into account) tend to be in contact more frequently than what expected by chance, while a disassortative graph has the opposite feature.
Studying a complex network by the action of a random walker producing a collection of time series of encounters with vertices has an advantage with respect to the simple computation of the static assortative indexes of the graph.
Indeed, the walker trajectories offer also a sampling of the paths distribution in the graph. This distribution is affected by the whole set of mutual relations of vertices at different scales, which are not fully appreciable by a single static snapshot of the network by means of classical network invariants.
In the same manner, we are able to gain an insight on the different scaling of the autocorrelation function and hence on the distribution of the corresponding observable across different locations and scales of the network.
In this study, we primarily focus on the protein contact networks (PCN) elaborated from the E. coli \cite{ecoli_graph__arxiv}. The networks consist of amino acid residues put in contact according to the 3D protein structure \cite{doi:10.1021/cr3002356}. We compare the properties of PCN with those of different known network and time series models.
In particular, we bias the study on their analogies and differences with a model of synthetic protein contact networks (PCN-S) as theorysed by \citet{bartoli2007effect}; PCN-S consist in coiled coils of polymers in which the probability of contact is a decreasing function of the distance between residues along the chain.
From our study, PCN clearly emerge as a particular class of networks whose time series exhibit properties in between mono and multi fractal signals.
The synthetic (simulated coiled coils) proteins, i.e., PCN-S, preserve sufficiently well the multifractality (thus the presence of different scaling relations at different scales or positions), while they seem to lose much of the strong persistence typical of PCN time series. Comparing the PCN with their synthetic counterparts allowed us to highlight a different level of order of the former not shown in the latter due to the pure synthetic polymer folding in space (i.e. no excluded volume contraints). Indeed, the additional freedom given by pure folding causes the loss of many peculiar properties of proteins, like allosteric effect (i.e. the global configurational change of the system upon ligand binding at a specific site) and rapid folding \cite{hilser2012structural,tsai2009protein}. Notably, the particular wiring of protein contact networks encompasses two different architectures fused in the same graph: a short-range contact pattern linking aminoacid residues within the same module, and a long-range contact pattern linking aminoacid residues very distant on the chain structure and pertaining to different modules. The long-range contacts are much less in number than short-range ones but allow for the rapid spreading of information across the structure in order to make the allosteric effect and rapid folding possible.
We successively focus on the embedding of multifractal spectra (MFS) domain and codomain in a vector space to offer further insights on the sensitivity of the coefficients describing the MFS with respect to fluctuation magnitudes (i.e., the small and large fluctuations of the time series, expressed by the $q$th-order moments of the variances).
The discriminating capabilities of different orders of fluctuations of the time series are highlighted by a principal component analysis (PCA) of the embedded MFS, which reduces the initial high-dimensional MFS profiles to four principal components (PC).
Interestingly, the result is that short and medium range observables (respectively, vertex degree and clustering coefficient) are influenced by the network class mainly in their large fluctuations, while the closeness centrality is sensitive to the topology at all fluctuation scales, given its global nature.
PCA of the MFS is hence proposed as a convenient post-processing method when the qualitative interpretation of the results is a primary objective in a multifractal analysis.
The remainder of this paper is structured as follows.
In Section \ref{sec:mfdfa} we provide the necessary technical background on the MFDFA.
In Section \ref{sec:data} we describe the data that we considered in this study.
Section \ref{sec:results} presents the results, which are organized in three subsections.
In Section \ref{sec:persistence} we study the persistence properties of the time series; Section \ref{sec:multifractal} offers a study on the MFS meant to provide a first insight about the multifractality of the time series; finally in Section \ref{sec:embedding} we discuss and interpret the statistical properties of the embedding space derived by means of PCA of the MFS.
Section \ref{sec:conclusions} provides the conclusions and offers some future directions.
\section{Multi-Fractal Detrended Fluctuation Analysis}
\label{sec:mfdfa}
It is known that many processes in Nature and society present long-term memory, manifested in primis as heavy tails in the autocorrelation function of the considered observables. This phenomenon, referred to as the \textit{persistence} of a process, can be characterized by the value of the Hurst exponent $H$, introduced in 1951 by the British hydrologist Harold Edwin Hurst \cite{hurst1951long}. The exponent normally assumes values in the range $[0,1]$ and is traditionally calculated with the R/S analysis, as shown in \cite{serinaldi2010use}. When the process corresponds to uncorrelated noise (e.g. Brownian motion) then the value of $H$ is 0.5, whereas if the process is persistent (correlated) or antipersistent (anticorrelated) it will be respectively greater than and less than 0.5. However, conventional methods employed to analyze the long-range correlation properties of a time series (e.g., spectral analysis, Hurst analysis \cite{serinaldi2010use,barunik2010hurst}) reveal to be
misleading when said time series is non-stationary.
In fact, in many cases it is important to distinguish fluctuations caused by trending behaviors of data at all time scales -- which in this context can be regarded as noise -- from the intrinsic fluctuations characterizing the dynamical process generating the time series.
One of the methods usually employed for this purpose is the Detrended Fluctuation Analysis (DFA), which has shown to be successful in a broad range of situations \cite{peng1995quantification}.
The DFA has been generalized in the so-called Multifractal Detrended Fluctuation Analysis (MFDFA) \cite{kantelhardt2002multifractal,PhysRevE.74.016103,bashan2008comparison,PhysRevLett.62.1327}, which accounts for multifractal scaling, that is, different correlation behaviors on different portions of data, which are thus identified by different sets of scaling exponents.
Among the many applications of MFDFA, it is possible to cite the analysis of human EEG \cite{zorick2013multifractal}, solar magnetograms \cite{makarenko2012multifractal}, human behavioral response \cite{ihlen2013multifractal}, hippocampus signals \cite{fetterhoff2014multifractal}, seismic series \cite{telesca2006measuring,telesca2005multifractal}, medical imaging \cite{lopes2009fractal}, financial markets \cite{schmitt1999multifractal,barunik2012understanding}, and written texts \cite{PhysRevE.86.031108}.
The MFDFA procedure is described thoroughly in \cite{kantelhardt2002multifractal} and it is reported briefly in the following. The method can be summarized in five steps, three of which are identical to the DFA version.
Given a time series $x_k$ of length $N$ with compact support, the MFDFA steps are:
\begin{itemize}
\item{\textit{Step 1} : Compute $Y(i)$ as the cumulative sum (profile) of the series $x_k$:
\begin{equation}
Y(i) \equiv \sum_{k=1}^i \left[ x_k - \langle x \rangle \right], \;\; i = 1, \dotsc, N.
\end{equation}
}
\item{\textit{Step 2} :
Divide $Y(i)$ in $N_s \equiv \text{int}(N/s)$ non-overlapping segments of equal
length $s$. Since the series length $N$ may not be a multiple of $s$, the last segment is likely to be shorter, so this operation is repeated in reverse order by starting from the opposite end of the series, thus obtaining a total of $2N_s$ segments.}
\item{\textit{Step 3} : Execute the local detrending operation by a suitable polynomial fitting on
each of the $2N_s$ segments. Then determine the variance,
\begin{equation}
F^2(\nu,s) \equiv \frac{1}{s} \sum_{i=1}^s \bigg\{ Y[(\nu-1)s+i] - y_\nu(i) \bigg\}^2,
\end{equation}
for each segment $\nu = 1,\dotsc,N_s$ and
\begin{equation}
F^2(\nu,s) \equiv \sum_{i=1}^s \bigg\{ Y[N-(\nu - N_s)s+i] - y_\nu(i)\bigg\}^2
\end{equation}
for $\nu = N_s +1,\dotsc, 2N_s$, where $y_\nu(i)$ is the fitted polynomial in segment $\nu$. The order $m$ of the fitting polynomial, $y_\nu(i)$, determines the capability of the (MF-)DFA in eliminating trends in the series, thus it has to be tuned according to the expected maximum trending order of the time series.
}
\item{\textit{Step 4} : Compute the $q$th-order average of the variance over all segments,
\begin{equation}
\label{eq:Fq}
F_q(s) \equiv \bigg\{ \frac{1}{2N_s} \sum_{\nu=1}^{2N_s} \left[ F^2(\nu,s)\right]^{q/2} \bigg\}^{1/q},
\end{equation}
with $q \in \mathbb{R}$. The $q$-dependence of the fluctuations function $F_q(s)$ serves the purpose of highlighting the contribute of fluctuations at different magnitude orders.
For $q > 0$ only the larger fluctuations contribute mostly to the average in Eq.~\ref{eq:Fq}; conversely, for $q < 0$ the magnitude of the smaller fluctuations is enhanced. For $q = 2$ the standard DFA procedure is obtained. The case $q = 0$ cannot be computed with the averaging form in Eq.~\ref{eq:Fq} and so a logarithmic form has to be employed,
\begin{equation}
F_0(s) = \exp \bigg\{ \frac{1}{2N_s} \sum_{\nu=1}^{2N_s} \ln \left[F^2(\nu,s)\right] \bigg\}.
\end{equation}
The steps 2 to 4 have to be repeated for different time scales $s$, where all values of $s$ have to be chosen such that $s \geq m+2$ to allow for a meaningful fitting of data. It is also convenient to avoid scales $s > N/4$ because of the statistical unreliability of such small numbers $N_s$ of segments considered.
}
\item{\textit{Step 5} : Determine the scaling behaviour of the fluctuation functions by analyzing log-log plots of $F_q(s)$ versus $s$ for each value of $q$. If the series $x_i$ is long-range power-law correlated, $F_q(s)$ is approximated (for large values of $s$) by the form
\begin{equation}
\label{eq:Fqshq}
F_q(s) \sim s^{h(q)}.
\end{equation}}
\end{itemize}
The exponent $h(q)$ is the generalized Hurst exponent; for $q=2$ and stationary time series, $h(q)$ reduces to the standard Hurst exponent, $H$. When the considered time series is monofractal, i.e., it shows a uniform scaling over all magnitude scales of the fluctuations, $h(q)$ is independent of $q$. On the contrary, when small fluctuations are different with respect to the large ones, the dependency of $h(q)$ on $q$ becomes apparent and the series can be considered multifractal.
Starting from Eq.~\ref{eq:Fq} and using Eq.~\ref{eq:Fqshq}, it is straightforward to obtain
\begin{equation}
\sum_{\nu=1}^{N/s} [ F(\nu,s)]^q \sim s^{qh(q) - 1},
\end{equation}
where, for simplicity, it has been assumed that the length $N$ of the series is a multiple of the scale $s$, such that $N_s = N/s$.
The exponent
\begin{equation}
\label{eq:tauq}
\tau(q) = qh(q) - 1
\end{equation}
corresponds to the MFDFA generalization of the multifractal mass exponent. In case of positive stationary and normalized time series, $\tau(q)$ corresponds to the scaling exponent of the $q$-order partition function $Z_q(s)$.
Another function that characterizes the multifractality of a series is the singularity spectrum, $D(\alpha)$, which is obtained via the Legendre transform of $\tau(q)$,
\begin{equation}
\label{eq:mutifractal_spectrum}
D(\alpha) = q\alpha - \tau(q),
\end{equation}
where $\alpha$ is equal to the derivative $\tau'(q)$ and corresponds to the H\"older exponent (also called singularity exponent). Using Eq.~\ref{eq:tauq} it is possible to directly relate $\alpha$ and $D(\alpha)$ to $h(q)$, obtaining:
\begin{equation}
\alpha = h(q) + qh'(q) \;\; \text{and} \;\; D(\alpha) = q[\alpha - h(q)] + 1.
\end{equation}
The singularity spectrum in Eq. \ref{eq:mutifractal_spectrum} -- also called multifractal spectrum (MFS) -- allows to infer important information regarding the ``degree of multifractality'' and the specific sensitivity of the time series to fluctuations of different magnitudes.
In fact, the width of the support of $D(\cdot)$ is an important quantitative indicator of the multifractal character of the series (the larger, the more fractal a series is).
Also the codomain of $D(\cdot)$ encodes useful information, since it corresponds to the dimension of the subset of the times series domain which is characterized by the singularity exponent $\alpha$.
\section{The Considered Data}
\label{sec:data}
The biological networks analysed in this work are partially linked to those analysed in our previous works \cite{mixbionets1__arxiv,ecoli_graph_complexity_arxiv}. We consider 400 E. coli protein contact networks (PCN) as the main object of study and we compare them to several models.
To this end, we generated 400 synthetic polymers (PCN-S) by employing the generation method presented by \citet{bartoli2007effect}, and by setting appropriate parameters in order to resemble the basic properties of each of the above PCN (i.e., the graph size).
More precisely, each E. coli protein {\tt JWxxxx} is juxtaposed with its synthetic counterpart, {\tt JWxxxx\,\_\,SYNTH}, having equal number of vertices and edges -- the four-digit number {\tt xxxx} stands for its unique identifier.
In addition, we consider 10 Erd\H{o}s-R\'{e}nyi networks (ER) and 10 scale-free networks generated using the Barab\'{a}si-Albert (BA) model, varying the number of vertices between 300 and 1200. The former are generated setting $p=\log(n)/n$ while for the latter we used a six-degree attachment scheme.
To allow the processing of such networks via the MFDFA procedure, we generate time series by means of stationary unbiased RWs, where at each step an observable is measured from the current vertex. Considering the size of the networks at hand, the RW length has been fixed at $10^5$ time instants; this length assures the coverage of all vertices for a statistically significant number of times and it is consistent with the recommendations in Ref. \cite{nicosia2013characteristic}.
We associate to each network three time series generated within the same RW. The first series considers vertex degree (VD) as observable; the second one the vertex clustering coefficient (VCL); the third one the vertex closeness centrality (VCL).
Those three observables account for, respectively, the short, medium, and long range information of the network from the point of view of a vertex.
The dataset is also composed by six classes of time series that act as probes, which are obtained directly from their generative models.
The herein considered time series are obtained from three fractional Brownian motion (FBM) processes with increasing Hurst coefficients, and three multifractal binomial cascades (BC), characterized by increasing MFS widths. FBMs have coefficients $H = 0.25,0.5,0.75$ and represent the poles of monofractality with increasing persistence. For each fixed value of $H$, we generated ten different time series (for a total of 30 FBMs) to account for the statistical variability.
On the other hand, BCs are deterministic multiplicative processes, which are generated with the partition coefficient $a = 0.6,0.7,0.8$. These series are inherently multifractal, although they possess different persistence levels. Notice that in this case there is no point in generating more than one instance of the BC processes for each value of $a$, since the process is deterministic; so only three BC time series are generated.
\section{Experimental Results}
\label{sec:results}
In this section we discuss the experimental results obtained by analyzing all considered time series. The results are organized in three main sections. In Sec. \ref{sec:persistence} we show the analysis of the persistence properties of the time series; Sec. \ref{sec:multifractal} provides the setting of the procedure employed to obtain the MFS coefficients, offering also an initial discussion on the multifractal properties of the considered time series; finally in Sec. \ref{sec:embedding} we discuss the embedding of the MFS in a suitable vector space.
\subsection{Analysis of persistence properties}
\label{sec:persistence}
The first property that we analyze is the Hurst coefficient that, as described above, quantifies the persistence of the time series.
In Figs.~\ref{fig:persistence_deg}, \ref{fig:persistence_cl}, and \ref{fig:persistence_cc} are shown the values of $H$ measured on each time series of the PCN, PCN-S, BA, and ER, for each of the three observables VD, VCL, and VCC, respectively, along with the Hurst exponents proper of the three classes of FBMs. Notice that, since FBM time series are not obtained as different observables yielded by a RW on a network, their Hurst exponents have been just replicated across the three figures.
\begin{figure}[ht!]
\centering
\subfigure[Time series of VD.]{
\includegraphics[viewport=70 0 340 190,scale=0.55,keepaspectratio=true]{./persistence_deg}
\label{fig:persistence_deg}}
~
\subfigure[Time series of VCL.]{
\includegraphics[viewport=20 0 300 190,scale=0.55,keepaspectratio=true]{./persistence_clust}
\label{fig:persistence_cl}}
\subfigure[Time series of VCC.]{
\includegraphics[viewport=75 0 340 240,scale=0.55,keepaspectratio=true]{./persistence_close}
\label{fig:persistence_cc}}\hspace{0.1cm}
~
\subfigure[Autocorr. for ``JW0058''.]{
\includegraphics[viewport=20 0 300 190,scale=0.53,keepaspectratio=true]{./autocorrelation_log-lin}
\label{fig:autocorrelation}}
\subfigure[Time series for ``JW0058''.]{
\includegraphics[viewport=0 0 341 177,scale=0.50,keepaspectratio=true]{./timeseries_JW0058}
\label{fig:tsjw0058}}
~
\subfigure[Time series for ``JW0058\_SYNTH''.]{
\includegraphics[viewport=0 0 341 177,scale=0.50,keepaspectratio=true]{./timeseries_JW0058_SYNTH}
\label{fig:tsjw0058s}}
\caption{Persistence of the series measured through the Hurst exponent for VD \subref{fig:persistence_deg}, VCL \subref{fig:persistence_cl}, and VCC \subref{fig:persistence_cc}. The PCN (red bands) show significant persistence for all the three observables. \subref{fig:autocorrelation} Sample autocorrelation function for the protein JW0058 and the corresponding synthetic polymer JW0058\_SYNTH. \subref{fig:tsjw0058} and \subref{fig:tsjw0058s} Sample time series. The higher persistence of the natural protein with respect to the synthetic analogue is particularly evident in the VCC series.}
\label{fig:persistence_level}
\end{figure}
As expected, BA and ER networks produce RWs consistent with an uncorrelated Brownian motion (i.e., $H=0.5$), since basically they are the result of an uncorrelated degree distribution.
Interestingly, from the persistence levels shown in Fig.~\ref{fig:persistence_level}, it is possible to observe that PCN (red bands) induce time series with strong persistence, regardless of the particular observable. It is also evident from Fig.~\ref{fig:persistence_level} that also synthetic polymers (green bands), similarly to the PCN, show positively correlated behaviours, even if they do not seem to capture this characteristic persistence in a satisfying manner. It is also important to mention that, when plotting the Hurst exponents of PCN-S as a function of Hurst exponents of their corresponding PCN (data not shown for brevity), no trending has been observed. Indeed, these two quantities are not proportional and thus PCN-S instances cannot be considered just as less-persistent versions of their corresponding PCN.
These results can be exploited to gain a more insightful view on the intrinsic characteristics of the PCN class, by relating the properties of the RWs to the topological properties of the corresponding graphs. In particular, the time series of VD show positive correlations, which in turn imply degree assortativity. This result is in agreement with the claims of \citet{bode2007network} and references therein, although we reached the same result by exploiting a different technique -- usually the degree assortativity is investigated through the method proposed by \citet{newman2002assortative}.
It is worth pointing out that, since PCN are known to be fractal networks (embedded into a three-dimensional space) \cite{granek2011proteins,ecoli_graph_complexity_arxiv}, the herein observed degree assortativity is not in agreement with the theoretical hypothesis of \citet{song2006origins}, which requires the degree distribution of fractal networks to be disassortative.
The high persistence of the clustering coefficient observed in the PCN is slightly more tricky to interpret in terms of topological properties.
Roughly speaking, the VCL of a vertex is proportional to the local connectivity of the subgraph formed by the vertex and its closest neighbors with respect to the whole graph.
It is known that PCN show a high degree of global modularity (see \cite{mixbionets1__arxiv,doi:10.1021/cr3002356}). Therefore, the persistence of the clustering coefficient can be interpreted as the tendency of vertices in the same module to be connected rather uniformly with the presence of medium-to-small hubs -- PCN do not have large hubs \cite{bode2007network,doi:10.1021/cr3002356}.
Another way to explain this property is to directly relate VCL to the persistence of VD time series.
To this end, Fig.~\ref{fig:degavgcl} shows the relation of the degree-dependent average VCL over the possible VD. Here we considered the whole PCN and PCN-S ensembles. The error bars (which are usually smaller than the marker) represent the standard deviation over the entire ensemble.
As it is possible to observe, while the two trends are substantially different, the standard deviation is very small in both cases for most values of the degree. This fact, along with the aforementioned degree assortativity, suggests a possible explanation for the persistence displayed by the VCL series of both PCN and PCN-S.
It is worth noting that, for PCN, the clustering coefficient remains high when increasing the degree, which can be interpreted as a sign of high global modularity.
\begin{figure}[ht!]
\centering
\subfigure[VCL vs VD.]{
\includegraphics[viewport=0 0 347 248,scale=0.60,keepaspectratio=true]{./degavgCL}
\label{fig:degavgcl}}
~
\subfigure[VCC vs VD.]{
\includegraphics[viewport=0 0 347 248,scale=0.60,keepaspectratio=true]{./degavgCC}
\label{fig:degavgcc}}
\caption{Ensemble average VCL \subref{fig:degavgcl} and VCC \subref{fig:degavgcc} as a function of the degree for all proteins (red points) and synthetic counterpart (green points). Relation with respect to the degree shows significant difference among the real proteins and the synthetic polymers.}
\label{fig:avg_vs_deg}
\end{figure}
While VD and VCL show similar characteristics, the behaviour of the VCC observable is considerably different. By looking at the plot in Fig.~\ref{fig:persistence_cc}, it is possible to observe that the Hurst coefficients of PCN are comparable and occasionally greater than one -- i.e., the corresponding time series are non-stationary. This might be considered as a symptom of different distributions of typical paths within the PCN.
This conjecture would be in line with the observation of \citet{yan2014construction}, where it is hypothesized that there are two characteristic distributions of paths within PCN, intra-module and inter-module paths, which is also a consequence of the PCN's high degree of modularity mentioned before.
On the other hand, PCN-S do not share this feature, confirming an intrinsically different configuration of the network topology at a global scale.
As for the VCL observable, the VCC persistence can be related to the VD persistence by inspecting Fig.~\ref{fig:degavgcc}.
By comparing the two trends, it can also be observed that the PCN-S have a broader distribution over the possible degree values, as a consequence of being small-world networks \cite{bartoli2007effect}, while PCN are neither small-world nor scale-free \cite{mixbionets1__arxiv,doi:10.1021/cr3002356}.
In Fig.~\ref{fig:autocorrelation} it is shown the autocorrelation function of the three time series for one randomly chosen protein and its synthetic counterpart.
First, it is worth noting that long-range correlations appear here in the form of heavy tails.
Additionally, the VCC autocorrelation function denotes a much heavier tail with respect to the other observables, which is justified by the higher persistence (see Fig. \ref{fig:persistence_cc}).
To conclude, in Fig.~\ref{fig:tsjw0058} and \ref{fig:tsjw0058s} are shown two excerpts of the time series generated by the same protein and its synthetic model for each observable. By visually comparing the two plots, in particular for the observables VCL and VCC, it is possible to note that the two networks generate RWs that are significantly different; the higher persistency of the PCN observables is also visually recognizable.
From these results it is clear that the PCN-S network models present significant discrepancies from their real counterparts, while still being distinguishable from other network models. These differences will be further analyzed in the following subsections.
\subsection{Analysis of multifractal properties}
\label{sec:multifractal}
After having calculated the persistence properties of the considered time series, we can now proceed to evaluate their degree of multifractality.
For each of the time series presented in Sec. \ref{sec:data}, we perform the MFDFA procedure exposed in Sec. \ref{sec:mfdfa} by executing the Matlab$^\circledR$ routine {\tt MFDFA1()}, written by Ihlen and described in detail in Ref. \cite{ihlen2012introduction}.
The input of the routine is the time series to analyze, a vector of the considered time scales (corresponding to the set of increasing length scales $s$ described in Sec. \ref{sec:data}), the range of $q$-orders to be considered for the analysis, and finally the polynomial order, $m$, for the detrending.
For the analysis of all time series, we used the following setting:
\begin{itemize}
\item{the time scales $s \in \{16,32,64,128,256,512,1024\}$;}
\item{the orders $q \in \{ -5, -4.8, -4.6, \dotsc, +4.8, +5\}$ for a total of 51 values;}
\item{the detrending order $m = 2$.}
\end{itemize}
The output produced by the routine, for all values of $q$, is the collection of (generalized) Hurst coefficients $H(q)$, mass exponents $\tau(q)$, singularity exponents $\alpha(q)$, dimension coefficients $D(\alpha(q))$, and scaling function $F(q)$.
Please note that since $D(\alpha(q))$ is returned by the procedure directly as a function of $q$, in the following we will denote $D(\alpha(q))$ as $D(q)$.
The width of the MFS is the extent of the $D(\cdot)$ support, which characterizes the degree of multifractality of a series.
Clearly, all these quantities are not independent with each other and thus, in order to reduce redundancies, we only considered the subset consisting of $\alpha(q)$ and $D(q)$ in the embedding discussed later in Sec. \ref{sec:embedding}.
In fact, as said before, the MFS, $D(\cdot)$, encodes all information regarding the multifractality of the time series.
Notice that all the networks are described simultaneously by three time series, corresponding to the three observables VD, VCL, and VCC, while the probe time series are expressed by the same realization of the process for all the three observables.
To gain a first insight on the multifractality of the considered time series, it is useful to relate this property to the persistence levels calculated in Sec. \ref{sec:persistence}.
In particular, we perform this analysis for the PCN and the PCN-S since they exhibit the highest values of $H$; we consider here also the six probes, i.e., the three FBMs and three deterministic BC.
Fig.~\ref{fig:spectra_deg}, \ref{fig:spectra_cl}, and \ref{fig:spectra_cc} show the plots of $H$ versus the width of the MFS, respectively for the observables VD, VCL, and VCC.
\begin{figure}[ht!]
\centering
\subfigure[Time series of VD.]{
\includegraphics[viewport=0 0 347 248,scale=0.52,keepaspectratio=true]{./spectra_deg_log-lin}
\label{fig:spectra_deg}}
~
\subfigure[Time series of VCL.]{
\includegraphics[viewport=0 0 347 248,scale=0.52,keepaspectratio=true]{./spectra_clust_log-lin}
\label{fig:spectra_cl}}
\subfigure[Time series of VCC.]{
\includegraphics[viewport=0 0 347 248,scale=0.52,keepaspectratio=true]{./spectra_close_log-lin}
\label{fig:spectra_cc}}
\caption{Hurst exponent vs MFS width of PCN and PCN-S. Both PCN and PCN-S show characteristics in-between mono and multi- fractal signals. Notice that the plot scale is log-lin for sake of clarity.}
\label{fig:mfs_width}
\end{figure}
By comparing in Fig. \ref{fig:mfs_width} the relative distances between the PCN points and the probes, it is possible to observe that most PCN exhibit MFS widths that could be considered in-between those of mono and multi-fractal signals.
Some proteins also show extremely wide MFS, while keeping the Hurst coefficient unaltered. Interestingly, PCN-S, while being less persistent, have a similar distribution of MFS widths. As observed for the persistence analysis, the VD and VCL observables behave very similarly also in terms of multifractality, while the VCC data points are more clustered and present slightly narrower spectra.
Once again, this can be attributed to the substantial difference between the types of observables.
Indeed, VD and VCL are short/medium range observables, so they can be influenced by the vertex position within the network at many distance scales. Instead, the VCC is mainly influenced by large scales (being a global topological descriptor), hence explaining why it shows less multifractal behaviour.
The variety of MFS herein observed justifies the experiments performed in the next section, which are focused on the analysis of the MFS projected in a suitable PCA space.
\subsection{Embedding of the multifractal spectra}
\label{sec:embedding}
As mentioned above, the MFS elaborated from the time series constitutes the principal hallmark of all multifractal features.
However, as first observed in Sec. \ref{sec:multifractal}, the spectra widths vary significantly even between members of the same class. Hence, there is no element that can be accounted for a meaningful representative of the whole class of proteins.
For this reason, we embed all considered MFS coefficients, i.e., $D(q)$ and $\alpha(q)$, in a suitable low-dimensional vector space derived by means of a PCA.
With the embedding into a PCA space, we are enabled to study the ensemble properties of each class without focusing on single elements alone, hence gaining an insight on the features that mostly characterize the particular typology of networks.
In such embedding, each time series is initially represented by a vector $v \in \mathbb{R}^n$. Here, $n = 300$ is the total number of coefficients retrieved by the MFDFA that we consider, which is composed by 50 values of $D(q)$ plus 50 values of $\alpha(q)$, for each of the three observables.
A given network $\mathcal{G}$, associated with time series $x^\text{VD}_\mathcal{G}(t), \, x^\text{VCL}_\mathcal{G}(t)$, and $x^\text{VCC}_\mathcal{G}(t)$, is thus represented by a vector $\vec{v}_\mathcal{G} \in \mathbb{R}^{300}$ with the form
\begin{align}
\label{eq:mfdfavec}
\begin{split}
\vec{v}_\mathcal{G}=
& \Big[ D_\text{VD}(-5),\,\dotsc\,, D_\text{VD}(+5),\; \alpha_\text{VD}(-5), \,\dotsc\,, \alpha_\text{VD}(+5), \Big. \\ & \;\;\;\;\; \drsh \; D_\text{VCL}(-5),\,\dotsc\,, D_\text{VCL}(+5),\; \alpha_\text{VCL}(-5), \,\dotsc\,, \alpha_\text{VCL}(+5), \\ & \;\;\;\;\; \drsh \;\Big. D_\text{VCC}(-5),\,\dotsc\,, D_\text{VCC}(+5),\; \alpha_\text{VCC}(-5), \,\dotsc\,, \alpha_\text{VCC}(+5)\Big]^\top, \scriptstyle
\end{split}
\end{align}
where $D_\mathcal{O}(q)$ and $\alpha_\mathcal{O}(q)$, with $\mathcal{O} \in \{ \text{VD, VCL, VCC} \}$, are respectively the dimension coefficient and the singularity coefficient associated to the time series $x^\mathcal{O}_\mathcal{G}(t)$ as a function of the order parameter $q$.
We stress that in our analysis $q$ assumes 51 equally-spaced values between -5 and 5, with a step size of 0.2; however, we do not consider the $q=0$ case since it yields trivial values for the MFS.
On the other hand, the probe time series (FBM and BC) are not derived from a network. To be consistent with the aforementioned vector representation, their MFDFA coefficients are simply replicated 3 times, giving a vector of the form:
\begin{align}
\label{eq:mfdfa_probe}
\begin{split}
\vec{v}_\text{probe}=
& \Big[ D(-5),\,\dotsc\,, D(+5),\; \alpha(-5), \,\dotsc\,, \alpha(+5), \Big. \\ & \;\;\;\;\; \drsh \; D(-5),\,\dotsc\,, D(+5),\; \alpha(-5), \,\dotsc\,, \alpha(+5), \\ & \;\;\;\;\; \drsh \;\Big. D(-5),\,\dotsc\,, D(+5),\; \alpha(-5), \,\dotsc\,, \alpha(+5)\Big]^\top. \scriptstyle
\end{split}
\end{align}
The 300-dimensional vector space described above is obviously unmanageable from the point of view of interpretation and, of course, visualization. For this reason, we perform a PCA to obtain a more synthetic description of the data.
The PCA does not allow only to reduce the dimensionality of the data, but it also allows to give a reasonable and more direct interpretation of the new reference framework, i.e., the PCs. This is the main reason why we opted for PCA instead of a more sophisticated, non-linear, dimensionality reduction technique. Notice also that the process has been operated on the standardized data (z-scores), which corresponds to the correlation-based PCA, instead of the covariance-based version.
As shown in Tab.~\ref{tbl:explvariance}, the first four PCs explain more than the 83\% of the entire variance.
For this reason, we will move our analysis to the considerably simpler four-dimensional space spanned by the first four PCs, which are shown in the plots of Fig.~\ref{fig:PCA_MFDFA}; we consider the two-dimensional subspaces derived by PC1--PC2 (\ref{fig:mfdfa_pc1-2}) and P3--PC4 (\ref{fig:mfdfa_pc3-4}), respectively.
\begin{table}
\caption{Explained variance of the first five PCs.}
\label{tbl:explvariance}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
& \textbf{PC1} & \textbf{PC2} & \textbf{PC3} & \textbf{PC4} & \textbf{PC5} \\
\hline
\textbf{Explained Variance(\%)} & 31.12 & 29.46 & 16.58 & 6.71 & 4.94 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[ht!]
\centering
\subfigure[PC1--PC2]{
\includegraphics[viewport=0 0 347 246,scale=0.60,keepaspectratio=true]{./MFDFA_PCA1-2}
\label{fig:mfdfa_pc1-2}}
~
\subfigure[PC3--PC4]{
\includegraphics[viewport=0 0 347 246,scale=0.60,keepaspectratio=true]{./MFDFA_PCA3-4}
\label{fig:mfdfa_pc3-4}}
\caption{PCA of the MFS extracted from all considered time series.}
\label{fig:PCA_MFDFA}
\end{figure}
To understand the meaning of the PCs just retrieved, we analyze their loadings. In Fig.~\ref{fig:factors} are shown the correlations of each original variable with the first four PCs, where the original variables are ordered as described in Eqs. \ref{eq:mfdfavec} and \ref{eq:mfdfa_probe}.
As it is possible to note in Tab.~\ref{tbl:explvariance}, the first two principal components, PC1 and PC2, are nearly equivalent in terms of explained variance and they are correlated, respectively, to the singularity exponent $\alpha(q)$ and spectrum $D(q)$.
In particular, they are both strictly related to the large fluctuations (positive $q$ orders) of the observables VD and VCL, and to almost all fluctuation orders of VCC -- see Fig. \ref{fig:fact1-2}.
Once again, there is a clear separation between the characteristics of short and medium range observables, VD and VCL, and the long-range observable VCC. In fact, the discriminating power of VD and VCL is limited to the structure of their larger fluctuations, which is due to their local nature (we stress that local here refers to the neighborhood extent of the corresponding vertex).
Arguably, the related small fluctuations behave just as a ``background noise'', providing little information on the relevant global properties of the networks.
On the other hand, the organization of large fluctuations of VD and VCL in a RW indicates the occurrence and distribution of significant events, i.e., those related to the global topology of the network, like for example jumps between modules or areas with different local topology, hub encounters, etc.
By following this interpretation, large fluctuations provide information that appears to play an important role in the discrimination of the network's class. The VCC, instead, is fundamentally different. In this case, as mentioned at the end of Sec. \ref{sec:multifractal}, the observable is much more sensitive since it is affected by the network topology at the largest distance scales. Hence, its variations are globally discriminating at all fluctuation orders.
The loadings of the third and fourth PC shown in Fig. \ref{fig:fact3-4} are easier to interpret. In fact, the first thing that is worth noting is the complete absence of influence of the VCC observable -- since it is almost completely loaded in the first two PCs. This first observation reconfirms that all fluctuation orders of VCC provide important information in terms of variance.
Interestingly, PC3 and PC4 are suitably allocated on the VCL and VD observables. At odds with what we have observed in Fig. \ref{fig:fact1-2}, PC3 and PC4 are characteristic of the small fluctuations only (of both $D(q)$ and $\alpha(q)$).
However, PC3 seems to be heavily influenced by the probe networks variance; Fig. \ref{fig:mfdfa_pc3-4} offers a visual understanding of this claim. Therefore, its contribution in the discrimination among the different network topologies is questionable, while PC4 provides a small but yet perceptible contribution in terms of variance.
\begin{figure}[ht!]
\centering
\subfigure[Loadings of PC1 and PC2]{
\includegraphics[viewport=0 0 350 246,scale=0.60,keepaspectratio=true]{./factors_f1-f2}
\label{fig:fact1-2}}
\subfigure[Loadings of PC3 and PC4]{
\includegraphics[viewport=0 0 350 246,scale=0.60,keepaspectratio=true]{./factors_f3-f4}
\label{fig:fact3-4}}
\caption{PC loadings. The three observables forming the overall MFS are differentiated by using diverse shaded colors.}
\label{fig:factors}
\end{figure}
\section{Conclusions}
\label{sec:conclusions}
In this paper, we have exploited the possibility to characterize protein contact networks by means of a multifractal analysis of suitable time series. Such time series have been generated by performing stationary, unbiased, random walks on the graph structures, recording at each vertex three different quantities: the degree, clustering coefficient, and closeness centrality of the vertex.
Those three observables capture, respectively, short, medium, and long range peculiarities of the considered networks.
Our analysis of the considered protein contact networks was compared with several probe data. Notably, we used a receipt to generate synthetic polymers designed to mimic random coiled cords, two well-known classes of random networks, and two models of time series embodying the archetypical monofractal and multifractal signals.
The herein presented study provided a number of results. First, persistence analysis of the time series showed that proteins, regardless of the considered vertex observable, generate strongly persistent signals.
When considering the degree as the observable, this can be translated into assortativity of the degree distribution. This first result is confirmed by the recent literature, although, to our knowledge, we are the first to asses such a property by means of time series analysis. We also pointed out that this result is in contrast with the recent hypothesis requiring disassortativity in fractal networks \cite{song2006origins,gallos2007review,zhang2007self}.
We also found that the assortativity of other observables can be linked to the assortativity of the vertex degree, since the degree basically controls the behaviour of the RW and thus influences to some extent all other measurements.
Then we moved to a first analysis inspecting the multifractal footprint proper of the considered time series.
Results showed that time series associated to protein contact networks -- again regardless of the observable -- should be considered as signals in-between the typical mono and multi- fractal behavior.
We further elaborated over those results by performing the interpretation of the entire multifractal spectrum via the embedding into a suitable vector space. Such a vector space has been derived by first associating each time series to a high-dimensional vector containing suitable samplings of the domain and codomain of the multifractal spectrum derived by the multifractal dentrended fluctuation analysis. Successively, we performed a principal component analysis, resulting in a four-dimensional vector space explaining large part of the original data variance.
The principal component analysis allowed us to perform a detailed interpretation regarding the importance of different fluctuation orders by analysing their loadings on the principal components.
Results showed that large (in magnitude) fluctuations of all observables are more important in terms of discrimination (variance) of the considered networks/time series. Along with these, small fluctuations of the closeness centrality observable were also recognized to be discriminating, fact that has been attributed to their long-range (global) nature. Small fluctuations of the degree and clustering coefficient, instead, are less informative, since they are more easily associated with background noise.
We conclude by arguing that the herein presented study for analyzing complex networks could be used also in different settings. Indeed, the techniques employed in this work never assume the knowledge of the global topology of the graph. In particular, this study might be of interest when the topology of the network under analysis is not directly observable, but can be gradually ``explored'' with suitable time-dependent measurements of the vertices.
Moreover, the comparison of the herein considered proteins with the corresponding synthetic versions highlighted important differences, which in turn strengthen the need to develop a more suitable generative model for protein contact networks. Such a generative model, if able to reproduce the co-existence of fast-lane communication by specific long-range contacts with the strong modularity of protein contact networks, could be of utmost importance in a number of technological applications.
\bibliographystyle{abbrvnat}
|
\section{Proofs of Main Theorems}
\label{sec:proofs}
\subsection{Proof of Lemma~\ref{lem:regret decomposition}}
\label{sec:proof regret decomposition}
Let $R_t = R(A_t, w_t)$ be the stochastic regret of ${\tt CombUCB1}$ at time $t$, where $A_t$ and $w_t$ are the solution and the weights of the items at time $t$, respectively. Furthermore, let $\mathcal{E}_t = \set{\exists e \in E: \abs{\bar{w}(e) - \hat{w}_{T_{t - 1}(e)}(e)} \geq c_{t - 1, T_{t - 1}(e)}}$ be the event that $\bar{w}(e)$ is outside of the high-probability confidence interval around $\hat{w}_{T_{t - 1}(e)}(e)$ for some item $e$ at time $t$; and let $\overline{\cE}_t$ be the complement of $\mathcal{E}_t$, $\bar{w}(e)$ is in the high-probability confidence interval around $\hat{w}_{T_{t - 1}(e)}(e)$ for all $e$ at time $t$. Then we can decompose the regret of ${\tt CombUCB1}$ as:
\begin{align*}
R(n) = \EE{\sum_{t = 1}^{t_0 - 1} R_t} +
\EE{\sum_{t = t_0}^n \I{\mathcal{E}_t} R_t} +
\EE{\sum_{t = t_0}^n \I{\overline{\cE}_t} R_t}\,.
\end{align*}
Now we bound each term in our regret decomposition.
The regret of the initialization, $\EE{\sum_{t = 1}^{t_0 - 1} R_t}$, is bounded by $K L$ because Algorithm~\ref{alg:initialization} terminates in at most $L$ steps, and $R_t \leq K$ for any $A_t$ and $w_t$.
The second term in our regret decomposition, $\EE{\sum_{t = t_0}^n \I{\mathcal{E}_t} R_t}$, is small because all of our confidence intervals hold with high probability. In particular, for any $e$, $s$, and $t$:
\begin{align*}
P(\abs{\bar{w}(e) - \hat{w}_s(e)} \geq c_{t, s}) \leq 2 \exp[-3 \log t]\,,
\end{align*}
and therefore:
\begin{align*}
\EE{\sum_{t = t_0}^n \I{\mathcal{E}_t}} \leq
\sum_{e \in E} \sum_{t = 1}^n \sum_{s = 1}^t P(\abs{\bar{w}(e) - \hat{w}_s(e)} \geq c_{t, s}) \leq
2 \sum_{e \in E} \sum_{t = 1}^n \sum_{s = 1}^t \exp[-3 \log t] \leq
2 \sum_{e \in E} \sum_{t = 1}^n t^{-2} \leq
\frac{\pi^2}{3} L\,.
\end{align*}
Since $R_t \leq K$ for any $A_t$ and $w_t$, $\EE{\sum_{t = t_0}^n \I{\mathcal{E}_t} R_t} \leq \frac{\pi^2}{3} K L$.
Finally, we rewrite the last term in our regret decomposition as:
\begin{align*}
\EE{\sum_{t = t_0}^n \I{\overline{\cE}_t} R_t} \stackrel{\text{(a)}}{=}
\sum_{t = t_0}^n \EE{\I{\overline{\cE}_t} \EE{R_t \,|\, A_t}} \stackrel{\text{(b)}}{=}
\EE{\sum_{t = t_0}^n \Delta_{A_t} \I{\overline{\cE}_t, \Delta_{A_t} > 0}}\,.
\end{align*}
In equality (a), the outer expectation is over the history of the agent up to time $t$, which in turn determines $A_t$ and $\overline{\cE}_t$; and $\EE{R_t \,|\, A_t}$ is the expected regret at time $t$ conditioned on solution $A_t$. Equality (b) follows from $\Delta_{A_t} = \EE{R_t \,|\, A_t}$. Now we bound $\Delta_{A_t} \I{\overline{\cE}_t, \Delta_{A_t} > 0}$ for any suboptimal $A_t$. The bound is derived based on two facts. First, when ${\tt CombUCB1}$ chooses $A_t$, $f(A_t, U_t) \geq f(A^\ast, U_t)$. This further implies that $\sum_{e \in A_t \setminus A^\ast} U_t(e) \geq \sum_{e \in A^\ast \setminus A_t} U_t(e)$. Second, when event $\overline{\cE}_t$ happens, $\abs{\bar{w}(e) - \hat{w}_{T_{t - 1}(e)}(e)} < c_{t - 1, T_{t - 1}(e)}$ for all items $e$. Therefore:
\begin{align*}
\sum_{e \in A_t \setminus A^\ast} \hspace{-0.1in} \bar{w}(e) +
2 \hspace{-0.1in} \sum_{e \in A_t \setminus A^\ast} \hspace{-0.1in} c_{t - 1, T_{t - 1}(e)} \geq
\sum_{e \in A_t \setminus A^\ast} \hspace{-0.1in} U_t(e) \geq
\sum_{e \in A^\ast \setminus A_t} \hspace{-0.1in} U_t(e) \geq
\sum_{e \in A^\ast \setminus A_t} \hspace{-0.1in} \bar{w}(e)\,,
\end{align*}
and $2 \sum_{e \in A_t \setminus A^\ast} c_{t - 1, T_{t - 1}(e)} \geq \Delta_{A_t}$ follows from the observation
that $\Delta_{A_t} = \sum_{e \in A^\ast \setminus A_t} \bar{w}(e) - \sum_{e \in A_t \setminus A^\ast} \bar{w}(e)$. Now note that $c_{n, T_{t - 1}(e)} \geq c_{t - 1, T_{t - 1}(e)}$ for any time $t \leq n$. Therefore, the event $\mathcal{F}_t$ in \eqref{eq:suboptimality event} must happen and:
\begin{align*}
\EE{\sum_{t = t_0}^n \Delta_{A_t} \I{\overline{\cE}_t, \Delta_{A_t} > 0}} \leq
\EE{\sum_{t = t_0}^n \Delta_{A_t} \I{\mathcal{F}_t}}\,.
\end{align*}
This concludes our proof.
\subsection{Proof of Theorem~\ref{thm:K43 one gap}}
\label{sec:proof K43 one gap}
By Lemma~\ref{lem:regret decomposition}, it remains to bound $\hat{R}(n) = \sum_{t = t_0}^n \Delta_{A_t} \I{\mathcal{F}_t}$, where the event $\mathcal{F}_t$ is defined in \eqref{eq:suboptimality event}. By Lemma~\ref{lem:K43 events} and from the assumption that $\Delta_{A_t} = \Delta$ for all suboptimal $A_t$, it follows that:
\begin{align*}
\hat{R}(n) =
\Delta \sum_{t = t_0}^n \I{\mathcal{F}_t} =
\Delta \sum_{t = t_0}^n \I{G_{1, t}, \Delta_{A_t} > 0} +
\Delta \sum_{t = t_0}^n \I{G_{2, t}, \Delta_{A_t} > 0}\,.
\end{align*}
To bound the above quantity, it is sufficient to bound the number of times that events $G_{1, t}$ and $G_{2, t}$ happen. Then we set the tunable parameters $d$ and $\alpha$ such that the two counts are of the same magnitude.
\begin{claim}
\label{claim:event 1} Event $G_{1, t}$ happens at most $\displaystyle \frac{\alpha}{d} K^2 L \frac{6}{\Delta^2} \log n$ times.
\end{claim}
\begin{proof}
Recall that event $G_{1, t}$ can happen only if at least $d$ chosen suboptimal items are not observed ``sufficiently often'' up to time $t$, $T_{t - 1}(e) \leq \alpha K^2 \frac{6}{\Delta^2} \log n$ for at least $d$ items in $\tilde{A}_t$. After the event happens, the observation counters of these items increase by one. Therefore, after the event happens $\frac{\alpha}{d} K^2 L \frac{6}{\Delta^2} \log n$ times, all suboptimal items are guaranteed to be observed at least $\alpha K^2 \frac{6}{\Delta^2} \log n$ times and $G_{1, t}$ cannot happen anymore.
\end{proof}
\begin{claim}
\label{claim:event 2} Event $G_{2, t}$ happens at most $\displaystyle \frac{\alpha d^2}{(\sqrt{\alpha} - 1)^2} L \frac{6}{\Delta^2} \log n$ times.
\end{claim}
\begin{proof}
Event $G_{2, t}$ can happen only if there exists $e \in \tilde{A}_t$ such that $T_{t - 1}(e) \leq \frac{\alpha d^2}{(\sqrt{\alpha} - 1)^2} \frac{6}{\Delta^2} \log n$. After the event happens, the observation counter of item $e$ increases by one. Therefore, the number of times that event $G_{2, t}$ can happen is bounded trivially by $\frac{\alpha d^2}{(\sqrt{\alpha} - 1)^2} L \frac{6}{\Delta^2} \log n$.
\end{proof}
\noindent Based on Claims~\ref{claim:event 1} and \ref{claim:event 2}, $\hat{R}(n)$ is bounded as:
\begin{align*}
\hat{R}(n) \leq
\left(\frac{\alpha}{d} K^2 + \frac{\alpha d^2}{(\sqrt{\alpha} - 1)^2}\right) L \frac{6}{\Delta} \log n\,.
\end{align*}
Finally, we choose $\alpha = 4$ and $d = K^\frac{2}{3}$; and it follows that the regret is bounded as:
\begin{align*}
R(n) \leq
\EE{\hat{R}(n)} + \left(\frac{\pi^2}{3} + 1\right) K L \leq
K^\frac{4}{3} L \frac{48}{\Delta} \log n + \left(\frac{\pi^2}{3} + 1\right) K L\,.
\end{align*}
\subsection{Proof of Theorem~\ref{thm:K43}}
\label{sec:proof K43}
Let $\mathcal{F}_t$ be the event in \eqref{eq:suboptimality event}. By Lemmas~\ref{lem:regret decomposition} and \ref{lem:K43 events}, it remains to bound:
\begin{align*}
\hat{R}(n) =
\sum_{t = t_0}^n \Delta_{A_t} \I{\mathcal{F}_t} =
\sum_{t = t_0}^n \Delta_{A_t} \I{G_{1, t}, \Delta_{A_t} > 0} +
\sum_{t = t_0}^n \Delta_{A_t} \I{G_{2, t}, \Delta_{A_t} > 0}\,.
\end{align*}
In the next step, we introduce item-specific variants of events $G_{1, t}$ \eqref{eq:event 1} and $G_{2, t}$ \eqref{eq:event 2}, and then associate the regret at time $t$ with these events. In particular, let:
\begin{align}
G_{e, 1, t} & = G_{1, t} \cap \left\{e \in \tilde{A}_t,
T_{t - 1}(e) \leq \alpha K^2 \frac{6}{\Delta_{A_t}^2} \log n\right\} \\
G_{e, 2, t} & = G_{2, t} \cap \left\{e \in \tilde{A}_t,
T_{t - 1}(e) \leq \frac{\alpha d^2}{(\sqrt{\alpha} - 1)^2} \frac{6}{\Delta_{A_t}^2} \log n\right\}
\end{align}
be the events that item $e$ is not observed ``sufficiently often'' under events $G_{1, t}$ and $G_{2, t}$, respectively. Then by the definitions of the above events, it follows that:
\begin{align*}
\I{G_{1, t}, \Delta_{A_t} > 0} & \leq
\frac{1}{d} \sum_{e \in \tilde{E}} \I{G_{e, 1, t}, \Delta_{A_t} > 0} \\
\I{G_{2, t}, \Delta_{A_t} > 0} &
\leq \sum_{e \in \tilde{E}} \I{G_{e, 2, t}, \Delta_{A_t} > 0}\,,
\end{align*}
where $\tilde{E} = E \setminus A^\ast$ is the set of subptimal items; and we bound $\hat{R}(n)$ as:
\begin{align*}
\hat{R}(n) \leq
\sum_{e \in \tilde{E}} \sum_{t = t_0}^n
\I{G_{e, 1, t}, \Delta_{A_t} > 0} \frac{\Delta_{A_t}}{d} +
\sum_{e \in \tilde{E}} \sum_{t = t_0}^n
\I{G_{e, 2, t}, \Delta_{A_t} > 0} \Delta_{A_t}\,.
\end{align*}
Let each item $e$ be contained in $N_e$ suboptimal solutions and $\Delta_{e, 1} \geq \ldots \geq \Delta_{e, N_e}$ be the gaps of these solutions, ordered from the largest gap to the smallest one. Then $\hat{R}(n)$ can be further bounded as:
\begin{align*}
\hat{R}(n)
& \leq \sum_{e \in \tilde{E}} \sum_{t = t_0}^n \sum_{k = 1}^{N_e}
\I{G_{e, 1, t}, \Delta_{A_t} = \Delta_{e, k}} \frac{\Delta_{e, k}}{d} +
\sum_{e \in \tilde{E}} \sum_{t = t_0}^n \sum_{k = 1}^{N_e}
\I{G_{e, 2, t}, \Delta_{A_t} = \Delta_{e, k}} \Delta_{e, k} \\
& \stackrel{\text{(a)}}{\leq} \sum_{e \in \tilde{E}} \sum_{t = t_0}^n \sum_{k = 1}^{N_e}
\I{e \in \tilde{A}_t, T_{t - 1}(e) \leq \alpha K^2 \frac{6}{\Delta_{e, k}^2} \log n,
\Delta_{A_t} = \Delta_{e, k}} \frac{\Delta_{e, k}}{d} + {} \\
& \quad\, \sum_{e \in \tilde{E}} \sum_{t = t_0}^n \sum_{k = 1}^{N_e}
\I{e \in \tilde{A}_t, T_{t - 1}(e) \leq \frac{\alpha d^2}{(\sqrt{\alpha} - 1)^2} \frac{6}{\Delta_{e, k}^2} \log n,
\Delta_{A_t} = \Delta_{e, k}} \Delta_{e, k} \\
& \stackrel{\text{(b)}}{\leq} \sum_{e \in \tilde{E}}
\frac{6 \alpha K^2 \log n}{d}
\left[\Delta_{e, 1} \frac{1}{\Delta_{e, 1}^2} + \sum_{k = 2}^{N_e} \Delta_{e, k}
\left(\frac{1}{\Delta_{e, k}^2} - \frac{1}{\Delta_{e, k - 1}^2}\right)\right] + {} \\
& \quad\, \sum_{e \in \tilde{E}}
\frac{6 \alpha d^2 \log n}{(\sqrt{\alpha} - 1)^2}
\left[\Delta_{e, 1} \frac{1}{\Delta_{e, 1}^2} + \sum_{k = 2}^{N_e} \Delta_{e, k}
\left(\frac{1}{\Delta_{e, k}^2} - \frac{1}{\Delta_{e, k - 1}^2}\right)\right] \\
& \stackrel{\text{(c)}}{<} \sum_{e \in \tilde{E}}
\left(\frac{\alpha}{d} K^2 + \frac{\alpha d^2}{(\sqrt{\alpha} - 1)^2}\right)
\frac{12}{\Delta_{e, \min}} \log n\,,
\end{align*}
where inequality (a) is by the definitions of events $G_{e, 1, t}$ and $G_{e, 2, t}$, inequality (b) is from the solution to:
\begin{align*}
\max_{A_1, \dots, A_n} \sum_{t = t_0}^n \sum_{k = 1}^{N_e}
\I{e \in \tilde{A}_t, T_{t - 1}(e) \leq \frac{C}{\Delta_{e, k}^2} \log n,
\Delta_{A_t} = \Delta_{e, k}} \Delta_{e, k}
\end{align*}
for appropriate $C$, and inequality (c) follows from Lemma 3 of Kveton \emph{et al.}~\cite{kveton14matroid}:
\begin{align}
\left[\Delta_{e, 1} \frac{1}{\Delta_{e, 1}^2} + \sum_{k = 2}^{N_e} \Delta_{e, k}
\left(\frac{1}{\Delta_{e, k}^2} - \frac{1}{\Delta_{e, k - 1}^2}\right)\right] <
\frac{2}{\Delta_{e, N_e}} =
\frac{2}{\Delta_{e, \min}}\,.
\label{eq:kveton2014}
\end{align}
Finally, we choose $\alpha = 4$ and $d = K^\frac{2}{3}$; and it follows that the regret is bounded as:
\begin{align*}
R(n) \leq
\EE{\hat{R}(n)} + \left(\frac{\pi^2}{3} + 1\right) K L \leq
\sum_{e \in \tilde{E}} K^\frac{4}{3} \frac{96}{\Delta_{e, \min}} \log n +
\left(\frac{\pi^2}{3} + 1\right) K L\,.
\end{align*}
\subsection{Proof of Theorem~\ref{thm:K one gap}}
\label{sec:proof K one gap}
The first step of the proof is identical to that of Theorem~\ref{thm:K43 one gap}. By Lemma~\ref{lem:regret decomposition}, it remains to bound $\hat{R}(n) = \sum_{t = t_0}^n \Delta_{A_t} \I{\mathcal{F}_t}$, where the event $\mathcal{F}_t$ is defined in \eqref{eq:suboptimality event}. By Lemma~\ref{lem:complete system} and from the assumption that $\Delta_{A_t} = \Delta$ for all suboptimal $A_t$, it follows that:
\begin{align*}
\hat{R}(n) =
\Delta \sum_{t = t_0}^n \I{\mathcal{F}_t} =
\Delta \sum_{i = 1}^\infty \sum_{t = t_0}^n \I{G_{i, t}, \Delta_{A_t} > 0}\,.
\end{align*}
Note that $\Delta_{A_t} > 0$ implies $\Delta_{A_t} = \Delta$. Therefore, $m_{i, t}$ does not depend on $t$ and we denote it by $m_i = \alpha_i \frac{K^2}{\Delta^2} \log n$. Based on the same argument as in Claim~\ref{claim:event 1}, event $G_{i, t}$ cannot happen more than $\frac{L m_i}{\beta_i K}$ times, because at least $\beta_i K$ items that are observed at most $m_i$ times have their observation counters incremented in each event $G_{i, t}$. Therefore:
\begin{align}
\hat{R}(n) \leq
\Delta \sum_{i = 1}^\infty \frac{L m_i}{\beta_i K} =
K L \frac{1}{\Delta} \left[\sum_{i = 1}^\infty \frac{\alpha_i}{\beta_i}\right] \log n\,.
\label{eq:ab regret bound}
\end{align}
It remains to choose $(\alpha_i)$ and $(\beta_i)$ such that:
\begin{itemize}
\item $\lim_{i \to \infty} \alpha_i = \lim_{i \to \infty} \beta_i = 0$;
\item Monotonicity conditions in \eqref{eq:condition 1} and \eqref{eq:condition 2} hold;
\item Inequality~\eqref{eq:condition 3} holds,
$\sqrt{6} \sum_{i = 1}^\infty \frac{\beta_{i - 1} - \beta_i}{\sqrt{\alpha_i}} \leq 1$;
\item $\sum_{i = 1}^\infty \frac{\alpha_i}{\beta_i}$ is minimized.
\end{itemize}
We choose $(\alpha_i)$ and $(\beta_i)$ to be geometric sequences, $\beta_i = \beta^i$ and $\alpha_i = d \alpha^i$ for $0 < \alpha, \beta < 1$ and $d > 0$. For this setting, $\alpha_i \to 0$ and $\beta_i \to 0$, and the monotonicity conditions are also satisfied. Moreover, if $\beta < \sqrt{\alpha}$, we have:
\begin{align*}
\sqrt{6} \sum_{i = 1}^\infty \frac{\beta_{i - 1} - \beta_i}{\sqrt{\alpha_i}} =
\sqrt{6} \sum_{i = 1}^\infty \frac{\beta^{i - 1} - \beta^i}{\sqrt{d \alpha^i}} =
\sqrt{\frac{6}{d}} \frac{1 - \beta}{\sqrt{\alpha} - \beta} \leq
1
\end{align*}
provided that $d \geq 6 \left(\frac{1 - \beta}{\sqrt{\alpha} - \beta}\right)^2$. Furthermore, if $\alpha < \beta$, we have:
\begin{align*}
\sum_{i = 1}^\infty \frac{\alpha_i}{\beta_i} =
\sum_{i = 1}^\infty \frac{d \alpha^i}{\beta^i} =
\frac{d \alpha}{\beta - \alpha}\,.
\end{align*}
Given the above, the best choice of $d$ is $6 \left(\frac{1 - \beta}{\sqrt{\alpha} - \beta}\right)^2$ and the problem of minimizing the constant in our regret bound can be written as:
\begin{align*}
\inf_{\alpha, \beta} & \quad 6 \left(\frac{1 - \beta}{\sqrt{\alpha} - \beta}\right)^2
\frac{\alpha}{\beta - \alpha} \\
\textrm{s.t.} & \quad 0 < \alpha < \beta < \sqrt{\alpha} < 1\,.
\end{align*}
We find the solution to the above problem numerically, and determine it to be $\alpha = 0.1459$ and $\beta = 0.2360$. For these $\alpha$ and $\beta$, $6 \left(\frac{1 - \beta}{\sqrt{\alpha} - \beta}\right)^2 \frac{\alpha}{\beta - \alpha} < 267$. We apply this upper bound to \eqref{eq:ab regret bound} and it follows that the regret is bounded as:
\begin{align*}
R(n) \leq
\EE{\hat{R}(n)} + \left(\frac{\pi^2}{3} + 1\right) K L \leq
K L \frac{267}{\Delta} \log n + \left(\frac{\pi^2}{3} + 1\right) K L\,.
\end{align*}
\subsection{Proof of Theorem~\ref{thm:K}}
\label{sec:proof K}
Let $\mathcal{F}_t$ be the event in \eqref{eq:suboptimality event}. By Lemmas~\ref{lem:regret decomposition} and \ref{lem:complete system}, it remains to bound:
\begin{align*}
\hat{R}(n) =
\sum_{t = t_0}^n \Delta_{A_t} \I{\mathcal{F}_t} =
\sum_{i = 1}^\infty \sum_{t = t_0}^n \Delta_{A_t} \I{G_{i, t}, \Delta_{A_t} > 0}\,.
\end{align*}
In the next step, we define item-specific variants of events $G_{i, t}$ \eqref{eq:event i} and associate the regret at time $t$ with these events. In particular, let:
\begin{align}
G_{e, i, t} = G_{i, t} \cap \left\{e \in \tilde{A}_t, T_{t - 1}(e) \leq m_{i, t}\right\}
\end{align}
be the event that item $e$ is not observed ``sufficiently often'' under event $G_{i, t}$. Then it follows that:
\begin{align*}
\I{G_{i, t}, \Delta_{A_t} > 0} \leq
\frac{1}{\beta_i K} \sum_{e \in \tilde{E}} \I{G_{e, i, t}, \Delta_{A_t} > 0}\,,
\end{align*}
because at least $\beta_i K$ items are not observed ``sufficiently often'' under event $G_{i, t}$. Therefore, we can bound $\hat{R}(n)$ as:
\begin{align*}
\hat{R}(n) \leq
\sum_{e \in \tilde{E}} \sum_{i = 1}^\infty \sum_{t = t_0}^n
\I{G_{e, i, t}, \Delta_{A_t} > 0} \frac{\Delta_{A_t}}{\beta_i K}\,.
\end{align*}
Let each item $e$ be contained in $N_e$ suboptimal solutions and $\Delta_{e, 1} \geq \ldots \geq \Delta_{e, N_e}$ be the gaps of these solutions, ordered from the largest gap to the smallest one. Then $\hat{R}(n)$ can be further bounded as:
\begin{align*}
\hat{R}(n)
& \leq \sum_{e \in \tilde{E}} \sum_{i = 1}^\infty \sum_{t = t_0}^n \sum_{k = 1}^{N_e}
\I{G_{e, i, t}, \Delta_{A_t} = \Delta_{e, k}} \frac{\Delta_{e, k}}{\beta_i K} \\
& \stackrel{\text{(a)}}{\leq} \sum_{e \in \tilde{E}} \sum_{i = 1}^\infty \sum_{t = t_0}^n \sum_{k = 1}^{N_e}
\I{e \in \tilde{A}_t, T_{t - 1}(e) \leq \alpha_i \frac{K^2}{\Delta_{e, k}^2} \log n,
\Delta_{A_t} = \Delta_{e, k}} \frac{\Delta_{e, k}}{\beta_i K} \\
& \stackrel{\text{(b)}}{\leq} \sum_{e \in \tilde{E}} \sum_{i = 1}^\infty
\frac{\alpha_i K \log n}{\beta_i}
\left[\Delta_{e, 1} \frac{1}{\Delta_{e, 1}^2} + \sum_{k = 2}^{N_e} \Delta_{e, k}
\left(\frac{1}{\Delta_{e, k}^2} - \frac{1}{\Delta_{e, k - 1}^2}\right)\right] \\
& \stackrel{\text{(c)}}{<} \sum_{e \in \tilde{E}} \sum_{i = 1}^\infty
\frac{\alpha_i K \log n}{\beta_i} \frac{2}{\Delta_{e, \min}} \\
& = \sum_{e \in \tilde{E}} K \frac{2}{\Delta_{e, \min}}
\left[\sum_{i = 1}^\infty \frac{\alpha_i}{\beta_i}\right] \log n\,,
\end{align*}
where inequality (a) is by the definition of event $G_{e, i, t}$, inequality (b) follows from the solution to:
\begin{align*}
\max_{A_1, \dots, A_n} \sum_{t = t_0}^n \sum_{k = 1}^{N_e}
\I{e \in \tilde{A}_t, T_{t - 1}(e) \leq \alpha_i \frac{K^2}{\Delta_{e, k}^2} \log n,
\Delta_{A_t} = \Delta_{e, k}} \frac{\Delta_{e, k}}{\beta_i K}\,,
\end{align*}
and inequality (c) follows from \eqref{eq:kveton2014}. For the same $(\alpha_i)$ and $(\beta_i)$ as in Theorem~\ref{thm:K one gap}, we have $\sum_{i = 1}^\infty \frac{\alpha_i}{\beta_i} < 267$ and it follows that the regret is bounded as:
\begin{align*}
R(n) \leq
\EE{\hat{R}(n)} + \left(\frac{\pi^2}{3} + 1\right) K L \leq
\sum_{e \in \tilde{E}} K \frac{534}{\Delta_{e, \min}} \log n +
\left(\frac{\pi^2}{3} + 1\right) K L\,.
\end{align*}
\subsection{Proof of Theorem~\ref{thm:gap-free}}
\label{sec:proof gap-free}
The key idea is to decompose the regret of ${\tt CombUCB1}$ into two parts, where the gaps are larger than $\epsilon$ and at most $\epsilon$. We analyze each part separately and then set $\epsilon$ to get the desired result.
By Lemma~\ref{lem:regret decomposition}, it remains to bound $\hat{R}(n) = \sum_{t = t_0}^n \Delta_{A_t} \I{\mathcal{F}_t}$, where the event $\mathcal{F}_t$ is defined in \eqref{eq:suboptimality event}. We partition $\hat{R}(n)$ as:
\begin{align*}
\hat{R}(n)
& = \sum_{t = t_0}^n \Delta_{A_t} \I{\mathcal{F}_t, \Delta_{A_t} < \epsilon} +
\sum_{t = t_0}^n \Delta_{A_t} \I{\mathcal{F}_t, \Delta_{A_t} \geq \epsilon} \\
& \leq \epsilon n + \sum_{t = t_0}^n \Delta_{A_t} \I{\mathcal{F}_t, \Delta_{A_t} \geq \epsilon}
\end{align*}
and bound the first term trivially. The second term is bounded in the same way as $\hat{R}(n)$ in the proof of Theorem~\ref{thm:K}, except that we only consider the gaps $\Delta_{e, k} \geq \epsilon$. Therefore, $\Delta_{e, \min} \geq \epsilon$ and we get:
\begin{align*}
\sum_{t = t_0}^n \Delta_{A_t} \I{\mathcal{F}_t, \Delta_{A_t} \geq \epsilon} \leq
\sum_{e \in \tilde{E}} K \frac{534}{\epsilon} \log n \leq
K L \frac{534}{\epsilon} \log n\,.
\end{align*}
Based on the above inequalities:
\begin{align*}
R(n) \leq \frac{534 K L}{\epsilon} \log n + \epsilon n + \left(\frac{\pi^2}{3} + 1\right) K L\,.
\end{align*}
Finally, we choose $\displaystyle \epsilon = \sqrt{\frac{534 K L \log n}{n}}$ and get:
\begin{align*}
R(n) \leq
2 \sqrt{534 K L n \log n} + \left(\frac{\pi^2}{3} + 1\right) K L <
47 \sqrt{K L n \log n} + \left(\frac{\pi^2}{3} + 1\right) K L\,,
\end{align*}
which concludes our proof.
\section{Technical Lemmas}
\label{sec:lemmas}
\begin{lemma}
\label{lem:little helper} Let $S_i$, $\bar{S}_i$, and $m_i$ be defined as in Lemma~\ref{lem:complete system}; and $|S_i| < \beta_i K$ for all $i > 0$. Then:
\begin{align*}
\sum_{i = 1}^\infty \frac{|\bar{S}_i \setminus \bar{S}_{i - 1}|}{\sqrt{m_i}} <
\sum_{i = 1}^\infty \frac{(\beta_{i - 1} - \beta_i) K}{\sqrt{m_i}}\,.
\end{align*}
\end{lemma}
\begin{proof}
The lemma is proved as:
\begin{align*}
\sum_{i = 1}^\infty |\bar{S}_i \setminus \bar{S}_{i - 1}| \frac{1}{\sqrt{m_i}}
& = \sum_{i = 1}^\infty (|S_{i - 1} \setminus S_i|) \frac{1}{\sqrt{m_i}} \\
& = \sum_{i = 1}^\infty (|S_{i - 1}| - |S_i|) \frac{1}{\sqrt{m_i}} \\
& = \frac{|S_0|}{\sqrt{m_1}} +
\sum_{i = 1}^\infty |S_i| \left(\frac{1}{\sqrt{m_{i + 1}}} - \frac{1}{\sqrt{m_i}}\right) \\
& < \frac{\beta_0 K}{\sqrt{m_1}} +
\sum_{i = 1}^\infty \beta_i K \left(\frac{1}{\sqrt{m_{i + 1}}} - \frac{1}{\sqrt{m_i}}\right) \\
& = \sum_{i = 1}^\infty (\beta_{i - 1} - \beta_i) K \frac{1}{\sqrt{m_i}}\,.
\end{align*}
The first two equalities follow from the definitions of $\bar{S}_i$ and $S_i$. The inequality follows from the facts that $|S_i| < \beta_i K$ for all $i > 0$ and $|S_0| \leq \beta_0 K$.
\end{proof}
\section{Conclusions}
\label{sec:conclusions}
The main contribution of this work is that we derive novel gap-dependent and gap-free upper bounds on the regret of ${\tt CombUCB1}$, a UCB-like algorithm for stochastic combinatorial semi-bandits. These bounds are tight up to polylogarithmic factors. In other words, we show that ${\tt CombUCB1}$ is sample efficient because it achieves near-optimal regret. It is well known that ${\tt CombUCB1}$ is also computationally efficient \cite{gai12combinatorial}, it can be implemented efficiently whenever the offline variant of the problem can be solved computationally efficiently. Therefore, we indirectly show that stochastic combinatorial semi-bandits can be solved both computationally and sample efficiently, by ${\tt CombUCB1}$.
Theorems~\ref{thm:K one gap} and \ref{thm:K} are proved quite generally, for any $(\alpha_i)$ and $(\beta_i)$ subject to relatively mild constraints. At the end of the proofs, we choose $(\alpha_i)$ and $(\beta_i)$ to be geometric sequences. This is sufficient for our purpose. But the choice is likely to be suboptimal and may lead to larger constants in our upper bounds than is necessary. We leave the problem of choosing better $(\alpha_i)$ and $(\beta_i)$ for future work.
We leave open several questions of interest. For instance, our $\Omega(K L (1 / \Delta) \log n)$ lower bound is derived on a problem where all suboptimal solutions have the same gaps. So technically speaking, our $O(K L (1 / \Delta) \log n)$ upper bound is tight only on this family of problems. It is an open problem whether our upper bound is tight in general.
Our $O(\sqrt{K L n \log n})$ upper bound matches the $\Omega(\sqrt{K L n})$ lower bound up to a factor of $\sqrt{\log n}$. We believe that this factor can be eliminated by modifying the confidence radii in ${\tt CombUCB1}$ \eqref{eq:confidence radius} along the lines of Audibert \emph{et al.}~\cite{audibert09minimax}. We leave this for future work.
\section{Experiments}
\label{sec:experiments}
In this section, we evaluate ${\tt CombUCB1}$ on a synthetic problem and demonstrate that its regret grows as suggested by our $O(K L (1 / \Delta) \log n)$ upper bound. We experiment with a stochastic longest-path problem on a $(m + 1) \times (m + 1)$ square grid (Figure~\ref{fig:main}b). The items in the ground set $E$ are the edges in the grid, $2 m (m + 1)$ in total. The feasible set $\Theta$ are all paths in the grid from the upper left corner to the bottom right corner that follow the directions of the edges. The length of these paths is $K = 2 m$. The weight of edge $e$ is drawn i.i.d. from a Bernoulli distribution with mean:
\begin{align*}
\bar{w}(e) =
\begin{cases}
0.5 + \sigma / 2 & \text{$e$ is a leftmost or bottomost edge} \\
0.5 - \sigma / 2 & \text{otherwise}\,,
\end{cases}
\end{align*}
where $0 < \sigma < 1$. The optimal solution $A^\ast$ is a path along the leftmost and bottommost edges (Figure~\ref{fig:main}b).
The sample complexity of our problem is characterized by $|\tilde{E}| = 2 m (m + 1) - 2 m$ gaps $\Delta_{e, \min}$ ranging from $2 \sigma$ to $2 m \sigma$. It is easy to show that the number of items $e$ where $\Delta_{e, \min} = i \sigma$ is at most $2 (i - 1)$. Therefore, we can bound the $(\log n)$-term in Theorem~\ref{thm:K} as:
\begin{align}
\sum_{e \in \tilde{E}} K \frac{534}{\Delta_{e, \min}} \log n
& < 1068 m \sum_{i = 2}^{2 m} \frac{2 i}{i \sigma} \log n \nonumber \\
& < \frac{4272 m^2 \log n}{\sigma}\,. \label{eq:grid regret bound}
\end{align}
Now we validate the dependence on $m$ and $\sigma$ empirically. We vary $m$ and $\sigma$, and run ${\tt CombUCB1}$ for $n = 10^5$ steps.
Our experimental results are reported in Figure~\ref{fig:main}c. We observe two trends. First, the regret of ${\tt CombUCB1}$ is linear in the number of items $L$, which depends quadratically on $m$ since $L = 2 m (m + 1)$. Second, the regret is linear in $1 / \sigma$. The dependence on $m$ and $1 / \sigma$ is the same as in our upper bound in \eqref{eq:grid regret bound}.
\section{Extensions}
\label{sec:extensions}
The computational efficiency of ${\tt CombUCB1}$ depends on the computational efficiency of the offline optimization oracle. When the oracle is inefficient, we suggest resorting to approximations. Let ${\tt ALG}$ be a computationally-efficient oracle that returns an approximation. Then ${\tt CombUCB1}$ can be straightforwardly modified to call ${\tt ALG}$ instead of the original oracle. Moreover, it is easy to bound the regret of this algorithm if it is measured with respect to the best approximate solution by ${\tt ALG}$ in hindsight.
Thompson sampling \cite{thompson33likelihood} often performs better in practice than ${\tt UCB1}$ \cite{auer02finitetime}. It is straightforward to propose a variant of ${\tt CombUCB1}$ that uses Thompson sampling, by replacing the UCBs in Algorithm~\ref{alg:ucb1} with sampling from the posterior on the mean of the weights. The frequentist analysis of regret in Thompson sampling \cite{agrawal12analysis} resembles the analysis of ${\tt UCB1}$. Therefore, we believe that our analysis can be generalized to Thompson sampling, and we hypothesize that the regret of the resulting algorithm is $O(K L (1 / \Delta) \log n)$.
\section{Introduction}
\label{sec:introduction}
A \emph{stochastic combinatorial semi-bandit} \cite{gai12combinatorial,chen13combinatorial} is an online learning problem where at each step a learning agent chooses a subset of ground items subject to combinatorial constraints, and then observes stochastic weights of these items and receives their sum as a payoff. The problem can be viewed as a learning variant of combinatorial optimization with a linear objective function and binary variables. Many classical combinatorial optimization problems have linear objectives \cite{papadimitriou98combinatorial}. Therefore, stochastic combinatorial semi-bandits have found many practical applications, such as learning spectrum allocations \cite{gai12combinatorial}, shortest paths \cite{gai12combinatorial}, routing networks \cite{kveton14matroid}, and recommendations \cite{kveton14matroid,kveton14learning}.
In our work, we study a variant of stochastic combinatorial semi-bandits where the learning agent has access to an \emph{offline optimization oracle} that can find the optimal solution for any weights of the items. We say that the problem is a $(L, K, \Delta)$ instance when $L$ is the cardinality of its ground set $E$, $K$ is the maximum number of chosen items, and $\Delta$ is the gap between the expected returns of the optimal and best suboptimal solutions. We also say that the problem is a $(L, K)$ instance if it is a $(L, K, \Delta)$ instance for some $\Delta$. Based on the existing bandit work \cite{auer02finitetime}, it is relatively easy to propose a UCB-like algorithm for solving our problem \cite{gai12combinatorial}, and we call this algorithm ${\tt CombUCB1}$. ${\tt CombUCB1}$ is a variant of ${\tt UCB1}$ that calls the oracle to find the optimal solution with respect to the upper confidence bounds on the weights of the items. Chen \emph{et al.}~\cite{chen13combinatorial} recently showed that the $n$-step regret of ${\tt CombUCB1}$ in any $(L, K, \Delta)$ stochastic combinatorial semi-bandit is $O(K^2 L (1 / \Delta) \log n)$.
Our main contribution is that we derive two upper bounds on the $n$-step regret of ${\tt CombUCB1}$, $O(K L (1 / \Delta) \log n)$ and $O(\sqrt{K L n \log n})$. Both of these bounds are significant improvements over Chen \emph{et al.}~\cite{chen13combinatorial}. Moreover, we prove two novel lower bounds, $\Omega(K L (1 / \Delta) \log n)$ and $\Omega(\sqrt{K L n})$, which match our upper bounds up to polylogarithmic factors. The consequence of these results is that ${\tt CombUCB1}$ is \emph{sample efficient} because it achieves near-optimal regret. It is well known that ${\tt CombUCB1}$ is also \emph{computationally efficient} \cite{gai12combinatorial}, it can be implemented efficiently whenever the offline optimization oracle is computationally efficient. So we close the problem of computationally and sample efficient learning in stochastic combinatorial semi-bandits, by showing that ${\tt CombUCB1}$ has both properties. This problem is still open in the adversarial setting (Section~\ref{sec:related work}).
Our analysis is novel. It is based on the idea that the event that ``many'' items in a chosen suboptimal solution are not observed ``sufficiently often'' does not happen ``too often''. The reason is that this event happens for ``many'' items simultaneously. Therefore, when the event happens, the observation counters of ``many'' items increase. Based on this observation, we divide the regret associated with the event among ``many'' items, instead of attributing it separately to each item as in the prior work \cite{gai12combinatorial,chen13combinatorial}. This is the key step in our analysis that yields tight upper bounds.
Our paper is organized as follows. In Section~\ref{sec:setting}, we introduce our learning problem and the algorithm for solving it. In Section~\ref{sec:discussion}, we summarize our results. In Section~\ref{sec:K43 upper bounds}, we prove a $O(K^\frac{4}{3} L (1 / \Delta) \log n)$ upper bound on the regret of ${\tt CombUCB1}$. In Section~\ref{sec:K upper bounds}, we prove a $O(K L (1 / \Delta) \log n)$ upper bound on the regret of ${\tt CombUCB1}$. In Section~\ref{sec:lower bounds}, we prove gap-dependent and gap-free lower bounds. In Section~\ref{sec:experiments}, we evaluate ${\tt CombUCB1}$ on a synthetic problem and show that its $n$-step regret grows as suggested by our gap-dependent upper bound. In Section~\ref{sec:related work}, we compare our results to prior work. In Section~\ref{sec:extensions}, we discuss extensions of our work. We conclude in Section~\ref{sec:conclusions}.
\section{Lower Bounds}
\label{sec:lower bounds}
In this section, we derive two lower bounds on the $n$-step regret in stochastic combinatorial semi-bandits. One of the bounds is gap-dependent and the other one is gap-free.
\begin{figure*}[t]
\centering
\hspace{-0.15in}
\includegraphics[width=4.4in, bb=1in 4.25in 6.5in 6.5in]{Figures/Figures}
\raisebox{0.055in}{\includegraphics[width=2.4in, bb=2.75in 4.5in 5.75in 6.5in]{Figures/GridTrends}}
\hspace{-0.4in} \makebox{} \\
\hspace{0.09in} (a) \hspace{1.925in} (b) \hspace{2.37in} (c)
\caption{{\bf a}. The $K$-path semi-bandit problem in Section~\ref{sec:lower bounds}. The red and blue nodes are the starting and end points of the paths, respectively. The optimal path is marked in red. {\bf b}. The grid-path problem in Section~\ref{sec:experiments}. The red and blue nodes are the starting and end points of the paths, respectively. The optimal path is marked in red. {\bf c}. The $n$-step regret of ${\tt CombUCB1}$ on the grid-path problem.}
\label{fig:main}
\end{figure*}
Our bounds are derived on a \emph{$K$-path semi-bandit} problem, which is illustrated in Figure~\ref{fig:main}a. The items in the ground set are $L$ path segments $E = \set{1, \dots, L}$. The feasible set $\Theta$ are $L / K$ paths, each of which contains $K$ unique items. Specifically, path $j$ contains items $(j - 1) K + 1, \dots, j K$. Without loss of generality, we assume that $L / K$ is an integer. The probability distribution $P$ over the weights of the items is defined as follows. The weights of the items in the same path are equal. The weights of the items in different paths are distributed independently. The weight of item $e$ is a Bernoulli random variable with mean:
\begin{align*}
\bar{w}(e) =
\begin{cases}
0.5 & \text{item $e$ belongs to path $1$} \\
0.5 - \Delta / K & \text{otherwise}\,,
\end{cases}
\end{align*}
where $0 < \Delta / K < 0.5$. Note that our problem is designed such that $\Delta_{e, \min} = \Delta$ for any item $e$ in path $j > 1$.
The key observation is that our problem is equivalent to a $(L / K)$-arm Bernoulli bandit whose returns are scaled by $K$, when the learning agent knows that the weights of the items in the same path are equal. Therefore, we can derive lower bounds for our problem based on the existing lower bounds for Bernoulli bandits \cite{auer02nonstochastic,bubeck12regret,lai85asymptotically}.
Our gap-dependent lower bound is derived for the class of consistent algorithms, which is defined as follows. We say that the algorithm is \emph{consistent} if for any stochastic combinatorial semi-bandit, any suboptimal $A$, and any $\alpha > 0$, $\EE{T_n(A)} = o(n^\alpha)$, where $T_n(A)$ is the number of times that solution $A$ is chosen in $n$ steps. The restriction to the consistent algorithms is without loss of generality. In particular, an inconsistent algorithm is guaranteed to perform poorly on some semi-bandit, and therefore cannot achieve logarithmic regret on all semi-bandits, as ${\tt CombUCB1}$.
\begin{proposition}
\label{prop:lower bound} For any $L$ and $K$ such that $L / K$ is an integer, and any $0 < \Delta / K < 0.5$, the regret of any consistent algorithm on the $K$-path semi-bandit problem is bounded from below as:
\begin{align*}
\liminf_{n \to \infty} \frac{R(n)}{\log n} \geq
\frac{(L - K) K}{4 \Delta}\,.
\end{align*}
\end{proposition}
\begin{proof}
The proposition is proved as follows:
\begin{align*}
\liminf_{n \to \infty} \frac{R(n)}{\log n}
& \stackrel{\text{(a)}}{\geq} K \sum_{k = 2}^{L / K} \frac{\Delta / K}{D_\mathrm{KL}(0.5 - \Delta / K \,\|\, 0.5)}
\nonumber \\
& = \left(\frac{L}{K} - 1\right) \frac{\Delta}{D_\mathrm{KL}(0.5 - \Delta / K \,\|\, 0.5)} \nonumber \\
& \stackrel{\text{(b)}}{\geq} \frac{(L - K) K}{4 \Delta}\,,
\end{align*}
where $D_\mathrm{KL}(p \,\|\, q)$ is the Kullback-Leibler (KL) divergence between two Bernoulli random variables with means $p$ and $q$. Inequality (a) follows from the fact that our problem is equivalent to a $(L / K)$-arm Bernoulli bandit whose returns are scaled by $K$. Therefore, we can bound the regret from below using an existing lower bound for Bernoulli bandits \cite{lai85asymptotically}. Inequality (b) is due to $D_\mathrm{KL}(p \,\|\, q) \leq \frac{(p - q)^2}{q (1 - q)}$, where $p = 0.5 - \Delta / K$ and $q = 0.5$.
\end{proof}
We also derive a gap-free lower bound.
\begin{proposition}
\label{prop:gap-free lower bound} For any $L$ and $K$ such that $L / K$ is an integer, and any horizon $n > 0$, there exists a $K$-path semi-bandit problem such that the regret of any algorithm is:
\begin{align*}
R(n) \geq \frac{1}{20} \min(\sqrt{K L n}, Kn)\,.
\end{align*}
\end{proposition}
\begin{proof}
The $K$-path semi-bandit problem is equivalent to a $(L / K)$-arm Bernoulli bandit whose payoffs are scaled by $K$. Therefore, we can apply Theorem 5.1 of Auer \emph{et al.}~\cite{auer02nonstochastic} and bound the regret of any algorithm from below by:
\begin{align*}
\frac{K}{20} \min(\sqrt{(L / K) n}, n) = \frac{1}{20} \min(\sqrt{K L n}, K n)\,.
\end{align*}
Note that the bound of Auer \emph{et al.}~\cite{auer02nonstochastic} is for the adversarial setting. However, the worst-case environment in the proof is stochastic and therefore it applies to our problem.
\end{proof}
\section{Summary of Main Results}
\label{sec:discussion}
We prove three upper bounds on the regret of ${\tt CombUCB1}$. Two bounds depend on the gap $\Delta$ and one is gap-free:
\begin{align*}
\text{Theorem~\ref{thm:K43}}: & \qquad O(K^\frac{4}{3} L (1 / \Delta) \log n) \\
\text{Theorem~\ref{thm:K}}: & \qquad O(K L (1 / \Delta) \log n) \\
\text{Theorem~\ref{thm:gap-free}}: & \qquad O(\sqrt{K L n \log n})\,.
\end{align*}
Both gap-dependent bounds are major improvements over $O(K^2 L (1 / \Delta) \log n)$, the best known upper bound on the $n$-step regret of ${\tt CombUCB1}$ \cite{chen13combinatorial}. The bound in Theorem~\ref{thm:K} is asymptotically tighter than the bound in Theorem~\ref{thm:K43}, but the latter is tighter for $K < (534 / 96)^3 < 173$.
One of the main contributions of our work is that we identify an algorithm for stochastic combinatorial semi-bandits that is both computationally and sample efficient. The following are our definitions of computational and sample efficiency. We say that the algorithm is \emph{computationally efficient} if it can be implemented efficiently whenever the offline variant of the problem can be solved computationally efficiently. The algorithm is \emph{sample efficient} if it achieves optimal regret up to polylogarithmic factors. Based on our definitions, ${\tt CombUCB1}$ is both \emph{computationally} and \emph{sample efficient}. We state this result slightly more formally below.
\begin{theorem}
\label{thm:efficiency} ${\tt CombUCB1}$ is computationally and sample efficient in any $(L, K)$ stochastic combinatorial semi-bandit where the offline optimization oracle $\arg\max_{A \in \Theta} f(A, w)$ can be implemented efficiently for any $w \in (\mathbb{R}^+)^E$.
\end{theorem}
\begin{proof}
In each step $t$, ${\tt CombUCB1}$ calls the oracle once, and all of its remaining operations are polynomial in $L$ and $K$. Therefore, ${\tt CombUCB1}$ is guaranteed to be computationally efficient when the oracle is computationally efficient.
${\tt CombUCB1}$ is sample efficient because it achieves optimal regret up to polylogarithmic factors. In particular, the gap-dependent upper bound on the $n$-step regret of ${\tt CombUCB1}$ in Theorem~\ref{thm:K} matches the lower bound in Proposition~\ref{prop:lower bound} up to a constant factor. In addition, the gap-free upper bound in Theorem~\ref{thm:gap-free} matches the lower bound in Proposition~\ref{prop:gap-free lower bound} up to a factor of $\sqrt{\log n}$.
\end{proof}
\section{$O(K^\frac{4}{3})$ Upper Bounds}
\label{sec:K43 upper bounds}
In this section, we prove two $O(K^\frac{4}{3} L (1 / \Delta) \log n)$ upper bounds on the $n$-step regret of ${\tt CombUCB1}$. In Theorem~\ref{thm:K43 one gap}, we assume that the gaps of all suboptimal solutions are the same. In Theorem~\ref{thm:K43}, we relax this assumption.
The \emph{gap} of solution $A$ is $\Delta_A = f(A^\ast, \bar{w}) - f(A, \bar{w})$. The results in this section are presented for their didactic value. Their proofs are simple. Yet they illustrate the main ideas that lead to the tight regret bounds in Section~\ref{sec:K upper bounds}.
\begin{theorem}
\label{thm:K43 one gap} In any $(L, K, \Delta)$ stochastic combinatorial semi-bandit where $\Delta_{A} = \Delta$ for all suboptimal solutions $A$, the regret of ${\tt CombUCB1}$ is bounded as:
\begin{align*}
R(n) \leq K^\frac{4}{3} L \frac{48}{\Delta} \log n + \left(\frac{\pi^2}{3} + 1\right) K L\,.
\end{align*}
\end{theorem}
The proof of Theorem~\ref{thm:K43 one gap} relies on two lemmas. In the first lemma, we bound the regret associated with the initialization of ${\tt CombUCB1}$ and the event that $\bar{w}(e)$ is outside of the high-probability confidence interval around $\hat{w}_{T_{t - 1}(e)}(e)$.
\begin{lemma}
\label{lem:regret decomposition} Let:
\begin{align}
\mathcal{F}_t = \set{\Delta_{A_t} \leq 2 \sum_{e \in \tilde{A}_t} c_{n, T_{t - 1}(e)}, \ \Delta_{A_t} > 0}
\label{eq:suboptimality event}
\end{align}
be the event that suboptimal solution $A_t$ is ``hard to distinguish'' from $A^\ast$ at time $t$, where $\tilde{A}_t = A_t \setminus A^\ast$. Then the regret of ${\tt CombUCB1}$ is bounded as:
\begin{align}
R(n) \leq \EE{\hat{R}(n)} + \left(\frac{\pi^2}{3} + 1\right) K L\,,
\label{eq:regret bound}
\end{align}
where:
\begin{align}
\hat{R}(n) = \sum_{t = t_0}^n \Delta_{A_t} \I{\mathcal{F}_t}.
\label{eq:Rhat}
\end{align}
\end{lemma}
\begin{proof}
The claim is proved in Appendix~\ref{sec:proof regret decomposition}.
\end{proof}
Now we bound the regret corresponding to the events $\mathcal{F}_t$, the items in a suboptimal solution are not observed ``sufficiently often'' up to time $t$. To bound the regret, we define two events:
\begin{align}
G_{1, t} = \bigg\{
& \text{at least $d$ items in $\tilde{A}_t$ were observed} \label{eq:event 1} \\
& \text{at most $\alpha K^2 \frac{6}{\Delta_{A_t}^2} \log n$ times}
\bigg\} \nonumber
\end{align}
and:
\begin{align}
G_{2, t} = \bigg\{
& \text{less than $d$ items in $\tilde{A}_t$ were observed} \label{eq:event 2} \\
& \text{at most $\alpha K^2 \frac{6}{\Delta_{A_t}^2} \log n$ times}, \nonumber \\
& \text{at least one item in $\tilde{A}_t$ was observed} \nonumber \\
& \text{at most $\frac{\alpha d^2}{(\sqrt{\alpha} - 1)^2} \frac{6}{\Delta_{A_t}^2} \log n$ times}
\bigg\}\,, \nonumber
\end{align}
where $\alpha \geq 1$ and $d > 0$ are parameters, which are chosen later. The event $G_{1, t}$ happens when ``many'' chosen items, at least $d$, are not observed ``sufficiently often'' up to time $t$, at most $\alpha K^2 \frac{6}{\Delta_{A_t}^2} \log n$ times.
Events $G_{1, t}$ and $G_{2, t}$ are obviously mutually exclusive. In the next lemma, we prove that these events are exhaustive when event $\mathcal{F}_t$ happens. To prove this claim, we introduce new notation. We denote the set of items in $\tilde{A}_t$ that are not observed ``sufficiently often'' up to time $t$ by:
\begin{align*}
S_t = \cset{e \in \tilde{A}_t}{T_{t - 1}(e) \leq \alpha K^2 \frac{6}{\Delta_{A_t}^2} \log n}\,.
\end{align*}
\begin{lemma}
\label{lem:K43 events} Let $\alpha \geq 1$, $d > 0$, and event $\mathcal{F}_t$ happen. Then either event $G_{1, t}$ or $G_{2, t}$ happens.
\end{lemma}
\begin{proof}
By the definition of $S_t$, the following three events:\small
\begin{align*}
G_{1, t} & = \set{|S_t| \geq d} \\
G_{2, t} & = \set{|S_t| < d, \ \left[\exists e \in \tilde{A}_t: T_{t - 1}(e) \leq
\frac{6 \alpha d^2 \log n}{(\sqrt{\alpha} - 1)^2 \Delta_{A_t}^2}\right]} \\
\bar{G}_t & = \set{|S_t| < d, \ \left[\forall e \in \tilde{A}_t: T_{t - 1}(e) >
\frac{6 \alpha d^2 \log n}{(\sqrt{\alpha} - 1)^2 \Delta_{A_t}^2}\right]}
\end{align*}\normalsize
are exhaustive and mutually exclusive. Therefore, to prove that either $G_{1, t}$ or $G_{2, t}$ happens, it suffices to show that $\bar{G}_t$ does not happen. Suppose that event $\bar{G}_t$ happens. Then by the assumption that $\mathcal{F}_t$ happens and from the definition of $\bar{G}_t$, it follows that:
\begin{align*}
\Delta_{A_t}
& \leq 2 \sum_{e \in \tilde{A}_t} \sqrt{\frac{1.5 \log n}{T_{t - 1}(e)}} \\
& = 2 \sum_{e \in \tilde{A}_t \setminus S_t} \sqrt{\frac{1.5 \log n}{T_{t - 1}(e)}} +
2 \sum_{e \in S_t} \sqrt{\frac{1.5 \log n}{T_{t - 1}(e)}} \\
& < 2 \underbrace{|\tilde{A}_t \setminus S_t|}_{\leq K}
\sqrt{\frac{1.5 \log n}{\alpha K^2 \frac{6}{\Delta_{A_t}^2} \log n}} + {} \\
& \quad\,\, 2 \underbrace{|S_t|}_{\leq d}
\sqrt{\frac{1.5 \log n}{\frac{\alpha d^2}{(\sqrt{\alpha} - 1)^2} \frac{6}{\Delta_{A_t}^2} \log n}} \\
& \leq \frac{\Delta_{A_t}}{\sqrt{\alpha}} + \frac{\Delta_{A_t} (\sqrt{\alpha} - 1)}{\sqrt{\alpha}}
= \Delta_{A_t}\,.
\end{align*}
This is clearly a contradiction. Therefore, event $\bar{G}_t$ cannot happen; and either $G_{1, t}$ or $G_{2, t}$ happens.
\end{proof}
Now we are ready to prove Theorem~\ref{thm:K43 one gap}.
\begin{proof}
A detailed proof is in Appendix~\ref{sec:proof K43 one gap}. The key idea is to bound the number of times that events $G_{1, t}$ and $G_{2, t}$ happen in $n$ steps. Based on these bounds, the regret associated with both events is bounded as:
\begin{align*}
\hat{R}(n) \leq
\left(\frac{\alpha}{d} K^2 + \frac{\alpha d^2}{(\sqrt{\alpha} - 1)^2}\right) L \frac{6}{\Delta} \log n\,.
\end{align*}
Finally, we choose $\alpha = 4$ and $d = K^\frac{2}{3}$, and substitute the above upper bound into inequality~\eqref{eq:regret bound}.
\end{proof}
Theorem~\ref{thm:K43 one gap} can be generalized to the problems with different gaps. Let $\Delta_{e, \min}$ be the minimum gap of any suboptimal solution that contains item $e \in \tilde{E}$:
\begin{align}
\Delta_{e, \min}
& = \min_{A \in \Theta: e \in A, \Delta_A > 0} \Delta_A \label{eq:min gap} \\
& = f(A^\ast, \bar{w}) - \max_{A \in \Theta: e \in A, \Delta_A > 0} f(A, \bar{w})\,, \nonumber
\end{align}
where $\tilde{E} = E \setminus A^\ast$ is the set of \emph{subptimal items}, the items that do not appear in the optimal solution. Then the regret of ${\tt CombUCB1}$ is bounded as follows.
\begin{theorem}
\label{thm:K43} In any $(L, K)$ stochastic combinatorial semi-bandit, the regret of ${\tt CombUCB1}$ is bounded as:
\begin{align*}
R(n) \leq \sum_{e \in \tilde{E}} K^\frac{4}{3} \frac{96}{\Delta_{e, \min}} \log n +
\left(\frac{\pi^2}{3} + 1\right) K L\,.
\end{align*}
\end{theorem}
\begin{proof}
A detailed proof is in Appendix~\ref{sec:proof K43}. The key idea is to define item-specific variants of events $G_{1, t}$ and $G_{2, t}$, $G_{e, 1, t}$ and $G_{e, 2, t}$; and associate $\frac{\Delta_{A_t}}{d}$ and $\Delta_{A_t}$ regret with $G_{e, 1, t}$ and $G_{e, 2, t}$, respectively. Then, for each item $e$, we order the events from the largest gap to the smallest, and show that the total regret is bounded as:
\begin{align*}
\hat{R}(n) < \sum_{e \in \tilde{E}}
\left(\frac{\alpha}{d} K^2 + \frac{\alpha d^2}{(\sqrt{\alpha} - 1)^2}\right)
\frac{12}{\Delta_{e, \min}} \log n\,.
\end{align*}
Finally, we choose $\alpha = 4$ and $d = K^\frac{2}{3}$, and substitute the above upper bound into inequality~\eqref{eq:regret bound}.
\end{proof}
\section{$O(K)$ Upper Bounds}
\label{sec:K upper bounds}
In this section, we improve on the results in Section~\ref{sec:K43 upper bounds} and derive $O(K L (1 / \Delta) \log n)$ upper bounds on the $n$-step regret of ${\tt CombUCB1}$. In Theorem~\ref{thm:K one gap}, we assume that the gaps of all suboptimal solutions are identical. In Theorem~\ref{thm:K}, we relax this assumption.
The key step in our analysis is that we define a cascade of infinitely-many mutually-exclusive events and then bound the number of times that these events happen when a suboptimal solution is chosen. The events are parametrized by two decreasing sequences of constants:
\begin{align}
1 = \beta_0 > \beta_1 & > \beta_2 > \ldots > \beta_k > \ldots \label{eq:condition 1} \\
\alpha_1 & > \alpha_2 > \ldots > \alpha_k > \ldots \label{eq:condition 2}
\end{align}
such that $\lim_{i \to \infty} \alpha_i = \lim_{i \to \infty} \beta_i = 0$. We define:
\begin{align*}
m_{i, t} = \alpha_i \frac{K^2}{\Delta_{A_t}^2} \log n
\end{align*}
and assume that $m_{i, t} = \infty$ when $\Delta_{A_t} = 0$. The events at time $t$ are defined as:
\begin{align}
G_{1, t} = \{
& \text{at least $\beta_1 K$ items in $\tilde{A}_t$ were observed} \label{eq:event i} \\
& \text{at most $m_{1, t}$ times}\}\,, \nonumber \\
G_{2, t} = \{
& \text{less than $\beta_1 K$ items in $\tilde{A}_t$ were observed} \nonumber \\
& \text{at most $m_{1, t}$ times}, \nonumber \\
& \text{at least $\beta_2 K$ items in $\tilde{A}_t$ were observed} \nonumber \\
& \text{at most $m_{2, t}$ times}\}\,, \nonumber \\
& \vdots \nonumber \\
G_{i, t} = \{
& \text{less than $\beta_1 K$ items in $\tilde{A}_t$ were observed} \nonumber \\
& \text{at most $m_{1, t}$ times}, \nonumber \\
& \dots, \nonumber \\
& \text{less than $\beta_{i - 1} K$ items in $\tilde{A}_t$ were observed} \nonumber \\
& \text{at most $m_{i - 1, t}$ times}, \nonumber \\
& \text{at least $\beta_i K$ items in $\tilde{A}_t$ were observed} \nonumber \\
& \text{at most $m_{i, t}$ times}\}\,, \nonumber \\
& \vdots \nonumber
\end{align}
The following lemma establishes a sufficient condition under which events $G_{i, t}$ are exhaustive. This is the key step in the proofs in this section.
\begin{lemma}
\label{lem:complete system} Let $(\alpha_i)$ and $(\beta_i)$ be defined as in \eqref{eq:condition 1} and \eqref{eq:condition 2}, respectively; and let:
\begin{align}
\sqrt{6} \sum_{i = 1}^\infty \frac{\beta_{i - 1} - \beta_i}{\sqrt{\alpha_i}} \leq 1\,.
\label{eq:condition 3}
\end{align}
Let event $\mathcal{F}_t$ happen. Then event $G_{i, t}$ happens for some $i$.
\end{lemma}
\begin{proof}
We fix $t$ such that $\Delta_{A_t} > 0$. Because $t$ is fixed, we use shorthands $G_i = G_{i, t}$ and $m_i = m_{i, t}$. Let:
\begin{align*}
S_i = \cset{e \in \tilde{A}_t}{T_{t - 1}(e) \leq m_i}
\end{align*}
be the set of items in $\tilde{A}_t$ that are not observed ``sufficiently often'' under event $G_i$. Then event $G_i$ can be written as:
\begin{align*}
\textstyle
G_i = \left(\bigcap_{j = 1}^{i - 1} \set{|S_j| < \beta_j K}\right) \cap \set{|S_i| \geq \beta_i K}\,.
\end{align*}
As in Lemma~\ref{lem:K43 events}, we prove that event $G_i$ happens for some $i$ by showing that the event that none of our events happen cannot happen. Note that this event can be written as:
\begin{align*}
\bar{G}
& = \overline{\bigcup_{i = 1}^\infty G_i} \\
& = \bigcap_{i = 1}^\infty \Bigg[\Bigg(\bigcup_{j = 1}^{i - 1} \set{|S_j| \geq \beta_j K}\Bigg) \cup
\set{|S_i| < \beta_i K}\Bigg] \\
& = \bigcap_{i = 1}^\infty \set{|S_i| < \beta_i K}\,.
\end{align*}
Let $\bar{S_i} = \tilde{A}_t \setminus S_i$ and $S_0 = \tilde{A}_t$. Then by the definitions of $\bar{S_i}$ and $S_i$, $\bar{S}_{i - 1} \subseteq \bar{S}_i$ for all $i > 0$. Furthermore, note that $\lim_{i \to \infty} m_i = 0$. So there must exist an integer $j$ such that $\bar{S}_i = \tilde{A}_t$ for all $i > j$, and $\tilde{A}_t = \bigcup_{i = 1}^\infty (\bar{S}_i \setminus \bar{S}_{i - 1})$. Finally, by the definition of $\bar{S}_i$, $T_{t - 1}(e) > m_i$ for all $e \in \bar{S}_i$. Now suppose that event $\bar{G}$ happens. Then:
\begin{align*}
\sum_{e \in \tilde{A}_t} \frac{1}{\sqrt{T_{t - 1}(e)}}
& < \sum_{i = 1}^\infty \sum_{e \in \bar{S}_i \setminus \bar{S}_{i - 1}} \frac{1}{\sqrt{m_i}} \\
& = \sum_{i = 1}^\infty \frac{|\bar{S}_i \setminus \bar{S}_{i - 1}|}{\sqrt{m_i}} \\
& < \sum_{i = 1}^\infty \frac{(\beta_{i - 1} - \beta_i) K}{\sqrt{m_i}}\,,
\end{align*}
where the last inequality is due to Lemma~\ref{lem:little helper} (Appendix~\ref{sec:lemmas}). In addition, let event $\mathcal{F}_t$ happen. Then:
\begin{align*}
\Delta_{A_t}
& \leq 2 \sum_{e \in \tilde{A}_t} \sqrt{\frac{1.5 \log n}{T_{t - 1}(e)}} \\
& < \sqrt{6 \log n} \sum_{i = 1}^\infty \frac{(\beta_{i - 1} - \beta_i) K}{\sqrt{m_i}} \\
& = \Delta_{A_t} \sqrt{6} \sum_{i = 1}^\infty \frac{\beta_{i - 1} - \beta_i}{\sqrt{\alpha_i}}
\leq \Delta_{A_t}\,,
\end{align*}
where the last inequality is due to our assumption in \eqref{eq:condition 3}. The above is clearly a contradiction. As a result, $\bar{G}$ cannot happen, and event $G_i$ must happen for some $i$.
\end{proof}
\begin{theorem}
\label{thm:K one gap} In any $(L, K, \Delta)$ stochastic combinatorial semi-bandit where $\Delta_{A} = \Delta$ for all suboptimal solutions $A$, the regret of ${\tt CombUCB1}$ is bounded as:
\begin{align*}
R(n) \leq K L \frac{267}{\Delta} \log n + \left(\frac{\pi^2}{3} + 1\right) K L\,.
\end{align*}
\end{theorem}
\begin{proof}
A detailed proof is in Appendix~\ref{sec:proof K one gap}. The key idea is to bound the number of times that event $G_{i, t}$ happens in $n$ steps for any $i$. Based on this bound, the regret due to all events $G_{i, t}$ is bounded as:
\begin{align*}
\hat{R}(n) \leq
K L \frac{1}{\Delta} \left[\sum_{i = 1}^\infty \frac{\alpha_i}{\beta_i}\right] \log n\,,
\end{align*}
where $\hat{R}(n)$ is defined in (\ref{eq:Rhat}). Finally, we choose $(\alpha_i)$ and $(\beta_i)$, and apply the above upper bound in inequality~\eqref{eq:regret bound}.
\end{proof}
Now we generalize Theorem~\ref{thm:K one gap} to arbitrary gaps.
\begin{theorem}
\label{thm:K} In any $(L, K)$ stochastic combinatorial semi-bandit, the regret of ${\tt CombUCB1}$ is bounded as:
\begin{align*}
R(n) \leq \sum_{e \in \tilde{E}} K \frac{534}{\Delta_{e, \min}} \log n +
\left(\frac{\pi^2}{3} + 1\right) K L\,,
\end{align*}
where $\Delta_{e, \min}$ is the minimum gap of suboptimal solutions that contain item $e$, which is defined in \eqref{eq:min gap}.
\end{theorem}
\begin{proof}
A detailed proof is in Appendix~\ref{sec:proof K}. The key idea is to introduce item-specific variants of events $G_{i, t}$, $G_{e, i, t}$, and associate $\frac{\Delta_{A_t}}{\beta_i K}$ regret with each of these events. Then, for each item $e$, we order the events from the largest gap to the smallest, and show that the total regret is bounded as:
\begin{align*}
\hat{R}(n) <
\sum_{e \in \tilde{E}} K \frac{2}{\Delta_{e, \min}}
\left[\sum_{i = 1}^\infty \frac{\alpha_i}{\beta_i}\right] \log n\,,
\end{align*}
where $\hat{R}(n)$ is defined in (\ref{eq:Rhat}). Finally, we choose $(\alpha_i)$ and $(\beta_i)$, and apply the above upper bound in inequality~\eqref{eq:regret bound}.
\end{proof}
We also prove a gap-free bound.
\begin{theorem}
\label{thm:gap-free} In any $(L, K)$ stochastic combinatorial semi-bandit, the regret of ${\tt CombUCB1}$ is bounded as:
\begin{align*}
R(n) \leq 47 \sqrt{K L n \log n} + \left(\frac{\pi^2}{3} + 1\right) K L\,.
\end{align*}
\end{theorem}
\begin{proof}
The proof is in Appendix~\ref{sec:proof gap-free}. The key idea is to decompose the regret of ${\tt CombUCB1}$ into two parts, where the gaps are larger than $\epsilon$ and at most $\epsilon$. We analyze each part separately and then set $\epsilon$ to get the desired result.
\end{proof}
\section{Related Work}
\label{sec:related work}
Gai \emph{et al.}~\cite{gai12combinatorial} proposed ${\tt CombUCB1}$ and analyzed it. Chen \emph{et al.}~\cite{chen13combinatorial} derived a $O(K^2 L (1 / \Delta) \log n)$ upper bound on the $n$-step regret of ${\tt CombUCB1}$. In this paper, we show that the regret of ${\tt CombUCB1}$ is $O(K L (1 / \Delta) \log n)$, a factor of $K$ improvement over the upper bound of Chen \emph{et al.}~\cite{chen13combinatorial}. This upper bound is tight. We also prove a gap-free upper bound and show that it is nearly tight.
COMBAND \cite{cesabianchi12combinatorial}, online stochastic mirror descent (OSMD) \cite{audibert14regret}, and follow-the-perturbed-leader (FPL) with geometric resampling \cite{neu13efficient} are three recently proposed algorithms for adversarial combinatorial semi-bandits. In general, OSMD achieves optimal regret but is not guaranteed to be computationally efficient, in the same sense as in Section~\ref{sec:discussion}. FPL does not achieve optimal regret but is computationally efficient. It is an open problem whether adversarial combinatorial semi-bandits can be solved both computationally and sample efficiently. In this paper, we close this problem in the stochastic setting.
Matroid and polymatroid bandits \cite{kveton14matroid,kveton14learning} are instances of stochastic combinatorial semi-bandits. The $n$-step regret of ${\tt CombUCB1}$ in these problems is $O(L (1 / \Delta) \log n)$, a factor of $K$ smaller than is suggested by our $O(K L (1 / \Delta) \log n)$ upper bound. However, we note that the bound of Kveton \emph{et al.}~\cite{kveton14matroid,kveton14learning} is less general, as it applies only to matroids and polymatroids.
Our problem can be viewed as a linear bandit \cite{auer02using,abbasi-yadkori11improved}, where each solution $A$ is associated with an indicator vector ${\bf x} \in \set{0, 1}^E$ and the learning agent observes the weight of each non-zero entry of ${\bf x}$. This feedback model is clearly more informative than that in linear bandits, where the learning agent observes just the sum of the weights. Therefore, our learning problem has lower sample complexity. In particular, note that our $\Omega(\sqrt{K L n})$ lower bound (Proposition~\ref{prop:gap-free lower bound}) is $\sqrt{K}$ smaller than that of Audibert \emph{et al.}~\cite{audibert14regret} (Theorem 5) for combinatorial linear bandits. The bound of Audibert \emph{et al.}~\cite{audibert14regret} is proved for the adversarial setting. Nevertheless, it applies to our setting because the worst-case environment in the proof is stochastic.
Russo and Van Roy \cite{russo14information}, and Wen \emph{et al.}~\cite{wen14efficient}, derived upper bounds on the \emph{Bayes regret} of \emph{Thompson sampling} in stochastic combinatorial semi-bandits. These bounds have a similar form as our gap-free upper bound in Theorem~\ref{thm:gap-free}. However, they differ from our work in two aspects. First, the Bayes regret is a different performance metric from regret. From the frequentist perspective, it is a much weaker metric. Second, we also derive $O(\log n)$ upper bounds.
\section{Setting}
\label{sec:setting}
Formally, a \emph{stochastic combinatorial semi-bandit} is a tuple $B = (E, \Theta, P)$, where $E = \set{1, \dots, L}$ is a finite set of $L$ items, $\Theta \subseteq 2^E$ is a non-empty set of feasible subsets of $E$, and $P$ is a probability distribution over a \mbox{unit cube $[0, 1]^E$.} We borrow the terminology of combinatorial optimization and call $E$ the \emph{ground set}, $\Theta$ the \emph{feasible set}, and $A \in \Theta$ a \emph{solution}. The items in the ground set $E$ are associated with a vector of stochastic \emph{weights} $w \sim P$. The $e$-th entry of $w$, $w(e)$, is the weight of item $e$. The expected \mbox{weights of the} items are defined as $\bar{w} = \E{w}{w \sim P}$. The \emph{return} for choosing solution $A$ under the realization of the weights $w$ is:
\begin{align*}
f(A, w) = \sum_{e \in A} w(e)\,.
\end{align*}
The maximum number of chosen items is defined as $K = \max_{A \in \Theta} \abs{A}$.
Let $(w_t)_{t = 1}^n$ be an i.i.d. sequence of $n$ weights drawn from $P$. At time $t$, the learning agents chooses solution $A_t \in \Theta$ based on its observations of the weights up to time $t$, gains $f(A_t, w_t)$, and observes the weights of all chosen items at time $t$, $\cset{(e, w_t(e))}{e \in A_t}$. The learning agent interacts with the environment $n$ times and its goal is to maximize its expected cumulative reward over this time. If the agent knew $P$ a priori, the optimal action would be to choose the \emph{optimal solution}\footnote{For simplicity of exposition, we assume that the optimal solution is unique.}:
\begin{align*}
\textstyle
A^\ast = \argmax_{A \in \Theta} f(A, \bar{w})
\end{align*}
at all steps $t$. The quality of the agent's policy is measured by its \emph{expected cumulative regret}:
\begin{align*}
R(n) = \EE{\sum_{t = 1}^n R(A_t, w_t)}\,,
\end{align*}
where $R(A_t, w_t) = f(A^\ast, w_t) - f(A_t, w_t)$ is the regret of the agent at time $t$.
\begin{algorithm}[t]
\caption{${\tt CombUCB1}$ for stochastic combinatorial semi-bandits.}
\label{alg:ucb1}
\begin{algorithmic}
\STATE {\bf Input:} Feasible set $\Theta$
\STATE
\STATE // Initialization
\STATE $(\hat{w}_1, t_0) \gets {\tt Init}(\Theta)$
\STATE $T_{t_0 - 1}(e) \gets 1 \hspace{1.815in} \forall e \in E$
\STATE
\FORALL{$t = t_0, \dots, n$}
\STATE // Compute UCBs
\STATE $U_t(e) \gets \hat{w}_{T_{t - 1}(e)}(e) + c_{t - 1, T_{t - 1}(e)} \hspace{0.455in} \forall e \in E$
\STATE \vspace{-0.08in}
\STATE // Solve the optimization problem
\STATE $A_t \gets \argmax_{A \in \Theta} f(A, U_t)$
\STATE \vspace{-0.08in}
\STATE // Observe the weights of chosen items
\STATE Observe $\cset{(e, w_t(e))}{e \in A_t}$, where $w_t \sim P$
\STATE \vspace{-0.08in}
\STATE // Update statistics
\STATE $T_t(e) \gets T_{t - 1}(e) \hspace{1.5in} \forall e \in E$
\STATE $T_t(e) \gets T_t(e) + 1 \hspace{1.405in} \forall e \in {A_t}$
\STATE $\displaystyle \hat{w}_{T_t(e)}(e) \gets
\frac{T_{t - 1}(e) \hat{w}_{T_{t - 1}(e)}(e) + w_t(e)}{T_t(e)} \hspace{0.1in} \forall e \in {A_t}$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[t]
\caption{${\tt Init}$: Initialization of ${\tt CombUCB1}$.}
\label{alg:initialization}
\begin{algorithmic}
\STATE {\bf Input:} Feasible set $\Theta$
\STATE
\STATE $\hat{w}(e) \gets 0 \hspace{2.04in} \forall e \in E$
\STATE $u(e) \gets 1 \hspace{2.065in} \forall e \in E$
\STATE $t \gets 1$
\WHILE{$(\exists e \in E: u(e) = 1)$}
\STATE $A_t \gets \argmax_{A \in \Theta} f(A, u)$
\STATE Observe $\cset{(e, w_t(e))}{e \in A_t}$, where $w_t \sim P$
\FORALL{$e \in A_t$}
\STATE $\hat{w}(e) \gets w_t(e)$
\STATE $u(e) \gets 0$
\ENDFOR
\STATE $t \gets t + 1$
\ENDWHILE
\STATE
\STATE {\bf Output:}
\STATE \quad Weight vector $\hat{w}$
\STATE \quad First non-initialization step $t$
\end{algorithmic}
\end{algorithm}
\subsection{Algorithm}
\label{sec:algorithm}
Gai \emph{et al.}~\cite{gai12combinatorial} proposed a simple algorithm for stochastic combinatorial semi-bandits. The algorithm is motivated by ${\tt UCB1}$ \cite{auer02finitetime} and therefore we call it ${\tt CombUCB1}$. At each time $t$, ${\tt CombUCB1}$ consists of three steps. First, it computes the \emph{upper confidence bound (UCB)} on the expected weight of each item $e$:
\begin{align}
U_t(e) = \hat{w}_{T_{t - 1}(e)}(e) + c_{t - 1, T_{t - 1}(e)}\,,
\label{eq:UCB}
\end{align}
where $\hat{w}_s(e)$ is the average of $s$ observed weights of item $e$, $T_t(e)$ is the number of times that item $e$ is observed in $t$ steps, and:
\begin{align}
c_{t, s} = \sqrt{\frac{1.5 \log t}{s}}
\label{eq:confidence radius}
\end{align}
is the radius of a confidence interval around $\hat{w}_s(e)$ at time $t$ such that $\bar{w}(e) \in [\hat{w}_s(e) - c_{t, s}, \hat{w}_s(e) + c_{t, s}]$ holds with high probability. Second, ${\tt CombUCB1}$ calls the \emph{optimization oracle} to solve the combinatorial problem on the UCBs:
\begin{align*}
\textstyle
A_t = \argmax_{A \in \Theta} f(A, U_t)\,.
\end{align*}
Finally, ${\tt CombUCB1}$ chooses $A_t$, observes the weights of all chosen items, and updates the estimates of $\bar{w}(e)$ for these items. The pseudocode of ${\tt CombUCB1}$ is in Algorithm~\ref{alg:ucb1}.
\subsection{Initialization}
\label{sec:initialization}
${\tt CombUCB1}$ is initialized by calling procedure ${\tt Init}$ (Algorithm~\ref{alg:initialization}), which returns two variables. The first variable is a weight vector $\hat{w} \in [0, 1]^E$, where $\hat{w}(e)$ is a single observation from the $e$-th marginal of $P$. The second variable is the number of initialization steps plus one.
The weight vector $\hat{w}$ is computed as follows. ${\tt Init}$ repeatedly calls the oracle $A_t = \argmax_{A \in \Theta} f(A, u)$ on a vector of auxiliary weights $u \in \set{0, 1}^E$, which are initialized to ones. When item $e$ is observed, we set the weight $\hat{w}(e)$ to the observed weight of the item and $u(e)$ to zero. ${\tt Init}$ terminates when $u(e) = 0$ for all items $e$. Without loss of generality, let's assume that each item $e$ is contained in at least one feasible solution. Then ${\tt Init}$ is guaranteed to terminate in at most $L$ iterations, because at least one entry of $u$ changes from one to zero after each call of the optimization oracle.
|
\section{Introduction}
The Galilean group $\Gal(n)$ is the Lie group of transformations between
reference frames in Newtonian mechanics in $n$ dimensional space. It is
generated by spatial rotations, translations in space and time, and
boosts, which correspond to changes in velocity (see \cite{mmcm}).
Elements of $\Gal(n)$
can be represented as $(n+2)\times (n+2)$ matrices
\begin{align}
\label{GalGroup}
\left(\begin{array}{c|c}
\mathlarger{\mathlarger{\rho}} & \begin{smallmatrix}v_1 & x_1 \\\vdots &
\vdots\\v_n & x_n\end{smallmatrix} \\
\hline
0 & \begin{smallmatrix} 1 &x_0\\0 & 1\end{smallmatrix}
\end{array}\right)
\end{align}
where $\rho$ is an element of $\Orth(n)$,
$\left(\begin{smallmatrix}x_1\\\vdots \\
x_n\end{smallmatrix}\right)$ the spatial translation, $x_0$ the time
shift, and $\left(\begin{smallmatrix}v_1\\ \vdots
\\ v_n\end{smallmatrix}\right)$
is the boost.
To get the action on $n+1$-dimensional spacetime, we first identify
affine $(n+1)$-space with the plane $x_{n+2} = 1$ in $\ensuremath{\mathbb{R}}^{n+2}$.
Then the action on a point $(X_1,\ldots,X_n,T) \in \ensuremath{\mathbb{R}}^{n+1}$ is:
\[
\left(\begin{array}{c|c}
\mathlarger{\mathlarger{\rho}} & \begin{smallmatrix}v_1 & x_1 \\\vdots &
\vdots\\v_n & x_n \end{smallmatrix}\\
\hline
0 & \begin{smallmatrix} 1 &x_0\\0 & 1\end{smallmatrix}
\end{array}\right)\left(\begin{smallmatrix}
X_1 \\
\vdots \\
X_n \\
T \\
1
\end{smallmatrix}\right)
= \left( \begin{smallmatrix}
X_1' \\
\vdots \\
X_n' \\
T' \\
1
\end{smallmatrix}\right)
\]
Taking derivatives, the Lie algebra $\ensuremath{\mathfrak{gal}}(n)$ of $\Gal(n)$ is identified with
the set of matrices of the form
\begin{align}
\label{GalAlgebra}
\left(\begin{array}{c|c}
\mathlarger{\mathlarger{K}} & \begin{smallmatrix}v_1 & x_1 \\\vdots & \vdots\\v_n & x_n\end{smallmatrix}\\
\hline
0 & \begin{smallmatrix} 0 &x_0\\0 & 0\end{smallmatrix}
\end{array}\right)
\end{align}
where $K$ is an element of $\so(n)$.
The Galilean group is an important example of a {\em contraction}.
Given a Lie algebra $\ensuremath{\mathfrak{g}} = (V,[\cdot,\cdot])$, and a subalgebra $\ensuremath{\mathfrak{h}} =
(U,[\cdot,\cdot])$, choose a complementary subspace $W$ so that
$V = U \oplus W$.
For $\epsilon > 0$, let $T_\epsilon$ be the linear operator\[
T_\epsilon (u + w) = u + \epsilon w \; \; \; \;(u \in U, w \in W)
\]
One can define a new bracket $[\cdot,\cdot]_\epsilon$ on $V$ by
\[
[X,Y]_\epsilon = T_\epsilon^{-1}[T_\epsilon X,T_\epsilon Y]
\] The Lie algebra
$(V,[\cdot,\cdot]_\epsilon)$ is isomorphic to $\ensuremath{\mathfrak{g}}$. As long as $\ensuremath{\mathfrak{h}}$ is a
subalgebra, we may define $[\cdot,\cdot]_0 = \lim\limits_{\epsilon
\rightarrow 0} [\cdot,\cdot]_\epsilon$. The resulting Lie algebra
$\ensuremath{\mathfrak{g}}_0 = (V,[\cdot,\cdot]_0)$ is in general not isomorphic to $\ensuremath{\mathfrak{g}}$, and is
known as an In\"on\"u-Wigner contraction \cite{gilmore}. Note that the
subalgebra $\ensuremath{\mathfrak{h}}_0 = (U, [\cdot,\cdot]_0)$ is isomorphic to $\ensuremath{\mathfrak{h}}$, and that
$(W,[\cdot,\cdot]_0)$ is abelian. The contraction process can be seen
as `flattening' $W$.
The Galilean group is a contraction of the
Poincar\'e group, and this reflects the fact that special relativity
becomes Newtonian mechanics in the limit as the speed of light goes to
infinity. The Poincar\'e group is itself a contraction, whose bi-invariant
differential operators were found in \cite{takiff} and \cite{fgonz88}.
The universal enveloping algebra $\ensuremath{\mathcal{U}}(\ensuremath{\mathfrak{g}})$ of a Lie algebra can be
identified with the
algebra of left-invariant differential operators on the corresponding Lie
group $\mathbf{D}(G)$. If $S(\ensuremath{\mathfrak{g}})$ is the symmetric algebra on $\ensuremath{\mathfrak{g}}$, there
is a linear bijection $\lambda : S(\ensuremath{\mathfrak{g}}) \rightarrow \mathbf{D}(G)$ defined by
\[
(\lambda(P)f)(g) = P(\partial_1,\ldots,\partial_n)f(g \exp(t_1X_1 +
\dots + t_nX_n))
\]
where $\partial_i = \partialby{}{t_i}$, and $X_1,\ldots,X_n$ is a
basis of $\ensuremath{\mathfrak{g}}$. \cite[ch II, Theorem 4.3]{helgGGA}
This bijection commutes with the adjoint action of $G$ on $S(\ensuremath{\mathfrak{g}})$ and
$\mathbf{D}(G)$, and so we have, for $P \in S(\ensuremath{\mathfrak{g}})^{G}$: \[
\Ad(g)\lambda(P) = \lambda(\Ad(g)P) = \lambda(P)
\]
which shows that the image of an $\Ad(G)$ invariant polynomial is itself
$\Ad(G)$ invariant. Since the $\Ad(G)$ invariant elements of $\ensuremath{\mathcal{U}}(\ensuremath{\mathfrak{g}})$ are
exactly $\ensuremath{\mathfrak{z}}(\ensuremath{\mathcal{U}}(\ensuremath{\mathfrak{g}}))$, we have identified $S(\ensuremath{\mathfrak{g}})^{G}$ with $\ensuremath{\mathfrak{z}}(\ensuremath{\mathcal{U}}(\ensuremath{\mathfrak{g}}))$.
Unfortunately, while $\lambda$ is a linear bijection, it isn't an isomorphism
of algebras, and so in general, $\lambda(PQ) \neq \lambda(P)\lambda(Q)$.
However we do have that: \begin{align}
\label{highdeg}
\deg(\lambda(PQ) - \lambda(P)\lambda(Q)) < \deg(\lambda(PQ))
\end{align}
This allows us to show inductively that if $\{P_1, \dots,P_m\}$ generate
$S(\ensuremath{\mathfrak{g}})^G$, then $\{\lambda(P_1),\ldots,\lambda(P_m)\}$ generate $\ensuremath{\mathfrak{z}}(\ensuremath{\mathcal{U}}(\ensuremath{\mathfrak{g}}))$.
The degree zero case is trivial.
If $D \in \ensuremath{\mathfrak{z}}(\ensuremath{\mathcal{U}}(\ensuremath{\mathfrak{g}}))$, then we can write \[
D = \lambda(q(P_1,\ldots,P_m))
\]
for some polynomial $q$. Then \[
D - q(\lambda(P_1),\ldots,\lambda(P_m))
\]
is a central element whose degree is less than $\deg(D)$ by \eqref{highdeg}.
By induction, $\ensuremath{\mathfrak{z}}(\ensuremath{\mathcal{U}}(\ensuremath{\mathfrak{g}}))$ is generated by
$\{\lambda(P_1,\ldots,\lambda(P_m)\}$.
Elements of $S(\ensuremath{\mathfrak{g}})$ can be viewed as polynomials on $\ensuremath{\mathfrak{g}}^*$, the dual
space of $\ensuremath{\mathfrak{g}}$. With this identification, $\Ad(g)P(X^*) =
P(\Ad^*(g^{-1})X^*)$,
where $\Ad^*$ is the coadjoint representation of $G$ on $\ensuremath{\mathfrak{g}}^*$.
A polynomial which is invariant under the adjoint representation will be
constant on coadjoint orbits. This allows us to use the tools of classical
invariant theory to solve the problem.
\section{Restriction of the Problem}
We would like to find a subspace $S$ of $\ensuremath{\mathfrak{g}}^*$ for which the set
$\{Ad^*(g)s | g \in G, s \in S\}$ is Zariski dense in $\ensuremath{\mathfrak{g}}^*$. In such a case,
any polynomial which is invariant under the coadjoint action is defined
by its values on that subspace. In particular, we would like to find
an $S$ which is {\em transversal} to the coadjoint orbits.
Since in general orbits
will intersect the transversal subspace multiple times, the restriction of
an invariant function will satisfy some further restrictions, namely that it be
invariant under the the action of the subgroup of $G$ which fixes the
subspace.
Any polynomial function on $\ensuremath{\mathfrak{g}}^*$ will restrict to a polynomial function
on a subspace, but the converse is not true. Therefore, after identifying
candidate polynomials, we must eliminate those which don't extend to a
polynomial function.
For a semisimple Lie group, the Killing form provides a nondegenerate inner
product with which we can identify the Lie algebra and its dual. Because the
Killing form is invariant with respect to the adjoint action, the coadjoint
action is essentially the same as the adjoint action. Unfortunately, $\Gal(n)$
does not have such a convenient inner product. Instead, we use the usual matrix
inner product $\dotp{A,B} = \tr(A^TB)$.
\section{The Coadjoint Action}
Recalling the matrix forms of the Galilean group and its Lie algebra given in
\eqref{GalGroup} and \eqref{GalAlgebra}, we must now describe the dual space
$\ensuremath{\mathfrak{gal}}^*$ and the coadjoint action of $\Gal(n)$ on $\ensuremath{\mathfrak{gal}}^*$.
To represent $\ensuremath{\mathfrak{gal}}(n)^*$, we use the usual matrix inner product
$A^*(B) = \tr(A^T B)$. We can then identify $\ensuremath{\mathfrak{gal}}(n)^*$ with
$\ensuremath{\mathbb{R}}^{(n+2)\times (n+2)}$ modulo $\ensuremath{\mathfrak{gal}}(n)^\perp$, where
\[
\ensuremath{\mathfrak{gal}}(n)^\perp = \left\{
\left(\begin{array}{c|c}
\mathlarger{\mathlarger{S}} & \begin{smallmatrix}0 & 0\\\vdots & \vdots\\
0 & 0\end{smallmatrix} \\ \hline
A & \begin{smallmatrix}a & 0 \\b & c\end{smallmatrix}
\end{array}\right) \mid
S \text{ symmetric}, A \in \ensuremath{\mathbb{R}}^{2\times n}, a,b,c \in \ensuremath{\mathbb{R}}
\right\}
\]
Using the fact that the trace is invariant under cyclic permutations,
\begin{align*}
(\Ad^*(g)A^*)(B) &= A^*(\Ad(g^{-1}B) \\
&= A^*(g^{-1}Bg) \\
&= \tr(A^Tg^{-1}Bg) \\
&= \tr(gA^Tg^{-1}B) \\
&= ((gA^Tg^{-1})^T)^*(B)
\end{align*}
Every element of $\Gal(n)$ can be written as $\tau\rho$, where
$\rho$ is a spatial rotation and $\tau$ is a space-time translation and boost.
Because of this, we may
consider the actions of rotations separately from boosts and translations.
If \[
\rho = \left(\begin{array}{c|c}
\rho & \begin{smallmatrix}0 & 0\end{smallmatrix} \\
\hline
0 & \begin{smallmatrix}1 & 0 \\ 0 & 1\end{smallmatrix}
\end{array}\right)
\]
and \begin{align}
\label{genelt}
A^* = \left(\begin{array}{c|c}
\mathlarger{\mathlarger{K^*}} & \begin{smallmatrix}v^*_1 & x^*_1 \\\vdots & \vdots\\v^*_n & x^*_n\end{smallmatrix}\\
\hline
0 & \begin{smallmatrix} 0 &x^*_0\\0 & 0\end{smallmatrix}
\end{array}\right)
\end{align}
then a straightforward calculation shows:
\[
\rho^{-1} = \left(\begin{array}{c|c}
\rho^{-1} & \begin{smallmatrix}0 & 0\end{smallmatrix} \\
\hline
0 & \begin{smallmatrix}1 & 0 \\ 0 & 1\end{smallmatrix}
\end{array}\right)
\]
and
\begin{align}
\label{rotate}
(\rho A^{*T}\rho^{-1})^T = \left(\begin{array}{c|c}
\rho K^*\rho^{-1} & \rho\smallmat{v_1^* & x_1^*\\ \vdots & \vdots \\
v_n^* & x_n^*} \\ \hline
0 & \begin{smallmatrix} 0 & x_0^* \\ 0 & 0\end{smallmatrix}
\end{array}\right)
\end{align}
For translations and boosts, \begin{align}
\label{transboost}
\begin{split}
\tau &= \left(\begin{array}{c|c}
I & \begin{smallmatrix}v_1 & x_1 \\ \vdots & \vdots \\ v_n & x_n\end{smallmatrix}
\\ \hline
0 & \begin{smallmatrix}1 & x_0 \\ 0 & 1\end{smallmatrix}
\end{array}\right) \\
\tau^{-1} &= \left(\begin{array}{c|c}
I & \begin{smallmatrix}-v_1 & v_1x_0 - x_1 \\ \vdots & \vdots \\
-v_n & v_nx_0 - x_n\end{smallmatrix} \\ \hline
0 & \begin{smallmatrix}1 & -x_0 \\ 0 & 1\end{smallmatrix}
\end{array}\right) \\
(\tau A^*\tau^{-1})^T &= \left(\begin{array}{c|c}
K^* + \smallmat{v_1^* & x_1^* \\ \vdots & \vdots \\v_n^* & x_n^*}
\smallmat{v_1 \dots v_n\\x_1 \dots x_n} &
\begin{smallmatrix}v_1^* + x_0x_1^* & x_1^* \\ \vdots & \vdots \\v_n^* + x_0x_n^* & x_n^*\end{smallmatrix} \\ \hline
* & \begin{smallmatrix}* & x_0^* + \sum v_ix_i^* \\ * & *\end{smallmatrix}
\end{array}\right)\\
&\equiv \left(\begin{array}{c|c}
K^* + \frac{1}{2}\smallmat{v_1^* & x_1^* \\ \vdots & \vdots \\v_n^* & x_n^*}
\smallmat{v_1 \dots v_n\\x_1 \dots x_n} -
\frac{1}{2} \smallmat{v_1 \dots v_n\\x_1 \dots x_n}^T
\smallmat{v_1^* & x_1^* \\ \vdots & \vdots \\v_n^* & x_n^*}^T &
\begin{smallmatrix}v_1^* + x_0x_1^* & x_1^* \\ \vdots & \vdots \\v_n^* + x_0x_n^* & x_n^*\end{smallmatrix} \\ \hline
0 & \begin{smallmatrix}0 & x_0^* + \sum v_ix_i^* \\ 0 & 0\end{smallmatrix}
\end{array}\right)
\end{split}
\end{align}
where the final congruence is modulo $\ensuremath{\mathfrak{gal}}(n)^\perp$
To find a transversal subspace of $\ensuremath{\mathfrak{gal}}(n)^*$, let $A^*$ be a generic
element written as as \eqref{genelt}. We would like to find a $g$ which can take
$(g^{-1}A^{*T}g)^T$ to our subspace.
First, use a rotation \eqref{rotate} to put $x^*$ in the form
$(A 0 \dots 0)^T$ and $v^*$
in the form $(C B 0 \dots 0)^T$. We can then use an $x_0$-only translation
element (equation \eqref{transboost} with $x_0 = -\frac{v_1^*}{x_1^*}$) to zero out the first element of $v^*$. Thus we can restrict our
attention to matrices of the form \[
\left(\begin{array}{c|c}
K^* & \begin{smallmatrix}0 & A\\ B & 0\\ 0 & 0 \\\vdots & \vdots\end{smallmatrix}
\\ \hline
0 & \begin{smallmatrix} 0 & x_0^* \\ 0 & 0\end{smallmatrix}
\end{array}\right)
\]
Let's now turn to the top-left quadrant. Referring again to \eqref{transboost},
\begin{align*}
\frac{1}{2}\smallmat{0 & A \\ B & 0 \\ 0 & 0 \\\vdots & \vdots}
\smallmat{v_1 & \dots & v_n \\ x_1 & \dots & x_n} -
\frac{1}{2}\left(\smallmat{0 & A \\ B & 0 \\ 0 & 0 \\\vdots & \vdots}
\smallmat{v_1 & \dots & v_n \\ x_1 & \dots & x_n}\right)^T \\
= \frac{1}{2}\smallmat{0 & Ax_2 - Bv_1 & Ax_3 & Ax_4 & \dots & Ax_n\\
Bv_1 - Ax_2 & 0 & Bv_3 & Bv_4 &\dots & Bv_n \\
-Ax_3 & -Bv_3 & 0 & 0 & \dots & 0\\
-Ax_4 & -Bv_4 & 0 & 0 & \dots & 0\\
\vdots &\vdots &\vdots &\vdots &\ddots & \vdots\\
-Ax_n & -Bv_n & 0 & 0 & \dots & 0}
\end{align*}
Appropriate choices for the $x_i$'s and $v_i$'s allow us to zero out the
topmost two rows and leftmost two columns of $K^*$. In particular, for $i \geq
3$, we set $x_i = -\frac{2K^*_{1i}}{A}$ and $v_i -\frac{2K^*_{2i}}{B}$. It is
important to note that there is a remaining degree of freedom when zeroing
$K^*_{12}$, as this will allow us to clear $x^*_0$ by setting $v_1 =
-\frac{x_0^*}{x_1^*}$ in \eqref{transboost}.
We are free to choose a rotation from the subgroup of $\Orth(n)$ which fixes
the first two coordinates. We can use this freedom to conjugate the upper-left
block to an element of a maximal torus in $\so(n-2)$.
Our transversal subspace is made of matrices of the form\[
\left(\begin{array}{c|c}
\begin{smallmatrix} 0 & 0 & \dots \\ 0 & 0 & \dots \\ \vdots & \vdots & K
\end{smallmatrix} &
\begin{smallmatrix} 0 & A \\ B & 0 \\ 0 & 0 \\ \vdots & \vdots
\end{smallmatrix} \\ \hline
0 & \begin{smallmatrix}0 & 0 \\ 0 & 0\end{smallmatrix}
\end{array}\right)
\]
where $A$ and $B$ are real numbers, and $K$ is an element of a $\so(n-2)$
maximal torus.
\section{Case $n = 1$}
When $n = 1$, the upper-left hand block is always zero.
We use $x_0$ to zero out the $v_1^*$ and $v_1$ to zero out $x_0^*$. Thus,
a generic element \[
\left(\begin{array}{c c c} 0 &v_1^* & x_1^* \\0 &0& x_0^*\\ 0 & 0 & 0
\end{array}\right)
\]
of $\ensuremath{\mathfrak{gal}}(1)^*$ can be conjugated to\[
\left(\begin{array}{c c c} 0 &0 & x_1^* \\0 &0& 0\\ 0 & 0 & 0
\end{array}\right)
\]
Since $\left(-1\right) \in \Orth(1)$, the algebra of $\Ad^*$-invariant
polynomials is generated by $X_1^2$
\section{Case $n = 2,3$}
\begin{theorem}
For $n \in \{2,3\}$, $S(\ensuremath{\mathfrak{gal}}(n))^{\Gal(n)}$ is generated by $\sum\limits_i
X_i^2$ and \linebreak
$\left(\sum\limits_i X_i^2\right)\left(\sum\limits_i V_i^2\right) -
\left(\sum\limits_i X_iV_i\right)^2$
\end{theorem}
\begin{proof}
When $n = 2$ or $3$, $K^*$ may not be zero, but
we can still zero out the upper-left block entirely. If \[\Xi =
\left(
\begin{array}{c|cc}
K^* &v^* & x^* \\ \hline
0 & 0 & t^* \\ 0 & 0 & 0\\
\end{array}\right)
\] is a matrix representing an element of $\ensuremath{\mathfrak{gal}}(n)^*$, then
Then there exists an element $g \in \Gal(n)$ such that
\[
\Ad(g)^*\Xi \equiv
\left(\begin{array}{c|cc}
0 & \begin{smallmatrix} 0 \\ B \end{smallmatrix} & \begin{smallmatrix} A \\ 0
\end{smallmatrix} \\ \hline 0 & 0 & 0\\
0 & 0 & 0 \\
\end{array}\right)
\]
where $A = \norm{x^*}$ and $B = \norm{\proj_{x^{*\perp}}(v^*)}$. Conjugating
by an element of $\Orth(n)$ can switch the signs of $A$ and $B$ independently,
so the
invariant polynomials on the transverse manifold are generated by $A^2$ and
$B^2$. Thus any polynomial $P$ which is invariant under the coadjoint
action of $\Gal(n)$ restricts to a polynomial $\overline{P} = F(A^2,B^2)$
defined on the transverse manifold.
Define polynomials $Q_1 : \ensuremath{\mathfrak{gal}}(n)^* \rightarrow
\ensuremath{\mathbb{R}}$ and $Q_2 : \ensuremath{\mathfrak{gal}}(n)^* \rightarrow \ensuremath{\mathbb{R}}$, which take an element of $\ensuremath{\mathfrak{gal}}(n)^*$
to the values of $A^2$ and $A^2B^2$ on the intersection of its $\Ad^*$ orbit
with the transversal submanifold. If $\Xi$ is a matrix representing
an element of $\ensuremath{\mathfrak{gal}}(n)^*$, we have:
\begin{align*}
Q_1(\Xi) &= \norm{x^*}^2 = \left(\sum\limits_iX_i^{2}\right)(\Xi) \\
Q_2(\Xi) &= \norm{x^*}^2\norm{\proj_{x^{*\perp}}v^*}^2 = \left(
\left(\sum\limits_i X_i^{2}\right)\left(\sum\limits_i V_i^{2}\right)
- \left(\sum\limits_i X_i V_i\right)^2\right)(\Xi)
\end{align*}
Because the restriction is injective on the space of $\Ad^*$ invariant
polynomials, we know that \begin{align}
P(\Xi) &= F\left(Q_1(\Xi),\frac{Q_2(\Xi)}{Q_1(\Xi)}\right) \\
\label{23Fbreakdown}
&= F_1(Q_1(\Xi),Q_2(\Xi)) + \frac{F_2(Q_1(\Xi),Q_2(\Xi))}{Q_1(\Xi)^\ell}
\end{align}
where $F_2$ is indivisible by its first argument.
While we are considering only real Lie algebras, if $P$ is a polynomial,
it must extend to a polynomial on the complexification of $\ensuremath{\mathfrak{gal}}(n)^*$. With
this in mind, consider \[
\Xi = \left(\begin{array}{cc|cc}
0 & 0 & 0 & i\\
0 & 0 & z & 1 \\
\hline
0 & 0 & 0 & 0\\
0 & 0 & 0 & 0
\end{array}\right)
\]
for $n = 2$ or
\[
\Xi = \left(\begin{array}{ccc|cc}
0 &0 & 0 & 0 & i\\
0 &0 & 0 & z & 1 \\
0 &0 & 0 & 0 & 0 \\
\hline
0 &0 & 0 & 0 & 0\\
0 &0 & 0 & 0 & 0
\end{array}\right)
\]
when $n = 3$. In either case, $Q_1(\Xi) = 0$, and $Q_2(\Xi) = z^2$.
\eqref{23Fbreakdown} then becomes \begin{align*}
P(\Xi) = F_1(0,z^2) + \frac{F_2(0,z^2)}{0^\ell}
\end{align*}
Since $P$ was assumed to be a polynomial, either $\ell = 0$, or $F_2(0,z^2) =
0$ for all $z \in \ensuremath{\mathbb{C}}$. Since $F_2$ was assumed to be indivisible by its first
element, this implies that either $\ell = 0$ or $F_2$ is the constant function
$0$. Therefore, $P$ is a polynomial in $Q_1$ and $Q_2$
\end{proof}
\section{Case $n > 3$}
\begin{theorem}
When $n > 3$,
$S(\ensuremath{\mathfrak{gal}}(n))^{\Gal(n)}$ is generated
by
the sums of determinants of $2k\times 2k$ submatrices formed by taking the
symmetric minors of \[
\left(\begin{array}{c|c}
\mathlarger{\mathlarger{K^*}} & \begin{smallmatrix}v^*_1 & x^*_1 \\\vdots & \vdots\\v^*_n & x^*_n\end{smallmatrix}\\
\hline
\begin{smallmatrix}-v^{*T}\\-x^{*T}\end{smallmatrix} &
\begin{smallmatrix} 0 &0\\0 & 0\end{smallmatrix}
\end{array}\right)
\] which include the last two rows and columns.
\end{theorem}
\begin{proof}
When $n >3$, the coadjoint action can no longer zero out the entire
upper-left block, only the uppermost two rows and leftmost two columns. The
transversal subspace $S$ is made of matrices of the form\[
\left(\begin{array}{c|c}
\begin{smallmatrix} 0 & 0 & \dots \\ 0 & 0 & \dots \\ \vdots & \vdots & K
\end{smallmatrix} &
\begin{smallmatrix} 0 & A \\ B & 0 \\ 0 & 0 \\ \vdots & \vdots
\end{smallmatrix} \\ \hline
0 & \begin{smallmatrix}0 & 0 \\ 0 & 0\end{smallmatrix}
\end{array}\right)
\]
where $K$ is an $\so(n-2)$ maximal torus.
Conjugating by elements of $\Orth(n)$ can change the signs of $A$ and
$B$ independently,
as well as permuting and changing signs of the elements of $K$.
The polynomials
on $S$ invariant under the subgroup of $\Gal(n)$ which fixes $S$ are generated
by $A^2$, $B^2$, and the coefficients of the characteristic polynomial of $K$
\cite{KNII}.
By
the same argument as in the $n = 2$ case, $B^2$ must be multiplied by $A^2$
to clear the denominator. This gives invariant polynomials $A^2 = \sum X_i^2$
and $A^2B^2 = (\sum X_i^2)(\sum V_i^2) - (\sum X_iV_i)^2$
We now turn to the polynomials which depend on the $K$ part. When conjugating
to $S$, one step was to zero out the rows and columns corresponding to
$x^*$ and $v^*$. This is equivalent to pre- and post- multiplying
by the matrix $P$,
the orthogonal projection to $\spn\{x^*,v^*\}^\perp$.
The resulting element of $\so(n-2)^*$ is subject to a $\Orth(n-2)$ action, and
it is well-known that the resulting invariant polynomials generated by
the coefficients of the characteristic polynomial. Thus we are looking for
$\charpoly(PK^*P)$
We will also make use of some facts about exterior algebras. Suppose
that $e_1,..,e_m$ and $f_1,\ldots,f_n$ are bases for vector spaces
$V_1$ and $V_2$ respectively.
If $A : V_1 \rightarrow V_2$ is a linear function, then define
$\bigwedge\nolimits^kA : \bigwedge\nolimits^kV_1 \rightarrow \bigwedge\nolimits^kV_2$ to be the map
$x_1\wedge x_2\wedge\cdots\wedge x_k \mapsto Ax_1\wedge Ax_2 \wedge \cdots
\wedge Ax_k$. Then the $(e_{i_1}\wedge\cdots\wedge e_{i_k},f_{j_1}\wedge\cdots
\wedge f_{j_k})$ element of the matrix of $\bigwedge\nolimits^k A$ is
the determinant of the $k\times k$ matrix formed by taking elements in rows
$i_1,\ldots,i_k$ and columns $j_1,\ldots,j_k$. In particular, the coefficient
of the $x^{n-k}$ term of $\charpoly(A)$ is $\tr(\bigwedge\nolimits^k A)$.
Let \[\omega = \frac{v\wedge x}{\norm{v\wedge x}} = \frac{1}{AB}v\wedge x\]
(the sign of $AB$ is ambiguous, WLOG we may assume it's positive),
and let
\[
K' = \left(\begin{array}{c|c}
\mathlarger{\mathlarger{K^*}} & \begin{smallmatrix}v^*_1 & x^*_1 \\\vdots & \vdots\\v^*_n & x^*_n\end{smallmatrix}\\
\hline
\begin{smallmatrix}-v^{*T}\\-x^{*T}\end{smallmatrix} &
\begin{smallmatrix} 0 &0\\0 & 0\end{smallmatrix}
\end{array}\right)
\]
Let $e_i$ be the vector (written vertically) with a $1$ in the $i$'th row
and zeros elsewhere.
If $y \in \spn\{e_1,\ldots,e_n\}^\perp$:\begin{align}
K'y = K^*y - (y\cdot v^*)e_{n+1} - (y\cdot x^*)e_{n+2}\end{align}
And we also have the following:
\begin{align}
\label{epart}
\left(\bigwedge\nolimits^2K'\right)e_{n+1}\wedge e_{n+2} = v^*\wedge x^* = AB\omega
\end{align} and
\begin{align}
\begin{split}
&\left(\bigwedge\nolimits^2K'\right)\omega \\ &= \frac{1}{AB}\left(\bigwedge\nolimits^2K'\right)v^*\wedge
x^* \\
&= \frac{1}{AB}(K^*v^* - (v^* \cdot v^*)e_{n+1} - (v^*\cdot x^*)e_{n+2})\wedge
(K^*x^* - (x^*\cdot v^*) e_{n+1} - (x^*\cdot x^*) e_{n+2}) \\
&= \frac{1}{AB}\left[K^*v^*\wedge(K^*x^* - (x^*\cdot v^*) e_{n+1} -
(x^*\cdot x^*) e_{n+2}) \right.\\
&+ (K^*v^* - (v^* \cdot v^*)e_{n+1} - (v^*\cdot x^*)e_{n+2})\wedge K^*x^* \\
&+\left((x^*\cdot x^*)(v^*\cdot v^*) - (x^*\cdot v^*)^2\right)e_{n+1}\wedge
e_{n+2}\left.\right]
\end{split}
\end{align}
In particular, note that the $e_{n+1}\wedge e_{n+2}$ term of
$(\bigwedge\nolimits^2K')\omega$ is $ABe_{n+1}\wedge e_{n+2}$.
Now consider any diagonal element $(\bigwedge\nolimits^{k+4}K')_{(I,I)}$ for which
$I = (i_1,\ldots,i_{k+2},n+1,n+2)$. By \eqref{epart},
$(\bigwedge\nolimits^{k+4}K')\alpha\wedge e_{n+1}\wedge e_{n+2} \in
\bigwedge\nolimits^{k+2}\ensuremath{\mathbb{R}}^{n+2}\wedge\omega$. Since we're only considering diagonal
elements, this means we can restrict our attention to the subspace
\[
\left(\bigwedge\nolimits^k\ensuremath{\mathbb{R}}^n\right)\wedge\omega\wedge e_{n+1}
\wedge e_{n+2}
\]
where by $\ensuremath{\mathbb{R}}^n$ refers to $\spn\{e_1,\ldots,e_n\} \subset \ensuremath{\mathbb{R}}^{n+2}$. Recall that
$P$ projects to $\spn\{x^*,v^*\}^\perp$. The we have, for $y_1,\ldots,y_k \in
\ensuremath{\mathbb{R}}^n$:
\begin{align}
\label{finalbit}
\begin{split}
&\left(\bigwedge\nolimits^{(k+4)}K'\right)y_1\wedge\cdots\wedge y_k\wedge\omega\wedge
e_{n+1}\wedge e_{n+2} \\
=& \left(\bigwedge\nolimits^{(k+4)}K'\right)Py_1\wedge\cdots\wedge Py_k\wedge\omega\wedge
e_{n+1}\wedge e_{n+2} \\
=& A^2B^2 K^*Py_1\wedge\cdots\wedge K^*Py_k\wedge\omega\wedge e_{n+1}\wedge
e_{n+2} \\
&\quad + \text{terms without $\omega\wedge e_{n+1}\wedge e_{n+2}$} \\
=& A^2B^2PK^*Py_1\wedge\cdots\wedge PK^*Py_k\wedge\omega\wedge e_{n+1}
\wedge e_{n+2} \\
&\quad + \text{terms without $\omega\wedge e_{n+1}\wedge e_{n+2}$}
\end{split}
\end{align}
By \eqref{finalbit}, \[
\sum\limits_{I = (i_1,\ldots,i_k+2,n+1,n+2)}
\left(\bigwedge\nolimits^{k+4}K'\right)_{(I,I)} =
A^2B^2\tr\left(\bigwedge\nolimits^{k}PK^*P\right)
\]
which implies that the sum of $(k+4)\times(k+4)$ subdeterminants
which include the last two
rows and columns is the $x^{n -k}$ coefficient of $A^2B^2\charpoly(PK^*P)$.
Because $PK^*P$ is skew-symmetric, the odd-$k$ terms will be zero.
To summarize, we have the polynomials $Q_1 = x^*\cdot x^*$, $Q_2 =
(v^*\cdot v^*)(x^*\cdot x^*) - (x^*\cdot v^*)^2$, and
$Q_3,\ldots,Q_{2+\floor{n/2}}$, which are $Q_2$ times the characteristic
polynomial coefficients of $PK^*P$
Finally, we must show that the $Q_i$ generate the invariant polynomials.
Suppose that $F(X)$ is an invariant polynomial not generated by the $Q_i$.
Restricting to the transversal subspace $S$, $\overline{F}$ is generated
by the restrictions of $Q_1, \frac{Q_2}{Q_1}, \frac{Q_3}{Q_2},\ldots,
\frac{Q_{2+\floor{n/2}}}{Q_2}$. By the injectivity of restriction to $S$,
this means that
\begin{align}
\label{ratbreakdown}\begin{split}
F(X) &= F'\left(Q_1(X), \frac{Q_2}{Q_1}(X), \frac{Q_3}{Q_2}(X),\ldots,
\frac{Q_{2+\floor{n/2}}}{Q_2}(X)\right)\\
&= F''(Q_1(X),Q_2(X),\ldots,Q_{2+\floor{n/2}}(X)) +
\frac{\tilde{F}(Q_1(X),Q_2(X),\ldots,Q_{2+\floor{n/2}}(X))}
{Q_1(X)^kQ_2(X)^\ell}
\end{split}
\end{align}
where $\tilde{F}$ is assumed not to be divisible by its first or second
arguments.
To show that $F$ is generated by the $Q_i$, we must show that $k = \ell = 0$.
While all of the preceeding work was done over $\ensuremath{\mathbb{R}}$, any polynomial extends
to a polynomial on $\ensuremath{\mathbb{C}}$. Consider matrices of the form\[
X_\Xi = \left(\begin{array}{cc|ccc|cc}
0&0&0 & \dots & 0 & 0 & 1\\
0&0&0& \dots & 0 & 1 & i\\
\hline
0&0&&& & 0 & 0\\
\vdots&\vdots& & \Xi & & \vdots & \vdots \\
0&0& & && 0 & 0\\ \hline
0&0&0 & \dots & 0 & 0 &0\\
0&0&0 & \dots & 0 & 0 &0
\end{array}\right)
\]
where $\Xi$ is an $(n-2)\times(n-2)$ skew-symmetric matrix.
For any $\Xi$, $Q_1(X_\Xi) = 0$ and $Q_2(X_\Xi) = 1$. For $3 \leq i \leq
2+\floor{n/2}$, we get the nonconstant
characteristic polynomial coefficients for
$\Xi$, which are known to be algebraically independent. Because of this
algebraic independence, and that $\tilde{F}$ was assumed not to be divisible
by its first argument, there exists a $\Xi$ such that the numerator of
\eqref{ratbreakdown} is nonzero, while $Q_1(X_\Xi)$ is zero. Therefore,
$k = 0$.
To show that $\ell = 0$, let \[
Y_\Xi = \left(\begin{array}{cc|ccc|cc}
0&0&0 & \dots & 0 & 0 & 1\\
0&0&0& \dots & 0 & 1 & 0\\
\hline
0&0&&& & i & 0\\
0 & 0 &&& & 0 & 0\\
\vdots&\vdots& & \Xi & & \vdots & \vdots \\
0&0& & && 0 & 0\\ \hline
0&0&0 & \dots & 0 & 0 &0\\
0&0&0 & \dots & 0 & 0 &0
\end{array}\right)
\]
For any $\Xi \in \so(n-2)$, $Q_1(Y_\Xi) = 1$ and $Q_2(Y_\Xi) = 0$. For $j
\geq 3$, $Q_j(Y_\Xi)$ is the sum of all $2j\times2j$
subdeterminants of \[
Y'_\Xi = \left(\begin{array}{cc|ccc|cc}
0&0&0 & \dots & 0 & 0 & 1\\
0&0&0& \dots & 0 & 1 & 0\\
\hline
0&0&&& & i & 0\\
0 & 0 &&& & 0 & 0\\
\vdots&\vdots& & \Xi & & \vdots & \vdots \\
0&0& & && 0 & 0\\ \hline
0&-1&-i & \dots & 0 & 0 &0\\
-1&0&0 & \dots & 0 & 0 &0
\end{array}\right)
\]
which include the leftmost two columns and bottom two rows.
Consider a skew-symmetric submatrix of $Y'_\Xi$ which includes the last
two rows and columns. If the submatrix doesn't also include the first row
and column, its leftmost column, and thus its determinant, will be zero.
Similarly, any nonzero such minor must also include at least one of the
second and third rows.
If the submatrix includes the second row and column, the corresponding
minor is
\[
\det\left(\begin{array}{cc|ccc|cc}
0&0&0 & \dots & 0 & 0 & 1\\
0&0&0& \dots & 0 & 1 & 0\\
\hline
0&0&&& & \alpha & 0\\
0 & 0 &&& & 0 & 0\\
\vdots&\vdots& & \hat{\Xi} & & \vdots & \vdots \\
0&0& & && 0 & 0\\ \hline
0&-1&-\alpha & \dots & 0 & 0 &0\\
-1&0&0 & \dots & 0 & 0 &0
\end{array}\right) = \det(\hat{\Xi})
\]
where $\hat{\Xi}$ is a $(2j - 4)\times(2j-4)$ submatrix of $\Xi$, and
$\alpha$ is either $0$ or $i$ depending on whether the third column is
included.
If the submatrix does not include the second row and column (which can
only happen for $n > 4$), it must
include the third, and the minor
is instead \[
\det\left(\begin{array}{cc|ccc|cc}
0&0&0 & \dots & 0 & 0 & 1\\
0&0&\xi_{1k_1}& \dots & \xi_{1k_{2k-4}} & i & 0\\
\hline
0&-\xi_{1k_1}&& && 0 & 0\\
\vdots&\vdots& & \hat{\Xi} & & \vdots & \vdots \\
0&-\xi_{1k_{2k-4}} && && 0 & 0\\ \hline
0&-i&0& \dots & 0 & 0 &0\\
-1&0&0 & \dots & 0 & 0 &0
\end{array}\right) = -\det(\hat{\Xi})
\]
where now $\hat{\Xi}$ is a $(2j-4)\times(2j-4)$ submatrix of $\Xi$ which {\em
does not include the first row and column}. Summing all of the $2j\times2j$
minors of $Y'_\Xi$ then yields the sum of $(2j -4)\times(2j-4)$ minors of
$\Xi$ which include the first row.
Suppose $\Xi$ is of the form:
\[
\Xi = \left(\begin{array}{c|cccc}
0 & \xi_1 & 0 & \dots & 0 \\\hline
-\xi_1 & & & & \\
0 & & & & \\
\vdots & & & \Xi' & \\
0 & & & &
\end{array}\right)
\]
Then $\charpoly(\Xi) = \lambda\cdot\charpoly(\Xi') -
\xi_1^2\cdot\charpoly(\Xi')$ Since the nonconstant coefficients of the
characteristic polynomial are the sums of the minors, and are algebraically
independent on $\so(m)$, this tells us that the sums of minors which include
the first row and column are also algebraically independent. Putting everything
together, this tells us that for any $\tilde{F}$ in \eqref{ratbreakdown},
we can find a $Y_\Xi$ such that $Q_2(Y_\Xi) = 0$ and $\tilde{F}(Y_\Xi) \neq 0$.
Therefore, $\ell = 0$ in \eqref{ratbreakdown}, and so every polynomial
on $\ensuremath{\mathfrak{gal}}(n)^*$ which invariant on the coadjoint orbits is generated by our
$Q_j$'s
\end{proof}
|
\section*{Abstract}
Conjecturing formulas and other symbolic relations occurs frequently in number theory and combinatorics. If we could automate conjecturing, we could benefit not only from speeding up, but also from finding conjectures previously out of our grasp. Grammatical evolution, a genetic programming technique, can be used for automated conjecturing in mathematics. Concretely, this work describes how one can interpret the Frobenius problem as a symbolic regression problem, and then apply grammatical evolution to it. In this manner, a few formulas for Frobenius numbers of specific quadruples were found automatically. The sketch of the proof for one conjectured formula, using lattice point enumeration method, is provided as well.
Same method can easily be used on other problems to speed up and enhance the research process.
\section{Introduction}
Unlike automation of theorem-proving, automated conjecturing is still a rare theme in mathematics. Probably the most significant example of a program used for automated conjecturing is Fajtlowicz's Graffiti \cite{graffiti}, which produced hundreds of conjectures in graph theory and inspired dozens of papers (e.\ g.\ \cite{Erdos}). Another example is Colton's HR \cite{Colton}. Colton considered mathematics as "a new domain for datamining" and used very simple algorithms on existing databases of mathematical knowledge, discovering, for example, a necessary condition for perfect numbers. Advances in artificial intelligence and its successes in other sciences suggest that, simply by treating examples and results of math experiments as data for machine learning problem, we could automatically gain new insights.
One method that could be very useful for mathematics is grammatical evolution (GE). GE is a genetic programming technique (i.\ e., used for automated programming), and as such, it can easily generate conjectures of deterministic type. On the other hand, it is a very general method, applicable to any problem which can be described using a context-free grammar. Indeed, it was successfully used in fields as various as computer-aided design (\cite{cad}) and computational finance (\cite{fin}). To author's knowledge, this is the first instance of using grammatical evolution in mathematics, so this work can be seen as a proof-of-concept demonstration -- grammatical evolution can be used for conjecturing in mathematics. We chose to show that for Frobenius problem. It is particularly suitable for this study because it consists of finding formulas, so GE is applied in its most usual form.
Frobenius problem is the following: given relatively prime natural numbers $a_1, \dots, a_n$, find the \emph{Frobenius number}, the largest natural number that is not representable as a non-negative integer linear combination of $a_1, \dots, a_n$. This number is denoted as $g(a_1, \dots, a_n)$.
For $n=2$, it is long known that $g(a_1, a_2) = a_1a_2-a_1-a_2$ (usually attributed to \cite{1882}).
For $n=3$, the quest for a similar simple formula has been the subject of extensive research. Curtis (\cite{msc12330}) has shown that this quest, in a way, cannot succeed -- there is no finite set of polynomials $\{ f_1, \dots, f_k \}$ such that, for each choice of $a_1, a_2, a_3$, there is some $i \in \{ 1, \dots, k \}$ such that $f_i(a_1, a_2, a_3) = g(a_1, a_2, a_3)$.
Still, efficient algorithms for calculating Frobenius number have been developed, as well as numerous lower and upper bounds (for an extensive review on these and other results regarding Frobenius number one can check \cite{alfonsinBook}).
For a fixed $n \geqslant 4$, there is a polynomial time algorithm, but it is impractical. There is no proven efficient algorithm, but various formulas for some specific cases were found, e.\ \nolinebreak g., for arithmetic and geometric sequences of arbitrary length. For $n=4$ more specific formulas were found, e.\ g.
\[ g(a, a+1, a+2, a+4) = (a+1)\left\lfloor \frac{a}{4} \right\rfloor + \left\lfloor \frac{a+1}{4} \right\rfloor + 2\left\lfloor \frac{a+2}{4} \right\rfloor - 1. \]
We show that grammatical evolution can be used for automated conjecturing in mathematics. The rest of this work is divided as follows.
The second section is the most important part -- it describes how one can interpret the Frobenius problem as a symbolic regression problem, and then apply grammatical evolution to it. In this manner, a few formulas for Frobenius number were conjectured, including the following one:
\[ \displaystyle g(3k+1, 3k+4, 6k+3, 6k+9) = (3k+1)\left(k - \left\lfloor \frac{3k+1}{21} \right\rfloor \right) - 1 \quad \forall k \geqslant 5. \]
This section also shows the briefest sketch of the proof of this formula, using lattice point enumeration method. More important is the fact that, simply by changing the input data, we easily generated conjectures for quadruples of different kind, and even for sextuples.
The last section discusses the benefits and limitations of using grammatical evolution as automated conjecturing method in mathematics.
\section{The Frobenius problem as Symbolic Regression}
We interpret the Frobenius problem as a symbolic regression problem, and then approach this problem using grammatical evolution. Grammatical evolution is a method of automatic programming which uses grammars and evolutionary approach. Binary chromosomes are used. \nolinebreak{\emph{Codons}}, groups of $8$ bits, represent integer numbers. Codons determine which rule from the grammar will be used.
Following the short background on grammars, we explain how we used grammatical evolution and results we obtained.
\subsection{Context free grammars and Backus--Naur form (BNF)}
\label{BNF}
Grammars describing programming languages are defined following the
formal languages, which were introduced as a scientific discipline by Noam Chomsky (e.\ g.~\cite{Chomsky1956})
during his search for simple rules capable of describing natural language.
Chomsky concluded that context-free grammars are sufficient for formalizing the grammar of the English language. These grammars form theoretical basis for most programming languages, but the importance of grammars is even greater since they provide the means for formalizing knowledge. If a finite set of simple rules can describe the grammar of English language, then such sets can describe many other phenomena, which in turn opens up a possibility for automated research of these phenomena.
The Backus--Naur form is the most commonly used notation for expressing the grammar of a language. BNF defines syntax rules (as in~\cite{naur:algol60:cacm:1963}),
and it formalizes syntactic expressions. By listing the set of production rules, we completely determine the grammar. In each rule there is exactly one variable on its left hand side, while on its right hand side there are expressions which can replace the given variable.
The example of BNF is given in the next subsection (it is the grammar used in this research).
\subsection{Interpreting the Problem as Symbolic Regression Problem}
We calculated Frobenius number for $40$ quadruples of the form $(x, x+3, 2x+1, 2x+7)$ with a relatively simple recursive algorithm (using dynamic programming). These quadruples and their Frobenius numbers serve as the input data for grammatical evolution. Form of these quadruples was chosen as a simple form for which Frobenius number was previously unknown.
\newpage
The input data then looks like this:
\begin{table}[H]
\begin{center}
\begin{tabular}{ | c c c c | c |}
\hline
x & x+3 & 2x+1 & 2x+7 & Frobenius number (target function)\\ \hline
3 & 6 & 7 & 13 & 11 \\ \hline
4 & 7 & 9 & 15 & 10 \\ \hline
5 & 8 & 11 & 17 & 14 \\ \hline
\dots & \dots & \dots & \dots & \dots \\ \hline
\end{tabular}}
\end{center}
\caption{Input data -- Frobenius number for quadruples $(x, x+3, 2x+1, 2x+7)$.}
\end{table}
A simple grammar was used (this is an example of BNF):
\begin{mygrammarrule}
<expr> ::= <expr> <op> <expr> \grammartag
\alt (<expr> <op> <expr>) \grammartag
\alt <expr> / <const> \grammartag
\alt (<expr> / <const>) \grammartag
\alt <var> \grammartag
\end{mygrammarrule}
\begin{mygrammarrule}
<op> ::= + \grammartag
\alt - \grammartag
\alt * \grammartag
\end{mygrammarrule}
\begin{mygrammarrule}
<var> ::= x \grammartag
\alt <const> \grammartag
\end{mygrammarrule}
\begin{mygrammarrule}
<const> ::= <const> <op> <const> \grammartag
\alt (<const> <op> <const>) \grammartag
\alt <const> / <const> \grammartag
\alt (<const> / <const>) \grammartag
\alt 1.0 \grammartag
\alt 3.0 \grammartag
\end{mygrammarrule}
Individual's chromosome is seen as a sequence of codons (integers). Example individual is $(120, 44, 42, 96, 189, 64, \ldots)$. We briefly explain how does this sequence map to an expression. Start symbol is <expr>, the first symbol in the grammar. There are five rules given which can replace <expr>. First codon determines which rule is applied. Regardless of their integer value, codons are not skipped, but considered modulo the number of possibilities. First codon value is $120$, so rule number $0$ is applied, because $120 \equiv 0\ (\textrm{mod}\ 5)$. In this case, <expr> $\rightarrow$ <expr><op><expr>. Now, second codon determines which rule will be applied to <expr> (first of the three symbols). Second codon value is $44$, so the rule number $4$ is applied ($44 \equiv 4\ (\textrm{mod}\ 5)$), i.\ e., <expr><op><expr> $\rightarrow$ <var><op><expr>. Third codon value, $42$, determines that <var> will be replaced by $x$ (grammar gives just $2$ rules for <var> and $42 \equiv 0\ (\textrm{mod}\ 2)$, so we use the rule number $0$), i.\ e., <var><op><expr> $\rightarrow$ x<op><expr>. Now, $x$ is a terminal symbol (grammar does not provide any rules to replace it), so the next symbol we concentrate on is <op>. Next codon $96 \equiv 0\ (\textrm{mod}\ 3)$, so x<op><expr> $\rightarrow$ x+<expr>. The remaining two codons result in $x+\<expr>$ $\rightarrow$ $x+\<var>$ $\rightarrow$ $x+x$.
The aim is to find an individual that will, using this grammar, map to the expression describing the given data. A given sample of Frobenius numbers is used for the evaluation of indviduals -- fitness for this problem is given by the sum, taken over 40 function values, of the error between the evolved and target function (Frobenius number). For more details on GE, one can check the first publication \cite{oneill:2001:TEC}.
This search is done by evolutionary algorithm. As the individuals are simple bit-vectors, we do not have to employ any specific crossover or mutation operators. Standard genetic operators of one-point crossover and one-bit mutation were used. Parameters of the evolution were: population had $500$ individuals, it took $100$ generations, crossover probability was $0.9$, mutation probability was $0.1$. It is not important to carefully optimize these parameters.
GE did not find a formula valid for all given quadruples, but it found one which had a very high fit, i.\ e., a very small error:
$$g(x, x+3, 2x+1, 3x+7) \approx x(x/3 - x/21) - 1.$$
It was not hard to notice that this formula is valid in the case of $x = 3k+1$ and $k \geqslant 5$, which gives us the following result.
\begin{thm}
Frobenius number of the quadruple $(3k+1, 3k+4, 6k+3, 6k+9)$, for all $k \geqslant 5$, is
$$g(3k+1, 3k+4, 6k+3, 6k+9) = (3k+1)\left(k - \left\lfloor \frac{3k+1}{21} \right\rfloor \right) - 1. \quad $$
\end{thm}
(Floor function appears as integer division in C language which was used.)
\subsection{Proof of the conjectured formula}
The proof follows the method developed in the article \emph{Frobenius numbers by lattice point enumeration} \cite*{einstein2007frobenius} and concrete details were found with help of the algorithm, i.\ e., Mathematica package they developed. It is very similar to their proof for quadratic sequences. This is why author provides only the sketch ot the proof for the case of $k=7a$, without explaining the lattice point enumeration method (for understanding the proof and terminology, sections $1.-4., 6. \,\&\, 17.$ of \cite{einstein2007frobenius} should suffice).
Table $1$ is basically the proof for the formula when $k=7a$ ($a \in \mathbb{N}$). One can observe that the proposed \emph{elbows} determine the domain whose volume is $3k+1=21a+1$ (volume these elbows determine is a linear function of $a$, so it suffices to check two cases), while \emph{protoelbows} below show that the corresponding elbows are not in the fundamental domain.
\begin{table}[H]
\begin{center}
\begin{tabular}{ | p{2.8cm} | c c c |}
\hline
Corner \mbox{weights} & $126a^2+27a$ & $105a+16$ & $105a+10$ \\ \hline
Corners & $(0, 0, 3a)$ & $(1, 1, 1)$ & $(1, 2, 0)$ \\ \hline
Elbows & $(0, 0, 3a+1)$ & $(0, 1, 2)$ $(2, 1, 0)$ & $(0, 3, 0)$ $(0, 2, 1)$ \\ \hline
Protoelbows & $(-2, 0, 3a+1)$ & $(-5, 1, 2)$ $(2, 1, -1)$ & $(-1, 3, 0)$ $(-3, 2, 1)$\\ \hline
\end{tabular}
\begin{tabular}{ | p{2.8cm} | c c c |}
\hline
Corner \mbox{weights} & $126a^2+27a-1$ & $126a^2+27a-2$ & $126a^2+27a-3$ \\ \hline
Corners & $(2, 0, 3a-1)$ & $(4, 0, 3a-2)$ & $(6, 0, 3a-3)$ \\ \hline
Elbows & $(1, 0, 3a)$ & $(3, 0, 3a-1)$ & $(5, 0, 3a-2)$ $\, (7, 0, 0)$ \\ \hline
Protoelbows & $(1, -2, 3a)$ & $(3, -1, 3a-1)$ & $\,\,\,(0, 0, 0)$ $\quad (7, 0, -3)$ \\ \hline
\end{tabular}
}
\end{center}
\caption{Description of fundamental domain for $(21a+1, 21a+4, 42a+3, 42a+9)$.}
\end{table}
\vskip 0.5em
The Frobenius corner here is $(0, 0, 3a)$, so Frobenius number is \[ 3a\cdot(6k+9)-(3k+1) = 3a\cdot(42a+9)-21a-1=126a^2+6a-1 = (21a+1)(7a-a)-1, \] \vskip -0.5em which is exactly what the formula gives for $k=7a$.
\subsection{Other conjectures}
Simply by changing the input data, system can generate other conjectures.
One can try to generalize the formulas for arithmetic and geometric sequences. Probably the simplest generalization of these two is linear recursive sequence, e.\ g., sequence defined by $a_{n+1} = 3a_{n}+2$.
GE system resulted with some conjectures in these cases -- for odd numbers $x$ (the quadruples are not relatively prime for even numbers):
\[ f(x = 4k+1, y = 3x+2, z = 3y+2, w = 3z+2) = \]
$$ 48k^2+16k-1-3(4k+1)\left(k+\left\lfloor\frac{2k}{13}\right\rfloor + \left\lfloor\frac{2k+6}{13}\right\rfloor - \left\lfloor\frac{2k+19}{26}\right\rfloor \right), $$
\[ f(x = 4k+3, y = 3x+2, z = 3y+2, w = 3z+2) = \]
$$ 48k^2+64k+19-3(4k+3)\left(k+\left\lfloor\frac{2k+1}{13}\right\rfloor + \left\lfloor\frac{2k+7}{13}\right\rfloor - \left\lfloor\frac{k+3}{13}\right\rfloor \right). $$
Even the size of the tuple is easily changed, e.\ g., system found the following conjecture as well:
\[ g(6k+1, 6k+4, 6k+7, 12k+3, 12k+9, 12k+15) = (6k+1)\left(k - \left\lfloor \frac{k}{13} \right\rfloor \right)-1 \quad \forall k \geqslant 5. \]
One can even change the problem (again, simply by changing the input data to examples for that problem), but that would take us out of the scope of this work.
\section{Benefits and limitations}
Grammatical evolution can generate solutions in an arbitrary language by a simple change of the grammar -- which is usually one short input file. This advantage emerges from the generality of context-free grammars, hence GE has broad generativity -- it can be applied not only to the search of a formula, but also to anything context-free grammars can describe.
Knowledge embedded in grammar's rules, speeds up the search process. In the present study, knowledge was basically the way expressions are created. Minor expectation of desired formula, the fact that division operator should appear only with constant as divisor, was also embedded in the grammar. The challenge in similar studies could be in embedding more complex expectations (e.\ g., recent results) in grammars. Constructing the grammar might require more creativity when one applies this method on different types of problems (say, constructing a grammar to find a graph with some particular property probably will not be such an easy task). However, given that usage of grammars is exactly the reason of great generalizability, the author believes that this advantage outweighs disadvantages.
Formulas found here could have been found using a more standard type of regression or using polynomial interpolation. However, because of the floor function, the formula proven here would be broken into seven polynomial formulas. Far more important, these methods require researcher to know or guess the model prior to applying them, and they are not so easy to generalize. On the other hand, fine tuning, i.\ e., getting the constants right, is a potential, but minor problem for GE. Formulas like the one proven here, ending with $-1$, can be missed by that $-1$ (in many runs of the program). Unlike computer, researcher can easily note and fix this type of problem.
\bibliographystyle{babplain}
|
\section{Electronic version of tables}
\setcounter{table}{0}
\begin{longtable}{lccccl}
\caption{Catalog of ETGs with MegaCam observations (online web version)} \label{tab:obsmeg} \\
\hline
Galaxy & Band & N & Integration & IQ & Background \\
& & & s & arcsec & ADUs \\
(1) & (2) & (3) & (4) & (5) & (6) \\ \hline \\
\endfirsthead
\multicolumn{6}{l}
{{\bfseries \tablename\ \thetable{} -- continued from previous page}} \\
&&\\ \hline
Galaxy & Band & N & Integration & IQ & Background \\ \hline
\endhead
\endfoot
\hline
\endlastfoot
NGC0448 & g' & 7 & 2415 & 0.64 & 704.29 \\
& r' & 7 & 2415 & 0.77 & 941.57 \\
& i' & 7 & 1610 & 0.65 & 1054.86 \\
\hline
NGC0474 & u* & 7 & 4900 & 1.16 & 223.43 \\
& g' & 7 & 2415 & 0.83 & 784.43 \\
& r' & 7 & 2415 & 0.66 & 935.43 \\
& i' & 14 & 3220 & 0.65 & 1080.79 \\
\hline
NGC0502 & g' & 7 & 2415 & 0.82 & 688 \\
& r' & 7 & 2415 & 0.79 & 981.29 \\
& i' & 7 & 1610 & 0.66 & 1626.29 \\
\hline
NGC0509 & g' & 6 & 2070 & 0.82 & 688 \\
& r' & 6 & 2070 & 0.79 & 981.29 \\
& i' & 6 & 1380 & 0.66 & 1626.29 \\
\hline
NGC0516 & g' & 4 & 1380 & 0.82 & 688 \\
& r' & 4 & 1380 & 0.79 & 981.29 \\
& i' & 4 & 920 & 0.66 & 1626.29 \\
\hline
NGC0524 & g' & 7 & 2415 & 1.28 & 773.43 \\
& r' & 7 & 2415 & 0.71 & 1182 \\
& i' & 7 & 1610 & 0.75 & 993.43 \\
\hline
NGC0525 & g' & 7 & 2415 & 1.28 & 773.43 \\
& r' & 7 & 2415 & 0.71 & 1182 \\
& i' & 7 & 1610 & 0.75 & 993.43 \\
\hline
NGC0661 & g' & 7 & 2415 & 0.89 & 586.71 \\
& r' & 7 & 2415 & 0.78 & 854.14 \\
& i' & 7 & 1610 & 0.6 & 868 \\
\hline
NGC0680 & g' & 6 & 2070 & 0.95 & 276.67 \\
& r' & 12 & 4140 & 0.71 & 725.75 \\
& i' & 6 & 1380 & 0.58 & 589 \\
\hline
NGC0770 & g' & 7 & 2415 & 0.6 & 723 \\
& r' & 7 & 2415 & 0.79 & 839 \\
& i' & 7 & 1610 & 0.55 & 990.29 \\
\hline
NGC0936 & g' & 7 & 2415 & 0.69 & 655.29 \\
& r' & 7 & 2415 & 0.62 & 860.14 \\
& i' & 7 & 1610 & 0.79 & 834.86 \\
\hline
NGC1023 & u* & 14 & 9800 & 1.08 & 293.29 \\
& g' & 7 & 2415 & 0.71 & 593.29 \\
& r' & 7 & 2415 & 0.86 & 760.43 \\
\hline
NGC1121 & g' & 7 & 2415 & 1.12 & 757.86 \\
& r' & 7 & 2415 & 1.18 & 827.29 \\
\hline
NGC1222 & g' & 7 & 2415 & 0.81 & 624.43 \\
& r' & 7 & 2415 & 1.16 & 862.86 \\
\hline
NGC1248 & g' & 7 & 2415 & 1.35 & 878.86 \\
& r' & 7 & 2415 & 0.95 & 754.43 \\
\hline
NGC1266 & g' & 7 & 2415 & 0.78 & 596 \\
& r' & 7 & 2415 & 0.57 & 1252.57 \\
\hline
NGC1289 & g' & 7 & 2415 & 1.54 & 964.43 \\
& r' & 7 & 2415 & 1.02 & 1022.43 \\
\hline
NGC1665 & g' & 7 & 2415 & 0.8 & 595.43 \\
& r' & 7 & 2415 & 0.93 & 754.71 \\
& i' & 10 & 2300 & 0.57 & 1710.4 \\
\hline
PGC016060 & g' & 7 & 2415 & 0.96 & 623.14 \\
& r' & 7 & 2415 & 1.17 & 1580.57 \\
& i' & 8 & 1840 & 0.58 & 1575.88 \\
\hline
UGC03960 & g' & 12 & 4140 & 1.61 & 311.33 \\
& r' & 12 & 4140 & 1.45 & 655.08 \\
& i' & 6 & 1380 & 0.53 & 1165.5 \\
\hline
NGC2481 & g' & 6 & 2070 & 0.87 & 430 \\
& r' & 6 & 2070 & 0.76 & 908.67 \\
& i' & 6 & 1380 & 0.89 & 669.67 \\
\hline
NGC2549 & g' & 7 & 2415 & 0.94 & 658.57 \\
& r' & 7 & 2415 & 0.93 & 804.57 \\
\hline
NGC2577 & g' & 7 & 2415 & 0.69 & 1007 \\
& r' & 7 & 2415 & 0.66 & 1060.57 \\
\hline
NGC2592 & g' & 7 & 2415 & 1.15 & 748.71 \\
& r' & 7 & 2415 & 0.93 & 891.43 \\
\hline
NGC2594 & g' & 7 & 2415 & 1.15 & 748.71 \\
& r' & 7 & 2415 & 0.93 & 891.43 \\
\hline
UGC04551 & g' & 7 & 2415 & 1.2 & 632.14 \\
& r' & 7 & 2415 & 0.79 & 740.86 \\
\hline
NGC2679 & g' & 7 & 2415 & 1.09 & 789.43 \\
& r' & 7 & 2415 & 0.6 & 1531 \\
\hline
NGC2695 & g' & 7 & 2415 & 0.7 & 720.43 \\
& r' & 7 & 2415 & 0.51 & 997.43 \\
\hline
NGC2685 & u* & 7 & 4900 & 1.21 & 243 \\
& g' & 14 & 4830 & 0.88 & 924.71 \\
& r' & 7 & 2415 & 0.81 & 987 \\
& i' & 7 & 1610 & 0.77 & 1211.14 \\
\hline
NGC2698 & g' & 5 & 1725 & 0.7 & 720.43 \\
& r' & 5 & 1725 & 0.51 & 997.43 \\
\hline
NGC2699 & g' & 5 & 1725 & 0.7 & 720.43 \\
& r' & 5 & 1725 & 0.51 & 997.43 \\
\hline
NGC2764 & g' & 7 & 2415 & 0.7 & 929.57 \\
& r' & 7 & 2415 & 0.5 & 1084.14 \\
\hline
NGC2768 & g' & 14 & 4830 & 1.01 & 534.14 \\
& r' & 7 & 2415 & 1.12 & 597.14 \\
\hline
NGC2778 & g' & 6 & 2070 & 0.81 & 297 \\
& r' & 6 & 2070 & 1.43 & 550 \\
& i' & 6 & 1380 & 1.06 & 491.83 \\
\hline
NGC2852 & g' & 7 & 2415 & 1.06 & 578.57 \\
& r' & 7 & 2415 & 0.7 & 752.57 \\
\hline
NGC2859 & g' & 7 & 2415 & 0.87 & 597.86 \\
& r' & 7 & 2415 & 0.85 & 850.14 \\
\hline
NGC2880 & g' & 7 & 2415 & 0.88 & 624.86 \\
& r' & 7 & 2415 & 0.71 & 834.71 \\
\hline
NGC2962 & g' & 7 & 2415 & 0.54 & 742.57 \\
& r' & 7 & 2415 & 0.56 & 1593.43 \\
\hline
NGC3032 & g' & 7 & 2415 & 0.73 & 725.14 \\
& r' & 7 & 2415 & 0.8 & 1218.29 \\
\hline
PGC028887 & g' & 7 & 2415 & 0.93 & 561.14 \\
& r' & 7 & 2415 & 1.18 & 706.43 \\
\hline
NGC3073 & g' & 7 & 2415 & 0.85 & 510.14 \\
& r' & 7 & 2415 & 0.99 & 792 \\
\hline
NGC3098 & g' & 7 & 2415 & 0.83 & 639.43 \\
& r' & 7 & 2415 & 0.56 & 1467.43 \\
\hline
UGC05408 & g' & 7 & 2415 & 1.14 & 705.86 \\
& r' & 7 & 2415 & 1.03 & 1133.86 \\
\hline
NGC3193 & g' & 7 & 2415 & 1.2 & 565.86 \\
& r' & 7 & 2415 & 0.92 & 786.57 \\
\hline
NGC3182 & g' & 14 & 4830 & 1.32 & 532.86 \\
& r' & 7 & 2415 & 0.71 & 882.29 \\
\hline &&&&&\\
NGC3226 & g' & 7 & 2415 & 1.17 & 690.14 \\
& r' & 7 & 2415 & 0.68 & 907.43 \\
\hline
NGC3230 & g' & 7 & 2415 & 0.97 & 705.86 \\
& r' & 7 & 2415 & 0.86 & 830.86 \\
\hline
NGC3245 & u* & 7 & 4900 & 0.97 & 208 \\
& g' & 7 & 2415 & 0.63 & 910 \\
& r' & 7 & 2415 & 0.81 & 672 \\
& i' & 7 & 1610 & 0.5 & 1151 \\
\hline
NGC3379 & g' & 12 & 4140 & 1.58 & 325.75 \\
& r' & 6 & 2070 & 1.24 & 421.83 \\
& i' & 6 & 1380 & 0.97 & 550.83 \\
\hline
NGC3384 & g' & 12 & 4140 & 1.58 & 325.75 \\
& r' & 6 & 2070 & 1.24 & 421.83 \\
& i' & 6 & 1380 & 0.97 & 550.83 \\
\hline
NGC3400 & g' & 4 & 1380 & 0.99 & 580.71 \\
& r' & 4 & 1380 & 0.81 & 720.29 \\
\hline
NGC3414 & g' & 7 & 2415 & 0.99 & 580.71 \\
& r' & 7 & 2415 & 0.81 & 720.29 \\
\hline
NGC3457 & g' & 6 & 2070 & 0.74 & 266.17 \\
& r' & 6 & 2070 & 1.08 & 519.5 \\
& i' & 6 & 1380 & 0.89 & 426.33 \\
\hline
NGC3489 & g' & 6 & 2070 & 1.53 & 275 \\
& r' & 6 & 2070 & 1.28 & 557.17 \\
& i' & 6 & 1380 & 0.9 & 342.67 \\
\hline
NGC3522 & g' & 7 & 2415 & 1.01 & 660.29 \\
& r' & 7 & 2415 & 0.84 & 778.43 \\
\hline
UGC06176 & g' & 7 & 2415 & 0.55 & 687.71 \\
& r' & 7 & 2415 & 1.16 & 851.57 \\
\hline
NGC3599 & g' & 6 & 2070 & 0.79 & 274.83 \\
& r' & 6 & 2070 & 1.33 & 551.67 \\
& i' & 6 & 1380 & 0.85 & 355.17 \\
\hline
NGC3605 & g' & 6 & 2070 & 0.79 & 274.83 \\
& r' & 6 & 2070 & 1.33 & 551.67 \\
& i' & 6 & 1380 & 0.85 & 355.17 \\
\hline
NGC3607 & g' & 6 & 2070 & 0.79 & 274.83 \\
& r' & 6 & 2070 & 1.33 & 551.67 \\
& i' & 6 & 1380 & 0.85 & 355.17 \\
\hline
NGC3608 & g' & 6 & 2070 & 0.79 & 274.83 \\
& r' & 6 & 2070 & 1.33 & 551.67 \\
& i' & 6 & 1380 & 0.85 & 355.17 \\
\hline
NGC3610 & g' & 7 & 2415 & 0.89 & 607 \\
& r' & 7 & 2415 & 0.64 & 831.57 \\
\hline
NGC3613 & g' & 6 & 2070 & 1.05 & 166.5 \\
& r' & 6 & 2070 & 0.96 & 339.17 \\
& i' & 12 & 2760 & 0.64 & 630.5 \\
\hline
NGC3619 & g' & 6 & 2070 & 1.05 & 166.5 \\
& r' & 6 & 2070 & 0.96 & 339.17 \\
& i' & 12 & 2760 & 0.64 & 630.5 \\
\hline
NGC3626 & g' & 7 & 2415 & 0.97 & 644.14 \\
& r' & 7 & 2415 & 1.17 & 957.86 \\
\hline
NGC3630 & g' & 7 & 2415 & 0.65 & 781.29 \\
& r' & 7 & 2415 & 0.84 & 1008.43 \\
\hline
NGC3640 & g' & 7 & 2415 & 0.65 & 781.29 \\
& r' & 7 & 2415 & 0.84 & 1008.43 \\
\hline
NGC3641 & g' & 7 & 2415 & 0.65 & 781.29 \\
& r' & 7 & 2415 & 0.84 & 1008.43 \\
\hline &&&&&\\ &&&&&\\
NGC3665 & g' & 7 & 2415 & 1.02 & 586.43 \\
& r' & 7 & 2415 & 0.95 & 982.86 \\
\hline
NGC3796 & g' & 7 & 2415 & 0.73 & 653 \\
& r' & 7 & 2415 & 0.62 & 897.57 \\
\hline
NGC3838 & g' & 7 & 2415 & 0.71 & 578.57 \\
& r' & 7 & 2415 & 0.68 & 850 \\
\hline
NGC3941 & g' & 7 & 2415 & 0.77 & 626.57 \\
& r' & 7 & 2415 & 0.56 & 1026.29 \\
\hline
NGC3998 & g' & 7 & 2415 & 0.72 & 670.57 \\
& r' & 7 & 2415 & 0.55 & 924.71 \\
\hline
NGC4026 & g' & 7 & 2415 & 0.93 & 506.86 \\
& r' & 7 & 2415 & 0.81 & 616.71 \\
\hline
NGC4036 & g' & 7 & 2415 & 0.98 & 413.43 \\
& r' & 7 & 2415 & 0.99 & 506.14 \\
\hline
NGC4119 & g' & 8 & 2760 & 0.98 & 916.38 \\
& r' & 7 & 2415 & 0.61 & 1158.57 \\
\hline
NGC4150 & g' & 7 & 2415 & 0.67 & 774.57 \\
& r' & 7 & 2415 & 0.66 & 1016.86 \\
\hline
NGC4191 & g' & 7 & 2415 & 0.77 & 805.29 \\
& r' & 7 & 2415 & 0.66 & 1016.43 \\
\hline
NGC4203 & g' & 7 & 2415 & 0.65 & 694.57 \\
& r' & 7 & 2415 & 0.6 & 883.86 \\
\hline
NGC4249 & g' & 7 & 2415 & 1.03 & 707.14 \\
& r' & 7 & 2415 & 0.75 & 1007 \\
\hline
NGC4259 & g' & 7 & 2415 & 1.03 & 707.14 \\
& r' & 7 & 2415 & 0.75 & 1007 \\
\hline
NGC4261 & g' & 7 & 2415 & 1.03 & 707.14 \\
& r' & 7 & 2415 & 0.75 & 1007 \\
\hline
NGC4264 & g' & 7 & 2415 & 1.03 & 707.14 \\
& r' & 7 & 2415 & 0.75 & 1007 \\
\hline
NGC4268 & g' & 7 & 2415 & 1.03 & 707.14 \\
& r' & 7 & 2415 & 0.75 & 1007 \\
\hline
NGC4270 & g' & 7 & 2415 & 1.03 & 707.14 \\
& r' & 7 & 2415 & 0.75 & 1007 \\
\hline
NGC4278 & g' & 6 & 2070 & 1.33 & 226.67 \\
& r' & 6 & 2070 & 1.45 & 546.83 \\
& i' & 12 & 2760 & 1.06 & 718.33 \\
\hline
NGC4283 & g' & 6 & 2070 & 1.33 & 226.67 \\
& r' & 6 & 2070 & 1.45 & 546.83 \\
& i' & 12 & 2760 & 1.06 & 718.33 \\
\hline
NGC4281 & g' & 7 & 2415 & 1.03 & 707.14 \\
& r' & 7 & 2415 & 0.75 & 1007 \\
\hline
NGC4382 & g' & 7 & 2415 & 0.63 & 689.43 \\
& r' & 7 & 2415 & 0.69 & 1291 \\
\hline
NGC4690 & g' & 7 & 2415 & 0.62 & 780.43 \\
& r' & 7 & 2415 & 0.64 & 1144.71 \\
\hline
NGC4753 & g' & 7 & 2415 & 0.8 & 776.14 \\
& r' & 7 & 2415 & 0.73 & 1484.43 \\
\hline
NGC5173 & g' & 6 & 2070 & 1.25 & 438 \\
& r' & 6 & 2070 & 1.06 & 740 \\
\hline
NGC5198 & g' & 7 & 2415 & 1.25 & 438 \\
& r' & 7 & 2415 & 1.06 & 740 \\
\hline
NGC5273 & g' & 7 & 2415 & 0.83 & 562.57 \\
& r' & 7 & 2415 & 0.54 & 800.29 \\
\hline
NGC5308 & g' & 7 & 2415 & 1.26 & 496.71 \\
& r' & 7 & 2415 & 0.55 & 807 \\
\hline &&&&&\\
NGC5322 & g' & 7 & 2415 & 1.19 & 613.14 \\
& r' & 7 & 2415 & 0.76 & 1056 \\
\hline
NGC5342 & g' & 6 & 2070 & 1.19 & 613.14 \\
& r' & 6 & 2070 & 0.76 & 1056 \\
\hline
NGC5379 & g' & 12 & 4140 & 1.25 & 182.58 \\
& r' & 12 & 4140 & 1.68 & 382.25 \\
& i' & 12 & 2760 & 1 & 509.5 \\
\hline
NGC5422 & g' & 7 & 2415 & 1.03 & 597.71 \\
& r' & 7 & 2415 & 1.17 & 845.29 \\
\hline
NGC5473 & g' & 6 & 2070 & 1.35 & 229.67 \\
& r' & 6 & 2070 & 1.32 & 563 \\
& i' & 6 & 1380 & 1.22 & 838.67 \\
\hline
NGC5481 & g' & 12 & 4140 & 1.54 & 213.67 \\
& r' & 12 & 4140 & 1.23 & 516.5 \\
& i' & 12 & 2760 & 1.09 & 668.17 \\
\hline
NGC5485 & g' & 6 & 2070 & 1.35 & 229.67 \\
& r' & 6 & 2070 & 1.32 & 563 \\
& i' & 6 & 1380 & 1.22 & 838.67 \\
\hline
PGC050395 & g' & 6 & 2070 & 1.35 & 229.67 \\
& r' & 6 & 2070 & 1.32 & 563 \\
& i' & 6 & 1380 & 1.22 & 838.67 \\
\hline
NGC5507 & g' & 14 & 4830 & 0.94 & 797.5 \\
& r' & 7 & 2415 & 0.57 & 1228.57 \\
\hline
NGC5557 & g' & 6 & 2070 & 1.37 & 190.67 \\
& r' & 6 & 2070 & 1.12 & 364.5 \\
& i' & 6 & 1380 & 0.95 & 621.17 \\
\hline
NGC5582 & g' & 7 & 2415 & 0.86 & 730.14 \\
& r' & 7 & 2415 & 1.13 & 890.43 \\
\hline
NGC5574 & g' & 14 & 4830 & 1.05 & 721.79 \\
& r' & 7 & 2415 & 0.93 & 981.14 \\
\hline
NGC5576 & g' & 14 & 4830 & 1.05 & 721.79 \\
& r' & 7 & 2415 & 0.93 & 981.14 \\
\hline
NGC5611 & g' & 7 & 2415 & 0.85 & 554.71 \\
& r' & 7 & 2415 & 0.79 & 945.43 \\
\hline
NGC5631 & g' & 6 & 2070 & 1.18 & 238.5 \\
& r' & 6 & 2070 & 1.33 & 387.17 \\
& i' & 6 & 1380 & 1.24 & 601.17 \\
\hline
NGC5638 & g' & 7 & 2415 & 1.08 & 707.71 \\
& r' & 7 & 2415 & 1.16 & 882.71 \\
\hline
IC1024 & g' & 7 & 2415 & 1.08 & 707.71 \\
& r' & 7 & 2415 & 1.16 & 882.71 \\
\hline
UGC09519 & g' & 7 & 2415 & 0.97 & 563.86 \\
& r' & 7 & 2415 & 1.03 & 675.57 \\
& i' & 7 & 1610 & 0.45 & 1139.14 \\
\hline
NGC5813 & g' & 7 & 2415 & 0.76 & 931.43 \\
& r' & 7 & 2415 & 0.69 & 1503.14 \\
\hline
NGC5831 & g' & 7 & 2415 & 0.87 & 700.14 \\
& r' & 7 & 2415 & 1.04 & 1057.14 \\
\hline
NGC5838 & g' & 7 & 2415 & 0.87 & 696.71 \\
& r' & 7 & 2415 & 0.84 & 1069.71 \\
\hline
NGC5845 & g' & 7 & 2415 & 0.9 & 863.43 \\
& r' & 7 & 2415 & 0.63 & 1622.57 \\
\hline
NGC5866 & g' & 6 & 2070 & 1.57 & 619.83 \\
& r' & 7 & 2415 & 0.92 & 684 \\
& i' & 7 & 1610 & 0.62 & 1298.86 \\
\hline &&&&&\\&&&&&\\&&&&&\\
NGC6014 & g' & 14 & 4830 & 0.96 & 588.36 \\
& r' & 7 & 2415 & 0.77 & 1144 \\
& i' & 7 & 1610 & 0.45 & 919.14 \\
\hline
NGC6017 & g' & 12 & 4140 & 0.96 & 588.36 \\
& r' & 6 & 2070 & 0.67 & 775.29 \\
& i' & 7 & 1610 & 0.45 & 919.14 \\
\hline
PGC056772 & g' & 14 & 4830 & 1.05 & 548.07 \\
& r' & 7 & 2415 & 0.94 & 771.29 \\
& i' & 7 & 1610 & 0.5 & 1290.29 \\
\hline
PGC058114 & g' & 7 & 2415 & 1.09 & 794.29 \\
& r' & 7 & 2415 & 0.9 & 878.29 \\
& i' & 7 & 1610 & 0.57 & 1520.43 \\
\hline
NGC6278 & g' & 12 & 4140 & 1.21 & 207.92 \\
& r' & 6 & 2070 & 1.02 & 399.17 \\
& i' & 12 & 2760 & 0.63 & 504.08 \\
\hline
NGC6547 & g' & 7 & 2415 & 1.28 & 856.57 \\
& r' & 7 & 2415 & 0.67 & 878.14 \\
& i' & 7 & 1610 & 0.5 & 1444.43 \\
\hline
NGC6548 & g' & 7 & 2415 & 1.45 & 571.71 \\
& r' & 7 & 2415 & 0.64 & 973.14 \\
& i' & 7 & 1610 & 0.64 & 1669.43 \\
\hline
NGC6703 & g' & 14 & 4830 & 1.33 & 949.64 \\
& r' & 7 & 2415 & 0.92 & 862.71 \\
\hline
NGC6798 & g' & 6 & 2070 & 0.84 & 196.5 \\
& r' & 6 & 2070 & 1.09 & 418.33 \\
& i' & 12 & 2760 & 0.95 & 416.67 \\
\hline
NGC7280 & g' & 12 & 4140 & 1.23 & 254.75 \\
& r' & 12 & 4140 & 0.72 & 506.33 \\
& i' & 12 & 2760 & 1.05 & 462.5 \\
\hline
NGC7332 & u* & 7 & 4900 & 0.77 & 226.43 \\
& g' & 6 & 2070 & 0.69 & 217.5 \\
& r' & 6 & 2070 & 0.96 & 375.17 \\
& i' & 6 & 1380 & 0.83 & 368 \\
\hline
NGC7457 & u* & 7 & 4900 & 1.03 & 211.71 \\
& g' & 14 & 4830 & 1.02 & 1020.36 \\
& r' & 7 & 2415 & 0.76 & 714.14 \\
& i' & 7 & 1610 & 0.61 & 1264 \\
\hline
NGC7454 & u* & 7 & 4900 & 1.09 & 209.57 \\
& g' & 14 & 4830 & 0.95 & 821.14 \\
& r' & 12 & 4140 & 0.73 & 478.08 \\
& i' & 6 & 1380 & 0.87 & 556 \\
\hline
NGC7465 & u* & 7 & 4900 & 1.09 & 209.57 \\
& g' & 12 & 4140 & 1.2 & 284.42 \\
& r' & 12 & 4140 & 0.73 & 478.08 \\
& i' & 6 & 1380 & 0.87 & 556 \\
\hline
NGC7693 & g' & 7 & 2415 & 1.22 & 828.71 \\
& r' & 7 & 2415 & 0.69 & 928 \\
& i' & 7 & 1610 & 0.51 & 1326.29 \\
\hline
NGC7710 & g' & 7 & 2415 & 0.71 & 736.71 \\
& r' & 7 & 2415 & 0.7 & 1030.43 \\
& i' & 7 & 1610 & 0.67 & 946.43 \\
\hline
\multicolumn{6}{l}{\parbox{8cm}{Notes: (3) Number of individual exposures (4) Total integration time (5) Image Quality: FWHM of the PSF (6) Background level}} \\
\end{longtable}
\newpage
\setcounter{table}{3}
\begin{longtable}{llp{0.7\textwidth}}
\caption{ETG classification based on the deep imaging (online web version)} \label{tab:class} \\
\hline
Galaxy & Class & Individual comments \\ \hline &&\\
\endfirsthead
\multicolumn{3}{l
{{\bfseries \tablename\ \thetable{} -- continued from previous page}} \\
&&\\ \hline
Galaxy & Class & Individual comments \\ \hline &&\\
\endhead
\endfoot
\hline
\endlastfoot
NGC0448 & I+s & The ETG is in a tidal interaction with a disturbed companion. \\
NGC0474 & M+s+r+ph & The ETG is surrounded by multiple concentric shells and hosts several radial streams. Its outer halo reaches the disk of the unperturbed companion spiral galaxy, NGC 0470. \\
NGC0502 & M+t?+r?+ah-wc-h & The stellar halo of the ETG is asymmetric, possibly due to the presence of a diffuse tidal tail and/or a shell. \\
NGC0509 & R-pc & \\
NGC0516 & R-pc & \\
NGC0524 & U-pc-h & The ETG is surrounded by galactic cirrus and extended halos from bright stars, preventing the detection of fine structures around it. \\
NGC0525 & R-pc-h & \\
NGC0661 & U+ah-pc & The ETG is surrounded by galactic cirrus and extended halos from bright stars, preventing the detection of fine structures around it. \\
NGC0680 & I+t+s+r+ph+wl-wc & The ETG is tidally disturbed, showing two extended tidal tails, and an asymmetric stellar halo. It has a bright edge-on companion in its vicinity. Whether the tidal tails result from this on-going interaction or a past major merger is unclear. \\
NGC0770 & I+t-pc & The ETG lies within a prominent tidal tail. It is likely a satellite of the massive perturbed spiral NGC 0772, and is currently tidally disrupted. \\
NGC0936 & C+s+wl & A stellar stream hosting a tidally disrupted companion wraps around the ETG. \\
NGC1023 & U+ah-h & The stellar halo of the ETG seems to be slightly disturbed, but the extended halos of two bright nearby stars hamper the classification. \\
NGC1121 & U-h & The ETG totally lies within the reflection halo of a bright star, preventing the detection of fine structures. \\
NGC1222 & M+t+r?+ph+pl & The ETG exhibits multiple signs of a relatively recent gas-rich merger: tidal tails, perturbed main body, dust lanes. \\
NGC1248 & R-pc-h & The ETG does not show any evident sign of disturbances although it makes a close pair with the undisturbed spiral galaxy Mrk 604. \\
NGC1266 & C+s?+wl-pc & The ETG has several low mass companions, with possibly a tidally disrupted one. Note however the high level of cirrus contamination. \\
NGC1289 & U+s?+wl-pc & Model subtraction of the ETG reveals a slightly perturbed central body with possibly a faint stream. However Galactic cirrus prevents a firm classification. \\
NGC1665 & U-h & The ETG is surrounded by a ring like structure probably made of old stars. Prominent Galactic cirrus is present in the field. \\
PGC016060 & C+d+pl & The galaxy is surrounded by a warped star-forming ring, and is possibly interacting with an early-type companion to the East. \\
UGC03960 & R-h & The ETG is apparently relaxed though the halo of a nearby star hampers the detection of faint tidal streams. \\
NGC2481 & I+t?-wc & The ETG and its disturbed companion NGC 2480 make an interacting pair. The prominent tidal tails shown by the system likely come from the companion galaxy. \\
NGC2577 & C+s-wc & The ETG has a regular main body with one radial stellar stream sticking out to the North, hosting a possible progenitor. \\
NGC2592 & C+s+r-pc & The ETG is surrounded by cirrus. However the filament to the East is most likely a stellar stream as it hosts a putative progenitor. \\
NGC2594 & U+wl-pc-h & Galaxy classification is hampered by the presence of a nearby bright star and cirrus. The long filament to the South is most likely a cirrus. \\
NGC2695 & R & \\
NGC2685 & M+t?+ph+d+pl & The ETG exhibits a prominent perturbed star-forming disk and dust lanes, indicative of a rather recent gas-rich merger. \\
NGC2698 & I+t+ah-h & The ETG is involved in a tidal interaction with another ETG, NGC 2699. A diffuse bridge links the two galaxies. Besides, a large diffuse tail or stream wraps around the galaxy. \\
NGC2699 & I+t+s+r+ph & The ETG is involved in a tidal interaction with another ETG, NGC 2698. A diffuse bridge links the two galaxies. The main body is highly perturbed with multiple tails and streams around it. \\
NGC2764 & M+t+r+ph+pl-h & The ETG exhibits multiple tidal tails, shells and dust lanes, indicative of a relatively recent major wet merger. \\
NGC2768 & C+s?+r?+ah-h & The main body of the ETG is pretty relaxed, though asymmetric, and exhibits to the South either a stellar stream or shells, telling about a past merger. \\
NGC2778 & R & The ETG does not show any evidence of a tidal perturbation though it makes a close pair with the massive companion NGC 2779. \\
PGC028887 & U-h & The ETG is apparently relaxed, though its stellar halo hosts two possible companions, one with a stream (but the physical association is unsure). The presence of an extended reflection halo is problematic for the galaxy classification. \\
NGC3073 & I+t?+ah & The ETG may be a satellite of the nearby massive edge-on spiral NGC 3079. It exhibits an asymmetric main stellar body. Model subtraction reveals a possible diffuse tidal tail. \\
UGC05408 & C+s+wl-h & Two streams wrap around the ETG, revealing one or two minor mergers. \\
NGC3193 & I-h & The ETG makes an interacting pair with the tidally perturbed galaxy NGC 3189. It is embedded in the reflection halo of a bright star, preventing the detection of tidal tails. \\
NGC3226 & I+t?+s?+r?+ph+wl & The ETG is in close tidal interaction with the strongly disturbed spiral galaxy NGC 3227. The system is surrounded with multiple tidal tails. \\
NGC3230 & R-wc-h & \\
NGC3245 & I+r & Multiple shells are revealed by the ETG model subtraction. The galaxy is likely in tidal interaction with the edge-on, slightly warped, spiral, UGC 5662. \\
NGC3379 & R+r? & The ETG has a regular main body. The model subtraction possibly reveals shells. The galaxy makes a pair with the ETG NGC 3384. \\
NGC3384 & I+ah & The main stellar body of the ETG is asymmetric. It makes a pair with the ETG NGC 3379 and is believed to have been involved in a fast encounter with M96 that possibly formed the huge HI ring surrounding the system (known as the Leo Ring). \\
NGC3400 & R & \\
NGC3414 & I+s+ph & The disturbed ETG is in interaction with the tidally disturbed companion NGC 3418. It is crossed by a very extended South-North stellar stream. Its progenitor is visible to the South. To the North, the stream ends in a shell-like structure. \\
NGC3457 & R+wl-wc & \\
NGC3489 & R+wl & The roundish red halo around this relaxed ETG is likely caused by an internal reflection of the bright nucleus. \\
NGC3522 & R+s? & A stream hosting a possible progenitor is visible 50 kpc North of this relaxed ETG. Whether it is a disrupted satellite is unclear. \\
NGC3599 & R & \\
NGC3605 & U-h & The ETG is observed towards the halo of its companion NGC 3607. This prevents the detection of fine structures. \\
NGC3607 & I+ah+wl-h & The stellar halo of the ETG is slightly asymmetric. It makes a compact group with the ETGs NGC 3608 and NGC 3605. \\
NGC3608 & I+r?-h & Possible fine structures are visible on the image with the ETG model subtracted. It makes a close physical pair with the ETG NGC 3607. \\
NGC3613 & C+s+r?+ph-h & The subtraction of the ETG model disclosed a prominent stream and several other fainter fine structures. The three objects to the North-East are background galaxies. \\
NGC3619 & M+s+r+ph+pl & The main body of the ETG is strongly perturbed. Multiple shells are visible as well as a stellar stream crossing the galaxy from South to North. \\
NGC4026 & R & The ETG exhibits a strong slightly warped bar. The roundish red halo around the ETG is likely an artefact caused by a reflection of the bright nucleus. \\
NGC4036 & C+s+wl & Remnants of minor mergers are visible around and towards the ETG. It is unclear whether it is weakly interacting with the spiral galaxy NGC 4041, which seems slightly perturbed. \\
NGC4278 & R-h & The companion of the galaxy, the ETG NGC 4283, is located towards its stellar halo, but the two galaxies have a large velocity offset and are likely not interacting. \\
NGC4283 & U-h & The ETG is located towards the stellar halo of the ETG NGC 4278, hampering its classification. The two galaxies have a large velocity offset and are likely not interacting. \\
NGC5173 & C+s?+ah & The ETG, located in a group, is slightly perturbed and exhibits weak signs of minor accretion. \\
NGC5198 & C+s+r & The ETG exhibits a 90 kpc long narrow stellar stream to the West, hosting a disrupted progenitor. \\
NGC5322 & M+r+ah & Several shells are disclosed by the ETG model subtraction. \\
NGC5342 & R+wl & \\
NGC5379 & I+t+ah+wl & The ETG is tidally disturbed by the interaction with the massive spiral NGC 5389. \\
NGC5422 & R+pl-h & The ETG hosts a prominent dusty edge-on disk. \\
NGC5473 & R-h & \\
NGC5481 & R-wc & The ETG is relaxed but makes a close pair with the undisturbed spiral galaxy NGC 5480. \\
NGC5485 & M+t?+s+ph+pl & The stellar body of the ETG is disturbed. It possibly exhibits diffuse tails, streams and prominent dust lanes. \\
PGC050395 & R & \\
NGC5507 & I+t+ah & The ETG is in tidal interaction with the perturbed spiral NGC 5506: its main body is warped. Extended diffuse emission is seen to the South and West. \\
NGC5557 & M+t+r+ph-h & The external regions of the ETG are tidally perturbed. Two long tails emanate from it. The tail to the East hosts confirmed metal-rich tidal dwarf galaxies. \\
NGC5582 & R+d-h & A LSB blue star-forming spiral disk surrounds the relaxed stellar bulge of the ETG. \\
NGC5574 & I+t+ph-h & The ETG is tidally perturbed by the interaction with the ETG NGC 5576. A prominent large tidal tail emanates from it. \\
NGC5576 & I+r+ph & The ETG is strongly perturbed, likely following the tidal interaction with the ETG companion NGC 5574. \\
NGC5631 & M+s+r+ah+wl & The ETG exhibits multiple internal shells, and a curved stream in its outskirts. \\
NGC5638 & C+s & A stream ending with a shell like structure is visible to the South of the ETG. \\
IC1024 & C+s+ph+pl & The ETG shows signs of perturbations and dust lanes. One or two stellar filaments wrap around it. \\
UGC09519 & R+d+pl & The ETG is surrounded by a LSB star-forming disk. It exhibits prominent dust lanes in its central regions. \\
NGC5838 & R-h & \\
NGC5866 & M+ph+wl-h & The stellar body of the ETG is perturbed, but no fine structure is clearly observed. \\
NGC6014 & U+s?+ah+wl-wc & The main body of the ETG exhibits some underlying ring-like substructures, dust lanes and is slightly asymmetric. A filament (stellar or cirrus) of unknown origin is visible to the East. \\
NGC6017 & R+wl & The central bar of the ETG is perpendicular to the main axis of the stellar halo. \\
PGC056772 & R+wl & \\
PGC058114 & R+pl-h & \\
NGC6278 & C+s?-wc & The ETG makes a close pair with the unperturbed spiral NGC 6276. A possible stream is seen to the North, and fainter ones are disclosed by the model subtraction. \\
NGC6547 & R-h & The morphological classification is hampered by the presence of numerous foreground stars. \\
NGC6548 & R-wc & \\
NGC6703 & U+wl-pc-h & The ETG is located within a galactic cirrus and cannot be classified. \\
NGC6798 & U-pc-h & The ETG is surrounded by multiple filamentary cirrus-like structures. \\
NGC7280 & C+s & The ETG makes a pair with star-forming companion UGCA 429. Two faint streams are visible to the South. \\
NGC7332 & I+ah+d & The disk of the ETG seems slightly warped. It is likely in tidal interaction with the nearby spiral galaxy NGC 7339. \\
NGC7457 & R-pc-h & The filament to the East of the ETG is most likely a cirrus and not a stellar stream. \\
NGC7454 & U-pc-h & The presence of multiple narrow filaments due to cirrus prevents any classification. \\
NGC7465 & I+t+ph+d+pl-wc-h & The ETG hosts a tidally perturbed star-forming disk. It is interacting with the irregular galaxy NGC 7464 and possibly the spiral NGC 7463. \\
NGC7693 & R-wc & The ETG is relaxed though it makes a close pair with an undisturbed spiral to the South of unknown redshift. \\
NGC7710 & R+pl-wc & The ETG has a remarkable thin dusty edge-on disk. \\
&&\\
\end{longtable}
\end{document}
\section{Introduction}
\label{sec:intro}
Our knowledge of nearby, well resolved, galaxies made a big leap during the last 40 years with the availability of multi wavelength data, whereas for the previous decades only optical data, mostly images, were at the astronomers' disposal. Nowadays a nearby galaxy can no longer be characterized without ultraviolet and far infrared data revealing its star--forming activity, a near infrared image constraining its stellar mass, radio and millimeter maps, providing information about its gas content. Consequently, the interest for pure optical imaging surveys of galaxies has dropped, unless such surveys cover large regions of the sky and provide statistical information -- like the Sloan Digital Sky Survey \citep[SDSS,][]{York00} --, reach high angular resolution and give insight into nuclear regions, or target the high redshift Universe which still largely lacks non-optical data.
However, several recent imaging surveys benefiting from innovative observing techniques and instruments have rejuvenated our regard for the optical regime and to familiar galaxies, disclosing around them so far unknown prominent but low surface brightness (LSB) stellar structures, such as extended stellar halos and tidal tails \citep[e.g.][]{Mihos05,Janowiecki10,Martinez-Delgado10,Roediger11,vanDokkum14}.
At the same time, the development of cosmological numerical simulations contributed to foster the interest for deep imaging of the nearby Universe.
Following the paradigm of the hierarchical model, they figure out that todays massive galaxies grew from series of galactic collisions that left around them various types of vestiges, including shells and tails \cite[e.g][]{Bullock05,Naab07,Oser10,Helmi11,Cooper14}. The technique known as galactic archaeology that makes a census of LSB collisional debris with star counts has so far mostly been applied to galaxies in the Local and very nearby groups
\cite[e.g][]{McConnachie09,Crnojevic13}. Further away, structures can no longer be resolved into stars and other excavation tools should be used, in particular the diffuse light.
Whereas the large field of view photographic plates used in the 1950s-1970s enhanced with amplified techniques \citep{Malin78} were capable of detecting the extended LSB component of galaxies, the early CCDs developed in the 1980s, with their limited field of view, proved to be much less efficient at that task.
The complex cameras made with multiple optical elements that were built to host these new sensitive detectors, while well fitted to detect distant galaxies, generate numerous artefacts, such as internal reflections that hide extended LSB structures.
Thus somehow ironically LSB science made a step backwards with the advent of CCDs, and the few extragalactic papers published on the topic during the period 1980--2000 mostly focussed on the outer stellar populations of spiral galaxies \cite[e.g.][]{Lequeux96}.
Only recently astronomers, among them amateurs observing in very dark sites with simple cameras, raised the challenge of detecting again the diffuse light \citep{Martinez-Delgado09}. The astonishing images they produced pushed professional astronomers to develop new techniques to eliminate the instrumental signature in their camera. This involves special coating of the detectors \citep{Mihos05}, LSB-optimized observing techniques with large field of view mosaic cameras \citep{Ferrarese12}, or even the construction of new LSB dedicated cameras \citep{vanDokkum14}.
The gain of deep imaging is for some galaxies tremendous, with the detection of networks of interlaced filaments \citep{Martinez-Delgado10,Paudel13}, revealing past mass accretion histories that were much more complex than initially thought. However the sample of nearby objects with available deep optical imaging remains limited and highly biased towards galaxies for which previous imaging surveys such as the SDSS already indicated the possible presence of collisional debris \citep{Miskolczi11}.
This was the motivation for carrying out a systematic deep imaging survey of a well selected large sample of galaxies, allowing us to determine how the properties of the outskirts of galaxies vary with morphology, colour, gas content and structural parameters such as the mass, size and dynamics.
A survey of several tens of objects implies observing with at least medium size telescopes rather than small amateur-type telescopes, to keep individual exposure times (and thus survey length) relatively small, typically one hour instead of a full night. Using professional facilities ensures as well good photometric accuracy and image quality.
As part of the \AD\ project \citep{Cappellari11}, we have carried out with the Canada-France-Hawaii Telescope (CFHT) and the MegaCam camera a deep multi-band imaging survey of nearby Early-Type galaxies (ETGs).
This paper presents a comprehensive image catalog of 92 objects, located in low to medium density environments. Among them, a few systems were devoted individual studies \citep{Michel-Dansac10b,Duc11,Serra13,Alatalo14,Duc14}.
We address here the global survey strategy, observing and data reduction technique, and discuss its limitation, including contamination by instrumental artefacts such as diffuse halos of bright stars and galactic nuclei, as well as by foreground Galactic cirrus. These effects are not necessarily specific to images obtained with MegaCam, and may affect any deep imaging survey.
The gain of several magnitudes in limiting surface brightness compared to previous generation of imaging surveys makes us enter into a new regime having specific stumbling blocks which cannot be ignored and should be investigated in detail.
Sect.~\ref{sec:obs} presents the observations at CFHT, the observing strategy and data reduction technique. Sect.~\ref{sec:perf} discusses the survey performance and limitations. Sect.~\ref{sec:prod} details the sample selection as well as the high level image production.
The image catalog consists of composite multi-band images with a true colour rendering, surface brightness plus colour maps, and residual images, i.e. images in which the galaxy has been modeled and subtracted.
Sect.~\ref{sec:class} discusses the need to (re-) classify galaxies based on the newly discovered structures: extended halos, discs, stellar streams and tails, shells.
Conclusions and perspectives are given in Sect.~\ref{sec:sum}.
The scientific analysis of the images and statistical results will be presented in future papers.
\section{Observations and data reduction}
\label{sec:obs}
Observations have been carried out at the Canada-France-Hawaii Telescope, with the MegaCam camera, as part of a series of regular multi--semester PI and snapshot programs in the framework of the \AD\ project (10BF11, 11AF06, 11AD89, 11BD93, 12AF04, 12AF99).
The galaxies presented in this image release were observed between 2010 and 2013.
\subsection{Context}
MegaCam \citep{Boulade03} is a wide-field imager, made of 36 CCDs and covering 1 square degree field of view. It was initially designed for point-source type science, stars and distant galaxies, where large scale structures of the image background are secondary to photometric accuracy across the field of view. As part of the CFHT Legacy Survey effort, refinements over the years led to a better than 1\% photometric accuracy across the one degree field-of-view, a precision of great importance when dealing with precision cosmology \citep{Regnault09,Betoule13}. All these advances were integrated in the CFHT MegaCam official pipeline \citep{Magnier04} over the years. Since the background is usually internally subtracted for point-source type science, the global gradient caused by diffuse sky background reflections in the optics, plus the cumulative radial effect of the illumination correction to reach the percent photometric precision over the entire field of view, is of little concern.
But since the instrument first light in 2003, efforts were made to also enable scientific programs dealing with large and faint extended components that are de facto lost within the background of the image. The overall effort aimed at recovering the true sky background, that is a purely flat response on top of the astronomical sources signal.
The nod-and-subtract technique used in near-infrared astronomy was an inspiration for the observing strategy and pipeline developed for MegaCam, called Elixir-LSB. In the near--infrared regime, though, the background dominates the signal and varies over relatively short time scales. In the optical the idea is to model a background stable in time and dominated with astronomical sources, with just a handful of exposures.
We note that the processing presented in the following does not alter the photometric accuracy at small scales, i.e. the 1\% photometry precision is retained throughout the background correction process.
\subsection{Dithering, sky stability and overheads}
\label{sec:strategy}
In the g-band and the r-band, studies have demonstrated that the sky background on Mauna Kea is stable at a fraction of a percent over timescales of 1 hour, as long as the Moon does not rise or twilights are close in time. Since we want to model a background map common to, and built from, a series of consecutive images, we must limit the observing sequence in time. It is also important to have a sky photon limited regime in each exposure and reach a reasonable balance with the observing overheads (the camera readout time is 40 seconds). This led us to a single exposure integration time of $\sim$5\,mn. To reach 29 magnitudes per square arcsecond with Elixir-LSB, early studies demonstrated that $\sim$40\,mn total integration is enough (beyond that point, the systematics dominate - see below for a discussion).
However, the Elixir-LSB technique requires at least 7 images to derive a proper map of the background, with a telescope offset between exposures large enough to skip over the largest features caused by the astronomical sources: extended galaxies, and reflection halos from bright stars (7 arcmin). At the same time we want the main target of interest to be integrated 100\% of the time: luckily with the large field of view of MegaCam and the apparent size of the galaxies in the \AD\ sample, all features are smaller than 10 arcmin, and can thus be moved within an extended central region of the mosaic while enabling the background subtraction process. This is different from the technique used for the Next Generation Virgo Cluster Survey (NGVS) where the need to map large areas led to a different observing strategy \citep{Cuillandre11,Ferrarese12}
A specific dithering pattern was implemented in the CFHT service observing system to serve our survey, but it also applies to many other programs with sources of similar physical scales. That pattern with offsets in RA and DEC ranging between 2 and 14 arcmin allows the galaxy to never occupy the same physical area of the CCD mosaic across the 7 exposures, and in consequence allows the construction of a background map.
The pattern is also designed such that the mosaic gaps are naturally removed when stacking the images. The offset pattern is shown in Fig.~\ref{fig:obs-seq}. An example of the observing sequence is illustrated in the same figure.
\begin{figure}
\includegraphics[width=\columnwidth]{MegaCam-LSB7-Images.pdf}
\caption{The adopted observing sequence. The offset pattern is shown on the middle panel. The adjacent panels display the 7 individual g--band sky-subtracted images before stacking.
The target galaxy, here NGC~5582, is surrounded by the green ellipse.}
\label{fig:obs-seq}
\end{figure}
Each MegaCam pointing covers one square degree, and the field resulting from the image recombination with our LSB-optimized technique would yield a much larger one, but we limit our stack sky coverage to the central 63 $\times$ 69 arcmin where the signal-to-noise is the highest, though still with a lower sensitivity at the edges. The resulting weight map is shown in Fig.~\ref{fig:obs-weight}.
\begin{figure}
\includegraphics[width=\columnwidth]{Elixir-LSB7-WeightMap-inv.pdf}
\caption{Weight map resulting from the stacking process. Darker regions have a higher pixel redundancy and thus sensitivity. The green square corresponds to the area kept for the final stacked image (63 $\times$ 69 arcmin).}
\label{fig:obs-weight}
\end{figure}
On the stacked image, the target galaxy is located close to the centre of the field, precisely at position +0.1',+2.1' with respect to the centre. Within the target field of view, other \AD\ galaxies, most often lying in the same group as the primary, may be present. In that case, observations were generally not duplicated. Depending on their location, these secondary targets have images with a slightly lower sensitivity (but see below); the precise location of each object within the original frame is shown in the on--line version of the catalog. Due to technical issues with the telescope scheduling system, a few galaxies were observed twice and their stacked images were made with 12--14 individual exposures instead of 7. Conversely, observations made during the first observing run consisted of only 6 individual exposures. The number was adjusted to 7 for the following runs to optimize the background subtraction.
\subsection{Background correction}
The basic idea is to derive a map of the background which is mostly a large radial pattern caused by the sky background reflection in the optics and the photometric illumination correction to deliver a photometric flatness across the field of view. The entry frames for the Elixir-LSB pipeline are the regular Elixir frames where that radial structure is left untouched. The amplitude of that radial gradient is nearly 15\% of the sky background.
A multitude of astronomical sources of different angular sizes are present in the background. With the pattern selected we ensure that a column of pixels across the 7 images at any given location on the CCD mosaic will mostly see the sky (no astronomical sources). Extensive testing has demonstrated that this is true for any sky area with a high enough galactic latitude.
The 7 frames are median-stacked and the resulting image is smoothed on a 4 arcmin scale to reject possible artefacts due to a higher number of astronomical sources in the column of images at a given location. This map is then scaled back to each image's sky background level, and subtracted. The result is an image with sky background perfectly flat, with residuals left from overcrowding over an area of the CCD mosaic across the 7 images. In that case, there will be a slight overcorrection of the background, leaving a dip in the corrected image. However in the worst case, the deviation is not more than $\sim$0.4\% of the sky background (max-min/background). We thus consider that 7 exposures achieve a satisfactory level.
At the end of the background correction process, each of the 7 images now has a background flat within 0.4\% versus the 15\% initially. This means the true sky background (flat, at least in g and r) has been restored. An example of a background subtracted individual image, illustrating the efficiency of the method, is shown in Fig.~\ref{fig:skysub}.
\begin{figure}
\includegraphics[width=\columnwidth]{NGC5557-sky.pdf}
\caption{The dramatic gain of the LSB-optimized observing strategy and data reduction. The left panel shows one individual g--band exposure of NGC~5557 for which a bias and standard flat-field correction has been applied. At the chosen intensity contrast, the central regions of the frame are polluted by an extended scattered emission. On the individual image processed by Elixir-LSB shown on the right panel, the central scattered light has disappeared and the faint LSB tidal features around the galaxy show up. The CCD patterns which are visible on this figure are removed by the final image stacking.}
\label{fig:skysub}
\end{figure}
\subsection{Image stacking}
The 7 images are run through the AstrOmatic astrometry package {\sc SCAMP} \citep{Bertin06}, along the other set of images of the given target to ensure a uniform solution. The {\sc SCAMP} output is used by the AstrOmatic resampling package {\sc SWARP} \citep{Bertin10} with the image sampling going from 0.19 arcsec per pixel to 0.54 in order to facilitate the processing speed, but mostly to boost the signal-to-noise ratio with the filtering effect of the 3 by 3 average binning. The overall astrometry precision is about 0.05 arcsec (1/10th of a pixel).
The sky background (now the true sky with a constant value across the image) is precisely measured and subtracted. The final step consists in stacking the frames using a standard signal-clipping approach to reject artefacts (satellite tracks, cosmic rays).
Stacking the frames by averaging areas of the mosaic which have seen different parts of the sky further decreases the artefacts left by "vertical" astronomical crowding in the column of images when building the background. Elixir-LSB stacks deliver for MegaCam images flat at the $\sim$0.2\% level of the sky background.
The absolute photometric calibration produced by Elixir is tracked through the entire process and the stacks are calibrated in the MegaCam AB natural magnitude system from where transformation to other systems such a Sloan are readily available.
\subsection{Limitations of the approach}
0.2\% of the sky background means it is possible to detect low surface brightness features some 7 magnitudes fainter than the sky background. That limit has proven to be the same across all 5 broad band filters available on MegaCam. For the g-band, this corresponds to a 28.5 magnitude per square arcsecond limit.
Because that limit of 0.2\% is the same across all bands, this hints that the approach is limited by systematic errors at that level. Indeed, deeper integration provides only minor gains in flatness. The origin of the problem lies in the optical nature of the instrument with a 4-lens wide-field corrector added to the regular optics of the camera, leading to a large number of possibilities of internal light reflection. Here the problem is not the sky background but the many sources seen in the camera beam that all create their own set of halos, most of them too faint to be seen as the ones caused by bright stars, but all contributing to making the background uneven enough between exposures. Effects of these halos on the science results are discussed in Sect.~\ref{sec:halos}.
There is an on-going effort in modeling the halos in MegaCam for each individual exposure and correct them to further boost the performance of Elixir-LSB (Cuillandre et al., in preparation). If that effort is conclusive surely the whole survey would gain in depth, possibly reaching the $\sim$30~${\rm mag\,\,arcsec^{-2 }}$\ limit.
\subsection{Filter selection and observing time}
The choice of filters was primarily driven by our aim to detect low--surface brightness features and extend the study of the stellar populations of ETGs to large radii (up to 10~\mbox{R$_{\rm e}$}). We have selected the filters that maximize the contrast between the sky background and the debris while keeping the exposure time reasonable: g' and r' \footnote{The u*, g', r', i', z' filters of MegaCam slightly differ from the Sloan u, g, r, i and z filters, hence their different naming. However, for simplicity, throughout the paper, we will refer to them as u, g, r and i.}. From the derived g-r colour map, constraints on the stellar age and metallicity gradients may be obtained, though with a large degeneracy.
As argued in Sect.~\ref{sec:FS}, colour information is also useful to distinguish tidal tails formed in major mergers, composed of mixed, metal--rich, stellar populations expelled from their parent spiral galaxies, to stellar streams that result from the disruption of low--mass satellites primary composed of low--metallicity stars.
In addition, i band images were obtained for a sub-sample of ETGs, allowing in particular a comparison of the colour profile with galaxies located in the Virgo Cluster which were also observed with MegaCam as part of the NGVS but for which the r band is missing.
Finally, u band images were acquired for a few ETGs with distances below 20~Mpc. The u band helps to identify star--forming regions in particular in gas--rich outer regions of ETGs such as discs and tidal debris. However, the primary motivation for the u band observations is the detection of globular clusters (GCs). Combining u with g, r and i helps to identify GCs against foreground stars and background distant galaxies (Lan\c{c}on et al., in prep).
These multi-band images were combined to compute the ``true colour" images shown in the catalog.
Given the observing strategy defined in Sect.~\ref{sec:strategy}, exposure times were typically 7 $\times$ 345 sec in g and r bands, 7 $\times$ 230 sec in the i band and 7 $\times$ 700 sec in the u band. The list of available bands, total exposure times, number of individual exposures, image quality and background level for the first galaxies in the catalog is given in Table~\ref{tab:obsmeg}. The full table with all observed galaxies is available in the online web version.
\begin{table}
\caption{Catalog of ETGs with MegaCam observations (first entries; full table available in the online web version)}
\begin{tabular}{lccccl}
\hline
Galaxy & Band & N & Integration time & IQ & Background \\
& & & s & arcsec & ADUs \\
(1) & (2) & (3) & (4) & (5) & (6) \\ \hline \\
\hline
NGC0448 & g' & 7 & 2415 & 0.64 & 704.29 \\
& r' & 7 & 2415 & 0.77 & 941.57 \\
& i' & 7 & 1610 & 0.65 & 1054.86 \\
\hline
NGC0474 & u* & 7 & 4900 & 1.16 & 223.43 \\
& g' & 7 & 2415 & 0.83 & 784.43 \\
& r' & 7 & 2415 & 0.66 & 935.43 \\
& i' & 14 & 3220 & 0.65 & 1080.79 \\
\hline
NGC0502 & g' & 7 & 2415 & 0.82 & 688 \\
& r' & 7 & 2415 & 0.79 & 981.29 \\
& i' & 7 & 1610 & 0.66 & 1626.29 \\
\hline
\multicolumn{6}{l}{\parbox{8cm}{Notes: (3) Number of individual exposures (4) Total integration time (5) Image Quality: FWHM of the PSF (6) Background level}} \\
\end{tabular}
\label{tab:obsmeg} \\
\end{table}
\subsection{Observing conditions and image quality}
All observations were obtained in dark time, far from the twilights, so as to maximize the signal-to-noise of LSB structures and avoid varying sky background.
A large fraction of our images were obtained with seeing better than 1 arcsec for u, g, r and 0.7 arcsec for i, i.e. with observing conditions that are not needed for LSB science. The Image Quality of the MegaCam survey enables to address additional scientific objectives, such as the identification of Globular Clusters around the ETGs or the estimate of their distance with the technique of surface brightness fluctuations.
Thus one of the main values of the MegaCam deep imaging project is to benefit from an observing strategy optimized for the detection of LSB structures while also keeping the excellent spatial resolution of MegaCam and seeing of the Mauna Kea.
\section{Deep imaging survey performances and limitations}
\label{sec:perf}
We discuss here the performances and limitations of our MegaCam survey. Note that some of the issues raised here, in particular the extended reflection halos around bright objects and Galactic cirrus emission, are of broad interest as they plague similar deep imaging surveys.
\subsection{Instrumental signatures}
Some instrumental signatures remain visible after the Elixir-LSB and stacking processes. They are invisible when cutting the images at 27 ${\rm mag\,\,arcsec^{-2 }}$\ , but become prominent at fainter surface brightness limits.
We discuss here how they affect the science analysis and propose ways to minimize them.
\begin{figure}
\includegraphics[width=\columnwidth]{NGC0474_tcolfi.pdf}
\caption{g--band image of the field around NGC~474. The chosen intensity scaling enhances the fainter extended low surface brightness features, including the extended reflection halos around the bright stars and other instrumental signatures, such as horizontal and vertical bands corresponding to lower sensitivity in the CCD gaps and at the edges of the recombined frames. On such images, non circular or rectangular features, and oblique filaments are real structures on the sky. The target galaxy lies at the centre of the white square, the size of which corresponds to 40 times the effective radius of the ETG. To better identify the objects, a composite g+r+i image has been superimposed.
The field of view of the large-scale image is 57 \mbox{$\times$} 64 arcmin. North is up and East left.
Similar images are provided for all our targets in the image catalog.}
\label{fig:inst}
\end{figure}
\subsubsection{CCD gaps}
MegaCam is a mosaic of 36 CCDs with gaps of 13 to 80 arcsec between them. On the stacked images, their signature are horizontal and vertical bands of higher noise (see examples in Fig~\ref{fig:inst}). Such geometrical patterns are easily identified and most often cannot be misidentified with real celestial objects. Besides, they do not affect much the photometric measurements.
\subsubsection{Stellar halos}
\label{sec:halos}
More detrimental are the halos of bright stars, as their circular shape gets closer to that of real galaxies, especially those seen face-on.
They are superimposed on the classical wings of the instrument PSF.
The outer envelope of these features resembling out of focus stellar images has a typical radius of 3.5 arcmin.
The halos are of low surface brightness: most of them have mean surface brightness in the g band fainter than 26.5~${\rm mag\,\,arcsec^{-2 }}$\ (see Fig.~\ref{fig:star-halos}).
However, in LSB optimized deep images, they occupy large areas. As an example, in the MegaCam image shown in Fig.~\ref{fig:inst}, visible halos cover about one fifth of the frame. As a consequence these extended structures often spatially overlap with the outer regions of galaxies and contaminate their photometry.
Such halos, present in many imaging instruments, are imprints of internal reflections between the CCDs and different optical elements of the camera, including the dewar window and filters \citep{Slater09}.
They have complex shapes: the multiple internal reflections produce several more or less concentric overlapping discs; the exact position of the central star with respect to them depends on its position on the frame.
Furthermore, their brightness depends on the filter used; in our survey they are mostly prominent in the r--band and appear in green on our composite g+r+i images (see Fig~\ref{fig:halos}).
Indeed, the reflectivity of MegaCam has been minimized in the blue domain. At longer wavelength (i--band and beyond), the halos are lost in the sky background, so that their brightness peaks in the r--band.
Subtracting these halos is a tedious but necessary task.
We have modeled and removed the most prominent stellar halos located close to the target in an empirical and interactive way. We fitted the external reflection pattern with a disc of constant brightness, which we subtracted from the image. We iterated fitting the residual image with another disc of smaller size and slightly different centre. The process was pursued until only the central brightest - saturated - spot remained on the residual.
This halo subtraction requiring manual interaction was particularly time consuming, but proved efficient in revealing the structures of galaxies even rather close to bright stars. It basically leaves only the imprint of the obstructing support spider of the telescope mirror (see Fig.~\ref{fig:star-halos}).
As mentioned earlier, the exact shape of the halos depends on the position of the star in the MegaCam field. Our final image is a combination of individual frames which were shifted on the sky by large offsets. This resulted in stellar halos with an even more complex structure than on an individual frame. Ideally, the halos should be removed in the 7 individual frames before stacking, multiplying by the same amount the time necessary to remove them. To avoid this, the stellar halos were computed from the final image and models were slightly blurred.
\begin{figure}
\includegraphics[width=\columnwidth]{NGC2764-halos-r.pdf}
\caption{Imprints of the internal reflections of bright stars on deep images with MegaCam. {\it Top-left:} composite g+r--band image of the field around NGC~2764. The position of the ETG is indicated by the white square. {\it Top-right:} empirical model of the reflections in the r--band. A surface brightness scale, ranging between 24 and 30 ${\rm mag\,\,arcsec^{-2 }}$\ , is used. {\it Bottom-left:} original r--band image. {\it Bottom-right:} resulting image at the same intensity scale after subtraction of the stellar halos.}
\label{fig:star-halos}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{halos.pdf}
\caption{Extended low surface brightness nuclear and stellar halos as seen on composite g+r+i images of the galaxies NGC~5473 (left) and NGC~3489 (right).
The nucleus of each galaxy generates its own ghost halo which has a size and colour roughly similar to that of the surrounding bright stars.}
\label{fig:halos}
\end{figure}
Some surveys have circumvented this difficulty by minimizing light scattering and reflections with light-absorbing material and antireflective coatings inside the telescope and the camera \citep{Mihos05,Slater09}, or developing concept cameras with radical different technologies \citep[see the Dragonfly Telephoto Array,][]{vanDokkum14}. Images generated by amateur cameras with simple optics also suffer much less from internal reflections and produce images that at first sight may be considered to be cleaner. An instructive comparison between the CFHT image of the field around the galaxies NGC~474 and NGC~467 obtained with CFHT/MegaCam and with a small 12 inch (0.3~m) telescope as part of a collaboration with amateur astronomers is presented in Fig.~\ref{fig:pro-amateur}. Most of the LSB features revealed by the MegaCam camera are visible on the amateur image, though the sky on the amateur image is not perfectly flat. Obviously, compared to cameras installed on amateur telescopes or just made of telephoto lenses, MegaCam on the CFHT benefits from the use of a 4-meter class telescope, allowing to reach a similar surface brightness limit in 30 times less observing time.
\begin{figure}
\includegraphics[width=\columnwidth]{NGC0474-irida-vs-megacam.pdf}
\caption{Images of the same field around NGC~474 (to the East) and NGC~467 (to the West) obtained with two different cameras and telescopes: MegaCam on the 3.6m CFHT (top) -- total exposure of 0.7 hour in the g-band -- and an ATIK 4000m CCD camera mounted on a 12" RC amateur telescope, located on the site of the Bulgarian National Astronomical Observatory Rohzen -- total exposure time of 21.5 hour with the Clear Luminance filter (L); Image credit: Irida telescope, Velimir Popov and Emil Ivanov -- Note the absence of large stellar halos in the image obtained by the Irida telescope.
}
\label{fig:pro-amateur}
\end{figure}
\subsubsection{Nuclear halos}
A potentially more devastating effect of internal reflections, which until recently has been often minimized, is the radial spreading of the brightest parts of galaxies, i.e. their nucleus, at large effective radius. This PSF far wing generates structures which may be mistaken for real stellar halos \citep{Michard02,Sandin14}. With shallow imaging surveys, the PSF signature may be seen in stacked images
\citep{deJong08,LaBarbera12,Dsouza14}. With our deep imaging survey, it is directly visible on individual images.
Fig.~\ref{fig:halos} displays composite g+r+i images of the fields surrounding NGC~5473 and NGC~3489. It is remarkable that the outer structures of these galaxies share the same roundish shape, size and even colour as the reflection halos due to the nearby bright stars. They have in fact the same origin. The compact galactic nuclei generate halos which are just a little more fuzzier than those produced by point--like stars.
In the case of NGC~5473, the fake reflection halo is easily identified as it is larger than the stellar halo of the galaxy and has a different almost monochromatic (green) colour. For NGC~3489, the ghost and stellar halos have about the same size, but because of the large ellipticity of the galaxy, the fake halo is still visible along the minor axis. When the apparent size of the galaxy exceeds that of the reflection halo -- typically 3.5 arcmin --, the latter is no longer directly distinguishable. Nonetheless it may still significantly affect the photometry of the outer regions of the galaxies, in particular falsify their colour. This is particularly evident in Fig.~\ref{fig:halo-color} showing the g-r colour map of NGC~3489. Whereas the galaxy has for most of its extent a uniform colour $g-r = 0.6$, it exhibits an outer annulus with $g-r = 1.1$.
This reddening is best explained by an internal instrumentation reflection: it reflects the red colour of the galactic nucleus and the fact that the r band is particularly sensitive to such reflections.
A significant number of the \AD\ galaxies do have a similar reddening which may then be considered as suspicious (Karabal, et al., in prep.).
It badly affects the interpretation of the outer colour gradients in terms of stellar populations.
In fact, such an effect may also be responsible for the reddening of the stellar halos described in a number of published studies but which is still controversial \citep{Bergvall10,Jablonka10,Zackrisson12}.
In particular the presence of very red halos on extremely deep images of some spiral galaxies has generated suggestions about possible unconventional stellar Initial Mass Functions \citep{Bergvall10,Bakos13} but was questioned as they could not be seen in nearby galaxies which are resolved into stars. Our MegaCam images provide an unambiguous confirmation of the instrumental origin for at least some of the red halos disclosed through their diffuse light.
Thus a proper study of the outer stellar populations requires to take into account the presence of ghost halos, especially in case the studied galaxy hosts a bright compact nucleus.
This requires a physical modeling of the reflections within the telescope and camera which may be achieved with ray--tracing experiments (Regnault et al., in prep.).
\begin{figure}
\centerline{\includegraphics[width=0.8\columnwidth]{halo-col.pdf}}
\caption{ g-r colour map of the galaxy NGC~3489. The reddening by 0.5~mag in the outer annulus is most likely due to the internal reflection of the red galaxy nucleus. The scale in mag is shown to the right.}
\label{fig:halo-color}
\end{figure}
\subsection{Sky pollution: Galactic cirrus}
In some fields, the detection of the low surface brightness stellar structures is hampered by the presence of contaminating features covering an even larger area than the reflection halos: Galactic cirrus.
Far infrared/millimetric emission from cold dust located in the Milky Way is well documented and appears as one of the principle hurdles in distant galaxy surveys in this wavelength regime as well as for the study of the Cosmic Microwave Background (CMB). It is less known that Galactic cirrus also affects the optical regime through their scattered light emission, although this light was clearly visible on high-contrast prints of the Palomar Sky Survey \cite[e.g.][]{Sandage76}.
At the depth of the MegaCam deep imaging survey, cirrus optical emission becomes prominent. In the image shown in Fig.~\ref{fig:cirrus}, they occupy about half of the frame. The close match between the diffuse optical structures and the 857 GHz (350~$\mu$m) emission as traced by the HFI camera on board the Planck telescope leaves no doubt on the origin of these structures.
At the resolution of MegaCam -- 300 times better than HFI --, Galactic cirrus shows up as multiple long, narrow filaments sharing the same orientation on the sky over fields of 10--30 arcmin. As seen in Fig.~\ref{fig:cirrus}, cirrus emission becomes prominent only at surface brightness in the the g band fainter than 26~${\rm mag\,\,arcsec^{-2 }}$\ , and its colour is globally uniform though locally colour gradients may be seen. In particular some filaments appear blue on the composite g+r image (see also Fig.~\ref{fig:cirrus-detail}).
\begin{figure*}
\includegraphics[width=\textwidth]{cirrus-wsb.pdf}
\caption{Contamination by Galactic cirrus.
{\it Middle:} Composite g+r MegaCam image of the field around NGC~2592 and NGC~2594. The position of each ETG is indicated by the white cross. The field of view is 51\x57 arcmin. {\it Left:} The same field observed by the HFI camera on board Planck at 857~GHz (350 $\mu$m). Note the very good match between the extended emission in the far IR and the optical, though the MegaCam image has a spatial resolution which is 300 times better than HFI. {\it Right:} Surface brightness map in the g-band. The scale in ${\rm mag\,\,arcsec^{-2 }}$\ is indicated to the right. Note that most cirrus emission shows up at surface brightness below 26~${\rm mag\,\,arcsec^{-2 }}$\ (shown in red). }
\label{fig:cirrus}
\end{figure*}
Unfortunately for the purpose of this survey, filamentary structures due to cirrus emission may resemble stellar tidal tails. A particularly striking example is shown on Fig.~\ref{fig:cirrus-detail}. The long feature emanating East of NGC~7457, which has the same colour as the ETG halo, resembles that of a stellar stream from a tidally disrupted satellite. Zooming out, one notes however that similar structures are present all over the field. Thus most likely the feature is a cirrus. Zooming in, an experienced observer will remark that real stellar streams are less striped than cirrus filaments.
Although cirrus is quite easily identified in most cases, either using complementary multi-wavelength data or examining in detail the MegaCam images, there is no way to subtract it, like it is done for instance to generate CMB maps, without at the same time erasing the real LSB stellar structures.
About 30~\% of the MegaCam images in our survey exhibit emission from cirrus. For about 15~\% of them, the contamination is so strong that the detection of low surface brightness stellar features becomes quasi impossible. The degree of contamination depends on the galaxy latitude and may be predicted from the Planck emission at 857~GHz. For that purpose we have extracted from the Planck archives the area covered by the MegaCam images and measured within them the average flux (mostly dominated by foreground cirrus emission).
In fields with a mean 857~GHz flux exceeding 1.5 MJy/sr, 100~\% of MegaCam images are contaminated. Between 1 and 1.2 MJy/sr, the contamination level is 50~\%. It is only below fluxes of 0.4 MJy/sr, that the contamination becomes less than 10~\%.
Whereas the optical emission of Galactic cirrus is a stumbling block for extragalactic research, it may be of great interest for the study of the interstellar medium. Indeed, the MegaCam images provide us with maps of Galactic cirrus at the unprecedented spatial resolution of about 1 arcsec whereas direct images in the FIR, for instance with Planck, have a resolution 300 times worse. A study dedicated to the cirrus emission will be published in a separate paper.
\begin{figure*}
\includegraphics[width=\textwidth]{tail-cirrus.pdf}
\caption{A filamentary structure near NGC~7457 resembling a tidal stellar stream but most likely due to Galactic cirrus, as suggested by the zoom out g--band image to the right exhibiting many similar structures all over the field. The extended halo of the nearby star has been subtracted from the composite g+r image to the left. }
\label{fig:cirrus-detail}
\end{figure*}
\subsection{Surface brightness limit and comparison with other surveys}
\label{sec:limit}
A key parameter for any deep imaging survey is the limiting surface brightness. Unfortunately its value is much more difficult to estimate than the limiting magnitude of traditional imaging surveys obtained with aperture photometry of point-like sources. This is largely due to the fact that the limiting factors are the systematics in the background.
The large range of limiting surface brightness quoted in the literature for LSB optimized surveys does not only reflect the range of exposure times, telescope size used and real achieved depth, but also the diversity of methods used to do the photometry.
To set the scene, the SDSS -- the current reference for imaging survey -- has a limiting surface brightness in the g--band of 26.4~${\rm mag\,\,arcsec^{-2 }}$\ \citep{York00,Kniazev04}, with a gain stacking the g, r and i bands \citep{Miskolczi11}, or for fields with repeated observations such as in the Stripe~82 calibration area \citep{Kim13}.
\noindent $\bullet$ \cite{Atkinson13} used the CFHTLS-Wide imaging survey (which was not LSB optimized) to probe tidal features around galaxies within the redshift range 0.04 -- 0.2. They estimated a limiting surface brightness of 27.7~$\pm$ 0.5 ~${\rm mag\,\,arcsec^{-2 }}$\ in the g--band based on histograms of rms variations of the sky noise estimated in multiple regions.\\
\noindent $\bullet$ \cite{Tal09} used pointed observations with the SMARTS 1m telescope to determine the frequency of tidal features in a sample of 55 elliptical galaxies. They heavily smoothed dark-sky and target images to the scale of typical tidal features and determined the 1$\sigma$ detection threshold of the latter.
They obtained a limiting surface brightness of 29~${\rm mag\,\,arcsec^{-2 }}$\ in the V(Vega)--band -- , and 27.7 (about 27.9 in the MegaCam g--band, and AB scale), using flatness-limited frames.\\
\noindent $\bullet$ The pilot survey of \cite{Martinez-Delgado10} of tidal streams around nearby spirals done with a luminance (L) broad filter and robotic amateur telescopes could not be directly photometrically calibrated, but using as a reference SDSS images, the authors estimated from the background fluctuation an equivalent V/g--band limiting magnitude of 28.5~$\pm$ 0.5~${\rm mag\,\,arcsec^{-2 }}$\ . \\
\noindent $\bullet$
The LSB--optimized NGVS at the CFHT \citep{Ferrarese12} reaches 29~${\rm mag\,\,arcsec^{-2 }}$\ in the g--band, a value checked comparing the surface brightness profile of the galaxy M49 with previously determined ones in the literature. \\
\noindent $\bullet$
\cite{Bridge10} used the deep component of the CFHTLS to estimate the merger rate of galaxies up to a redshift of 0.2. They claim a very low limiting surface brightness in the i--band of 29~${\rm mag\,\,arcsec^{-2 }}$\ (thus about 30~${\rm mag\,\,arcsec^{-2 }}$\ in g), but do not state in their paper the method used to estimate it. \\
\noindent $\bullet$
\cite{Sheen12} probed post-merger signatures in nearby cluster galaxies on images obtained with the Blanco 4m telescope supposedly reaching 30~${\rm mag\,\,arcsec^{-2 }}$\ in r (about 30.7 in g).\\
\noindent $\bullet$
Finally\footnote{A more comprehensive census of previous imaging surveys may be found in Table~1 of \cite{Atkinson13}.},
\cite{vanDokkum14} testing his LSB-dedicated camera -- the Dragonfly Telephoto Array -- on the spiral M101 holds the current record of claimed limiting surface brightness with an amazing value of 32~${\rm mag\,\,arcsec^{-2 }}$\ in g, estimated looking at the shape and extent of the galaxy radial profile. If confirmed, such a depth would reach that obtained with star counts in the Local Group, offering a tremendous opening for galactic archeology in the nearby Universe.
How does our survey compare with previous ones? As explained in Sect.~\ref{sec:obs}, the Elixir-LSB pipeline achieves a sky flattening of 0.2\%, nearly 7 magnitudes fainter than the sky background.
We estimated a nominal detection limit at 28.5~${\rm mag\,\,arcsec^{-2 }}$\ in g, determining the pixel value difference between the bump and holes on clean sky regions.
Whether this nominal value is realistic can be checked looking at the 2D surface brightness maps and 1D profile of real galaxies.
Examples of surface brightness maps are shown in Fig.~\ref{fig:sb} with the colours of the intensity scale chosen to highlight structures fainter than 26~${\rm mag\,\,arcsec^{-2 }}$\ , i.e. structures that are not typically seen in regular imaging surveys such as the SDSS.
In these images, all features until at least 28.5~${\rm mag\,\,arcsec^{-2 }}$\ are ``real": they do not have the shape of instrumental signatures (see above). This gives an upper limit to our limiting surface brightness sensitivity.
Surface brightness profiles will be discussed in another paper.
\begin{figure*}
\includegraphics[width=\textwidth]{sbmaps-re.pdf}
\caption{Surface brightness maps in the g-band of several ETGs. The brightness scale used, shown at the top--right -- ladder-type in grey scale below 26~${\rm mag\,\,arcsec^{-2 }}$\ , linear and red above --, enhances the low surface brightness structures disclosed by the deep MegaCam images. The galaxies, resp. NGC~516, NGC~509, NGC~2695, NGC~502, NGC~474 and NGC~2577 are ordered by relative increased contribution of this LSB component to the total light. 26~${\rm mag\,\,arcsec^{-2 }}$\ \ roughly corresponds to the surface brightness limit of regular imaging surveys such as the SDSS. Each bar corresponds to 10~kpc at the distance of the galaxy.}
\label{fig:sb}
\end{figure*}
\section{Data products}
\label{sec:prod}
\subsection{Sample presented in this paper and selection biases}
\label{sec:sample}
The \AD\ sample is volume limited: it includes galaxies located at a distance below 42~Mpc, classified as early-type based on a visual classification criterion -- the absence of discs and prominent dust lanes on shallow images --, with absolute K--band magnitude brighter than -21.5 (i.e. a stellar mass above $6 \mbox{$\times$} 10^{9}~\mbox{M$_{\odot}$}$) -- thus dwarf ETG galaxies are excluded --, and further restrictions on declination -- the galaxies should be visible from northern facilities -- and Galactic latitude \citep[see details in][]{Cappellari11}.
All galaxies in the \AD\ sample benefit from a wealth of ancillary multi-wavelength data, including the internal stellar kinematics derived from SAURON observations, based on which they have been classified as slow or fast rotators \citep{Emsellem11}.
We present here an image catalog for a sub--sample of 92 galaxies which have been observed with MegaCam through at least the g and r filters. Among those 59~\%, have also i--band images and 9~\%, u--band images.
The ETGs presented in this paper are located in environments with low to medium galaxy density, excluding the Virgo Cluster.
This is illustrated in Fig.~\ref{fig:loc} showing for the entire \AD\ sample and the sub-sample discussed here the stellar mass versus $\rho_{\rm 10}$, i.e. the volume density in Mpc$^{3}$ of galaxies inside a sphere of radius r$_{\rm 10}$ centred on a galaxy, which includes N$_{\rm gal}=10$ nearest neighbours \citep{Cappellari11b}.
This sub-sample of 92 galaxies -- already by far the largest sample of early-type galaxies with available deep images -- spans all the \AD\ range of masses, but high mass and extended galaxies are over-represented (see Fig.~\ref{fig:MRe}). As a consequence of the selection criteria used for the MegaCam runs, a large fraction of the rather rare slow rotators have deep images, while the numerous fast rotators are much less sampled (see Fig.~\ref{fig:rot}). Finally, the sample is also biased towards gas--rich objects, i.e. ETGs for which H\,{\sc i}\ and CO line emission has been detected.
The CFHT Large program MATLAS aims at eluding such selection effects. When completed, this survey done with the same observing strategy as the one described here, will provide deep images at similar depth for all 260 \AD\ galaxies.
Note that the ETGs located in the Virgo Cluster -- about one fourth of the \AD\ sample -- were already observed with MegaCam as part of the CFHT Next Generation Virgo Cluster Survey \citep[NGVS,][]{Ferrarese12}. Their images will be published in the framework of the NGVS collaboration (Duc et al., in prep.).
Consequently this paper focuses on general issues faced by deep images and addresses effects which are not influenced by selection biases while the in-depth analysis statistical analysis is postponed to future papers.
\begin{figure}
\includegraphics[angle=-90,width=\columnwidth]{M-ro.pdf}
\caption{Selected ETGs in this catalog plotted on a diagram showing the dynamical mass, M$_{\rm JAM}$, derived from modeling of the stellar kinematics and the isophotal structure of each galaxy, versus the volume density diagram, $\rho_{\rm 10}$, in log (Mpc$^{-3}$).
Open circles correspond to the full \AD\ sample, and filled discs are galaxies with available deep imaging. Note that all galaxies located in regions with density above log($\rho_{\rm 10})=-0.4$ belong to the Virgo Cluster and were observed as part of the NGVS project (filled grey discs). They will be presented in a separate paper. Relaxed galaxies showing no sign of tidal perturbations even with the deep imaging are shown in green.}
\label{fig:loc}
\end{figure}
\begin{figure}
\includegraphics[angle=-90,width=\columnwidth]{M-Re.pdf}
\caption{Selected ETGs in this catalog plotted on a diagram showing the dynamical mass, M$_{\rm JAM}$, derived from modeling of the stellar kinematics and the isophotal structure of each galaxy versus the effective radius, \mbox{R$_{\rm e}$}\ \citep[see][]{Cappellari13}. Open circles correspond to the full \AD\ sample, and filled discs are galaxies with available deep imaging (grey ones belong to Virgo and were not analyzed here). Relaxed galaxies showing no sign of tidal perturbations even with the deep imaging are shown in green. }
\label{fig:MRe}
\end{figure}
\begin{figure}
\includegraphics[angle=-90,width=\columnwidth]{eps-LR.pdf}
\caption{Selected ETGs in this catalog plotted on a diagram showing the specific angular momentum at the effective radius, $\lambda_{\rm R_e}$, versus the ellipticity, $\epsilon$. The solid line divides the fast (top) and slow (bottom) rotators according to the \AD classification scheme \citep{Emsellem11}. Open circles correspond to the full \AD\ sample, and filled discs are galaxies with available deep imaging (grey ones belong to Virgo and were not analyzed here). Relaxed galaxies showing no sign of tidal perturbations even with the deep imaging are shown in green. }
\label{fig:rot}
\end{figure}
\subsection{High level image production}
\subsubsection{True colour images}
True colour images obtained combining multi-wavelength bands have not only an aesthetic value\footnote{Images from the MegaCam survey have appeared in calendars, in an art gallery and the Astronomical Picture of the Day (see http://apod.nasa.gov/apod/ap140105.html )}, they are particularly efficient at synthesizing information.
Depending on the filters available, the composite images were either produced using the g (blue channel), r (green channel) and i (red channel) or only with the g and r bands. In the absence of a third band, a fair true colour rendering may be obtained using as the middle green channel the combination of the blue and red image (here g and r). An arcsinh intensity scale was applied to each channel in order to decrease the dynamical scale and make visible both the central and outer regions. The same weighting for the red, green and blue channels was applied to each galaxy, enabling a qualitative comparison of the colours between objects observed with the same filter set. Final image combination was carried out with {\sc STIFF} \citep{Bertin11}.
\subsubsection{Surface brightness maps}
Surface brightness maps with an intensity scale in ~${\rm mag\,\,arcsec^{-2 }}$\ \ are computed on sky--subtracted images. As the Elixir-LSB process already produces flat images, a simple constant was subtracted.
To compute the average sky value, a histogram of the pixel values over the full MegaCam stacked image was determined and the mode of the distribution was used. Note that with this method, the sky level might have been overestimated in frames contaminated by extended cirrus emission or full of bright stars.
In order to increase the contrast of LSB structures, an adaptive smoothing algorithm was applied to the data, using the software {\sc ADAPTMOOTH} of \cite{Zibetti10}.
Pixels are grouped together and averaged to keep the same S/N over the whole image. Contrary to regular smoothing techniques, the spatial scale of bright objects, i.e. the central regions of galaxies, stars, etc., is preserved with this scheme.
\subsubsection{Residual images}
\label{sec:model}
Historically, the presence of fine structures in galaxies, in particular ripples and shells, was discovered within their diffuse stellar halos, and was disclosed with various techniques of contrast enhancement such as unsharp masking \citep{Malin77,Malin83,Schweizer88}.
Our sensitivity enables the direct detection of LSB structures well outside the outer galactic halos. Removing the latter allows us however to connect the outer and inner fine structures and make their complete census.
The galaxy modeling required for this was done with two techniques: a multi-component parametrization of the host galaxy with {\sc GALFIT} \citep{Peng02}, and ellipse fitting using the eponymous package within the {\sc IRAF} software \citep{Jedrzejewski87}. We present in this catalog residual images obtained subtracting galaxies modeled by the ellipse fitting algorithm.
Both techniques generally give similar results, except in the very central regions, for which residual images are less noisy with GALFIT (as it may model non axisymmetric components), and in the very outer regions, best subtracted with the ellipse model. As the focus of our project is these external regions, we made the choice of preferentially using the ellipse models in our analysis.
As a first step, point-like objects in the field were identified with {\sc SExtractor} \citep{Bertin96}. Faint stars were removed replacing them with the local background determined from surrounding pixels. The extended reflection halos due to the bright nearby stars have been subtracted with the technique described in Sect.~\ref{sec:halos}. Remaining extended objects not associated to the central ETG, like bright stars, background or companion galaxies, were masked. The ellipse fitting was then performed, leaving whenever possible the central position, axis ratio and angle parameters free. In a few cases, those parameters had to be kept constant to avoid divergence. The resulting ellipse model was then subtracted from the image. The process was iterated to remove additional stars that were not identified in the original image.
Fig.~\ref{fig:ellipse} presents two examples of images with the host galaxy subtracted. Such residual images are especially useful to reveal the shape of fine structures that were barely visible on the original images.
\begin{figure*}
\includegraphics[width=\textwidth]{ellipse-subtracted.pdf}
\caption{Examples of fine structures being revealed after subtracting a galaxy model made with ellipse fitting.
Residual g--band image of UGC~05408 (left) and NGC~3245 (right). The insets show true colour g+r images of the central regions before the galaxy subtraction. }
\label{fig:ellipse}
\end{figure*}
\subsection{Image catalog}
\label{sec:cat}
We present in this catalog images of 92 ETGs. Derived products, such as the surface brightness and colour profiles, will be published elsewhere.
The full atlas, only available in electronic form (see Fig.~\ref{fig:cat} for the first galaxy in the catalog)\footnote{The true colour images, surface brightness and colour maps, image residuals may also be explored on line with the {\sc Google} navigation tool API (including zooming in/out and quick channel change capabilities). Access to the database is available through the \AD\ site http://purl.org/atlas3d.
Raw fits data are available from the Canadian Astronomy Data Centre (CADC) which hosts the CFHT archives.}, consists of a series of images:\\
\begin{itemize}
\item {\it Top left:} g--band image of the original MegaCam stacked frame hosting the ETG. It is shown in grey linear scale with a small cut range so as to increase the contrast of low surface brightness features. The displayed field of view is 57.0 \mbox{$\times$} 57.0 arcmin.
To identify the underlying objects -- stars and galaxies --, a composite g+r or g+r+i image has been superimposed.
The target galaxy is indicated by a white square. Its size corresponds to 40 times the effective radius of the galaxy, as determined in \cite{Cappellari11}. This delineates the area shown in all the other sub-panels in the figure, except that zooming into the central regions.
\item {\it Top right:} a ``true colour" (g+r+i, or g+r) image centered on the galaxy. Contrary to some true-colour images shown throughout the paper, no local correction was done on these images, and the signature of instrumental artefacts was kept.
\item {\it Middle right:} same true colour image zoomed on the central regions by a factor of 6.
\item {\it Bottom left:} g--band surface brightness map.
Between 23.5 and 26~${\rm mag\,\,arcsec^{-2 }}$\ \, the intensity scale used -- a ladder scale, with step of 0.5~mag and grey colours -- discloses the global elliptical shape of the ETG. Above 26~${\rm mag\,\,arcsec^{-2 }}$\ \, the intensity scale is linear and a red colour is used to highlight the new structures revealed by the deep MegaCam survey. The horizontal bar corresponds to 10~kpc at the distance of the galaxy.
\item {\it Bottom middle:} idem but for the r--band. The intensity scale has been shifted by 0.5~mag to take into account the colour of the galaxy.
\item {\it Bottom right:} corresponding g-r colour map. The colour bar, here displayed with a linear scale, is indicated to the right. Pixels below the detection limit are shown in white.
\item {\it Middle left:} residual image obtained subtracting from the g--band a galaxy model made with the ellipse fitting procedure described in Sect.~\ref{sec:model}
\end{itemize}
In all maps, North is up and East left.\\
\begin{figure*}
\includegraphics[width=\textwidth]{NGC0448-cat.pdf}
\caption{Set of images for the first entry in the on--line catalog, NGC~448. See Sect.~\ref{sec:cat} in main text for a detailed description of each panel. }
\label{fig:cat}
\end{figure*}
\section{(Re)-classifying galaxies with deep imaging}
\label{sec:class}
Galaxy classification is now often done based on:\\
-- their global colour: galaxies may belong to the red sequence, green valley or blue cloud.\\
-- ability or not to form stars: galaxies are passive, quiescent, active, star-bursting.\\
-- internal kinematics: galaxies are fast or slow rotators.\\
The kinematics criteria is at the base of the classification scheme proposed by the Sauron survey \citep{Emsellem07} and refined by the \AD\ project \citep{Cappellari11b}, which perhaps tells more about their origin and evolution than other properties.
However, apparent morphologies -- the importance of the bulge vs disc component, the presence or absence of spiral arms and bars, the degree of tidal perturbation -- remain an unavoidable criterion in the taxonomy of galaxies.
The Hubble sequence likely gives a misleading view of the diversity of galaxies but continues to be widely used, even at high redshift.
In that respect although astronomy has for long entered in a multi-wavelength area, optical imaging remains key. In times when the credibility of any discovery relies on its assessment on large samples -- especially in the nearby Universe --, galaxy classification relies on wide but relatively shallow surveys, such as the SDSS. The various galaxy zoo projects \citep{Lintott08} involving hundreds of thousands of citizen scientists inspecting millions of SDSS images has been instrumental in the classification of galaxies.
However, how much were these tremendous efforts impacted by the use of images reaching a limited surface brightness? Our deep imaging survey of already a substantial number of galaxies may address any putative bias. Initial results are presented here.
\subsection{Global morphologies}
We first address the global appearance of the galaxies imaged by MegaCam. The 260 Early-Type Galaxies in the \AD\ sample were initially selected after a visual inspection of a volume limited parent sample of 871 galaxies of all types. Following the classic criteria which define the revised Hubble classification scheme outlined by \cite{Sandage61}, those lacking spiral structure or prominent dust lanes were considered as ETGs. This selection was done based on SDSS images, whenever available, and images from the Digital Sky Survey (DSS). The classification was further checked with images acquired at the Isaac Newton Telescope having the same depth as the SDSS ones. Would the result have been the same if instead the much deeper MegaCam images had been used?
We re-examine here the principle components that supposedly disentangle spirals from lenticulars and ellipticals, but also relaxed, evolved systems from perturbed ones that do not fit on the Hubble sequence.
\subsubsection{Presence of spiral and ring structures}
\begin{figure*}
\includegraphics[width=\textwidth]{disks.pdf}
\caption{Examples of ETGs for which deep optical images reveal the presence of low surface brightness, blue, star-forming discs, rings or spiral structures around them. Composite true colour g+r+i i or g+r images of (clockwise) NGC~2685, NGC~5582, PGC~016060, NGC~7465 and UGC~09519. Field of views and physical scales which vary from one galaxy to the other may be found in the image catalog. Manual local corrections erasing instrumental signatures and bright halos as well as colour enhancements were performed for a better display. Unprocessed images can be inspected in the catalog. }
\label{fig:disk}
\end{figure*}
Fig.~\ref{fig:disk} presents composite true colour images of 5 ``ETGs'', for which MegaCam revealed the presence of an extended blue low surface brightness component surrounding the main (reddish) body. Spiral arms are visible in these structures which are presumably star--forming. Among those only NGC~2685, a polar--ring galaxy nicknamed Spindle or Helix, was already known to host an outer stellar ring \citep[e.g.][]{Sarzi06,Jozsa09}. Note that NGC~5582, classified as ``E'' in the RC3 catalog, would have been classified as a spiral galaxy based on our deep photometry alone, if one considers the disc component extending well outside the galaxy optical radius.
The presence of star--forming regions in the outskirts of some ETGs was already known from UV observations \citep[e.g.][]{Jeong07,Donovan09,Marino11}, with possibly H\,{\sc i}\ fueling them \citep[e.g.][]{vanDriel91,Serra12}. However, their optical counterpart had so far been elusive.
Besides, the previously UV--luminous rings documented in the literature \citep{Salim12} are most often located within the main body of the galaxy, contrary to those revealed by MegaCam located at typically 5 $\mbox{R$_{\rm e}$}$.
In the examples shown on Fig.~\ref{fig:disk} (see in particular UGC~09519), the ``red and dead" component of the ETG is spatially separated from the more active, blue one. This is best seen on the colour maps of these galaxies shown on Fig.~\ref{fig:disk-col}, showing an abrupt blueing of the colour profile at radial distances above 10~kpc.
\begin{figure}
\includegraphics[width=\columnwidth]{disks-col.pdf}
\caption{g-i colour maps of two ETGs with blue discs around them: NGC~2685 (left) and NGC~7465 (right). The colour scale in ${\rm mag\,\,arcsec^{-2 }}$\ \ is indicated to the right.}
\label{fig:disk-col}
\end{figure}
The possible external origin of these apparently decoupled structures will be investigated coupling H\,{\sc i}\ \citep{Serra12}, CO \citep{Davis11} and optical observations.
\subsubsection{Inner disturbances: merger remnants}
\label{sec:mergers}
A scenario for the formation of massive ellipticals through the merger of spiral galaxies was proposed decades ago after the first numerical simulations of collisions had been performed \citep{Toomre77a,Schweizer82,Barnes92}, and has since been a passionate subject of discussion. In any case, if this hypothesis is correct, at least a fraction of ETGs should be post-mergers, and exhibit their emblematic traces: tidally disturbed morphology and prominent dust lanes, also revealing the presence of accreted gas.
Fig.~\ref{fig:mergers} displays 3 examples of ETGs showing on the MegaCam images unambiguous signs of a past major merger event. The dusty ETGs NGC~1222 and NGC~2764 were already classified as peculiar in the NASA/IPAC Extragalactic Database. This was not the case for NGC~5557, classified as an E1 in the RC3 catalog: its post-merger nature revealed by the deep imaging was discussed in \cite{Duc11}.
\begin{figure*}
\includegraphics[width=\textwidth]{mergers.pdf}
\caption{Examples of ETGs with post- major merger signatures: inner dust lanes, strongly perturbed morphology and prominent tidal tails. From left to right, composite true colour images of NGC~5557, NGC~1222 and NGC~2764.}
\label{fig:mergers}
\end{figure*}
We have presented here cases of galaxies presenting a {\it global} disturbed morphology, indicative of a (relatively) recent merger. Many apparently relaxed ETGS in the sample exhibit around them fine structures, including tidal tails, which likely trace on-going interactions, but for some of them could be the vestiges of old collisions. They are addressed in Sect.~\ref{sec:FS}.
\subsubsection{Relaxed systems}
The images presented so far in this paper may give a misleading impression of this deep imaging exercise. Not all galaxies in our sample exhibit external star-forming discs, tidal tails and shells. Even at the depth of the MegaCam survey, a large fraction of them remain the featureless ``red and dead" galaxies that they were thought to be. The images of three of them are shown in Fig.~\ref{fig:relaxed}.
\begin{figure*}
\includegraphics[width=\textwidth]{relaxed.pdf}
\caption{Examples of ETGs which appear totally relaxed/regular even on the deep MegaCam images, and do not shown any fine structure in their vicinity. From left to right, composite true colour images of NGC~3457, NGC~3599 and PGC050395. The high Image Quality of the MegaCam survey is illustrated by these images showing the globular cluster population around the galaxies and background distant clusters of galaxies.}
\label{fig:relaxed}
\end{figure*}
Cosmological simulations made in the framework of the hierarchical model of the Universe predict that each galaxy is the result of multiple minor and major mergers and should thus be surrounded by vestiges of such collisions \cite[e.g][]{Bullock05,Helmi11}. This should be even more the case for ellipticals as they are believed to be the end--product of the mass assembly process at the galactic scale. Determining the fraction of ETGs that do not exhibit such collisional debris, and understanding why -- a property specific to a sub--class of old ETGs, a lack of sensitivity or issues in the models -- are among the major goals of our survey.
\subsection{Fine structures}
\label{sec:FS}
The prime original motivation of most deep imaging surveys is the detection of the so--called ``fine structures". The literature is rather elusive in the definition of this class of astronomical objects which gathers tails, streams, shells, ripples, etc. Authors tend to use different names for the same physical objects or reversely use the same name for objects that have a different intrinsic nature or origin. This is especially problematic for their census and classification, and later on use as archeological probes
\citep[see the reviews on collisional debris by][]{Duc13,Duc13b}.
The following section is an attempt to give a phenomenological description of each type of fine structure and provide a unique name for each different physical class. The ambition is not only to determine the fraction of galaxies having merger vestiges in their vicinity, as usually done, but to make the census of fine structures and distribute them within different sub--classes, as each of them is a probe of different past events: major, minor, wet or dry merger.
Obviously when confronted with noisy images, making an unambiguous distinction between what we later call a tail or a stream may be considered hopeless. However the intrinsic errors in such exercice may at least be compensated by the large number of objects and of volunteers making the classification.
\subsubsection{Tidal tails}
Tidal tails are the generic names of the elongated structures shaped by any tidal interaction. We restrict here the definition to the structures solely made during major mergers. They consist of material expelled from the primary galaxy, following an interaction with a companion massive enough to have significantly perturbed it. As a consequence, the outer regions of the primary -- here the ETG -- and the tails emanating from them share the same properties. This means similar age and metallicity, translating to similar colours, for the stellar tidal tails and galactic halos.
Usually tails are rather prominent, may extend to radial distances above 100~kpc, before gradually dispersing themselves, becoming diffuse, or falling back on the host galaxy.
Examples of prominent tidal tails are shown in Fig.~\ref{fig:tails}.
\begin{figure*}
\includegraphics[width=\textwidth]{interacting.pdf}
\caption{Examples of ETGs currently involved in a tidal interaction with a nearby massive companion and exhibiting prominent tidal tails. Clockwise, composite true colour images of NGC~770, NGC~680, NGC~2698/99, NGC~5507 and NGC~5574/76. The location of the \AD\ ETGs is indicated by a cross.}
\label{fig:tails}
\end{figure*}
\begin{figure}
\centerline{\includegraphics[width=0.6\columnwidth]{NGC3226_HI_tcol_cl2.pdf}}
\caption{The H\,{\sc i}\ and stellar tidal tails of NGC~3226/27. H\,{\sc i}\ contours from the WSRT \citep{Serra12} are superimposed on a composite g--band plus true colour image of the system. }
\label{fig:HI-tails}
\end{figure}
Such structures either trace on-going tidal interactions with a massive companion (the examples shown in Fig.~\ref{fig:tails}) or a past merger.
Collisions between pressure supported bodies are inefficient at forming tidal tails. Thus their detection implies that their progenitors had initially a dynamically cold, and in most cases gas--rich disc component. As a consequence, tidal tails are usually tracers of wet, major mergers. Gas clouds, mainly H\,{\sc i}\, but also molecular gas \citep{Braine01}, are expected to be associated with the stellar component of the tails, and were indeed detected in some stellar tails of ETGs \citep{Duc11}. In that case, such clouds may locally collapse, form stars or even (tidal) dwarf galaxies that will appear as blue condensations within tails of redder colour mainly composed of old stars \citep{Duc14}.
However, the close physical connection between a stellar and H\,{\sc i}\ tail, although frequent -- see the example of NGC~3226/27 in Fig.~\ref{fig:HI-tails} --, is not compulsory. Indeed, in addition to tidal forces that affect stars and gas the same way, the gaseous component may react to additional environmental effects, such as ram pressure. As a result the H\,{\sc i}\ and stellar tails may be separated \citep{Mihos01}. Besides, depending on local conditions both components do not have the same life expectancy, and one may become invisible before the other.
\subsubsection{Tidal streams}
We refer to tidal streams the structures that emanate from a low mass companion which is currently orbiting the primary galaxy or is being ingested by it. We thus adopt the point of view of the most massive galaxy with that definition. Contrary to the tidal tails, the material in the tidal streams differs from that of the main galaxy. Having been expelled from the companion, their stars have thus generally a lower metallicity. As a consequence, on the deep images, streams should be bluer than tails, though in practice the lack of signal in these low surface brightness features makes the comparison rather difficult.
Besides, streams appear as narrow and possibly very long filaments. Their length and shape are not only the result of tidal forces. They trace the orbit of the satellite around its massive host.
Examples of tidal streams are presented in Fig.~\ref{fig:streams}. In all of them, but IC~1024, a bright condensation is visible. It most likely corresponds to the remnant of the tidally disrupted companion.
The linear shape of the streams around NGC~2592 and NGC~5198 suggests a radial recent collision, while the curved shape of the streams around NGC~936 and IC~1024 indicate that the companion has wrapped around its host for a much longer time.
\begin{figure*}
\includegraphics[width=\textwidth]{streams.pdf}
\caption{Examples of ETGs hosting tidally disrupted satellites, as indicated by the presence of stellar streams around them. Clockwise, true colour images of NGC~2592, NGC~5198, NGC~3414, IC~1024 and NGC~936. Prominent cirrus emission is observed in the field of NGC~2592.}
\label{fig:streams}
\end{figure*}
With the adopted definition, streams trace past minor mergers, with a possible bias towards dry ones, given that the satellites of early-type galaxies tend to be gas--poor, having lost their gas by internal -- feedback -- or external -- ram pressure -- processes before they were accreted \citep[e.g.][]{Geha06}. Consequently, contrary to tidal tails, tidal streams are generally not expected to have any H\,{\sc i}\ counterpart and host star--forming regions.
\subsubsection{Shells}
Shells, also known as ripples, are identified by their circular shapes and shape-edged inner structure. Whenever present around a galaxy, they are usually numerous and for at least one class of shells, concentric. Despite an abundant literature devoted to shells and multiple models aimed at reproducing them, their origin is not yet fully understood. Indeed different types of collisions may produce them \citep[see the review by ][and references therein]{Struck99}. It is relevant for our study to note that the physical conditions in intermediate mass mergers favour the formation of shells.
Examples of galaxies in our sample showing prominent shells around them are presented in Fig.~\ref{fig:shells}. Several of the galaxies in our sample, in particularly the truly spectacular galaxy NGC~474, were already known to exhibit shells \citep{Turnbull99}. Indeed, having not a particularly low surface brightness, they were visible in previous generations of imaging surveys. Our deep imaging survey reveals new shells, located in the outermost regions, as well as so far unknown radial linear structures which may have formed at the same time. These structures might provide new constraints for numerical models. The MegaCam images also disclose a variety of colours from one shell to the other -- see in particular NGC~474 --, which should be taken into account when modeling the collision.
\begin{figure*}
\includegraphics[width=\textwidth]{shells.pdf}
\caption{Examples of ETGs exhibiting shells around them. From left to right, true colour images of NGC~474, NGC~502 and NGC~3619.}
\label{fig:shells}
\end{figure*}
Unsharp masking techniques have traditionally been used to disclose the inner shells embedded in the body of early-type galaxies. We have rather used here an ellipse fitting modeling and subtraction (see Sect.~\ref{sec:model}) to reveal them. The technique offers the advantage of enhancing the shells while preserving the more diffuse tidal tails and streams \citep{Forbes92}.
Residual images of the shell galaxies presented in Fig.~\ref{fig:shells} are shown in Fig.~\ref{fig:shells-sub}. Besides the ripples, some of them also show radial streams.
\begin{figure*}
\includegraphics[width=\textwidth]{shells-sub.pdf}
\caption{Shells disclosed by subtracting a galaxy model with ellipse fitting. From left to right, residual g--band images of NGC~474, NGC~502 and NGC~3619. Note also the presence of radial structures, formed either together with the shells during the merger of the companion (likely the case for NGC~474) or during a previous or late independent accretion event (NGC~3619). }
\label{fig:shells-sub}
\end{figure*}
Shell identification is usually unambiguous, especially because of their specific shape and multiplicity. Nonetheless in some cases, they may be confused with wrapping streams, or rings.
\subsection{Eye classification}
Automatic algorithm parametrizing galaxies with coefficients such as the CAS \citep{Conselice03c}, Gini \citep{Abraham03}, or M20 \citep{Lotz04}, have been widely used for morphological classification of large number of galaxies, especially distant ones.
Whereas they may identify highly perturbed galaxies, such as the post-mergers presented in Sect.~\ref{sec:mergers}, they are not sensitive to the LSB components of galaxies, and thus are not able to distinguish the presence or absence of fine structures around apparently relaxed systems. Furthermore, as argued above, identifying each type of fine structure is particularly instructive and allows us, confronting the results with predictions of numerical simulations, to trace back the various types of mergers responsible for the mass assembly of galaxies.
Machines would have a hard time dealing with the subtlety of for instance disentangling a stream resulting from a minor merger from a tail, vestige of a major merger. In such conditions, eye classification remains unavoidable.
In a second step, the use of machine learning algorithms \cite[e.g.][]{Huertas-Company13} that are trained to reproduce the output of the eye classification, may be considered.
The Galaxy zoo project has pioneered the eye classification of large numbers of galaxies with equally large numbers of volunteers filling web--based polls \citep{Lintott08,Willett13}. More recently the CANDELS project also used on--line tools, but with a smaller number of classifiers consisting of professional astronomers \citep{Kartaltepe14}.
We used the latter approach to exploit our survey.
Team members were invited to browse with on--line navigation tools a set of images similar to that presented in the catalog, consisting of true colour images, surface brightness, colour maps and residual images, to identify and whenever possible count the external features.
On average, each galaxy was classified by 8 team members. A statistical analysis was then performed, and galaxies were assigned a type characterizing the properties of their outermost regions, i.e. shape of their halos, frequency and type of fine structures.
The first capital letter of the assigned type summarizes the global status of the galaxy, indicating whether it is fully regular ('R'), involved in an on-going tidal interaction with a massive companion ('I'), in an on-going or past collision with a low-mass companion (C), or show evidence for being a post-merger (M).
Table~\ref{tab:code-status} lists the criteria used to assign the galaxy types and gives references to the figures illustrating them.
In addition, the presence of various types of features around the galaxies is coded: a '+' indicates the presence of stellar structures like ripples (+r), tails (+t), shells (+s), discs (+d) or perturbed halos (+h); a '-' indicates contamination by the halos of nearby bright objects (stars, companion galaxies, -h) or Galactic cirrus (-c).
Further details are given in Table~\ref{tab:code-feature}.
As an example, a galaxy classified as I+t+ph-h like NGC~5574 (see Fig.~\ref{fig:tails}) is involved in an on-going interaction with a massive companion, exhibits tidal tails, a perturbed halo and is embedded within the halo of a bright object; a galaxy classified as R-sc like NGC~509 (see Fig.~\ref{fig:sb}) is apparently fully relaxed but strong cirrus emission is present in its vicinity.
In this classification scheme, an on-going interaction (I) takes precedence over a post-merger (M) which itself takes precedence over a minor merger (C).
Thus a galaxy believed to be involved in an on-going interaction, like NGC~3414 (see Fig.~\ref{fig:streams}), but also showing streams from a disrupted companion, is classified as 'I' rather than 'C'. This scale of priorities corresponds to a decreasing level of expected tidal perturbations on the stellar populations of the galaxy.
\begin{table}
\caption{Criteria used for the galaxy classification}
\begin{tabular}{llp{0.4\columnwidth}l}
\hline
Code & Type & Description & Illustration \\ \hline
R & Fully relaxed & Regular halo; no fine structure & Fig.~\ref{fig:relaxed}\\
C & Minor merger & Regular halo; streams or shells from an accreted low-mass companion & Fig.~\ref{fig:streams} \\
M & Major merger & Strongly perturbed halo; dust lanes; tidal tails; no massive companion & Fig.~\ref{fig:mergers} \\
I & Interacting & Perturbed halo; prominent tails due to a tidal interaction with a massive companion & Fig.~\ref{fig:tails} \\
U & Undetermined & Too close to a bright halo or Galactic cirrus to assign a type & \\ \hline
\end{tabular}
\label{tab:code-status}
\end{table}
\begin{table}
\caption{Coding of LSB structures and contaminants}
\begin{tabular}{lp{0.65\columnwidth}l}
\hline
Code & Features / contaminants & Illustration \\ \hline
+s & stream & Fig.~\ref{fig:streams} \\
+r & shells / ripples & Fig.~\ref{fig:shells} \\
+t & tail & Fig.~\ref{fig:tails} \\
+d & external star-forming disc / ring & Fig.~\ref{fig:disk} \\
+ah & asymmetric halo & \\
+ph & perturbed halo & \\
+wl & weak central dust lanes & \\
+pl & prominent dust lanes & \\ \hline
-h & galaxy embedded in the halo of a nearby star or galaxy & Fig.~\ref{fig:halos} \\
-wc & weak Galactic cirrus in the field & \\
-pc & prominent Galactic cirrus & Fig.~\ref{fig:cirrus} \\ \hline
? & presence of a given feature is uncertain & \\ \hline
\end{tabular}
\label{tab:code-feature}
\end{table}
Table~\ref{tab:class} lists the adopted classification and provides individual comments for the first galaxies in the catalog. The full table is available in the online web version.
Rough statistics on the classification -- number of galaxies and percentage -- is provided by Table~\ref{tab:stat}.
In the sample presented here, half of the ETGs show some sort of tidal perturbation whereas 35~\%\ appear to be regular, even at the depth of the survey. These fully relaxed galaxies are highlighted in green on the diagrams stellar mass vs volume density (Fig.~\ref{fig:loc}), dynamical mass vs effective radius (Fig.~\ref{fig:MRe}) and specific angular momentum vs ellipticity (Fig.~\ref{fig:rot}).
A tendency emerges from these figures: the least massive and fast rotating galaxies seem to be less perturbed than the massive, slow rotating galaxies. A detailed analysis of the trends is postponed to other papers and the completion of the survey. Despite the large number of galaxies presented here with respect to other deep imaging surveys (92), they only corresponds to 35~\% of the \AD\ volume limited sample, and the catalog has selection biases that need to be taken into account in the interpretation of the data (see Sect.~\ref{sec:sample}).
\begin{table*}
\caption{ETG classification based on the deep imaging (first entries; full table available in the online web version)}
\begin{tabular}{llp{0.7\textwidth}}
\hline
Galaxy & Class & Individual comments \\
\hline
NGC0448 & I+s & The ETG is in a tidal interaction with a disturbed companion. \\
NGC0474 & M+s+r+ph & The ETG is surrounded by multiple concentric shells and hosts several radial streams. Its outer halo reaches the disk of the unperturbed companion spiral galaxy, NGC 0470. \\
NGC0502 & M+t?+r?+ah-wc-h & The stellar halo of the ETG is asymmetric, possibly due to the presence of a diffuse tidal tail and/or a shell. \\
\hline
\end{tabular}
\label{tab:class}
\end{table*}
\begin{table}
\caption{Statistics on galaxy classification}
\begin{tabular}{lcc}
\hline
Class / features & Number of ETGs & Fraction \\ \hline
Relaxed (R) & 32 & 35~\% \\
Minor merger (C) & 15 & 16~\% \\
Major merger (M) & 11 & 12~\% \\
Interacting (I) & 20 & 22~\% \\
Undetermined (U) & 14 & 15~\% \\ \hline
with stream(s) & 26 & 28~\% \\
with tail(s) & 16 & 17~\% \\
with shell(s) & 19 & 21~\% \\
with a perturbed/asymmetric halo & 32 & 35~\% \\
with a star forming disc & 6 & 7~\% \\
\hline
\end{tabular}
\label{tab:stat}
\end{table}
\subsection{The structural parameters revisited}
\subsubsection{The contribution of the outer halo to the total luminosity}
Some of the extended structures around ETGs revealed by the MegaCam images look rather prominent on surface brightness maps (see Fig.~\ref{fig:sb}). In fact, they are of low luminosity and do not contribute much to the total mass of the galaxy.
We computed from the model galaxies (obtained with the ellipse fitting) the relative faction of g--band flux enclosed by the isophote 26~${\rm mag\,\,arcsec^{-2 }}$\ and that corresponding to the minimum isophote level below which the ellipse fitting algorithm failed (on average 27.7~${\rm mag\,\,arcsec^{-2 }}$\ ) \footnote{This is not the limiting surface surface brightness, which is deeper. Besides galaxies for which the ellipse fitting failed already below 27~${\rm mag\,\,arcsec^{-2 }}$\ were excluded from the analysis: they correspond to galaxies contaminated by the halos of bright stars which could not be properly subtracted.}.
We found that this fraction of extra g--band light revealed by our survey corresponds on average to only 5~\% of the total light of the galaxy. The maximum is 16~\%. In the r--band, the excess light, below the isophote 25.5~${\rm mag\,\,arcsec^{-2 }}$\ (to take into account the color term) is similar: 4.5~\%, with a maximum of 14~\%.
This value does not include the fine structures -- streams, tails or shells --, that are de facto excluded by the ellipse fitting algorithm. Their contribution may be determined from aperture photometry, albeit with very large error bars given their very low surface brightness.
The rather low contribution of the outer halo to the total luminosity implies that the structural parameters of the ETGs in our sample -- mass, effective radius -- initially derived from the shallow SDSS and INT images should not change dramatically when measured on the MegaCam images.
\subsubsection{Changes in effective radii from deep imaging}
We compared the effective radius directly determined from the ellipse model of the ETG, R$_{\rm e,meg}$\footnote{based on the MegaCam r--band image, along the major axis. The total luminosity was obtained within the last isophote fitted by the ellipse procedure.} with the effective radius computed from the MGE models by \cite{Scott13}, R$_{\rm e,MGE}$\footnote{tabulated in Table 1, column 10 of
\cite{Cappellari13}. Like for the MegaCam analysis, these values are derived from the observed images, without extrapolating the photometry to infinite radii. For a consistent comparison, they were not multiplied by the factor of 1.35 used to reconcile the MGE value with the original one in \cite{Cappellari11}.}. On average, the MegaCam values are just 11~\% larger. As seen in Fig.~\ref{fig:Re}, below $5 \mbox{$\times$} 10^{10}~\mbox{M$_{\odot}$}$, the old and revised measures of \mbox{R$_{\rm e}$}\ agree, with a scatter of about 0.1, i.e. within the typically 20\% measurement error of \mbox{R$_{\rm e}$} \citep{Cappellari13}.
The galaxies with unchanged measures are those we classified as ``relaxed (R)'' in our morphological analysis.
Above a transition mass of $\sim 10^{11}~\mbox{M$_{\odot}$}$, the MegaCam values of \mbox{R$_{\rm e}$}\ are systematically higher, with a maximum excess of a factor of 1.7.
For these galaxies, the larger fraction of stellar light below the SDSS surface brightness limit detected by MegaCam accounts for the increase of \mbox{R$_{\rm e}$}.
This is also related to the observation that massive galaxies are best fitted by larger \cite{Sersic68} indices \citep[e.g.][]{Caon93,Kormendy09}.
Note that similar transition masses are also observed for other ETG properties such as the internal kinematics
\citep[see Fig.~14 in][]{Cappellari13b}.
\begin{figure}
\includegraphics[angle=-90,width=\columnwidth]{nRe.pdf}
\caption{Ratio between the effective radius estimated with the MegaCam r band (with an ellipse fitting modeling of the ETG) and that determined from the SDSS/INT images \citep[with MGE modeling,][]{Cappellari13} as a function of the dynamical mass, M$_{\rm JAM}$. Relaxed galaxies showing no sign of tidal perturbations even with the deep imaging are shown in green.}
\label{fig:Re}
\end{figure}
\section{Summary and perspectives}
\label{sec:sum}
We have presented an image atlas of 92 early-type galaxies taken from the \AD\ volume limited sample. The ETGs have been observed with the MegaCam camera on the Canada-France-Hawaii-Telescope. The observing strategy and data reduction were optimized for the detection of extended low surface brightness (LSB) structures.
The resulting images have a limiting surface brightness in the g--band at least 2 mag fainter than so far available for this sample.
The catalog (available in electronic format) includes ETGs located in a low to medium density environment (thus excluding the Virgo Cluster) and spans the whole range of masses and kinematical properties probed by the \AD\ survey.
g and r band images are available for all galaxies in this catalog, while a fraction of them have also been observed in the u and i band.
The atlas consists of true colour composite images, surface brightness and colour maps, residual images resulting from the subtraction of a model of the ETG obtained with ellipse fitting. An image of the large scale environment of each galaxy covering one square degree is also shown.
These images reveal a number of LSB structures around the ETGs which were not visible on the previous generation of images, including the SDSS.
We discuss how deep imaging may change or not the way we see and classify these galaxies. In particular the presence of blue star--forming spiral structures around several objects so far considered as red and dead may appear troublesome. It reminds us about the fragility of a morphological classification based on inspection of images having different depth and of the need to add other criteria, such as the internal kinematics as proposed by the Sauron and \AD\ collaborations.\\
The previously unknown fine structures were arranged in different classes:\\
$\bullet$ prominent tidal tails, that may be gas--rich, and share the same colour as their host apart from local condensations where in-situ star-formation proceeded. They trace on going tidal interactions with massive companions or past major mergers.\\
$\bullet$ long and narrow tidal streams, sometimes wrapping around the ETG. The progenitor of their stars is likely a gas--poor low--mass disrupted companion. Its remnant is often still visible somewhere along the filament. These streams made of low metallicity material are expected to have a different colour than the host. They trace minor mergers.\\
$\bullet$ shells or ripples, that surround the ETG, often as a series of concentric circles. Such sharp-edge structures, which are best disclosed subtracting the host, were usually already visible on shallow images. However our survey revealed new ones at larger radii and associated radial linear structures. At least a fraction of them trace intermediate mass mergers.
The census of each type of fine structures was carried out by a visual inspection made by the team members. We discuss the ambiguities of such identification and present the prospects of automatic identification and classification.
We note that a rather large fraction of early-type galaxies -- about one third in our sample -- remains fully regular even at the depth of our imaging survey. Whether, given our sensitivity limit, their lack of external LSB structures and vestiges of past collisions is compatible with predictions from the cosmological simulations should be further investigated.
For most ETGs, the extra diffuse component disclosed by the survey does not increase significantly the total stellar budget. On average, isophotes fainter than 26~${\rm mag\,\,arcsec^{-2 }}$\ in the g--band contribute to about 5 percent, and a maximum of 16 percent of the total luminosity of the galaxy, not taking into account the external fine structures. An increase by on average 11\% of the effective radius of the galaxies is obtained with MegaCam with respect to the previously published values. Above a transition mass of $\sim 10^{11}~\mbox{M$_{\odot}$}$, the excess is systematic, with a maximum factor of 1.7.
The detailed statistical analysis of the results will be presented in future papers. They will address correlations between the degree of tidal perturbation with both the large scale environment and the properties of ETGs (internal kinematics and structure, gas content, etc...), which are already known thanks to the wealth of multi-wavelength data collected as part of the \AD\ project. Besides, the follow-up and on-going CFHT Large Programme MATLAS, which has exactly the same observing strategy as the one presented here, will complete the deep imaging survey. Multi-band images of hopefully all 260 \AD\ galaxies will be acquired. The galaxies located in the Virgo cluster -- about one fourth of the \AD\ sample -- were already observed as part of the Next Generation Virgo Cluster Survey (NGVS) and an image atlas similar to the one presented here for the field ETGs will be presented in the NGVS paper series.
Several teams, some including amateur astronomers, are currently carrying out similar deep, LSB optimized, imaging surveys. Ours does not necessarily stand out by the depth reached: about 28.5~${\rm mag\,\,arcsec^{-2 }}$\ in the g--band. Projects like that announced by \cite{vanDokkum14} claim significantly deeper limiting surface brightness, though as argued in this paper, determining its real value is not straightforward, as it is set by the background variations rather than by photon noise statistics. Observing with a medium size telescope, rather than with a small telescope as usually done in deep imaging experiments, allowed us to limit the integration time to less than one hour per band, instead of a full night or more, and thus to target a large number of galaxies. Besides, the great Image Quality of MegaCam and the observations from a site exceptional for its low seeing -- the Mauna Kea --, enables to address additional scientific questions, such as the detection of globular clusters and determination of their distribution around the ETGs. This is another archeological probe of the assembly of galaxies \citep{Zhu14,Brodie14}.
Nonetheless, a survey made with a complex camera like MegaCam faces a number of difficulties, some being intrinsic to any LSB study. We have presented a number of them, and on-going efforts to minimize them. In particular internal reflections within the camera imprint on the CCDs extended ghost halos which badly affect the colour of galaxies. Cumulated, they may contribute to a large fraction of the background and being the limiting factor for LSB science from the ground.
Besides the exploration of the LSB extragalactic Universe is hampered by the presence of extended emission from Galactic cirrus. At the depth of the survey, the scattered light from dust clouds becomes prominent, even at relatively large Galactic latitude. Their filamentary structure and colour are similar to that of tidal tails and streams, and thus cannot be easily subtracted. On the positive side, this scattered light offers the opportunity to study the Milky Way cold dust distribution at unprecedented high spatial resolution.
Instrumental artefacts and cirrus contamination should however not prevent the exploitation of diffuse light for galactic archeology, which has so far been underused. Even in the area of extremely large telescopes, the number of galaxies for which resolved stellar populations is accessible will remain limited. Beyond 10-20 Mpc, the LSB component of galaxies might be one of the most promising tracer of the mass assembly of galaxies, and its study provides key checks for numerical cosmological simulations.
\bibliographystyle{mn2e}
|
\section{Introduction}
\label{Sec:Introduction}
Landau-Zener (LZ) transitions occur when two energy levels cross, or more accurately experience an avoided crossing, as some external parameter is varied in time \cite{Landau,Zener,Stueckelberg,Majorana}. The system can then either stay in the same energy level that it occupied before the crossing, or it can undergo a transition to the other level. Such a universal phenomenon is ubiquitous and has applications in various areas of quantum physics. Among the new areas where the physics of LZ transitions can play an important role are adiabatic quantum computation (AQC) \cite{Farhi} and amplitude spectroscopy in nanoscale circuits \cite{Berns,Shevchenko,Johansson} and in nitrogen-vacancy centers in diamond \cite{Huang}. It could also play a role in the inter-molecular energy transfer in biological light-harvesting systems. In AQC, the parameters of a physical system (which can be called a quantum computer or annealer) are varied slowly such that the system transforms from an easy-to-prepare ground state into a ground state that contains the answer to a physical problem (or even a computational problem of non-physical nature). In biological light-harvesting systems, energy transfer between different parts of a molecule could be governed by molecular changes that act as driving fields for electronic motion.
The LZ problem in a closed system was solved soon after it was formulated over eighty years ago \cite{Landau,Zener,Stueckelberg,Majorana}. Physical systems, however, invariably interact with a surrounding environment. There have been numerous studies on the LZ problem in the presence of an environment \cite{Kayanuma,Gefen,Ao,Shimshony,Nishino,Pokrovsky2003,Sarandy,Wubs,AshhabAQC,Lacour,Pokrovsky2007,Amin,Nalbach,Dodin,Xu,Haikka}, and some methods have produced accurate results in their regimes of validity. However, there is no method that is valid and computationally efficient for all parameter regimes. In particular, the different methods typically have underlying assumptions justifying the validity of their mathematical formulation based on physical arguments. For example, one could make the assumption of a very short correlation or memory time in the environment's degrees of freedom and use a Markovian approach. This approach would, however, break down when the environment's correlation time is not short compared to the LZ timescale, a situation that could occur when dealing with low-frequency noise.
Here we take a different approach. We numerically solve a rather simple physical problem that involves a single two-level system (to which we also refer as the qubit) coupled to a single harmonic oscillator. We can therefore be confident that our numerical calculations provide an accurate description of the problem as formulated. After obtaining the numerical results for the relatively simple problem, we comment on the physical significance of these results and how they could apply for a system where the single harmonic oscillator is replaced by an environment with a large number of degrees of freedom.
The remainder of this paper is organized as follows: In Sec.~\ref{Sec:Hamiltonian} we describe the basic setup and introduce the corresponding Hamiltonian. In Sec.~\ref{Sec:NumericalCalculations} we describe our numerical calculations. In Sec.~\ref{Sec:Results} we present the results of these calculations and discuss the interpretation of the results. Section \ref{Sec:Conclusion} contains some concluding remarks.
\section{Model system and Hamiltonian}
\label{Sec:Hamiltonian}
We consider the basic LZ problem where the system of interest possesses only two quantum states. As such, it can be described using the Pauli matrices $\hat{\sigma}_{\alpha}$ with $\alpha=x,y$ or $z$. We use the basis states defined by the relations $\hat{\sigma}_{z}\ket{\uparrow}=\ket{\uparrow}$ and $\hat{\sigma}_{z}\ket{\downarrow}=-\ket{\downarrow}$.
In an isolated system, the LZ Hamiltonian is given by
\begin{equation}
H = -\frac{vt}{2} \hat{\sigma}_{z}-\frac{\Delta}{2} \hat{\sigma}_{x},
\label{Eq:LZHamiltonianClosedSystem}
\end{equation}
where the time variable $t$ goes from $-\infty$ to $+\infty$, $v$ is the sweep rate and $\Delta$ is the minimum energy gap between the ground and excited states, which occurs at $t=0$. At large negative times the ground and excited states asymptotically coincide with the states $\ket{\downarrow}$ and $\ket{\uparrow}$, respectively. The roles of these states are reversed at large positive times. At $t=0$, the instantaneous ground and excited states are equal superpositions of the states $\ket{\uparrow}$ and $\ket{\downarrow}$. The LZ formula, which for example gives the probability for a system prepared in its ground state at $t\rightarrow -\infty$ to end up in the excited state at $t\rightarrow \infty$, is given by $P_{\rm LZ} = \exp\{-\pi\Delta^2/(2v)\}$. In particular, for a slow sweep (i.e.~$v/\Delta^2\ll 1$), $P_{\rm LZ}\rightarrow 0$ and a system that is initially prepared in the ground state has a high probability of remaining in the ground state.
The LZ problem can be generalized in order to take into account the effects of an uncontrolled external environment. Early studies on this problem used somewhat ad hoc quantum master equations in order to incorporate dissipative processes in the dynamics \cite{Kayanuma}. Subsequent studies generally started with a specific model of the environment and derived approximate equations of motion for the system under certain approximations (see e.g.~Refs.~\cite{Ao,Wubs,Nalbach,Dodin}). Because of computational convenience and physical relevance, the environment is commonly modeled as a large set of harmonic oscillators, even if the microscopic details of the environment are not known. This approach has been applied successfully to the study of the LZ problem in a number of regimes. It is not possible, however, to obtain analytic results for this problem, and approximations that are valid for specific regimes are commonly made in order to numerically calculate the effect of the large number of harmonic oscillators on the LZ probability. The strong-coupling and low-temperature regimes are particularly challenging for these methods.
Here we take a different approach to studying the effects of the environment on the LZ problem. We consider an environment composed of a single harmonic oscillator. Clearly this simple model will not be able to capture all the effects that occur in a complex environment. However, the simplicity of the model allows us to have confidence in the results of standard numerical simulations. Rather than having to make assumptions concerning the behaviour of the system at the beginning of the calculation, the difficult task is then shifted to the step of interpreting the numerical results and identifying in these results patterns and tendencies that one can expect to apply for a large environment. It is also worth mentioning here that there can be cases where the largest environmental effects are caused by a single mode in the environment, in which case the results of this simple model become particularly relevant. Another advantage of treating such a minimal model is the fact that it allows us to discuss physical processes rather clearly.
We would like to note here that a related system, namely an LZ problem of a qubit coupled to a harmonic oscillator and an environment, was recently considered in Ref.~\cite{Zueco}. In that work, however, there is no minimum-gap term in the qubit's Hamiltonian, and the avoided crossings arise as a result of the coupling between the qubit and the oscillator, rendering the system qualitatively different from the one that we consider in this paper.
\begin{figure}[h]
\includegraphics[width=8.0cm]{LandauZenerWithOscillatorFigEnergyLevelDiagram.eps}
\caption{(color online) Energy level diagram of a coupled qubit-oscillator system with the qubit bias conditions varied according to the LZ protocol.}
\label{Fig:EnergyLevelDiagram}
\end{figure}
The Hamiltonian of the LZ problem with a single-mode environment and linear coupling is given by:
\begin{equation}
H = -\frac{vt}{2} \hat{\sigma}_{z} - \frac{\Delta}{2} \hat{\sigma}_{x} + \hbar \omega \hat{a}^{\dagger} \hat{a} + g \sigma_z \otimes \left( \hat{a} + \hat{a}^{\dagger} \right),
\label{Eq:LZHamiltonianSingleEnvironmentMode}
\end{equation}
where $\omega$ is the characteristic frequency of the harmonic oscillator, $\hat{a}$ and $\hat{a}^{\dagger}$ are, respectively, the oscillator's annihilation and creation operators, and $g$ is the qubit-oscillator coupling strength. The energy level diagram of this problem is illustrated in Fig.~\ref{Fig:EnergyLevelDiagram}.
We are interested in particular in the case of slow, nearly adiabatic passage. This case corresponds to the desired condition for obtaining a high transfer probability in adiabatic passage protocols; it is also the relevant regime for maximizing the success probability in an adiabatic quantum computation.
\section{Numerical calculations}
\label{Sec:NumericalCalculations}
We numerically solve the time-dependent Schr\"odinger (or Liouville-von Neumann) equation using the Hamiltonian given in Eq.~(\ref{Eq:LZHamiltonianSingleEnvironmentMode}). In these calculations we set the sweep rate $v$ to the value that gives $P_{\rm LZ}=0.1$ (i.e., starting from the ground state, the two-level system ends up in the ground state with 90\% probability). In other words, we choose a sweep rate that is close to the adiabatic limit in the absence of the coupling to the oscillator. We take three different values for the oscillator frequency: $\omega/\Delta=0.2$ (low-frequency oscillator), 1 (intermediate regime), and 5 (high-frequency oscillator). We vary the coupling strength from $g/\Delta=0$ to $g/\Delta=2$, and we vary the temperature $T$ from $k_BT/\Delta=0$ to $k_BT/\Delta=5$, where $k_B$ is the Boltzmann constant.
In order to incorporate the finite temperature into the calculation, the simulations are started in thermal equilibrium at a large negative value for the time variable. In this limit, the qubit and resonator are effectively decoupled from each other, except for simple mean-field shifts that they induce on each other. Furthermore, the qubit's energy splitting is very large in the limit $t\rightarrow -\infty$. As a result, the qubit starts initially in its ground state $\ket{\downarrow}$. The harmonic oscillator starts in a mixed thermal state according to the Boltzmann probability distribution with an average number of quanta $k_BT/(\hbar\omega)$ for high temperatures (Note that the Boltzmann probability distribution extends up to several times this value). This estimate provides a minimum number of basis states that need to be included in the simulations, and it also sets a limit to the highest temperatures that can be reached in simulations with a given size of the Hilbert space. In particular for the lowest oscillator frequency and highest temperature that we consider, we use a Hilbert space with 1000 basis states. Note that the initial state of the oscillator is the only part of the calculation where the finite temperature of the environment enters the calculation.
After setting the initial state according to the Boltzmann distribution, we evolve the density matrix of the combined system in time according to the Schr\"odinger equation. Note that this evolution is unitary, which is the reason why we can say that, in contrast to most other methods, we do not make any approximations or assumptions concerning the internal dynamics of the environment. The evolution is stopped at a sufficiently large and positive value of the time variable, such that further evolution would not have any noticeable effect on the occupation probabilities of the different states. At this final time, we examine the occupation probabilities of the different quantum states, from which we can easily calculate the probability that the qubit remains in its ground state.
\begin{figure}[h]
\includegraphics[width=7.5cm]{LandauZenerWithOscillatorFigColorPlot02.eps}
\includegraphics[width=7.0cm]{LandauZenerWithOscillatorFigVaryTemperature02.eps}
\includegraphics[width=7.0cm]{LandauZenerWithOscillatorFigVaryCouplingStrength02.eps}
\caption{(color online) Top: Qubit's final excited-state probability $P$ as a function of temperature $k_BT$ and coupling strength $g$, both measured relative to the qubit's minimum gap $\Delta$. Middle: $P$ as a function of $k_BT/\Delta$ for four different values of $g/\Delta$: 0.1 (red solid line), 0.3 (green dashed line), 1 (blue dotted line) and 2 (magenta dash-dotted line). Bottom: $P$ as a function of $g/\Delta$ for three different values of $k_BT/\Delta$: 1 (red solid line), 3 (green dashed line), and 5 (blue dotted line). In all the panels, the harmonic oscillator frequency is $\hbar\omega/\Delta=0.2$. The sweep rate is chosen such that $P_{\rm LZ}=0.1$, and this value is the baseline for all of the results plotted in this figure.}
\label{Fig:ExcitationProbability02}
\end{figure}
\begin{figure}[h]
\includegraphics[width=7.5cm]{LandauZenerWithOscillatorFigColorPlot10.eps}
\includegraphics[width=7.0cm]{LandauZenerWithOscillatorFigVaryTemperature10.eps}
\includegraphics[width=7.0cm]{LandauZenerWithOscillatorFigVaryCouplingStrength10.eps}
\caption{(color online) Same as in Fig.~\ref{Fig:ExcitationProbability02} but for $\hbar\omega/\Delta=1$.}
\label{Fig:ExcitationProbability10}
\end{figure}
\begin{figure}[h]
\includegraphics[width=7.5cm]{LandauZenerWithOscillatorFigColorPlot50.eps}
\includegraphics[width=7.0cm]{LandauZenerWithOscillatorFigVaryTemperature50.eps}
\includegraphics[width=7.0cm]{LandauZenerWithOscillatorFigVaryCouplingStrength50.eps}
\caption{(color online) Same as in Fig.~\ref{Fig:ExcitationProbability02} but for $\hbar\omega/\Delta=5$.}
\label{Fig:ExcitationProbability50}
\end{figure}
\section{Results}
\label{Sec:Results}
The probability for the qubit to end up in the excited state at the final time as a function of temperature and coupling strength is plotted in Figs.~\ref{Fig:ExcitationProbability02}-\ref{Fig:ExcitationProbability50}. As expected from known results \cite{Wubs}, the final excited-state occupation probability $P$ remains equal to 0.1 whenever the temperature or the coupling strength is equal to zero. Otherwise, the coupling to the oscillator causes this probability to increase. A common, and somewhat surprising, trend for all values of $\hbar\omega/\Delta$ is the non-monotonic dependence on the coupling strength $g$. As the coupling strength is increased from zero to finite but small values, $P$ increases. But when the coupling strength is increased further, $P$ starts decreasing. Based on the results that are plotted in Figs.~\ref{Fig:ExcitationProbability02}-\ref{Fig:ExcitationProbability50}, one can expect that in the limit of large $g/\Delta$ (and assuming not-very-large values of $k_BT/\Delta$) the excited-state occupation probability will go back to its value in the uncoupled case, i.e.~$P=0.1$. This phenomenon is probably a manifestation of the superradiance-like behaviour in a strongly coupled qubit-oscillator system \cite{AshhabSuperradiance}. In the superradiant regime (i.e.~the strong-coupling regime), the ground state is highly entangled exactly at the symmetry point (which corresponds to the bias conditions at $t=0$ in the LZ problem), but even small deviations from the symmetry point can lead to an effective decoupling between the qubit and resonator with the exception of some state-dependent mean-field shifts. Indeed the maximum values of $P$ reached in Figs.~\ref{Fig:ExcitationProbability10} and \ref{Fig:ExcitationProbability50} occur at coupling strength values that are comparable to the expression for the uncorrelated-to-correlated crossover value, namely $g\sim\hbar\omega$ (and we have verified that the near-linear increase in peak location as a function of oscillator frequency continues up to $\hbar\omega/\Delta=20$). This relation does not apply in the case $\hbar\omega/\Delta=0.2$, shown in Fig.~\ref{Fig:ExcitationProbability02}. In this case, the peak occurs when the coupling strength $g$ is comparable to the minimum gap $\Delta$. It is in fact quite surprising that the excitation peak in the case $\hbar\omega/\Delta=0.2$ occurs at a higher coupling strength than that obtained in the case $\hbar\omega/\Delta=1$. In order to investigate this point further, we tried values close to $\hbar\omega/\Delta=1$ and found that this value gives a minimum in the peak location (i.e.~the peak in $P$ when plotted as a function of $g/\Delta$).
Another feature worth noting is the temperature dependence of $P$ close to zero temperature. As can be seen clearly in Figs.~\ref{Fig:ExcitationProbability10} and \ref{Fig:ExcitationProbability50}, the initial increase in $P$ with temperature is very slow, indicating that it probably follows an exponential function that corresponds to the probability of populating the excited states in the harmonic oscillator (and the same dependence is probably present but difficult to see because of the scale of the $x$ axis in Fig.~\ref{Fig:ExcitationProbability02}). After this initial slow rise, and in particular when $k_BT \gtrsim \hbar\omega$, we see a steady rise that in the case of Fig.~\ref{Fig:ExcitationProbability02} can be approximated as a linear increase in $P$ with increasing $T$. Importantly, the slope of this increase can be quite large for intermediate $g$ values. From the results shown in Figs.~\ref{Fig:ExcitationProbability02}-\ref{Fig:ExcitationProbability50}, we find that the maximum slope $[dP/d(k_BT/\Delta)]_{\rm max}=0.18\times(\hbar\omega/\Delta)^{-0.57}$, and results for other parameter values extending up to $\hbar\omega/\Delta=20$ follow this dependence. The implication of this result can be seen clearly in the middle panel of Fig.~\ref{Fig:ExcitationProbability02}: even when the temperature is substantially smaller than the qubit's minimum gap $\Delta$, the initial excitation of the low-frequency oscillator (stemming from the finite temperature) can cause a large increase in the qubit's final excited-state probability. This result is in contrast with the exact result of Ref.~\cite{Wubs} stating that at zero temperature the qubit's final excited-state probability is given by $P_{\rm LZ}$ regardless of the value of $g$. The typical temperature scale at which deviations from the LZ formula occur can therefore be much lower than $\Delta/k_B$. This result is relevant for adiabatic quantum computing, because it contradicts the expectation that having a minimum gap that is large compared to the temperature might provide automatic protection for the ground state population against thermal excitation. Another point worth noting here is that when $\hbar\omega<\Delta$, there is no point in time where the qubit and oscillator are resonant with each other, yet the initial thermal excitation of the oscillator can result in exciting the qubit at the final time. The excitations in the oscillator are in some sense up-converted into excitations in the qubit as a result of the sweep through the avoided crossing.
\begin{figure}[h]
\includegraphics[width=7.0cm]{LandauZenerWithOscillatorFigVaryInitialState02.eps}
\caption{(color online) The final excited state probability $P$ as a function of the number of excitation quanta $n$ present in the initial state of the oscillator. Here we take $\hbar\omega/\Delta=0.2$. The different lines correspond to different values of the coupling strength: $g/\Delta=0.1$ (red solid line), 0.5 (green dashed line), 1 (blue dotted line) and 2 (magenta dash-dotted line).}
\label{Fig:ExcitationProbabilityAsFunctionOfInitialOscillatorExcitationNumber}
\end{figure}
We can also see in Fig.~\ref{Fig:ExcitationProbability02} that for $g/\Delta\gtrsim 1$ the temperature dependence is non-monotonic. In particular, for low temperatures we obtain the intuitively expected increase in excitation probability with increasing temperature, but this trend reverses for higher temperatures. In order to investigate this feature further, we calculate the qubit's final excited-state probability as a function of the number $n$ of excitation quanta present in the initial state of the oscillator (Note that this calculation differs from the ones described above in that here we do not use the Boltzmann distribution for the oscillator's initial state). The results are plotted in Fig.~\ref{Fig:ExcitationProbabilityAsFunctionOfInitialOscillatorExcitationNumber}. These results explain the non-monotonic dependence on temperature. For intermediate values of $g/\Delta$ (e.g.~for $g/\Delta=1$), there is a peak at a small but finite excitation number followed by a steady decrease. As the temperature is increased from zero, the qubit's final excited-state probability samples the probabilities for increasingly high excitation numbers, and a peak at intermediate values of temperature is obtained. Note that for large excitation numbers, the increase in $P$ as a function of $n$ resumes, and this increase will also be reflected in the temperature dependence.
We note in this context that recent theoretical studies \cite{Nalbach,Dodin} have reported non-monotonic dependence of the excitation probability as a function of the sweep rate $v$. However, that dependence was generally oscillatory, and we suspect that it has a different origin from the behaviour obtained in the present study. We expect that similar oscillatory behaviour would be obtained if we varied $v$ in our calculations. As mentioned in Sec.~\ref{Sec:Hamiltonian}, however, here we are mainly interested in the almost-adiabatic regime, and we have therefore not analyzed the $v$ dependence in our calculations.
\begin{figure}[h]
\includegraphics[width=7.5cm]{LandauZenerWithOscillatorFigColorPlotIncoherent02.eps}
\includegraphics[width=7.5cm]{LandauZenerWithOscillatorFigColorPlotIncoherent10.eps}
\includegraphics[width=7.5cm]{LandauZenerWithOscillatorFigColorPlotIncoherent50.eps}
\caption{(color online) Qubit's final excited state probability $P$ obtained from the semiclassical calculation as a function of temperature $k_BT$ and coupling strength $g$, both measured relative to the minimum qubit gap $\Delta$. The different panels correspond to different values of the harmonic oscillator frequency: $\hbar\omega/\Delta=0.2$ (top), 1 (middle) and 5 (bottom).}
\label{Fig:ExcitationProbabilityFromIncoherentCalculation}
\end{figure}
In addition to solving the Schr\"odinger equation, we have performed semiclassical calculations where we assume that there is no quantum coherence between the different LZ processes. (Note here that when we replace the isolated qubit with the coupled qubit-oscillator system the single avoided crossing is replaced by a complex network of avoided crossings.) Under this approximation, we only need to calculate the occupation probabilities of the different states, and these probabilities change (according to the LZ formula) only at the points of avoided crossing. This approach greatly simplifies the numerical calculations because the locations and gaps for the different avoided crossings can be determined easily (see e.g.~Fig.~\ref{Fig:EnergyLevelDiagram}). The results are shown in Fig.~\ref{Fig:ExcitationProbabilityFromIncoherentCalculation}. The results of this calculation agree generally well with those obtained by solving the Schr\"odinger equation when $\hbar\omega/\Delta=1$. For $\hbar\omega/\Delta =5$, the semiclassical calculation consistently underestimates the excited-state probability, but the overall dependence on temperature and coupling strength is remarkably similar to that shown in Fig.~\ref{Fig:ExcitationProbability50}. We should note that higher values of $\hbar\omega$ (not shown) exhibit more pronounced deviations, with side peaks appearing in the dependence of $P$ on $g/\Delta$. The most striking deviation from the results of the fully quantum calculation is seen in the case $\hbar\omega/\Delta=0.2$ (i.e.~the case of a low-frequency oscillator). In the semiclassical calculation, there is a rather high peak at a small value of the coupling strength (and sufficiently high temperatures), and the excited-state probability starts decreasing when the coupling strength $g$ becomes larger than $\hbar\omega$. In the fully quantum calculation, however, the peak is located at a much higher value, somewhere between 0.5 and 1 depending on the temperature.
The fact that the semiclassical calculation generally gives results different from those given by the fully quantum calculation is an indication that quantum coherence and interference between multiple LZ processes play a role in determining the final occupation probabilities. In this context we note that the avoided crossings occur at instances separated by time intervals $\tau_{\rm separation}=\hbar\omega/v$ (with an infinite number of avoided crossings occurring simultaneously at each one of these instances), and the time duration over which an LZ mixing process occurs (in the almost-adiabatic regime) is given by $\tau_{\rm LZ}\sim\Delta/v$ \cite{Shevchenko,MinimumGapFootnote}. The ratio between these two time scales is then given by $\tau_{\rm separation}/\tau_{\rm LZ}\sim\hbar\omega/\Delta$. In other words, when $\hbar\omega/\Delta$ is small the different LZ processes will overlap in time, and it is not too surprising that the semiclassical calculation gives incorrect predictions in this case. It is somewhat surprising, however, that when $\hbar\omega/\Delta=1$ the two calculations agree quite well and then in the regime $\hbar\omega/\Delta>1$ the effect of quantum interference between the LZ processes can again be seen in the final occupation probabilities.
We now take another look at our results presented above from the point of view of how they might apply in the case of a large environment containing a large number of degrees of freedom with no single dominant environmental mode. Note here that the coupling between the qubit and the environment can in principle be strong, even if the coupling to each individual mode in the environment is weak. We first consider our results in the regime of strong qubit-environment coupling. We have found that strong coupling to a single mode results in a reduced effect of that mode on the final occupation probabilities. It is unlikely that this result will apply to the case where the qubit is coupled strongly to an uncontrolled environment containing a large number of independent modes with the coupling to each individual mode in the environment being weak. The weakening of the environmental effects with increased coupling strength in the case of a single mode is most likely related to the energy level structure and the possible paths that the system can follow while it traverses the network of avoided crossings. The energy level structure and the possible paths are vastly different when the strong coupling to the environment is caused by the large number of modes in the environment. It would be more plausible that in this case one can make statements concerning large environments using the following approach: focus on the small $g/\Delta$ region of the results discussed above, take the contributions of the individual environment modes and add up these small contributions. In this case an increase in coupling strength would result in an increase in the excited state probability, as would be intuitively expected. We therefore expect that the result of non-monotonic behaviour with increasing coupling strength should be thought of as a result pertaining to the case with a single dominant mode in the environment. Another area where we can try to extract from our results statements concerning a large environment occurs in the regime of low temperatures, which can be particularly relevant in the context of AQC. As a side note, we mention here that one of the central questions in the field of AQC is the scaling of the minimum gap with system size. It is known that the minimum gap decreases with increasing system size, and there are ongoing studies on the exact scaling law. This minimum-gap scaling is typically discussed in relation to the time needed to ensure adiabatic evolution of the quantum annealer, and the minimum running time is calculated based on the well-known LZ formula given in Sec.~\ref{Sec:Hamiltonian}. An independent question is the resistance of the AQC success probability to environmental noise. The facts that at finite temperatures the excitation probability increases above the base value $P_{\rm LZ}$ and that the excitation probability can be substantially larger than $P_{\rm LZ}$ even at temperatures much lower than the minimum gap mean that the coupling to the low-frequency modes in the environment needs to be considered with extra care in the low-temperature regime. In a previous work \cite{AshhabAQC}, we discussed the scaling of the noise amplitude with system size, with the main message being that the noise amplitude increases with increasing system size. The present work complements our earlier work in that it provides a quantitative analysis of the effect of the environment on a system driven using an adiabatic passage protocol, as is the case in AQC.
\section{Conclusion}
\label{Sec:Conclusion}
We have investigated the problem of a two-level system undergoing an LZ passage through an avoided crossing while it interacts with a finite-temperature harmonic oscillator. We have found a number of counter-intuitive results, including non-monotonic dependence of the final-time excitation probability as a function of temperature or qubit-oscillator coupling strength. We have provided physical explanations for these phenomena. The physical mechanisms at play include modifications to the avoided crossing structure related to the formation of highly correlated energy eigenstates as well as quantum coherence between multiple LZ processes.
Our original motivation for analyzing a system with a single qubit and a single additional degree of freedom was to use the obtained results in order to make statements relevant for a large environment, and we have indeed attempted to make such an extrapolation of results. We emphasize, however, that our results are of interest even in relation to the single-oscillator case, both because they pertain to a model system that allows a clear discussion of the physical mechanisms involved and because certain systems in nature are accurately described by the model of a single qubit coupled to a single oscillator. In other words, in addition to the general principles that we have deduced concerning general environments, our results can have direct applicability to qubit-oscillator systems such as cavity electrodynamics or some molecular systems.
|
\section{Introduction and Summary}\label{introsec}
\subsection{Motivation and antecedents}\label{motivationsubsec}
One of the useful ways to inquire into the behavior of a field theory is to couple an external probe to it and deduce the response of the fields. The vevs of line operators associated with such probes serve as diagnostics of the phase of the field theory. Furthermore, even for a simple theory such as Maxwell's,
once one considers probes with generic motion, one encounters the fascinating phenomenon of radiation, with far-reaching implications both on a theoretical and a practical level. For generic interacting quantum field theories, the analytic treatment of detailed properties of radiation, like its frequency and angular distribution, its broadening, or the damping it induces on the probe, is mostly based on perturbation theory and therefore limited to the realm of weak coupling. In recent years, it has been appreciated that for quantum field theories with additional symmetries ({\it e.g.}, conformal invariance and/or supersymmetry) and for suitably chosen probes, there are a number of techniques that allow us to address these questions for strongly-coupled field theories. In very fine-tuned examples, and for very specific questions, we can even obtain exact answers, something quite unusual for a field theory in more than two dimensions.
Depending on the questions we want to address, we sometimes need to be able to handle probes with arbitrary timelike trajectories, while for selected questions having control over very specific trajectories is enough. As a first illustration of this point, let us start by recalling the definition of one of the most interesting quantities associated to a heavy probe, the cusp anomalous dimension. Consider a probe at rest that receives a sudden kick and afterwards moves with constant speed; its worldline presents a cusp at the event of the kick. The evaluation of the corresponding Wilson line will feature a logarithmic divergence \cite{polyakov}
\begin{equation}
\vev{W}\sim e^{-\Gamma_{\mbox{\tiny cusp}}(\varphi)\log\frac{L}{\epsilon}}~,
\end {equation}
with $\varphi$ the boost parameter, and $L,\epsilon$ IR and UV cutoffs, respectively. The function $\Gamma_{\mbox{\scriptsize cusp}}(\varphi)$ is the cusp anomalous dimension. This logarithmic divergence will be on top of other divergences that the vev of a similar Wilson line with a smooth contour might have. While the determination of this function for generic boosts is an interesting problem, its evaluation at either very large or very small boosts is already quite rewarding physically. Let us focus in particular in the small boost limit, by performing a Taylor expansion of the cusp anomalous dimension in terms of the boost parameter,
\begin{equation}
\Gamma{\mbox{\scriptsize cusp}}(\varphi)=B(\lambda,N) \varphi^2+{\cal O}(\varphi^4)~.
\end{equation}
The coefficient of $\varphi^2$ was called the Bremsstrahlung function in \cite{correa}, and as indicated, it depends on the rank $N$ of the gauge group and the coupling $\lambda=g_{YM}^2 N$. At this point, we can start to illustrate the power of focusing on theories with additional symmetries. It was argued in \cite{correa} that for probes coupled to arbitrary four-dimensional conformal field theories (CFTs), a number of physically interesting quantities are completely determined up to a coefficient, and for all these quantities, the undetermined coefficients are the same and equal to the Bremsstrahlung function. In particular, the energy loss of a probe at small velocities is
\begin{equation}
E=2\pi B\; \int dt (\dot v)^2~.
\end{equation}
This identification is valid for any line operator and any conformal field theory.
A second example of a very interesting specific trajectory is that of a probe with constant proper acceleration. Its worldline is a branch of a hyperbola, which translates into a circle in Euclidean signature. If the probe is coupled to a conformal field theory, from the vev of the circular Wilson loop and the one-point function of the stress-energy tensor in the presence of such Wilson line, one can obtain an alternative derivation of the Bremsstrahlung function \cite{fgl,lewkomalda}, the momentum difussion coefficient of the accelerated probe \cite{fgt}, or the change in entanglement entropy of a spherical region when adding a static probe \cite{lewkomalda}. In the case of a 1/2 BPS probe of ${\cal N}=4$ $SU(N)$ super-Yang-Mills (SYM), it was argued in \cite{correa} that the Bremsstrahlung function is very simply related to the vev of a circular Wilson loop,
\begin{equation}
B=\frac{1}{2\pi^2}\lambda \partial_\lambda \hbox{log }\vev{W_{\circledcirc}}~.
\label{bfromw}
\end{equation}
Relation (\ref{bfromw}) allows us to derive the exact Bremsstrahlung function from the exact vev of a circular Wilson loop \cite{esz}, which can be evaluated by means of a matrix model computation \cite{dg} that ultimately finds its justification through localization arguments \cite{pestun}. One then finds for a fundamental 1/2 BPS probe coupled to $U(N)$ \cite{correa}
\begin{equation}
B=\frac{\lambda}{16\pi^2}\frac{L_{N-1}^2\left(-\frac{\lambda}{4N}\right)
+L_{N-2}^2\left(-\frac{\lambda}{4N}\right)}{L_{N-1}^1\left(-\frac{\lambda}{4N}\right)}~,
\label{exactb}
\end{equation}
where the $L_n^\alpha$ are generalized Laguerre polynomials. This expression was also deduced by a different chain of reasoning in \cite{fgl, lewkomalda}.
While the response of the theory to probes following these two particular trajectories is quite interesting, it is certainly far from covering all the rich physics of accelerating probes, and for other questions we need to be able to handle probes with arbitrary trajectories. In what follows we will concentrate on the study of (locally) 1/2-BPS probes of the vacuum state of ${\cal N}=4$ $SU(N)$ SYM by means of the AdS/CFT correspondence, so we find it appropriate to start by reviewing some known results.
Consider first a probe in the fundamental representation of the $SU(N)$ gauge group, i.e., a quark. In the context of the AdS/CFT correspondence \cite{malda,gkpw}, its dual is a fundamental string embedded in AdS$_5\times {\mathbf{S}}^5$ and governed by the Nambu-Goto action. Mikhailov \cite{mikhailov} considered such a probe following an arbitrary timelike trajectory, and found the corresponding string embedding in the supergravity regime ($N\gg 1$, $\lambda\gg 1$). The energy of the string at a given time was shown to contain both the expected instantaneous energy of the quark \cite{dragtemp}, $E_{\mbox{\scriptsize q}}=\gamma m$, and the energy that the quark has radiated over all of its previous history \cite{mikhailov},
\begin{equation}\label{radenergyfund}
E_{\mbox{\scriptsize rad}}= \frac{\sqrt{\lambda}}{{2\pi}}\int dt_r \, a^\mu a_\mu
=\frac{\sqrt{\lambda}}{{2\pi}}\int dt_r \, \gamma^6\left(\vec a^2-|\vec v\times \vec a|^2\right)~.
\end{equation}
Notice this says that, remarkably, the total radiated power by the fundamental probe in this strongly-coupled non-Abelian theory is given (up to the coefficient) by the familiar Li\'enard formula from classical electrodynamics. The coefficient of (\ref{radenergyfund}) informs us that the Bremsstrahlung function for a 1/2-BPS particle in the fundamental representation is given in the supergravity limit by
\begin{equation}
B=\frac{\sqrt{\lambda}}{4\pi^2}~.
\end{equation}
matching the result first obtained in \cite{Kruczenski:2002fb}. In hindsight, it could have also been deduced by applying (\ref{bfromw}) to the vev of the corresponding circular Wilson loop \cite{bcfm,dgo}
\begin{equation}
\vev{W_{\circledcirc}}=e^{\sqrt{\lambda}}~.
\end{equation}
Alternatively, the same Bremsstrahlung function can be deduced from the computation of the one-point function of the Lagrangian density \cite{dkk,cg,trfsq,fgl} or energy density \cite{liusynchrotron,iancuradiation,lewkomalda,tmunu} in the presence of a quark, or the momentum diffusion coefficient of the same probe following a trajectory with constant proper acceleration \cite{xiao,brownian}.
When we move on to probes in higher-rank representations of the gauge group, the totally antisymmetric representations turn out to be the easiest generalization. In the supergravity regime, the dual of an antisymmetric $k$-quark is given \cite{yamaguchi,gp} by a D5-brane with $k$ units of worldvolume electric flux (and consequently, string charge). The brane wraps an ${\mathbf{S}}^4\subset{\mathbf{S}}^5$ at polar angle $\vartheta_k$ such that \cite{baryon,pr}
\begin{equation}
\sin \vartheta_k \; \cos \vartheta_k-\vartheta_k= \frac{\pi k}{N}~.
\end{equation}
Combining a very general result due to Hartnoll \cite{hartnoll} with the string solutions of Mikhailov, it is immediate to obtain the D5-brane solution dual to an antisymmetric probe following an arbitrary trajectory. This allows for the determination of the energy loss \cite{fg},
\begin{equation}
E_{\mbox{\scriptsize rad}}^{{\cal A}_k}=\frac{N\sqrt{\lambda}}{{3\pi^2}}\sin^3 \vartheta_k \; \int dt_r \,\gamma^6 \left(\vec a^2-|\vec v\times \vec a|^2\right)~.
\label{radenergyant}
\end{equation}
We observe that again the radiated power is given by a Li\'enard-type formula. The corresponding Bremsstrahlung function in the supergravity regime can be read off from (\ref{radenergyant}) to be
\begin{equation}
B_{{\cal A}_k}=\frac{N\sqrt{\lambda}}{{6\pi^3}}\sin^3 \vartheta_k~.
\end{equation}
Similarly to what happens in the fundamental representation, the exact vevs of the circular Wilson loops for probes in the antisymmetric representations have been computed exactly \cite{Fiol:2013hna}, and from them one can deduce the corresponding Bremsstrahlung function using (\ref{bfromw}).
Let us now turn to probes in the totally symmetric representation of ${\cal N}=4$ $SU(N)$ SYM, which will be the main focus of the present work.\footnote{Since the configurations we study lie at a fixed position on the ${\mathbf{S}}^5$, our results make no use of the internal dimensions and should be relevant for other four-dimensional CFTs, whose gravity dual would involve AdS$_5$ times a different compact manifold.} In the AdS/CFT correspondence, a symmetric $k$-quark is dual to a D3-brane with $k$ units of electric flux, which is embedded fully within AdS$_5$ and reaches the boundary of this spacetime at the worldline of the probe \cite{gp,gp2}. Contrary to what happens for probes in the fundamental and antisymmetric representations, up until now the D3-brane embedding dual to an arbitrary timelike worldline was not available. This limitation was bypassed in \cite{fg} by considering a D3-brane dual to a probe with constant proper acceleration, essentially obtained in \cite{df}. The computation of the energy loss for that particular trajectory yielded the Bremsstrahlung function
\begin{equation}
B_{{\cal S}_k}=\frac{k\sqrt{\lambda}}{4\pi^2}\sqrt{1+\frac{k^2\lambda}{16\pi^2N^2}}~.
\label{bksym}
\end{equation}
Again, since the Bremsstrahlung function for this 1/2-BPS probe appears in various physical quantities, it is also possible to determine it by considering the momentum difussion coefficient \cite{fgt} of the probe following accelerated motion, or the one-point function of the Lagrangian density in the presence of a static probe \cite{fgl}, which in turn is related by supersymmetry to the one-point function of the $\Delta=2$ chiral primary operator computed in \cite{Giombi:2006de}. The exact result for the circular Wilson loop in the totally symmetric representation was determined recently in \cite{Fiol:2013hna}, but it is given in a form that makes it very hard to carry out a systematic large $N$ expansion. The first subleading correction in $1/N$ was worked out and successfully matched to the D3-brane description very recently in \cite{leo}.
As we have stressed, there are physically interesting questions where one needs to have a handle on probes following arbitrary timelike trajectories. In the case at hand this means finding a D3-brane that reaches the boundary of AdS$_5$ at such trajectories. Besides the intrinsic interest of studying probes in higher-rank representations, there is an additional motivation that is specific to the totally symmetric case. Upon setting $k=1$ in the supergravity result (\ref{bksym}) for the $k$-symmetric Bremsstrahlung function, one obtains the result in the fundamental representation, corrected by an infinite series in $\sqrt{\lambda}/N$. What is striking is that these corrections precisely match those found in the \emph{exact} result (\ref{exactb}) for the fundamental, in the limit where one takes $\lambda,N\rightarrow \infty$ with $\sqrt{\lambda}/N$ fixed. This was first noticed in the computation of the vev of the 1/2 BPS circular Wilson loop \cite{df}, and it is true in spite of the fact that, a priori, $k=1$ lies outside the regime of validity of the D3-brane calculation \cite{fg}. The reasons behind this better than expected behavior are not known, and in particular it is not clear what symmetries the trajectory must preserve for the D3-brane to correctly capture these subleading terms. It then seems worth exploring to what extent one can use D3-brane probe results to obtain $1/N$ corrections to other properties of probes in the fundamental representation. In order to do so, the first ingredient we need are the relevant D3-brane probes, and to those we turn next.
\subsection{Outline and main results}\label{resultssubsec}
In this paper we construct the D3-brane solution dual to a symmetric $k$-quark with arbitrary motion in the vacuum of ${\cal N}=4$ SYM. We start in Section \ref{lightstringsec} by reviewing Mikhailov's embedding (\ref{mikhsolnc}) for the string. Whereas the emphasis in many of the early applications (see, e.g., \cite{dragtemp,lorentzdirac,damping}) was on the reliance of his construction on null geodesics on the string worldsheet, which are only known after the embedding has been specified, we observe here that these curves turn out to also be null geodesics directly in spacetime (see also \cite{iancuradiation}). This implies that Mikhailov's solution can be constructed by shooting light rays in from the AdS boundary. Concretely, the string worldsheet is obtained by tracing an inbound null geodesic from each point on the quark's worldline, with tangent determined by the quark's velocity.
In Section \ref{d3subsec} we generalize this technique to the case of the D3-brane. This requires that from each point on the $k$-quark's trajectory we shoot not one but infinitely many light rays, spanning the surrounding ${\mathbf{S}}^2$. This ultimately leads to the 4-dimensional embedding (\ref{D3covariant}), which, when combined with a worldvolume field strength chosen to be proportional to the induced metric, as specified in (\ref{Fg}), quite remarkably turns out to satisfy all of the (highly nonlinear) equations of motion. We thus achieve our main goal in the paper, to obtain the heretofore unknown embeddings dual to $k$-symmetric probes with an arbitrary timelike worldline. We illustrate the general construction with the explicit example of a probe in circular motion.
The eminently geometric character of our construction method suggests that it should have wider relevance, and indeed, in Section \ref{d5subsec} we show that the known D5-brane embeddings dual to $k$-antisymmetric probes are correctly reproduced by tracing light rays in from the AdS boundary and selecting the field strength according to (\ref{Fg}). Many other extensions seem possible and desirable, and we intend to pursue them in future work.
As a first application of our D3-brane solutions, we compute their total energy and show that, much as in \cite{mikhailov,dragtemp}, it cleanly splits into two contributions attributable respectively to the $k$-quark's intrinsic and radiated energy. For the former, we find
\begin{equation}
E_{k\mbox{\scriptsize q}}=k\gamma m~,
\label{intenergysym}
\end{equation}
meaning that even for arbitrary motion our probe continues to be a threshold bound state. For the latter, we obtain
\begin{equation}
E_{\mbox{\scriptsize rad}}^{{\cal S}_k}=\frac{k\sqrt{\lambda}}{2\pi}\sqrt{1+\frac{k^2\lambda}{16\pi^2N^2}} \int dt_r
\left[\gamma^6\left(\vec a^2-|\vec v\times \vec a|^2\right) \right]~.
\label{radenergysym}
\end{equation}
which has the by now familiar Li\'enard-type form, and confirms the Bremsstrahlung function (\ref{bksym}) deduced previously for a probe with proper constant acceleration. This shows in particular that, at least for this observable, the relation between the fundamental and totally symmetric representations extends to arbitrary probe trajectories.
\section{Strings of Light}\label{lightstringsec}
We focus on the duality between maximally supersymmetric Yang-Mills theory (MSYM) in $3+1$ dimensions, with gauge group $SU(N)$, and Type IIB string theory on 5-dimensional anti-de Sitter (AdS) spacetime cross a 5-sphere. To examine the gauge theory on Minkowski spacetime, we work in the Poincar\'e patch of AdS$_5$, where the metric reads
\begin{equation}\label{metric}
G_{mn}dx^m dx^n=\frac{L^2}{z^2}\left(\eta_{\mu\nu}dx^{\mu}dx^{\nu}+dz^2\right)
=\frac{L^2}{z^2}\left(-dt^2+dx_1^2+dx_2^2+dx_3^2+dz^2\right)~.
\end{equation}
There are also a constant dilaton $e^{\phi}=g_s$ and $N$ units of flux of the self-dual Ramond-Ramond 5-form field strength. The AdS component of the corresponding
potential can be taken to be
\begin{equation}\label{C4}
C_{0123}=-\frac{L^4}{z^4}~.
\end{equation}
The ratio of the AdS radius of curvature $L$ to the string length $\ls$ is related to the MSYM 't~Hooft coupling through
\begin{equation}\label{L}
\frac{L}{\ls}=\lambda^{1/4}~.
\end{equation}
An external (infinitely massive) quark in MSYM is dual to a fundamental string extending from the AdS boundary at $z=0$ to the Poincar\'e horizon at $z\to\infty$. More precisely, it is the endpoint of the string that is dual to the quark, whereas the body of the string encodes the profile of the gluonic (and other) field(s) sourced by the quark. Given a quark trajectory $x^{\mu}(\tau)$, the dual string embedding $X^{\mu}(\tau,z)$ is found by extremizing the usual Nambu-Goto action, subject to the condition that the string endpoint follow the same path as the quark, $X^{\mu}(\tau,0)=x^{\mu}(\tau)$. Remarkably, Mikhailov \cite{mikhailov} was able to solve this problem for \emph{arbitrary} quark motion. His solution takes the surprisingly simple form
\begin{equation}\label{mikhsol}
X^{\mu}(\tau,z)=x^{\mu}(\tau)+v^{\mu}(\tau)z~,
\end{equation}
where $\tau$ is chosen to be the proper time of the quark, and $v^{\mu}\equiv dx^{\mu}/d\tau$ denotes its 4-velocity. This solution is \emph{retarded} or \emph{purely outgoing}, i.e., it is appropriate for the situation of primary physical interest, where excitations of the MSYM fields propagate outward from the quark toward infinity. Flipping the sign in front of $v^{\mu}$ gives instead an advanced or purely ingoing solution.\footnote{The more generic embedding would be a nonlinear superposition of retarded and advanced contributions, but its form is not known.}
The physics of the embedding (\ref{mikhsol}) is more easily unpacked by rewriting it in noncovariant notation,
\begin{eqnarray}\label{mikhsolnc}
t(t_r,z)&=&t_r+\gamma(t_r)z~,
\\
\vec{X}(t_r,z)&=&\vec{x}(t_r)+\gamma(t_r)\vec{v}(t_r)z~.
\nonumber
\end{eqnarray}
The information of interest here is the position at time $t$ of the string bit at radial depth $z$, which essentially codifies the configuration at that time of the gluonic field a distance $\sim z$ away from the quark. We see from the second equation that this information is determined by the position $\vec{x}$ and 3-velocity $\vec{v}$ of the quark/endpoint at the earlier, \emph{retarded} time $t_r$ defined by the first equation, in close analogy with the Li\'enard-Wiechert story in classical electromagnetism.\footnote{The fact that, unlike in classical electromagnetism, the gluonic field is a nonlinear medium that can rescatter signals is accounted for in the dual gravity description by the fact that fields at the AdS boundary receive contributions from each and every point along the string \cite{cg,trfsq}. Even so, 1-point functions of gluonic field observables turn out to have some surprising features, such as beaming \cite{liusynchrotron,veronika,beaming}, lack of radial/temporal broadening \cite{iancuaspects,iancuradiation,trfsq} and a `near'-field tail reaching out to infinity \cite{tmunu,lewkomalda}.}
The curves at constant $t_r$ (or $\tau$) are null geodesics on the string worldsheet, and this property played an important role in Mikhailov's construction. But what will be of interest to us in this paper is the observation that they are also null geodesics in spacetime, as is clear from (\ref{mikhsolnc}), where we see that they are straight lines in the Poincar\'e coordinates (\ref{metric}). We can therefore interpret the Mikhailov embedding as the surface obtained by shooting light rays from the AdS boundary into the bulk, with the starting point and slope of each ray determined respectively by the position and velocity of the quark, according to (\ref{mikhsol}) or (\ref{mikhsolnc}). The nontrivial fact that the ruled surface so obtained is a classical string worldsheet (i.e., a solution of the Nambu-Goto system) was proven somewhat indirectly in \cite{mikhailov}, by arguing that it extremizes the action, and later verified in \cite{dampingtemp} by directly plugging it into the equations of motion.
A connection between null geodesics and strings was discussed previously in studies of light parton energy loss in a thermal gluon plasma \cite{gubserlight,jensenlight1,jensenlight2,arnold}. Light rays were employed there to estimate the trajectory of the endpoint of the string dual to a light quark or gluon, as it falls towards the black hole horizon (see however \cite{horowitz}). In contrast, we are here finding that (at least in pure AdS) the entire worldsheet of the string dual to a heavy quark can be generated by throwing light rays in from the boundary.\footnote{Another interesting use of light rays in the holographic description of quarks appeared in the beaming proposal of \cite{veronika} to determine the gravitational backreaction of the string (see also \cite{beaming}). A relation between light rays and the string worldsheet was also postulated in \cite{zahed} in the finite temperature context, but it was not shown whether the ansatz in question satisfies the equation of motion.} It is perhaps worth emphasizing that the motion of internal points of the string, unlike that of its endpoint, depends on the choice of worldsheet coordinates. To uncover the geometric description of the string in terms of light rays, it is crucial then that in (\ref{mikhsolnc}) (or (\ref{mikhsol})) we are choosing to parametrize the worldsheet in terms of the (proper) retarded time $t_r$ ($\tau$) at the boundary.
\section{Branes of Light}\label{lightbranesec}
Having understood the purely geometric origin of the Mikhailov solution for the string, it is natural to wonder whether it might be possible to construct other brane embeddings by similarly shooting light from the AdS boundary. In this section we will examine this question for the specific case of a D3-brane with $k$ units of electric flux, which is known to be dual to a $k$-quark in the totally symmetric representation of the $SU(N)$ group \cite{df,gp,gp2,dgrt}. As explained in the Introduction, this case is of particular interest because up to now the appropriate D3-brane embeddings are only known for a couple of specific $k$-quark worldlines, in contrast with the case of the totally antisymmetric representation. The latter is dual \cite{baryon,yamaguchi,gp,draggluon} to a D5-brane wrapped on a ${\mathbf{S}}^4\subset {\mathbf{S}}^5$ \cite{pr,cpr}, whose embedding in AdS turns out to exactly coincide \cite{hartnoll}, for arbitrary $k$-quark trajectory, with the string solution (\ref{mikhsol}) dual to a quark with the same trajectory. We will come back to that case at the end of the section.
\subsection{D3-brane}\label{d3subsec}
Let us begin by considering the case of a static $k$-quark, $\vec{x}(t)=\mbox{constant}$. Our goal is to generate the 4-dimensional worldvolume of the dual D3-brane by shooting light rays into the bulk as we move along this worldline at the AdS boundary. In Poincar\'e coordinates, each light ray is a straight line, and will correspond to a fixed retarded time $t_r$. By symmetry, it is clear that identical light rays must be shot along all directions $(\theta,\phi)$ of the ${\mathbf{S}}^2$ that surrounds the location of the $k$-quark. We thus obtain an ansatz for the D3 embedding that is closely analogous to (\ref{mikhsolnc}):
\begin{eqnarray}\label{D3static}
\vec{X}(t_r,z,\theta,\phi)&=&\vec{x}(t_r)+\kappa z\vec{n}~,
\\
t(t_r,z,\theta,\phi)&=&t_r+\sqrt{1+\kappa^2}\,z~,
\nonumber
\end{eqnarray}
where $\vec{n}$ for now denotes the unit vector $(\cos\theta,\sin\theta\cos\phi,\sin\theta\sin\phi)$, $\kappa$ is a number expressing the common slope of all of the straight lines, and the $\kappa$-dependence of the last equation has been fixed with the requirement that these lines be null. To be able to represent a bound state of $k$ quarks, we expect this putative D3-brane to carry $k$ units of fundamental string charge along the AdS radial direction. In other words, the embedding should be sustained by $k$ units of the worldvolume electric flux associated with $F_{t_r z}$ (this will be made more explicit in (\ref{Pitz}) below, after we write down the D3-brane action). The slope $\kappa$ must depend on $k$, because for $k=1$ (and $N\to\infty$) all lines must come together as the D3-brane embedding closes up into the fundamental string (\ref{mikhsolnc}) known to be dual to a static quark.
The D3-brane solution for the static case was obtained in \cite{reyyee, df}, and indeed happens to agree with our ansatz (\ref{D3static}), with
\begin{equation}\label{kappa}
\kappa=\frac{k\sqrt{\lambda}}{4N}~, \nonumber
\end{equation}
and
\begin{equation}\label{Ftzstatic}
F_{t_r z}=\frac{\sqrt{\lambda}}{2\pi}\frac{1}{z^2}~.
\end{equation}
This serves then as a first successful test of our approach.
Notice that (\ref{D3static}) can be rewritten as
\begin{equation}\label{D3constantv}
X^{\mu}(\tau,z,\theta,\phi)=x^{\mu}(\tau)+\left[\sqrt{1+\kappa^2}\,v^{\mu}+\kappa n^{\mu}(\theta,\phi)\right]z~,
\end{equation}
with $v^{\mu}=(1,0,0,0)$ the 4-velocity of the static $k$-quark and $n^{\mu}\equiv(0,\vec{n})$. By Lorentz covariance, we are assured then that this same equation gives the correct D3-brane embedding dual to a $k$-quark translating uniformly, as long as we take $v^{\mu}$ to be the corresponding velocity and $n^{\mu}$ the appropriately boosted 4-vector. Note that the latter automatically satisfies the two constraints $n^2=1$ and $n\cdot v=0$, and so has two independent components, which we are choosing to parametrize with the original rest frame angles $(\theta,\phi)$. The resulting embedding is shown at various times in Fig.~\ref{vfig}, along with some of the null geodesics that generate the D3-brane worldvolume. The corresponding field strength is
\begin{equation}\label{Ftauz}
F_{\tau z}=\frac{\sqrt{\lambda}}{2\pi}\frac{1}{z^2}~,
\end{equation}
i.e., $F_{t_r z}=\sqrt{\lambda}/2\pi\gamma z^2$. Thanks to our choice of worldvolume coordinates, no magnetic components are turned on in spite of the boost.
\begin{figure}[hbt]
\begin{center}
\includegraphics[width=8cm]{vplot.pdf}
\setlength{\unitlength}{1cm}
\begin{picture}(0,0)
\put(-4.3,0.9){ $x_1$}
\put(-5.1,0.5){\vector(2,1){0.8}}
\put(-8.1,1.7){ $x_2$}
\put(-7.3,0.7){\vector(-1,2){0.4}}
\put(-8.9,5.2){ $z$}
\put(-8.4,4.2){\vector(-1,4){0.2}}
\end{picture}
\end{center}
\vspace*{-0.8cm}
\caption{Successive snapshots (at $t=-4,0,4$) of the D3-brane solution dual to a $k$-quark moving at constant velocity along the $x^1$ direction (color online), with $v=0.9$ and $\kappa=1/2$. The embedding is symmetric under rotations in the $x^2$-$x^3$ plane, and the boosted conical surfaces shown here are just the sections at fixed azimuthal angle $\phi=0$. The dashed blue lines are light rays emitted into AdS from the boundary at $t=-4$. The entire worldvolume is generated by such ingoing rays, emitted from all points along the trajectory of the $k$-quark.
\label{vfig}}
\end{figure}
In Mikhailov's solution (\ref{mikhsol}), the string embedding depends only on the quark's position and velocity. It is natural to suspect that the same is true for our case, and we are led then to conjecture that allowing for a time-dependent velocity in (\ref{D3constantv}),
\begin{equation}\label{D3covariant}
X^{\mu}(\tau,z,\theta,\phi)=x^{\mu}(\tau)+\left[\sqrt{1+\kappa^2}\,v^{\mu}(\tau)+\kappa n^{\mu}(\tau,\theta,\phi)\right]z~,
\end{equation}
yields the correct D3-brane embedding dual to a $k$-quark with an \emph{arbitrary} trajectory. The $\tau$-dependence in $n^{\mu}$ is due to the fact that this 4-vector has to be suitably transported along the worldline. To avoid confusion, we will henceforth denote the rest frame unit 3-vector by $\vec{n}_R$, and let $\vec{n}$ stand for the spatial part of the Lorentz-transformed 4-vector $n^{\mu}$. Notice that, due to the properties of this 4-vector stated in the previous paragraph, the proposed embedding (\ref{D3covariant}) satisfies $(X-x)^2=-z^2$, which is a restatement of the fact that light fronts are being emitted from $z=0$ at each $x^{\mu}(\tau)$. This equation is also satisfied by Mikhailov's string embedding, (\ref{mikhsol}).
To visualize our ansatz more explicitly, we can spell it out when the motion is purely
along direction $x\equiv x^1$:
\begin{eqnarray}\label{D31D}
t&=&t_r+\gamma z(\sqrt{1+\kappa^2}+\kappa v\cos\theta )~,
\nonumber\\
X^1&=&x+\gamma z(\sqrt{1+\kappa^2}v+\kappa\cos\theta)~,
\nonumber\\
X^2&=&\kappa z\sin\theta\cos\phi~,
\\
X^3&=&\kappa z\sin\theta\sin\phi~.
\nonumber
\end{eqnarray}
For the specific case of uniform proper acceleration $A$, $x(t)=\sqrt{A^{-2}+t^2}$,
this can be compared against the solution obtained in \cite{df,fg}, which was found to be described by
\begin{equation}\label{D3accel}
\left(-t^2+\vec{X}^2+z^2-A^{-2}\right)^2+4 A^{-2}\left(X_2^2+X_3^2\right)-4\kappa^2 A^{-2}z^2=0~.
\end{equation}
And indeed, the ansatz (\ref{D31D}) can be verified to satisfy this equation, which constitutes a second successful test of our construction. The embedding at various times is shown in Fig.~\ref{afig}, together with some of the corresponding light rays. In parallel with the quark/string case examined in \cite{noline,eprer}, the retarded D3-brane solution (\ref{D31D}) associated with a uniformly accelerated $k$-quark terminates at a finite radial position, and can only be smoothly completed by a suitable \emph{advanced} $k$-antiquark solution. More details on this are given in Appendix~\ref{accelapp}.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=8cm]{aplot.pdf}
\setlength{\unitlength}{1cm}
\begin{picture}(0,0)
\put(-4.6,0.3){ $x_1$}
\put(-5.4,0.9){\vector(2,-1){0.8}}
\put(-0.5,1.7){ $x_2$}
\put(-0.7,0.7){\vector(1,2){0.4}}
\put(-8.9,5.2){ $z$}
\put(-8.4,4.2){\vector(-1,4){0.2}}
\end{picture}
\end{center}
\vspace*{-0.8cm}
\caption{Successive snapshots (at $t=0,2,4$) of the D3-brane solution dual to a $k$-quark and $k$-antiquark undergoing back-to-back uniform acceleration along $x^1$, with $A=1$ and $\kappa=1/2$. The embedding is invariant under azimuthal rotations, and the banana-shaped surfaces shown here are just the $\phi=0$ sections. The dashed blue (red) lines are light rays emitted into AdS from (absorbed from AdS into) the $k$-quark ($k$-antiquark) tip of the brane, at $t=2$. The full worldvolume is generated by all such ingoing (outgoing) rays.
The black circle marks the locus where the retarded quark and advanced antiquark embeddings smoothly join together, which is seen to move at the speed of light in the negative $x^1$ direction (see Appendix~\ref{accelapp}).
\label{afig}}
\end{figure}
The worldvolume $U(1)$ gauge field should still encode $k$ units of the electric flux associated with $F_{t_r z}$. But for this we get no guidance from Mikhailov's solution, so it is \emph{a priori} unclear whether (\ref{Ftauz}) continues to be the only nonvanishing component of the field strength, or if other components are turned on, due to some acceleration-dependent contributions. To settle this point, we must work out and enforce the relevant equations of motion.
Let us turn then to the D3-brane action, which includes both the Dirac-Born-Infeld and Wess-Zumino terms,
\begin{equation}\label{D3action}
S_{\text{D3}}=T_{\text{D3}}\int d^4\xi\, \left(-\sqrt{-\det(g_{\alpha\beta}+2\pi \ls^2 F_{\alpha\beta})}+c_{0123}\right)~,
\end{equation}
where
$T_{\text{D3}}=1/(2\pi)^3\gs\ls^4=N/2\pi^2 L^4$ is the tension of the D3-brane,
$g_{\alpha\beta}=\partial_{\alpha} X^m \partial_{\beta} X^{n}G_{mn}$ is the induced metric,
and $c$ is likewise the pullback of the 4-form $C$ onto the worldvolume,
$c_{\alpha\beta\gamma\delta}=\partial_{\alpha} X^m\partial_{\beta} X^n\partial_{\gamma} X^p\partial_{\delta} X^q C_{mnpq}$.
The worldvolume coordinates are denoted collectively by $\xi^{\alpha}$, and, based on the preceding analysis, will be chosen by us to be
$t_r$, $z$, $\theta$, and $\phi$.
{}From this action we can work out the momentum density conjugate to each of the embedding fields,
$P^{\alpha}_m\equiv \partial\mathcal{L}_{\text{D3}}/\partial(\partial_{\alpha} X^m)$, and to each of the gauge field components, $\Pi^{\alpha\beta}\equiv\partial\mathcal{L}_{\text{D3}}/\partial(\partial_{\alpha} A_{\beta})=\partial\mathcal{L}_{\text{D3}}/\partial F_{\alpha\beta}$. The generic expressions are long and unenlightening, so we will not write them here. Since the fields $X^{\mu}$ and $A_{\alpha}$ do not appear undifferentiated in $S_{\text{D3}}$, their equations of motion are just the statement that these momenta are conserved:
\begin{equation}\label{eom}
\partial_{\alpha} P^{\alpha}_{\mu}=0~,\qquad\partial_{\alpha}\Pi^{\alpha\beta}=0~.
\end{equation}
Consider first the case of motion purely along one dimension. Using our ansatz (\ref{D31D}) for the embedding, and assuming for the time being that (\ref{Ftauz}) is the only nonvanishing field strength component, we obtain
\begin{equation}\label{Pitz}
\Pi^{t_r z}=\frac{k\sin\theta}{4\pi}~.
\end{equation}
Upon angular integration, this gives the desired integer value for the fundamental string charge (along the radial AdS direction) carried by the D3, $\int d\theta d\phi\,\Pi^{t_r z}=k$, thereby validating (\ref{D31D}) and (\ref{Ftauz}). In fact, using (\ref{D31D}) one finds that in order for (\ref{Pitz}) to hold, all magnetic components of the field strength must be set to zero, $F_{z\theta}=F_{z\phi}=F_{\theta\phi}=0$, but the electric components $F_{t_r \theta}$ and $F_{t_r \phi}$ can be arbitrary. Independently of the values of the latter, we find that $\Pi^{t_r \theta}$ and $\Pi^{t_r \phi}$ vanish, just as they should.
Allowing for these electric components of $F$ to be arbitrary, and continuing to use (\ref{D31D}) and (\ref{Ftauz}), we can then notice that
$\Pi^{\theta\phi}=0$ automatically, but
\begin{eqnarray}\label{Pismagnetic}
\Pi^{z\theta}&=&\frac{4N\sqrt{1+\kappa^2}\gamma^2\sin\theta}{4\pi \lambda\kappa}
\left[\frac{2\pi}{\gamma^2}F_{t_r\theta}-\sqrt{\lambda}\kappa a\sin\theta \right]~,
\nonumber\\
\Pi^{z\phi}&=&\frac{2N\sqrt{1+\kappa^2}\csc\theta}{\lambda\kappa}F_{t_r\phi}~,
\end{eqnarray}
with $a$ the acceleration of the $k$-quark.
These momenta must be set to zero, because they would represent unwanted D1-brane charge densities on the D3, and more importantly, because if they do not vanish then the gauge field equations of motion (\ref{eom}) are not satisfied. Notice that this forces us to turn on $F_{t_r\theta}$, with a value proportional to the acceleration, which is consistent with the fact that this component was not found to be excited when the $k$-quark moves with constant velocity.
Altogether, we have deduced then that, in the case of one-dimensional motion, the worldvolume field strength must take the form
\begin{eqnarray}\label{F1D}
F_{t_r z}&=&\frac{\sqrt{\lambda}}{2\pi}\frac{1}{\gamma z^2}~,
\nonumber\\
F_{t_r \theta}&=&\frac{\sqrt{\lambda}}{2\pi}\kappa\gamma^2 a\sin\theta~,
\\
F_{t_r\phi}&=&0~,
\nonumber\\
F_{z\theta}&=&F_{z\phi}=F_{\theta\phi}=0~.
\nonumber
\end{eqnarray}
And as a definitive, highly nontrivial test of our overall ansatz, one can verify that all of the equations of motion (\ref{eom}) are correctly satisfied upon assuming (\ref{D31D}) and (\ref{F1D}).
It will prove useful to note that the field strength (\ref{F1D}) is closely related to the induced metric on the D3-brane,
\begin{equation}\label{Fg}
2\pi \ls^2 F_{\alpha\beta}=-\frac{g_{\alpha\beta}}{\sqrt{1+\kappa^2}}\qquad\forall\quad \alpha<\beta~.
\end{equation}
This relation, the fact that $g_{zz}=0$ (which is nothing but the requirement that each line at constant $t_r,\theta,\phi$ be null), and the vanishing of all off-diagonal space-space components of the induced metric are crucial features of the solution. Together, they imply the drastic simplifications
\begin{equation}\label{bisimp}
\sqrt{-\det(g_{\alpha\beta}+2\pi \ls^2 F_{\alpha\beta})}=\frac{\kappa}{\sqrt{1+\kappa^2}}\sqrt{-\det g_{\alpha\beta}}
\end{equation}
and
\begin{equation}\label{ngsimp}
\sqrt{-\det g_{\alpha\beta}}=-g_{t_r z}L^2 \kappa^2 \mbox{vol}({\mathbf{S}}^2)~,
\end{equation}
where $\mbox{vol}({\mathbf{S}}^2)=\sin\theta$.
Let us now move on to the case of arbitrary 3-dimensional motion. To give meaning to (\ref{D3covariant}), we need a precise specification of the unit vector $n^{\mu}(\tau,\theta,\phi)$, or in other words, a choice of labels $(\theta,\phi)$ for the light rays shot along the different directions of the ${\mathbf{S}}^2$ that surrounds $\vec{x}(\tau)$.
The timelike vector at rest $v_R^{\mu}=(1,0,0,0)$ is converted into an arbitrary 4-velocity $v^{\mu}$ by the Lorentz transformation
\begin{equation}\label{boost}
\Lambda^{\mu}{}_{\nu}=
\left(
\begin{array}{cccc}
\gamma & \gamma v_1 & \gamma v_2 & \gamma v_3 \\
\gamma v_1 & 1+\frac{\gamma^2 v_1^2}{1+\gamma} & \frac{\gamma^2 v_1 v_2}{1+\gamma} & \frac{\gamma^2 v_1 v_3}{1+\gamma} \\
\gamma v_2 & \frac{\gamma^2 v_1 v_2}{1+\gamma} & 1+\frac{\gamma^2 v_2^2}{1+\gamma} & \frac{\gamma^2 v_2 v_3}{1+\gamma} \\
\gamma v_3 & \frac{\gamma^2 v_1 v_3}{1+\gamma} & \frac{\gamma^2 v_2 v_3}{1+\gamma} & 1+\frac{\gamma^2 v_3^2}{1+\gamma}
\end{array} \right)~,
\end{equation}
where $v_i$ denotes the components of the 3-velocity. A subtlety in the case of accelerated motion along more than one dimension, however, is that if we use this canonical boost based on $v^{\mu}(\tau)$ to define $n^{\mu}(\tau,\theta,\phi)$ at each point along the worldline of the $k$-quark, the outcomes at successive points would not be related purely by a boost, but would also include a rotation (this fact is at the root of the Thomas precession). The most natural choice, then, is to avoid this spurious rotation on the ${\mathbf{S}}^2$, by demanding that $n(\tau,\theta,\phi)$ be transported along the worldline as dictated by the Fermi-Walker equation (see, e.g., \cite{misner})
\begin{equation}\label{fw}
\partial_{\tau}n^{\mu}=(n\cdot\partial_{\tau}v) v^{\mu}-(n\cdot v)\partial_{\tau}v^{\mu}~,
\end{equation}
from which the second term drops out because $n\cdot v=0$. In more detail, at some initial time $\tau_0$ we apply the boost (\ref{boost}) to define $n^{\mu}(\tau_0,\theta,\phi)=\Lambda^{\mu}{}_{\nu}n^{\mu}_R(\theta,\phi)$, and then evolve from there using (\ref{fw}) to obtain $n^{\mu}$ at arbitrary times. Note that the result generally depends on the probe's velocity \emph{and} acceleration.
For use below, it is useful to notice that for each $\tau$, the 4 vectors $v^{\mu}$, $n_1^{\mu}\equiv n^{\mu}$, $n_2^{\mu}\equiv \partial_{\theta}n^{\mu}$ and $n_3^{\mu}\equiv\partial_{\phi}n^{\mu}/\sin\theta$ form an orthonormal tetrad,
\begin{equation}\label{tetrad}
v^2=-1~,\qquad v\cdot n_I=0~,\qquad n_I\cdot n_J =\delta_{IJ}~,
\end{equation}
with each $n_I$ transported according to (\ref{fw}), and with the additional property
\begin{equation}\label{tetradlevi}
\epsilon_{\mu \nu \lambda \rho} \; v^\mu n_1^\nu n_2^\lambda n_3 ^\rho=1~.
\end{equation}
Physically, the implication is that these four vectors define a nonrotating reference frame.
It is easy to check that (\ref{tetrad}) and (\ref{tetradlevi}) hold in the rest frame, and they then follow in general from the fact that the angular derivatives commute both with the boost (\ref{boost}) and with the transport (\ref{fw}). Finally, it follows from the definition (\ref{fw}) of Fermi-Walker transport that
\begin{equation}
\partial_{t_r}n_I\cdot n_J =0= \partial_{t_r}v\cdot\partial_{t_r}n_I~,
\end{equation}
and we will make repeated use of these relations in what follows.
The calculation of the induced metric on the embedding (\ref{D3covariant}) is presented in Appendix~\ref{inducedapp}.
Upon examining the momentum densities conjugate to the gauge field, the situation is found to be exactly the same as in the 1-dimensional case: all magnetic components of the worldvolume field strength must be set to zero in order to reproduce the expected $k$ units of string charge (\ref{Pitz}), and all other components of $\Pi^{\alpha\beta}$ vanish automatically except for $\Pi^{z\theta}$ and $\Pi^{z\phi}$. Setting these to zero we thus obtain again two equations for the two unknown electric components of $F$. Just as for 1-dimensional motion, the solution takes the form (\ref{Fg}). More explicitly, we must set
\begin{eqnarray}\label{F3D}
F_{t_r z}&=&\frac{\sqrt{\lambda}}{2\pi}\frac{1}{\gamma z^2}~,
\nonumber\\
F_{t_r \theta}&=&-\frac{\sqrt{\lambda}}{2\pi}\kappa
\partial_{t_r}v\cdot n_2~,
\\
F_{t_r\phi}&=&-\frac{\sqrt{\lambda}}{2\pi}\kappa\sin\theta
\partial_{t_r}v\cdot n_3~,
\nonumber\\
F_{z\theta}&=&F_{z\phi}=F_{\theta\phi}=0~.
\nonumber
\end{eqnarray}
Again it is true that the $zz$ and off-diagonal space-space components of the induced metric vanish, so (\ref{bisimp}) and (\ref{ngsimp}) still hold (see Appendix~\ref{inducedapp}).
And again, the ultimate test is the verification that with (\ref{D3covariant}) and (\ref{Fg}) (or (\ref{F3D})) all equations of motion are correctly satisfied.
We have thus succeeded in constructing our desired D3-brane embeddings by shooting light in from the AdS boundary.
As an example, consider the case of motion with constant angular velocity $\omega$ on a circle of radius $R$ on the $x^2$-$x^3$ plane: $v^{\mu}=\gamma(1,0,-\omega R\sin\omega t,\omega R\cos\omega t)$. Solving the Fermi-Walker equation (\ref{fw}) one finds \cite{misner}
\begin{eqnarray}\label{fwcircular23}
n^0&=&\gamma\omega R \sin \theta \sin (\phi-\gamma \omega t)~,
\nonumber\\
n^1&=& \cos \theta~,
\\
n^2&=&\sin \theta\left[\cos \omega t \cos (\phi-\gamma \omega t)-\gamma \sin \omega t \sin(\phi-\gamma \omega t)\right]~,
\nonumber\\
n^3&=&\sin\theta \left[\sin \omega t \cos (\phi-\gamma \omega t)+\gamma \cos \omega t \sin(\phi-\gamma \omega t)\right]~.
\nonumber
\end{eqnarray}
The resulting embedding (\ref{D3covariant}) is depicted in Fig.~\ref{circularfig}.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=8cm]{circularplot3.pdf}
\setlength{\unitlength}{1cm}
\begin{picture}(0,0)
\put(-7.8,1.0){$x_2$}
\put(-6.4,0.3){\vector(-4,3){0.8}}
\put(-1.6,1.0){$x_3$}
\put(-2.6,0.3){\vector(4,3){0.8}}
\put(-0.3,5.2){$z$}
\put(-0.4,4.2){\vector(1,4){0.2}}
\end{picture}
\end{center}
\vspace*{-0.8cm}
\caption{Snapshot at $t=0$ of the D3-brane solution dual to a $k$-quark undergoing uniform circular motion on the $x^2$-$x^3$ plane, with radius $R=1$, angular frequency $\omega=7/8$ and $\kappa=0.1$. The plot omits $x^1$ and shows only the section of the embedding at polar angle $\theta=\pi/3$, over the full range of the azimuthal angle $\phi$. The dashed blue lines are light rays emitted into AdS from the $k$-quark at $t=0$. The full worldvolume is generated by all such ingoing rays.
\label{circularfig}}
\end{figure}
\subsection{D5-brane}\label{d5subsec}
Encouraged by our success with the D3-brane, it is natural to suspect that the essential idea of the method, tracing out the embedding of interest using null spacetime geodesics, should be equally useful to construct branes of various types in diverse dimensions, probably even on backgrounds other than pure AdS. As a first step in this direction, we now show that the D5-brane solutions dual to a $k$-quark in the totally antisymmetric representation of $SU(N)$ do in fact fit naturally within the same framework.
For the static case, the embeddings we have in mind were deduced in \cite{pr,cpr}, and then understood in \cite{draggluon} to be a special case of the larger family of solutions constructed earlier in \cite{baryon}. Their generalization for arbitrary $k$-quark motion was obtained in \cite{hartnoll}. They are characterized by the fact that the D5-brane wraps an ${\mathbf{S}}^4\subset{\mathbf{S}}^5$ at fixed polar angle $\vartheta\equiv\vartheta_1$, in the usual coordinatization where the 5-sphere metric reads
\begin{equation}\label{metricsphere}
G_{MN}dx^{M} dx^{N}=L^2(d\vartheta_1^2+\sin^2\vartheta_1 d\vartheta_2^2 +\ldots + \sin^2\vartheta_1\sin^2\vartheta_2\sin^2\vartheta_3\sin^2\vartheta_4 d\vartheta_5^2)~.
\end{equation}
If we are given a $k$-quark trajectory $x^{\mu}(\tau)$ and set out to construct a D5-brane of this type using our light ray method, by analogy with the D3-brane story we are led to pick $\xi^{\alpha}=(t_r,z,\vartheta_2,\vartheta_3,\vartheta_4,\vartheta_5)$ as a convenient choice of worldvolume coordinates. We are then required to shoot light inward from the AdS boundary, from each worldline point $x^{\mu}(t_r)$, along each of the directions on the ${\mathbf{S}}^4$, thereby tracing out null geodesics parametrized by $z$. Since we are specifically looking for embeddings with $\vartheta=$constant, these geodesics lie purely within AdS, and have no motion whatsoever along the ${\mathbf{S}}^5$. But then they have no choice but to coincide with the by now familiar straight lines (\ref{mikhsolnc}), meaning that the AdS part of the D5-brane embedding is \emph{identical} to the fundamental string embedding dual to a quark with the same trajectory. And indeed, this is precisely the form of the embeddings deduced in \cite{hartnoll}.
To have a complete solution, we still need to determine the profile of the gauge field. Taking the hint from (\ref{Fg}), we propose a proportionality relation with the induced metric,
\begin{equation}\label{FgD5}
2\pi \ls^2 F_{\alpha\beta}=f g_{\alpha\beta}\qquad\forall\quad \alpha<\beta~,
\end{equation}
with $f$ to be determined. Using (\ref{mikhsolnc}), this says that only the electric component $F_{t_r z}$ is nonvanishing, and it takes a value proportional to $g_{t_r z}=-L^2/\gamma z^2$. (Analogs of the simplifications (\ref{bisimp}) and (\ref{ngsimp}) are then in play.) This proposal indeed can be seen to satisfy the equation of motion for $\vartheta$, but only if the proportionality constant is fixed to $f=-\cos\vartheta$. We thus learn that
\begin{equation}\label{FtzD5}
2\pi \ls^2 F_{t_r z}=\frac{L^2\cos\vartheta}{\gamma z^2}~.
\end{equation}
All that remains is to enforce the condition that the fundamental string charge carried by the D5-brane be integer and equal to the required value, $\int d^4\vartheta\,\Pi^{t_r z}=k$. In other words, the solution must satisfy
\begin{equation}\label{PitzD5}
\Pi^{t_r z}=\frac{k\,\mbox{vol}({\mathbf{S}}^4)}{8\pi^2/3}~,
\end{equation}
with $\mbox{vol}({\mathbf{S}}^4)=\sin^3\vartheta_2\sin^2\vartheta_3\sin\vartheta_4$.
(This is the direct analog of (\ref{Pitz}), and all other $\Pi^{ab}$ are found to correctly vanish.) This condition determines $\vartheta$ in terms of $k$,
\begin{equation}\label{thetak}
\sin\vartheta\cos\vartheta-\vartheta=\frac{\pi k}{N}~.
\end{equation}
Remarkably, (\ref{mikhsolnc}), (\ref{FtzD5}) and (\ref{thetak}) precisely agree with the solutions of \cite{pr,cpr,hartnoll}.
Now that we have obtained with our method both the D3-brane and D5-brane solutions respectively dual to the totally symmetric and antisymmetric representations of $SU(N)$, the same procedure is guaranteed to cover the case of $k$-quarks in \emph{arbitrary} irreducible representations. Indeed, it was shown in \cite{gp} that the dual of a source labelled by an arbitrary Young diagram can be understood as a collection of either D3-branes (if the diagram is decomposed into rows) or D5-branes (if it is decomposed into columns). At leading order in the large $N$ expansion, these branes do not interact with one another, and are therefore exactly as we have described in this section.
The fact that the D5-brane turns out to have the same structure as the D3-brane strengthens the hope that our light ray method will have more general applicability. For instance, we would expect it to also account successfully for the more general antisymmetric $k$-quark embeddings obtained in \cite{baryon}, of which the solutions described in this subsection are a special case. The situation there is more interesting because $\vartheta$ becomes a function of $z$, meaning that the relevant null geodesics will necessarily differ from (\ref{mikhsolnc}). We could similarly consider cases where $\vartheta$ depends in addition (or alternatively) on $t_r$.
The supersymmetric loops of \cite{zarembo,dgrt} are other notable examples associated with a variable direction on the ${\mathbf{S}}^5$.
It is also interesting to note that the connection established in \cite{hartnoll} between D5-brane and string embeddings does not require that the spacetime be pure AdS, which again suggests that the null geodesic construction could hold in more general backgrounds. We leave an exploration of these and other extensions to future work.
\section{Intrinsic and Radiative Energy}
As an application of our newly found D3-brane solutions, in this section we will compute their total energy and show that, similarly to \cite{mikhailov,dragtemp}, it cleanly splits into 2 contributions that represent the intrinsic energy of the $k$-quark and the energy that has already been carried away by gluonic (and scalar) radiation. This exercise will thus lead in particular to a holographic prediction for the rate of radiation of a(n infinitely massive) symmetric $k$-quark at strong coupling.
\subsection{D3-brane energy density}
We wish to compute the energy $E$ of the D3-brane at a fixed observation time $t$. For this purpose, unlike what we did in the previous section, it is convenient to choose as worldvolume coordinates $(t,z,\theta,\phi)$. As before, we use indices $\mu,\nu$ for the Minkowski directions of the target AdS space, $\alpha,\beta$ for worldvolume coordinates, and $a,b$ for spatial worldvolume indices ($a,b=z,\theta,\phi$). Starting from the D3-brane action (\ref{D3action}), the energy density in our chosen parametrization is found to be \cite{fg}
\begin{equation}
{\cal E}=T_{D3}\left(
\frac{L^2}{z^2}\frac{\det(g_{ab}+2\pi \ls^2 F_{ab})}{\sqrt{-\det(g_{\alpha\beta}+2\pi \ls^2 F_{\alpha\beta})}}-
\frac{L^4}{z^4}\frac{\partial X^{\mu}}{\partial z} \frac{\partial X^{\nu}}{\partial \theta} \frac{\partial X^{\lambda}}{\partial \phi} \epsilon_{0\mu\nu\lambda}
\right)~.
\label{energydens}
\end{equation}
To be clear, the numerator in the first term is a $3\times 3$ determinant, with spatial indices only.
The full Born-Infeld determinant that appears in the denominator is available to us from Appendix~\ref{inducedapp}, albeit in the $(t_r,z,\theta, \phi)$ parametrization. Applying to (\ref{bi}) a change of variables to our desired set $(t,z,\theta,\phi)$ yields
\begin{equation}
\det(g_{\alpha\beta}+2\pi\ls^2 F_{\alpha\beta})=-\frac{L^8}{z^4}\left(\frac{\partial t_r}{\partial t}\right)^2 \kappa^6 \gamma^{-2}\sin^2 \theta~.
\label{thebidet}
\end{equation}
Let us now compute the spatial determinant in the numerator of (\ref{energydens}). To do so, define $M=g+2\pi\ls^2F$ and notice that
$\det M_{ab}=M^{00}\det M_{\alpha\beta}$,
where $M$ with indices above denotes the inverse matrix. Our strategy is to compute the matrix inverse to $M$ in the coordinates $(t_r,z,\theta,\phi)$, since in these coordinates the calculation is simpler, and then obtain $M^{00}$ in the $(t,z,\theta,\phi)$ coordinates by a change of variables.
This gives
\begin{eqnarray}
\left(\frac{\partial t}{\partial t_r}\right)^2 \det(g_{ab}+2\pi \ls^2 F_{ab})&=&
g_{\theta \theta} g_{\phi \phi} \partial_z t\left(g_{t_r t_r} \partial_z t-2g_{t_r z}\partial _{t_r}t\right)
\\
{}&{}&
-\frac{\kappa^2}{1+\kappa^2}g_{\theta \theta}\left(g_{t_r\phi}\partial_zt-g_{t_r z}\partial_\phi t\right)^2
\nonumber\\
{}&{}&
-\frac{\kappa^2}{1+\kappa^2}g_{\phi \phi}\left(g_{t_r\theta}\partial_zt-g_{t_r z}\partial_\theta t\right)^2~.
\nonumber
\end{eqnarray}
A rather long computation then leads to\footnote{To arrive at the simplified expression (\ref{subdeter}) we have used the fact that any 4-vector
can be decomposed in terms of the tetrad (\ref{tetrad}), which implies for instance that
$\partial_{t_r}v^{\mu}=(\partial_{t_r}v\cdot n_1)n^{\mu}_1+(\partial_{t_r}v\cdot n_2)n^{\mu}_2+(\partial_{t_r}v\cdot n_3)n^{\mu}_3$.\label{tetradfoot}}
\begin{eqnarray} \label{subdeter}
\left(\frac{\partial t}{\partial t_r}\right)^2 \det(g_{ab}+2\pi \ls^2 F_{ab})=
\frac{L^6\kappa^4 \sin^2\theta}{z^2}\Bigg[(1+\kappa^2)-\kappa^2 \gamma^{-2}\left( (n_1^0)^2+(n_2^0)^2+(n_3^0)^2\right)
\nonumber\\
+ 2z \partial_z t \gamma^{-1}\left( \left(\partial_{t_r}v^0+\frac{\kappa}{\sqrt{1+\kappa^2}}\partial_{t_r} n^0\right)+\gamma \frac{\kappa}{\sqrt{1+\kappa^2}}\partial_{t_r} n\cdot v\right)
+z^2(\partial_z t)^2 \left(\partial_{t_r}v \right)^2 \Bigg] .\quad
\end{eqnarray}
Note that the first line above contains terms independent of the acceleration, while the second line contains terms linear and quadratic in the acceleration.
The last piece in (\ref{energydens}) is the contribution from the Wess-Zumino term. We start by noticing that
\begin{equation}
\epsilon_{0\nu\lambda\rho} \partial_z X^\nu \partial_\theta X^\lambda \partial_\phi X^\rho=
\kappa^2 z^2\left(\frac{\partial t_r}{\partial t}\right) \sin\theta \;
\epsilon_{\mu \nu\lambda\rho} \partial_{t_r}X^\mu \partial_z X^\nu n_2^\lambda n_3^\rho~.
\end{equation}
Then we make use of property (\ref{tetradlevi}) to arrive at
\begin{equation}\label{wzterm}
\frac{\partial X^\mu}{\partial z} \frac{\partial X^\nu}{\partial \theta} \frac{\partial X^\lambda}{\partial \phi} \epsilon_{0 \mu \nu \lambda} =
\kappa^2 z^2 \left(\frac{\partial t_r}{\partial t}\right) \sin\theta \; (\kappa \gamma^{-1}+z\partial_{t_r}n_1\cdot v)~.
\end{equation}
Plugging (\ref{thebidet}), (\ref{subdeter}) and (\ref{wzterm}) into the expression for the energy density (\ref{energydens}), we find
\begin{eqnarray}\label{enerdens}
{\cal E}&=&\frac{N\kappa}{2\pi^2 z^2} \left(\frac{\partial t_r}{\partial t}\right) \sin \theta \, \Bigg[
\gamma +\kappa^2 \gamma^{-1}\left(\gamma^2-1-(n_1^0)^2-(n_2^0)^2-(n_3^0)^2\right)
\nonumber\\
{}&{}&\qquad\qquad
+2z \partial_z t \left( \left(\partial_{t_r}v^0+\frac{\kappa}{\sqrt{1+\kappa^2}}\partial_{t_r} n^0\right)+ \frac{\kappa}{\sqrt{1+\kappa^2}}\gamma \partial_{t_r} n\cdot v\right)-\kappa z\partial_{t_r} n\cdot v
\nonumber\\
{}&{}& \qquad\qquad+\gamma z^2 (\partial_z t)^2 \left(\partial_{t_r}v\right)^2 \Bigg].
\end{eqnarray}
The first, second and third line correspond respectively to terms whose explicit dependence on the acceleration is nonexistent, linear and quadratic (there is also implicit dependence in the $n_I$). In the first line there appears the combination
$\gamma^2-1-(n_1^0)^2-(n_2^0)^2-(n_3^0)^2$, which can be seen to vanish due to (\ref{fw}): differentiate with respect to $\tau$ to show that it is constant; then by boundary conditions it is zero.
\subsection{Total energy and rate of radiation}
We consider first the integral for terms that are independent of or linear in the acceleration,
\begin{eqnarray}\label{Ea}
E_{1}&=&\frac{N\kappa}{2\pi^2} \int dz d\Omega \, \frac{\partial_t t_r}{z^2}
\left[\gamma+2z\partial_z t\left(\left(\partial_{t_r}v^0+\frac{\kappa}{\sqrt{1+\kappa^2}}\partial_{t_r} n^0\right)
\right.\right.
\\
{}&{}&\qquad\qquad\qquad\qquad\qquad\qquad\qquad
+\frac{\kappa}{\sqrt{1+\kappa^2}}\gamma \partial_{t_r} n\cdot v\bigg)-\kappa z\partial_{t_r} n\cdot v\bigg]~,
\nonumber
\end{eqnarray}
with $d\Omega=\sin\theta d\theta d\phi$.
It is convenient to change the integration variable $z\rightarrow t_r$, using
\begin{equation}\label{ztotr}
\frac{\partial z}{\partial t_r}=-\frac{1}{(\partial_t t_r)( \partial_z t)}~.
\end{equation}
This cancels the $\partial_t t_r$ prefactor in (\ref{Ea}). We must also use (\ref{D3covariant}) to substitute $z$ in terms of $t_r$ in the integrand, but for compactness we keep it as a shorthand.
In the string analysis of \cite{mikhailov}, at this stage a total derivative, $d(\gamma z^{-1})/dt_r$, was discarded from the integrand, which was later reinstated in \cite{dragtemp} and shown to have physical significance: upon integration, it yields the intrinsic energy of the quark at time $t$. Inspired by this, we observe that in our D3 context too the derivative
\begin{equation}\label{totalder}
\frac{\partial (\gamma z^{-1})}{\partial t_r}=\frac{\gamma}{z^2\partial_z t}+\frac{\partial_{t_r}\gamma}{z}+\frac{\sqrt{1+\kappa^2}}{z \partial_z t}\gamma\left(\partial_{t_r}v^0+\frac{\kappa}{\sqrt{1+\kappa^2}}\partial_{t_r} n^0\right)~,
\end{equation}
is similar to the integrand we have, but differs by terms proportional to $n^{\mu}$:
\begin{equation}\label{Ea2}
E_{1}=\frac{N\kappa}{2\pi^2} \int dt_r d\Omega \,
\left[\frac{\partial (\gamma z^{-1})}{\partial t_r}
+\frac{\kappa}{t-t_r}\left(n^0\partial_{t_r}\gamma-(\gamma^2-1)n\cdot\partial_{t_r}v\right)\right]~.
\end{equation}
Remembering that $n^{\mu}$ is obtained from its rest frame counterpart $n^{\mu}_R$ by a boost followed by Fermi-Walker transport (\ref{fw}), we see that each of its components is linear in the components of $n^{\mu}_R$ (with coefficients that depend on the 4-velocity and 4-acceleration), so their angular integral at fixed $t_r$ vanishes,
\begin{equation}\label{nintegral}
\int d\Omega \; n^{\mu}_R \; = 0 \quad {\Longrightarrow} \quad \int d\Omega \; n^{\mu}=0~.
\end{equation}
For the remaining, total derivative term, it is evidently convenient to carry out the $t_r$ integral first, to be left with just a surface term at $t_r=t$. The result is in fact divergent, and we choose to regularize it by cutting off the D3-brane at a fixed radial position $z_{\mbox{\scriptsize min}}$ (so the cutoff in $t_r$
depends on $\theta,\phi$).\footnote{Alternatively, one can choose to cut off the integral at $t_r=t-\epsilon$, with $\epsilon$ constant, in which case the corresponding $z_{\mbox{\scriptsize min}}$ would depend on $\theta,\phi$. The results of the two regularizations agree in the limit where the cutoff is removed.}
The leading contribution as the cutoff is removed is then just $\gamma(t)/z_{\mbox{\scriptsize min}}$, which is spherically symmetric, so the angular integral is trivial. We are thus left with
\begin{equation}\label{Easimp}
E_{1}=\lim_{z_{\mbox{\tiny min}}\to 0}\frac{2N\kappa\gamma(t)}{\pi z_{\mbox{\scriptsize min}}}=k m\gamma(t)~,
\end{equation}
where $m=\sqrt{\lambda}/2\pi z_{\mbox{\scriptsize min}}$ denotes the rest mass of a quark ($k=1$) with the same UV cutoff.
Consider now the integral of the terms quadratic in the acceleration. After the change of variable $z\rightarrow t_r$ it can be rewritten as
\begin{equation} \label{Easq}
E_{2}=\frac{N\kappa}{2\pi^2}\int dt_r d\Omega \, \gamma\partial_z t \, (\partial_{t_r} v)^2~.
\end{equation}
Performing the integral over $\theta,\phi$ and using (\ref{nintegral}) again, we finally obtain
\begin{equation}\label{Easqsimp}
E_{2}=\frac{2N}{\pi}\kappa\sqrt{1+\kappa^2}\int dt_r\,\gamma^2 \, (\partial_{t_r} v)^2
=\frac{2N}{\pi}\kappa\sqrt{1+\kappa^2}\int dt_r\,\gamma^6(\vec a^2-|\vec v\times \vec a|^2)~.
\end{equation}
As promised, we have thus found by explicit computation that, much as in \cite{mikhailov,dragtemp}, the total energy of the D3-brane can be understood as the sum of 2 contributions, (\ref{Easimp}) and (\ref{Easqsimp}), that admit a pleasant and direct gauge-theoretic interpretation. $E_1$, which depends only on the state of motion of the probe at the observation time $t$, is the energy attributable to the $k$-quark itself, while $E_2$, which depends on the entire previous history, is the total energy that has been radiated away. The fact that these quantities can be determined analytically in the strong coupling regime constitutes yet another illustration of the power of the AdS/CFT correspondence.
{}From (\ref{Easimp}), which agrees with the result (\ref{intenergysym}) that we had advertised in the Introduction, we see that the $k$-quark has an intrinsic energy that, even for arbitrary motion, equals that of $k$ individual quarks, i.e., it is a threshold bound state. This is a direct consequence of its pointlike nature, Lorentz invariance, and the fact that it must obey the BPS bound.\footnote{It is worth noting that, as the UV cutoff is removed to obtain (\ref{Easimp}), there is a subleading, order $\epsilon^0$ correction to the intrinsic energy of both the quark and the $k$-quark. This contribution, proportional to $\vec{v}\cdot\vec{a}$, is closely related to the results of \cite{dragtemp,lorentzdirac,damping} for a finitely-massive quark, and also, to the surprising fact that the `near field' of the probe has a tail that reaches out to infinity \cite{iancuradiation,tmunu}.}
Expression (\ref{Easqsimp}), which we had advertised in (\ref{radenergysym}), amounts to a prediction for the rate of radiation of a totally symmetric color source in the strongly-coupled gauge theory: it is given by the usual Li\'enard formula, with a numerical coefficient that matches the findings of \cite{fg,fgl} in the context of a more indirect calculation which examined only the specific case of uniform acceleration. Now that we have access to the D3-brane dual to a symmetric $k$-quark undergoing arbitrary motion, we are able to verify that the functional form of the rate of energy loss is unmodified, and the corresponding Bremsstrahlung function is indeed given by (\ref{bksym}) in the supergravity limit. For $k=1$, and at leading order in the $1/N$ expansion, (\ref{Easqsimp}) reproduces Mikhailov's result (\ref{radenergyfund}) for the rate of radiation of a quark. And, since the former coincides with the result of \cite{fg,fgl}, it shares the remarkable property that the entire series of corrections in powers of $\kappa\propto \sqrt{\lambda}/N$ matches onto the exact result obtained in \cite{correa,lewkomalda} for a probe in the fundamental representation, in the limit where $N,\lambda\to\infty$ with $\sqrt{\lambda}/N$ fixed.
\section*{Acknowledgements}
We are grateful to Mariano Chernicoff, Roberto Emparan, Jaume Garriga, Simone Giombi, Veronika Hubeny, Igor Klebanov, Aitor Lewkowycz and David Mateos for useful discussions, and to Mariano Chernicoff for suggestions on the manuscript. BF would like to thank the organizers of Mextrings 2014 for the welcoming atmosphere that ignited this collaboration. He is also grateful to the Crete Center for Theoretical Physics, the Theory Division at CERN and also CERN-Korea Theory Collaboration funded by National Research Foundation (Korea), for hospitality during the course of this work. The research of BF is supported by MEC FPA2010-20807-C02-02, CPAN CSD2007-00042, within the Consolider-Ingenio2010 program, and AGAUR 2009SGR00168.
AG is partially supported by Mexico's National Council of Science and Technology (CONACyT) grant 104649, DGAPA-UNAM grant IN110312, and sabbatical fellowships from DGAPA-UNAM and CONACyT. He would also like to thank the Department of Physics of Princeton University, and Igor Klebanov in particular, for hosting his sabbatical.
JFP is partially supported by the National Science Foundation under Grant No. PHY-1316033 and by Perimeter Institute for Theoretical Physics. Research at Perimeter Institute is funded by the Government of Canada through Industry Canada and by the Province of Ontario through the Ministry of Research and Innovation.
|
\section{First proof of Theorem~\ref{1}}
\begin{proof}[First proof of Theorem~\ref{1}.]
Let $b \ge 2$ be an integer. Recently, \citet{Bugeaud:2008}
constructed a class ${\cal C}$
of real numbers whose irrationality exponent can be read off from their base-$b$ expansion.
The class ${\cal C}$ includes the real numbers of the form
\[
\xi_{\bf n} = \sum_{j \ge 1} \, b^{-n_j},
\]
for a sequence
%
%
${\bf n} = (n_j)_{j \ge 1}$ of positive integers satisfying
$n_{j+1} / n_j \ge 2$
for every large integer $j$.
To obtain good rational approximations to $\xi_{\bf n}$, we
simply truncate the above sum. Thus, we set
\[
\xi_{{\bf n}, J} = \sum_{j = 1}^J \, b^{-n_j} = {p_J \over b^{n_J}}, \quad J \ge 1.
\]
It then follows from
\[
\left|\xi_{\bf n} - \frac{p_J}{b^{n_J}}\right| < \frac{2}{(b^{n_J})^{n_{J+1}/n_J}},
\quad J \ge 1,
\]
that the irrationality exponent $\mu (\xi_{\bf n})$ of $\xi_{\bf n}$ satisfies
\[
\mu (\xi_{\bf n}) \ge \limsup_{j \to \infty} \, \frac{n_{j+1}}{n_j}\;.
\]
\cite{Shallit:1982}
proved that the continued fraction expansion of some rational translate
of any such $\xi_{\bf n}$ can be given explicitly, and \cite{Bugeaud:2008}
proved that its irrationality exponent is given precisely by
\[
\mu (\xi_{\bf n}) = \limsup_{j \to \infty} \, \frac{n_{j+1}}{n_j}\;,
\]
and hence can be read off from its expansion in base $b$.
This means that the denominators of the best rational approximations
to $\xi_{\bf n} $ are (except finitely many of them) powers of~$b$.
Consequently, given a real number $a\geq 2$ for which there is a computable sequence $(a_j)_{j\geq
0}$ of rational numbers such that $\limsup_{j\to \infty} a_j=a$, it is sufficient to construct a
computable strictly increasing sequence ${\bf n} = (n_j)_{j \ge 1}$ of positive integers
satisfying $n_{j+1} / n_j \ge 2$ and
\[
\limsup_{j \to \infty} \, \frac{n_{j+1}}{n_j} = a,
\]
which we do as follows. By substituting $2$ for any smaller values, we may assume that each $a_j$
is greater than or equal to $2$. We construct the desired sequence ${\bf n}$ by induction as
follows. Let $n_1=2$. Given $n_1,\dots,n_j$, let $n_{j+1}$ be the least $n$ such that $n/n_j\geq
a_{j+1}$. By construction, for all $j$, $n_{j+1}/n_j \ge 2$. Consequently, $n_j\geq 2^j$.
Since $(n_{j+1}-1)/n_j<a_{j+1}$,
$n_{j+1}/n_j -a_{j+1}$ is less than or equal to $1/2^j$. It follows directly that
$\limsup_{j\to\infty}n_{j+1}/n_j$ is equal to $\limsup_{j\to \infty} a_j=a$.
\end{proof}
\section{Second proof of Theorem~\ref{1}}
For each real number $a$ greater than $2$, \cite{Jar31} gave a Cantor-like construction of a fractal
subset~$K$ of $[0,1]$ such that the uniform measure $\nu$ on $K$ has the property that the set of
real numbers with irrationality exponent equal to $a$ has $\nu$-measure equal to~$1$. Thus, for all
real numbers $b$ greater than~$a$, the set of real numbers in $K$ with irrationality exponent equal
to $b$ has $\nu$-measure equal to~$0$.
\begin{lemma}[\protect{\cite{Jar31}}] \label{5} For every real number $a$ greater
than or equal to $2$, the set of numbers with irrationality exponent equal to~$a$ has Hausdorff
dimension~$2/a$.
\end{lemma}
We note that \citet{Jar29} and \citet{Bes34} independently established that the set of real numbers
with irrationality exponent greater than or equal to~$a$ has Hausdorff dimension~$2/a$. Actually,
Lemma~\ref{5} is not explicitly stated in \citet{Jar31}; however, it is an immediate consequence of
the results of that paper.
In the following and throughout this text, we denote by $|I|$ the length of the interval~$I$.
\begin{lemma}[Mass Distribution Principle]\label{6}
Let $\nu$ be a finite measure, $d$ a positive real number and $X$ a set with Hausdorff dimension
less than~$d$. Suppose that there is a positive real number $C$ such that for every interval $I$,
$\nu(I)<C \ |I|^d$. Then we have $\nu(X)=0$.
\end{lemma}
\begin{lemma}\label{7}
Let $\vec{a}$ be a strictly decreasing sequence of rational numbers greater than $2$ which is
computable in~$0'$ and has limit equal to $a$, greater than~$2$.
There is a Cantor-like construction of a fractal $K$, with uniform measure $\nu$,
and a function $C$, computable in~$0'$, from ${\mathbb{Q}}\cap(0,2/a)$ to ${\mathbb{Q}}$ such
that for each rational number $d<2/a$, for every interval $I$,
$\nu(I) \leq C(d) |I|^d$.
\end{lemma}
\begin{proof}
We follow the proof of Jarn\'{\i}k's Theorem as presented in \cite{Fal03}.
Let $\vec{a}$ be $(a_j)_{j\geq 0}$. Fix a computable doubly-indexed sequence $(a(j,s))_{j,s\geq 0}$ of rational numbers
such that for all $j$, $\lim_{s\to\infty}a(j,s)=a_j$. Without loss of generality, we assume that
for every~$s$, the sequence $a(j,s)_{j\geq0}$ is strictly decreasing, $a(0,0) $ is an
integer and for all $s$, $a(0,s)=a(0,0)$. Further, we fix a rational number $\beta$ greater than
$2$ and assume that $\beta$
is a lower bound for the numbers $a(j,s)$.
We fix some notation to be applied in the course of our eventual construction. For a positive
integer $q$ and a real number $b$ greater than $\beta$, let
\[
G_q(b)=\left\{x\in \left(\frac{1}{q^{b}}, 1-\frac{1}{q^{b}}\right): \exists p\in{\mathbb{Z}},
\left|\frac{p}{q} - x\right|< \frac{1}{q^b}\right\}.
\]
For $M$ a sufficiently large positive integer according to $\beta$, and $p_1$ and $p_2$ primes
such that $M< p_1<p_2<2M$, the sets
$G_{p_1}(b)$ and $G_{p_2}(b)$ are disjoint and the distance between any
point in $G_{p_1}(b)$ and any point in $G_{p_2}(b)$ is greater than or equal to
\[
\frac{1}{4M^2}-\frac{2}{M^b} \geq \frac{1}{8 M^2}.
\]
For such $M$ the set
\[
H_M(b)=\bigcup_{\substack{p \text{ prime}\\M<p< 2M}} G_p(b)
\]
is the disjoint union of the intervals composing
the sets $G_p(b)$, so $H_{M}(b)$ is made up of intervals of
length less than or equal to $2/M^{b}$ which are separated by gaps of length at least
$1/(8M^2)$. If $I\subseteq[0,1]$ is any interval with $3/|I|< M< p< 2M$ then at least $p|I|/3
> M |I|/3$ of the intervals in $G_p(b)$ are completely contained in $I$. By the prime number
theorem, for sufficiently large $M$ the number of primes between $M$ and $2M$ is at least $M/(2\log
M)$. Thus, for such $M$ and $I$, at least $M^2|I|/(6\log M)$ intervals of $H_M(b)$ are contained in
$I$. With $M_1$ sufficiently large as above and larger than $3 \times 2^{a(0,0)}$, let
\[
M_k= M_{k-1}^k = M_1^{k!}, \ \ \ (k \ge 1).
\]
For a positive integer $k$, let $j$ be the least integer less than $k$ such that $a(j+1,k)\neq
a(j+1,k-1)$, if such exists, and let $j$ be $k-1$, otherwise. That is, $j$ is the greatest index
less than $k$ such that the approximation to $\vec{a}$ remains unchanged at positions less than or
equal to $j$ from step $k-1$ to step $k$. Let
\[
b_k= a(j,k).
\]
Let $E_0=[0,1]$ and for $k=1,2,\ldots$ let $E_k$ consist of those intervals of $H_{M_k}(b_k)$ that
are completely contained in $E_{k-1}$. By discarding intervals if necessary, we arrange that
all intervals in $E_{k-1}$ are split into the same number of intervals in $E_k$. The intervals of
$E_k$ are of length at least $1/(2M_k)^{b_k}$ and are separated by gaps of length at least
\[
g_k=\frac{1}{8 M_k^2}.
\]
Thus, each interval of $E_{k-1}$ contains at least $m_k$ intervals of $E_k$ where $m_1=1$ and
\[
m_k=\frac{M_k^2 }{ (2 M_{k-1})^{b_k} 6\log M_k } \geq \frac{c M_k^2 }{( M_{k-1})^{b_k} \log M_k },
\]
if $k\geq 2$ and $c= 1/(2^{a(0,0)} 6)$.
Let
\[
K=\bigcap_{k\geq 1} E_k.
\]
Define a mass distribution $\nu$ on $K$ by assigning a mass of
$1/(m_1 \times \ldots \times m_{k})$
to each of the $m_1 \times \ldots \times m_{k}$ many $k$-level intervals.
Let $S$ be a subinterval of $[0,1]$. For a lighter notation we write $2\epsilon$ to denote
the length $|S|$ of $S$.
We estimate $\nu(S)$.
Let $k$ be the integer such that
$g_k\leq 2 \epsilon < g_{k-1}$.
The number of $k$-level intervals that intersect $S$ is
\begin{itemize}
\item at most $m_{k}$, since $S$ intersects at most one $(k-1)$-level interval.
\item at most $1+ 2\epsilon/g_{k} \leq 4\epsilon/g_{k}$ since the $k$-level intervals have gaps of at least $g_k$ between them.
\end{itemize}
Each $k$-level interval has measure $1/(m_1\times\ldots\times m_k)$ so that
\[
\nu(S)\ \leq \ \frac{\min(4\epsilon/g_k,m_k)}{m_1\times\ldots\times m_k} \ \leq \
\frac{ (4 \epsilon/g_k)^{ s } \ m_{k}^{1-s}}{m_1\times\ldots\times m_k},
\]
for every $s$ between $0$ and $1$.
Hence,
\[
\nu(S)\ \leq\ \frac{2^s}{(m_1\times \ldots\times m_{k-1}) \ m_k^s g_k^s} (2\epsilon)^s.
\]
Thus, $\nu(S)$ is at most
\begin{align*}
&
1
\frac
{ M_{1}^{b_2} \log M_2 }
{c M_2^2 }\ \
\frac{M_2^{b_{3} }\log M_{3} }
{c M_{3}^2 }
\ldots
\frac
{ M_{k-2}^{b_{k-1}} \log M_{k-1} }
{c M_{k-1}^2 }
\ \
\frac{2^s}{m_k^s g_k^s} \ (2\epsilon)^s =
\\
&
\frac
{ M_{1}^{b_2} \log M_2 }
{c M_2^2 }\ \
\frac{ M_2^{b_{3} }\log M_{3} }
{c M_{3}^2 }
\ldots
\frac
{ M_{k-2}^{b_{k-1}} \log M_{k-1} }
{c M_{k-1}^2 }
\left(\frac{ M_{k-1}^{b_k} \log M_k }{c M_k^2 }\right)^s
({8 M_k^2})^s
2^s\
(2\epsilon)^s =
\\
& \big(\log M_2 \ldots \log M_{k-1}\big)
\big(M_{1}^{b_2} M_{2}^{b_3-2} \ldots M_{k-2}^{b_{k-1}-2}\big)
( \log M_k)^s
(16 )^s
c^{-k+2-s} M_{k-1}^{b_ks-2 }
(2\epsilon)^s.
\end{align*}
We want to verify that
for every $j$ and for every $s< 2/a_j$
there is a $C$ such that
$\nu(S) < C (2\epsilon)^s$.
It suffices to show that there is a $C$ such that for every $k$,
\begin{equation}
\label{{*}}
\tag{*}
\big(\log M_2 \ldots \log M_{k-1}\big)
\big(M_{1}^{b_2} M_{2}^{b_3-2} \ldots M_{k-2}^{b_{k-1}-2}\big)
( \log M_k)^s (16 )^s \ c^{-k+2-s} < \ C \ M_{k-1}^{2-b_ks}.
\end{equation}
Fix $k_0$ such that for every $k\geq k_0,$ $a(j+1,k)=a(j+1,k_0)$.
Thus, for every $k\geq k_0,$ $a(j+1,k)=a_{j+1}$.
Then, define $\delta>0$ as follows so that
for every $k\geq k_0$,
\[
2-\left( b_k \frac{2}{a_j}\right) \geq 2- 2\left( a_{j+1} \frac{2}{a_j}\right) \geq 2 - 2\frac{a_{j+1}}{a_{j}} = \delta.
\]
By the choice of $k_0$ and the definition of $b_k$, for all $k>k_0$, it holds that $b_k<a_{j+1}$.
Hence the left hand side of the inequality \eqref{{*}} is at most a constant multiple of
\[
\big(\log M_2 \ldots \log M_{k-1}\big)
\big(M_{1}^{a_{j+1}} M_{2}^{a_{j+1}-2} \ldots M_{k-2}^{a_{j+1}-2}\big)
(\log M_k)^s (16 )^s \ c^{-k+2-s}.
\]
Furthermore, there is a constant $C$ such that
\[
\big(\log M_2 \ldots \log M_{k-1}\big)
\big(M_{1}^{a_{j+1}} M_{2}^{a_{j+1}-2} \ldots M_{k-2}^{a_{j+1}-2}\big)
(\log M_k)^s (16 )^s \ c^{-k+2-s}
< \ C \ M_{k-1}^{\delta}.
\]
The above inequality follows by noticing that $M_\ell = M_1^{\ell!}$ for $\ell \ge 1$,
taking logarithms on each side and recognizing that
the contribution of $M_{k-1}$ is the dominating term for sufficiently large $k$.
The value of $C$ is determined by the value $k_0,$ which is computable in~$0'$
as a function of $j$.
\end{proof}
\begin{proof}[Second proof of Theorem~\ref{1}.]
Let $a$ be a real number right-computably enumerable in~$0'$ and greater than
$2$ (for $a$ equal $2$, taking $x$ equals $\sqrt{2}$ suffices). Fix a computable doubly-indexed sequence
$(a(j,s))_{j,s\geq 0}$ of rational numbers
satisfying property (3) of Lemma~\ref{4}.
That is, we assume that $\lim_{j\to\infty}
\lim_{s\to\infty}a(j,s)=a$, for all $s$ the sequence $(a(j,s))_{j\geq 0}$ is strictly decreasing,
for all $j\geq 0$ the sequence $(a(j,s))_{s\geq 0}$ is eventually constant,
for all $s$, we have $a(0,s)=a(0,0)$ and $a(1,s)=a(1,0)$.
The last condition gives an appropriate
initialization of the construction. Let $K$ be the fractal with measure $\nu$ and $C$ be the
function associated with this approximation of $a$ in Lemma~\ref{7}. Fix a computable function
$C(r,s): {\mathbb{Q}}\times{\mathbb{N}}\to {\mathbb{Q}}$ such that for every $r$ in ${\mathbb{Q}}$, $(C(r,s))_{s \ge 0}$ is eventually
equal to $C(r)$. We
may also assume that for all $s$, $C(a(1,s),s)=C(a(1,0),0)$.
We compute a real number $x$ in $K$. By recursion on $s$ we construct a sequence of nested
intervals $(I(s))_{s\geq 0}$ such that if $I(s)$ is different from $I(s-1)$ then $I(s)$ is an
element of the $s$-level of $K$. We define an auxiliary function $\ell(s)$, with infinite limit,
to approximate the convergence of the sequence $a(j,s)$. We also define an auxiliary
integer-valued function $q(j,s)$ where $j$ is an integer in $[0,\ell(s))$, with the intention that
$x$ avoids approximation by rational numbers with denominator $q$ greater than or equal to
$q(j,s)$ within $1/q^{a(j,s)}$.
This intention will be realized in the construction at step $s$ onwards provided that
at every step $t\geq s$, $\ell(t)$ is greater than $j$; in particular,
provided that $a(j,s)$ and $C(a(j,s),s)$ have reached their limit values relative to~$s$.
We will employ the following estimate. For a natural number $q_0$ and a real number~$b$ greater
than or equal to~$2$, let
\[
V(q_0,b)= \bigcup_{q\geq q_0} \left\{x\in \left(\frac{1}{q^{b}}, 1-\frac{1}{q^{b}}\right): \exists p\in{\mathbb{Z}},
\left|\frac{p}{q} - x\right|< \frac{1}{q^b}\right\}.
\]
Suppose that $b_1>b_2>a$. By Lemma~\ref{7}, we can estimate $\nu(V(q_0,b_1))$ by
\begin{align*}
\nu\big(V(q_0,b_1)\big)&\leq \sum_{q\geq q_0}\sum_{0<p<q} C\big({2}/{b_2}\big)\Big(\frac{2}{q^{b_1}}\Big)^{2/b_2}\\
&\leq 2 C\big({2}/{b_2}\big)\sum_{q\geq q_0} q \Big(\frac{1}{q^{b_1}}\Big)^{2/b_2}\\
&\leq 2 C\big({2}/{b_2}\big)\sum_{q\geq q_0} \frac{1}{q^{2 b_1/b_2-1}}.
\end{align*}
Thus, for any $\epsilon>0$ there is a $q_0$, uniformly computable from $\epsilon$, $b_1,$ $b_2$
and $C(2/b_2)$, such that $\nu(V(q_0,b_1))$ is less than $\epsilon$.
\begin{enumerate}
\item[] {\em Initial step $0$}. Start with $I(0)$ equal to the unit interval and $\ell(0)=0.$
\item[] {\em Step $s$, greater than $0$}.
Let $\ell(s)$ be the least $j$ less than or equal to $s$ such that
\[
a(j+1,s-1)\neq a(j+1, s) \text{ or } C\big({2}/{a(j+1,s)}, s-1\big) \neq C\big({2}/{a(j+1,s)}, s\big)
\]
if such exists; otherwise, let $\ell(s)$ be $s$. By our assumptions on $a(j,s)$ and
$C(a(1,0),s),$ for every $s>0$, we have that $\ell(s)\geq 1.$
Let $m(s)$ be the $\nu$-measure given to a level-$s$ interval in~$K$. We find $h(s)$ so that the
following inequality holds for each $j$ such that $0\leq j<\ell(s)$,
\[
2 C\big({2}/{a(j,s)}, s\big) \sum_{q\geq h(s)} {1}/{q^{\frac{2 a(j,s)}{a(j+1,s)}-1}}<\frac{1}{s}\ \frac{m(s)}{ 2^{s}}.
\]
We define $q(j,s)$ for each $j\in [0,\ell(s))$ as follows: if $q(j,s-1)$ is defined then let $q(j,s)=q(j,s-1)$; otherwise, let $q(j,s)=h(s)$.
Let $I(s)$ be the leftmost level-$s$ interval in $K$ that is included in $I(s-1)$ and satisfies
\[
\nu\Big( I(s)\cap \bigcup_{0\leq j< \ell(s)}
V\big( q(j, s) , a(j,s)\big)\setminus V\big(h(s) , a(j,s)\big) \Big) < m(s) - 2 \frac{m(s) }{2^{s}}
\]
if such exists; otherwise, let $I(s)$ be $I(s-1)$.
Note that $m(s) \le \nu (I(s))$.
\end{enumerate}
We now verify that the construction works. Define ${\ell_{\text{\it min}}}(s)=\min_{t\geq s}\ell(t)$. We show by
induction on $s$ that
\[
\nu\Big(I(s)\cap\bigcup_{0\leq j<{\ell_{\text{\it min}}}(s)}V\big(q(j,s),a(j,s)\big)\Big) \leq \nu(I(s))\Big(1-\frac{1}{2^s}\Big).
\]
Since ${\ell_{\text{\it min}}}(0)=0$, the inductive claim holds for $s=0$.
Assume the inductive claim for $s-1$:
\[
\nu\Big(I(s-1)\cap\bigcup_{0\leq j<{\ell_{\text{\it min}}}(s-1)}V\big(q(j,s-1),a(j,s-1)\big)\Big) \leq\nu\big(I(s-1)\big)\Big(1-\frac{1}{2^{s-1}}\Big).
\]
Consider those integers $j$ such that $j<{\ell_{\text{\it min}}}(s)$. By the definition of ${\ell_{\text{\it min}}}$, we have
$a(j,s)=\lim_{t\to\infty}a(j,t)$ and $C(2/a(j,s),s)=\lim_{t\to\infty}C(2/a(j,s),t)=C(2/a(j,s)).$
Further, by the discussion above,
\[
\nu\Big(V\big(h(s),a(j,s)\big)\Big)\leq 2 C\big({2}/{a(j,s)}, s\big) \sum_{q\geq h(s)} 1/q^{\frac{2a(j,s)}{a(j+1,s)}-1}.
\]
In the construction we choose $h(s)$ so that for each $j$ less than $\ell(s)$, the term
on the right side of this inequality is less than $m(s)/(s 2^s)$.
This ensures that for each $j$ less than ${\ell_{\text{\it min}}}(s)$,
the same upper bound holds for $\nu(V(h(s),a(j,s)))$.
Now, consider the action of the construction during step $s$.
If $I(s)$ is equal to $I(s-1)$, then
\begin{align*}
& I(s)\cap\bigcup_{0\leq j<{\ell_{\text{\it min}}}(s)}V\big(q(j,s),a(j,s)\big) =\\
& \qquad \Big(I(s)\,\cap\bigcup_{0\leq j<{\ell_{\text{\it min}}}(s-1)}V\big(q(j,s-1),a(j,s-1)\big)\Big)\ \cup \\
& \qquad \Big(I(s)\,\cap\bigcup_{{\ell_{\text{\it min}}}(s-1)\leq j<{\ell_{\text{\it min}}}(s)}V\big(q(j,s),a(j,s)\big)\Big).
\end{align*}
The first component of the union has $\nu$-measure at most $\nu(I(s))(1-1/2^{s-1})$ and the second
component has $\nu$-measure at most $m(s)/2^s$. The union has measure at most
$\nu(I(s))(1-1/2^{s})$, as required.
Otherwise, $I(s)$ is a proper subinterval of $I(s-1)$ and satisfies
\[
\nu\Big( I(s)\cap \bigcup_{0\leq j< \ell(s)} V\big( q(j, s) , a(j,s)\big)\setminus V\big(h(s),
a(j,s)\big) \Big) < m(s) - 2 \frac{m(s) }{2^{s}}.
\]
Then,
\begin{align*}
&I(s)\cap\bigcup_{0\leq j<{\ell_{\text{\it min}}}(s)} V\big(q(j,s),a(j,s)\big)=\\
&\qquad \Big(I(s)\cap \bigcup_{0\leq j< {\ell_{\text{\it min}}}(s)} V\big( q(j, s) , a(j,s)\big)\setminus V\big(h(s) ,
a(j,s)\big) \Big)\ \ \cup\\
&\qquad \Big(I(s)\,\cap\bigcup_{0\leq j<{\ell_{\text{\it min}}}(s)}V\big(h(s),a(j,s)\big)\Big).
\end{align*}
The $\nu$-measure of the first component of the union is less than
\[
m(s) - 2 \frac{m(s) }{2^{s}} =\nu\big(I(s)\big)\Big(1 - \frac{1}{2^{s-1}}\Big).
\]
As in the previous case, the $\nu$-measure of the second component is
less than $\nu(I(s))/2^s$. Again, the union has measure at most $\nu(I(s))(1-1/2^{s})$, as required.
It remains to show that there are infinitely many $s$ such that $I(s)$ is a proper subinterval of
$I(s-1)$. Consider an $s$ such that $\ell(s)$ is equal to ${\ell_{\text{\it min}}}(s)$. Since
\[
\nu\Big(I(s-1)\cap\bigcup_{0\leq j<{\ell_{\text{\it min}}}(s-1)}V\big(q(j,s-1),a(j,s-1)\big)\Big)<\nu\big(I(s-1)\big)\Big(1 - \frac{1}{2^{s-1}}\Big),
\]
we may fix an $s$-level subinterval $I$ of $I(s-1)$ such that
\[
\nu\,\Big(I\,\cap\,\bigcup_{0\leq j<{\ell_{\text{\it min}}}(s-1)}V\big(q(j,s-1),a(j,s-1)\big)\Big)< \nu(I)\Big(1 - \frac{1}{2^{s-1}}\Big).
\]
For this $I$,
\begin{align*}
&I\cap \bigcup_{0\leq j< \ell(s)}V\big( q(j, s), a(j,s)\big)\setminus V\big(h(s),
a(j,s)\big) \subseteq \\
&\qquad \Big(I\cap \bigcup_{0\leq j< {\ell_{\text{\it min}}}(s-1)} V\big( q(j, s), a(j,s)\big)\Big) \ \cup \\
&\qquad \Big(I\cap \bigcup_{{\ell_{\text{\it min}}}(s-1)\leq j< \ell(s)} V\big( q(j, s), a(j,s)\big)\setminus V\big(h(s), a(j,s)\big) \Big).
\end{align*}
For each $j$ such that $\ell(s-1)\leq j< \ell(s)$, $q(j,s)$ is equal to $h(s)$, so the second
component of the union is empty. Thus,
\begin{align*}
\nu\Big( I\cap \bigcup_{0\leq j< \ell(s)}
V\big( q(j, s) , a(j,s)\big)\setminus V\big(h(s) , a(j,s)\big) \Big) &<\ \nu(I)\Big(1 - \frac{1}{2^{s-1}}\Big)
\\& = \ m(s) - 2 \frac{m(s) }{2^{s}}.
\end{align*}
Hence, the conditions for the construction to define $I(s)$ to be a proper subinterval of $I(s-1)$
are satisfied, as required.
Consider the sequence given by the closures of the intervals $I(s), {s\geq0}$. This is a
computable nested sequence of intervals whose lengths approach zero in the limit. Let $x$ be the
unique real number in their intersection. By construction, $x$ is computable (as is its base-$b$
expansion, for every integer $b$ greater than or equal to $2$.)
We now prove that the irrationality exponent of $x$ is equal to~$a$. For each $j\geq 0$, let $b_j=\lim_{s\to
\infty} a(j,s)$. The sequence $(b_j)_{j\geq 0}$ is strictly decreasing with limit~$a$. The
construction ensures that for every $j$, there is a step $s$ such that $I(s)$ is a level-$s$
interval of $K$ containing real numbers that have at least one rational approximation $p/q$ within
$1/q^{b_j}$. Thus, the real number $x$ has irrationality exponent greater than or equal to $a$. We
now show it can not be greater than $a$. Suppose that $b$ is greater than $a$. Let $j$ be such
that $b$ is greater than $b_j$ and let $s$ be such that ${\ell_{\text{\it min}}}(s)$ is greater than $j$. Then, for
all $t>s$, $a(j,t)=a(j,s)=b_j$ and $q(j,t)=q(j,s)$. Further, for any $t>s$,
$\nu(I(t)\setminus V(q(j,t),b_j))$ is positive. If there were an integer $q>q(j,s)$ and an integer $p$ such
that
\[
\left|x-\frac{p}{q}\right|<\frac{1}{q^b},
\]
then there would be a $t$ greater than $s$ such that
\[
I(t)\subset
\Big (\frac{p}{q}-\frac{1}{q^{b_j}},\frac{p}{q}+\frac{1}{q^{b_j}}\Big).
\]
But then $I(t)\setminus V(q(j,t),b_j)$ would be empty, a
contradiction with the fact that it has positive measure.
\end{proof}
\bigskip
\bigskip
\noindent
{\bf Acknowledgements.}
The authors worked on this problem while they were visiting the Institute for Mathematical Sciences, National University of Singapore, in 2014. V. Becher is a member of
Laboratoire International Associ\'e INFINIS, Universit\'e Paris Diderot CNRS - Universidad de Buenos
Aires-CONICET. Slaman's research was partially supported by the National Science Foundation under grant number DMS-1301659.
|
\section{'normal' and 'anomalous' averages }
Consider an average of the form
\begin{eqnarray}\label{1}
\bar{s}=\sum_{n=1}^N s_n P_n,\quad \sum_n P_n=1
\end{eqnarray}
where $s_1>s_2...>s_N$. We will call $\bar{s}$ {\it normal} if it lies between $s_1$ and $s_N$,
$s_1\ge \bar{s} \ge s_N$, and {\it anomalous} otherwise. It is readily seen that
$\bar{s}$ is always normal if $P_n\ge 0$, $n=1,2,..N$, and to be anomalous it requires that
at least one of the $P$'s is negative. To see how an anomalous average may be produced, consider
$N=2$, $s_{1,2}=\pm1$, $P_1=1001$, $P_2=-1000$, $P_1+P_2=1$, and find that
$\bar{s}=2001$. Thus, a large anomalous value would occur where the moduli of $P_n$ are large,
but the sum of all $P_n$ is unity due to a very precise cancellation.
Multiplying the $P_n$'s by $s_n$ destroys the cancellation, so that the resulting $\bar{s}$ is unduly large.
A more detailed example is given in \cite{ANN}, where a similar effect is found responsible for what appears
as super-luminal transmission of a wave packet.
So where else is one likely to encounter anomalous mean values?
\newline
\section{Quantum weak values}
The place to look is in quantum mechanics.
Consider a two level quantum system (a spin $1/2$) with a hamiltonian $H$, pre- and post-selected (observed) in some states
$\psi$ and $\phi$ at $t=0$ and $t=T$, respectively. Choose an operator $S$ with eigenstates $|s_{1,2}{\rangle}$ and eigenvalues
$\pm1$. Inserting, at some $0<t'<T$ the unity $I=|s_{1}{\rangle}{\langle} s_1|+|s_{2}{\rangle}{\langle} s_2|$ into the transition amplitude
${\langle} \phi|\exp(-iHT)|\psi{\rangle}$, shows that the spin can reach the final state via two virtual
routes, $\psi\to s_1 \to \phi$ and $\psi \to s_2 \to \phi$. Putting for simplicity $H=0$, for the two corresponding amplitudes we
have
\begin{eqnarray}\label{1a}
A_{1,2}={\langle} \phi |s_{1,2}{\rangle}{\langle} s_{1,2}|\psi{\rangle}.
\end{eqnarray}
To see what actually happens at $t'$
we may employ a von Neumann pointer with position $f$,
initially
decoupled from the spin.
We set the pointer at some $f'$ by preparing in in a state $|M_{f'}{\rangle}=\int df G(f-f') |f{\rangle}$, where $G(f)$ is a function
peaked around $0$ with a width $\Delta f$, such that $\int|G(f)|^2=1$.
For $t' <t<t+\tau$ the pointer briefly interacts with the spin via
$H_{int}=-i\tau^{-1}\partial_f A$, and then its final position is measured exactly.
Now $\Delta f $ determines what we know about the initial position of the pointer and, therefore, the accuracy of the measurement.
For $\Delta f<<1$, the accuracy is good, and the interference between two routes is destroyed. With $f'=0$ he pointer's readings are narrowly grouped around the values $f=\pm 1$
which occur with the probabilities $P_{1,2}=|A_{1,2}|^2/(|A_{1}|^2+|A_{2}|^2)$, respectively.
Thus, the average of $S$ at $t'$ is given by
\begin{eqnarray}\label{2}
\bar{s}_{acc}(\phi|\psi)=(|A_{1}|^2-|A_{2}|^2)/(|A_{1}|^2+|A_{2}|^2),
\end{eqnarray}
With $P_{1,2}\ge 0$ it is always normal, and lies between $-1$ and $1$.
\newline
Our measurement can be made less accurate it two different ways. A {\it classical} uncertainty
is added when we still have an accurate meter which destroys interference between the two
routes, but cannot, for some reason, set it exactly to zero. With its initial
position being $f'$ with a probability $W(f')=W(-f')$, peaked around $0$ with a width $\delta f'$, the initial meter state is a mixture,
\begin{eqnarray}\label{3}
\rho_M=\int|M_{f'}{\rangle} W(f'){\langle}|M_{f'}|df', \quad \int W(f')df'=1.
\end{eqnarray}
If $\delta f'$ is large, the final pointer's readings are widely spread, and we must associate a reading of $100$ not with
the actual value of the spin's component, but rather with the shortcomings of our measurement.
Note that after averaging over many trials we still find that the mean pointer reading is given by Eq.(\ref{2}).
In other words, as far as the averages are concerned, we still have the same information,
it is just harder to obtain.
\newline
The other possibility is more interesting.
A {\it quantum} uncertainty can be introduced by preparing the meter in a pure state (\ref{3}) and selecting $\Delta f\to \infty$.
We still do not know the pointer's initial position, but in a quite different sense: there is no probability for it to be $f'$.
Now the pointer's readings are spread over a range $\sim \Delta f$, and it still seems unwise to associate the
value of $100$ with an intrinsic property of the spin.
The central result of \cite{AHAR} (see also \cite{ME}) is that
the mean pointer reading is given not by (\ref{2}), but by the real part of an expression similar to (\ref{2}), with $|A_n|^2$
replaced by the amplitudes themselves
\begin{eqnarray}\label{4}
\bar{s}_{weak}(\phi|\psi)=Re\{(A_{1}-A_{2})/(A_{1}+A_{2})\}=Re \frac{{\langle} \phi|A|\psi{\rangle}}{{\langle} \phi|\psi{\rangle}}.\quad
\end{eqnarray}
Note that since the final meter's readings $f$ are distributed between $-\infty$ and $\infty$ with non-negative
probability density $p(f)=\ge 0$, there is nothing strange about $\bar{s}_{weak}$ taking any real value, however large.
It can, alternatively, be seen as an average of the variable $s$ numbering our two routes, and taking the values $+1$ and
$-1$, $\bar{s}_{weak}(\phi|\psi)=\sum_{n=1,2}s_n P_n$, with $P_{1,2}=Re \{A_{1,2}/(A_1+A_2)\}$.
Since $A_1$ and $A_2$ are arbitrary complex quantities, $P_{1,2}$ may or not be non-negative, and
the average $\bar{s}_{weak}$ may be both normal or anomalous. For example, it is always normal
for $\phi=\psi$ and $A_{1,2}=|{\langle} s_{1,2}|\psi{\rangle}|^2\ge 0$, or anomalous, e.g., for for $\phi=\psi$, or $100$ for $ReA_{1,2}=0$, $A_2/A_1=-99/101$.
Thus, initial meter's position is uncertain in the quantum sense, $\bar{s}_{weak}(\phi|\psi)$ no longer gives one an indication that we started with
exactly two routes labelled $1$ and $-1$. The question is now of of interpreting this result, and here we disagree with the authors of \cite{AHAR}.
\section{Anomalous values and the uncertainty principle}
The answer goes back to one of basic postulates of quantum mechanics. A very inaccurate 'weak' measurement
does not destroy interference between the two routes \cite{ME}. Accordingly, there are only the amplitudes, and not the probabilities, assigned to the two
virtual routes, and used in Eq.(\ref{4}).
Further, the Uncertainty Principle states that two interfering routes cannot be told apart and should
be considered a single pathway \cite{Feyn}. If so, the question {\it "what was, on average, the value of $S$ if we hadn't destroyed coherence between the two routes?"} should not have meaningful answer \cite{ME}. Although a mean value of $100$ for a spin 1/2 looks like a wrong result produced by a malfunctioning meter, yet it cannot simply be 'corrected', since the 'correct' answer just doesn't exist. This is a purely quantum dilemma: do we ascribe the same degree of importance and 'reality' to the results of the accurate and the inaccurate 'weak' measurements, just because weak measurements can and have been made.
It is worth arguing in favor of accurate 'strong' measurements.
For any choice of $\psi$ and $\phi$, an accurate pointer would only yield the values $1$ and $-1$. This is how one knows that spin of $1/2$ is an intrinsic property of the electron,
and is later able to write its wave function as a Pauli spinor in situations which has nothing to do with the original Stern-Gerlach experiment.
One who only has access to the averages (\ref{2}), but not the distributions, may at least note that for any $\psi$ and $\phi$ $\bar{s}_{strong}(\phi|\psi)$ lies between $-1$ and $1$.
This is still a property of the electron, and not just of the chosen transition. But from $\bar{s}_{weak}(\phi|\psi)$ one can only deduce that, depending on $\psi$ and $\phi$,
the expectation value of $S$ may take an arbitrary value. Such value is a property of the particular transition, and does not have a more general meaning.
The answer to a question which should not have an answer is, in this case, {\it "anything at all"}.
\section{Where did Ferrie and Combes go wrong?}
A classical theory operates with strictly non-negative probabilities which are ascribed to all
observable scenarios from the very beginning.
Consequently, all
observable averages must be of the normal type, and lie between the smallest and the largest values the variable
of interest may take.
The authors of \cite{PRL} claim to have demonstrated the contrary using a simple classical model, but their analysis contains an error.
The classical model in \cite{PRL} has two initial states $\psi$, two final states $\phi$, and two intermediate states $s$, all labelled by $\pm1$.
With the choice $\psi=1$ and $\phi=-1$ there are two pathways $1\to \pm 1 \to -1$ travelled with the probabilities [cf. Eq.(27) of Ref.\cite{PRL}]
\begin{eqnarray}\label{5}
P_{1,2}=(1\pm\lambda-\delta)/2(1-\delta).
\end{eqnarray}
Here $0<\lambda<1$ is a small parameter, and $\delta$ is chosen so that $P_{1,2}$ are strictly positive,
$0<\delta<1-\lambda$. With this choice the authors find that
\begin{eqnarray}\label{6}
\overline{s/\lambda}=\sum_{n=1,2}(-1)^nP_n/\lambda=1/1-\delta
\end{eqnarray}
can be made arbitrarily small, provided $\lambda$ is small, and $\delta$ is close to unity.
This, claim the authors of Ref.\cite{PRL}, is an anomalous weak value obtained in a purely
classical context, and this is where they are wrong. The mean in (\ref{6}) is the mean not of the $s$ taking the values of $\pm1$, but of the
variable $s/\lambda=\pm1/\lambda$, and $1 <\overline{s/\lambda} < 1/\lambda$ is a perfectly normal average,
well within the allowed interval $[-1/\lambda,1/\lambda]$. The same criticism applies to the 'classical protocol' illustrated in Fig.1, and described below the Eq.(34),
since it relies on Eqs.(\ref{5}) and (\ref{6})
\newline
To put it plain, with $\lambda =1/200$ the authors of \cite{PRL} tell Alice-at-the-end-of-the-line to write down
$200$ if he/she receives the coin showing up heads, or $-200$ otherwise.
Adding the numbers and dividing by the number of trials may now yield a value of $100$,
which is then shown to the reader as the proof that there is 'a simple classical model which exhibits anomalous weak values' \cite{PRL}. What the authors of \cite{PRL} should have been saying, in answering their own question, is:
'It can, if you call 'heads' - '200 heads', and 'tails' - '200 tails'. But this observation doesn't add much to our understanding of weak measurements or weak values .
\section{Conclusions}
In summary, appearance of anomalous weak values is a strictly quantum phenomenon.
Such values arise in a situation where the Uncertainty Principle forbids gaining required
information about a quantum system under observation.
Since for an average to be anomalous the 'probabilities' in Eq.(\ref{1}) must change sign,
anomalous weak values cannot arise for any observable quantity in
classical statistics.
|
\section{Introduction}
Apart from other methods, our current understanding of QCD in the
nonperturbative regime is strongly based
on lattice gauge theory and
effective models \cite{Leupold:2011zz}.
These complementary approaches
are compared to each other
for reasons
of crosscheck and systematic improvement \cite{Pawlowski:2010ht,Herbst:2013ufa}.
Despite all efforts the order of the chiral phase
transition of QCD with two
massless flavors has not been rigorously determined yet, and
the interest in a reliable prediction
remains strong.
The case of two massless (or light, respectively) flavors at vanishing baryonic chemical potential
is of particular interest for lattice studies due to
the comprehensive predictions of effective models \cite{Cossu:2013uua,Bhattacharya:2014ara}.
The possible existence of a second-order chiral phase transition, as well
as the corresponding universality class, can be investigated from
the effective theory for the chiral
condensate \cite{Pisarski:1983ms,Berges:1997eu,Berges:1998sd,Butti:2003nu,Calabrese:2004uk,Braun:2010vd,Fukushima:2010ji,Grahl:2013pba,Pelissetto:2013hqa,Aoki:2013zfa,Meggiolaro:2013swa,Nakayama:2014sba,Fejos:2014qga}.
Using the $[\bar{2},2]+[2,\bar{2}]$
representation of $SU(2)_L \times SU(2)_R$ \cite{Paterson:1980fc},
we can take into account the scalar mesons ($\sigma$ and $\vec{a}_0$)
as well as the pseudoscalar mesons ($\eta$ and $\vec{\pi}$) by
writing down the most general Lagrangian invariant under chiral symmetry.
For the full symmetry, $U(2)_A \times U(2)_V \simeq U(1)_A \times U(1)_V \times
\left[ SU(2)/Z(2) \right]_L \times \left[ SU(2)/Z(2) \right]_R$,
taking into account all linearly independent invariants up to eighth
polynomial order in the fields,
this Lagrangian can be written as \cite{Pisarski:1983ms,Berges:1997eu,GrahlDiss,Patkos:2012ex}
\begin{gather}
\mathscr{L} = \frac{Z}{2} {\rm Tr} (\partial_{\mu} \Phi^{\dagger} )
(\partial_{\mu} \Phi) +r {\rm Tr} \Phi^{\dagger} \Phi
+ g_1 ({\rm Tr} \Phi^{\dagger} \Phi)^2 + g_2 \xi + g_3 ({\rm Tr} \Phi^{\dagger} \Phi)^3 \nonumber\\
+g_4 ({\rm Tr} \Phi^{\dagger} \Phi) \xi+ g_5 {\rm Tr} ( \Phi^{\dagger} \Phi)^4
+g_6 ({\rm Tr} \Phi^{\dagger} \Phi)^4 +g_7 ({\rm Tr} \Phi^{\dagger} \Phi)^2 {\rm Tr} ( \Phi^{\dagger} \Phi)^2 \; , \label{lsp}
\end{gather}
where
$ \Phi = \left(\sigma + i \eta \right) t_0 + \vec{t} \cdot \left(\vec{a}_0
+ i \vec{\pi} \right)$, with $t_a$ denoting the generators of $U(2)$ normalized such that ${\rm Tr}(t_a t_b) \equiv 1$ \cite{Grahl:2013pba}.
Furthermore,
\begin{gather*}
{\rm Tr} \Phi^{\dagger} \Phi = \sum_{i} \phi_i^2 \equiv 2 \rho\;, \ \phi_i \equiv \sigma,\vec{\pi},\eta,\vec{a}_0 \;, \\
\frac{1}{2}{\rm Tr} ( \Phi^{\dagger} \Phi)^2 - \rho^2 = \left(\sigma^2 +\vec{\pi}^2 \right) \left(\eta^2 +
\vec{a}_0^2 \right) - \left(\sigma \eta - \vec{\pi} \cdot \vec{a}_0 \right)^2 \equiv \xi\;.
\end{gather*}
\noindent
We omit derivate couplings since we will only discuss the local-potential
approximation (LPA, $Z=1$) and, respectively, its minimal extension allowing for a
field-independent wave-function renormalization factor $Z$ (LPA').
We note that the invariants $({\rm Tr} \Phi^{\dagger} \Phi) {\rm Tr} ( \Phi^{\dagger} \Phi)^3$,
$\left({\rm Tr} ( \Phi^{\dagger} \Phi)^2 \right)^2$, and ${\rm Tr} ( \Phi^{\dagger} \Phi)^3$
do not yield further linearly independent contributions
to Eq.\ (\ref{lsp}). \\
In this paper we focus on the case where the axial $U(1)_A$ symmetry, which is anomalously broken
at vanishing temperature, has already been restored at the critical temperature $T_c$.
Therefore we do not take account of $U(1)_A$-breaking terms in Eq.\ (\ref{lsp}).
For studies concerning the opposite scenario in which the anomaly remains present at $T_c$ we
refer to Refs.\ \cite{Grahl:2013pba,Pawlowski:1996ch,Fischer:2011pk,Pelissetto:2013hqa,Schaefer:2013isa,Mitter:2013fxa,Aoki:2013zfa,Meggiolaro:2013swa}.
The long-standing question which of the both scenarios is actually realized
is subject to an ongoing debate. The latest lattice results are quite controversial: whereas
the case of restored anomaly is advocated by Refs.\ \cite{Aoki:2012yj,Cossu:2013uua}, the opposite
scenario is favored by Refs.\ \cite{Sharma:2013nva,Bhattacharya:2014ara}.
The predictions of effective theories for the chiral condensate are summarized in the following. \\
The existence of an infrared-stable (IR-stable) fixed point in the RG flow
of the effective theory for the order parameter is a necessary
condition for a second-order phase transition to occur.
If this scenario is realized or not depends on the initial values
for the parameters in the ultraviolet (UV) limit determined by
the underlying microscopic theory. Therefore, the RG analysis
serves to either rule out the existence of a second-order phase transition
or to confirm its possible existence. \\
If the anomaly strength exceeds the cut-off scale, a phase transition
of second order in the $O(4)$ universality class is
predicted \cite{Pisarski:1983ms,Jungnickel:1995fp,Grahl:2013pba}.
The case of small anomaly strength is subtle. The anomaly yields two independent quadratic
mass terms. In Landau theory, i.e., at mean-field
level, it is evident that such a situation corresponds to a multicritical point with at least two relevant
scaling variables. This is used as an argument in Ref.\ \cite{Pelissetto:2013hqa} to rule out
a second-order phase transition with temperature being the only relevant scaling variable.
However, in consistence with Refs.\ \cite{Grahl:2013pba,GrahlDiss}, we argue that the inclusion of
fluctuations can, in principle, lead to a IR-stable fixed point corresponding to exactly such a scenario.
Although associated with unphysical masses in the approximation considered, there in fact exists
an (unphysical) $SU(2)_A \times U(2)_V$-symmetric, IR-stable fixed point
exemplifying our consideration.
This observation extends the critical reinvestigation of the
standard criterions used for ruling out continuous transitions presented in Ref.\ \cite{Przy}. The
latter particularly points out that the irreducibility of a representation is not strictly ruling
out a second-order phase transition associated with a single relevant scaling variable.
In the absence of the anomaly there is strong evidence
from Refs.\ \cite{Pelissetto:2013hqa,Nakayama:2014sba} for the
existence of a second-order phase transition belonging to the
$U(2)_V \times U(2)_A$ universality class. The existence and properties of the corresponding
fixed point will be discussed in the remainder of this paper. \\
Ref.\ \cite{Pelissetto:2013hqa} uses a resummed loop expansion at fixed spatial dimension, $D=3$, based on
the $\overline{MS}$ and the \textit{MZM} scheme, respectively. The discovered IR-stable, $U(2)_V \times U(2)_A$-symmetric
fixed point corresponds to an anomalous dimension of $\eta \sim 0.12$. Previous studies in the
framework of the $\epsilon$-expansion ($\epsilon = 4-D$) failed to find the fixed point \cite{Pisarski:1983ms,Calabrese:2004uk}.
It is an important question why this is the case.
A plausible explanation is given in Ref.\ \cite{Pelissetto:2013hqa}: the
fixed point only exists near $D=3$. One might wonder, however, if the
resummation scheme and the loop-order also play a role.
With our FRG investigation presented in Secs.\ \ref{frgres}--\ref{staban} we demonstrate that
the existence not only depends on the fixed spatial dimension, but also on the
way how nonperturbative corrections are included. \\
Due to the converging correlation length at a second-order phase transition we can
work in the dimensionally reduced theory \cite{DimRed2ndorder}.
\section{Fixed points from FRG}
\label{frgres}
Assuming an homogeneous condensate, and using the Litim regulator,
the Wetterich equation for the potential of the truncation (\ref{lsp})
is given by
\begin{gather}
\frac{\partial U_k}{\partial k} = \frac{2\pi^{D/2} k^{D+1} Z_k}{D \ \Gamma(D/2) (2\pi)^D} \left(1- \frac{\eta}{2+D}\right) \sum_{i}
\frac{1}{Z_k k^2 + M_i^2} \; , \label{fleq}
\end{gather}
\noindent
where $\mathscr{L}_{k} = \frac{1}{2} Z_k Tr (\partial_{\mu} \Phi^{\dagger}) (\partial_{\mu} \Phi)+U_k$,
with $\mathscr{L}_{k=\Lambda}=\mathscr{L}$ defining the bare Lagrangian in the UV limit.
$M_i^2$ denote the eigenvalues of the mass matrix
\begin{gather}
M_{ij} \equiv \frac{\partial^2 U_k}{\partial \phi_i \partial \phi_j}
\; , \; \; i,j=1,\ldots,8 \; . \label{mm}
\end{gather}
\noindent
The anomalous dimension, $\eta$, is determined from the relation
\begin{gather}
\eta_k = - Z_k^{-1} k \frac{\partial Z_k}{\partial k} \; , \; \lim_{k \to 0} \eta_k = \eta\;.
\end{gather}
The flow equation for $Z_k$ is derived from the second derivative of the effective action with
respect to the fields and evaluated at the global minimum of the potential \cite{Kopietz:FRG}.
For our purposes we can restrict our discussion of the LPA' to the truncation
$U_k (\rho, \xi) \equiv V(\rho) + W(\rho) \xi$, which is suited up to sextic
truncation order ($g_5 = g_6 = g_7 = 0$).
Setting $D=3$, in agreement with Ref.\ \cite{Berges:2000ew} we obtain
\begin{gather}
\eta_k = \frac{2}{3 \pi^2 [1+\bar{V}_k^{'}(\bar{\rho}_{0,k})]^2 } \left( \frac{4 \bar{\rho}_{0,k}
\bar{W}_k(\bar{\rho}_{0,k})^2}{ [1+ 4 \bar{W}_k(\bar{\rho}_{0,k}) \bar{\rho}_{0,k} + \bar{V}_k^{'}(\bar{\rho}_{0,k})]^2 }
+ \frac{ \bar{\rho}_{0,k} \bar{V}_k^{''}(\bar{\rho}_{0,k})^2 }{
[1+ \bar{V}^{'}(\bar{\rho}_{0,k}) +2\bar{\rho}_{0,k} \bar{V}_k^{''}(\bar{\rho}_{0,k})]^2 } \right) \; ,
\end{gather}
\noindent
where we introduced rescaled variables (labeled by a bar),
\begin{gather*}
\bar{U} = k^{-D} U \; , \; \bar{\rho}= Z k^{2-D} \rho \; , \; \bar{\xi} = Z^2 k^{4-2D} \xi
\; , \; \bar{V} = k^{-D} V \; , \; \bar{W} = Z^{-2} k^{D-4} W \; ,
\end{gather*}
\noindent
and denoted the global minimum of $U_k$ by $\rho_0$ (assuming $\xi_0 = 0$). \\
The flow equations for the rescaled parameters of Eq.\ (\ref{lsp}) are
derived similar to Refs.\ \cite{Fukushima:2010ji,Grahl:2013pba},
not listed explicitly here.
The numerically determined fixed points for sextic truncation order are listed in Table \ref{tabfp6}, those
for octic truncation order in Table \ref{tabfp8}. We proceed with a detailed analysis of their stability
properties and the resultant implications in Sec.\ \ref{staban}.
\begin{table}
\caption{\label{tabfp6}Fixed points in sextic truncation order (for the LPA denoted by $F_i^{(6)}$, for the
LPA' by $F_i'^{(6)}$). $D=3$.}
\begin{ruledtabular}
\begin{tabular}{l|l l l l l | l }
$F$ & $\bar{r}$ & $\bar{g}_1$ & $\bar{g}_2$ & $\bar{g}_3$ & $\bar{g}_4$ & $\eta$ \hspace{1cm} \\ \hline
$F_0^{(6)}$, $F_0'^{(6)}$ & 0 & 0 & 0 & 0 & 0 & 0 \\
$F_1^{(6)}$ & -0.1316 & 0.0827 & 0.8586 & 0.2091 & 0.2161 & 0 \\
$F_1'^{(6)}$ & -0.1251 & 0.0795 & 0.8447 & 0.1981 & 0.1876 & 0.0334 \\
$F_2^{(6)}$ & -0.103 & 0.3334 & -0.9411 & 0.307 & -0.7154 & 0 \\
$F_2'^{(6)}$ & -0.0938 & 0.3151 & -0.8981 & 0.2634 & -0.6024 & 0.0529 \\
$F_3^{(6)}$ & -0.1355 & 0.2132 & 0 & 0.1285 & 0 & 0 \\
$F_3'^{(6)}$ & -0.1317 & 0.2103 & 0 & 0.123 & 0 & 0.0195 \\
\end{tabular}
\end{ruledtabular}
\begin{ruledtabular}
\begin{tabular}{l|c | c }
$F$ & stability-matrix eigenvalues & nonzero $\bar{M}_i^2$ \\ \hline
$F_0^{(6)}$, $F_0'^{(6)}$ & \{-2,-1,-1,0,0\} & -- \\
$F_1^{(6)}$ & \{15.6603,0.6245+3.5342 i,0.6245-3.5342 i,1.6306,-1.3743\} & \{0.8246,0.6434,0.6434,0.6434\} \\
$F_1'^{(6)}$ & \{14.5059,0.5839+3.2722 i,0.5839-3.2722 i,1.5485,-1.3614\} & \{0.7823,0.6261,0.6261,0.6261\} \\
$F_2^{(6)}$ & \{13.2219,1.1882+2.1481 i,1.1882-2.1481 i,-1.5108,1.3732\} & \{0.4750,-0.2707,-0.2707,-0.2707\} \\
$F_2'^{(6)}$ & \{11.6716,1.0622+1.874 i,1.0622-1.874 i,-1.5279,1.3464\} & \{0.4272,-0.2502,-0.2502,-0.2502\} \\
$F_3^{(6)}$ & \{12.9247,8.125,1.5092,-1.3798,-0.5034\} & \{0.6442\} \\
$F_3'^{(6)}$ & \{12.3598,7.7931,1.4673,-1.3745,-0.4802\} & \{0.6233\} \\
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{table}
\caption{\label{tabfp8}Fixed points for the LPA in octic truncation order. $D=3$. }
\begin{ruledtabular}
\begin{tabular}{l|l l l l l l l l }
$F$ & $\bar{r}$ & $\bar{g}_1$ & $\bar{g}_2$ & $\bar{g}_3$ & $\bar{g}_4$ & $\bar{g}_5$ & $\bar{g}_6$ & $\bar{g}_7$ \\ \hline
$F_0^{(8)}$ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
$F_1^{(8)}$ & -0.0153 & 0.0274 & 0.1007 & -0.0020 & -0.1529 & -0.0432 & -0.0143 & 0.0321 \\
$F_2^{(8)}$ & -0.0141 & 0.0567 & -0.1151 & -0.0485 & 0.1676 & -0.0472 & -0.0997 & 0.1429 \\
$F_3^{(8)}$ & -0.0148 & 0.0414 & 0 & -0.0258 & 0 & 0 & -0.0118 & 0 \\
$F_4^{(8)}$ & -0.1721 & 0.2192 & 0 & 0.1828 & 0 & 0 & 0.1006 & 0 \\
\end{tabular}
\end{ruledtabular}
\begin{ruledtabular}
\begin{tabular}{l|c }
$F$ & stability-matrix eigenvalues \\ \hline
$F_0^{(8)}$ & \{-2, \ -1, \ -1, \ 1, \ 1, \ 1, \ 0, \ 0\} \\
$F_1^{(8)}$ & \hspace{2cm} \{4.1374, \ 2.1731, \ -2.0064, \ 1.3755, \ -1.1678, \ -0.9261, \ 0.2814, \ 0.1298\} \hspace{2cm} \\
$F_2^{(8)}$ & \{4.2123, \ 2.3531, \ -2.0072, \ 1.5805, \ -1.1886, \ -0.9334, \ 0.2274, \ 0.1216\} \\
$F_3^{(8)}$ & \{3.6611, \ 2.8933, \ -2.0029, \ 1.685, \ -1.1024, \ -1.0081, \ 0.1351, \ -0.0505\} \\
$F_4^{(8)}$ & \{34.5986, \ 26.9475, \ 12.7525, \ 9.3877, \ 5.1825, \ 1.3058, \ -1.1215, \ -0.65\}
\end{tabular}
\end{ruledtabular}
\end{table}
\section{Stability analysis}
\label{staban}
In order to determine the stability properties of the fixed points one can analyze the flow
in their neighborhood where it is governed by the linearized system.
For this purpose one calculates the eigenvalues of the
stability matrix
\begin{gather}
\label{stm}
(S_{ij}) \equiv \left( \frac{\partial \beta_i}{\partial \bar{p}_j}
\right) \Bigl\vert_{\bar{p}=\bar{p}*} \ ,
\end{gather}
\noindent
where we denote the $n$ rescaled parameters of the Lagrangian by
$\bar{p} = \{\bar{p}_i \}$, a fixed point by $\{ \bar{p}_i^* \}$, and
the beta functions are given by
$\beta_i(\bar{p}) \equiv k \partial_k \bar{p}_i$.
In general one obtains $n_s$ eigenvalues with positive real part,
$n_u$ with negative real part, and $n_m$ with vanishing real part.
The corresponding eigenvectors give rise to invariant subspaces of the
parameter space inside which the flow stays if one starts within them \cite{regev2006chaos}.
In case of distinct eigenvalues there is a $n_s$-dimensional
invariant subspace (called critical manifold) inside which the flow
is attracted towards the fixed point in the infrared limit $k=0$.
Respectively, there exists a $n_u$-dimensional invariant subspace (called
unstable manifold) inside which the flow is repelled, and a $n_m$-dimensional
invariant subspace (called marginal manifold) inside which the flow has no
direction at all.
Here we note that complex valued eigenvalues always appear as conjugate pairs.
Referring to the real and imaginary parts of the associated complex eigenvectors as
eigenvectors, too, the critical manifold is spanned by $n_s$ eigenvectors, the
unstable manifold is spanned by $n_u$ eigenvectors, and the marginal manifold is
spanned by $n_m$ eigenvectors.
Therefore, if $n_m=0$, one can reach the critical
manifold by tuning $n_u$ parameters starting anywhere in parameter space.
Hence, a second-order phase transition with
respect to a single scaling variable (temperature) can only exist
if we have exactly $n_u=1$. In this case we speak of an IR-stable fixed point. \\
The stability-matrix eigenvalues are listed for each fixed point in
Table \ref{tabfp6}--\ref{tabfp8}.
We begin with discussing the LPA in sextic truncation order.
$F_1^{(6)}$ and $F_2^{(6)}$ are different, IR-stable,
$U(2)_A \times U(2)_V$-symmetric spiral fixed points. $F_1^{(6)}$ is associated
with physical mass-matrix eigenvalues whereas $F_2^{(6)}$ is not.
Their existence is highly nontrivial since they do not exist at quartic
truncation order, neither in the LPA \cite{Fukushima:2010ji,Fejos:2014qga}, nor in the LPA' \cite{Berges:2000ew}.
The critical exponent, $\nu \sim 1/1.3614 \sim 0.7345$, associated with
$F_1'^{(6)}$ is in unexpectedly good agreement with the values reported in Ref.\ \cite{Pelissetto:2013hqa}
($\nu \sim 0.71$ for the MZM scheme, $\nu \sim 0.76$ for the $\overline{MS}$ scheme). This
agreement is most likely accidental. The value for the anomalous dimension is actually significantly
smaller ($\eta \sim 0.0334$ compared to $\eta \sim 0.1$).
$F_3^{(6)}$ is an unstable $O(8)$-symmetric fixed point. All fixed points
are also present in the LPA' without qualitative changes ($F_i^{(6)}$ corresponds
to $F_i'^{(6)}$). \\
Of particular interest to us are the marginal eigenvalues encountered for the
Gaussian fixed points, $F_0$, which will be discussed next.
From a merely mathematical
standpoint one can decide whether marginal eigenvalues are relevant
or not by going beyond the linear order utilized in Eq.\ (\ref{stm}). For this purpose
one can either use the second derivatives of the beta functions, or one has to perform
a more general Lyapunov analysis. However, this is not meaningful in our case
because in presence of
marginal eigenvalues one has to consider a change in the
fixed-point structure at higher polynomial truncation order.
In general, such a change cannot be excluded by a nonlinear
stability analysis at lower order.
The occurrence of the marginal eigenvalues, however, can be explained as follows.
In general the beta functions for a rescaled mass parameter $\bar{m}^2$,
a rescaled quartic coupling
$\bar{g}_4$, and a rescaled sextic coupling $\bar{g}_6$, respectively, are given by
\begin{gather}
\beta_{m^2} = (-2 + \eta) \bar{m}^2 + f_{2}(\bar{p}) \; , \;
\beta_4 = (D-4+2 \eta) \bar{g}_4 + f_{4}(\bar{p}) \; , \;
\beta_6 = (2D-6+3 \eta) \bar{g}_6 + f_{6} (\bar{p}) \; ,
\end{gather}
\noindent
where the $f_i (\bar{p})$ denote nonlinear functions of the rescaled parameters.
In FRG the polynomial order of these functions depends on the truncation order
of the effective action, whereas in RG approaches based on a loop expansion
it depends on the loop order. Since these functions as well as the anomalous
dimension, $\eta$, vanish at the Gaussian fixed point, we can conclude that (for $D=3$)
$\bar{m}^2$ and $\bar{g}_4$ are relevant parameters with respect to this fixed point.
They yield stability matrix eigenvalues $-2$ and $-1$, respectively.
Similarly, the sextic coupling contributes a vanishing eigenvalue at the
Gaussian fixed point, and higher order couplings yield positive eigenvalues.
We conclude that the marginal eigenvalues in
Table \ref{tabfp6}--\ref{tabfp8} do not render the stability
analysis inconclusive. However, in the remainder of this section, we will
argue why the LPA' remains inconclusive, pointing out general differences between
FRG and other RG approaches first.
For a more fundamental comparison between both approaches we refer
to Refs.\ \cite{Litim:2002xm,Codello:2013bra}. \\
In the framework of the $\epsilon$-expansion or other loop expansions
at fixed spatial dimension $D$, one usually argues
that also in case
of non-Gaussian fixed points the canonical scaling dimension determines
if a coupling can affect stability \cite{herbut2007modern}.
Accordingly, depending on the sign of their
canonical scaling dimension, one speaks of relevant, marginal, and irrelevant
parameters. Obviously, especially marginal eigenvalues are sensitive to the
loop order. Therefore, one has to consider the possibility that higher-order loop corrections
change the marginal eigenvalue into a nonvanishing one.
It is important to note that if a marginal eigenvalue
for a certain fixed point turns nonzero
at higher order, this can also change the stability properties of
the other fixed points. This is for example the case in the $O(N=4)$ model
with di-icosahedral anisotropy. The $\epsilon$-expansion
of this model has been derived in Ref.\ \cite{Toledano:1985},
pointing out that the case of $N=4$ is special. In the presence of
an anisotropy, the $O(4)$-symmetric fixed point acquires a marginal eigenvalue
at one-loop order in the $\epsilon$-expansion whereas the anisotropic fixed
point is IR-unstable. At two-loop order, however, the anisotropic fixed point
can become the IR-stable one. We reinvestigated the situation using
the FRG in LPA and
found that the anisotropic fixed point also becomes IR-stable when going
beyond the quartic truncation order \cite{GrahlDiss}. \\
However, a change of stability can
occur even in the absence of any marginal eigenvalues. A famous example is the
$O(N)$ model with cubic anisotropy for $D=3$ \cite{Varnashev:1999ze,Pelissetto:2000ek}. The model exhibits
an $O(N)$-symmetric (isotropic) fixed point as well as a cubic fixed point.
For $N > N_c$ the cubic fixed point is the IR-stable one, the isotropic
fixed point being IR-unstable, and vice versa for $N < N_c$.
The value for $N_c$ depends on the loop order as well as on the resummation scheme and
is still under debate. \\
In comparison to loop expansions, the stability matrix eigenvalues
are much more sensitive to the polynomial truncation
order in the FRG formalism.
Using FRG, the accuracy of the critical exponents heavily depends
on irrelevant couplings \cite{Litim:2002cf}.
This is explained by
the fact that fluctuations are taken into account differently in both approaches.
Irrelevant couplings can be safely ignored in the loop expansion and
nonperturbative effects are captured by using resummation.
In contrast, if we were able to solve the FRG equation without
truncating the effective action, we would obtain exact results.
In the LPA
at quartic truncation order, however,
one generically reproduces the one-loop epsilon-expansion
results when setting the mass parameter to zero \cite{Kopietz:FRG,Fukushima:2010ji}. \\
Our conclusions are as follows.
Naively, one would trust the utilized approximation scheme
since no marginal eigenvalues appear for the non-Gaussian fixed points. However, we argued that
even in this case the fixed-point structure can change at higher truncation order.
Especially the
presence of the unphysical fixed point advises caution. In fact, the spiral fixed points become
unstable fixed points ($F_1^{(8)}$ and $F_2^{(8)}$, respectively) at
octic truncation order (see Table \ref{tabfp8}). Interestingly,
at this order one finds two unstable $O(8)$-symmetric fixed points.
Going to any higher (finite) polynomial order in the LPA' will not clarify the situation.
If an IR-stable fixed point were found at higher order, one could not rule out its
disappearance beyond that order. And in the opposite case the
discrepancy with Ref.\ \cite{Pelissetto:2013hqa} would require to go beyond the LPA' as well.
Therefore it is necessary to include derivative couplings in order to decide whether the
$U(2)_A \times U(2)_V$-symmetric fixed point is stable or not. In addition, novel criteria to
assess the conclusiveness of truncation schemes need to be developed.
\section{Conclusions}
\label{conclusions}
We further investigated the possibility that the two-flavor chiral phase
transition can be of second order in the absence of the
axial anomaly, using the FRG method
in the LPA as well as in the LPA'. \\
We found two IR-stable, $U(2)_A \times U(2)_V$-symmetric fixed points
at sextic polynomial truncation order, one of them
associated with unphysical masses.
The value for the critical exponent, $\nu \sim 0.7345$, calculated
for the one associated with physical masses is in (most likely accidental)
agreement with the result reported in Ref.\ \cite{Pelissetto:2013hqa}.
At higher polynomial order both fixed
points become unstable.
Nevertheless, the results of our research provide further evidence for the existence
of the IR-stable, $U(2)_A \times U(2)_V$-symmetric fixed point from
an independent perspective. \\
The fact that an $U(2)_A \times U(2)_V$-symmetric
fixed point appears by simply including
sextic invariants demonstrates that its existence not only
depends on the spatial dimension but also on the
way nonperturbative corrections are taken into account. In the
framework of a resummed perturbative expansion this concerns the
resummation scheme and the perturbative order. \\
Our main conclusion is that the LPA' is not capable to unambiguously
clarify the stability of the fixed points.
Since the fixed-point structure of the dimensionally
reduced theory controls the behavior near $T_c$, previous
finite-temperature studies \cite{Berges:2000ew,GrahlDiss,Fejos:2014qga} remain inconclusive, too.
We expect clarification beyond the LPA' taking into account derivative
couplings. \\
Finally, the simultaneous occurrence of two IR-stable fixed points
(although one of them being unphysical, and the truncation is
not reliable) is interesting
regarding the universality hypothesis. The example illustrates that, in principle, it is possible that
two systems sharing (a) the same spatial dimension, (b) the same number of order parameter components,
and (c) the same symmetry properties can be attracted to different IR-stable fixed points (here $F_1'^{(6)}$
and $F_2'^{(6)}$, respectively). Both associated universality classes are characterized by the same
representation of the same symmetry group. However, we state clearly that the given
example has to be regarded as an artifact of the utilized truncation. A similar situation, although to our
knowledge not strictly ruled out, is commonly not believed to appear in a physical setting.
\section*{Acknowledgment}
The author would like to thank HIC for FAIR for funding. The author would further like
to thank J{\"urgen} Eser, Francesco Giacosa, Mario Mitter, Dirk-Hermann Rischke, and
Bernd-Jochen Schaefer for valuable discussions.
\bibliographystyle{unsrt}
|
\section{Introduction}
\label{sec:intro}
In the context of General Relativity all kinds of exotic spacetimes
are allowed. With the appropriate stress-energy tensor $T_{\mu \nu}$,
following Einstein's equations, the spacetime can contain wormholes
and allow superluminal travel and the construction of ``time
machines''. However, in quantum field theory, there are restrictions
on $T_{\mu \nu}$. Two examples of these are the energy conditions and
the quantum inequalities. Pointwise energy conditions bound the
stress-energy tensor at each spacetime point, but they are easily
violated, since quantum field theory allows arbitrary negative
energies (e.g., in the Casimir effect). On the other hand, averaged
energy conditions bound the stress-energy tensor integrated along a
complete geodesic and quantum inequalities bound a weighted time
average of the total energy. These have been proven to hold in a
variety of spacetimes.
Ford \cite{Ford:1978qya} introduced quantum inequalities to prevent
the violation of the second law of thermodynamics. After that, quantum
inequalities were derived for various spacetimes and fields. The
majority of these results are for free fields on flat spacetimes
without boundaries, while a few are for interacting fields in
spacetimes with less than four dimensions
\cite{math-ph/0412028,arXiv:1304.7682}. For spacetimes with boundaries
there are difference quantum inequalities, which bound the difference
of $T_{\mu \nu}$ between some state and a reference state. But these
inequalities cannot be used to rule out exotic spacetimes arising from
vacuum energies.
Energy conditions have been used to address the possibility of exotic
spacetimes. Specifically, Ref.~\cite{Graham:2007va} showed that the
achronal averaged null energy condition (achronal ANEC) is sufficient
to rule out most known spacetimes with exotic curvature. In previous
work \cite{Kontou:2012ve}, we proved achronal ANEC for spacetimes with
a classical source. However, to do that we assumed that with a
timescale small compared to any curvature radius the quantum
inequality for flat spacetime still holds with small corrections. Ford,
Pfenning and Roman \cite{Ford:1995wg, Pfenning:1997wh} also have
suggested that the flat-space quantum inequalities can be used in
spacetimes with small curvature. However none of these results have
been explicitly proven.
Fewster and Smith \cite{Fewster:2007rh} proved an absolute quantum
inequality (i.e., one without dependence on a reference state) that
applies to spacetimes with curvature. Their bound involves the Fourier
transform of differentiated terms of the Hadamard series up to fifth
order. In recent work \cite{Kontou:2014eka}, we used their result to
provide a bound for flat spacetimes with a background potential. In
the same paper we also showed that is sufficient to consider only
terms up to first order, which makes Fewster and Smith's result more
practical. Using this result we will now show, in accordance with our
past conjecture and previous work, that in spacetimes with small
curvature, the quantum inequality is the same as in flat space with
small corrections that depend on the curvature.
The present paper closely follows Ref.~\cite{Kontou:2014eka}. We
begin by stating the general absolute quantum inequality of Fewster
and Smith \cite{Fewster:2007rh} in Sec.~\ref{sec:FSQEI}. The
inequality bounds the time averaged, renormalized energy density using
the Fourier transform of a point-split energy operator applied to
$\tilde{H}$, which is a combination of the Hadamard series and the
advanced-minus-retarded Green's function. In Sec~\ref{sec:tsplit} we
discuss and simplify this operator. In Sec.~\ref{sec:iE} we compute
the Green's function to first order for a spacetime with curvature,
and in Sec.~\ref{sec:H} we use that result to calculate
$\tilde{H}$. In section \ref{sec:tsplitH} we apply the point-split
energy operator and compute $\tilde{H}$. Finally we perform the
Fourier transform, and find the resulting quantum inequality in
Sec.~\ref{sec:QI}. We conclude in Sec.~\ref{sec:conclusions}.
We use the sign convention $(-,-,-)$ in the classification of Misner,
Thorne and Wheeler \cite{MTW} . Indices $a,b,c, \dots$ denote all
spacetime coordinates while $i,j,k \dots$ only spatial coordinates.
\section{Absolute Quantum Energy Inequality}
\label{sec:FSQEI}
We consider a massless, minimally-coupled scalar field with the usual
classical stress-energy tensor,
\be
T_{ab}=\nabla_a \Phi \nabla_b \Phi-\frac{1}{2} g_{ab} g^{cd} \nabla_c
\Phi \nabla_d \Phi \,.
\ee
Let $\gamma$ be any timelike geodesic parametrized by proper time $t$,
and let $g(t)$ be any any smooth, positive, compactly-supported
sampling function. In flat spacetime, Fewster and Eveson
\cite{Fewster:1998pu} showed that
\be\label{flatQI}
\int_{-\infty}^\infty dt \, T_{tt}(\gamma(t)) g(t)^2 \ge
-\frac{1}{16\pi^2}\int_{-\infty}^\infty dt \, g''(t)^2 \,.
\ee
We will generalize Eq.~(\ref{flatQI}) to geodesics in curved
spacetime.
First we construct Fermi normal coordinates \cite{Manasse:1963zz} in
the usual way: We let the vector $e_0(t)$ be the unit tangent to the
geodesic $\gamma$, and construct a tetrad by choosing arbitrary
normalized vectors $e_i(0), i=1,2,3$, orthogonal to $e_0(0)$ and to
each other, and define $\{e_i(t)\}$ by parallel transport along
$\gamma$. The point with coordinates $(x^0,x^1,x^2,x^3)$ is found by
traveling unit distance along the geodesic given by $x^i e_i(x^0)$
from the point $\gamma(x^0)$.
We work only in first order in the curvature and its derivatives, but
don't otherwise assume that it is small. We assume that the
components of the Ricci tensor in any Fermi coordinate system, and
their derivatives, are bounded,
\bea \label{Rmax}
|R_{ab}| \leq \Rmax \qquad |R_{ab,cd}| \leq \Rmax'' \qquad |R_{ab,cde}| \leq \Rmax'''\,.
\eea
Eqs.~(\ref{Rmax}) are intended as universal bounds which hold without
regard to the specific choice of Fermi coordinate system above. We
will not need a bound on the first derivative. The reason that we
bound the Ricci tensor and not the Riemann tensor is that, as we will
prove, the additional terms of the quantum inequality do not depend on
any other components of the Riemann tensor. We will discuss this
result further in the conclusions.
Following Ref.\cite{Fewster:2007rh}, we define the renormalized energy
density
\be \label{Tren}
\langle \Tren_{tt} \rangle \equiv \lim_{x\to x'} \Tsplit \left( \langle \phi(x)\phi(x') \rangle-H(x,x') \right)-Q+C_{tt}\,,
\ee
with quantities appearing in Eq.~(\ref{Tren}) defined as follows. $\Tsplit$ is the point-split energy density operator,
\be \label{tsplit}
\Tsplit=\frac{1}{2}\sum_{a=0}^3 e^\alpha_a \nabla_\alpha \otimes
e_a^{\beta'} \nabla_{\beta'}=\frac{1}{2}\sum_{a=0}^3 \partial_a
\partial_{a'}\,.
\ee
where $\partial_a f$ or $f_{,a}$ denotes the gradient of a function
$f$ with respect to $x$ in the direction of $e_a(x)$, and
$\partial_{a'} f$ or $f_{,a'}$ the same with $x'$ in place of $x$.
We renormalize the energy density according to the procedure of Wald
\cite{Wald:qft}, by taking the difference between the two point
function, $\langle \phi(x)\phi(x') \rangle$, and the Hadamard series,
\be \label{hadamard}
H(x,x')=\frac{1}{4\pi^2} \left[ \frac{\Delta^{1/2}}{\sigma_+(x,x')}+\sum_{j=0}^{\infty}v_j(x,x') \sigma_+^j (x,x') \ln(\sigma_+(x,x'))+\sum_{j=0}^{\infty}w_j (x,x')\sigma^j (x,x') \right] \,,
\ee
where $\sigma$ is the squared invariant length of the geodesic between
$x$ and $x'$, negative for timelike distance. In flat space
\be
\sigma(x,x')=-\eta_{ab} (x-x')^a (x-x')^b \,.
\ee
By $F(\sigma_+)$, for some function $F$, we mean the distributional
limit
\be
F(\sigma_+)=\lim_{\epsilon \to 0^+} F(\sigma_{\epsilon}) \,,
\ee
where
\be
\sigma_{\epsilon}(x,x')=\sigma(x,x')+2i \epsilon(t(x)-t(x'))+\epsilon^2 \,.
\ee
In some parts of the calculation it is possible to assume that both
points lie on the geodesic, so we define
\be
\tau=t-t'
\ee
and write
\be
F(\sigma_+)=F(\tau_-)=\lim_{\epsilon \to 0} F(\tau_\epsilon) \,,
\ee
where
\be
\tau_\epsilon=\tau-i\epsilon \,.
\ee
The function $\Delta$ is the van Vleck-Morette determinant bi-scalar, given by
\be \label{delta}
\Delta(x,x')=-\frac{\det (-\nabla_a \otimes \nabla_{b'} \sigma(x,x'))}{\sqrt{-g(x)}\sqrt{-g(x')}} \,.
\ee
The term $Q$ is the one introduced by Wald to preserve the
conservation of the stress-energy tensor. Wald \cite{Wald:1978pj}
calculated this term in the coincidence limit,
\be \label{Q}
Q=\frac{1}{12\pi^2}w_1(x,x) \,.
\ee
The term $C_{tt}$ handles the ambiguities in the definition of the
stress-energy tensor $T$ in curved spacetime. We will adopt the
axiomatic definition given by Wald \cite{Wald:qft}, but there remains
the ambiguity of adding local curvature terms with arbitrary
coefficients. From Ref.~\cite{Birrellbook} we find that these terms
include
\bml\label{eqn:12H}
\bea
^{(1)}H_{ab} &=& 2R_{;ab} -2g_{ab}\Box R - g_{ab}R^2/2 + 2RR_{ab} \label{1H}\\
^{(2)}H_{ab} &=& R_{;ab} -\Box R_{ab} -g_{ab}\Box R/2
- g_{ab}R^{cd} R_{cd}/2 + 2R^{cd}R_{acbd} \label{2H}\,.
\elea
Thus in Eq.~(\ref{qinequality}) we must include a term given by a
linear combination of Eqs.~(\ref{1H}) and (\ref{2H}) to first order in $R$,
\be \label{localc}
C_{tt}=a\, ^{(1)}\!H_{tt}+b\, ^{(2)}\!H_{tt}=2aR_{,ii}-\frac{b}{2}(R_{tt,tt}+R_{ii,tt}-3R_{tt,ii}+R_{ii,jj}) \,,
\ee
where $a$ and $b$ are undetermined constants.\footnote{There are also
ambiguities corresponding to adding multiples of the metric and the
Einstein tensor to the stress tensor. The first can be considered
renormalization of the cosmological constant and the second
renormalization of Newton's constant. We will assume that these
renormalization have been performed, and that the cosmological
constant is considered part of the gravitational sector, so neither
of these affects $T_{ab}$.}
From Ref.~\cite{Fewster:2007rh} we have the definition
\be \label{tilde}
\tilde{H}(x,x')=\frac{1}{2} \left[ H(x,x')+H(x',x)+iE(x,x') \right] \,,
\ee
where $iE$ is the antisymmetric part of the two-point function, which
we calculate in Sec.~\ref{sec:iE}. We will use the
Fourier transform convention
\be \label{Fourier}
\hat{f}(k) \text{ or } f^{\wedge}[k]=\int_{-\infty}^\infty dxf(x) e^{ixk} \,.
\ee
We can now state the quantum inequality of Ref.~\cite{Fewster:2007rh},
\bea \label{qinequality}
\int_{-\infty}^\infty d\tau \, g(t)^2 \langle \Tren_{tt} \rangle_{\omega} (t,0) &\geq& -\int_0^\infty\frac{d\xi}{\pi} \left[ g \otimes g (\theta^* \Tsplit \tilde{H}_{(5)})(t,t') \right]^{\wedge} (-\xi,\xi) \nonumber\\
&&+\int_{-\infty}^\infty dt \, g^2(t) (-Q+C_{tt}) \,,
\eea
where the operator $\theta^*$ denotes the pullback of the function to
the geodesic,
\be
(\theta^*\Tsplit \tilde{H}_{(5)})(t,t') \equiv (\Tsplit \tilde{H}_{(5)})(\gamma(t),\gamma(t'))\,,
\ee
and the subscript $(5)$ means that we include
only terms through $j = 5$ in the sums of
Eq.~(\ref{hadamard}). However, as we proved in
Ref.~\cite{Kontou:2014eka}, terms of order $j >1$ make no contribution
to Eq.~(\ref{qinequality}).
Thus we can write Eq.~(\ref{qinequality}) in our case as
\be\label{qinequality2}
\int_{-\infty}^\infty d\tau\,g(t)^2 \langle \Tren_{tt} \rangle (t,0)
\geq -B\,,
\ee
where
\be\label{B}
B = \int_0^\infty\frac{d\xi}{\pi}\hat F(-\xi,\xi)
+\int_{-\infty}^\infty dt\,g^2(t) \left(Q-2aR_{,ii}-\frac{b}{2}(R_{tt,tt}+R_{ii,tt}-3R_{tt,ii}+R_{ii,jj})\right)\,,
\ee
\be\label{F}
F(t,t') = g(t)g(t')\Tsplit \tilde H_{(5)}((t,0),(t',0))\,,
\ee
and $\hat F$ denotes the Fourier transform in both arguments according to
Eq.~(\ref{Fourier}).
\section{Simplification of $\Tsplit$}
\label{sec:tsplit}
The $\Tsplit$ operator, Eq.~(\ref{tsplit}), can be written
\be \label{spti}
\Tsplit=\frac12\left[\partial_t \partial_{t'}+\sum_{i=1}^3\partial_i \partial_{i'}\right]\,.
\ee
To simplify it, we will define the following operator,
\be\label{barxd}
\nabla_{\bar{x}}^2=\nabla_x^2+2\sum_{i=1}^3\partial_i\partial_{i'}+\nabla_{x'}^2\,,
\ee
which in flat space would be the derivative with respect to the center point.
Then Eqs.~(\ref{spti}) and (\ref{barxd}) give
\bea
\Tsplit&=&\frac{1}{2}\left[\partial_t \partial_{t'}+\frac{1}{2}
\left(\nabla_{\bar{x}}^2-\nabla_x^2-\nabla_{x'}^2\right)
\right]\nonumber\\
&=&\frac{1}{4}\left[\nabla_{\bar{x}}^2+\Box_x-\partial_t^2
+\Box_{x'}-\partial_{t'}^2+2\partial_t\partial_{t'}\right]\,,
\eea
where $\Box_x$ and $\Box_{x'}$ denote the D'Alembertian operator with
respect to $x$ and $x'$. Because we are using Fermi coordinates and
are on the generating geodesic, the D'Alembertian and Laplacian
operators have the same form with respect to Fermi coordinates as they
do in flat space. Then using
\be\label{bartd}
\partial_\tau^2=\frac14\left[\partial_t^2-2\partial_t \partial_{t'}+\partial_{t'}^2\right]\,,
\ee
we can write
\be
\Tsplit \tilde{H}=\frac{1}{4}\left[ \Box_x \tilde{H}
+ \Box_{x'} \tilde{H}+\nabla_{\bar{x}}^2 \tilde{H} \right]-\partial_\tau^2 \tilde{H}\,.
\ee
Consider the first term. The function $H(x,x')$ obeys the equation of
motion in $x$ and so does $E(x,x')$. Thus
\be \label{eqmd}
\Box_x \tilde{H}=\frac{1}{2}\Box_x H(x',x)\,.
\ee
The only asymmetrical part of $H$ comes from the $w_j$, so
\be
H(x',x) = H(x,x') + \frac{1}{\pi^2}\sum_j(w_j(x',x)-w_j(x,x'))\sigma^j(x,x')\,.
\ee
and so we have
\be\label{EofMH1}
\Box_x \tilde{H}
= \frac{1}{8\pi^2}\Box_x\sum_j(w_j(x',x)-w_j(x,x'))\sigma^j(x,x')\,.
\ee
Similarly,
\be\label{EofMH2}
\Box_{x'} \tilde{H}
= \frac{1}{8\pi^2}\Box_{x'}\sum_j(w_j(x,x')-w_j(x',x))\sigma^j(x,x')\,.
\ee
Adding together Eqs.~(\ref{EofMH1}) and (\ref{EofMH2}), we get
something which is symmetric in $x$ and $x'$ and vanishes in the
coincidence limit. Following the analysis of \S3A of
Ref.~\cite{Kontou:2014eka}, such a term makes no contribution to
Eq.~(\ref{B}) and for our purposes we can take
\be \label{T}
\Tsplit \tilde{H}=\left[\frac{1}{4}\nabla_{\bar{x}}^2-\partial_\tau^2
\right] \tilde{H}\,.
\ee
\section{General computation of $E$}
\label{sec:iE}
The function $E$ is the advanced minus the retarded Green's function,
\be \label{theE}
E(x,x')=G_A(x,x')-G_R(x,x')\,,
\ee
and $iE$ is the imaginary, antisymmetric part of the two-point
function. The Green's functions satisfy
\be \label{green}
\Box G(x,x')=\frac{\delta^{(4)}(x-x')}{\sqrt{-g}} \,.
\ee
Following Poisson, et al. \cite{Poisson:2011nh} and adjusting for different
sign and normalization conventions,
\be
G(x,x')=\frac{1}{4\pi} \left(2U(x,x')\delta(\sigma)+V(x,x')\Theta(-\sigma)\right)\,,
\ee
where $U(x,x') = \Delta^{1/2}(x,x')$ and $V(x,x')$ are smooth
biscalars.
For points $y$ null separated from $x'$, $V$ is called $\check V$
\cite{Poisson:2011nh} and satisfies
\be \label{vcheck}
\check V_{,a}\sigma^{,a}+\left[\frac{1}{2}\Box\sigma+2\right]\check V=-\Box U \,,
\ee
with all derivatives with respect to $y$. Now $\check V$ is first
order in the curvature, so we will do the rest of the calculation as
though we were in flat space. Under this approximation, we will
neglect coefficients which depend on the curvature, and also evaluate
curvature components at locations that would be relevant if we were in
flat space. The distance between these locations and the proper
locations is first order in the curvature, so the overall inaccuracy
will always be second order in the curvature and its derivatives.
Thus we use $\sigma^{,a}=-2(y-x')^a$ and
$\Box\sigma=-8$ in Eq.~(\ref{vcheck}) to get
\be \label{vcheck2}
(y-x')^a \check V_{,a}(y)+\check V(y)=\frac{1}{2} \Box U(y) \,.
\ee
Now suppose we want to compute $\check V$ at some point $x''$. We
need to integrate along the geodesic going from $x'$ to $x''$. So let
$y=x'+\lambda(x''-x')$ and observe that
\be
\frac{d(\lambda \check V(y))}{d \lambda}
= \lambda\frac{ d \check V(y)}{d\lambda} + \check V(y)
= \lambda(x''-x')^a \check V_{,a} + \check V(y)
= (y-x')^a \check V_{,a} + \check V(y)
= \frac{1}{2} \Box U(y)\,,
\ee
so
\be\label{vcheckfinal}
\check V(x'',x') = \frac{1}{2} \int_0^1 d\lambda \Box U(y)\,.
\ee
The function $V$ obeys \cite{Poisson:2011nh}
\be
\Box V(x,x')=0 \,.
\ee
Consider points $x$ and $x'$ on the geodesic $\gamma$, which in the
flat-space approximation means they are separated only in time, and
let $\bar x = (x+x')/2$. Then $V(x,x')$ can be found in terms of $V$
and its derivatives evaluated at the time $\bar t$ (the time component
of $\bar x$) using Kirchhoff's formula,
\be \label{Vu1}
V(x,x')=\frac{1}{4\pi} \int d\Omega \left[\check
V(x'')+\frac{\tau}2\frac{\partial}{\partial r}\check
V(x'')
+\frac{\tau}2\frac{\partial}{\partial t}\check V(x'') \right] \,,
\ee
where $\int d\Omega$ means to integrate over all unit vectors
$\hat\Omega$, and we now set
\be
x'' = \bar{x}+(\tau/2)\Omega
\ee
with the 4-vector $\Omega$ given by $\hat\Omega$ with unit time component.
Let us establish null-spherical coordinates $(u,v,\theta,\phi)$
with $u=t+r$, $v=t-r$, and the origin at $x'$. Then $x''$ has
$u=\tau$, $v = 0$. The derivative $\partial/\partial u$ can be
written $(\partial/\partial t + \partial/\partial r)/2$ and so
\be \label{Vu}
V(x,x')= \frac{1}{4\pi} \int d\Omega \, \frac{d}{du}\left[
u \check V(u\Omega/2,x')\right]_{u=\tau} \,.
\ee
From Eq.~(\ref{vcheckfinal}),
\be
u \check V\big(\frac{u}2\Omega,x'\big)
= \frac{1}{2} \int_0^u du'(\Box U)(u'\Omega/2,x')
\ee
and so
\be\label{Vint}
V(x,x')=\frac{1}{8\pi} \int d\Omega\left[(\Box U)(\tau\Omega/2,x') \right] \,.
\ee
We are only interested in the first order of curvature, so we can
expand U, which is just the square root of the Van Vleck determinant,
to first order. From Ref.~\cite{Visser:1992pz},
\be \label{deltaexp}
\Delta^{1/2}(x, x')=1-\frac{1}{2}
\int_0^1 ds (1-s)s R_{ab}(sx+(1-s)x')(x-x')^a (x-x')^b+O(R^2) \,,
\ee
so in the case at hand we can use
\be
U(x'') = \Delta^{1/2}(x'')=1-\frac{1}{2} \int_0^1 ds (1-s)s
R_{ab}(y) X^a X^b \,
\ee
where $y=sx'' = (su'',sv'',\theta'', \phi'')$ is a point between 0 and
$x''$, and the tangent vector $X=dy/ds$. We are interested in
$\Box_{x''} U(x'',0)$. To bring the $\Box$ inside the integral, we
define $Y=sX = (su'',sv'',0,0)$, and
\be
\Box U(x'', 0) = -\frac{1}{2} \int_0^1 ds (1-s)s \Box_{x''}[R_{ab}(y) X^a X^b]
= -\frac{1}{2} \int_0^1 ds (1-s)s \Box_{y}[R_{ab}(y) Y^a Y^b] \,.
\ee
For the rest of this section, all occurrences of $u$, $v$, $\theta$,
$\phi$, and derivatives with respect to these variables will refer to
these components of $y$ or $Y$.
Now we expand the D'Alembertian, in terms of angular derivatives,
derivatives in $u$, and derivatives in $v$,
\be
\Box=4\frac{\partial^2}{\partial v \partial u}-\frac{4}{u-v} \left(\frac{\partial}{\partial u}-\frac{\partial}{\partial v} \right)-\nabla^2_\Omega \,,
\ee
with
\be
\nabla^2_\Omega = \frac{4}{(v-u)^2 \sin{\theta}}
\frac{\partial}{\partial \theta} \left(\sin{\theta}
\frac{\partial}{\partial \theta} \right)+\frac{4}{(v-u)^2
\sin{\theta}^2}\frac{\partial^2}{\partial \phi^2} \,.
\ee
The angular integration in Eq.~(\ref{Vint}) annihilates the results of
$\nabla^2_\Omega$, so we have
\be \label{Vs}
V(x,x')=-\frac{1}{4\pi} \int d\Omega \int_0^1 ds s(1-s) \left[\partial_u \partial_v-\frac{1}{u-v} \left(\partial_u-\partial_v \right) \right](R_{ab}(y) Y^a Y^b) \,.
\ee
Outside the derivatives, we can take $v=0$ and change variables to $u=
s\tau$, giving
\bea
V(x,x')&=&-\frac{1}{4\pi\tau^3} \int d\Omega \int_0^\tau du (\tau-u)\left[
u\partial_u \partial_v- \partial_u+\partial_v \right](R_{ab}(y) Y^a Y^b)\\
&=&-\frac{1}{4\pi\tau^3} \int d\Omega \int_0^\tau du (\tau-u)
\partial_u[(u \partial_v- 1)(R_{ab}(y) Y^a Y^b)] \,.
\eea
We can integrate by parts with no surface contribution, giving
\bea
V(x,x')&=&\frac{1}{4\pi\tau^3} \int d\Omega \int_0^\tau du
(1-u \partial_v)(R_{ab}(y) Y^a Y^b)\\
&=&\frac{1}{4\pi\tau^3} \int d\Omega \int_0^\tau du
u^2 \left[ -u R_{uu,v}(y)-2 R_{uv}(y)+R_{uu}(y) \right]\nonumber \,.
\eea
Now
\be \label{RG}
R_{ab} = G_{ab}- (1/2)g_{ab}G \,,
\ee
where $G_{ab}$ is the Einstein tensor and $G$ its trace. Thus
\be\label{VG}
V(x,x')=\frac{1}{4\pi\tau^3} \int d\Omega \int_0^\tau du\,
u^2 \left[ -u G_{uu,v}(y)-2 G_{uv}(y)+(1/2)G(y)+G_{uu}(y) \right] \,.
\ee
Now define a vector field $Q_a(y) = G_{ab}(y)Y^b$. Then
\be
Q_{a;c} =G_{ab;c}(y)Y^b+G_{ab}(y){Y^b}_{;c}\,.
\ee
We write the covariant derivative only because we are working in
null-spherical coordinates, rather than because of spacetime
curvature, which we are ignoring because we already have first order
quantities.
Since the covariant divergence of $G$ vanishes,
\be
g^{ac}Q_{a;c} =g^{ac} G_{ab}(y){Y^b}_{;c} \,.
\ee
In Cartesian coordinates, $Y^b = y^b$, and ${y^b}_{;c} = \delta^b_c$,
which means that (in any coordinate system).
\be
g^{ac}Q_{a;c} = G \,.
\ee
Explicit expansion gives
\be
g^{ac}Q_{a;c} = 2 (Q_{v,u} + Q_{u,v})
-\frac{4}{u-v} (Q_u-Q_v)
- \frac{4}{(v-u)^2}\left[ \frac1{\sin\theta}
\frac{\partial}{\partial \theta} (\sin{\theta} Q_\theta)
+\frac{1}{\sin\theta^2}Q_{\phi,\phi}\right] \,,
\ee
but the angular terms vanish on integration. Now we expand
the derivatives in $u$ and $v$ and set $v=0$, giving
\blea
Q_{v,u}& = & uG_{uv,u}+ G_{uv}\\
Q_{u,v}& = & uG_{uu,v}+ G_{uv} \,,
\elea
so
\be\label{Qfinal}
\int d\Omega \,\left(2uG_{uv,u} + 2uG_{uu,v}+8G_{uv}- 4G_{uu}\right) =
\int d\Omega \, G \,.
\ee
Substituting Eq.~(\ref{Qfinal}) into Eq.~(\ref{VG}), we find
\be
V(x,x')=\frac{1}{4\pi\tau^3} \int d\Omega \int_0^\tau du \,
u^2 \left[u G_{uv,u}(y)+2 G_{uv}(y)-G_{uu}(y) \right]
\ee
and integration by parts yields
\be
V(x,x')=\frac{1}{4\pi} \int d\Omega \left[G_{uv}(x'') -
\frac{1}{\tau^3} \int_0^\tau du \, u^2 \left(G_{uv}(y)+G_{uu}(y) \right)\right] \,.
\ee
Now
\bea
\int d\Omega \int_0^\tau du \, u^2 \left(G_{uv}(y)+G_{uu}(y) \right)
&=&\frac12\int d\Omega\int_0^\tau du \, u^2 \left(G_{tt}(y)+G_{tr}(y) \right)\nonumber\\
&=& \frac12\int d\Omega\int_0^\tau du \, u^2 \left(G^{tt}(y)-G^{tr}(y) \right)\,,
\eea
which is 4 times the total flux of $G^{ta}$ crossing inward through
the light cone. Since this quantity is conserved, ${G^{ta}}_{;a}=0$,
we can integrate instead over a ball at constant time $\bar t$, giving
\be
4\int d\Omega\int_0^{\tau/2} dr\, r^2 G^{tt}(\bar x + r\Omega )
= \frac{\tau^3}2\int d\Omega\int_0^1 ds\, s^2 G^{tt}(\bar x + s(\tau/2)\Omega)
\ee
so
\be
V(x,x')=\frac{1}{8\pi} \int d\Omega \left[\frac{1}{2} \left[ G_{tt}(x'')-G_{rr}(x'') \right]
-\int_0^1 ds\, s^2 G_{tt}(x''_s)\right] \,,
\ee
where $x''_s=\bar{x}+s(\tau/2)\Omega$, and
\bea
G_R(x,x')=\Delta^{1/2}(x,x') \frac{\delta(\sigma)}{2\pi}+\frac{1}{32 \pi^2} \int d\Omega\bigg\{\frac{1}{2} \left[ G_{tt}(x'')-G_{rr}(x'') \right]-\int_0^1 ds \, s^2 G_{tt}(x''_s) \bigg\} \,,
\eea
\bea \label{finaliE}
E(x,x')=\Delta^{1/2}(x,x') \frac{\delta(\tau-|\bx-\bx'|)-\delta(\tau+|\bx-\bx'|)}{4\pi|\bx-\bx'|}&+&\frac{1}{32 \pi^2} \int d\Omega\bigg\{ \frac{1}{2} \left[ G_{tt}(x'')-G_{rr}(x'') \right]\nonumber\\
&&-\int_0^1 ds \, s^2 G_{tt}(x''_s) \bigg\} \sgn{\tau} \,.
\eea
\section{Computation of $\tilde{H}$}
\label{sec:H}
We now need to compute $\tilde H(x,x')$ and apply $\Tsplit$. First we
consider the term in $\tilde H(x,x')$ that has no dependence on the
curvature. It has the same form as it would in flat space
\cite{Fewster:2007rh,Kontou:2014eka},
\be\label{H-1}
\tilde H_{-1}(x,x')=H_{-1}(x,x')=\frac{1}{4\pi^2\sigma_+(x,x')}\,.
\ee
In Sec.~\ref{sec:tsplitH}, we will apply the fully general $\Tsplit$
from Eq.~(\ref{T}) with $\nabla_{\bar{x}}$ defined in
Eq.~(\ref{barxd}) to $\tilde H_{-1}(x,x')$.
All the remaining terms that we need are first order in the curvature,
so for these it is sufficient to take $\nabla_{\bar{x}}$ as the
flat-space Laplacian with respect to the center point, $\bar{x}$.
For this we only need to compute $\tilde H$ at positions given by time
coordinates $t$ and $t'$ but the same spatial position.
As we discussed, we only need to keep terms in $\tilde H$ with powers
of $\tau$ up to $\tau^2$, but we need $E$ exactly. The terms from $H$
alone give a function whose Fourier transform does not decline fast
enough for positive $\xi$ for the integral in Eq.~(\ref{B}) to
converge. Thus we extract the leading order terms from $iE$ and
combine these with the terms from $H$. This combination gives a
result that has the appropriate behavior after the Fourier transform.
Following the notation of Ref~\cite{Kontou:2014eka}, we let
$H_j(t,t')$, $j=0,1,\ldots$, denote the term in $H$ involving
$\tau^{2j}$ (with or without $\ln\tau$), and $H_{(j)}$ denote the sum
of all terms from $H_{-1}$ through $H_j$. We will split up $E(x,x')$
in similar fashion, define a ``remainder term''
\be
R_j = E - \sum_{k=-1}^j E_k\,,
\ee
and let
\blea
\tilde
H_{j}(x,x') &=& \frac12\left[H_j(x,x') + H_j(x',x) + iE_j(x,x')\right]\\
\tilde H_{(j)}(x,x') &=& \frac12\left[H_{(j)}(x,x') + H_{(j)}(x',x) + iE(x,x')\right]\,.
\elea
\subsection{Terms with no powers of $\tau$}
First we want to calculate the zeroth order of the Hadamard
series. The Hadamard coefficients are given by the Hadamard recursion
relations, which are the solutions to
\be
\Box H(x,x')=0, w_0=0 \,.
\ee
The recursion relations for the massless field in a curved background are \cite{Fewster:2007rh}
\be \label{recursion1}
\Box \Delta^{1/2}+2 v_{0,a} \sigma^{,a}+4 v_0+v_0 \Box \sigma=0 \,,
\ee
\be \label{recursion2}
\Box v_j +2(j+1) v_{j+1,a}\sigma^{,a}-4j(j+1)v_{j+1}+(j+1)v_{j+1}\Box \sigma=0 \,.
\ee
To find the zeroth order of the Hadamard series we need only
$v_0(x,x')$, which we find by integrating Eq.~(\ref{recursion1}) along
the geodesic from $x'$ to $x$. Since we are computing a first-order
quantity, we can work in flat space by letting $y'=x'+\lambda(x-x')$
and using the first-order formulas $\Box \sigma=-8$ and
$\sigma^{,a}=-2 (y'-x')^a$. From Eq.~(\ref{recursion1}), we have
\be
(y'-x')^a v_{0,a}+v_0=\frac{1}{4} \Box \Delta^{1/2} (y',x') \,,
\ee
and thus
\be\label{v01}
v_0(x,x')=\frac{1}{4} \int_0^1 d\lambda (\Box \Delta^{1/2}) (x'+\lambda(x-x'),x') \,.
\ee
by the same analysis as Eq.~(\ref{vcheckfinal}).
Using the expansion for $\Delta^{1/2}$ from Eq.~(\ref{deltaexp}) gives
\bea
v_0(x,x')&=&-\frac{1}{8}\int_0^1 d\lambda \int_0^1 ds (1-s)s
\Box_{y'}[R_ {ab}(sy'+ (1 -s)x') (y' -x') ^a (y' -x') ^b]\nonumber\\
&=&-\frac{1}{8} \int_0^1 d\lambda \int_0^1 ds(1-s)s \bigg[ (\lambda s)^2 (\Box R_{ab})(x'+s\lambda(x-x'))(x-x')^a (x-x')^b\nonumber\\
&&+2\lambda s R_{,b} (x'+s\lambda(x-x')) (x-x')^b+2 R(x'+s\lambda (x-x')) \bigg] \,.
\eea
We can combine the $s$ and $\lambda$ integrals by defining a new variable $\sigma=s\lambda$
\bea \label{lambdasigma}
&&\int_0^1 d\lambda \int_0^1 ds (1-s)s f(\lambda s)=\int_0^1 d\lambda \int_0^\lambda d\sigma \left(\frac{\sigma}{\lambda^2}-\frac{\sigma^2}{\lambda^3} \right) f(\sigma) \\
&&=\int_0^1 d\sigma \, f(\sigma) \int_\sigma^1 d\lambda \left(\frac{\sigma}{\lambda^2}-\frac{\sigma^2}{\lambda^3} \right)=\int_0^1 d\sigma \, f(\sigma) \left[ -\frac{\sigma}{\lambda}+\frac{\sigma^2}{2\lambda^2} \right]_\sigma^1=\frac{1}{2}\int_0^1 d\sigma \, f(\sigma) (1-\sigma)^2 \nonumber \,.
\eea
Then, changing $\sigma$ to $s$, we find
\bea\label{v0s}
v_0(x,x')&=&-\frac{1}{16}\int_0^1 ds (1-s)^2 \bigg[s^2 (\Box R_{ab})(x'+s(x-x'))(x-x')^a (x-x')^b\nonumber\\
&&+2 s R_{,b} (x'+s(x-x')) (x-x')^b+2 R(x'+s (x-x')) \bigg] \,.
\eea
or when the two points are on the geodesic,
\be
v_0(t,t')=-\frac{1}{16} \int_0^1 ds (1-s)^2 \bigg[ s^2 (\Box R_{tt})(x'+s \tau) \tau^2+4s \eta^{cd} R_{ct,d}(x'+s \tau)\tau+2 R(x'+s \tau) \bigg] \,.
\ee
In the second term we use the contracted Bianchi identity, $\eta^{cd}
R_{ct,d}= R_{,t}/2$, giving
\bea
&&2\int_0^1 ds (1-s)^2 s\tau R_{,t}(x'+s\tau)= 2\int_0^1 ds (1-s)^2 s \frac{d}{ds} R(x'+s\tau) \nonumber\\
&&=-2\int_0^1 ds (1-s)(1-3s)R(x'+s\tau)\,,
\eea
so the final expression for $v_0$ is
\be \label{v0}
v_0(t,t')=-\frac{1}{16} \int_0^1 ds (1-s) \bigg[ s^2 (1-s) \Box R_{tt} (\bar{x}+(s-1/2)\tau) \tau^2+4 s R(\bar{x}+(s-1/2)\tau) \bigg] \,.
\ee
To calculate $H_0$ we only need the zeroth order in $\tau$ from
$v_0$, so the first term does not contribute. In the second term, we
make a Taylor series expansion,
\be\label{Rt1}
R(\bar{x}+(s-1/2)\tau)=R(\bar{x})+R_{,t}(\bar{x})\tau (s-1/2)\tau+\frac{1}{2}R_{,tt}(\bar{x})\tau^2 (s-1/2)^2+O(\tau^3) \,,
\ee
but only the first term is relevant here. Thus
\be\label{v0final}
v_0(t,t')=-\frac{1}{4}\int_0^1 ds (1-s)s R(\bar{x})=-\frac{1}{24} R(\bar{x}) \,.
\ee
We also need to expand the Van Vleck determinant appearing in the the
Hadamard series. From Eq.~(\ref{deltaexp}),
\be \label{vanvleck}
\Delta^{1/2}(t,t')=1-\frac{1}{12}R_{tt}(\bar{x})\tau^2-\frac{1}{480}R_{tt,tt}(\bar{x})\tau^4+O(\tau^6)\,.
\ee
Keeping the first order term from Eq.~(\ref{vanvleck}) and using
Eq.~(\ref{v0final}), we have
\be \label{had0}
H_0(x,x')=\frac{1}{48\pi^2} \left[R_{tt}(\bar{x})-\frac{1}{2}R(\bar{x}) \ln{(-\tau_-^2)} \right] \,.
\ee
Now we can add the $H_0(x',x)$ which is the same except that $t$ and $t'$ interchange
\be \label{H0}
H_0(x,x')+H_0(x',x)=\frac{1}{24\pi^2} \left[R_{tt}(\bar{x})-R(\bar{x}) \ln{|\tau|} \right] \,.
\ee
Next we must include $E$ from Eq.~(\ref{finaliE}). We can expand the
components of the Einstein tensor around $\bar{x}$,
\be
G_{ab}(x'')=G_{ab}(\bar{x})+G_{ab}^{(1)}(x'') \,,
\ee
where $G_{ab}^{(1)}$ is the remainder of the Taylor series
\be \label{rem1}
G_{ab}^{(1)}(x'')=G_{ab}(x'')-G_{ab}(\bar{x})=\int_0^{\tau/2} dr \,G_{ab,i}(\bar{x}+r\Omega)\Omega^i \,.
\ee
Then from Eq.~(\ref{finaliE}) and using $\int d\Omega\, \Omega^i=0$ and
$\int d\Omega\, \Omega^i \Omega^j=(4\pi/3) \delta^{ij}$ we have
\bea \label{E0}
E_0(x,x')&=&\frac{1}{8 \pi} \left\{ \frac{1}{2}G_{tt}(\bar{x})-\frac{1}{6}G_{ii}(\bar{x})-\int_0^1 ds \, s^2 G_{tt}(\bar{x}) \right\} \sgn{\tau}\nonumber\\
&=& \frac{1}{48 \pi} G(\bar{x})\sgn{\tau} =-\frac{1}{48 \pi} R(\bar{x})\sgn{\tau}
\eea
and
\bea \label{R0}
R_0(x,x')=\frac{1}{32\pi^2}\int d\Omega \left\{ \frac{1}{2}\left[G_{tt}^{(1)}(x'')-G_{rr}^{(1)}(x'')\right]-\int_0^1 ds\,s^2 G_{tt}^{(1)}(x''_s) \right\} \sgn{\tau} \,.
\eea
Using
\be \label{log}
2\ln{|\tau|}+\pi i\sgn{\tau}=\ln{(-\tau_-^2)} \,,
\ee
we combine Eqs.~(\ref{H0}) and (\ref{E0}) to find
\be \label{Ht0}
\tilde{H}_0(t,t')=\frac{1}{48\pi^2} \left[ R_{tt}(\bar{x})-\frac{1}{2}R(\bar{x}) \ln{(-\tau_-^2)} \right] \,.
\ee
Combining all terms through order 0 gives
\be \label{Ht-10}
\tilde{H}_{(0)}(t,t')=\tilde{H}_{-1}(t,t')+\tilde{H}_0(t,t')+\frac{1}{2}i R_0(t,t') \,.
\ee
\subsection{Terms of order $\tau^2$}
Now we compute the terms of order $\tau^2$ in $H$ and $E$. To find
$v_0$ at this order we take Eqs.~(\ref{v0}) and (\ref{Rt1}) and
include terms through second order in $\tau$,
\bea \label{v0ab}
v_0(x,x')&=&-\frac{1}{24}R(\bar{x})-\frac{1}{16} \int_0^1 ds (1-s) \left[ s^2(1-s) \Box R_{tt}(\bar{x})+2s(s-1/2)^2 R_{,tt}(\bar{x})\right]\tau^2+ \dots \nonumber\\
&=&-\frac{1}{24}R(\bar{x})-\frac{1}{480} \left(\Box R_{tt}(\bar{x})+\frac{1}{2}R_{,tt}(\bar{x}) \right)\tau^2+\dots \,.
\eea
Next we need $v_1$ but since it is multiplied by $\tau^2$ in $H$ we
need only the $\tau$ independent term. From Eq.~(\ref{recursion2})
\be
\Box v_0+2v_{1,a} \sigma^{,a}+v_1 \Box \sigma=0 \,,
\ee
At $x=x'$, $ \sigma^{,a}=0$ so
\be\label{v1lim}
v_1(x,x)=\frac{1}{8} \lim_{x \to x'} \Box_x v_0(x,x') \,.
\ee
Using Eq.~(\ref{v0s}) in Eq.~(\ref{v1lim}), the only terms that
survive in the coincidence limit are those that have no powers of
$x-x'$ after differentiation, so
\be \label{v1}
v_1(x,x) = -\frac{1}{16} \int_0^1 ds (1-s)^2 s^2 \Box R(\bar{x})=-\frac{1}{480} \Box R(\bar{x}) \,.
\ee
Equations~(\ref{v0}), (\ref{v0ab}) and (\ref{v1}) agree with
Ref.~\cite{Decanini:2005gt} if we note that their expansions are
around $x$ instead of $\bar{x}$.
The $w_1$ at coincidence is given by Ref.~\cite{Wald:1978pj},
\be \label{w1}
w_1(x,x)=-\frac{3}{2} v_1(x,x)=\frac{1}{320} \Box R(\bar{x}) \,.
\ee
Combining Eqs.~(\ref{v0ab}), (\ref{v1}), and (\ref{w1}), and the
fourth order term from the Van Vleck determinant of
Eq.~(\ref{vanvleck}), and keeping in mind that $\sigma=-\tau^2$ when
both points are on the geodesic, we find
\be
H_1(x,x')=\frac{1}{640\pi^2} \bigg[\frac{1}{3}R_{tt,tt}(\bar{x})-\frac{1}{2} \Box R(\bar{x})
-\frac{1}{3}\left(\Box R_{ii}(\bar{x})+\frac{1}{2}R_{,tt}(\bar{x})\right)\ln{(-\tau_-^2)} \bigg]\tau^2 \,.
\ee
Then $H_1(x',x)$ is given by symmetry so
\be \label{H1}
H_1(x,x')+H_1(x',x)=\frac{1}{160\pi^2} \left[\frac{1}{6}R_{tt,tt}(\bar{x})-\frac{1}{4}\Box R(\bar{x})-\frac{1}{3}\left(\Box R_{ii}(\bar{x})+\frac{1}{2}R_{,tt}(\bar{x})\right)\ln{|\tau|} \right]\tau^2 \,.
\ee
The calculation of $E_1$ is similar to $E_0$, but now we have to include more terms to the Taylor expansion,
\be
G_{ab}(x'')=G_{ab}(\bar{x})+\frac{\tau}{2}G_{ab,i}(\bar{x})\Omega^i+\frac{\tau^2}{8}G_{ab,ij}\Omega^i \Omega^j(\bar{x})+G_{ab}^{(3)}(x'') \,,
\ee
where the remainder of the Taylor series is
\be \label{rem3}
G_{ab}^{(3)}(x'')=\frac{1}{2}\int_0^{\tau/2} dr G_{ab,ijk}(\bar{x}+r\Omega) \left( \frac{\tau}{2}-r\right)^2 \Omega^i \Omega^j \Omega^k \,.
\ee
Then from Eq.~(\ref{finaliE}) and using that $\int d\Omega \, \Omega^i=\int d\Omega \, \Omega^i \Omega^j \Omega^k=0$, $\int d\Omega \, \Omega^i \Omega^j=4\pi/3 \delta^{ij}$ and $\int d\Omega \, \Omega^i \Omega^j \Omega^k \Omega^l=(4\pi/15) (\delta^{ij} \delta^{kl}+\delta^{ik}\delta^{jl} +\delta^{il}\delta^{jk})$ we have
\bea
E_1(x,x')&=&-\frac{1}{192 \pi} \left[ \frac{1}{10} G_{ii,jj}(\bar{x})+\frac{1}{5} G_{ij,ij}(\bar{x}) -\frac{1}{2} G_{tt,ii}(\bar{x})+\int_0^1 ds \, s^4 G_{tt,ii}(\bar{x}) \right]\tau^2 \sgn{\tau} \nonumber\\
&=&-\frac{1}{320\pi} \left[ \frac{1}{6} G_{ii,jj}(\bar{x})+\frac{1}{3} G_{ij,ij}(\bar{x})-\frac{1}{2} G_{tt,ii} (\bar{x}) \right] \tau^2 \sgn{\tau} \,.
\eea
Using the conservation of the Einstein tensor, $0 = \eta^{ab} G_{ia,b} = G_{it,t} - G_{ij,j}$ and
$0 = \eta^{ab} G_{ta,b} = G_{tt,t} - G_{it,i}$ we can write
\be
G_{ij,ij}(\bar{x})=G_{tt,tt}(\bar{x}) \,.
\ee
So
\bea \label{E1}
E_1(x,x')&=&-\frac{1}{960\pi} \left(\frac{1}{2}G_{ii,jj}(\bar{x})+ G_{tt,tt}(\bar{x})-\frac{3}{2} G_{tt,ii} (\bar{x})\right) \tau^2 \sgn{\tau}\nonumber\\
&=&-\frac{1}{960\pi} \left( \Box R_{ii}(\bar{x})+\frac{1}{2} R_{,tt}(\bar{x}) \right) \tau^2 \sgn{\tau}
\eea
and
\be \label{R1}
R_1(x,x')=\frac{1}{32\pi^2} \int d\Omega \bigg\{ \frac{1}{2} \left[G_{tt}^{(3)}(x'')-G_{rr}^{(3)}(x'') \right]-\int_0^1 ds \, s^2 G_{tt}^{(3)}(x''_s) \bigg\} \sgn{\tau} \,.
\ee
To calculate $\tilde{H}_1$, we combine Eqs.~(\ref{H1}) and (\ref{E1})
and use Eq.~(\ref{log}) to get
\be \label{Ht1}
\tilde{H}_1(x,x')=\frac{\tau^2}{640\pi^2} \left[ \frac{1}{3}R_{tt,tt}(\bar{x})-\frac{1}{2} \Box R(\bar{x})-\frac{1}{3} \left( \Box R_{ii}(\bar{x})+\frac{1}{2} R_{,tt}(\bar{x}) \right) \ln{(-\tau_-^2)} \right] \,.
\ee
All terms through order 1 are then given by
\be \label{Ht-11}
\tilde{H}_{(1)}(t,t')=\tilde{H}_{-1}(t,t')+\tilde{H}_0(t,t')+\tilde{H}_1(t,t')+\frac{1}{2}iR_1(t,t') \,.
\ee
\section{The $\Tsplit \tilde{H}$}
\label{sec:tsplitH}
We can easily take the derivatives of $\tilde{H}_0$ and $\tilde{H}_1$
using Eq.~(\ref{T}), because they are already first order in
$R$. However in the case of the term $\nabla_{\bar{x}}^2
\tilde{H}_{-1}$ we have to proceed more carefully. From
Eqs.~(\ref{barxd}) and (\ref{H-1}) we have
\bea \label{TH-1}
\nabla_{\bar{x}}^2 \tilde{H}_{-1}&=&\frac{1}{4\pi^2} \sum_{i=1}^3 \left( \frac{\partial^2}{\partial (x^i)^2}+2\frac{\partial}{\partial x^i} \frac{\partial}{\partial x'^i}+\frac{\partial^2}{\partial (x'^i)^2} \right)\left(\frac{1}{\sigma_+}\right)\nonumber\\
&=&-\frac{1}{4\pi^2\sigma_+^2} \sum_{i=1}^3 \left( \frac{\partial^2\sigma}{\partial (x^i)^2}+2\frac{\partial^2 \sigma}{\partial x^i\partial x'^i}+\frac{\partial^2 \sigma}{\partial (x'^i)^2} \right) \,,
\eea
where we used $\partial \sigma/ \partial x^i=\partial \sigma/
\partial x'^i=0$ when the two points are on the geodesic. From \cite{Decanini:2005gt}, after we shift the Taylor series so that the Riemann tensor is evaluated at $\bar{x}$, we have
\bml\label{d2sigma}\bea
\frac{\partial^2 \sigma}{\partial (x^i)^2}&=&=-2\eta_{ii}-\frac{2}{3}R_{itit}(\bar{x})\tau^2-\frac{1}{2} R_{itit,t}(\bar{x})\tau^3-\frac{1}{5} R_{itit,tt}\tau^4+O(\tau^5) \label{sigma1}\\
\frac{\partial^2 \sigma}{\partial (x'^i)^2}&=&=-2\eta_{ii}-\frac{2}{3}R_{itit}(\bar{x})\tau^2+\frac{1}{2} R_{itit,t}(\bar{x})\tau^3-\frac{1}{5} R_{itit,tt}\tau^4+O(\tau^5) \label{sigma2}\\
\frac{\partial^2 \sigma}{\partial x^i\partial x'^i} &=&2\eta_{ii}-\frac{1}{3}R_{itit}(\bar{x})\tau^2-\frac{7}{40} R_{itit,tt}\tau^4+O(\tau^5)\label{sigma3} \,.
\elea
From Eqs.~(\ref{TH-1}) and (\ref{d2sigma}), and using $R_{itit}=-R_{tt}$ we have
\be \label{TH}
\nabla_{\bar{x}}^2 \tilde{H}_{-1}=-\frac{1}{4\pi^2} \left[\frac{2}{\tau^2}R_{tt}(\bar{x})+\frac{3}{4} R_{tt,tt}(\bar{x}) \right] \,.
\ee
From Eqs.~(\ref{F}) and (\ref{T}), we need to compute
\be
\int_0^\infty \frac{d\xi}{\pi}\hat F(-\xi,\xi')\,,
\ee
where
\be
F(t,t') = g(t) g(t')\left[\frac{1}{4}\nabla_{\bar{x}}^2 \tilde H_{(0)}(t,t')
-\partial_\tau^2 \tilde{H}_{(1)}(t,t') \right]\,.
\ee
Using
Eqs.~(\ref{H-1}), (\ref{R0}), (\ref{Ht0}), (\ref{Ht-10}), (\ref{R1}), (\ref{Ht1}), (\ref{Ht-11}) and (\ref{TH})
we can combine all terms in $F$ to write
\be
F(t,t') = g(t) g(t')\sum_{i=1}^6 f_i(t,t')\,,
\ee
with
\blea
f_1&=&\frac{3}{2\pi^2\tau_-^4}\\\
f_2&=&\frac{1}{48\pi^2\tau_-^2}[R_{ii}(\bar{x})-7 R_{tt}(\bar{x})]\\
f_3&=&\frac{1}{384\pi^2} \left[ \frac{1}{5} R_{tt,tt}(\bar{x})+\frac{1}{5}R_{ii,tt}(\bar{x})-R_{tt,ii}(\bar{x})+\frac{3}{5}R_{ii,jj}(\bar{x}) \right] \ln{(-\tau_-^2)}\\
f_4&=&\frac{1}{320\pi^2}\bigg[-\frac{43}{3}R_{tt,tt}(\bar{x})+\frac{7}{6} R_{tt,ii}(\bar{x})-\frac{1}{2}R_{ii,jj}(\bar{x}) \bigg] \\
f_5&=&\frac{1}{256\pi^2} \int d\Omega\,\nabla_{\bar{x}}^2 \left\{ \frac{1}{2} \left[ G_{tt}^{(1)}(x'')-G_{rr}^{(1)}(x'') \right] -\int_0^1 dss^2 G_{tt}^{(1)}(x''_s) \right\} i\sgn{\tau} \label{f5}\\
f_6&=& -\frac{1}{64\pi^2} \int d\Omega\,\partial_{\tau}^2 \bigg\{ \frac{1}{2} \left[G_{tt}^{(3)}(x'')-G_{rr}^{(3)}(x'') \right]-\int_0^1 ds s^2 G_{tt}^{(3)}(x''_s) \bigg\} i\sgn{\tau} \label{f6} \,.
\elea
\section{The quantum inequality}
\label{sec:QI}
We want to calculate the quantum inequality bound $B$, given
by Eq.~(\ref{B}). We can write it
\be
B=\sum_{i=1}^8 B_i\,,
\ee
where
\blea
B_i&=&\int_0^\infty \frac{d\xi}{\pi} \int_{-\infty}^\infty dt
\int_{-\infty}^\infty dt' g(t) g(t') f_i(t,t') e^{i\xi(t'-t)}\nonumber\\
&=&\int_0^\infty \frac{d\xi}{\pi} \int_{-\infty}^\infty d\tau
\int_{-\infty}^\infty d\bar{t}\,
g(\bar{t}-\frac{\tau}{2})g(\bar{t}+\frac{\tau}{2})f_i(\bar t,\tau)e^{-i\xi \tau}
\qquad i = 1\ldots6\\
\label{B7}B_7 &=& \int_{-\infty}^\infty dt\,g^2(t) Q(t)
= \frac{1}{3840\pi^2}\int_{-\infty}^\infty dt\,g^2(t)\Box R(\bar{t})\\
\label{B8}
B_8 &=& -\int_{-\infty}^\infty dt\,g^2(t)\left[2aR_{,ii}(\bar{x})-\frac{b}{2}(R_{tt,tt}(\bar{x})+R_{ii,tt}(\bar{x})-3R_{tt,ii}(\bar{x})+R_{ii,jj}(\bar{x}))\right]
\elea
using Eqs.~(\ref{Q}), (\ref{localc}), (\ref{B}) and (\ref{w1}). The first 6 terms
have exactly the same $\tau$ dependence as the corresponding terms in
Ref.~\cite{Kontou:2014eka}. So the Fourier transform proceeds in the
same way, except that instead of dependence on the potential and its
derivatives, we have dependence on the Ricci tensor and its
derivatives. After the Fourier transform, we see that $B_4$ and $B_7$
have exactly the same form so we merge them in one term. Thus
\be\label{BforR}
B = \frac{1}{16\pi^2}\left[ I_1
+\frac{1}{12} I_2
-\frac{1}{12} I_3
+\frac{1}{240} I_4
+\frac{1}{16\pi} I_5
-\frac{1}{4\pi} I_6\right]
-I_7\,,
\ee
where
\bml\label{I}\bea
I_1&=&\int_{-\infty}^{\infty} dt\,g''(t)^2 \\
I_2&=&
\int_{-\infty}^\infty d\bar{t}\,[R_{ii}(\bar{x})-7R_{tt}(\bar{x})]
(g(\bar{t}) g''(\bar{t}) - g'(\bar{t}) g'(\bar{t}))\\
\label{I3}I_3 &=&\int_{-\infty}^{\infty} d\tau\,\ln{|\tau|}\sgn{\tau}
\int_{-\infty}^{\infty} d\bar t\,\bigg[\frac{1}{5} R_{tt,tt}(\bar{x})+\frac{1}{5}R_{ii,tt}(\bar{x})-R_{tt,ii}(\bar{x}) \nonumber \\
&&+\frac{3}{5}R_{ii,jj}(\bar{x})\bigg] g(\bar{t}-\frac{\tau}{2})g'(\bar{t}+\frac{\tau}{2})\\
\label{I4}I_4 &=& \int_{-\infty}^{\infty}d\bar{t}\,
g(\bar{t})^2\bigg[-171 R_{tt,tt}(\bar{x})-R_{ii,tt}(\bar{x})+13R_{tt,ii}(\bar{x})-5R_{ii,jj}(\bar{x})\bigg]
\\
\label{I5}I_5 &=& \int_{-\infty}^\infty d\tau \, \frac{1}{\tau}
\int_{-\infty}^\infty d\bar{t}\,g(\bar{t}-\tau/2)g(\bar{t}+\tau/2) \int d\Omega\,\nabla_{\bar{x}}^2 \bigg\{ \frac{1}{2}\left[G_{tt}^{(1)}(x'')-G_{rr}^{(1)}(x'')\right] \nonumber\\
&&-\int_0^1 ds \, s^2 \left[G_{tt}^{(1)}(x''_s)\right] \bigg\} \sgn{\tau}
\\
\label{I6}I_6&=&\int_{-\infty}^\infty d\tau
\int_{-\infty}^\infty d\bar t\,
\partial_\tau^2 \left[\frac{1}{\tau} g(\bar{t}-\tau/2)g(\bar{t}+\tau/2) \right]
\int d\Omega\,\bigg\{ \frac{1}{2} \left[G_{tt}^{(3)}(x'')-G_{rr}^{(3)}(x'') \right]\nonumber\\
&&-\int_0^1 ds \, s^2 G_{tt}^{(3)}(x''_s) \bigg\}\sgn{\tau}\\
\label{I7}I_7 &=&\int_{-\infty}^\infty dt\,g^2(t)\left[2aR_{,ii}(\bar{x})-\frac{b}{2}(R_{tt,tt}(\bar{x})+R_{ii,tt}(\bar{x})-3R_{tt,ii}(\bar{x})+R_{ii,jj}(\bar{x}))\right]\,.
\elea
If we only know that the Ricci tensor and its derivatives are bounded,
as in Eq.~(\ref{Rmax}), we can restrict the magnitude of each term of
Eq.~(\ref{BforR}). We start with the second term
\bea
|I_2| &\leq& \int_{-\infty}^\infty d\bar{t}\, \left|R_{ii}(\bar{x})-7R_{tt}(\bar{x})\right|
|g(\bar{t}) g''(\bar{t}) - g'(\bar{t}) g'(\bar{t})| \nonumber\\
&\leq& 10 \Rmax \int_{-\infty}^\infty d\bar{t} [g(\bar{t}) |g''(\bar{t})| + g'(\bar{t})^2] \,.
\eea
Terms $I_3$, $I_4$ and $I_7$ are similar. Since Eq.~(\ref{Rmax}) holds
regardless of rotation, we can write $G_{rr}$ in terms of radial and
as azimuthal components of $R$ to find $|G_{rr}| < 2\Rmax$, and
similarly $|G_{tt}|<2\Rmax$. Using these results and Eq.~(\ref{rem1})
for the remainder we have
\bea
&& \left| \int d\Omega\,\nabla_{\bar{x}}^2 \bigg\{ \frac{1}{2}\left[G_{rr}^{(1)}(x'')-G_{tt}^{(1)}(x'')\right]+\int_0^1 ds \, s^2 G_{tt}^{(1)}(x''_s) \bigg\} \right| \nonumber\\
&& \leq \frac{|\tau|}{2} \int d\Omega \left\{ \frac{1}{2} \left[
|\nabla^2 G_{rr,i}(\bar x)|+|\nabla^2 G_{tt,i}(\bar x)| \right]+\int_0^1 ds \, s^3 |\nabla^2 G_{tt,i}(\bar x)| \right\} |\Omega^i| \nonumber\\
&& \leq \Rmax''' \frac{15|\tau|}{4} \sum_i \int d\Omega |\Omega^i|=\frac{45\pi}{2} |\tau| \Rmax''' \,.
\eea
For $I_6$ we use Eq.~(\ref{rem3}) for the remainder
\bea
&&\left| \int d\Omega\,\bigg\{ \frac{1}{2} \left[G_{rr}^{(3)}(x'')-G_{tt}^{(3)}(x'') \right]+\int_0^1 ds \, s^2 G_{tt}^{(3)}(x''_s) \bigg\} \right| \nonumber\\
&& \leq \frac{|\tau|^3}{48} \int d\Omega \left\{ \frac{1}{2} \left[ |G_{rr,ijk}(\bar x)|+|G_{tt,ijk}(\bar x)| \right]+\int_0^1 ds \,s^5 |G_{tt,ijk}(\bar x)| \right\} |\Omega^i||\Omega^j||\Omega^k| \nonumber\\
&& \leq \Rmax''' \frac{7 |\tau|^3}{144} \sum_{i,j,k}\int d\Omega |\Omega^i| |\Omega^j| |\Omega^k|=\frac{7( 2 \pi+1)}{24} |\tau|^3 \Rmax''' \,.
\eea
After we bound all the terms and calculate the derivatives in $I_6$ we can define
\bml\label{J17}\bea
J_2&=&\int_{-\infty}^\infty dt\left[g(t)|g''(t)|+g'(t)^2\right]\\
J_3&=&\int_{-\infty}^\infty dt \int_{-\infty}^\infty dt' |g'(t')|g(t) |\!\ln{|t'-t|}| \\ J_4&=&\int_{-\infty}^\infty dt\,g(t)^2\\
J_5&=&\int_{-\infty}^\infty dt \int_{-\infty}^\infty dt' g(t)g(t') \\
J_6&=&\int_{-\infty}^\infty dt \int_{-\infty}^\infty dt' |g'(t')|g(t)|t'-t|\\
J_7&=&\int_{-\infty}^\infty dt \int_{-\infty}^\infty dt'
\left [g(t)|g''(t')| +g'(t)g'(t')\right] (t'-t)^2
\elea
and find
\blea
|I_2| &\le & 10 \Rmax J_2\\
|I_3| &\le &\frac{46}{5} \Rmax'' J_3\\
|I_4| &\le & 258 \Rmax''J_4\\
|I_5|&\le & \frac{45\pi}{2} \Rmax''' J_5\\
|I_6|&\le &\frac{7(2\pi+1)}{48}\Rmax'''\left(4J_5+4J_6+J_7\right)\\
|I_7|&\le & (24|a|+11|b|)\Rmax''J_4\,.
\elea
Thus the final form of the inequality is
\bea
\label{final}
\int_{\mathbb{R}} d\tau\,g(t)^2\langle T^{ren}_{tt}\rangle_{\omega}
(t,0) \geq- \frac{1}{16\pi^2} \bigg\{& &I_1+\frac{5}{6}\Rmax J_2\\
&+&\Rmax''\left[\frac{23}{30}J_3+\left(\frac{43}{40}+16\pi^2(24|a|+11|b|)\right) J_4\right]
\nonumber\\
&+&\Rmax''' \left[\frac{163\pi+14}{96\pi}J_5
+\frac{7(2\pi+1)}{192\pi}(4J_6+J_7)\right] \bigg\}\,.\nonumber
\eea
Once we have a specific sampling function $g$, we can compute the
integrals of Eqs.~(\ref{J17}) to get a specific bound. In the case of
a Gaussian sampling function,
\be\label{Gaussiang}
g(t)=e^{-t^2/t_0^2}\,,
\ee
we computed these integrals numerically in Ref.~\cite{Kontou:2014eka}.
Using those results the right hand side of Eq.~(\ref{final}) becomes
\be\label{finalGaussian}
-\frac{1}{16 \pi^2 t_0^3} \left\{ 3.76+2.63 \Rmax t_0^2+[3.42+ 197.9(24|a|+11|b|)] \Rmax'' t_0^4+ 6.99 \Rmax''' t_0^5 \right\}\,.
\ee
The leading term is just the flat spacetime bound of
Ref.~\cite{Fewster:1998pu} for $g$ given by Eq.~(\ref{Gaussiang}).
The possibility of curvature weakens the bound by introducing
additional terms, which have the same dependance on $t_0$ as in
Ref.~\cite{Kontou:2014eka}, with the Ricci tensor bounds in place of
the bounds on the potential.
\section{Conclusion}
\label{sec:conclusions}
In this work, using a general quantum inequality of Fewster and Smith
\cite{Fewster:2007rh} we derived an inequality for a minimally-coupled
quantum scalar field on spacetimes with small curvature. We calculated
the necessary Hadamard series terms and the Green's function for this
problem to first order in the curvature. Combining these terms gives
$\tilde{H}$ and taking the Fourier transform gives a bound in terms of
the Ricci tensor and its derivatives.
If we know the spacetime explicitly, Eqs.~(\ref{qinequality2}),
(\ref{BforR}), and (\ref{I}) give an explicit bound on the weighted
average of the energy density along the geodesic. This bound depends
on integrals of the Ricci tensor and its derivatives combined with the
weighting function $g$.
If we do not know the spacetime explicitly but we know that the Ricci
tensor and its first 3 derivatives are bounded, Eqs.~(\ref{J17}) and
(\ref{final}) give a quantum inequality depending on the bounds and the
weighting function. If we take a Gaussian weighting function,
Eq.~(\ref{finalGaussian}) gives a bound depending on the Ricci tensor
bounds and the width of the Gaussian, $t_0$.
As expected, the result shows that the corrections due to curvature
are small if the quantities $\Rmax t_0^2$, $\Rmax'' t_0^4$, and
$\Rmax''' t_0^5$ are all much less than 1. That will be true if the
curvature is small when we consider its effect over a distance equal
to the characteristic sampling time $t_0$ (or equivalently if $t_0$ is
much smaller than any curvature radius), and if the scale of variation
of the curvature is also small compared to $t_0$.
In all bounds, there is unfortunately an ambiguity resulting from the
unknown coefficients of local curvature terms in the gravitational
Lagrangian. This ambiguity is parametrized by the quantities $a$ and
$b$.
Ford and Roman \cite{Ford:1995wg} have argued that flat-space quantum
inequalities can be applied in curved spacetime, so long as the radius
of curvature is small as compared to the sampling time. The present
paper explicitly confirms this claim and calculates the magnitude of
the deviation. The curvature must be small not only on the path where
the quantum inequality is to be applied but also at any point that is
in both the causal future of some point of this path and the causal
past of another. All such points are included in the integrals in
Eq.~(\ref{I5}) and (\ref{I6}).
Is is interesting to consider vacuum spacetimes, i.e., those whose
Ricci tensor vanishes. These include, for example, the Schwarzschild
and Kerr spacetimes, and those consisting only of gravitational waves.
In such spacetimes, the flat-space quantum inequality will hold to
first order without modification. There are, of course, second-order
corrections. For the Schwarzschild spacetime, for example, these were
calculated explicitly by Visser
\cite{Visser:1996iw,Visser:1996iv,Visser:1997sd}.
In Ref.~\cite{Kontou:2012ve} we proved a theorem ruling out achronal
ANEC violation, given a conjecture that paths with small acceleration
in spacetimes with small curvature obey the same null-projected
timelike-averaged quantum inequality as in flat space
\cite{Fewster:2002ne}, with corrections of the form discussed here.
The present result is a step toward proving that conjecture. In
future work we intend to extend the present result to null-projected
instead of timelike-projected quantum inequalities and to handle
slightly non-geodesic curves.
\section*{Acknowledgments}
We thank Larry Ford for helpful discussions. This research was
supported by grant RFP3-1014 from The Foundational Questions Institute
(fqxi.org). E.-A. K. gratefully acknowledges support from a John
F. Burlingame Graduate Fellowship in Physics."
|
\section{Introduction}
Radiation transfer (RT) has been long recognized as a indispensable
ingredient in numerically simulating many astrophysical phenomena
including the reionization of intergalactic medium (IGM) in the early
universe, radiative feedback during the galaxy formation, and others. So
far, varieties of numerical schemes for solving the RT in three
dimensions are proposed during the last two decades \citep{iliev2006},
and some of them can be coupled with the hydrodynamic simulations
\citep{iliev2009} thanks to not only the increase of the available
computational resources, but also the improvement of numerical
algorithms to solve the RT in many astrophysical conditions.
Most of the numerical schemes for RT can be divided into two groups: one
is the moment-based schemes which solve the moment equation of the RT
equation instead of solving the RT equation directly, and the other is
the ray-tracing schemes. As for the moment-based schemes, the important
advantage is that the computational costs scale with the number of mesh
grids, $N_{\rm m}$ and hence can be easily coupled with hydrodynamic
simulations. The flux-limited diffusion (FLD) scheme, which adopts the
closure relation valid in the diffusion limit, is the most common among
the moment-based schemes, while there are a number of more sophisticated
schemes which close the moment equations with the optically thin
variable Eddington tensor approximation \citep{gnedin2001} and the
locally evaluated Eddington tensor (the M$_1$ model)
\citep{gonzalez2007, skinner2013, kanno2013}. The accuracy and validity
of the moment-based schemes are, however, problem-dependent. For
example, the FLD scheme has a problem in handling shadows formed behind
opaque objects \citep{gonzalez2007}. While schemes with M$_1$ model are
capable of simulating shadows sucessfully, they cannot solve the
crossing of multiple beamed lights, where the beamed lights unphysically
merge into one beam \citep{rosdahl2013}. Therefore, the ray-tracing
schemes are naturally chosen for solving the RT in situations that we
are considering in the studies of galaxy formation and cosmic
reionization, in which there exist a number of radiation sources.
In ray-tracing schemes, emission and absorption of radiation are
followed along the light-rays that extend through the computational
domain. As for the long-characteristics schemes
\citep{abel1999,sokasian2001} in which light-rays between all radiation
sources and all other relevant meshes are considered, the computational
cost scales with $N_{\rm m}^{2}$ in general cases and $N_{\rm
m}^{4/3}N_{\rm s}$ when we consider only the RT from point radiating
sources, where $N_{\rm s}$ is the number of point sources. On the other
hand, for the short-characteristics schemes \citep{kunasz1988,stone1992}
which are similar to the long-characteristics schemes but integrate the
RT equation only along paths connecting nearby mesh grids, the
computational cost scales with $N_{\rm m}^{5/3}$ in general and $N_{\rm
m}N_{\rm s}$ for the RT from point sources. Ray-tracing schemes are in
principle versatile for any physical settings but computationally much
more expensive than the moment-based schemes. Due to such huge
computational costs, RT simulations with the ray-tracing schemes have
been applied only to static conditions or snapshots of hydrodynamical
simulations in a post-process manner in many previous studies.
Some of the ray-tracing schemes are now coupled with hydrodynamical
simulations adopting smoothed particle hydrodynamics (SPH) codes
\citep{susa2006, hasegawa2010, pawlik2011} and mesh-based codes
\citep{rijkhorst2006, wise2011}, and they can handle the RT and its
hydrodynamical feedback in a self-consistent manner. Majority of these
radiation hydrodynamics codes, however, consider the transfer of
radiation only from point sources and ignore the effect of radiation
transfer from spatially extended diffuse sources, such as the
recombination radiation emitted from ionized regions and infrared
radiation emitted by dust grains, since the computational costs for
computing the transfer of diffuse radiation is prohibitively large.
Specifically, in the numerical RT calculations of the hydrogen ionizing
radiation, we usually adopt the on-the-spot approximation in which one
assumes that the ionizing photons emitted by radiative recombinations in
ionized regions are absorbed by neutral atoms in the immediate vicinity
of the recombining atoms. However, adopting the on-the-spot
approximation can fail to notice the important effects of diffuse
recombination radiation in some situations. The roles of ionizing
recombination photons in the epoch of cosmic reionization is discussed
by a number of works \citep{ciardi2001, miralda-escude2003, dopita2011,
rahmati2013a}. \citet{dopita2011} proposed the recombination photons
produced in the fast accretion shocks in the structure formation in the
universe as an possible source of ionizing photons responsible for the
cosmic reionization, though \citet{wyithe2011} showed that its impact on
the cosmic reionization is not very significant. It is also reported
that the recombination radiation plays an important role at transition
regions between highly ionized and self-shielded regions
\citep{rahmati2013a}. As for the effect of recombination photons on the
galaxy-size scales, \citet{inoue2010} showed that the recombination
radiation produces the Lyman-`bump' feature in the spectral energy
distributions of high-$z$ galaxies, and also that the escaping ionizing
photons from high-$z$ galaxies are to some extent contributed by the
recombination radiation. \citet{rahmati2013b} also pointed out that the
recombination radiation makes the major contribution to the
photo-ionization at regions where the gas is self-shielded from the UV
background radiation.
The RT of infrared diffuse radiation emitted by dust grains plays an
important roles in the evolution of star-forming galaxies, in which the
radiation pressure exerted by multi-scattered infrared photons drives
stellar winds. In most of numerical simulations of galaxy formation,
however, such momentum transfer is treated only in a phenomenological
manner (e.g. Okamoto et al. 2014).
In this paper, we present a new ray-tracing scheme to solve the RT of
diffuse radiation from spatially extended radiating sources efficiently
on processors with highly parallel architectures such as graphics
processing units (GPUs) and multi-core CPUs which are recently popular
or available in near future. The basic idea of the scheme is based on
the scheme presented by \citet{razoumov2005} and `Authentic Radiation
Transfer' (ART) scheme \citet{nakamoto2001b}. Generally speaking,
development of such numerical schemes with high concurrency is of
critical importance because the performance improvement of recent
processors are achieved by the increase of the number of processing
elements or CPU cores integrated on a single processor chip rather than
the improvement of the performance of individual processing elements.
The rest of the paper is organized as follows. Section 2 is devoted to
describe the numerical scheme to simulate the radiation transfer. In
section 3, we present our implementation of the scheme suitable to
highly parallel architectures such as GPUs and CPUs with multi-core
architectures. We present the results of numerical test suits of RT of
diffuse radiation in Section 4. The computational performance of our
implementation is shown in Section 5. Finally, we summarize our results
in Section 6.
\section{Methodology}
In this section, we describe our ray-tracing scheme of diffuse radiation
transfer. Generally, the radiation field can be decomposed into two
components. One is the direct incident radiation from point radiation
sources, and the other is the diffuse radiation emerged from spatially
extended regions. In our implementation, the RT of photons emitted by
point radiation sources is computed separately from that of diffuse
radiation. Throughout in this paper, we consider the RT of hydrogen
ionizing photons emitted by point radiation sources, and recombination
photons emerged from the ionized regions as the diffuse radiation. We
use the steady state RT equation for a given frequency $\nu$:
\begin{equation}
\frac{dI_{\nu}}{d\tau_{\nu}} = -I_{\nu} + \mathcal{S}_{\nu} ,
\label{eq:rad_tr1}
\end{equation}
where $I_{\nu},\tau_{\nu}$ and $\mathcal{S}_{\nu}$ are the specific
intensity, the optical depth and the source function, respectively. The
source function is given by $\mathcal{S}_{\nu} =
\varepsilon_{\nu}/\kappa_{\nu},$ where $\kappa_{\nu}$ and
$\varepsilon_{\nu}$ are the absorption and emission coefficients,
respectively. The formal solution of this equation is given by
\begin{equation}
I_{\nu}(\tau_{\nu}) = I_{\nu}(0)\,e^{-\tau_{\nu}} + \int^{\tau_{\nu}}_{0} \mathcal{S}_{\nu}(\tau'_{\nu}) e^{-(\tau_{\nu}-\tau'_{\nu})}d\tau'_{\nu} ,
\label{eq:rad_tr2}
\end{equation}
where $\tau'_{\nu}$ is the optical depth at a position along the ray.
When we adopt the ``on-the-spot'' approximation in which recombination
photons emitted in ionized regions are assumed to be absorbed where they
are emitted, we neglect the source function, $\mathcal{S}_\nu$, and the
formal solution is simply reduced to
\begin{equation}
I_{\nu}(\tau_{\nu}) = I_{\nu}(0)\,e^{-\tau_{\nu}} .
\label{eq:on_the_spot}
\end{equation}
\subsection{RT from point radiation sources\label{sub:rt_point}}
To solve the RT from point radiation sources, we compute the optical
depth between each pair of a point radiation source and a target mesh
grid, i.e. an end point of each light-ray (see
Figure~\ref{fig:long}). Instead of solving equation
(\ref{eq:on_the_spot}), we compute the radiation flux density at the
target mesh grid as
\begin{equation}
f_\alpha(\nu) = \frac{L_\alpha(\nu)}{4\pi r_\alpha^2}\,\exp\left[-\tau_\alpha(\nu)\right],
\end{equation}
where $L_\alpha(\nu)$ is the intrinsic luminosity of the $\alpha$-th
point radiation source, and $r_\alpha$ and $\tau_\alpha(\nu)$ are the
distance and the optical depth between the point radiation source and the
target mesh grid, respectively. Then, the photo-ionization and
photo-heating rates of the $i$-th species contributed by the $\alpha$-th
point radiation source are computed by
\begin{equation}
\Gamma^\alpha_{i,\gamma} = \int^\infty_{\nu_i}
\frac{f_\alpha(\nu)}{h\nu}\sigma_i(\nu)\,d\nu,
\end{equation}
and
\begin{equation}
\mathcal{H}^\alpha_{i,\gamma} = \int^\infty_{\nu_i} \frac{f_\alpha(\nu)}{h\nu}(h\nu-h\nu_i)\sigma_i(\nu)\,d\nu
\end{equation}
respectively, where $\sigma_i(\nu)$ and $\nu_i$ are the ionization cross
section and the threshold frequency of the $i$-th species,
respectively. In the test simulations desribed in this paper, we compute
these photo-ionization and photo-heating rates in a photon-conserving
manner \citep{abel1999} as described in appendix
~\ref{app:photon_conserving}.
\begin{figure}
\centering \includegraphics[width=5cm]{f1.eps}
\caption{Schematic illustration of the ray-tracing method for the
radiation emitted by a point radiation source in the
two-dimensional mesh grids.\label{fig:long}}
\end{figure}
For a single point radiation source, the number of rays to be calculated
is $N_{\rm m}$, and the number of mesh grids traveled by a single
light-ray is in the order of $N^{1/3}_{\rm m}$. Thus, the computational
cost for a single point radiation source is proportional to
$N^{4/3}_{\rm m}$. Therefore, the total computational cost scales as
$N_{\rm m}^{4/3} N_{\rm s}$, where $N_{\rm s}$ is the number of point
radiation sources. For a large number of point radiation sources, we can
mitigate the computational costs by adopting more sophisticated scheme
such as the ARGOT scheme \citep{okamoto2012} in which a distant group of
point radiation sources is treated as a bright point source located at
the luminosity center with a luminosity summed up for all the sources in
the group to effectively reduce the number of radiation sources and
hence the computational cost is proportional to $N^{4/3}_{\rm m} \log
N_{\rm s}$.
\subsection{RT of the diffuse radiation \label{sub:rt_rec}}
We solve the equation~(\ref{eq:rad_tr2}) to compute the RT of the
diffuse radiation. The numerical scheme we adopt in this work is based
on the method developed by \citet{razoumov2005} and ART scheme
\citep{nakamoto2001b, iliev2006}, which is reported to have little
numerical diffusion in the searchlight beam test like the long
characteristics method although its computational cost is proportional to
$N_{\rm m}^{5/3}$ similarly to that of the short characteristic method
\citep{nakamoto2001b}. In this scheme, we solve the
equation~(\ref{eq:rad_tr2}) along equally spaced parallel rays as
schematically shown in Figure~\ref{fig:parallel_ray}.
For a given incoming radiation intensity $I^{\rm in}_\nu$ along a
direction $\hat{\mathbf{n}}$, the outgoing radiation intensity $I^{\rm
out}_\nu$ after getting through a path length $\Delta L$ of a single
mesh is computed by integrating equation~(\ref{eq:rad_tr2}) as
\begin{equation}
I^{\rm out}_\nu (\hat{\mathbf{n}}) = I^{\rm in}_\nu(\hat{\mathbf{n}})\,e^{-\Delta\tau_\nu} + \mathcal{S}_\nu
(1-e^{-\Delta\tau_\nu}),
\label{eq:rt}
\end{equation}
where $\Delta\tau_\nu$ is the optical depth of the path length $\Delta
L$ (i.e. $\Delta\tau_\nu = \kappa_\nu \Delta L$), and $\mathcal{S}_\nu$
and $\kappa_\nu$ are the source function and the absorption coefficient
of the mesh grid, respectively.
The intensity of the incoming radiation averaged over the path length
$\Delta L$ across a single mesh grid can be calculated as
\begin{equation}
\bar{I}^{\rm in}_\nu(\hat{\mathbf{n}}) = \frac{1}{\Delta L}\int_0^{\Delta L} I^{\rm
in}_\nu (\hat{\mathbf{n}})e^{-\kappa_\nu l}\,dl = I_\nu^{\rm in}(\hat{\mathbf{n}})\frac{1-e^{-\Delta
\tau_\nu}}{\Delta \tau_\nu}.
\end{equation}
In addition to this, we have a contribution to the radiation intensity
from the source function which we set constant in each mesh grid, and
the total intensity averaged over the path length is given by
\begin{equation}
\bar{I}_\nu(\hat{\mathbf{n}}) = \bar{I}^{\rm in}_\nu(\hat{\mathbf{n}}) + \mathcal{S}_\nu = I_\nu^{\rm in}(\hat{\mathbf{n}}_i)\frac{1-e^{-\Delta
\tau_\nu}}{\Delta \tau_\nu} + \mathcal{S}_\nu
\end{equation}
For those mesh grids through which multiple parallel light-rays pass,
the averaged intensity can be given by
\begin{equation}
\bar{I}_\nu^{\rm ave}(\hat{\mathbf{n}}) =
\frac{\sum_j \Delta \tau_{\nu,j} \bar{I}_{\nu,j}(\hat{\mathbf{n}})}{\sum_j \Delta
\tau_{\nu,j}}
= \bar{I}_{\nu}^{\rm ave, in}(\hat{\mathbf{n}}) + \mathcal{S}_\nu,
\label{eq:path_averaged_intensity}
\end{equation}
where $\bar{I}_{\nu,i}$ and $\Delta \tau_{\nu, i}$ are the intensity averaged
over the $i$-th light-ray and the optical depth of $i$-th light-ray in
the mesh grids, respectively, $\bar{I}^{\rm ave, in}_\nu$ is a
contribution from the incoming radiation given by
\begin{equation}
\bar{I}^{\rm ave, in}_\nu(\hat{\mathbf{n}}) = \frac{\sum_j \Delta \tau_{\nu,j}
\bar{I}^{\rm in}_{\nu,j}(\hat{\mathbf{n}})}{\sum_j \Delta
\tau_{\nu,j}},
\label{eq:averaged_intensity}
\end{equation}
and the summation is over all the parallel light-rays in the same mesh
grid. Then, the mean intensity can be computed by averaging
$\bar{I}^{\rm ave}_\nu$ described above over all the directions as,
\begin{equation}
J_\nu = \frac{1}{N_{\rm d}}\sum_{i=1}^{N_{\rm d}} \bar{I}_{\nu}^{\rm
ave}(\hat{\mathbf{n}}_i) = J^{\rm in}_\nu + \mathcal{S}_\nu,
\label{eq:mean_intensity}
\end{equation}
where $\hat{\mathbf{n}}_i$ describes a vector toward the $i$-th
direction and $N_{\rm d}$ is the number of directions of light-rays to
be considered, $\bar{I}_{\nu}^{\rm ave}(\hat{\mathbf{n}}_i)$ is the
averaged intensity along the $i$-th direction calculated with
equation~(\ref{eq:path_averaged_intensity}), and $J^{\rm in}_\nu$ is
given by
\begin{equation}
J^{\rm in}_\nu=\frac{1}{N_{\rm d}}\sum_{i=1}^{N_{\rm d}} \bar{I}^{\rm
ave, in}_{\nu}(\hat{\mathbf{n}}_i).
\end{equation}
Then, the photo-ionization and photo-heating rates of the $i$-th species
contributed by the diffuse radiation in each mesh grid can be computed
as
\begin{equation}
\Gamma_{i,\gamma}^{\rm diff} = 4\pi\int_{\nu_i}^\infty \frac{J_\nu}{h\nu}\sigma_i(\nu)\,d\nu
\end{equation}
and
\begin{equation}
\mathcal{H}_i^{\rm diff} = 4\pi \int_{\nu_i}^\infty
\frac{J_\nu}{h\nu}(h\nu-h\nu_i) \sigma_i(\nu)\,d\nu
\end{equation}
As for the recombination radiation of ionized hydrogen (HII) regions,
the number of recombination photons to the ground state per unit time
per unit volume, $\dot{N}^{\rm rec}$, can be expressed in terms of the
emissivity coefficient $\varepsilon_\nu$ as
\begin{equation}
\dot{N}^{\rm rec}=4\pi \int_{\nu_0}^{\infty}\frac{\varepsilon_\nu}{h\nu}\,d\nu =
[\alpha_{\rm A}(T)-\alpha_{\rm B}(T)]n_{\rm e}n_{\rm HII},
\end{equation}
where $\nu_0$ is the Lyman limit frequency, $\alpha_{\rm A}(T)$ and
$\alpha_{\rm B}(T)$ are the recombination rates of HII as functions of
temperature $T$ in the case-A and case-B approximations, respectively,
and $n_{\rm e}$ and $n_{\rm HII}$ are the number densities of the
electrons and HII, respectively.
In this work, we adopt a rectangular functional form of
$\varepsilon_\nu/(h\nu)$ as
\begin{equation}
\frac{\varepsilon_\nu}{h\nu} = \left\{
\begin{array}{ll}
\displaystyle \frac{\Delta\alpha(T)n_{\rm
e}n_{\rm HII}}{4\pi \Delta\nu_{\rm th}} &
(\nu_0\le\nu\le\nu_0+\Delta\nu_{\rm th}) \\
0 & (\mbox{otherwise}),
\end{array}
\right.
\end{equation}
where $\Delta\alpha(T)=\alpha_{\rm A}(T)-\alpha_{\rm B}(T)$ and
$\Delta\nu_{\rm th}$ is the frequency width of the recombination
radiation and given by $\Delta\nu_{\rm th} = k_{\rm B}T/h$. Thus, the
source function is given by
\begin{equation}
\mathcal{S}_\nu = \frac{\varepsilon_\nu}{\kappa_\nu} =
\left\{
\begin{array}{ll}
\displaystyle \frac{\Delta\alpha(T)n_{\rm e}n_{\rm HII}h\nu}{4\pi n_{\rm
HI}\sigma_{\rm HI}(\nu) \Delta\nu_{\rm th}} &
(\nu_0\le\nu\le\nu_0+\Delta\nu_{\rm th} )\\
0& ({\rm otherwise}).
\end{array}
\right.
\end{equation}
This spectral shape is the same as adopted in \citet{kitayama2004} and
\citet{hasegawa2010}, the results of which are compared with our results
to verify the validity of our scheme. Note that for the typical
temperature of HII regions, $T=10^4\,\,{\rm K}$, we have $\Delta\nu_{\rm
th}\ll \nu_0$. Calculations of the mean intensity based on
equations~(\ref{eq:rt}) to (\ref{eq:mean_intensity}) are done in a
monochromatic manner at the Lyman limit frequency $\nu_0$. In computing
photo-ionization and photo-heating rates, we assume that the mean
radiation intensity $J^{\rm in}_\nu$ has a rectangular functional form
as
\begin{equation}
J^{\rm in}_\nu = \left\{
\begin{array}{ll}
\displaystyle \mathcal{J}^{\rm in} &
(\nu_0\le\nu\le\nu_0+\Delta\nu_{\rm th}) \\
0 & (\mbox{otherwise}).
\end{array}
\right.
\end{equation}
Therefore, the photo-ionization and photo-heating rates of neutral
hydrogen can be rewritten as
\begin{equation}
\Gamma^{\rm diff}_{\rm HI,\gamma} = 4\pi\mathcal{J}^{\rm
in}\int_{\nu_0}^{\nu_0+\Delta\nu_{\rm th}}\frac{\sigma_{\rm HI}(\nu)}{h\nu} d\nu + \frac{\Delta\alpha(T)n_{\rm
e}n_{\rm HII}}{n_{\rm HI}},
\label{eq:gamma_diff}
\end{equation}
and
\begin{equation}
\mathcal{H}^{\rm diff}_{\rm HI,\gamma} = 4\pi\mathcal{J}^{\rm
in}\int_{\nu_0}^{\nu_0+\Delta\nu_{\rm th}}\left(1-\frac{\nu_0}{\nu}\right)\,\sigma_{\rm HI}(\nu)\,d\nu +
\frac{\Delta\alpha(T)n_{\rm e}n_{\rm HII}}{2n_{\rm HI}}h\Delta\nu_{\rm
th},
\label{eq:heat_diff}
\end{equation}
respectively. In the test simulations described in
section~\ref{sec:test}, we fix the frequency width $\Delta\nu_{\rm th}$
by assuming temperature of HII regions to be $10^4$ K, and the integrals
in equations~(\ref{eq:gamma_diff}) and (\ref{eq:heat_diff}) can be
estimated prior to the simulations. For the transfer of diffuse
radiation with more general spectral form, we can easily extend our
method described above by adopting a nonparametric functional form of
radiation spectra as
\begin{equation}
I_{\nu}=\sum_i \mathcal{I}^i \Pi(\nu-\nu_i,\Delta\nu),
\end{equation}
where $\Pi(x,y)$ is the rectangular function given by
\begin{equation}
\Pi(x,y) = \left\{\begin{array}{ll}
1 & -y/2 \le x \le y/2\\
0 & {\rm otherwise} \\
\end{array}\right.,
\end{equation}
and $\nu_i$ is the central frequency of the $i$-th frequency bin.
The number of light-rays parallel to a specific direction is
proportional to $N_{\rm m}^{2/3}$, and the number of mesh grids
traversed by a single light-ray is in the order of $N_{\rm
m}^{1/3}$. Therefore, the total computational cost is proportional to
$N_{\rm m}N_{\rm d}$.
\begin{figure}
\centering \includegraphics[width=5cm,clip]{f2.eps}
\caption{Schematic illustration of the ray-tracing scheme for the
diffuse radiation in the two-dimensional mesh grid. For a given
direction, equally-spaced parallel light-rays are cast from boundaries
of the simulation volume and travel to the other boundaries. Note that
gray mesh grids are traversed by multiple parallel light-rays, while
the subsets of light-rays depicted by blue or red get through them only
once. } \label{fig:parallel_ray}
\end{figure}
\subsection{Angular resolution for RT of the diffuse radiation}
\label{sub:direction}
The number of the directions of light-rays, $N_{\rm d}$, determines
angular resolution of the RT of the diffuse radiation. In order to
guarantee that light-rays from a mesh grid on a face of the simulation
box reach all the mesh grids on the other faces, $N_{\rm d}$ should be
in the order of $N_{\rm m}^{2/3}$. In the case that the mean free path
of the diffuse photons is sufficiently shorter the simulation box size,
however, such a large $N_{\rm d}$ is redundant because only a small
fraction of diffuse photons reach the other faces, and we can reduce the
total computational cost by decreasing the number of directions, $N_{\rm
d}$, while keeping the reasonable accuracy of the diffuse RT. Thus, the
number of directions should be flexibly changed depending on the
physical state.
To achieve this, we use the HEALPix (Hierarchical Equal Area isoLatitude
Pixelization) software package \citep{gorski2005} to set up the
directions of the light-rays. The HEALPix is suitable to our purposes in
the sense that each direction corresponds to exactly the same solid angle
and that the directions are nearly uniformly sampled. Furthermore, it
can provide a set of directions with these properties in arbitrary
resolutions, each of which contains $12N_{\rm side}^2$ directions, where
$N_{\rm side}$ is an angular resolution parameter. Since it is larger
than the number of mesh grids on six faces of a cube with a side length
of $N_{\rm side}$ mesh spacings, $6N_{\rm side}^2$, it is expected that
a set of light-rays originated from a single point with directions
generated by the HEALPix with an angular resolution parameter of $N_{\rm
side}$ get through all the mesh grids within a cube centered by the
point with a side length of $N_{\rm side}$ mesh spacings. Thus, the
optimal number of directions should be chosen so that the mean free path
of the recombination photons is sufficiently shorter than $N_{\rm
side}\Delta H$, where $\Delta H$ is the mesh spacing.
\subsection{Chemical reactions and radiative heating and cooling}
With photo-ionization and photo-heating rates computed with the
prescription described above, time evolutions of chemical compositions
and thermal states of gas are computed in the same manner as adopted in
\citet{okamoto2012}. Details of the numerical schemes are briefly
described in appendices~\ref{app:reaction},\ref{app:heating_cooling} and
\ref{app:timestep}.
The chemical reaction rates and radiative cooling rates adopted in this
paper are identical to those adopted in \citet{okamoto2012}, and the
literatures from which we adopt these rates are summarized in Table
\ref{tab:rate}.
\begin{table}
\caption{Rates of chemical reactions and radiative cooling processes
adopted in this paper. Reference for radiative recombination rates (RR)
of HII, HeII and HeIII in the case-A and case-B approximation;
collisional ionization rates (CIR) of HI, HeI, and HeII; recombination
cooling rates (RCR) of HII, HeII and HeIII in the case-A and case-B
approximation; collisional ionization cooling rates (CICR) of HI, HeI
and HeII; collisional excitation cooling rates (CECR) of HI, HeI and
HeII; bremsstrahlung cooling rate; inverse Compton cooling rate (CCR);
photoionization cross sections (CS) of HI, HeI and HeII. \label{tab:rate}}
\centering
\begin{tabular}{ll}
\hline
physical process & literature \\
\hline
RR (case-A) & (1), (1), (2) \\
RR (case-B) & (3), (3), (3) \\
CIR & (7), (7), (1) \\
RCR (case-A) & (2), (2), (2) \\
RCR (case-B) & (3), (5), (3) \\
CICR & (2), (2), (2) \\
CECR & (2), (2), (2) \\
BCR & (4) \\
CCR & (6) \\
CS & (8), (8), (8)\\
\hline
\end{tabular}
\begin{flushleft}
(1) \citet{abel1997};
(2) \citet{cen1992};
(3) \citet{hui1997};
(4) \citet{hummer1994};
(5) \citet{hummer1998};
(6) \citet{ikeuchi1986};
(7) \citet{janev1987};
(8) \citet{agnagn};
\end{flushleft}
\end{table}
\section{Implementation on Highly Parallel Architectures}
In this section, we describe the details of the implementation of the RT
calculation of the diffuse radiation which performs effectively on
recently popular processors with highly parallel architecture, such as
GPUs, multi-core CPUs, and many-core processors. Throughout this paper,
we present the results using the implementation with the
\verb|OpenMP| and \verb|CUDA| technologies. The former is supported by
most of the multi-core processors, and the many-core processors such as
the Intel Xeon-Phi processor, while the latter is the parallel
programming platform for GPUs by NVIDIA.
In the implementation on GPUs with the \verb|CUDA| platform, the fluid
dynamical and chemical data in all the mesh grids are transferred from
the memory attached to CPUs to those of GPUs prior to the RT
calculations. After the RT calculations, ionization rates and heating
rates in all the mesh grids computed on GPUs are sent back to the CPU
memory.
\subsection{Ray Grouping}\label{ss:ray_grounping}
In the calculations of the transfer of the diffuse radiation described
in the previous section, many parallel light-rays travel from boundaries
of the simulation volume until they reach the other boundaries. On
processors with highly parallel architecture, a straightforward
implementation is to assign a single thread to compute the RT along each
light-ray and calculate the RT along multiple light-rays in
parallel. Such a simple implementation, however, does not work because
some mesh grids are traversed by multiple parallel light-rays (see gray
mesh grids in Figure~\ref{fig:parallel_ray}), and in computing
equation~(\ref{eq:path_averaged_intensity}), multiple computational
threads write data to the identical memory addresses. Thus,
equation~(\ref{eq:path_averaged_intensity}) has to be computed not in
parallel but in a exclusive manner using the ``atomic operations''. The
use of the atomic operations, however, significantly degrade the
parallel efficiency and computational performance in many architectures.
To avoid such use of atomic operations and the deterioration of the
parallel efficiency, we split the parallel light-rays into several
groups so that parallel light-rays in each group do not traverse any
mesh grid more than once. For example, in two-dimensional mesh grids in
Figure~\ref{fig:parallel_ray}, parallel light-rays are split into two
groups each of which are depicted by blue and red arrows. One can see
that light-rays in each groups do not intersect any mesh grids more than
once. We can extend this technique to the three-dimensional mesh grids
by splitting the parallel light-rays into four groups, where the
light-rays in each group are cast from the two-dimensionally interleaved
mesh grids as depicted by the same color in Figure~\ref{fig:grouping}.
\begin{figure}
\centering \includegraphics[width=6.9cm]{f3.eps}
\caption{Schematic illustration of light-ray grouping for the
three-dimensional mesh grids. Light-rays in each group start from
boundary faces of mesh grids painted with the same color. Only the
light-rays in one group are shown in this figure. \label{fig:grouping}}
\end{figure}
\subsection{Efficient Use of Multiple External Accelerators}
Many recent supercomputers are equipped with multiple external
accelerators such as GPUs on a single computational node, each of which
has an independent memory space. To attain the maximum benefit of the
multiple accelerators, we decompose calculations of the diffuse
radiation transfer according to the directions of the light-rays, and
assign the decomposed RT calculation to the multiple accelerators. After
carrying out the RT calculation for the assigned set of directions, and
computing the mean intensity with equation (\ref{eq:mean_intensity})
averaged over the partial set of directions on each external
accelerator, the results are transferred to the memory on the hosting
nodes. Then, we obtain the mean intensity averaged over all directions.
\subsection{Node Parallelization}
In addition to the thread parallelization within processors, we
implement the inter-node parallelization using the Message Passing
Interface (MPI). In the inter-node parallelization, the simulation box
is evenly decomposed into smaller rectangular blocks with equal volumes
along the Cartesian coordinate.
For the inter-node parallelization of the calculations of the diffuse
radiation transfer, we adopt the multiple wave front (MWF) scheme
developed by \citet{nakamoto2001}, in which light-ray directions are
classified into eight groups according to signs of their three direction
cosines, and for each group of light-ray directions, the RT calculations
along each direction are carried out in parallel on a ``wave front'' in
the node space, while the RT for different directions are computed on
the other wave fronts simultaneously. By transferring the radiation
intensities at the boundaries from upstream nodes to downstream ones,
one can sequentially compute the RT of diffuse radiation along all the
directions in each group of light-ray directions. See
\citet{nakamoto2001} for more detailed description of MWF scheme.
\section{Test Simulation}
\label{sec:test}
In this section, we present a series of test simulations to validate our
RT code. All the test simulations are carried out with $128^3$ mesh
grids and the angular resolution parameter of $N_{\rm side}=8$ unless
otherwise stated.
\subsection{Test-1 : HII region expansion}
The first test is the simple problem of a HII region expansion in a
static homogeneous gas which consists of only hydrogen around a single
ionizing source. We adopt the same initial condition as that of Test-2
in Cosmological Radiative Transfer Codes Comparison Project I
\citep{iliev2006}, where the hydrogen number density is $n_{\rm
H}=10^{-3}\,{\rm cm}^{-3}$ and the initial gas temperature is
$T=100\,{\rm K}$. The ionizing source emits the blackbody radiation with
an effective temperature of $10^5\,{\rm K}$, and $5\times 10^{48}$
ionizing photons per second and located at a corner of simulation box
with a side length of 6.6 kpc. In this initial condition, the
recombination time is $t_{\rm rec}=122.4{\rm\, Myr}$ and the
Str\"{o}mgren radius is estimated to be 5.4 kpc.
Figure~\ref{fig:test-1} shows the radial profiles of ionization fraction
and gas temperature at $t=30{\rm Myr}$, $100{\rm Myr}$ and $500{\rm
Myr}$. The solid lines with and without circles indicate the results
with and without the on-the-spot approximation (OTSA) , respectively.
In the calculation with the effect of recombination radiation, the
ionized regions are more extended than those computed with the
on-the-spot approximation , especially at later stages ($t=100{\rm Myr}$
and $500{\rm Myr}$) because of the additional ionization of hydrogen by
the recombination photons.
To verify the validity of our scheme for the transfer of diffuse
recombination radiation, we compare our results with the ones obtained
with the one-dimensional spherically symmetric RT code by
\citet{kitayama2004}, which also incorporates the transfer of
recombination photons emitted by the ionized hydrogen using the
impact-parameter method. We find that the one-dimensional results with
the effect of recombination radiation denoted by dashed lines show a
good agreement with our three-dimensional results, indicating that our
treatment of diffuse radiation transfer is consistent with that of
well-established impact-parameter method.
\begin{figure}
\centering
\includegraphics[width=15cm]{f4.eps}
\caption{Test-1: Radial profiles of neutral and ionized fractions of
hydrogen and gas temperature at $t=30{\rm Myr}$, $100{\rm Myr}$ and
$500{\rm Myr}$. Solid lines with and without circles indicates the
results with and without the on-the-spot approximation (OTSA),
respectively. Dashed lines show the results obtained with
one-dimensional spherically symmetric code without the OTSA presented
in \citet{kitayama2004}. \label{fig:test-1}}
\end{figure}
\subsection{Test-2 : Shadow by a dense clump}
\label{sub:angular_resolution}
In the second test, we compute the RT from point radiation source in the
presence of a dense gas clump. A point radiation source is located at
the center of the simulation box with the same side length as the Test-1
(6.6kpc), and surrounded by the ambient uniform gas with the same
hydrogen number density and temperature as the Test-1 ($n_{\rm
H}=10^{-3}\,{\rm cm}^{-3}$ and $T=100\,{\rm K}$, respectively). In
addition, we set up a spherical dense gas clump with a radius of
0.56\,kpc centered at 0.8 kpc apart from the point radiation source
along the $x$-direction. We set the density of the dense clump to 200
times higher than that of the ambient gas, and the temperature is set to
100\,K. The spectrum and luminosity of the point radiation source is the
same as that in Test-1.
In Figure~\ref{fig:test-2-map}, maps of the neutral fraction of hydrogen
in the mid-plane of the simulation volume at $t=30\,{\rm Myr}$, 100Myr
and 500Myr are shown. One can see that the ionizing photons are strongly
absorbed by the dense gas clump and conical shadows are created behind
the gas clump in the both runs with and without the effect of
recombination radiation. In the run without the on-the-spot approximation
(upper panels of Figure~\ref{fig:test-2-map}), the recombination photons
emitted by the ionized gas gradually ionize the neutral gas behind the
dense gas clump. On the other hand, in the run with the on-the-spot
approximation, the boundaries of neutral and ionized regions are kept
distinct because of the lack of recombination photons.
This test is identical to Test-6 in \citet{hasegawa2010} calculated with
the START code. In the START code, the RT is solved with a ray-tracing
scheme based on the SPH technique, and the transfer of diffuse
recombination radiation can be handled by allowing each SPH paricle to
radiate recombination photons. Figure~\ref{fig:test-2-section} shows
profiles of gas temperature and hydrogen neutral fraction along the
lines across the conical shadow shown in Figures~\ref{fig:test-2-map}
and \ref{fig:test-2-t} as well as the results obtained with the START
code. One can see that both the results with and without the OTSA are in
fairly good agreement with each other, which supports the validity of
our scheme for diffuse radiation transfer.
\begin{figure}
\centering \includegraphics[width=10cm]{f5.eps}
\caption{Test-2: Maps of neutral fraction of hydrogen in the mid-plane
of the simulation box at $t=30{\,\rm Myr}$, $100{\,\rm Myr}$ and
$500{\,\rm Myr}$. The lower and upper panels show the results with and
without the on-the-spot approximation (OTSA), respectively. Dashed
vertical lines indicate the location along which the profiles of
temperature and neutral fractions are presented in
Figure~\ref{fig:test-2-section}. \label{fig:test-2-map}}
\end{figure}
\begin{figure}
\centering \includegraphics[width=10cm]{f6.eps}
\caption{Test-2: Same as Figure~\ref{fig:test-2-map} but shows the gas
temperature maps in the mid-plane of the simulation box. Dashed
vertical lines indicate the location along which the profiles of
temperature and neutral fractions are presented in
Figure~\ref{fig:test-2-section}.\label{fig:test-2-t}}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=10cm]{f7.eps}
\caption{Profiles of temperatures and hydrogen neutral fractions along
the dashed lines shown in Figures~\ref{fig:test-2-map} and
\ref{fig:test-2-t} with and without the OTSA. Results in
\citet{hasegawa2010} are also shown for
comparison.\label{fig:test-2-section}}
\end{figure}
In this test, we also perform runs with various angular resolution
parameter $N_{\rm side}$ in the RT calculation of diffuse radiation to
see the effect of angular resolution.
Figure~\ref{fig:test-2-resolution} shows maps of neutral fraction of
hydrogen in the mid-plane of the simulation box at $t=30\,{\rm Myr}$
with angular resolution parameter $N_{\rm side}$ of 16, 4 and 1. The
results with $N_{\rm side}=16$ and $4$ are in good agreement with one
with $N_{\rm side}=8$ in Figure~\ref{fig:test-2-map}, indicating that
the angular resolution with $N_{\rm side}=4$ is sufficient for the
current RT calculations. The results with $N_{\rm side}=1$, however,
have spurious features in the map of neutral fraction. These numerical
artifacts can be ascribed to the low angular resolution of light-rays by
the comparison of the mean free path of the diffuse recombination
photons and $N_{\rm side}\Delta H$. As described in
\S~\ref{sub:angular_resolution}, the mean free path of the diffuse
photons should be sufficiently smaller than $N_{\rm side}\Delta H$ to
compute the RT of diffuse photons accurately. For the recombination
photons emitted by ionized hydrogens in the current setup, the mean free
path in the neutral ambient gas is estimated as
\begin{equation}
\lambda_{\rm mfp} = \frac{1}{n_{\rm HI}\sigma_{\rm HI}(\nu_0)}=51.4
\left(\frac{n_{\rm HI}}{10^{-3} {\rm cm}^{-3}}\right)^{-1} {\rm pc},
\end{equation}
and the mesh spacing is $\Delta H=6.6{\,\rm kpc}/128=51.5{\,\rm pc}$.
Thus, it is quite natural to have strong numerical artifacts in the
results with $N_{\rm side}=1$, because the mean free path is almost
equal to $N_{\rm side}\Delta H$, and the condition for the accurate RT
calculation ($\lambda_{\rm mfp} \ll N_{\rm side}\Delta H$) is not
satisfied.
\begin{figure}
\centering \includegraphics[width=10cm]{f8.eps}
\caption{Test-2: Maps of neutral fraction of hydrogen in the mid-plane
of the simulation box at $t=30{\,\rm Myr}$ for different angular
resolution parameter, $N_{\rm side}=16$, 4 and 1 from left to right. \label{fig:test-2-resolution}}
\end{figure}
\subsection{Test-3 : Ionization front trapping and shadowing by a dense clump}
The third test computes the transfer of ionizing radiation incident to a
face of the rectangular simulation box and the propagation of ionized
region into a spherical dense clump. This test is indentical to the
Test-3 in \citet{iliev2006}. The size of the simuation box is $6.6$ kpc,
and hydrogen number density and initial temperature are set to $n_{\rm
H}=2\times 10^{-4}$ cm$^{-3}$ and $T=8000$ K, except that a spherical
dense clump with a radius of 0.8 kpc located at 1.7 kpc apart from the
center of the simulation volume has a uniform hydrogen number density of
$n_{\rm H,c}=200n_{\rm H}=0.04$ cm$^{-3}$ and a temperature of $T_{\rm
c}=40$ K. The ionizing radiation has the blackbody spectrum with a
tempearature of $T=10^5 {\rm K}$ and constant ionizing photon flux of
$F=10^6$ s$^{-1}$ cm$^{-2}$ at a boundary of the simulation box.
\begin{figure}[htbp]
\centering \includegraphics[width=10cm]{f9.eps}
\caption{Test-3: Maps of neutral fraction of hydrogen in the mid-plane
of the simulation box at $t=1$ Myr, 3 Myr and 15 Myr. The lower and
upper panels show the results with and without the on-the-spot
approximation, respectively.\label{fig:test-3-fmap}}
\end{figure}
\begin{figure}[htbp]
\centering \includegraphics[width=10cm]{f10.eps}
\caption{Test-3: Same as Figure~\ref{fig:test-3-fmap} but shows the gas
temperature maps in the mid-plane of the simulation
box.\label{fig:test-3-tmap}}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=12cm]{f11.eps}
\caption{Test-3: Profiles of hydrogen neutral and ionized fractions and
gas temperature along the axis of symmetry at $t=1$ Myr, 3 Myr and 15
Myr. Solid lines with and without circles indicate the results with and
without the on-the-spot approximation, respectively. Dotted lines shows
the results with RSPH code~\citep{susa2006} presented in
\citet{iliev2006} with the on-the-spot approxiamtion. \label{fig:test-3-prof}}
\end{figure}
Figures~\ref{fig:test-3-fmap} and \ref{fig:test-3-tmap} shows the maps
of hydrogen neutral fraction and gas temperature in the mid-plane of the
simulation volume at $t=1$ Myr, 3 Myr and 15 Myr from left to right,
where the ionizing photons enter from the left boundary of the
figures. We show the results with and without the on-the-spot
approximation in the lower and upper panels, respectively.
At $t=1$ Myr, the ionization front enters the spherical clump and a
cylindrical shadow is formed behind the clump. At $t=3$ Myr and 15 Myr,
the spherical clump is slightly ionized and the boundary of the shadow
is ionized and photo-heated by the hard photons which penetrate the edge
of the clump. These overall ionization and temperature structures with
the on-the-spot approximation are consistent with the ones presented in
\citet{iliev2006}. The effect of the recombination radiation is clearly
seen in the results at 15 Myr, in which the cylindrical shadow is
significantly ionized and heated by the recombination photons emitted at
the ambient ionized region.
Figure~\ref{fig:test-3-prof} shows the profiles of ionized fraction and
gas temperature along the axis of symmetry at $t=1$ Myr, 3 Myr and 15
Myr, where we also plot the results of the \verb|RSPH| code
\citep{susa2006} which are computed with the on-the-spot approximation
and presented in \citet{iliev2006}. Our results with the on-the-spot
approximation are consistent with the ones computed with the \verb|RSPH|
code in \citet{iliev2006}. The effect of the recombination radiation is
siginificant at $t=3$ Myr and 15 Myr in the ionized fraction profiles,
in which the recombination photons accelerate the propagation of the
ionization front in the run without the on-the-spot approximation.
\section{Performance}
In this section, we show the performance of our RT calculations of
diffuse radiation. The code for the transfer of diffuse radiation is
designed so that it can be run both on multi-core CPUs and GPUs produced
by NVIDIA. The performance is measured on the HA-PACS system installed
in Center for Computational Sciences, University of Tsukuba. Each
computational node of the HA-PACS system consists of two sockets of 2.6
GHz Intel Xeon processor E5-2670 with eight cores based on the
Sandy-Bridge microarchitecture and four GPU boards of NVIDIA Tesla
M2090, each of which is connected to the CPU sockets through PCI Express
Gen2 $\times$ 16 link. Thus, a single computational node provides 2.99
Tflops (0.33 Tflops by CPUs and 2.66 Tflops by GPUs) of computing
capability in double precision.
The upper panel of Figure~\ref{fig:time_single} shows wallclock time for
a iteration of the diffuse RT calculation on a single node with various
numbers of CPU cores and GPU boards. The wallclock times are measured
for $N_{\rm m}=64^3$, $128^3$ and $256^3$. The angular resolution
parameter $N_{\rm side}$ is set to $N_{\rm side} = N_{\rm m}^{1/3}/16$
so that $N_{\rm side}\Delta H$ is kept constant. Note that the wallclock
times are nearly proportional to $N_{\rm m}^{5/3}$ as theoretically
expected. The lower panel of Figure~\ref{fig:time_single} shows the
performance gain of the diffuse RT calculation with multiple CPU cores
and GPU boards relative to the performance with a single CPU core and a
single GPU board, respectively. Use of the multiple CPU cores and
multiple GPU boards provides the efficient performance gains nearly
proportional to the adopted numbers of CPU cores and GPU boards for
$N_{\rm m}=128^3$ and $256^3$ except for the fact that those with 16 CPU
cores (2 CPU sockets) is not very impressive even for $N_{\rm m}=256^3$
because of the relatively slow memory access across the CPU sockets. On
the other hand, the performance gain for $N_{\rm m}=64^3$ is somewhat
degraded because of the overheads for invoking the multiple threads and
communication overhead for data exchange between CPUs and GPUs. The
performance with the aid of four GPU boards is nearly 7 times better
than that with 16 CPU cores for $256^3$ mesh grids, while it is only 3.5
times better for $64^3$ mesh grids due to the communication overhead
between CPUs and GPUs.
\begin{figure}
\centering\includegraphics[width=7cm]{f12.eps}
\caption{Wallclock times of diffuse RT calculation with various numbers
of CPU cores and GPU boards for $N_{\rm m}=64^3$, $128^3$ and $256^3$
are shown in the upper panel. A dotted line indicate the dependence of
computational cost on a number of mesh grids, $\propto N_{\rm
m}^{5/3}$. In the lower panel, we present the performance gains of
diffuse RT calculation with multiple CPU cores and GPU boards relative
to the performance with a single CPU core and GPU board,
respectively. Horizontal dotted lines indicates the performance gains
of 2, 4, 8 and 16 from bottom to top. \label{fig:time_single}}
\end{figure}
We compare the performance of our diffuse RT calculations with and
without the ray grouping technique on GPUs. In the implementation
without the ray grouping technique, we utilize the atomic operation
provided by the \verb|CUDA| programming platform in computing the
averaged radiation intensity
(equation~(\ref{eq:averaged_intensity})). Figure~\ref{fig:time_atomic}
shows the performance gains obtained by the use of the ray grouping
technique, where the individual performance is measured with a single
computational node and four GPUs. One can see that the use of the ray
grouping technique significantly improves the performance of diffuse RT
more than by a factor of two irrespective of the number of mesh grids.
\begin{figure}[tbp]
\centering \includegraphics[width=7cm]{f13.eps}
\caption{Performance gains obtained by the use of ray grouping for
$N_{\rm m}=64^3$, $128^3$ and $256^3$ on GPUs. The performances are
measured on a single computational node and four
GPUs. \label{fig:time_atomic}}
\end{figure}
\begin{figure}
\centering \includegraphics[width=7cm]{f14.eps}
\caption{Wallclock times for diffuse RT calculation with various number
of mesh grids computed with 1, 8 and 64 nodes are shown in the upper
panel. The results with and without the use of four GPU boards are
depicted. The dashed line indicates an analytic scaling of
computational cost for the RT calculation of diffuse radiation,
$N_{\rm m}^{5/3}$. \label{fig:time_multi}}
\end{figure}
\begin{figure}[htbp]
\centering \includegraphics[width=7cm]{f15.eps}
\caption{Upper panel: Wallclock times for MPI communication in diffuse
radiation transfer calculations for various number of mesh grids with 8
and 64 nodes. A scaling relation of $N_{\rm m}^{4/3}$ is shown in a
dashed line. Lower panel: fractions of wallclock times for MPI
communication relative to total wallclock time elapsed in the
calculations of diffuse radiation transfer in the runs with the use of
GPUs.\label{fig:time_comm}}
\end{figure}
The upper panel of Figure~\ref{fig:time_multi} shows the wallclock time
of diffuse RT calculation performed on a single and multiple
computational nodes with and without GPU boards, where we invoke one MPI
process on each computational node. In the runs without the use of GPUs,
each MPI process invokes 16 \verb|OpenMP| threads, while in the runs
with the aid of GPUs, we utilize four GPU boards on each computational
node. We measure the wallclock time consumed for a single iteration of
diffuse RT calculation for $64^3$--$1024^3$ mesh grids on 1, 8, and 64
computational nodes. The lower panel depicts the performance gain of the
runs with 8 and 64 computational nodes relative to those with 1 and 8
computational nodes, respectively, where the ideal performance gain of 8
is shown by a dotted line. As for the runs without the use of GPUs, the
parallel efficiency is reasonable when $N_{\rm m}/N_{\rm node} \ge
64^3$, where $N_{\rm node}$ is the number of computational nodes in
use. For a given number of computational nodes, the runs with the use of
GPUs have poorer performance gains than those without it, mainly
beacause the computational times in the runs with GPUs are significantly
shorter than those withtout GPUs, and the MPI data communication, as
well as the commnication between CPUs and GPUs, gets more salient. Such
communication overhead is proportional to the number of light-rays
getting through the surface of the decomposed computational domains,
$\propto N_{\rm m}^{2/3}N_{\rm d}\propto N_{\rm
m}^{4/3}$. Figure~\ref{fig:time_comm} shows that the time consumed by
the MPI communication is nearly proportional to $N_{\rm m}^{4/3}$, and
that it occupies a significant fraction of the total wallclock time for
a small $N_{\rm m}$. This scaling with respect to $N_{\rm m}$ has weaker
dependence on $N_{\rm m}$ than the computational costs, $\propto N_{\rm
m}^{5/3}$. Therefore, the overhead can be concealed for a sufficient
number of mesh grids, and we have better parallel efficiency for a
larger $N_{\rm m}/N_{\rm node}$.
\section{Summary \& Discussion}
In this paper, we present a new implementation of the RT calculation of
diffuse radiation field on three-dimensional mesh grids, which is
suitable to be run on recent processors with highly-parallel
architecture such as multi-core CPUs and GPUs. The code is designed to
be run on both of ordinary multi-core CPUs and GPUs produced by NVIDIA
by utilizing the
\verb|OpenMP| application programming interface and the \verb|CUDA|
programming platform, respectively.
Since our RT calculation is based on the ray-tracing scheme, the RT
calculation itself can be carried out concurrently by assigning the RT
calculation along each light-ray to individual software threads. To
avoid the atomic operations in computing the averaged intensity
(equation~(\ref{eq:path_averaged_intensity})) which can potentially
degrade the efficiency of the thread parallelization, we devise a new
scheme of the RT calculations in which a set of parallel light-rays are
split into 4 groups so that parallel light-rays in each group do not get
through any mesh grids more than once. As well as the thread
parallelization inside processors or computational nodes, we also
parallelize our code on a multi-node system using the MWF scheme
developed by \citet{nakamoto2001}.
We perform several test simulations where the transfer of photo-ionizing
radiation emitted by a point radiating source and recombination
radiation from ionized regions as diffuse radiation are solved. We
verify the validity of our RT calculation of the diffuse radiation by
comparing our results with the effect of recombination radiation and the
ones with other two independent codes, the one-dimensional spherical
code by \citet{kitayama2004} and START code by \citet{hasegawa2010}. We
also clarify the condition of the required angular resolution in our
diffuse radiation transfer scheme based on the mean free path of the
diffuse photons and the mesh spacing.
We show good parallel efficiency of our implementation for intra- and
inter-node parallelizations. As for the intra-node parallelization, the
performance scales well with the number of CPU cores and GPU boards in
use, except for the one in the case that multiple CPU sockets are used
as a single shared-memory system. The scalability of the inter-node
parallelization with the MWF scheme is also measured for $64^3$ to
$1024^3$ mesh grids on up to 64 computational nodes and it is found that
the inter-node parallelization is efficient when we have a sufficient
number of mesh grids per node, $N_{\rm m}/N_{\rm node}\ge 128^3$ and
$N_{\rm m}/N_{\rm node}\ge 64^3$ for the runs with and without GPUs,
respectively. The ray-grouping technique described in
\ref{ss:ray_grounping} is effective and significantly improves the
performance of our RT calculations by a factor of more than two, at
least on GPUs (NVIDIA Tesla M2090).
With our implementation presented in this paper, we are able to perform
the diffuse RT calculations in a reasonable wallclock time comparable to
that of other physical processes such as hydrodynamical
calculations. This means that the calculations of the diffuse radiation
transfer can be coupled with hydrodynamic simulations and we are able to
conduct radiation hydrodynamical simulations with the effect of diffuse
radiation transfer as well as the radiation transfer from point
radiating sources in three-dimensional mesh grids. Currently, we are
developing such a radiation hydrodynamic code and, based on this, we
will address astrophysical problems in which diffuse radiation transfer
plays important roles.
It should be noted that, though we present the implementations and the
performance on the multi-core CPUs and GPUs produced by NVIDIA, our
approaches presented in this paper can be readily applied to other
processors with similar architecture, such as the Intel Xeon-Phi
processor or GPUs by other vendors. In addition, our approach can be
easily extended to adaptively refined mesh grids using the prescription
described in \citet{razoumov2005}, although we present the
implementation for uniform mesh grids in this paper.
\bigskip
We would like to thank Tetsu Kitayama for providing us with his
radiation transfer code. We are also grateful to the anonymous referee
for helpful comments. Numerical simulations in this work have been
carried out on the HA-PACS supercomputer system under the
``Interdisciplinary Computational Science Program'' in the Center for
Computational Sciences, University of Tsukuba. This work was partially
supported by Japan Society for the Promotion of Science (JSPS)
Grant-in-Aid for Scientific Research (S) (20224002). TO acknowledges the
financial support of JSPS Grant-in-Aid for Young Scientists (B:
24740112). KH acknowledges the support of MEXT SPIRE Field 5 and JICFuS
and the financial support of JSPS Grain-in-Aid for Young Scientists (B:
24740114).
|
\section{Introduction}
Quantum chromodynamics (QCD) undergoes a phase transition from the hadronic phase to
the quark gluon plasma (QGP) phase at high temperature.
Recently, a beam energy scan (BES) program at the Relativistic Heavy Ion Collider
has reported valuable data for the long-standing issue of identifying the phase boundary
in the QCD phase diagram by using heavy ion collisions with different collision energies, centralities, etc.
\cite{Aggarwal:2010wy,Adamczyk:2013dal,Adamczyk:2014fia}.
There the probability distribution of conserved charges has been constructed by measuring them for each collision.
Extensive efforts have been invested to understand event-by-event fluctuations of those charges, as
they are expected to be useful observables for locating the critical end point~\cite{Stephanov:1998dy,Hatta:2003wn,Stephanov:2008qz,Asakawa:2009aj,Stephanov:2011pb}.
The setup in BES experiments, where a part of fireballs
is accessible for measurements, resembles a grand canonical ensemble
in statistical mechanics~\cite{BraunMunzinger:2011dn,Morita:2012kt,Garg:2013ata},
and this parallelism enables us to study the probability distribution
of the net baryon number $n_{\rm B}$ theoretically\comment{:}
Consider a grand canonical ensemble for
\comment{a single particle species}.
The grand canonical partition function $Z(\mu)$ is expanded in terms of the number of
particles $n$ as $Z(\mu) = \sum_n Z_n e^{n \mu/T}$.
Here $Z_n$ is a canonical partition function, which depends on temperature $T$
but not on the chemical potential $\mu$.
For given $\mu$ and $T$, $Z_n e^{n \mu/T}$ is proportional to the probability
of observing an $n$-particle state in the grand canonical system.
The BES experiments have so far measured the net proton multiplicity and reported \cite{Adamczyk:2013dal} that
it closely follows the Skellam distribution for several collision energies and centralities.
This observation is consistent with the hadron resonance gas (HRG) model,
in which the net baryon multiplicity is approximately given by a Skellam distribution~\cite{BraunMunzinger:2011dn}.
On the other hand, in lattice QCD simulations the canonical approach
has been proposed as a tool to circumvent the sign problem associated with a finite chemical potential
~\cite{Barbour:1988ax,Barbour:1991vs,Hasenfratz:1991ax,deForcrand:2006ec,Kratochvila:2005mk,Kratochvila:2004wz,Ejiri:2008xt,Li:2010qf,Danzer:2012vw,Nagata:2012tc,Nakamura:2013ska}.
In previous studies~\cite{Nagata:2012tc,Nakamura:2013ska},
two of the authors (K. N. and A. N.) found that
the canonical partition functions follow the Gaussian distribution
of the baryon number at high temperatures and that
the Lee-Yang zeros obtained from the canonical partition functions of the Gaussian type
exhibit a behavior consistent with a Roberge-Weiss (RW) phase transition~\cite{Roberge:1986mm}.
This result has several implications\comment{:}
The connection between the Gaussian behavior of the net baryon number distribution and the RW phase transition
can be used as an experimental probe indicating the QGP phase, since the RW phase transition is a
phenomenon specific to the QGP phase.
The result is also interesting in the context of the Lee-Yang zero analysis~\cite{Yang:1952be,Lee:1952ig},
since the distribution of Lee-Yang zeros is known for some limited cases, e.g.~\cite{Lee:1952ig,Biskup:2000prl1}.
Despite the aforementioned importance, the determination of Lee-Yang zeros
in Monte Carlo simulations is \comment{a} difficult task.
Canonical partition functions suffer from a phase fluctuation configuration by
configuration due to the sign problem.
This problem may be reduced by using sophisticated approaches proposed in e.g.~\cite{deForcrand:2006ec,Li:2010qf}.
However, the phase fluctuation becomes more severe as the baryon number increases,
and the truncation of the fugacity polynomial at a certain order is inevitable.
Such methodological artifacts might sensitively affect the thermodynamic behavior of the Lee-Yang zeros
that are the roots of the truncated polynomial.
The purpose of the present work is to determine the canonical partition functions and
Lee-Yang zeros in QCD at high temperatures and to reexamine their relationships.
To this end, we present an alternative analytical calculation and
assess the reliability of the lattice results in light of the former.
Specifically, we first derive the canonical partition functions and Lee-Yang zeros at high temperature
by utilizing the fact that the free energy is then given as a simple quartic
function of the quark chemical potential.
The canonical partition function is defined as a Fourier integral
of the grand canonical partition function with pure imaginary chemical potential.
At high temperature, this integral can be
evaluated in a saddle point approximation,
yielding the Gaussian function.
Accordingly, the grand canonical partition function is expressed
as a Jacobi theta function.
Using the property of the zeros of the theta function, we show that the Lee-Yang zeros are located on
the negative real axis on the complex plane of the baryon fugacity.
Because of the RW periodicity, these zeros are aligned on three radial lines on the complex plane
of the quark fugacity.
This elucidates the close connection between the Gaussian behavior of the canonical partition functions
and the RW phase transition.
In reexamining the results from the lattice QCD simulations,
some of which have been already presented in ~\cite{Nakamura:2013ska},
we newly address the issue of the convergence of the fugacity polynomial and Lee-Yang zeros,
and perform a bootstrap analysis for their distribution.
We find that the Lee-Yang zeros related to the RW phase transition are not sensitive
to the truncated part of the fugacity polynomial.
We also find an agreement between the lattice data and analytic calculation.
This paper is organized as follows.
In the next section, we explain the canonical approach and Lee-Yang zero theorem.
In Sec. \ref{sec:jtheta}, we explain some features of QCD.
Using those features, we derive the canonical partition functions and Lee-Yang zeros.
In Sec. \ref{sec:simulation}, we compute the canonical partition functions
and Lee-Yang zeros in lattice QCD simulation.
We also discuss implications and reliability of the results.
The final section is devoted to a summary.
\section{Canonical approach and Lee-Yang zeros}
In this section, we explain the canonical approach and
the Lee-Yang zero theorem.
The grand canonical partition function is defined by
\begin{align}
Z(\mu)= {\rm tr}\, e^{ - (\hat{H}- \mu \hat{N})/T},
\end{align}
where $\hat{H}$ and $\hat{N}$ denote the Hamiltonian and the quark number operator in QCD,
and $T$ and $\mu$ the temperature and the quark chemical potential.
We refer to $V$ as the spatial volume of the system.
Using the eigenstates of the number operator, $Z(\mu)$ can be expanded
in powers of fugacity $\xi=e^{\mu/T}$,
\begin{align}
Z(\mu) = \lim_{N\to\infty} \sum_{n=-N}^N Z_n \xi^n.
\label{Eq:2014Mar16eq0}
\end{align}
Here, $Z_n= {\rm tr}( e^{-\hat{H}/T}\,\delta_{\hat{N},n})$
is the canonical partition function at a fixed quark number $n$,
which is the eigenvalue of $\hat{N}$.
$Z_n$ is real and positive for any $n$, and satisfies $Z_n = Z_{-n}$.
$N$ is the maximum number of quarks that can be supported on the system.
The maximum number $N$ is finite on the lattice and diverges in the thermodynamic limit.
\commentb{$Z_n$ is related to the Helmholtz free energy density $f_{\rm H} (n)$ as
$Z_n = \exp(-V f_{\rm H} (n) /T)$~\cite{Kratochvila:2005mk,Ejiri:2008xt,Alexandru:2010yb}, }
and converges in the thermodynamic limit.
By extending
$\mu$ to pure imaginary values,
$\mu=i \mu_I, \mu_I \in \mathbb{R}$, Eq.~(\ref{Eq:2014Mar16eq0})
is regarded as a Fourier expansion of $Z(\mu)$ with the
Fourier coefficients $Z_n$.
The latter can be expressed as the Fourier transformation
\begin{equation}
Z_n
= \int d \theta Z (\theta) e^{i n \theta}
= \int d \theta e^{ - V f(\theta)/T} e^{ i n \theta},
\label{Eq:2014Dec04eq1}
\end{equation}
where $\theta = \mu_I/T$, and we have used $f(\mu)=-(T/V) \ln Z(\mu)$ to
obtain the right-hand side~\footnote{We use Eq.~(\ref{Eq:2014Dec04eq1}) for analytic calculation,
while we a fugacity expansion formula for numerical simulations. See Sec. IV A}.
\commentb{Here $f(\mu)$ denotes the Gibbs free energy density}.
The domain of the Fourier integral is usually $0$ to $2\pi$.
In QCD, we need to take into account the Roberge-Weiss periodicity to determine
the domain of the integral,
as we elaborate in the next section.
For real $\mu$ (i.e.~ real and positive $\xi$),
$Z(\mu)$ can never have zeros
since its coefficient $Z_n$ is real and positive for any $n$.
However, $Z(\mu)$ can have zeros for complex $\mu$.
Using the roots $\xi_i$ of $Z(\mu)$ in the complex $\xi$ plane,
it is expressed in a factorized form\comment{~\cite{Yang:1952be,Lee:1952ig}}
\begin{align}
Z(\mu) = \lim_{N\to \infty} Z_{-N} \xi^{-N} \prod_{i=1}^{2N}
\left(1-\frac{\xi}{\xi_i}\right).
\label{Eq:2014Mar16eq1}
\end{align}
The roots $\{\xi_i\}$ are referred to as Lee-Yang zeros.
Because of the symmetry $Z_n = Z_{-n}$,
any root $\xi_i$ inside the unit circle is accompanied by another root $1/\xi_i$ outside.
In spite of its general importance, it is practically
difficult to obtain Lee-Yang zeros for different models.
Lee and Yang ~\cite{Lee:1952ig} showed that Lee-Yang zeros in Ising models are distributed only on
the unit circle on the complex plane of $e^h$; see Fig.~\ref{Fig:2014Sep21fig1}.
\comment{
To relate Lee-Yang zeros to thermodynamic singularities, it is useful to
recall an electrostatic analogue
proposed by the very founders
\cite{Lee:1952ig}
and later used in the context of QCD
in e.g.~\cite{Blythe:2003aaa,Stephanov:2006dn,Ejiri:2014oka}.
Considering the free energy as an analytic function on the complex $\xi$-plane, we
denote its real part by $\phi \equiv \Re\, f$, which is written as
\begin{align}
\phi(\xi)
= - \frac{T}{V} \sum_{i=1}^{2N} \ln|\xi-\xi_i| - \frac{T}{V} \ln Z_{N} + \frac{NT}{V} \ln |\xi|.
\end{align}
Here the third term comes from a multiplicative factor of the grand canonical partition function,
and is irrelevant to Lee-Yang zeros.
This also provides a constant contribution to the number density because $\ln \xi=\mu/T$, and
is irrelevant to phase transitions.
Therefore, we can safely ignore this term~\cite{Stephanov:2006dn,Blythe:2003aaa}.
Taking the derivatives of $\phi$ with respect to $\xi$, we obtain
\begin{align}
\nabla_\xi^2
\phi(\xi) = - 2\pi \frac{T}{V} \sum_{i=1}^{2N}
\delta^{(2)}
(\xi-\xi_i),
\label{Eq:2015Jan04eq1}
\end{align}
where $
\nabla_\xi
\equiv ( \partial/(\partial\,\Re\,\xi), \partial/(\partial\,\Im\,\xi))$.
We have used
\begin{align}
\nabla^2
\ln |z| = \left( \frac{\partial^2 }{\partial x^2} +
\frac{\partial^2}{\partial y^2} \right) \ln | x+ iy| = 2 \pi \delta(x) \delta(y).
\end{align}
Equation~(\ref{Eq:2015Jan04eq1})
is just the Poisson's equation
for a two-dimensional
electrostatic potential problem; the
real part of the free energy $\phi$ is interpreted as the electrostatic potential,
its derivative $\nabla_\xi \phi$
as the electric field, and Lee-Yang zeros as the location of
charges.
In an electrostatic problem
of charges that
accumulate
e.g.~on a circle,
the electric field is discontinuous across the circle,
while the potential is continuous.
Analogously to this problem, if Lee-Yang zeros accumulate on a curve in the thermodynamic limit,
then the electric field $
\nabla_\xi \phi$ is discontinuous across the curve.
As
$\nabla_\xi \phi$ is proportional to the
(complexified) number density,
the discontinuity in it signals
the first order phase transition
taking place across the curve.
}
\section{Lee-Yang zeros in QCD at high temperature}
\label{sec:jtheta}
\comment{In this section, we derive the Lee-Yang zeros of QCD.
We start by recapitulating
some features of QCD at pure imaginary chemical potential
and of its free energy at high temperature.
Then we calculate canonical partition functions by
using the saddle point approximation,
leading to the analytical determination of Lee-Yang zeros.}
\subsection{QCD at pure imaginary chemical potential}
\comment{Pure SU($N_c$) Yang-Mills theory, such as quenched QCD, has
a center symmetry
$\mathbb{Z}_{N_c},$ where $N_c$ is the number of colors, and $N_c=3$ in QCD.
A
transition from the hadronic phase to the QGP phase occurs at high temperature.
The Polyakov loop defined by
\begin{align}
L(\vec{x})= \frac13\,\tr\, {\rm P} \biggl\{ \exp\biggl(i g \oint_0^{1/T} A_4(\vec{x},\tau) d \tau \biggr)\biggr\},
\end{align}
is related to the free energy of static quarks and can be used
as an order parameter of the deconfinement transition.
Here ${\rm P}$ denotes the path-ordering~\cite{Peskin:1995aaa},
$\vec{x}$ the spatial coordinate,
$\tau$ the imaginary time,
$g$ the gauge coupling constant of QCD,
and $A_4$ the $\mu=4$ component of the gauge field $A_\mu$ in QCD.
$A_\mu$ is an su(3) matrix, and the trace is taken over color indices.
Since the Polyakov loop is $\mathbb{Z}_3$-variant, nonzero values of the Polyakov
loop $\langle L \rangle$ mean that the center symmetry is spontaneously broken, where
the bracket $\langle \cdots \rangle$ denotes the average over gauge fields.
In the deconfinement phase, QCD has three degenerate vacua according to the center
symmetry, and one of them is favored to break the $\mathbb{Z}_3$ symmetry spontaneously.
In the presence of quarks, the center symmetry is explicitly broken.
The deconfinement phase transition turns into a smooth crossover without
discontinuity in thermodynamic quantities.
Nevertheless, the Polyakov loop is still used to distinguish the confinement
and deconfinement phases.
Roberge and Weiss found~\cite{Roberge:1986mm} that even in the presence of the quarks,
the grand canonical partition function of SU($N_c$) gauge theory is invariant under the shift of
$\mu_I=\Im \mu$ as
\begin{align}
Z\biggl( \frac{\mu_I}{T} \biggr) = Z\biggl( \frac{\mu_I}{T} + \frac{2\pi}{N_c} \biggr).
\label{Eq:2014Nov27Eq1}
\end{align}
This states that the grand canonical partition function
is periodic with respect to $\mu_I/T$ with the period of $2\pi/N_c$.
This is referred to as the Roberge-Weiss (RW) periodicity.
Roberge and Weiss also found a first-order phase transition at
$\mu_I/T= (2k+1) \pi/N_c, (k=1, 2, \cdots, N_c)$ under the increase of $\mu_I$ at high temperature~\cite{Roberge:1986mm}.
The argument of the Polyakov loop $\omega = \arg\langle L\rangle$ is often used
as an order parameter of the
phase transition.
In the deconfinement phase and in the presence of quarks with $\mu=0$,
$\omega$ takes zero in the ground state, while the other two
$\mathbb{Z}_3$ vacua are local minima.
As $\mu_I$ is increased at high temperature,
the Polyakov loop jumps from one of
the $\mathbb{Z}_3$ vacua
to another at $\mu_I/T=\pi/3, \pi, 5\pi/3$, and
the argument of the Polyakov loop changes discontinuously;
\begin{align}
\omega = \left\{
\begin{matrix}
0 & (0 \le \frac{\mu_I}{T} \le \frac{\pi}{3}, \frac{5\pi}{3} \le \frac{\mu_I}{T} \le 2\pi), \\
\frac{4\pi}{3} & (\frac{\pi}{3} \le \frac{\mu_I}{T} \le \pi), \\
\frac{2\pi}{3} & (\pi \le \frac{\mu_I}{T} \le \frac{5 \pi}{3}).
\end{matrix}
\right.
\end{align}
This transition is referred to as the
RW phase transition.
Roberge and Weiss found
the presence of the RW phase transition by using a perturbative analysis,
and its absence
at low temperature using a strong coupling analysis.
This was later confirmed non-perturbatively in lattice QCD simulations
for several different setups~
\cite{deForcrand:2002ci,D'Elia:2002gd,D'Elia:2004at,D'Elia:2007ke,D'Elia:2009tm,Cea:2010md,deForcrand:2010he,Nagata:2011yf,Cea:2012ev,Bonati:2014kpa}.
It was also studied in effective models, see e.g.~\cite{Kashiwa:2013rm,Sakai:2008py,Morita:2011jva}.
A nonzero value of $\omega$ means that the gauge field $A_4$ acquires
an expectation value $\omega T/g$.
As we will show in the next subsection, this effect plays a
crucial role in realizing
the RW periodicity of the free energy at high temperature.
In this work, we assume that the system is homogeneous and in equilibrium.
We also assume a priori that
the thermal fluctuation of $\omega$ and $A_4$ is negligibly small.
Under this assumption we fix $A_4$ at its classical value
in a background field method~\cite{Peskin:1995aaa}. }
\comment{Using the RW periodicity,
canonical partition functions $Z_n$ are classified in terms of
the value of $n \; \mod \; N_c$, which is referred to as the triality
for QCD.
We also refer $\{n| n\equiv 0\; \mod \; 3 \}$ as the triality sector,
and the other two sectors, $\{n| n\not\equiv 0 \;\mod \;
3 \} $ as the non-zero triality sector.
Using Eqs.~(\ref{Eq:2014Mar16eq0}) and (\ref{Eq:2014Nov27Eq1}), we can show
~\cite{Kratochvila:2006jx} that
\begin{align}
Z_n = 0,\ n \not\equiv 0
\; \mod\ 3.
\label{Eq:2014Jan08eq1}
\end{align}
This means that only the triality sector contributes to $Z(\mu)$ due to the RW periodicity,
while the nonzero triality sector does not.
Using Eq.~(\ref{Eq:2014Jan08eq1}), the fugacity polynomial Eq.~(\ref{Eq:2014Mar16eq0}) is expressed
in terms of the baryon number instead of the quark number as
\begin{align}
Z(\mu) &= \sum_{
n \equiv 0 \; \mod \; 3 } Z_n \xi^n
= \sum_{n_{\rm B}} Z_{n_{\rm B}} \xi_{\rm B}^n,
\label{Eq:2014Dec31eq1}
\end{align}
where $\xi_{\rm B}=\exp(3 \mu/T)=\xi^3$ and $n_{\rm B} = n/3$.
We rewrite the canonical partition functions $Z_n$ for quark number $n=3n_{\rm B}$
as those for baryon number $Z_{n_{\rm B}}$ to obtain the right hand side.
}
\subsection{Free energy at temperatures above $T_c$}
At high temperature, the free energy density of QCD
eventually approaches a quartic polynomial with even powers of $\mu/T$~\cite{Kapusta:2006aaa,Allton:2003vx}:
\begin{align}
- \frac{f(\mu)}{T^4} = c_0 + c_2 \biggl( \frac{\mu}{T}\biggr)^{2}
+ c_4 \biggl( \frac{\mu}{T}\biggr)^{4}.
\label{Eq:2014Apr24eq1}
\end{align}
The minus sign is conventionally introduced.
\comment{
The question arises as to whether higher order coefficients survive
in cases other than the Stefan-Boltzmann (SB) limit,
such as in the presence of interaction
or at moderate temperature, etc.
Lattice QCD simulations suggested that the free energy indeed approaches
Eq.~(\ref{Eq:2014Apr24eq1}) at temperature slightly
above the pseudo critical temperature $T_c$.
The sixth order term $c_6$ has been calculated in lattice QCD with
different setups: the p4-improved staggered fermions with the bare quark mass $m/T=0.4$
on a $16^3\times 4$ lattice~\cite{Allton:2005gk},
at two different pion masses $m_\pi=220$ and $770$ MeV\cite{Miao:2008sz},
the clover-improved Wilson fermions on $8^3\times 4$ with
the pion mass about $800-1000$ MeV~\cite{Nagata:2012pc}.
A common feature is that $c_6$ rapidly decreases with temperature for $T>T_c$ and
vanishes at a certain temperature.
The vanishing temperature of $c_6$ was estimated to be
$T=(1.1-1.2)T_c$ in~\cite{Allton:2005gk,Miao:2008sz,Nagata:2012pc}.
Usually, Taylor coefficients are calculated using
the so-called noise method~\cite{Ejiri:2009hq},
which associates with a randomly generated vector.
This method becomes less useful
for higher order terms due to the numerical
uncertainty caused by the random noise vector.
On the other hand, in~\cite{Nagata:2012pc}, we used a reduction formula~\cite{Nagata:2010xi}.
The reduction formula provides a method to evaluate the Taylor coefficients
without the random noise vector, although
its applicability has been limited to small lattice volumes.
Using the formula,
the Taylor coefficients was calculated up to tenth order,
and the values of $c_8$ and $c_{10}$ were indeed
consistent with zero at $T>1.1 T_c$ within error bars.
Thus, the lattice QCD simulations are unanimous in supporting
\cite{Allton:2005gk,Miao:2008sz,Nagata:2012pc} that
the free energy takes the quartic form at temperature
$(1.1-1.2) T_c$ or above.}
\comment{
Lattice QCD simulations also showed that $c_2$ is
larger compared to $c_4$ at high temperature.
Below, we will use a saddle point approximation, which requires $c_2$
to be sufficiently large compared to $c_4$.
Here, we estimate the validity range of this approximation.
$c_2$ and $c_4$ have been well studied in lattice QCD simulations,
e.g.~\cite{Allton:2005gk,Miao:2008sz,Ejiri:2009hq,Nagata:2012pc}. They are comparable
in magnitude at $T\sim T_c$, and
$c_2$ ($c_4$) increases (decreases) monotonically
and rapidly as the temperature is raised above $T_c$,
so that $c_2$ surpasses $c_4$
at temperature to some extent higher than $T_c$.
The lattice QCD simulations reported ~\cite{Allton:2005gk,Miao:2008sz,Nagata:2012pc,Ejiri:2009hq}
that $c_2$ is about 10 times larger than $c_4$ at temperatures
$T=(1.1-1.2) T_c$.
As we will see below, this
justifies the use of the saddle point approximation.
Note that $c_2$ and $c_4$
deviate from those at the SB limit in lattice QCD simulations
due to lattice artifacts.
The discussion below is valid if the free energy
satisfies the above two conditions,
$c_6=c_8=\cdots=0$ and $c_2\gg c_4$,
regardless of whether
the SB limit is reached or not.
\subsection{Canonical partition functions and Lee-Yang zeros}
Having explained two features of the free energy of QCD at high
temperature,
we limit our discussion to the case where these conditions hold
and
study the canonical partition functions and Lee-Yang zeros.
First, we extract $Z_n$ by substituting Eq.~(\ref{Eq:2014Apr24eq1}) into Eq.~(\ref{Eq:2014Dec04eq1}).
However, the free energy (\ref{Eq:2014Apr24eq1}) is obtained for $\Im \mu=0$.
As we have remarked,
the argument of the Polyakov loop $\omega$, i.e. the field $A_4$,
acquires a nonzero expectation
value via the RW phase transition as $\mu_I$ is increased.
}
We assume $A_4$ to be constant, and absorb this
condensate into the shift of the imaginary part of the
chemical potential : $\mu + i g A_4 = \mu +i \omega T$.
By taking into account the contributions from \comment{the} three domains,
$Z_n$ is given as
\begin{align}
Z_n &= \int_{-\pi/3}^{\pi/3} e^{-V f(\theta)/T} e^{in\theta} \frac{d \theta}{2\pi} + \int_{\pi/3}^{\pi}
e^{-V f(\theta - 2\pi/3)/T} e^{in\theta} \frac{d \theta}{2\pi} \nonumber \\
& + \int_{\pi}^{5\pi/3} e^{-V f(\theta-4\pi/3)/T} e^{in\theta} \frac{d\theta}{2\pi} ,
\end{align}
where $\theta = \mu_I/T$. Using the RW periodicity~\cite{Roberge:1986mm}, this reads
\begin{align}
Z_n = \int_{-\pi/3}^{\pi/3} e^{-V f(\theta)/T} e^{in\theta} (1 + e^{ -i 2 \pi n /3} + e^{ -i 4 \pi n / 3})
\frac{d\theta}{2\pi}.
\label{Eq:2014Apr22eq1}
\end{align}
This ensures $Z_n = 0$ for $n \not\equiv 0 \ \mod\ 3$,
so that Eq.~(\ref{Eq:2014Apr22eq1}) is expressed as
\begin{align}
Z_n = \frac{3}{2\pi} \int_{-\pi/3}^{\pi/3} d \theta \; e^{T^3 V g(\theta) + i n\theta}\ \ (n \equiv 0 \ \mod\ 3).
\label{Eq:2014Mar16eq3}
\end{align}
where $g(\theta) = c_0 - c_2 \theta^2 + c_4 \theta^4$.
The function $g(\theta)$ has one maximum at $\theta=0$ and two minima at $\theta = \pm \sqrt{c_2/(2 c_4)}$.
\comment{
At high temperature, those two minima are outside of the
integration domain.
As we have mentioned above, $c_2$ and $c_4$ are comparable at $T=T_c$
and $c_2$ rapidly increases and $c_4$ rapidly decreases above $T_c$.
For instance, $\sqrt{c_2/(2 c_4)} = \sqrt{2}\pi$ in the SB limit~\cite{Kapusta:2006aaa,Allton:2003vx,Ejiri:2009hq}
and $\sqrt{c_2/(2 c_4)} \sim \sqrt{5}$ for $T=(1.1-1.2)T_c$
\cite{Allton:2005gk,Miao:2008sz,Nagata:2012pc,Ejiri:2009hq}.
$g(\theta)$ is a concave function for $\theta \in [-\pi/3, \pi/3]$ with the maximum
at $\theta=0$.}
It has a sharper peak for larger volume, and the integral
in Eq.~(\ref{Eq:2014Mar16eq3}) is dominated by $\theta=0$ for large $V$.
This allows for the use of the saddle point approximation
\commentb{(SPA)} to Eq.~(\ref{Eq:2014Mar16eq3}) so that $Z_n$ is reduced to
\begin{align}
Z_n = C e^{ - n^2/(4T^3V c_2)} \ \ (n\equiv 0\ \mod\ 3),
\label{Eq:2014Apr21eq1}
\end{align}
where $C= \frac{3}{2\pi} \sqrt{\frac{\pi}{T^3 V c_2}} e^{T^3 V c_0}$.
\commentb{The quartic term in $g(\theta)$ vanishes according to the saddle point approximation.}
In transition from Eq.~(\ref{Eq:2014Mar16eq3}) to (\ref{Eq:2014Apr21eq1}), we approximated
an incomplete gamma function by a complete counterpart (see the Appendix).
The validity of our approximation is estimated by the condition
$c_2 (\mu/T)^2 \gg c_4 (\mu/T)^4$ so that the free energy is dominated by the second-order term.
To evaluate the applicable range of Eq.~(\ref{Eq:2014Apr21eq2}),
we have plotted
$(f(\mu)-f(0))/T^4$ for the saddle point approximation
and the original one (\ref{Eq:2014Apr24eq1})
in Fig.~\ref{Fig:2014May29fig1}.
It indicates that the
saddle point approximation is valid for small $\mu/T$ $\lesssim 0.5.$
\begin{figure}[htbp]
\includegraphics[width=7cm]{checkspa.eps}
\caption{Comparison of the saddle point approximation (SPA) and exact result
for the free energy density. We use values of $c_2$, $c_4$ and $VT^3$ used
in our simulations in the next section ($c_2=2.20$, $c_4=0.27$\comment{).}}
\label{Fig:2014May29fig1}
\end{figure}
Assuming the validity of Eq.~(\ref{Eq:2014Apr21eq1}), $Z(\mu)$ is reconstructed as
\begin{align}
Z(\mu) &= C \sum_{{n_{\rm B}}=-\infty}^\infty e^{- 9 {n_{\rm B}}^2/(4T^3 V c_2) + 3 {n_{\rm B}} \mu/T}.
\label{Eq:2014Apr21eq2}
\end{align}
The zeros of the grand canonical partition function are readily obtained
by recognizing that Eq.~(\ref{Eq:2014Apr21eq2}) is equal to the Jacobi
theta function $\vartheta(z, \tau) = \sum_{n=-\infty}^{\infty} e^{\pi i n^2 \tau + 2 \pi i n z}$,
\begin{align}
Z(\mu) \propto \vartheta(z, \tau),
\end{align}
where $z$ and $\tau$ are given by
\begin{align}
2\pi i z = 3 \frac{\mu}{T},
\quad
\pi i \tau = - \frac{9}{4T^3 V c_2}.
\end{align}
Thus the Lee-Yang zeros of Eq.~(\ref{Eq:2014Apr21eq2}) are given by the zeros
of $\vartheta(z,\tau)$ located at
\begin{align}
\frac{\mu}{T} = \frac{(2k+1) \pi i}{3} - \frac{3(2 \ell +1)}{4 T^3 V c_2},
\label{Eq:2014Apr28eq1}
\end{align}
where $k$ and $\ell$ take all integer values
as a consequence of the pseudodouble periodicity of the theta function.
On the complex plane of the baryon fugacity $\xi_{\rm B} = \xi^3$,
all zeros in Eq.~(\ref{Eq:2014Apr28eq1}) are located on the negative real axis.
On the complex $\xi$ plane, Eq.~(\ref{Eq:2014Apr28eq1})
lies on three radial lines at arguments
$\arg \xi =\Im (\mu/T) = \pi/3, \pi,$ and $5\pi/3$.
The RW phase transition occurs at the points
$(\Re(\mu/T), \Im(\mu/T)) =(0, (2k+1)\pi/3)$~\cite{Roberge:1986mm}.
The Lee-Yang zeros closest to these points are given by
\begin{align}
\frac{\mu}{T} = \frac{(2k+1) \pi i}{3} \pm \frac{3}{4 T^3 V c_2}.
\label{Eq:2014Mar22eq1}
\end{align}
Each of them approaches the corresponding RW phase transition point
in \comment{the} thermodynamic limit as $1/V$.
This explains the first-order nature of the RW phase transition according
to the Lee-Yang zero theorem.
In addition, Eq.~(\ref{Eq:2014Apr28eq1}) also indicates that the RW phase transition
occurs at $\mu_I/T = \pi/3$ even for ${\rm Re}(\mu/T)\neq 0$,
as long as the saddle point approximation is valid.
We note that it is possible to obtain the Lee-Yang zeros of Eq.~(\ref{Eq:2014Mar22eq1})
directly from the free energy by using a method \comment{presented} in~\cite{Biskup:2000prl1}.
\section{Lattice QCD simulations}
\label{sec:simulation}
\subsection{Method and setup}
In this section, we reexamine the data of our previous lattice QCD simulations,
in which three quantities, the RW phase transition~\cite{Nagata:2011yf},
Taylor coefficients of the free energy~\cite{Nagata:2012pc},
canonical partition functions and Lee-Yang zeros~\cite{Nagata:2012tc,Nakamura:2013ska},
were calculated in the same lattice setup.
Below we recapitulate the calculation of $Z_n$ and Lee-Yang zeros on the lattice,
and summarize the setup of the simulations to make the paper self-contained.
The grand canonical partition function of lattice QCD is given by
\begin{align}
Z(\mu) = \int {\cal D} U (\det \Delta(\mu))^{N_f} e^{-S_g},
\label{Eq:2014Apr22eq2}
\end{align}
where $U$, $\Delta(\mu)$, and $S_g$ denote link variables, fermion matrix, and gauge action, respectively.
We employ a clover-improved Wilson fermion action with $N_f=2$ and renormalization-group
improved gauge action~\cite{AliKhan:2000iz}.
We calculate $Z_n$ using a Glasgow method~\cite{Barbour:1991vs,Hasenfratz:1991ax}.
We expand the fermion determinant in powers of $\xi$ using a reduction formula of the Wilson fermion
determinant~\cite{Nagata:2010xi,Gibbs:1986hi,Hasenfratz:1991ax,Adams:2003rm,Borici:2004bq,Alexandru:2010yb},
\begin{align}
(\det \Delta (\mu))^{N_f} = \sum_{n=-2N_f N_s^3 }^{2 N_f N_s^3} d_n \xi^n,
\label{Eq:2014May03eq1}
\end{align}
which provides the fugacity expansion of $Z(\mu)$.
Since $d_n$ is complex, it is not possible to use $d_n$ as a measure
for Monte Carlo simulations.
Instead we use Ferrenberg-Swendsen reweighting for the fermion determinant\comment{: }
$\det \Delta(\mu) = (\det \Delta(\mu)/\det \Delta(0)) \det \Delta(0)$
and express $Z(\mu)$ as an expectation value of the operator
$\det \Delta(\mu)/\det \Delta(0)$ averaged over gauge configurations generated at $\mu=0$.
Then $Z_n$ is given by
\begin{align}
Z_n & = \int {\cal D}U \frac{d_n}{(\det \Delta(0))^{N_f}} (\det \Delta(0))^{N_f} e^{-S_g}, \nonumber \\
& = Z_0 \left\langle \frac{d_n}{(\det \Delta(0))^{N_f}} \right\rangle_0,
\label{Eq:2014May14eq1}
\end{align}
where $Z_0 = \int {\cal D}U (\det \Delta(0))^{N_f} e^{-S_g}$, and
$\langle \cdots \rangle_0$ denotes the expectation value obtained from
gauge configurations generated at $\mu=0$ with reweighting.
The simulation was performed in the following setup\comment{:}
The lattice volumes were $N_s^3 \times N_t = 8^3 \times 4$ and $10^3\times 4$ with
spatial and temporal lattice sizes $N_s$ and $N_t$.
The simulation was performed along the line of constant physics with
$m_{\pi}/m_{\rho} = 0.8$~\cite{Ejiri:2009hq}.
We considered two temperatures $T/T_c=0.99$\; ($\beta=1.85$) and $1.20\;(1.95)$,
where $\beta =6/g^2$ is the bare lattice coupling constant, and $T_c$ is
the pseudocritical temperature at $\mu=0$.
The RW phase transition point was estimated to be at $\beta=1.92$~\cite{Nagata:2011yf}
and 11 000 HMC trajectories were simulated for each parameter set.
The observables were calculated using 400 configurations with 20-trajectory intervals
after removing the initial 3000 trajectories for thermalization.
Lee-Yang zeros are obtained by using a method based on the Cauchy integral theorem
with a divide-and-conquer algorithm and multiprecision arithmetic~\cite{Nakamura:2013ska,web:FMlib}.
\comment{
We apply the RW periodicity to $Z(\mu)$ and express it in terms of the baryon number,
as shown in Eq.~(\ref{Eq:2014Dec31eq1}).
We solve $Z(\mu)=0$ for Eq.~(\ref{Eq:2014Dec31eq1}), and obtain Lee-Yang zeros for $\xi_{\rm B}$.
They are transformed into
the zeros for $\xi$ by using $\xi=\xi_{\rm B}^{1/3}$.
Thus, obtained Lee-Yang zeros automatically satisfy
the $\mathbb{Z}_3$ symmetry on the complex $\xi$ plane.
}
For further detail, see \cite{Nagata:2012pc,Nagata:2012tc,Nakamura:2013ska}.
\subsection{Canonical partition functions}
\begin{figure}[htbp]
\includegraphics[width=8cm]{taylorcoeff.eps}
\caption{Second- and fourth-order coefficients of Taylor expansion of
the free energy, $c_2$ and $c_4$, for $N_s=8$ and $N_s=16$.
The data for $N_s=8$ and $N_s=16$ are taken from \cite{Nagata:2012pc} and \cite{Ejiri:2009hq}, respectively.}
\label{Fig:2014Mar17fig1}
\end{figure}
The $T$ and $V$ dependences of Taylor coefficients
$c_2$ and $c_4$ are plotted in Fig.~\ref{Fig:2014Mar17fig1}.
At $T/T_c=0.99$, $c_2$ and $c_4$ are comparable in magnitude,
while $c_2$ is several times larger than $c_4$ at $T/T_c=1.20$.
We observed in ~\cite{Nagata:2012pc} that higher-order coefficients $c_6, c_8,$
and $c_{10}$ are consistent with zero at $T/T_c=1.20$ so that $f(\mu)$
approaches the quartic function of $\mu$ as expected in (\ref{Eq:2014Apr24eq1}).
We also observe that the coefficients $c_2$ and $c_4$ are insensitive to
the lattice volume.
Thus, the conditions used in the saddle point approximation are satisfied
at $T/T_c=1.20$ or higher temperatures.
\begin{figure}[htbp]
\includegraphics[width=8cm]{TriZn1850.eps}
\includegraphics[width=8cm]{TriZn1950.eps}
\caption{Canonical partition function as a function of the
baryon number for $T/T_c=0.99$ (top) and $1.20$ (bottom).
The data are obtained from the canonical formalism,
while the solid curves are obtained from the saddle point approximation.
Note that only positive $Z_n$'s are shown.
}
\label{Fig:2014Mar17fig2}
\end{figure}
We plot the canonical partition functions $Z_{n_{\rm B}}$ ($n=3n_{\rm B}$) in Fig.~\ref{Fig:2014Mar17fig2}.
The squares and circles indicate the values obtained from the canonical approach,
Eq.~(\ref{Eq:2014May14eq1}).
Since the average of $Z_n$ can be negative for large $n$ due to the overlap problem,
we plot the values of $Z_{n_{\rm B}}$ up to $\pm n_{\rm B}$ below which
the partition functions are all positive.
The solid curves represent the Gaussian functions (\ref{Eq:2014Apr21eq1})
with $c_2$ obtained from the lattice simulation.
We observe that $Z_{n_{\rm B}}$ with relatively small $n_{\rm B}$
follows the Gaussian function at
$T/T_c=1.20$ as expected, while they fail to match at $T/T_c=0.99$.
\begin{figure}[htbp]
\includegraphics[width=8cm]{deviation1950_1850.eps}
\caption{The difference of the canonical partition functions
between the data $(Z_{n_{\rm B}}^{(\rm lat)})$ and Gaussian fit $(Z_{n_{\rm B}}^{(\rm fit)})$.
Circle and square symbols denote results for $N_s=8$ and $10$, respectively.
Open (red) and closed (blue) symbols are for $T/T_c=1.20$ and $0.99$, respectively.
}
\label{Fig:2014Sep03fig1}
\end{figure}
In order to quantify the consistency, we plot the relative difference between the lattice data and
Gaussian fit in Fig.~\ref{Fig:2014Sep03fig1}.
The data and Gaussian functions show better agreement for higher temperature and larger volume.
However, the Gaussian function systematically deviates from the data for large $n$ even at high $T$.
This deviation is partly caused by the smallness of the lattice volume,
as a better agreement is observed for larger volume.
It is likely that the deviation may originate from the breakdown of the saddle point approximation,
as its validity is limited to small $\mu/T$.
Below, we shall examine how the deviation affects the distribution of Lee-Yang zeros.
\comment{Kratochvila and de Forcrand showed the agreement of the free energy obtained from the canonical approach
and Taylor expansion using a staggered fermion action~\cite{Kratochvila:2005mk,Kratochvila:2006jx}. }
\subsection{Lee-Yang zeros}
In this work we calculate Lee-Yang zeros for $T/T_c=1.20$.
Before proceeding to numerical results, we remark on the
numerical instability of the canonical partition functions at large $n$.
The fugacity coefficients $d_n$ take complex values for each configuration.
The phase of $d_n$ fluctuates more rapidly for larger $n$, because its
modulus is exponentially suppressed as $n$ is increased.
Beyond a certain value of $n$, $Z_n$ becomes negative,
and the inclusion of such $Z_n$ would yield unphysical zeros of $Z(\mu)$ and
cause unphysical nonanalyticity for the free energy, even in a finite volume.
Accordingly, we are obliged to truncate the fugacity polynomial at
$|n_{\rm B}|=n_0$ so that all $Z_{n_{\rm B}}$'s are positive for
$|n_{\rm B}| \le n_0$.
\comment{In the following we will consider three different cases
of truncation:
(a)
$n_0=37$ and $N_s=10$, (b) $N_s=8$ and $n_0=32$ and (c)
$N_s=8$ and $n_0=19$.
For a larger spatial lattice of size $N_s=10$,
it is natural to take
(a) the maximal permissible value $n_0=37$
as seen from Fig.~\ref{Fig:2014Mar17fig2}.
For a smaller spatial lattice $N_s=8$,
we try the following two alternatives:
(b) maximal permissible value $n_0=32$ as seen from Fig.~\ref{Fig:2014Mar17fig2}, and
(c) $n_0=19\simeq 37\times(8/10)^3$ so that the truncation order is proportional to the lattice volume
as compared to the case with $N_s=10$.
}
Below, we examine the convergence of the fugacity polynomial by comparing these two choices.
\begin{figure}[htbp]
\includegraphics[width=8cm]{zerob1950.eps}
\includegraphics[width=8cm]{zerob1950L.eps}
\caption{Lee-Yang zeros on the complex fugacity plane for $\beta=1.95, T/T_c=1.20$.
Top panel: Zeros inside unit circle. Bottom panel: Zeros on
or in the vicinity of the negative real axis.
Red, green, and blue symbols denote Lee-Yang zeros for
(a) $n_0=37$ and $N_s=10$, (b) $N_s=8$ and $n_0=32$, and (c)
$N_s=8$ and $n_0=19$, respectively.
Note that zeros also exist outside the unit circle with symmetry $\xi\leftrightarrow 1/\xi$.}
\label{Fig:2014Mar17fig3}
\end{figure}
Figure \ref{Fig:2014Mar17fig3} shows the distributions of the Lee-Yang zeros on the complex plane of the quark fugacity $\xi$,
corresponding to the cases (a) red, (b) green, and (c) blue, respectively.
The distributions on the baryon fugacity plane are readily obtained by
using the relation $\xi_{\rm B}=\xi^3$.
Near the unit circle, the Lee-Yang zeros are located on three
radial lines with arguments $\arg \xi = \pi/3, \pi,$ and $5\pi/3$.
This behavior is qualitatively consistent with the prediction in Eq.~(\ref{Eq:2014Apr28eq1})
\commentb{and indicative of the RW phase transition.}
As the origin is approached, each line branches to two curves.
\comment{Barbour et
al.~\cite{Barbour:1991vs} obtained
this behavior of the Lee-Yang zeros in
their pioneering study of finite density lattice QCD simulations.
Specifically they found the zeros on the twelve radial lines on the
$e^{\mu}$-plane.
In this work, we take a further step
to confirm this interpretation by examining
the volume scaling and the asymptotic convergence, and
by comparing them with the analytic calculation.}
The zeros near the unit circle $(0.6 < |\xi| <1)$ are stable
as $n_0$ is increased from 19 to 32 for $N_s=8$,
which indicates the convergence of the fugacity polynomial.
On the other hand, the increment of $n_0$ affects the location of the zeros for large chemical potential (small $\xi$).
We also observe that under a shift of $n_0$ as proportional to the volume
($n_0=19$ for $N_s=8$ and $n_0=37$ for $N_s=10$), the Lee-Yang zeros are located on the
common trajectories, and the density of the zeros are doubled.
\begin{figure}[htbp]
\includegraphics[width=8cm]{on_axis.eps}
\includegraphics[width=8cm]{on_axis_scaled.eps}
\caption{Top: $|\Re\, \xi|$ for zeros on the negative real axis near the unit circle,
\comment{where $\xi=\exp(\mu/T)$.}
Red, green, and blue symbols denote Lee-Yang zeros for
(a) $n_0=37$ and $N_s=10$, (b) $N_s=8$ and $n_0=32$, and (c)
$N_s=8$ and $n_0=19$, respectively.
The curves represent predictions from the Jacobi theta function
$\exp(-3(2\ell+1)/ (4T^3V c_2))$ for $N_s=10$ (red) and $N_s=8$ (green and blue).
Bottom: Volume-independent combination $|\Re \, \xi|^{VT^3}$.
The curve represents $\exp(-3(2\ell+1)/ (4 c_2))$.}
\label{Fig:2014Sep03fig2}
\end{figure}
In order to make a quantitative comparison between lattice and analytic results,
in Fig.~\ref{Fig:2014Sep03fig2} we plot $|{\rm Re}\, \xi|$ (top) and $|{\rm Re}\, \xi|^{VT^3}$ (bottom)
for several Lee-Yang zeros near the unit circle on the negative real axis.
Statistical errors of the Lee-Yang zeros in the plot are estimated
with a bootstrap method as follows\comment{: }
For each bootstrap sample, we calculate $Z_n$ up to $n_0$ and locate the Lee-Yang zeros.
Since we are interested in the zeros relevant to the RW phase transition,
we pick up some zeros near the unit circle
and label them as $\ell=1, 2, \ldots$ in the order of modulus.
For each label $\ell$, statistical errors are estimated
as the variance over 1000 bootstrap samples.
Note that the zeros at $|\xi| \lesssim 1$ shown
in Fig.~\ref{Fig:2014Sep03fig2} indicate no
fluctuation in the imaginary part, while the zeros with smaller $|\xi|<0.6$
fluctuate both in their real and imaginary parts.
We observe that each Lee-Yang zero calculated in the simulation is
systematically smaller in magnitude than the zero of the corresponding order
predicted in the saddle point approximation.
In principle, there could be two possible origins for this deviation:
slow convergence of the fugacity expansion (\ref{Eq:2014May03eq1}) and/or
deviation of the canonical partition functions from the Gaussian function.
As there is no systematic difference between
the two choices of the truncation order $n_0=19$ and $32$,
we can exclude the former origin and safely conclude that
the deviation is caused by the deviation of $Z_n$
in the large-$n$ sector from the Gaussian function.
Despite this systematic deviation, the saddle point approximation well explains the features
of the lattice data, such as trajectory of zeros, spacing between zeros, and volume dependence.
\begin{figure*}[htb]
\includegraphics[width=13cm]{Lee-Yangzeros.eps}
\caption{Schematic figures for the distribution of Lee-Yang zeros in
several cases. (a) Ising models on the complex plane for $e^{h}$, where
$h$ is the external magnetic field in Ising models,
(b) QCD on the complex plane for baryon fugacity,
and (c) QCD on the complex quark fugacity plane. Dotted circles denote
the unit circle.
Case (b) can be generalized to free fermion theories.}
\label{Fig:2014Sep21fig1}
\end{figure*}
The result is indicative in the viewpoint of the Lee-Yang zero theorem.
As we have discussed, if the canonical partition function is Gaussian,
then zeros are located on the negative real axis (for the baryon fugacity).
Thus, theories with the Gaussian type of canonical partition functions, such as
a gas of free fermions at small chemical potential, are
exceptional cases of the Lee-Yang zero circle theorem (Fig.\ref{Fig:2014Sep21fig1}).
QCD is expected to be an exceptional case of the Lee-Yang zero circle theorem if it undergoes
a phase transition at $\Re\,\mu \neq 0$.
This can be trivially proven for the case with even number of flavors: Then, the
Boltzmann weight is real and positive on the unit circle on the complex fugacity plane,
and zeros cannot exist on the unit circle.
A concrete example of the Lee-Yang zeros provided in the present work
would help to deepen our understanding of the Lee-Yang zero theorem.
\comment{
We obtained the Lee-Yang zeros as the roots of the fugacity polynomial.
An alternative method
would be to calculate zeros of the grand partition function
with a reweighting method.
However, the validity of the latter is controversial because
Lee-Yang zeros appear when the average of a reweighting factor vanishes.
This seems to imply the breakdown of the reweighting method.
Moreover this problem tends to become severer
for
\comment{larger} volume,
which makes it difficult to distinguish physical zeros from
contamination of the
sign problem~\cite{Ejiri:2005ts}.
The same problem does indeed occur even in the canonical approach;
the sign problem sometimes makes $Z_n$ negative, which allows $Z(\mu)$ to vanish
for real quark chemical potentials even in a finite volume.
Such zeros are unphysical ones
originated from the sign problem.
In this work, we have circumvented this problem by the truncation of the fugacity
polynomial so that all $Z_n$ are positive.
Although the truncated large-$n$ part of the canonical partition functions
suffers from a severe sign problem, we confirmed that the zeros relevant to the RW phase transitions
are insensitive to the truncated terms.
A further question still arises as to how the infinite sum of the truncated terms
affects the location of Lee-Yang zeros, or whether the fugacity polynomial really converges.
We \comment{claim} that Lee-Yang zeros near the unit circle are not affected
even in the limit $n_0 \to \infty$
in which the cutoff is removed.
To support this claim, we estimate the magnitude of the truncated terms for a simple case
where $Z_n$ is well approximated by a Gaussian function $Z_n^{(\rm Gauss)}$ up to $N$;
namely $\delta_n = Z_n - Z_n^{(\rm Gauss)}\ll 1$ for $|n|< N$,
and $\delta_n$ is not negligible for $|n|\ge N$.
Denoting $Z=\sum_n Z_n \xi^n$, and $G=\sum_n Z_n^{(\rm Gauss)} \xi^n$,
the deviation is given by $Z-G \, \simeq\, \sum_{|n|\ge N} (Z_n - Z_n^{(\rm Gauss)}) \xi^n$.
For $\xi$ real and negative, $\xi^n$ is positive for even $n$ and negative for odd $n$.
Our lattice data exhibited in Fig.~\ref{Fig:2014Sep03fig1}
suggests that $\delta_n$ is dominated by $Z_n$, which decreases \comment{monotonically}
as long as there is no phase transition at high temperature.
Then the individual deviation $\delta_n$ monotonously decreases as $n$ increases,
so that the overall deviation is bounded as $|Z-G|\le \delta_N |\xi|^N$ due to the cancellation
between even and odd terms.
The point is that the sum of the truncated terms is bounded by the deviation at $n=N$,
and does not increase as $N$ is increased.
}
\subsection{Discussion}
\label{sec:discussion}
\comment{Our results} have several implications for theoretical and experimental studies.
The connection between the Gaussianity of the canonical partition functions
and the RW phase transition is worth emphasizing\comment{; }
the former can be extracted from an experimentally
measurable quantity, and the latter is a phenomenon specific to the QGP phase.
\comment{
Here we emphasize that the Gaussianity of the canonical partition function
is a sufficient condition, but not a necessary one, for the existence of the RW phase transition; if
the canonical partition function is the Gaussian with regard to
the baryon number, then it implies the existence of the RW phase transition.
However, the converse is not always true; the Gaussian distribution is
merely one type of realization of the RW phase transition.}
Although there have been many studies on QCD at the imaginary chemical potential and the RW phase transition~\cite{D'Elia:2014poslat,deForcrand:2002ci,D'Elia:2002gd,D'Elia:2004at,D'Elia:2007ke,D'Elia:2009tm,Cea:2010md,deForcrand:2010he,Nagata:2011yf,Cea:2012ev,Bonati:2014kpa,Kashiwa:2013rm,Sakai:2008py,Morita:2011jva},
\commentb{we are not aware of any literature presenting a way to relate the RW phase transition to
the canonical partition functions, which can be constructed by measuring number of the baryon in
the system.}
We predict that the canonical partition functions obtained from the probability
distribution of the net baryon number inferred from the multiplicity of baryons
(or three-quark states) created at high temperature
be well approximated by the Gaussian function and associate the RW phase transition.
\comment{Such measurements, however, may be difficult at present
because observed hadrons in heavy ion collisions are generated at the freeze-out temperature.}
This observation, together with HRG,
might serve as a basis to interpret experimental data obtained in BES
experiments and help to distinguish deviations driven by the critical end point.
The distribution of the Lee-Yang zeros indicates that the RW phase transition
persists at $\mu_I/T=\pi/3$ even in the presence of the real part of
the quark chemical potential.
An analytic form of the canonical partition \comment{of} the Lee-Yang zero distribution obtained in this work
can be used as a reference for future finite density QCD studies.
Recently the STAR Collaboration \cite{Adamczyk:2013dal} reported that the net proton multiplicity closely follows
the Skellam distribution for several collision energies and centralities.
The multiplicity of the net baryon number is also approximately given by a Skellam
distribution in the HRG model~\cite{BraunMunzinger:2011dn}.
In \cite{Nakamura:2013ska}, we applied the Lee-Yang zero theorem to
the net proton multiplicity data in BES experiments, and found that
they did not imply the RW phase transition.
This
is consistent with the common understanding that the freeze-out temperature is lower
than the temperature at which the RW phase transition takes place.
However, there may be several controversies in deriving the above conclusion, namely,
the assumption of the equilibrium \comment{or} the use of the net proton multiplicity
as a substitute for net baryon multiplicity~\cite{Hatta:2003wn,Kitazawa:2012at}.
In addition, the probability distribution has so far been measured for a limited
range of the net proton number.
One of the interesting topics is an end point of the RW phase
transition~\cite{deForcrand:2010he,D'Elia:2009qz,Bonati:2014kpa}.
The RW-like behavior appears when the free energy approaches
the quartic function at high temperature,
while it does not at $T\approx T_c.$
In this sense, the RW phase transition is likely an indication of the
completion of the deconfinement transition.
It may be interesting to examine the relation between the RW end point and
canonical partition functions, which may provide us with a possibility to study
the latter experimentally. We leave this problem for future study.
Admittedly, we need to clarify the subtleties involved in numerical
evaluation of canonical partition functions and Lee-Yang zeros.
The present QCD simulation inevitably contains several lattice artifacts
originated from coarse lattices, large quark masses, and small lattice volumes.
However, we consider that the present results \comment{are} robust and \comment{are} likely to
hold for another lattice setup in general,
since the RW-like behavior is based only on a few assumptions,
namely, on the quartic form of the free energy.
Our numerical results suggest the deviation \comment{of} the Lee-Yang zeros from the RW-like
behavior at large $\mu$,
\comment{where the analytic calculation also breaks down.}
We do not understand if this deviation is physical or not
as it lies beyond the applicable range of the present work.
Since one ordinarily expects that there is no phase transition for the quark chemical potential
in the QGP phase,
we conjecture that the behavior smoothly changes from
the RW-like one to
a region in which the $c_4$ term in Eq.~(\ref{Eq:2014Apr24eq1}) dominates.
It may be interesting to investigate whether
there is a nontrivial Lee-Yang zero structure at large $\mu$.
\section{Summary}
In summary, we studied the canonical partition functions and Lee-Yang zeros in QCD at high temperature.
We analytically derived them from the free energy in the
Stefan-Boltzmann limit using the saddle point approximation.
The canonical partition functions in QCD follow the Gaussian function at high temperature
and at small chemical potential.
We pointed out that the grand canonical partition function is approximately expressed
as a Jacobi theta function, which enables us to determine all
Lee-Yang zeros analytically.
These Lee-Yang zeros are located on the negative real axis on the complex plane
of the baryon fugacity. They are translated into three
radial lines on the complex plane of the quark fugacity
owing to the RW periodicity. The zeros exhibit the first-order RW phase transition.
We also performed lattice QCD simulations. To remove numerical subtleties, we
examined the convergence of the fugacity polynomial and performed the bootstrap analysis of the
distribution of Lee-Yang zeros.
The analytic calculations well explain the results obtained from the lattice QCD simulations.
The novelties of the present study are the analytic solution of the canonical partition functions and
Lee-Yang zeros, examination of the convergence of the fugacity polynomial, and bootstrap
analysis for the distribution of Lee-Yang zeros.
Additionally, we pointed out that the gas of free fermions
provides an exceptional case of the Lee-Yang zero circle theorem.
We leave some problems for future studies\comment{: } Namely,
the confirmation of nontrivial
behavior of Lee-Yang zeros observed at large quark chemical potentials
and the determination of the RW end point in the canonical approach
are worth pursuing.
\begin{acknowledgements}
K. N. thanks Sinya Aoki, Teiji Kunihiro, Akira Ohnishi for discussions,
Etsuko Itou for advice on error analysis and analytic calculations,
and Shoji Hashimoto for comments on an early version of the manuscript.
This work is supported in part by JSPS Grants-in-Aid for Scientific Research (Kakenhi)
Grants No.\
00586901 (K. N.),
No. 26-1717 (K. K.),
No. 24340054 and No. 26610072 (A. N.), and
No. 25400259 (S. M. N.).
K. N. is also supported by MEXT SPIRE and JICFuS.
The lattice simulations were mainly performed on SX9 at RCNP and CMC at Osaka University.
The error analysis was done on the HPC system at RCNP.
This work is
also supported by HPCI System Research project (hp130058)
and RICC system at RIKEN.
\end{acknowledgements}
|
\section{Introduction}
Scanned probe imaging is a versatile tool for studying, with high spatial resolution, many interesting physical phenomena (magnetism, surface roughness, conductivity, etc.). The toolkit of scanned probe techniques has been expanded by the use of a scanned proximal probe for imaging and local perturbation in conjunction with simultaneous electrical transport measurements. For example, quantized conductance and universal conductance fluctuations have been mapped by scanned gate imaging \cite{topinka_imaging_2000, berezovsky_imaging_2010}, Kelvin probe microscopy has been used to characterize charge traps in AlGaN/GaN HEMTs \cite{cardwell_nm-scale_2012}, and a Hall cross was used to quantify and map the magnetic field of an MFM cantilever \cite{panchal_magnetic_2013}. One goal of this article is to provide a guide to the fundamental steps necessary to reproduce an SPM with such capabilities. Commercially-available SPM systems facilitate simple and quick sample analysis, providing high resolution and scan rates, but are typically focused on measurements of passive sample properties (topography, e.g.). In order to expand these capabilities to monitor active devices, the measurement sequence must interface with external hardware and allow end-user modification for flexible measurement protocols. Synchronization of imaging and acquisition of transport data is of particular importance. In addition to electrical connectivity for devices, some varieties of transport measurements also require specific environments and functionalities, such as magnetic field application, vacuum, and cryogenic compatibility. Here we describe an integrated, low cost, custom-built system that combines all of these capabilities in an instrument that provides both conventional scanning probe force microscopy and mapping of electrical transport properties as a function of probe position.
The primary challenges for in-operando SPM are the precise positioning of the scanned probe relative to the device, connecting and managing electrical connections to a sample, and flexibility and precision for combining and synchronizing transport and scanned probe measurements and manipulations. To further complicate matters, these tasks may need to be accomplished within the constraints imposed by magnetic field application (the confined space of an electromagnet), in vacuum, and at cryogenic temperatures. Proper cantilever-sample positioning requires: accurately locating and positioning the active area of the device, whose size is often small relative to the total substrate area, determining and controlling scan height, and minimizing vibrational noise and drift (thermal or magnetic field-influenced). Managing electrical connectivity involves enabling relatively easy and reconfigurable wiring of samples in a geometry that provides cantilever access to the device, electrostatic-discharge protection of sensitive devices, and minimization of electrical crosstalk and noise. To permit flexibility of the measurement protocol, an open hardware/software architecture that can be easily modified by the user is necessary. In this article, we present our solutions to these challenges.
A broadly useful system will allow the operator easy access and flexibility with regards to both hardware connections and software processing. This was the primary factor in our selection of the FPGA and LabVIEW-based microscope control solution. The FPGA (Field-Pro\-gram\-ma\-ble Gate Array) provides user-defined reprogrammable processing, while ensuring the fast, deterministic control necessary for scanned probes. Our system relies on open interfaces---industry standard BNC connections and GPIB communications---which allow easy expansion and flexible combination with standard measurement equipment. In-operando imaging could use a variety of inputs as the imaging parameter: cantilever frequency or amplitude, device conductance, Hall voltage, cantilever tip bias (as in Kelvin probe microscopy \cite{girard_electrostatic_2001}), and others. Our hardware/software framework makes it easy to add imaging parameters due to its reprogrammability and LabVIEW's extensive library of third-party device drivers. For example, incorporation of dedicated current and voltage sources and meters is straightforward with GPIB communication. The microscope user can then select the most useful imaging parameter or collect multiple parameters simultaneously.
The accessibility and edit-ability of the software, written in the widely-used LabVIEW programming language, offers the user precise control of measurement sequencing. Scanned probe operations can be synchronized with transport measurements, scan parameters (e.g. tip bias, or scan height) can be tuned in response to transport, and actions or procedures can be easily added to the measurement protocol. This is a shift of focus relative to commercial SPM solutions, which are typically optimized for ease-of-use and provide less software and hardware re-configurability for customized measurement protocols.
Lastly, while home-built SPM controllers often require several custom analog circuits (PID controllers, cantilever phase/frequency detection circuitry), our FPGA/LabVIEW architecture is a complete microscope control solution, requiring no additional home-built circuitry. This also permits much more versatility than hardwired circuits, which may require rebuilding if different measurement parameters are required (for example, interchanging cantilevers with different resonant frequencies). This combination of features and processing tools ensures the flexibility and performance to implement and execute a large variety of measurements.
\section{Experimental Setup}
Figure \ref{fig:MicroscopeDesign} shows various levels of detail of the microscope design. The probe head incorporates a cantilever mounted $10^{\circ}$ from horizontal, fiber interferometric detector for cantilever position sensing, and sample mount with electrical contacts. The sample mount (a patterned printed circuit board) sits atop the scanning hardware---a piezo tube for fine scanning and attocube \cite{attocube} micro-positioners for coarse motion. The entire scan head is mounted on and enclosed by gold-plated, oxygen-free high thermal conductivity copper (to enable cryogenic operation; the gold prevents oxidation of the copper surface). All components of the SPM are made of non-magnetic materials for operation in an external magnetic field (see Fig. \ref{fig:HardDriveMFM}(b)). The scan head measurements were chosen to fit inside a standard electromagnet (maximum pole spacing = 3 in).
\begin{figure*}[htb]
\centering
\includegraphics[width=\linewidth]{MicroscopeDesign_v3.png}
\caption{(a) Microscope scan head. (b) Isometric view of cantilever and electrical sample holder. (c) View from optical window of sample-cantilever alignment. Both lithographic and wirebond leads to device are visible. The cantilever, outlined in red, is also visible. (d) View of entire experiment, including optical table and vibrationally-isolated rotatable electromagnet.}
\label{fig:MicroscopeDesign}
\end{figure*}
At its widest, the scan head measures 1.72 in (43.69 mm) in diameter in order to fit inside the 2 in outer diameter (1.87 in inner diameter) vacuum can. Operation in vacuum reduces damping of the oscillating cantilever, provides a clean environment (especially for surface-sensitive graphene samples), and slows the oxidation of ferromagnetic contacts. The space constraints imposed by the electromagnet and vacuum can however make cantilever-sample alignment more challenging. We therefore use a vacuum can with an optical window. An example of the area viewable through this window is shown in Fig. \ref{fig:MicroscopeDesign}(c). Using a CMOS camera mounted on a Nikon long-working distance optical microscope (SMZ1500), we are able to focus on and monitor the cantilever position relative to the sample during coarse positioning tasks. It is critical to achieve good initial coarse alignment to a device. Acquiring and piecing together many fine scan AFM images in order to locate a device can be very time consuming, especially if the initial scans do not contain obvious device features.
\subsection{Instrumentation}
\subsubsection{Electronics}
Central to the microscope operation is a National Instruments \cite{nationalInstruments} PXI-7851R FPGA and data acquisition card (DAQ). At its front end, this card is responsible for all analog-to-digital (A/D) conversions of the various input signals and digital-to-analog (D/A) conversions of output control signals (see Table \ref{tab:FPGAIO}). The A/D inputs and D/A outputs (8 of each, $\pm$\SI{10}{\volt} range) are digitized with 16-bit resolution, providing \SI{305}{\micro\volt} precision. The inputs are sampled at \SI{750}{\kilo\hertz}; outputs at \SI{1}{\mega\hertz}. In addition to voltage I/O, the FPGA provides high speed signal processing, with a primary clock rate of \SI{40}{\mega\hertz}.
An FPGA is a reprogrammable logic chip. Unlike a CPU, the logic executed by an FPGA is not determined on-the-fly by software. Instead, the desired code is downloaded to the chip, internally rewiring it in order to construct a hardware implementation of the specific logic. This enables the code to execute deterministically and reliably, with no added latency due to variable resource demand (as in software processes executed by a PC CPU). Furthermore, multiple processes can be executed truly in parallel by the FPGA: each process has dedicated circuitry associated with it, and so adjacent processes have no bearing on one another's execution. The FPGA card is housed in a PXIe-1073 chassis, which communicates via a high speed bridge (MXI data link, 250 MB/s) to the host PC's PCI bus. The PXI system was chosen to allow system expandability, as it has slots for four additional DAQ or bus interface cards. Furthermore a dedicated chassis for DAQ cards provides cleaner power and reduces electromagnetic noise as compared to a PCI solution housed in the host PC. To access the voltage I/O of the FPGA, the card connects to an SCB-68 screw terminal connector block via a shielded 68-pin SCSI cable. We installed the terminal block into a custom breakout box which connects the screw terminals to BNC-style panel-mount connectors, enabling easier connectivity. Table \ref{tab:FPGAIO} describes all of the FPGA inputs and outputs.
\noindent
\begin{table*}[htp]
\caption{FPGA Inputs and Outputs}
\begin{tabular}{ | l | l | }
\hline
\textbf{\textit{FPGA Inputs}} & \textbf{\textit{FPGA Outputs}} \\
\hline
AI0 - AC-coupled cantilever interferometer signal & AO0 - self-excitation signal \\
AI1 - DC-coupled cantilever interferometer signal & AO1 - Piezo Tube +X \\
AI2 - available & AO2 - Piezo Tube -X \\
AI3 - available & AO3 - Piezo Tube +Y \\
AI4 - available & AO4 - Piezo Tube -Y \\
AI5 - available & AO5 - Piezo Tube Z \\
AI6 - available & AO6 - Electromagnet control voltage \\
AI7 - available & AO7 - available \\
\hline
\end{tabular}
\label{tab:FPGAIO}
\end{table*}
\subsubsection{Positioning}
For coarse positioning, we use attocube piezo stepper stages ANPx101 (for x and y motion) and ANPz101, each with 5 mm of travel. An attocube ANC-150 controller produces the voltage pulses to move the stages. The piezo tube which we use for fine scanning (EBL \#4, length = 1 in, diameter = 0.25 in, wall thickness = 0.02 in) has a lateral range of \SI{30}{\micro\meter} and a vertical range of $\sim$\SI{3}{\micro\meter} at room temperature, given the $\pm$\SI{500}{\volt} range of our high voltage amplifiers (Trek \cite{trek} 601C). The FPGA card outputs control voltages of $\pm$\SI{10}{\volt} for each of the 5 electrodes of the piezo tube ($\pm$x, $\pm$y, z). The parallelism of the FPGA enables simultaneous control of all piezo tube axes, and the step size resolution (given the 16-bit digital resolution of the D/A) is \SI{0.4}{\nano\meter} in x,y and \SI{0.05}{\nano\meter} in z. Each of the 5 channels is amplified by a high voltage amplifier (gain = 50) in order to bias the piezo tube appropriately. Note that a simple op-amp inverter circuit (unity gain) could be used to generate the control voltages for the $-x$ and $-y$ quadrants of the piezo tube if additional FPGA analog outputs are required.
\begin{figure}[htp]
\centering
\includegraphics[width=\linewidth]{MicroscopeBlockDiag_v2.pdf}
\caption{Block diagram of the microscope and control system. The cantilever position signal from the fiber optic interferometer is fed to the A/D converter of the FPGA. After bandpass filtering, this signal is sent both to the host for further processing (for explanation of frequency detection algorithm, see Sec. \ref{sec:FreqDet} and reference \citenum{obukhov_real_2007}) and to a phase shift function before subsequent output for cantilever self-excitation. The host PC also calculates scanning voltages, which are sent to the FPGA for D/A conversion and output.}
\label{fig:MicroscopeBlockDiag}
\end{figure}
\subsubsection{Cantilever Measurement and Control}
We use a fiber-optic interferometer \cite{rugar_force_1988} to measure the cantilever displacement as a function of time. Such a system has several implementation advantages for low temperature operation (not demonstrated here) and for use in a confined vacuum space. However, any system that measures the cantilever position versus time could be substituted for the interferometry system (four-quadrant position-sensing detector, piezoresistive cantilever, etc.), so long as a voltage signal (representing cantilever position) can be provided to the frequency detection software via the DAQ.
The interferometer photoreceiver voltage is fed to the A/D input of the FPGA, which digitizes at a maximum rate of \SI{750}{\kilo\hertz}. We know by the Nyquist-Shannon theorem \cite{shannon_communication_1949} that the maximum detectable cantilever frequency should therefore be \SI{375}{\kilo\Hz}. In practice, sampling $\sim$10 times faster than the cantilever frequency of $\sim$\SI{75}{\kilo\Hz} is preferred to achieve our desired frequency measurement precision. Additionally, because the input range of the A/D is fixed, it is useful to adjust the amplitude of the cantilever photodiode signal (using a pre-amplifier, e.g. Stanford Research \cite{stanfordResearch} SR560) to maximize use of the $\pm$\SI{10}{\volt} range.
While collecting cantilever position data, the FPGA simultaneously provides a periodic voltage to excite the cantilever at its self-determined resonance (see Sec. \ref{sec:SelfExcite}). This voltage ($\sim$200 mV) is applied to a piezo disc (EBL \#4; 0.25'' diameter, 0.08'' thick) which is mechanically coupled to the cantilever. Figure \ref{fig:ScanningHardware} shows the various electrical connections for cantilever self-excitation and sample scanning.
The accessibility of the microscope signal inputs and outputs (photodiode signal, self-excitation drive signal, etc.) affords straightforward interfacing with external hardware. For example, the user can easily connect the voltage output of the interferometer to a spectrum analyzer and oscilloscope. Alternatively, the LabVIEW environment makes software implementation of such functionality simple.
\begin{figure}[htp]
\centering
\includegraphics[width=\linewidth]{ScanningHardware.pdf}
\caption{Hardware and electronics for cantilever self-excitation, and sample positioning and scanning.}
\label{fig:ScanningHardware}
\end{figure}
\subsubsection{Transport Measurements and Sample Wiring}
For electrical measurements of transport devices, we use a Keithley \cite{keithley} 6221 AC/DC current source and 2425 DC SourceMeter. A Stanford Research SR850 lock-in amplifier provides sensitive lock-in detection. These instruments interface with the PC via GPIB-USB. Depending on the user's preference for stand-alone measurement hardware versus software-based implementation, these instruments could be implemented by appropriate PXI cards and LabVIEW-based software algorithms, such that the entire microscope could be run from a single chassis and host-PC, providing a compact, relatively inexpensive control system. Table \ref{tab:FPGAIO} shows that there are several remaining inputs and outputs available which could also be utilized for these purposes.
Transport measurements of a FET device require reconfigurable electrical wiring (choice of source, drain, gate, voltage probes, etc.). To this end, we have developed a compact sample stage which allows 16 electrical contacts be made to a device without obstructing cantilever access (Fig. \ref{fig:MicroscopeDesign}(b)). This design---made from standard printed circuit board (PCB)---provides sample interchangeability without placing much stress on the piezo tube during mounting and dismounting. The shape of the PCB enables maximum scan range without colliding with the scan head support structure. Electrical contact is made to a transport device by wirebonding from the copper traces on the PCB to the device. The layout of the copper traces was chosen to provide ample clearance between the cantilever and device wirebonds. Wires extend from the PCB and are connected to an interconnect held just above the microscope scan head in the vacuum space. This intermediate interconnect is used for ease of exchanging samples and to provide wire management. From this interconnect, wires run to a vacuum feedthrough (Fig. \ref{fig:MicroscopeDesign}(d)). On the outside of the vacuum can, wires connect the vacuum feedthrough pins to a custom BNC-style breakout box containing the 16 connections. This breakout box makes circuit reconfigurability simple, as the desired pins can be connected to electrical sources and meters, as well as the FPGA I/Os on its home-built breakout box. Each pin is controlled by a toggle switch to connect the device lead to either ground or to the connected instrument. This provides device protection against electrostatic discharge when physically changing the circuit configuration.
An example of a transport measurement obtained with a device mounted in our microscope is shown in Fig. \ref{fig:4ptGDR}. The device-under-test is a graphene field-effect transistor (FET), patterned in a Hall bar geometry (with multiple pairs of Hall contacts). Connecting the device in a conventional four-point measurement scheme, we obtain the Dirac-like dependence of the graphene resistivity versus applied back gate voltage \cite{novoselov_electric_2004}.
\begin{figure}[htp]
\centering
\includegraphics[width=\linewidth]{4ptGDrho.pdf}
\caption{Four point gate-dependent graphene resistivity measurement, acquired in the microscope. Inset: cartoon of measurement configuration.}
\label{fig:4ptGDR}
\end{figure}
\subsection{Software}
The microscope control software package contains two sub-programs, one of which is compiled and downloaded for execution on the FPGA, and a second which is executed on the host computer. The FPGA program provides A/D, D/A, and any time-sensitive, high-speed data processing. This includes interferometer signal digitization, filtering, cantilever self-excitation, and output of piezo tube control voltages. The host program serves as the user-interface and performs tasks that are less time-sensitive, such as cantilever signal processing (frequency and amplitude determination), calculation of piezo tube scan voltages (raster and line scan) and other control voltages, various feedback loops, real-time data display, and on-demand data file saving.
At runtime, to obtain digitized data or send control commands, the host program communicates with the FPGA via DMA (direct memory access) transfers that have been established by the FPGA code. These transfers utilize the MXI data link between the PXI chassis and the PCI bus of the PC. See Fig. \ref{fig:MicroscopeBlockDiag} for an overview of the host PC and FPGA tasks and connectivity. The user does not interact directly with the FPGA program at runtime, except through control variables that have been mapped to the host program. Still, the FPGA program can be modified and recompiled offline if additional functionality or changes to operation are desired.
\subsubsection{Cantilever Self-Excitation}
\label{sec:SelfExcite}
Cantilever self-excitation is a positive feedback method which uses the cantilever's own oscillation to create the signal that feeds back to drive the cantilever. In this manner, the cantilever is always driven at its self-determined resonance frequency, enabling sensitive frequency shift force detection (see Sec. \ref{sec:FreqDet}). Conventional amplitude or phase shift detection of forces, where the cantilever is driven at a fixed frequency, experiences a loss of force sensitivity if large forces shift the cantilever response away from the drive frequency by an amount larger than the cantilever bandwidth ($\Delta f \sim f_0/Q$). There is no risk of this with self-excitation.
The digitization and processing speed of the FPGA make cantilever self-excitation simple and reliable. The interferometer signal representing the cantilever position is digitized, delayed by an integer number of FPGA clock-cycles by a $z^{-n}$ Discrete Delay function in the FPGA code, multiplied by a gain factor, and output as a voltage to drive a piezo disc to which the cantilever is mechanically coupled. (The FPGA clock runs at \SI{40}{\mega\hertz}, enforcing only that the signal be delayed by an integer number of \SI{25}{\nano\second} clock ticks). When the drive voltage sinusoid is $\pi/2$ out-of-phase with the cantilever position, self-excitation is achieved \cite{albrecht_frequency_1991}. The user can control the delay time in order to fine-tune the drive phase relative to the cantilever oscillation until the cantilever amplitude is maximized. The user can also control the drive amplitude, as well as an interrupt time (duration during which the drive is turned off). In addition, this interrupt time can be controlled by active feedback in order to maintain steady cantilever amplitude during scanning and imaging tasks.
\subsubsection{Cantilever Frequency Detection}
\label{sec:FreqDet}
The frequency of cantilever oscillations is affected by tip-sample interaction and can provide topographic, magnetic, and/or electrostatic information. The general form of a force acting on a cantilever is given by ${\bf F} = -{\bf \nabla} U$, where $U$ is the potential energy of the cantilever as a function of its position. The specific form of $U$ will depend on the interaction (topographic, magnetic, electrostatic, etc.). The force(s) in question will shift the resonance frequency of the cantilever according to
\begin{equation}
\delta f = -\frac{f_0}{2k} \frac{\partial F(z_0)}{\partial z}
\end{equation}
\noindent for a cantilever with spring constant $k$, resonant frequency $f_0$, and equilibrium position $z_0$. By detecting this frequency shift as a function of lateral position above the sample, we map the interaction force (although quantitatively extracting the force is difficult). Using cantilever frequency detection as the imaging modality enables fast and sensitive imaging with high Q cantilevers \cite{albrecht_frequency_1991}. There is no need to wait for cantilever ``ring-up,'' as in amplitude detection. The frequency detection algorithm we use works by utilizing the straightforward relationship between a sinusoidal signal and its second derivative; namely that the second derivative is $-\omega^2$ times the original signal.
\begin{equation}
\frac{\partial^2 \mbox{sin}(\omega t)}{\partial t^2} = -\omega^2 \mbox{sin}(\omega t)
\end{equation}
\noindent Therefore, by calculating the second derivative ($\partial^2 z(t)/\partial t^2$) of the input signal ($z(t)$), and fitting a line to $\partial^2 z(t)/\partial t^2$ vs. $z(t)$, the signal frequency can be determined \cite{obukhov_real_2007}.
This algorithm provides a computationally efficient way to detect the first harmonic of a sinusoidal signal. It makes use of the entire time record of the oscillatory signal. This is in contrast to frequency determination by measurement the period between zero-crossings, which discards a majority of the data and is particularly susceptible to noise. Furthermore, because we are interested in only a single frequency component of the signal, this method is much more precise than performing an FFT on the same time record. The frequency resolution of an FFT is $1/T$, where $T$ is the length of the entire signal record in seconds. Therefore, if attempting to obtain \SI{1}{\milli\hertz} frequency resolution, one would need 1000 seconds of data. This is obviously impractical for scanning applications, where each pixel requires a precise frequency measurement. We have achieved sub-\SI{1}{\milli\hertz} resolution of a \SI{75}{\kilo\hertz} signal using approximately \SI{3}{\milli\second} of data.
To test the limits of our frequency detection algorithm, we generated a frequency modulated sine wave using a Stanford Research DS345 function generator. The baseline frequency was set to \SI{75213.833}{\hertz} (on the order of many commercially-available cantilevers). The frequency modulation depth was set to \SI{1}{\milli\hertz} with a modulation frequency of \SI{10}{\hertz}. The input signal is digitized at \SI{690}{\kilo\hertz} (Fig \ref{fig:FreqPSD}(a)), and 2048 data points are used for each frequency measurement, corresponding to a frequency measurement time of 3 ms, suitable for most scanning applications. The input signal is filtered by a bandpass of 20 Hz width around the baseline frequency. Frequency detection results are shown in Fig \ref{fig:FreqPSD}(b,c). The detected frequency vs. time record (panel (b)), clearly demonstrates the imposed frequency modulation. In panel (c)---an FFT of (b)---the 1 mHz modulation ``signal'' rises above the noise floor baseline ($\sim$\SI{10}{\micro\hertz\per\hertz^{1/2}}) with an SNR of nearly 100.
Strictly speaking, the frequency measurement time (\SI{3}{\milli\second}) is shorter than the settling time of the \SI{20}{\hertz} bandpass, resulting in attenuation of the full \SI{1}{\milli\hertz} modulation amplitude. In practice, we find that the improvement in frequency noise with such a strict filter is justified. Depending on the measurement, the user can fine-tune the filtering and the scan rate to optimize frequency noise, scan speed, and spatial resolution. The number of time record data points used for each frequency calculation can also be reduced if necessary and should be determined empirically, depending on the source and magnitude of noise. Note that presently in our imaging experiments, frequency shift sensitivity is limited not by the frequency detection algorithm, but by the thermal noise of the cantilever $\delta f_{\rm{th}} = F_{\rm{th}} f_0/(2 k x_0)$, where $\delta f_{\rm{th}}$ is the thermally-induced frequency noise (in \SI{}{\hertz\per\hertz^{1/2}}). The thermal force noise (in \SI{}{\newton\per\hertz^{1/2}}) is calculated as $F_{\rm{th}} = \sqrt{4 k k_B T / 2 \pi f_0 Q}$. Our system operates in high vacuum ($<$\SI{0.01}{\milli\torr}), at T = \SI{300}{\kelvin}, with a cantilever Q$\approx$10,000 and spring constant $k \sim$\SI{2.8}{\newton\per\meter}. For comparison purposes, we plot $\delta f_{\rm{th}}$ for this cantilever at \SI{300}{\kelvin} and \SI{4}{\kelvin} in Fig. \ref{fig:FreqPSD}(c) (red and blue dashed lines, respectively). For ultra-high-Q cantilevers where the thermal noise floor would lie below the FPGA detection noise floor presented in Fig \ref{fig:FreqPSD}(c), the SNR of the input sine wave would need to be improved (for example, by reducing cable noise). If the input sine wave SNR exceeds 65,536 ($=2^{16}$) the limiting factor becomes the 16-bit digitization noise of the A/D conversion.
\begin{figure}[htp]
\centering
\includegraphics[width=\linewidth]{FreqPSD_v2.pdf}
\caption{(a) Sinusoidal signal at \SI{75213.833}{\hertz} generated by Stanford Research DS345 function generator, digitized and filtered (bandpass width = 20 Hz), showing the 690 kHz sampling rate. (b) Frequency vs. time record of the 75,213.833 Hz signal, showing clear frequency modulation: $\sim$1 mHz oscillations at 10 Hz. Each frequency measurement is calculated from a 2048-data point (3 ms) time record of the sinusoidal signal. (c) Frequency noise power spectral density (Fourier transform) of the frequency vs. time record, showing the large SNR of the modulation signal at \SI{10}{\hertz}---a 1 mHz peak rising well above the noise floor ($\sim$\SI{10}{\micro\hertz\per\hertz^{1/2}}). For comparison, the thermal noise limit for a 75 kHz cantilever, oscillation amplitude 20 nm, Q = 10,000 is shown at 300 K and 4 K.}
\label{fig:FreqPSD}
\end{figure}
The actual code which executes the frequency detection algorithm is shared between the FPGA and host computer. Because it is memory-intensive to store long arrays of high-precision data, the limited resources of the FPGA are not well-suited to executing the entire algorithm. Furthermore, a host PC can perform all necessary calculations with double-precision (64-bit) floating point numbers. This offers immense improvement as compared to the fixed-point and single-precision (32-bit) floating point numbers handled by the FPGA. As such, the host computer performs the bulk of array manipulations.
\subsubsection{Host Code}
The host code provides the graphical user interface (GUI) with which the user primarily interacts. Furthermore, much of the instrument control functionality and data processing are provided by the host. The microscope control software was developed and successfully demonstrated on a standard desktop PC with an Intel Core i7 (2.93GHz) CPU and 8GB of RAM.
In addition to performing the calculation of the cantilever frequency, the host code is also responsible for calculating the piezo tube voltages necessary for sample approach and raster scanning, performing cantilever amplitude and frequency feedback (if desired), plotting the scanned images (frequency vs. position, for example), and performing on-demand file saving. To ensure proper sample mapping during imaging tasks, scanning operations must be properly sequenced with measurements of the imaging parameter (e.g. cantilever frequency). Therefore, LabVIEW queues and notifiers are used extensively for data handling in what is known as a ``producer/consumer'' program design. This also enables parallelism on the host: data acquired in one loop (producer) can be stored in a queue and accessed by a parallel loop (consumer). The consumer loop can then perform data processing tasks that would hinder the processing speed of the producer loop if it had been responsible for acquisition \textit{and} processing. Measurement sequencing is also enforced through the use of interrupts, which are used to notify the host when the FPGA has completed a task (e.g. scanned to a setpoint voltage). In order to showcase the high-degree of control the user has over the scan sequence, we describe our scanning algorithm below and in Fig. \ref{fig:ScanSequence}.
\begin{enumerate}
\item Mode selection - Raster, Line Scan, Manual Positioning, or Field Scan
\item Read Scan Parameters (start, end, step)
\item Calculate estimated scan time
\item Go to initial position (e.g. set the piezo tube voltage to $(x_0, y_0, z_0)$)
\item Begin scan (see Fig. \ref{fig:ScanSequence})
\end{enumerate}
\begin{figure}[htp]
\centering
\includegraphics[width=\linewidth]{ScanSequence_v3_AI.pdf}
\caption{A schematic representation of the LabVIEW scan sequence used for a 2D raster scan. It is simple to add procedures, data collection and manipulations, etc. to the sequence.}
\label{fig:ScanSequence}
\end{figure}
The scanning is handled by a sequence of nested ``for'' loops. The outside loop is responsible for stepping through the slow axis, the intermediate loop selects the trace or retrace scan, and the innermost loop iterates through the fast axis positions. The user can easily incorporate acquisition of the desired imaging parameter into this sequence. After each slow axis step (completion of trace and retrace of the fast axis scan), the results of the trace and retrace line scans are plotted, and the latest scan is added to the composite 2D image. Because a sample is often mounted with an average tilt, the 2D data is also fed to a real-time plane fitting algorithm in order to flatten the acquired image. Using LabVIEW's General Linear Fit function, we calculate the best fit plane to the composite 2D image after each line scan is acquired. This best-fit plane is then subtracted from the raw image data in order to provide a ``flattened'' image. See Fig. \ref{fig:RasterScan_FrontPanel} for an example of the GUI display during image acquisition.
\begin{figure*}[htb]
\centering
\includegraphics[width=\linewidth]{RasterScan_FrontPanel_v6.png}
\caption{The host GUI showing the Raster Scan controls and indicators. Visible in three columns from left to right are the scan \textbf{Settings}, 1D \textbf{Real Time Line Scan} (trace and retrace) graphs, and composite \textbf{2D Image}. The scan settings include the origin; scan size, resolution, angle (rotation of fast and slow axes with respect to the piezo tube x and y axes), and rate; slope correction controls; choice of imaging variable; and estimated scan time indicator. The real time line scan shows each fast axis trace and retrace scan with real time indicators of the applied voltages to the piezo tube, and a percent complete indicator. The 2D composite image updates after the completion of each 1D line scan. The user can view either the raw trace or retrace images, or the ``flattened'' images. A ``Save Raster?'' control for on-demand file saving of the 2D data is seen below the ``Raster Scan'' (engage) and ``Abort'' buttons. Also visible are tabs for other scan modes: Line Scan, Manual Positioning (which sets the piezo tube to a specific location (x,y,z)), and a Field Scan option, which sweeps a control voltage for the electromagnet.}
\label{fig:RasterScan_FrontPanel}
\end{figure*}
\section{Results and Experiments}
\subsection{Standard SPM Operation}
We performed several tests to calibrate the piezo tube motion and demonstrate the performance and capabilities of our microscope.
\subsubsection{Sample approach}
In order to calibrate the piezo tube extension (in nm per applied volt), we record the interferometer DC level as the sample is raised towards and brought into contact with the cantilever. Continued extension of the tube causes upward deflection of the cantilever, and a reduction in interferometer cavity length (distance between cantilever and fiber end). This allows one to observe interferometric oscillations according to $\mbox{sin}((4\pi/\lambda)({\delta z/\delta V})\Delta V)$, where $\lambda$ is the interferometer wavelength (\SI{1550}{\nano\meter}), $\Delta V$ is the applied piezo tube voltage, and $\delta z/\delta V$ is the tube's distance/voltage coefficient \cite{rugar_improved_1989}. A sinusoidal fit (black dashed line) to the approach curve in Fig. \ref{fig:SampleApproach}(a) yields a calibration coefficient of 3.08 nm/V at room temperature.
Locating the sample surface is a critical step to setting up an imaging scan. A coarse tip-sample distance determination is performed as follows: while continuously driving the cantilever, the attocube z motor is stepped upward (towards the cantilever) until the sample comes into gentle contact with the cantilever, causing the oscillations to disappear. The sample is then retracted by less than \SI{1.5}{\micro\meter}. We then perform a piezo tube z scan while monitoring either the cantilever frequency (self-excited cantilever, Fig. \ref{fig:SampleApproach}(b), red), or DC level (undriven cantilever, Fig. \ref{fig:SampleApproach}(b), black). The frequency-measurement approach provides an approximate idea of piezo z voltage necessary to bring the sample into contact with the cantilever. The DC level approach, because it detects cantilever snap down and snap off, provides a more accurate measurement.
\begin{figure}[htp]
\centering
\includegraphics[width=\linewidth]{SampleApproach_v2.pdf}
\caption{Cantilever-sample approach curves. (a) Piezo tube voltage-distance calibration. Approach and retract curves are shown, demonstrating snap-down and off. The broad sinusoidal oscillation is an interferometer fringe, which can be used for piezo voltage-distance calibration, for which we obtain \SI{3.08}{\nano\meter\per\volt}. (b) Comparison of frequency-shift (oscillating cantilever) and DC level (undriven cantilever) approach curves. On a different sample than shown in (a), we acquire frequency shift (red) and interferometer DC level (black) approach traces, taken to establish the sample position, and set a desired scan height. Note that, due to the oscillations of the cantilever, the frequency shift approach experiences snap down sooner than the DC level approach. The frequency measurement becomes inaccurate after snap down has occurred.}
\label{fig:SampleApproach}
\end{figure}
\subsubsection{Non-contact AFM imaging of a calibration grating}
\label{sec:AFM}
We can perform PID control of the cantilever frequency with standard LabVIEW functions, controlling the piezo tube z voltage (extension) with the feedback output. For samples that interact with the tip purely by Van der Waals forces, this imaging mode corresponds to constant tip-sample spacing. We imaged a standard AFM calibration grating (MikroMasch \cite{mikromasch} TGX01) to demonstrate this non-contact AFM capability (see Fig. \ref{fig:CalibrationGrating}). This grating also allows us to calibrate the lateral motion of the piezo scan tube: \SI{28.3}{\nano\meter\per\volt}.
\begin{figure}[htp]
\centering
\includegraphics[width=\linewidth]{CalibrationGrating_v4_AI.pdf}
\caption{(a) Non-contact AFM image of MikroMasch TGX01 AFM calibration grating (\SI{3}{\micro\meter} pitch), with frequency feedback. The tip-sample distance is kept constant by adjusting the piezo z voltage until the cantilever frequency reaches the desired setpoint. This mode is useful for samples with tall features, or when tip-sample contact is not desirable. The image was acquired at a rate of 50 ms/pixel at a tip-sample distance of 73 nm and oscillation amplitude of 47 nm. (b) SEM image of calibration grating \cite{schaefer_calibration}.}
\label{fig:CalibrationGrating}
\end{figure}
This image was acquired at a scan speed of 50 ms/pixel. Since the host PC is responsible for calculating the cantilever frequency, it also handles the frequency feedback. Iterative loops, as are used in PID, execute much more slowly on a host PC than on the FPGA. If the cantilever frequency could be calculated on the FPGA (by use of an FPGA card with greater processing capability than the 7851R), feedback could also be executed on the FPGA itself. This would result in much faster feedback and scanning.
\subsubsection{MFM imaging of a hard disk drive}
With a magnetically-coated Bruker \cite{brukerAFM} MESP cantilever ($f_0$ = \SI{79198}{\hertz}), we can detect sample magnetization. Figure \ref{fig:HardDriveMFM}(a) shows a frequency-shift image for a magnetic hard disk drive, showing two tracks of magnetic bit data. Because magnetic forces can be either attractive or repulsive, the simple frequency feedback mode demonstrated in Sec. \ref{sec:AFM} cannot be used with magnetic samples. This image was therefore taken at nominally-fixed piezo z voltage. As a result, some topographic signal may mix into such an image if the tip-sample distance changes during scanning. A more sophisticated MFM imaging algorithm (e.g. a lift mode) could be incorporated into our software. Presently, we do incorporate sample tilt correction, which adjusts the piezo z voltage as a linear function of both x and y position. This tilt can be calibrated by performing sample approaches at three different (x,y) positions.
\begin{figure}[htp]
\centering
\includegraphics[width=\linewidth]{HardDriveMFM_v3_AI.pdf}
\caption{(a) Frequency-shift image of hard-disk drive showing two tracks of data (base cantilever frequency = 79,198 Hz, tip-sample distance = 150 nm, oscillation amplitude = 126 nm). (b) Frequency-shift image of hard-disk drive at zero applied magnetic field (top) and at 500G (bottom). The magnetization of the cantilever coating has reversed direction (at ~\SI{400}{\gauss}) causing a reversal of the frequency-shift contrast colors.}
\label{fig:HardDriveMFM}
\end{figure}
The microscope was designed to operate in external magnetic fields. We use a GMW \cite{GMW} 5403 electromagnet atop a rotating stage, in order to provide magnetic field in any in-plane direction (see Fig. \ref{fig:MicroscopeDesign}(d)). The magnet current is supplied by a Kepco \cite{kepco} 15 V/20 A BOP 15-20M voltage-controlled bipolar operational power supply. In Fig. \ref{fig:HardDriveMFM}(b), we show a region of the hard drive sample imaged at \SI{0}{\gauss} and \SI{500}{\gauss}. Cantilever magnetometry \cite{stipe_magnetic_2001} data (not pictured) show that the cantilever magnetization undergoes switching at about \SI{400}{\gauss}. Such a magnetization reversal will change the sign of the force exerted on the cantilever by an unchanged sample magnetization. As a result, the color contrast of the frequency shift data in the \SI{0}{\gauss} and \SI{500}{\gauss} images are inverted. A dashed line is shown as a guide to the eye to show a particular feature where this is evident.
\subsection{In-Operando Imaging: Electrostatic Force Microscopy}
\label{sec:EFM}
We acquired scanned electrostatic force microscopy (EFM) images of an electrically-biased graphene field effect transistor (CVD graphene on SiO2(300nm)/n-Si) in order to demonstrate the ability to perform integrated scanning and current-voltage measurements. In an EFM measurement, the cantilever frequency is shifted by the capacitive interaction between tip and sample \cite{girard_electrostatic_2001}.
\begin{equation}
\delta f = - \frac{1}{2} \left( \frac{d^2 C}{dz^2} \right) V^2
\end{equation}
\noindent where $C$ is the capacitance of the tip-sample system, and $V$ is the potential difference between tip and surface.
\begin{figure*}[htb]
\centering
\includegraphics[width=\linewidth]{EFM_v5_AI.pdf}
\caption{EFM images of a biased graphene hall cross. (a) \SI{100}{\micro\ampere} applied with left contact at $V = +V_0$, right contact at ground. (c) \SI{100}{\micro\ampere} applied with right contact at $V = +V_0$, left contact at ground. Panel (b) shows a the line-cut indicated by white dashes in (a), while (d) represents the line-cut from (c).}
\label{fig:SKFPM}
\end{figure*}
In Fig. \ref{fig:SKFPM} we show the measured cantilever frequency shift for two configurations of the graphene hall cross: electrical current flowing left to right, or right to left. The cantilever is sensitive only to voltage differences between itself and the sample. The cantilever (which has a conductive chromium coating) is grounded, so the grounded electrical contact causes a negligible frequency shift. By contrast, the biased electrical contact shows a dramatic 2 Hz frequency shift relative to the cantilever's natural frequency. By incorporating a closed-loop Kelvin probe microscopy KPM technique, where the cantilever bias $V_{\rm{probe}}$ is adjusted in order to null the sample's contact potential difference $V_{\rm{CPD}}$ (instead of using a grounded cantilever, as presently), the sample's surface potential (in units of volts) could be directly measured. The technique can also be used to characterize the quality of contact to graphene \cite{yu_tuning_2009}. Even without a KPM controller, Fig. \ref{fig:SKFPM} makes evident the voltage drops due to contact resistance between the gold electrodes and the graphene. Voltage drops in the graphene itself are also visible, particularly in the narrow regions, but appear smaller than those from the contact resistance due to the $V^2$ dependence of $\delta f$. Three-point current-voltage measurements of the gold/graphene contacts found contact resistances of \SI{11.5}{\kilo\ohm} and \SI{10.3}{\kilo\ohm} for the left and right contacts, respectively. The total two-point resistance between the pair was found to be 36.6 k$\Omega$, leaving the graphene channel with a \SI{14.8}{\kilo\ohm} resistance. Because the graphene channel extends far above and below the image ($\sim$\SI{100}{\micro\meter}), current spreading is also observed. The inhomogeneity in local voltage requires further investigation.
We have also acquired images under varying back gate conditions, as shown in Fig. \ref{fig:GateDependence}. The gate potential exerts a force on the cantilever, shifting its frequency. The image contrast is provided by shielding of the applied back gate by the grounded device (graphene and gold contacts), above which the cantilever experiences almost no frequency shift. The force on the cantilever does not depend on the sign of the voltage difference (since it is proportional to $V^2$). For example, an attractive force (negative frequency shift) is evident above the unshielded gate for both \SI{-10}{\volt} and +\SI{15}{\volt}. Because of the long range of electrostatic forces as compared to topographic forces, these images suggest a means for helping to locate a device \cite{li_self-navigation_2011}.
\begin{figure}[htb]
\centering
\includegraphics[width=\linewidth]{GateDependence_v3_AI.pdf}
\caption{EFM images of the same graphene hall cross as Fig. \ref{fig:SKFPM}, under varying gate bias. The force on the cantilever is sensitive only to $V^2$, and insensitive to the sign of the gate bias. When held above the graphene (and gold electrodes), the cantilever is shielded from the applied back gate, and experiences relatively little frequency shift. This technique could be used to easily locate a device, since the electrostatic force is fairly long-ranged \cite{li_self-navigation_2011}. All images, with the exception of $V_G$ = \SI{15}{\volt}, were taken at a nominal scan height of 340 nm. The cantilever was retracted to a height of 970 nm in order to acquire the image with $V_G$ = \SI{15}{\volt}, due to the increased electrostatic force causing tip-sample contact.}
\label{fig:GateDependence}
\end{figure}
\section{Further Improvements}
Because of the flexibility of the software and hardware we are using, it would be straightforward to expand the imaging and measurement modes of this instrument beyond cantilever force detection. For example, using a biased conducting tip, the device conductance could be monitored as a function of tip position (scanned gate imaging) \cite{topinka_imaging_2000, topinka_coherent_2001, berezovsky_imaging_2010}. Alternatively, the Hall voltage of a device such as that used in Fig. \ref{fig:SKFPM} could be monitored in order to quantify the stray magnetic field of an MFM tip \cite{panchal_magnetic_2013}. We plan to use the microscope to create a spin map of a lateral spin valve by monitoring the effect of a magnetic cantilever tip on the device's non-local voltage, as in scanned spin-precession imaging \cite{bhallamudi_imaging_2012, bhallamudi_experimental_2013}. Additional measurement protocols could be added in order to perform sensitive measurements. For example, the cantilever oscillations could be used as a lock-in reference, while monitoring the device voltage. Again, the modularity of the PXI chassis, the reprogrammability of the FPGA and control software, and the ease of interfacing LabVIEW with external instrumentation will make such measurements possible and relatively easy to implement.
\section{Conclusion}
Scanned probe microscopy plays a central role in micro- and nano-scale characterization of samples and devices. Combining scanned probes with operational devices is not a capability well-supported by commercial SPM solutions. The variety of transport measurements and effects that can be studied puts a premium on reconfigurability. LabVIEW places measurement control in the hands of the experimenter, enabling nearly endless combinations of scanning and transport measurements and protocols. FPGA-based operation is a perfect fit for imaging tasks because of its speed and deterministic hardware execution. As such, the combination of FPGA and LabVIEW enables low-cost, versatile in-operando SPM solutions. Using this platform, we have demonstrated fast and accurate frequency shift detection and imaging, straightforward and reliable incorporation of transport measurements, and the flexibility to pursue unique and innovative measurement schemes for investigation of electronic and spintronic systems.
\section{Acknowledgments}
This research was supported by funding from the Center for Emergent Materials: an NSF MRSEC, Award Number DMR-1420451 and by the Army Research Office, ARO Award No. W911NF-12-1-0587. Technical support was provided by the NanoSystems Laboratory at OSU.
\bibliographystyle{unsrtnat}
|
\section{Introduction}
Two new models piece together a story of the evolution of the Solar System. They are the Grand Tack and Nice models.
The Grand Tack model addresses the formation of the inner Solar System (Walsh et al 2011; see also Morbidelli et al 2012; Raymond et al 2013; O'Brien et al 2014). It invokes an inward-then-outward migration of Jupiter and Saturn consistent with hydrodynamical simulations (Masset \& Snellgrove 2001; Morbidelli \& Crida 2007). The change in the direction of migration, or "tack", is set to occur when Jupiter is at 1.5 AU, so that the sculpted inner disk of planetary embryos can reproduce the terrestrial planets, in particular the large Earth/Mars mass ratio (Wetherill 1991; Hansen 2009; Raymond et al 2009).
The Nice model proposes that after the gas disk dispersal, the outer Solar System underwent a delayed instability triggered by interactions between the giant planets and an outer disk of planetesimals. In the original version of the model, the giant planets started on an arbitrary but more closely-packed orbital configuration. The instability was triggered by Jupiter and Saturn crossing their mutual 2:1 resonance (Tsiganis et al 2005). In more recent incarnations, the giant planets' orbital configuration was sculpted by an earlier phase of gas-driven migration (Morbidelli et al 2007), with initial conditions that are consistent with the Grand Tack scenario. Angular momentum exchange between the giant planets and the planetesimal disk can act to extract the giant planets from a resonant chain and trigger the instability (Levison et al 2011).
The Nice model has been invoked to explain the Late Heavy Bombardment (Gomes et al. 2005). It can reproduce the current architecture of the giant planets (Nesvorny 2011; Nesvorny \& Morbidelli 2012), the orbital distribution of Jupiter's Trojan asteroids (Morbidelli et al. 2005;
Nesvorny et al. 2013), the irregular satellites of Jupiter (Nesvorny et
al. 2014), Saturn, Uranus and Neptune (Nesvorny et al. 2007), and the structure of the Kuiper belt in broad strokes (Levison et al. 2008, Batygin et al. 2012).
Jupiter and Saturn's gas-driven migration in 3:2 resonance has been confirmed by several studies (Masset \& Snellgrove 2001; Morbidelli \& Crida 2007; Pierens \& Nelson 2008; Pierens \& Raymond 2011). Jupiter and Saturn naturally converge into 3:2 resonance and then migrate outward. However, each of these studies used an isothermal disk with a standard value for the aspect ratio $h\sim 0.05$.
In this paper we present the outcome of hydrodynamical simulations of the planets embedded in gaseous disks that are subject to viscous heating and radiative cooling. We show that in relatively low-mass ($M_{disk} \lesssim M_{MMSN}$, where MMSN is the minimum-mass solar nebula model of Hayashi 1981) and low-viscosity (viscous stress parameter $\alpha = 10^{-4}$ or $10^{-5}$; see Shakura \& Sunyaev 1973) disks, Jupiter and Saturn can migrate outward in 2:1 resonance. This arises because for a purely viscously heated disk, a low-disk mass and/or a small viscosity give rise to a very small disk aspect ratio $h\sim 0.02-0.03$, which tends to favor outward migration. The advantage of having Jupiter and Saturn in 2:1 rather than in 3:2 resonance is that it can increase the success rate of the Nice model instability. Nesvorny \& Morbidelli (2012) indeed showed that if Jupiter and Saturn start in a 2:1 resonance,
there is a better chance that their period ratio jumps to the correct value.
Our paper is structured as follows. In Sect. 2, we describe the hydrodynamical model and discuss the results
of our simulations in Sect. 3, and the effects of stellar heating on these results in Sect. 4. Nice model simulations that
study the evolution after disk dispersal are presented in Sect. 5. Finally, we draw our conclusions in Sect. 6.
\begin{figure}
\includegraphics[width=0.8\columnwidth]{bilan2.eps}
\caption{ Results of the $16$ simulations ({\it symbols}) we performed as a function of the $\alpha$ viscous stress parameter
and the normalized initial disk surface density at $1$ AU.}
\label{fig:bilan}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.33\textwidth]{ae32.eps}
\includegraphics[width=0.33\textwidth]{p32.eps}
\includegraphics[width=0.33\textwidth]{angle32.eps}
\includegraphics[width=0.33\textwidth]{ae21.eps}
\includegraphics[width=0.33\textwidth]{p21.eps}
\includegraphics[width=0.33\textwidth]{angle21.eps}
\includegraphics[width=0.33\textwidth]{aeiso.eps}
\includegraphics[width=0.33\textwidth]{iso-period.eps}
\includegraphics[width=0.33\textwidth]{anglesiso.eps}
\caption{{\it Upper panel:} time evolution of the planets' semi-major axes $a$, perihelia $q$ and aphelia $Q$
(left) for a disk model with $f=3$ and $\alpha=10^{-4}$. The middle panel
shows the evolution
of the period ratio and the right panel the evolution of the resonant angles
$\theta_1=3\lambda_S-2\lambda_J-\varpi_S$ (black) and $\theta_2=3\lambda_S-2\lambda_J-\varpi_J$ (red)
associated with the 3:2 resonance, where $\lambda_J$ ($\lambda_S$) and $\varpi_J$ ($\varpi_S$) are the
longitude and pericenter of Jupiter (Saturn).
{\it Middle panel:} Same but for a disk model with $f=0.3$ and $\alpha=10^{-4}$. The middle right panel shows the evolution of the resonant angles
$\theta_1=2\lambda_S-\lambda_J-\varpi_S$ (black) and $\theta_2=2\lambda_S-\lambda_J-\varpi_J$ (red)
associated with the 2:1 resonance.
{\it Lower panel:} Same but for an isothermal run with $\alpha=10^{-4}$ and $h=0.02$.}
\label{fig:at}
\end{figure*}
\begin{figure}
\includegraphics[width=\columnwidth]{hsr.eps}
\includegraphics[width=\columnwidth]{den.eps}
\caption{{\it Upper panel:} Disk aspect ratio as a function of radius for the models with $f=0.3$, $\alpha=10^{-4}$ (black) and $f=3$,
$\alpha=10^{-3}$ (red). Here Jupiter and Saturn are held on circular orbits with $a_J=2$ and $a_S=3.4$ AU.
{\it Lower panel:} Surface density profile for the same models.}
\label{fig:dh}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.33\textwidth]{a21js.eps}
\includegraphics[width=0.33\textwidth]{ej.eps}
\includegraphics[width=0.33\textwidth]{es.eps}
\caption{Time evolution of Jupiter and Saturn semi-major axes (left panel), Jupiter's eccentricity (middle panel) and
Saturn's eccentricity (right) for simulations that led to the formation of a stable 2:1 resonance.}
\label{fig:models21}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{nice.eps}
\includegraphics[width=0.8\columnwidth]{nice2.eps}
\caption{{\it Upper panel:} Time evolution of planets' semi major axes for a five-planet Nice model run
with initial configurations resulting from the disk model with $f=0.3$, $\alpha=10^{-4}$. The five planets
were started in a (2:1, 3:2, 3:2, 3:2) resonant chain and the mass of the planetesimal disk is $m_{disk}=20$ $M_\oplus$.
{\it Lower panel:} Period ratio between between Jupiter and Saturn as a function of time.}
\label{fig:nice}
\end{figure}
\section{Hydrodynamical model}
Simulations were performed using the GENESIS (De Val-Borro et al. 2006) numerical code that solves the equations governing the disk evolution on a polar grid $(R,\phi)$. The code's energy equation reads:
\begin{equation}
\frac{\partial e}{\partial t}+\nabla \cdot (e{\bf v})=-p(\nabla\cdot{\bf v})+Q^+_{visc}-Q^-_{rad}
\label{eq:energy}
\end{equation}
where $\bf v$ is the gas velocity, $e$ the thermal energy density, $\gamma$ the adiabatic index (set to $\gamma=1.4$).
$p=(\gamma -1) e$ is the pressure which is related to the disk temperature $T$ and surface density $\Sigma$ as
$p={\cal R} \Sigma T/\mu$, where ${\cal R}$ is the ideal gas constant and $\mu=2.35$ the mean molecular weight. $Q^+_{visc}$ is the viscous heating term where viscous stresses are modeled using the standard 'alpha' prescription for the disk viscosity
$\nu=\alpha c_s H$, where $c_s=(\gamma {\cal R} T/\mu)^{1/2}$ is the sound speed and $H$ the disk scale height which is
related to the angular velocity $\Omega$ and the isothermal sound speed $c_{s,iso}=c_s/\sqrt{\gamma}$ as
$H=c_{s,iso}/\Omega$.
$Q^-_{rad}=2\sigma_B T_{eff}^4$ is the local radiative cooling term,
where $\sigma_B$ is the Stephan-Boltzmann constant and $T_{eff}$ is the effective temperature which
is computed using the opacity law of Bell \& Lin (1994). In this work, effects resulting from stellar heating
are not considered in the energy budget.
We employ $N_R=896$ radial grid cells uniformly distributed between $R_{in}=0.25$ and $R_{out}=12$ AU, and $N_\phi=700$ azimuthal grid cells. At the inner edge, we use a viscous outflow boundary condition (see Pierens \& Nelson 2008). At the outer edge, we employ a wave-killing zone for $R>11.4$ to avoid wave reflections.
The initial surface density profile is
$\Sigma=f\;\Sigma_{MMSN} ( R/1 AU)^{-3/2}$,
where
$\Sigma_{MMSN}=2\times 10^{-4}$ in dimensionless units is the surface density at $1$ AU of the MMSN, and $f$ is
an enhancement factor. The initial temperature profile is such that $T\propto R^{-1}$. Due to the action of source terms in Eq. $1$, however, this initial temperature quickly evolves until an equilibrium state is reached.
Jupiter and Saturn initially evolve on circular orbits with semi-major axes $a_J=2$ and $a_S=3.4$ AU respectively, just exterior to their mutual $2:1$ resonance. To give them sufficient time to open a gap, the planets are held on fixed circular orbits for $\sim 500$ orbits, and then are released and evolve under the action of disk torques. When calculating the disk torques, we exclude the material contained within a distance $0.6R_H$ from the planets
using a Heaviside filter (Crida et al. 2009) , where $R_H$ is the Hill radius.
In our simulations we tested the effect of two parameters: the disk's surface density and viscosity. We tested $\alpha=10^{-5}, 10^{-4},10^{-3}, 10^{-2}$, and for each value of $\alpha$ we performed four different simulations with $f=0.3, 1, 3, 10$, for a total of 16 simulations.
\section{Results}
The results of our simulations are illustrated in Fig. \ref{fig:bilan}. There are four qualitatively different outcomes
shown with different symbols. Runs that are labelled "3:2 resonance" (downwards triangles) produced outward migration of Jupiter and Saturn in a 3:2 resonance (Masset \& Snellgrove 2001; Morbidelli \& Crida 2007). The top panels of Fig. \ref{fig:at} show one such outcome, for the simulation with $\alpha=10^{-4}$ and $f=3$. The planets migrate convergently and become trapped in a 2:1 resonance at $t\sim 100$ orbits (see top middle panel of Fig. \ref{fig:at}). Due to Saturn's fast migration, the planets break free from the 2:1 resonance and become trapped in a $3:2$ resonance at $t\sim 300 $ orbits. From that time on, Jupiter and Saturn evolve in a common gap and the
resulting change in the torque balance, together with the flux of gas from the outer disk across the gap, makes them migrate outward. The 3:2 resonance is maintained throughout (resonant angles shown in the top right panel of Fig. \ref{fig:at}).
The middle panels of Fig. \ref{fig:at} show the evolution of a simulation with $f=0.3$ and $\alpha = 10^{-4}$. Jupiter and Saturn were captured in 2:1 resonance. This triggered outward migration that was maintained to the end of the simulation. As for the 3:2 resonance, outward migration was maintained by a continuous flux of gas across the planets' common gap. This flux acts to replenish the inner disk and also to exert a positive corotation torque on the planets. The main difference between the two simulations illustrated in Fig.~\ref{fig:at} is that the planets' eccentricities are much higher migrating outward in 2:1 resonance. This occurs because the gap is wider and deeper in the simulation with the gas giants in 2:1 resonance, resulting in a weaker damping of the planets' eccentricities.
Jupiter and Saturn should be caught in stable 2:1 resonance when Saturn's migration is slower than a critical rate. The critical velocity for capture in the 2:1 resonance $\sim (M_J/M_*)^{4/3} a_S \Omega_S$ (D'Angelo \& Marzari 2012), where $\Omega_S$ is the Saturn's angular velocity. Assuming an isothermal type I migration rate for Saturn (Tanaka et al. 2002) and that Jupiter does not migrate, D'Angelo \& Marzari (2012) found that Jupiter and Saturn would become trapped in 2:1 resonance if the surface density at $2$ AU is $\lesssim 650$ $g.cm^{-2}$. Although we use a non-isothermal
disk model here, Fig. \ref{fig:bilan} shows that our simulations are consistent with this estimate, since all runs with $f=0.3$ formed a stable 2:1 resonance. An alternative possibility, which is the one that actually happens in the
simulations, is that both the disk aspect ratio and viscosity are small enough for Saturn undergo a slow Type II migration (Lin \& Papaloizou 1993). In that case, capture in 2:1 resonance is expected provided that the condition for gap opening is satisfied (Crida et al. 2006):
\begin{equation}
1.1\left(\frac{q_S}{h^3}\right)^{-1/3}+\frac{50\nu}{q_Sa_S^2\Omega_S}<1
\end{equation}
where $q_S=M_S/M_*$ is the Saturn mass ratio. For a thin disk with $h=0.03$, $\alpha \lesssim 3\times 10^{-3}$ whereas
for $h=0.05$, capture in the 2:1 resonance should arise for $\alpha \lesssim 4\times 10^{-4}$. This estimate also agrees with our simulations since runs with disc masses corresponding to the MMSN and $\alpha \leq 10^{-4}$ produced a stable 2:1 resonance.
When viscous stresses are the only heating process, a low disk mass (small value of $f$) and a modest viscosity (small $\alpha$) produces a thin disk. Figure \ref{fig:dh} (top panel) shows the aspect ratio as a function of radius for two disks; a low-mass, low-viscosity disk with $f=0.3$, $\alpha=10^{-4}$ and a high-mass, modest-viscosity disk with $f=3$, $\alpha=10^{-3}$. Here, Jupiter and Saturn are held on circular orbits at $2$ and $3.4$ AU, respectively (exterior to 2:1 resonance). In the high-mass disk the aspect ratio is $h\sim 0.04$ ($h\sim 0.05$) at the location of Jupiter (Saturn). In contrast, in the low-mass disk $h\sim 0.02$ ($h\sim 0.03$) at Jupiter's (Saturn's) orbit. For the low-mass, low-viscosity disk, the gaps opened by the planets are wider and deeper than in the high-mass disk (see bottom panel of Fig. \ref{fig:dh}). In the low-mass, low-viscosity disk the gas near Saturn's orbit is significantly depleted, imparting a strong positive torque on Jupiter which may favor outward migration (Morbidelli \& Crida 2007).
A small aspect ratio thus appears to be the key factor in causing outward migration with Jupiter and Saturn in 2:1 resonance. In principle, outward migration should also occur in isothermal disk models with small aspect ratios. To test
this hypothesis, we performed a simulation using an isothermal equation of state with $h=0.02$ and $\alpha=10^{-4}$
and which
indeed resulted in outward migration with Jupiter and Saturn in a 2:1 resonance, as illustrated in the lower panel of
Fig. \ref{fig:at}.
Fig. \ref{fig:models21} shows the results of radiative simulations that produced stable 2:1 resonances. For a MMSN model, the condition for outward migration is $\alpha\lesssim 10^{-4}$ whereas for $f=0.3$, outward migration occurs for $\alpha \lesssim 10^{-3}$. This is consistent with the expectation that the aspect ratio in radiative disks should be a function of the
product $ \nu \Sigma$ (Bitsch et al. 2014). Outward migration tends to be faster for smaller $f$ and smaller $\alpha$. The run with $f=0.3$ and $\alpha=10^{-5}$, which exhibits a smaller outward migration rate than in the case with $\alpha=10^{-4}$, is an exception; this probably occurs because the inner edge of Jupiter's gap is located further away from the planet in that case.
\section{Effect of stellar irradiation}
Results from the provious section suggest that the 2:1 resonance can lead to outward migration in disks with very small aspect ratios $h\sim 0.02$. For a radiative disk subject to viscous heating only, low values of the aspect ratio correspond to small disk masses and/or viscosities and therefore to low accretion rates. In that case, stellar irradiation may contribute significantly to the disk temperature structure and possibly dominate over viscous heating, but this depends strongly on the disk
metallicity (Bitsch et al. 2014). Recent hydrodynamical simulations of
the structure of disks with constant accretion rate (Bitsch et al. 2014) show that in the late stages of evolution, a metallicity of $0.05$ can lead to an aspect ratio $h\sim 0.03$ in the inner parts, and it is expected that even smaller values for $h$ to
be reached for lower metallicities. Moreover, 1D modeling of viscous protoplanetary disks subject to
stellar irradiation (Bailli\'e \& Charnoz, in prep.) also shows that at very late times, the aspect ratio can be typically $h\sim 0.02$ at $5$ AU. Both of these studies therefore suggest that the conditions for Jupiter and Saturn to migrate outward
in 2:1 resonance may be fulfilled under certain conditions. It remains to be seen, however, whether outward migration of Jupiter and Saturn in 2:1 resonance triggered in a part of the disk with a low aspect ratio could be maintained if the planets migrated into an outer, thicker part of the disk. Clearly, more sophisticated hydrodynamical simulations are needed to definitely assess the effect of stellar heating on these results.
\section{Evolution after gas disk dispersal}
We now turn our attention to the later evolution of the outer Solar System. Our goal is to test whether simulations showing outward migration of Jupiter and Saturn in 2:1 resonance are compatible with the current architecture of the Solar System.
We selected three disk models that produced outward migration in 2:1 resonance at different rates (with $f=0.3$, $\alpha=10^{-4}$; $f=0.3$, $\alpha=10^{-3}$ and $f=1$, $\alpha=10^{-5}$). We restarted each simulation and artificially dissipated the disk by forcing the gas surface density to decay exponentially with an e-folding time $t_{dis}=10^4$ yr. We then used the outputs of these runs, rescaled such that the initial semi-major
axis of Jupiter was $\sim 5$ AU, as initial conditions for N-body simulations of the evolution of the outer Solar System after dispersion of the gas disk.
Following Nesvorny (2011) and Nesvorny \& Morbidelli (2012), we performed simulations with Uranus and Neptune initially located in a resonant chain with the giant planets. The resonant chain is generated using both hydrodynamic
and N-body simulations to identify the resonant configurations that are compatible with Jupiter
and Saturn in 2:1 resonance. We also considered a five-planet case where an additional planet with mass comparable to that of Uranus/ Neptune was placed between the orbits of Saturn and Uranus. In each case, the orbit of one planet was purposely shifted by $180^\circ$ in mean anomaly to artificially disrupt the resonant chain and trigger an instability (see Levison et al 2011).
An outer planetesimal disk was included with masses of $m_{disk}=20, 35, 50, 100 \;M_\oplus$. For each case, we performed $30$ simulations with different, randomly-generated, radial ranges and surface density profiles of the planetesimal disk.
In the four-planet case, the initial states deduced from our three disk models did not produce good Solar System analogs. Low disk masses typically led to final systems with fewer than four planets. For high disk masses the ice giants migrated too far and the giant planets end up on too-circular orbits due to too strong disk-induced eccentricity damping (Nesvorny 2011). We obtained better results in simulations with an extra ice giant. This is broadly consistent with Nesvorny \& Morbidelli (2012), who found a much higher success rate in reproducing the current architecture of the Solar System in the five-planet case. This was true for Jupiter and Saturn in both 3:2 and 2:1 resonance.
Figure \ref{fig:nice} shows the results of a successful five-planet run with initial configuration from the model with $f=0.3$, $\alpha=10^{-4}$. The planets were started in a (2:1, 3:2, 3:2, 3:2) resonant chain and the planetesimal disk mass was $m_{disk}=20$ $M_\oplus$. The extra ice giant (in blue) was ejected from the system after $\sim 6$ Myr, and the subsequent migration of the giant planets was marginal. This run satisfies most of the constraints from Nesvorny \& Morbidelli (2012). The final semi major axes of the planets are $5.20$, $9.33$, $18.01$ and $28.40$ AU and their final eccentricities are $0.029$, $0.045$, $0.058$ and $0.03$. Neptune is a little too close to the Sun and Jupiter's eccentricity is slightly smaller than its current value, but the lower panel of Fig. \ref{fig:nice} shows that the final Jupiter-Saturn period ratio is almost perfect. When the ice giant is ejected at $\sim 6$ Myr, the Jupiter-Saturn period ratio jumps from $(a_S/a_J)^{1.5}/<2.1$ to $(a_S/a_J)^{1.5}/> 2.3$ in $< 1$ Myr such that the $\nu_5$ secular resonance jumps, rather than sweeps, across the inner Solar System (Brasser et al 2009; Morbidelli et al 2010; called constraint D by Nesvorny 2011 and Nesvorny \& Morbidelli 2012). In this example the secular amplitude ${\it e_{55}}$ is also close to the current value.
Although we did not perform a broad statistical study as in Nesvorny \& Morbidelli (2012), the existence of successful cases such as the one from Fig.~\ref{fig:nice} shows that the outward migration of Jupiter and Saturn in 2:1 resonance is consistent with the late evolution of the outer Solar System.
\section{Discussion and Conclusion}
We have shown that a Grand Tack can occur with Jupiter and Saturn in 2:1 resonance for a limited range of the parameter space defined by the disk's mass and viscosity. Outward migration in 2:1 takes place in relatively low-mass ($M_{disk} \lesssim M_{MMSN}$), low-viscosity ($\alpha \lesssim 10^{-3}$) disks that tend to have very small aspect ratios ($h \sim 0.02$). Compared with outward migration in 3:2 resonance, the biggest difference in the evolution when Jupiter and Saturn are in 2:1 resonance is that their eccentricities are higher, with $e_{J,S} \approx$0.05-0.2.
Outward resonance of Jupiter and Saturn in 2:1 resonance is consistent with a later, Nice model instability. We used the simulations in which the gas giants migrated outward in 2:1 resonance as inputs to N-body simulations of the evolution of the outer Solar System after dispersal of the gas disk. Instabilities including an outer planetesimal disk produced good Solar System analogs, matching most of the criteria derived in Nesvorny (2011). Therefore, a scenario in which Jupiter "tacked" at $\sim 1.5$ AU when Saturn caught up and was trapped in a 2:1 resonance may also
explain the evolution of both the inner and outer Solar System.
Outward migration in 2:1 resonance is possible in disk models that have a very small aspect ratio, typically
$h\sim 0.02-0.03$ at the location of Jupiter. For such a value of $h$, the disk is substantially depleted at Saturn's orbit, creating a strong positive torque exerted on Jupiter and favoring outward migration. This result was confirmed by locally isothermal runs which resulted in a similar outcome for $h=0.02$ and $\alpha=10^{-4}$ . We note that more
realistic protoplanetary disks models subject to
stellar irradiation are thought to have similar aspect ratios at late stages, especially if their
metallicity is small (Bitsch et al. 2014). This suggests that a Grand Tack scenario with Jupiter and Saturn
in 2:1 resonance is possible in evolved protoplanetary disks, but this needs to be checked with 3D
radiative hydrodynamical simulations.
\acknowledgements
Computer time for this study was provided by HPC resources of Cines under allocation c2014046957 made
by GENCI (Grand Equipement National de Calcul Intensif). We thank the Agence Nationale pour la
Recherche under grant ANR-13-BS05-0003-002 (project MOJO).
|
\section{Introduction}
Current understanding of the role of the Luminous Blue Variable (LBV) stage in the evolution
of the most massive stars is very limited. Important questions are how the eruptions of several
$\rm M_{\odot}$ are triggered, and what role binarity plays in the properties of the LBV
stage and in shaping the nebulae observed around LBVs. For example, is most of the
mass lost by an LBV star due to a steady radiatively-driven stellar wind, or is the
H-rich envelope removed by punctuated eruption-driven mass loss? There are two
confirmed LBV binaries: $\eta$ Car in the Galaxy, and HD~5980 in the SMC.
The extremely luminous star MWC 314 has recently been recognized ($P_{\rm orb}$ = 60.8 d and $e$ = 0.23) as
an eclipsing massive binary system \citep{2013AA...559A..16L}. Indications that MWC 314 is
a (dormant) LBV are strong.
The star is very luminous \citep*[log $L$/$\rm L_{\odot}$ = 5.9; see also][]{1998AAS..131..469M}
with an optical spectrum and SED nearly identical to the
canonical LBV P Cygni. MWC 314 is surrounded by a bipolar H$\alpha$
nebula that may be the result of an eruption of MWC 314 more than
100,000 years ago \citep{2008AA...477..193M}.
MWC 314 is a single-lined, eclipsing binary system with masses of 40$+$26 $\rm M_{\odot}$, and
radii of 87$+$20 $\rm R_{\odot}$. The largest diameter of the binary system is $\sim$1.1 AU.
\begin{figure}
\begin{center}
\includegraphics[width=6.cm, height=12cm, angle=270]{Lobel_MWC314_HeI_3.ps}
\caption{{\bf Left-hand panel}: orbital radial velocity curve of MWC 314. Two orbital eclips phases B and D when the primary is in front of the secondary ($\phi$$\sim$0), and two quadrature phases A and C ($\phi$$\sim$0.8) are marked. {\bf Right-hand panels}: the profiles of three He~{\sc i} lines are shown for orbital phases marked A, B, C, and D. DACs are observed in the violet wings of the lines ({\it marked with arrows}) during the orbital phases B and D. XMM-Newton observed X-rays on 6 May 2014 (D).
}
\label{fig1}
\end{center}
\end{figure}
\section{Radial velocity curve and DACs in He~{\sc i} P Cyg lines}
We measure the radial velocity curve in Fig. 1 from selected absorption
lines of the primary star. Two spectra marked with A and C are observed in the
quadrature phase close to minimum radial velocity of the primary
(0.75 $<$ $\phi$ $<$ 0.85). Two spectra observed when the primary is in front of the
secondary are marked with B and D (0.95 $<$ $\phi$ $<$ 1.05). Three He~{\sc i} lines reveal
P Cyg profiles with wind expansion velocities to $\sim$1200 $\rm km\,s^{-1}$.
On 5 May 2012 (B) we observe two discrete absorption components in the
extended violet wings of the three lines. The DACs are observed at wind
velocities of $\sim$100 $\rm km\,s^{-1}$ and $\sim$180 $\rm km\,s^{-1}$
in the line profiles ({\em marked with arrows}).
The DACs are not clearly observed in the quadrature phases of 29 June 2010 (A)
and 24 June 2012 (C). We observe four DACs on 6 May 2014 (D) at wind
velocities between $\sim$100 $\rm km\,s^{-1}$ and $\sim$600 $\rm km\,s^{-1}$.
\section{Conclusions}
Our high-resolution spectroscopic monitoring program of MWC 314
with Mercator-HERMES during the last 5 years reveals DACs in the orbital phases when
the primary is in front of the secondary star on 5 Sep 2009, 5 May 2012, and 6 May 2014.
The high-excitation temperatures of the He~{\sc i}
lines signal expanding wind regions of enhanced density and variable outflow velocity.
The DACs can form close to the primary's surface (of B0-type) in high-temperature and
density-enhanced wind regions in front of its orbit. The recurrence of the DACs
in orbital phases when the primary is in front of the secondary can result from wave
propagation which is physically linked to the orbital motion near the (low-velocity $<$ 150 $\rm km\,s^{-1}$)
wind base. Alternatively, the DACs can form in dynamical (high-temperature) wind regions
confined between the binary stars at the shock interface of a colliding wind region. We expect
orbital X-ray variability in MWC 314 (similar to the LBVs $\eta$ Car and HD~5980) during
planned XMM-Newton observations in the quadrature phase of Oct. 2014.
\bibliographystyle{iau307}
|
\section{Introduction}
Convolutional neural networks (ConvNets) are attracting much attention largely due to their impressive empirical performance on large scale visual recognition tasks (c.f.~\cite{krizhevsky2012imagenet, zeiler2014visualizing, sermanet2013overfeat, taigman2013deepface, szegedy2014going}). The ConvNet architecture has the capacity to model large learning problems that include thousands of categories, while employing prior knowledge embedded into the architecture. The ConvNet capacity is controlled by varying the number of layers (depth), the size of each layer (breadth), and the size of the convolutional windows (which in turn are based on assumptions on local image statistics). The learning capacity is controlled using over-specified networks (networks that are larger than necessary in order to model the problem), followed by various forms of regularization techniques such as Dropout~(\cite{hinton2012improving}).
Despite their success in recent years, ConvNets still fall short of reaching the holy grail of human-level visual recognition performance. Scaling up to such performance levels could take more than merely dialing up network sizes while relying on prior knowledge to compensate for what we cannot learn. It may be worthwhile to challenge the basic ConvNet architecture, in order to obtain more compact networks for the same level of accuracy, or in other words, in order to increase the abstraction level of the basic network operations.
A few observations have motivated our work. The first is that the ConvNet architecture has not changed much since its early introduction in the 1980s~(\cite{lecun1995convolutional}) -- there were some attempts to create other types of deep-layered architectures (cf.~\cite{riesenhuber1999hierarchical, bruna2013invariant, poon2011sum}), but these are not commonly used compared to ConvNets. Arguably, the empirical success that ConvNets have witnessed in recent years is mainly fueled by the ever-growing scale of available computing power and training data, with the contribution of algorithmic advances having secondary importance. Our second observation is that although there were attempts to use unsupervised learning to initialize ConvNets (c.f.~\cite{hinton2006fast, bengio2007greedy, vincent2008extracting}), it has since been observed that these schemes have little to no advantage over carefully selected random initializations that do not use data at all (see for example~\cite{krizhevsky2012imagenet,zeiler2014visualizing,sermanet2013overfeat}). We nevertheless believe that unsupervised initialization has an important role in scaling up the capacity of deep learning, and therefore find interest in deep architectures that give rise to natural initializations using unsupervised data. The third observation that motivated our work is that the ConvNet learning paradigm completely took over classification engines developed in the 1990s like Support Vector Machines (SVM) and kernel machines in general. These machine learning methods were well suited for ``flat'' architectures, and while attempts to apply them to deep layered architectures have been made~(\cite{cho2009kernel}), they did not keep up with the performance levels of the layered ConvNet architecture. It may be beneficial to develop a deep architecture that includes the body of work on kernel machines, but which still has the capacity to model large learning problems like the ConvNet architecture.
In this paper we introduce a new family of layered networks we call SimNets (similarity networks). The general idea is to ``lift'' the classical ConvNet architecture into something more general, a multilayer kernel network architecture, which carries several attractive features. First, the architecture bridges the decades-old ConvNets with the statistical learning machinery of the last decade or so. Second, it provides a higher level of abstraction than the convolutional and pooling layers of ConvNets, thus potentially providing more compact networks for the same level of accuracy. Third, the architecture is endowed with a natural initialization based on unlabeled data, which also has the potential for determining the number of channels in each layer based on variance analysis of patterns generated from the previous layer. In other words, the structure of a SimNet can potentially be determined automatically from (unlabeled) training data.
The SimNet architecture is based on two operators. The first is analogous to, and generalizes, the convolutional operator in ConvNets. The second, as special cases, plays the role of ReLU activation~(\cite{nair2010rectified}) and max pooling in ConvNets, but in addition, has capabilities that make SimNets much more than ConvNets. In a set of limited experiments on CIFAR-10 dataset~(\cite{krizhevsky2009learning}) using a small number of layers, we achieved better or comparable performance to state of the art ConvNets with the same number of layers, and the specialized network studied in~\cite{coates2011analysis}, using 1/9 and 1/5, respectively, of the number of learned parameters.
In the following sections, we introduce the two operators that the SimNet architecture comprises, and describe its special cases and properties. The experiments section is still preliminary but demonstrates the power of SimNets and their potential for high capacity learning. Additional experiments with deeper SimNets are underway, but those require extensive optimization and coding infrastructure in order to apply to large scale settings.
\section{The SimNet architecture } \label{sec:simnet_arch}
The SimNet architecture consists of two operators -- a ``similarity" operator that generalizes the inner-product operator found in ConvNets, and a soft max-average-min operator called MEX that replaces the ConvNet ReLU activation~(\cite{nair2010rectified}) and max/average pooling layers, and allows additional capabilities as will be described below.
The similarity operator matches an input $\x\in\R^d$ with a template $\z\in\R^d$ and a weight vector $\uu\in\R_+^d$ ($\R_+^d$ stands for the non-negative orthant of $\R^d$) through $\uu^\top\phi(\x,\z)$, where $\phi:\R^d\times\R^d\to\R^d$ is a similarity mapping. We will consider two forms of similarity mappings: the ``linear'' form $\phi(\x,\z)_i=x_i z_i$, such that when setting $\uu=\1$ we obtain an inner-product operator, and the ``$l_p$'' form $\phi(\x,\z)_i=-\abs{x_i-z_i}^p$ defined for $p>0$.
In a layered architecture, a similarity layer is illustrated in fig.~\ref{fig:layers_nets}(a), where the similarity operator is applied to patches $\x_{ij}\in\R^{hwD}$ of width $w$, height $h$ and depth $D$, with the indexes $(i,j)$ describing the location of the patch within the layer's input. Given $n$ templates $\z_1,...,\z_n\in\R^{hwD}$ and weights $\uu_1,...,\uu_n\in\R_+^{hwD}$, the layer's output at coordinates $(i,j,l)$ becomes $\uu_l^\top\phi_l(\x_{ij},\z_l)$, where we use index $l$ in $\phi_l$ to indicate that the similarity mapping may differ across channels. As customary with ConvNets, the width and height of the layer's output depends on the ``stride'' setting, which determines the step-size between input patches, e.g. with horizontal and vertical strides of $s$, the spatial dimensions of the output become $\lfloor(H-h)/s\rfloor+1$ and $\lfloor(W-w)/s\rfloor+1$. Note that using the linear-similarity mapping with unit weights($\uu_l=\1$) reduces the similarity layer to a standard convolutional layer where $\z_l$ are the convolution kernels, whereas for $l_p$-similarity with fixed $p=2$, the output at coordinates $(i,j,l)$ measures the weighted Euclidean (Mahalanobis) distance between the input patch $\x_{ij}$ and the template $\z_l$ with every pixel weighted through the entries of the weight vector $\uu_l$. When using $l_p$-similarity in general, fixing the order $p$ is not obligatory -- the order can be learned based on training data, either globally or independently for each output channel.
We will see later on that, when setting unit weights, the (unweighted) linear and $l_p$ similarity mappings correlate with kernel-SVM methods of statistical machine learning (through special cases of the SimNet architecture), and that the view of $\z_l$ as templates allows natural unsupervised initialization of networks using conventional statistical estimation methods.
The MEX operator, whose name stands for Maximum-minimum-Expectation Collapsing Smooth (with ``CS'' pronounced as ``X''), is responsible for the role of activation functions, max or average pooling (both spatially and across channels), and weights necessary for classification. The operator is defined as follows:
\begin{equation}
\mexu{\xi}{i=1,...,n}\{c_i\}:=\frac{1}{\xi}\log\left(\frac{1}{n}\sum_{i=1}^n\exp\{\xi{\cdot}c_i\}\right)
\label{eqn:mex_def}
\end{equation}
with the alternative notation $MEX_\xi\{c_i\}_{i=1}^n$ used interchangeably. The parameter $\xi\in\R$ spans a continuum between maximum, expectation (mean), and minimum:
\begin{eqnarray*}
MEX_\xi\{c_i\}_{i=1}^n&\underset{\xi\to+\infty}{\longrightarrow}&\text{max}\{c_i\}_{i=1}^n \\
MEX_\xi\{c_i\}_{i=1}^n&\underset{\xi\to0}{\longrightarrow} &\text{mean}\{c_i\}_{i=1}^n \\
MEX_\xi\{c_i\}_{i=1}^n&\underset{\xi\to-\infty}{\longrightarrow} &\text{min}\{c_i\}_{i=1}^n
\end{eqnarray*}
Moreover, for a given value of $\xi$, the operator is smooth and exhibits the ``collapsing'' property defined below:
\begin{eqnarray}
&MEX_\xi\{ MEX_\xi\{c_{ij}\}_{1\leq j\leq m}\}_{1\leq i\leq n}& \nonumber \\
&=MEX_\xi\{c_{ij}\}_{1\leq j\leq m,1\leq i\leq n}& \label{eqn:mex_collapse}
\end{eqnarray}
In a layered architecture, the MEX operator is used to define the MEX layer -- see illustration in fig.~\ref{fig:layers_nets}(b). In the MEX layer, the input is divided into (possibly overlapping) blocks, each mapped to a single output element. The output value associated with the $t$'th input block is given by:
$$out(t)=MEX_{\xi_t}\left\{\left\{inp(s)+b_{ts}\right\}_{s\in block(t)}, c_t\right\}$$
where the index $s$ runs though the input block, the offsets $b_{ts}\in\R$ serve various roles as will be described later, and $c_t\in\R$ are optional (may or may not be used). The MEX layer can realize two standard ConvNet layers -- the ReLU activation and the max-pooling layer. To realize ReLU activation, one should set the input blocks to be single entries, have the output dimensions equal to the input dimensions, set $b_{ts}=0,c_t=0$, and let ${\xi_t}\rightarrow+\infty$, and as a result $out(t)=\max\{inp(t),0\}$ as required (see fig.~\ref{fig:layers_nets}(c)). To realize a max-pooling layer, set the input blocks to cover a 2D area, set the depth of the output equal to that of the input, set $b_{ts}=0$, omit $c_t$, and set ${\xi_t}\rightarrow+\infty$. As a result $out(i,j,l)=\max\{inp(i',j',l)\}_{(i',j')\in pool(i,j)}$ (see fig.~\ref{fig:layers_nets}(d)). Note that by setting ${\xi_t}\rightarrow 0$ one obtains an average-pooling layer, and moreover, the parameters $\xi_t$ can be learned (optimized) as part of the training process, allowing additional flexibility.
To recap, the layers corresponding to the two operators of the SimNet architecture -- similarity and MEX, can realize conventional ConvNets as follows:
\begin{itemize}
\item \emph{Convolutional layer}: use similarity layer with linear form $\phi_l(\x,\z)_i=x_i z_i$ and unit weights $\uu_l=\1$.
\item \emph{ReLU activation}: use MEX layer with $b_{ts}=0,c_t=0,\xi_t\rightarrow+\infty$ and single-entry input blocks.
\item \emph{Max pooling layer}: use MEX layer with $b_{ts}=0,\xi_t\rightarrow+\infty$, $c_t$ omitted, and 2D input blocks.
\item \emph{Dense layer}: use similarity layer with the entire input as the only patch, linear form $\phi_l(\x,\z)_i=x_i z_i$ and unit weights $\uu_l=\1$.
\end{itemize}
Next, we make wider use of the two SimNet layers, taking us beyond classical ConvNets, exploring connections to classical statistical learning models with kernel machines.
\begin{figure*}
\includegraphics[width=\textwidth,height=\textheight,keepaspectratio]{layers_nets}
\vspace{-1cm}
\caption{\footnotesize (a)~SimNet similarity layer~~(b)~SimNet MEX layer~~(c)~ConvNet ReLU activation layer~~(d)~ConvNet max/average pooling layer~~(e)~SimNet MLP with multiple outputs~~(f)~SimNet with locality, sharing and pooling~~(g)~Patch-labeling SimNet~~(h)~SimNet $l_p$-similarity layer followed by MEX layer}
\label{fig:layers_nets}
\end{figure*}
\section{SimNets and kernel machines} \label{sec:simnets_ksvm}
So far, we set the architectural choices of SimNets to realize classical ConvNets, which form a rudimentary special case of the available possibilities. In particular, we did not make use of the $l_p$ similarity, and of the offsets $\{b_{ts}\}$ in the MEX layer. In the following subsection, we consider a ``multi-layer perceptron'' (MLP) construction consisting of a single hidden layer in addition to input and output layers. In the subsection that follows, we will study the case where the input layer is processed by patches, the hidden layer involves locality and sharing, and a pooling operation follows the hidden layer -- a structure prevalent in classical ConvNets.
\subsection{MLP analogy: input $\rightarrow$ hidden layer $\rightarrow$ output} \label{subsec:mlp}
The Similarity and MEX operators give straightforward generalizations of the convolution and max/average pooling layers in ConvNets. As we now show, they create something of greater consequence when applied one after the other in succession. To make the point as succinctly as possible, consider a MLP construction consisting of $d$ input nodes (making up the input vector $\x\in\R^d$), $n$ hidden units, and a single output unit. The value $h(\x)$ of the output unit is a result of a mapping $\R^d\to\R$, defined by the two SimNet operators applied in succession ($n$ similarity operators with different templates and shared mapping $\phi$, followed by MEX with offsets):
\begin{equation}
h(\x)=MEX_\xi\left\{\uu_l^\top\phi(\x,\z_l)+b_l\right\}_{l=1,...,n} \label{eqn:hx}
\end{equation}
A straightforward analogy to existing work is obtained by setting unit weights $\uu=1$, linear similarity $\phi(\x,\z)_i=x_i z_i$ and $\xi\to\infty$, resulting in $h(\x)=\max\left\{\z_l^\top\x+b_l\right\}_{l=1}^n$ -- a maxout operation~\cite{goodfellow2013maxout}. There were other attempts to generalize maxout, notably the recently proposed $L_p$ unit~\cite{gulcehre2014learned}, which is defined by $\left(\frac{1}{n}\sum_{l=1}^n\abs{\z_l^\top\x+b_l}^p\right)^{1/p}$. When $p\to\infty$, this reduces to $\max_l\left\{\abs{\z_l^\top\x+b_l}\right\}$. The differences between this and the SimNet generalization of maxout are: (i) the $L_p$ unit generalizes maximum of \emph{absolute} values (rather than the values themselves), and (ii) the $L_p$ unit tries to create a maxout in a single operation whereas the SimNet creates $h(\x)$ over a succession of two operators -- similarity followed by MEX.
Next, consider the case of fixed $\xi>0$ and unweighted ($\uu_l=\1$) linear similarity ($\uu_l^\top\phi(\x,\z_l)=\x^\top\z_l$) or unweighted $l_p$ similarity ($\uu^\top\phi(\x,\z)=-\norm{\x-\z}_p^p$) with fixed $0<p\leq2$. We will show below that in this case, the output $h(\x)$ is the result of a non-linear monotone activation function applied to the inner-product between a mapping of the input $\x$ and a vector ${\mathbf w}$ in some high-dimensional feature space $\R^F$. More formally, we will show that $h(\x)=\sigma(\inprod{{\mathbf w}}{\psi_\phi(\x)})$, where the mapping $\psi_\phi:\R^d\to\R^F$ depends on the choice of similarity mapping $\phi$, $\sigma$ is a non-linear monotone activation function, and ${\mathbf w}=\sum_{l=1}^n\alpha_l\psi_\phi(\z_l)$ for some $\alpha_1,...,\alpha_n\in\R$. We thus conclude that the output unit is a ``neuron'' in the classical sense, but in a high-dimensional feature space. To prove this assertion, we notice that $h(\x)$ can be expressed as follows:
\begin{eqnarray}
h(\x)
&=&MEX_\xi\left\{\uu_l^\top\phi(\x,\z_l)+b_l\right\}_{l=1,...,n}\nonumber \\
&=&\frac{1}{\xi}\ln\left(\frac{1}{n}\sum_{l=1}^n\alpha_l\cdot\exp\left\{\xi\sum_{i=1}^d\phi(\x,\z_l)_i\right\}\right) \nonumber\\
&=&\sigma\left(\sum_{l=1}^n\alpha_{l}\cdot K_\phi (\x,\z_l)\right) \label{eqn:hx_2}
\end{eqnarray}
where $\alpha_l:=e^{\xi\cdot b_l}$ and $\sigma$ is a non-linear monotone activation function given by $\sigma(t)=\frac{1}{\xi}\ln\left(\frac{t}{n}\right)$. We use the notation $K_\phi(\x,\z):=\exp\left\{\xi\sum_{i=1}^d\phi(\x,\z)_i\right\}$ to indicate that, under the similarity mappings considered, the function is a kernel on $\R^d$. In particular, for the linear and $l_p$ similarities we have:
\begin{eqnarray*}
K_{lin}(\x,\z) &=& \exp\left\{\xi\cdot\x^\top\z \right\} \\
K_{l_p}(\x,\z) &=& \exp\left\{-\xi\norm{\x-\z_l}_p^p\right\}
\end{eqnarray*}
As shown in~\cite{scholkopf2002learning}, $K_{lin}$ and $K_{l_p}$ are kernels on $\R^d$ (note that for $p>2$, the expression above for $K_{l_p}$ is not a kernel). We refer to them as the ``Exponential'' kernel and the ``Generalized Gaussian'' kernel\footnote{When $p=2$, this reduces to the well-known Gaussian (radial basis function) kernel. When $p=1$, it reduces to the Laplacian kernel.} respectively. Since $\sigma$ is monotonically increasing, $h(\x)$ realizes a 2-class ``reduced'' kernel-SVM decision rule, with $\z_l$ being the (reduced) support-vectors\footnote{We use the term ``reduced'' to refer to the case where the number of support-vectors is predetermined and they are not constrained to lie in the training set. This setting was studied in~\cite{wu2006direct} in the context of binary (2-class) classification. The extension to multiclass~(\cite{crammer2002algorithmic}) is straightforward.} and $\alpha_l\geq0$ being the coefficients associated with the support-vectors. Let $\psi_\phi:\R^d\to\R^F$ be a feature mapping corresponding to $K_\phi$, i.e. $K_\phi(\x,\z)=\inprod{\psi_\phi(\x)}{\psi_\phi(\z)}$. Eqn.~\ref{eqn:hx_2} can now be expressed as $h(\x)=\sigma(\inprod{{\mathbf w}}{\psi_\phi(\x)})$, where ${\mathbf w}:=\sum_{l=1}^n \alpha_l\psi_\phi(\z_l)$. This shows that the output unit $h(\x)$ is a ``neuron is feature space'', as stated above.
In the case of weighted ($\uu_l$ are learned) $l_p$-similarity, the hypothesis space realized by the output unit $h(\x)$ is no longer representable by a kernel-SVM. Moreover, the view of $h(\x)$ as a neuron in feature space with learned vector ${\mathbf w}$ no longer applies. This is stated formally below (proof in app.~\ref{app:proof_l_p_sim_wgt_no_K}):
\begin{theorem}
\label{thm:l_p_sim_wgt_no_K}
For any dimension $d\in\N$, and constants $c>0$ and $p>0$, there are no mappings $Z:\R^d\to\R^d$ and $U:\R^d\to\R_+^d$ and a kernel $K:(\R^d\times\R_+^d)\times(\R^d\times\R_+^d)\to\R^d\times\R_+^d$, such that for all $\z,\x\in\R^d$ and $\uu\in\R_+^d$:
\begin{equation}
K\left([Z(\x),U(\x)],[\z,\uu]\right)=\exp\left\{-c\sum_{i=1}^d u_i\abs{x_i-z_i}^p\right\} \label{eqn:l_p_sim_wgt_K_req}
\end{equation}
\end{theorem}
We now turn to consider a straightforward extension to the setup above, which includes $k$ output units. The MLP will now consist of an input signal $\x\in\R^d$, a set of $n$ hidden units defined by similarity functions over $\x$ (all based on the same similarity mapping $\phi$), and a set of $k$ output units defined by MEX operators (all having the same parameter $\xi$) with offsets $b_{rl}$ where $l\in\{1,...,n\}$ runs over the hidden units and $r\in\{1,...,k\}$ is the index of the output unit. Fig.~\ref{fig:layers_nets}(e) illustrates this basic operation. If we consider the output nodes as predicting a label associated with the input $\x$, the chosen label being the index of the node with maximal activation, then running the two operators, similarity and MEX, one following the other, produces the classification rule below:
\begin{equation}
\hat{y}(\x)=\argmax_{r=1,...,k}MEX_{\xi}\{\uu_l^\top\phi(\x,\z_l)+b_{rl}\}_{l=1}^n
\label{eqn:mlp_multiclass}
\end{equation}
This classification measures weighted similarities to $n$ templates, with class-dependent offsets. The role of the MEX operators is to combine the weighted similarities (with offsets) of the input $\x$ to the templates. For example, when $\xi\to+\infty$, the classification rule is attracted to the most similar template where offsets assign relevancy of templates to classes. Let $h_r(\x)$, $r=1,...,k$, be the value of output unit $r$ when the MLP is fed with input $\x$. Following the lines of the derivation carried out for the single-output MLP, when working with unweighted linear similarity or unweighted $l_p$ similarity with $0<p\leq2$, it holds that:
$$ h_r(\x)=\sigma\left(\inprod{{\mathbf w}_r}{\psi_\phi(\x)}\right)$$
where ${\mathbf w}_r:=\sum_{l=1}^n\alpha_{rl}\psi_\phi(\z_l)$ and $\alpha_{rl}:=e^{\xi\cdot b_{rl}}$. Moreover, the decision rule in eqn.~\ref{eqn:mlp_multiclass} can be expressed as:
\begin{eqnarray*}
\hat{y}(\x)
&=& \argmax_{r\in\{1,...,k\}}h_r(\x) \\
&=& \argmax_{r\in\{1,...,k\}}\inprod{{\mathbf w}_r}{\psi_\phi(\x)} \\
&=& \argmax_{r\in\{1,...,k\}}\sum_{l=1}^n\alpha_{rl} K_\phi(\x,\z_l)
\end{eqnarray*}
where $K_\phi$ is a kernel (Exponential or Generalized Gaussian) on $\R^d$. This classification realizes the hypothesis space of a reduced multiclass kernel-SVM\footnote{Note that the coefficients $\alpha_{rl}$ are positive, whereas in classical multiclass SVM they may be any real numbers. This however does not limit generality, as we can always add a common offset to all coefficients after SVM training is complete.}.
To summarize so far, we have shown that with linear similarity, the ``MLP'' construction consisting of input $\to$ hidden layer $\to$ output, gives rise to the hypothesis space of a (reduced) SVM with the Exponential kernel. Replacing the linear similarity with unweighted $l_p$ similarity having fixed order $p$, gives rise to a kernel-SVM if and only if $p\leq2$, in which case the underlying kernel is the Generalized Gaussian kernel (the special cases of Gaussian and Laplacian kernels are obtained for $p=2$ and $p=1$ respectively). With these similarities that give rise to kernel machines, a unit generated by similarity operators followed by MEX with offsets is a ``neuron in feature space''. Finally, with weighted ($\uu_l$ are learned) $l_p$ similarity, the framework is no longer representable by a kernel-SVM.
To obtain a sense of the network's abstraction level, i.e. its ability to capture concept (category) distributions in input space, consider the classification rule in eqn.~\ref{eqn:mlp_multiclass} in the case $\xi\to+\infty$:
$$\hat{y}(x)=\argmax_{r=1,...,k}\max_{l=1,...,n}\{\uu_l^\top\phi(\x,\z_l)+b_{rl}\}$$
For any $r\in\{1,...,k\}$, denote by $A_r$ the decision region in input space that corresponds to class $r$, i.e. $A_r:=\{\x\in\R^d:\hat{y}(\x)=r\}$. To understand the shape of $A_r$, we make the following definitions:
\begin{eqnarray}
&A_{r,l}^{r',l'}:=& \label{eqn:decision_region_basic_shape} \\
&\{\x\in\R^d:\uu_l^\top\phi(\x,\z_l)+b_{rl}\geq\uu_{l'}^\top\phi(\x,\z_{l'})+b_{r'l'}\}& \nonumber
\end{eqnarray}
$$ A_{r,l}:=\bigcap_{(r',l')\neq(r,l)} A_{r,l}^{r',l'} $$
where the class index $r'$ ranges over $\{1,...,k\}$, and the template indexes $l,l'$ range over $\{1,...,n\}$. One can readily see that up to boundary conditions:
$$ A_{r} = \bigcup_{l\in\{1,...,n\}}A_{r,l} $$
Consider first the setting of linear similarity ($\phi(\x,\z)_i=x_i z_i$). In this case $A_{r,l}^{r',l'}$ are half-spaces and $A_{r,l}$ are intersections of half-spaces (polytopes). The decision region $A_r$ is thus a union of $n$ polytopes. As we now show, this is the same type of decision regions as obtained with unweighted $l_2$ similarity ($\uu_l=\1$, $\phi(\x,\z)_i=-\abs{x_i-z_i}^2$). Indeed, in this case the term $\norm{\x}_2^2$ in both sides of the inequality defining $A_{r,l}^{r',l'}$ cancels-out, and we obtain again a half-space. This in turn implies that as before, $A_{r,l}$ are polytopes and $A_r$ is a union of polytopes. We conclude that with the MLP structure of: input $\to$ hidden layer $\to$ output units, the setting that realizes a Gaussian kernel machine (unweighted $l_2$ similarity), is qualitatively equivalent to the ``ConvNet'' (linear similarity) setting that realizes an Exponential kernel machine. The difference in kernels does not account for any material difference in the network's hypothesis space, i.e. its abstraction level.
Remaining with $l_2$ similarity, we now consider the weighted setting, i.e. the setting in which $\uu_l$ are not fixed. Thm.~\ref{thm:l_p_sim_wgt_no_K} tells us that in this case the hypothesis space is no longer governed by kernel-SVM. From the decision region point-of-view, it is not difficult to see that in this case $A_{r,l}^{r',l'}$ is no longer a half-space, but a region defined by a second-order hyper-surface. This implies that the set $A_{r,l}$ is no longer a polytope, and in particular is not necessarily convex. The possible shapes that the decision region $A_r$ can take are thus enriched. We conclude that unlike in the case of unweighted $l_2$ similarity, the setting of weighted $l_2$ similarity is characterized by an abstraction level higher than that induced by linear similarity (convolutional operator).
In the general setting of $l_p$ similarity, the sets $A_{r,l}^{r',l'}$ are more complex, and may be governed by non-convex non-smooth separating hyper-surfaces. The full analysis is outside the scope of this paper, but an informal illustration of how the space is divided for $p=1$ and $d=2$ ($\R^d$ is the 2D plane) is given in fig.~\ref{fig:l1_decision_regions}. Under this specific setup, the 2D plane is divided into two by a piece-wise linear separating boundary. The unweighted case (uniform weights) is shown in fig.~\ref{fig:l1_decision_regions}(a). In this case the space is divided equally (up to a shift caused by the offsets $b_{rl},b_{r'l'}$) based on the $l_1$ (Manhattan) distance metric. Adding weights deforms the boundary line, where the higher the weights associated with a template ($\z_l$ or $\z_{l'}$) are, the less space is allocated to that template. For example, in fig.~\ref{fig:l1_decision_regions}(d) the weights associated with the template $z_{l'}$ are uniformly high, thereby creating a small aperture in the 2D plane around that template. Given that $A_{r,l}^{r',l'}$ is highly non-convex in the weighted setting, we expect weighted $l_1$ similarity to provide a higher abstraction level than that of linear similarity (convolutional operator).
\begin{figure*}
\includegraphics[width=\textwidth,height=\textheight,keepaspectratio]{l1_decision_regions}
\vspace{-3mm}
\caption{\footnotesize Illustration of $A_{r,l}^{r',l'}$ (defined in eqn.~\ref{eqn:decision_region_basic_shape}) in the setting of $l_1$ similarity and $d=2$ ($\R^d$ -- the 2D plane). Each panel shows the location of the templates $\z_l$ and $\z_{l'}$, and on the top the values of the corresponding similarity weights.~~(a):~Unweighted setting (uniform weights). The 2D plane is divided equally between the two templates (up to a shift resulting from the offsets $b_{r,l},b_{r',l'}$.~~(d):~Here $\z_{l'}$ is associated with high weights, thus the portion of the plane allocated to this template is ``shrinked''.}
\label{fig:l1_decision_regions}
\end{figure*}
\subsection{A basic 3-layer SimNet with locality, sharing and pooling} \label{subsec:local_share_pool}
Next, we analyze a 3-layer SimNet with locality, sharing and pooling. The network's input is processed by patches (locality), with the same templates and weights applied to all patches (sharing), thereby creating a stack of feature maps (channels) -- one for each template. Spatial regions of each feature map are then pooled together to reduce dimensionality, and finally, a classification output layer predicts the label of the input. We will show that such a network, consisting of input $\to$ feature maps $\to$ pooling $\to$ output, also corresponds to a kernel-SVM, with the kernels designed for a ``patch-based'' representation of the input signal.
Locality, sharing and pooling are realized in the conventional manner. Namely, the input is divided into (possibly overlapping) patches $\x_{ij}\in\R^d$ where $d=h\cdot w\cdot D$, with $h,w$ being the height and width of the patches. A similarity layer as illustrated in fig.~\ref{fig:layers_nets}(a), but with the same similarity mapping $\phi$ for all channels, matches the patch $\x_{ij}$ with the template $\z_l\in\R^d$ (which is now a template representing a local patch in the layer's input) using the weights $\uu_l\in\R_+^d$, and the resulting value $\uu_l^\top\phi(\x_{ij},\z_l)$ is stored in coordinates $(i,j,l)$ of the layer's output.
The mapping from the similarity layer to the $k$-node classification output is realized through two MEX layers. The first MEX layer implements a pooling layer as follows. Let $q(i,j)=(q_h(i),q_w(j))$ be a (contraction) mapping of the 2D coordinate system in the similarity layer to the 2D coordinate system in the pooling layer. Normally, a 2D coordinate in the pooling layer corresponds to a 2D window in the similarity layer. The value assigned to an element in the pooling layer is simply a MEX operation taken over the corresponding 2D window (in the respective channel $l\in\{1,...,n\}$) in the similarity layer:
$$pool(p_h,p_w, l)=MEX_{\xi_1}\{sim(i,j,l)\}_{i,j: q(i,j)=(p_h,p_w)}$$
All MEX operators in the layer have the same parameter -- $\xi_1$. When $\xi_1\rightarrow+\infty$ for instance, we obtain max-pooling as implemented in conventional ConvNets.
The second MEX layer implements a dense mapping from the pooling layer to the $k$ output nodes, which includes offsets. The value of the $r$'th output node is given by:
$$out ( r )= MEX_{\xi_2}\{pool(p_h,p_w,l)+b_{r l p_h p_w}\}_{p_h,p_w,l}$$
where $(p_h,p_w)$ runs over the 2D coordinates of the pooling layer and $l$ runs over the pooling layer's channels (which correspond to templates). Note that here too all MEX operators have the same parameter -- $\xi_2$. The offsets $b_{r l p_h p_w}$ depend on the output node $r$ and on the 3D coordinates of the pooling layer $(p_h,p_w,l)$, i.e. for every output node there is an offset for each coordinate in the pooling layer.
The SimNet we obtain is illustrated in fig.~\ref{fig:layers_nets}(f). It is a basic similarity-pooling-output network that employs locality and sharing, as conventional ConvNets do. The point to be made next, is that in the special case where $\uu_l=\1$, $\xi_1,\xi_2$ are fixed to a constant $\xi>0$, and $\phi$ is set to linear form or $l_p$ form with fixed $p\leq2$, the classification resulting from this network is a kernel-SVM, where the kernel is designed for a ``patch-based'' representation of the input signal. First, by concatenating the three steps -- similarity, pooling and output, and assuming that $\xi_1=\xi_2=\xi>0$, the decision rule associated with the network becomes:
\begin{equation}
\hat{y}(inp)=\argmax_{r=1,...,k}\mexu{\xi}{i,j,l}\left\{\uu_l^\top\phi(\x_{ij},\z_l)+b_{r,l,q(i,j)}\right\}
\label{eqn:simnet_patch_class}
\end{equation}
This follows from the identities below:
\begin{eqnarray*}
&\mexu{\xi}{p_h,p_w,l}\left\{\mexu{\xi}{i,j: q(i,j)=(p_h,p_w)}\left\{\uu_l^\top\phi(\x_{ij},\z_l)\right\}+b_{r l p_h p_w}\right\}& \\
&=\mexu{\xi}{p_h,~p_w,~l,~i,j:q(i,j)=(p_h,p_w)}\left\{\uu_l^\top\phi(\x_{ij},\z_l)+b_{r l p_h p_w}\right\}& \\
&=\mexu{\xi}{i,j,l}\left\{\uu_l^\top\phi(\x_{ij},\z_l)+b_{r,l,q(i,j)}\right\}&
\end{eqnarray*}
where for the first equality, we used the collapsing property of the MEX operator described in eqn.~\ref{eqn:mex_collapse}. The classification described in eqn.~\ref{eqn:simnet_patch_class} is similar to that described in eqn.~\ref{eqn:mlp_multiclass}, but has two important distinctions: \emph{(i)} the templates $\z_l$ are local and similarity is applied locally (hence the ``locality'' and ``sharing''), and \emph{(ii)} the offsets are region-based (hence the ``pooling''), i.e. each collection of input patches $\x_{ij}$ ascribed to the same pool is associated with a single set of offsets (per-class and per-template). To see the kernel structure associated with this classification, we perform the following manipulations to the rule given in eqn.~\ref{eqn:simnet_patch_class}:
\begin{eqnarray}
&\hat{y}(inp)=& \nonumber \\
&\argmax\limits_{r=1,...,k}\mexu{\xi}{i,j,l}\left\{\uu_l^\top\phi(\x_{ij},\z_l)+b_{r,l,q(i,j)}\right\}=& \nonumber \\
&\argmax\limits_{r=1,...,k}\sum\limits_{i,j,l}e^{\xi\cdot(\uu_l^\top\phi(\x_{ij},\z_l)+b_{r,l,q(i,j)})}=& \nonumber \\
&\argmax\limits_{r=1,...,k}\sum\limits_{i,j,l}\underbrace{e^{\xi\cdot b_{r,l,q(i,j)}}}_{:=\alpha_{r,l,q(i,j)}}\cdot e^{\xi\cdot\uu_l^\top\phi(\x_{ij},\z_l)}=& \nonumber \\
&\argmax\limits_{r=1,...,k}\sum\limits_{p_h,p_w,l}\alpha_{r l p_h p_w}\sum\limits_{i,j:q(i,j)=(p_h,p_w)} e^{\xi\cdot\uu_l^\top\phi(\x_{ij},\z_l)}& \label{eqn:simnet_patch_class_2}
\end{eqnarray}
Setting $\uu_l=\1$, and referring to subsec.~\ref{subsec:mlp}, we denote $K_\phi(\x_{ij},\z_l):=\exp\{\xi\cdot\1^\top\phi(\x_{ij},\z_l)\}$, emphasizing that this function is a kernel for the similarity mappings we consider (Exponential kernel for linear similarity, Generalized Gaussian kernel for $l_p$ similarity with fixed $p\leq2$). Eqn.~\ref{eqn:simnet_patch_class_2} then becomes:
\begin{eqnarray}
&\hat{y}(inp)=& \label{eqn:simnet_patch_class_3} \\
&\argmax\limits_{r=1,...,k}\sum\limits_{p_h,p_w,l}\alpha_{r l p_h p_w}\sum\limits_{i,j:q(i,j)=(p_h,p_w)}K_\phi(\x_{ij},\z_l)& \nonumber \\
&=\argmax\limits_{r=1,...,k}\sum_{p_h,p_w,l}\alpha_{r l p_h p_w}\K_\phi(X,Z_{l p_h p_w})& \nonumber
\end{eqnarray}
where $X$ contains the concatenation of all the input patches $\x_{ij}$, and $\Z_{l p_h p_w}$ is a structure containing copies of $\z_l$ in locations corresponding to the pool index $(p_h,p_w)$ -- the details, including definition of $\K_\phi$ and proof that it is indeed a kernel, are given in app.~\ref{app:patch_ksvm}.
\section{Other SimNet settings -- global average pooling } \label{sec:other_simnets}
In subsec.~\ref{subsec:local_share_pool} we introduced the SimNet basic building chain of the form: input $\to$ similarity $\to$ pooling $\to$ output, whose structure follows the line of classical ConvNets. We noted that the basic building chain realizes a kernel-SVM hypothesis space, where the templates in the similarity layer correspond to the (reduced) support-vectors, and the offsets in the last MEX layer (from pooling to output) are related to the SVM coefficients. The SVM hypothesis space is realized when the similarity operator is set to linear form or $l_p$ form with fixed $p\leq2$, and is unweighted ($\uu_l=\1$). Using weighted $l_p$ similarity (weights are not applicable to linear similarity) has the potential of providing a richer hypothesis space than kernel-SVM (at the expense of doubling the number of parameters in the similarity layer). Indeed, experiments we conducted (reported in sec.~\ref{sec:exp}) validate the power of $l_p$ similarity weighting, showing that it matters more than merely the added number of parameters to the model.
In this section, we introduce another SimNet building chain with two MEX layers, designed in such a way that when the MEX parameters are equal, the chain collapses into the one presented above (decision rule in eqn.~\ref{eqn:simnet_patch_class}), but when the MEX parameters are determined separately -- either learned using training data or set manually, the SimNet chain allows for new possibilities (without additional parameters). For example, setting the MEX parameter of the first layer to $1$ and that of the second layer to $0$ gave rise to the best experimental performance we encountered.
The idea is to switch the roles of the two MEX layers -- rather than having the first play the role of pooling and the second the role of classification (using the offsets $b_{rl}$), we start with a MEX layer with offsets and finish with a MEX for pooling. The interpretation of such a structure is that each input patch $\x_{ij}$ undergoes classification in the first MEX layer. The second MEX layer performs a majority voting over all the patch-based classification results to form a final classification decision. This approach follows the line of the ``global average pooling'' structure recently suggested in the context of ConvNets, which has been shown to outperform the traditional ``dense classification'' paradigm~(\cite{lin2013network,szegedy2014going}). To enforce spatial consistency in the labeling characteristics of patches, we constrain the first MEX layer's offsets to be uniform inside predetermined spatial regions. The resulting SimNet, which we refer to as a ``patch labeling'' network, is illustrated in fig.~\ref{fig:layers_nets}(g) (note that all channels in the similarity layer share the same similarity mapping, and that both MEX layers have global parameters $\xi_1,\xi_2$). Its classification rule takes the following form:
$$ \hat{y}(inp)=\argmax_{r=1,...,k}out(r) $$
with:
\begin{eqnarray*}
&out(r)=& \\
&\mexu{\xi_2}{i,j}\{\mexu{\xi_1}{l}\{\uu_l^\top\phi(\x_{ij},\z_l)+b_{r,l,q(i,j)}\}\}&
\end{eqnarray*}
The variables which can be learned here are the offsets $b_{r l p_h p_w}\in\R$ (with $r$ ranging over the classes, $l$ over the templates and ($p_h,p_w)$ over the regions in which offsets are shared), the templates $\z_l\in\R^{hwD}$, the similarity weights $\uu_l\in\R_+^{hwD}$, the order $p>0$ in case $l_p$ similarity is chosen, and the MEX parameters $\xi_1,\xi_2\in\R$. Assume we constrain the MEX parameters to be equal: $\xi_1=\xi_2=\xi$. The MEX collapsing property (eqn.~\ref{eqn:mex_collapse}) then applies, and the classification rule becomes:
$$ \hat{y}(inp)=\argmax_{r=1,...,k}MEX_{\xi}\{\uu_l^\top\phi(\x_{ij},\z_l)+b_{r,l,q(i,j)}\}_{i,j,l} $$
which is identical to the decision rule in eqn.~\ref{eqn:simnet_patch_class}. However, there is no reason to have the MEX parameters equal to each other. We can estimate their value during training, or set them manually. For example, during our experimentation we found that the case of equal MEX parameters -- $\xi_1=\xi_2=\xi$, is significantly outperformed by the setting $\xi_2\to0$, which corresponds to the following classification rule:
\begin{eqnarray*}
&\hat{y}(inp)=& \\
&\argmax\limits_{r=1,...,k}\sum_{i,j}\mexu{\xi_1}{l\in\{1,...,n\}}\left\{\uu_l^\top\phi(\x_{ij},\z_l)+b_{r,l,q(i,j)}\right\}&
\end{eqnarray*}
\section{Initializing parameters using unsupervised learning} \label{sec:init}
For classical ConvNets, various schemes of initializing a network based on unlabeled data (unsupervised initialization) have been proposed (c.f.~\cite{hinton2006fast,bengio2007greedy,vincent2008extracting}). Over time, however, these were taken over by carefully selected random initializations that do not use data at all (see for example~\cite{krizhevsky2012imagenet,zeiler2014visualizing,sermanet2013overfeat}). No initialization scheme to-date is sufficient on its own for overcoming the hardness of training. Indeed, successful training of ConvNets typically requires designing an over-specified network (i.e. a network that is much larger than necessary in order to represent the true hypothesis space). While the latter has been shown to produce good training results, it bares a computational price, and also aggravates the problem of overfitting. The enhanced susceptibility to overfitting has led to various regularization techniques and heuristics (Dropout~(\cite{hinton2012improving}) being the most prominent), which nowadays form an art that one must master in order to properly train ConvNets. In this section, we discuss a natural unsupervised initialization scheme for SimNets, which is based on statistical estimation. Such a scheme may provide a more effective local minima in the process of training a SimNet, thereby reducing the need for over-specification, supporting smaller networks that are more efficient computationally, and less prone to overfit. Experiments we conducted (reported in sec.~\ref{sec:exp}) validate this conjecture, showing that the SimNet unsupervised initialization scheme indeed improves performance over random initializations, especially in the case of small networks.
Recall from sec.~\ref{sec:simnet_arch} that measuring weighted similarities to templates forms the similarity layer -- a basic building block of the SimNet architecture (see fig.~\ref{fig:layers_nets}(a)). Focusing on the case of $l_p$ similarity mappings ($\phi(\x,\z)_i = -\abs{x_i-z_i}^p$), we show how the application of statistical estimation methods to unlabeled training data can produce initialization values for the layer's templates $\z_1,...,\z_n$, weights $\uu_1,...,\uu_n$ and orders $p_1,...,p_n$. Consider a probability distribution over $\R^d$ defined by a mixture of $n$ Generalized Gaussian distributions, each having independent coordinates with a shared shape parameter and separate scales and means:
$$P(\x)=\sum_{l=1}^n\lambda_l\prod_{i=1}^d\frac{\beta_l}{2\alpha_{l,i}\Gamma(1/\beta_l)}
\exp\left\{-\left(\frac{\abs{x_i-\mu_{l,i}}}{\alpha_{l,i}}\right)^{\beta_l}\right\}$$
In the above, $\lambda_l$ stands for the prior probability of component $l$ ($\lambda_l\geq0,\sum_l\lambda_l=1$), $\beta_l>0$ stands for the shape parameter of all coordinates in component $l$, $\alpha_{l,i}>0$ stands for the scale of coordinate $i$ in component $l$, $\mu_{l,i}\in\R$ stands for the mean of coordinate $i$ in component $l$, and $\Gamma$ is the Gamma function, defined by $\Gamma(s)=\int_0^{\infty}e^{-t}t^{s-1}dt$. The log-probability that a vector drawn from this distribution is equal to $\x$ and originated from component $l$ is:
$$\log P(\x\land \text{component}~l)=-\sum_{i=1}^d\alpha_{l,i}^{-\beta_l}\abs{x_i-\mu_{l,i}}^{\beta_l}+c_l$$
where $c_l:=\log\left\{\lambda_l\prod_{i=1}^d\frac{\beta_l}{2\alpha_{l,i}\Gamma(1/\beta_l)}\right\}$ is a constant that depends on the component only (not on $\x$). Setting the layer's templates by $z_{l.i}=\mu_{l,i}$, its weights by $u_{l,i}=\alpha_{l,i}^{-\beta_l}$ and its $l_p$ orders by $p_l=\beta_l$, would give:
$$ \uu_l^\top\phi_l(\x_{ij},\z_l)=\log P(\x\land \text{component}~l)-c_l $$
This implies that if we assume input patches follow a Generalized Gaussian mixture as described, initializing the similarity layer's templates, weights and orders as above would result in channel $l$ of the layer's output holding, up to a constant, the probabilistic heat map of component $l$ and the patches. This observation suggests estimating the parameters (shapes $\beta_l$, scales $\alpha_{l,i}$ and means $\mu_{l,i}$) of the Generalized Gaussian mixture using unlabeled input patches (via standard statistical estimation methods, such as that presented in~\cite{bazi2007image}), and initializing the similarity layer accordingly.
Consider now the case where the initialized $l_p$ similarity layer is followed by a MEX layer with learned offsets (see fig.~\ref{fig:layers_nets}(h), where for convenience, the linear index $t$ is used to refer to elements of the MEX layer's 3D output array). We now assume that not only do input patches come from a mixture of Generalized Gaussian components as above, but also that each input patch location corresponds to a different mixture (priors) of these components. This makes sense, as certain templates that are likely to appear in the center of an image for example, may be less likely to appear on the top-left corner of the image, for example. Using our estimates of the global components obtained during the initialization of the similarity layer, we can estimate a mixture for a certain input patch location, by applying an estimation method to patches only from that location, with the component shapes, means and scales held fixed. We may then calculate offsets for the $n$ elements of the similarity layer's output that correspond to that location, such that the probabilistic heat maps will take into account the location-dependent statistics, and will be precise (not up to a constant). For example, if there is a region in the input for which a certain template is very unlikely to appear, that template's heat map in the aforementioned region will be suppressed. The offsets we compute may serve for initialization of the MEX layer's offsets.
\section{Experiments} \label{sec:exp}
\begin{figure*}
\includegraphics[width=\textwidth,height=\textheight,keepaspectratio]{exp_nets}
\vspace{-1cm}
\caption{\footnotesize Networks evaluated on CIFAR-10.~~(a)~Patch labeling SimNet~~(b)~Comparable ConvNet~~(c)~Comparable ``single-layer'' network studied in~\cite{coates2011analysis}.}
\label{fig:exp_nets}
\end{figure*}
\begin{figure*}
\includegraphics[width=\textwidth,height=\textheight,keepaspectratio]{benchmark_results_templates}
\vspace{-5mm}
\caption{\footnotesize CIFAR-10 cross-validation accuracies plotted against the number of templates in the networks (denoted $n$ in fig.~\ref{fig:exp_nets}). `SimNet l1-weight' and `SimNet l2-weight' correspond to the network structure illustrated in fig.~\ref{fig:exp_nets}(a), with $l_1$ and $l_2$ similarities respectively; 'SimNet l1-unweight' and 'SimNet l2-unweight' correspond to the same networks, but with the weight vectors $\uu_l$ held fixed during training; 'ConvNet' corresponds to the network illustrated in fig.~\ref{fig:exp_nets}(b); 'Coates network' corresponds to the network illustrated in fig.~\ref{fig:exp_nets}(c).}
\label{fig:benchmark_results_templates}
\end{figure*}
\begin{figure*}
\includegraphics[width=\textwidth,height=\textheight,keepaspectratio]{benchmark_results_params}
\vspace{-5mm}
\caption{\footnotesize CIFAR-10 cross-validation accuracies plotted against the number of learned parameters in the networks. This is merely a different display of the results given in fig.~\ref{fig:benchmark_results_templates}. Notice how with weighted similarities, the SimNet reaches approximately the same level of performance as the competition, using much smaller networks.}
\label{fig:benchmark_results_params}
\end{figure*}
We implemented the ``patch labeling'' SimNet discussed in sec.~\ref{sec:other_simnets}, and experimented with the specific architectural settings illustrated in fig.~\ref{fig:exp_nets}(a). The network consists of a $l_p$-similarity layer with $p$ fixed at $1$ or $2$, followed by two MEX layers. Implementation and evaluation of deeper SimNets is currently under work, and will be reported at a later time. For the experiments reported here, we used the CIFAR-10 dataset~(\cite{krizhevsky2009learning}), which consists of $60,000$ color images ($50,000$ for training and $10,000$ for testing) of size $32\times32$ partitioned into $10$ classes, with $6,000$ images per class ($5,000$ for training, $1,000$ for testing). The network's input is an RGB image ($32\times32\times3$ array), processed by patches of size $6\times6\times3$ with a single-pixel stride between them. For a given number of templates in the similarity layer (denoted by $n$), the SimNet's learned parameters are the templates $\z_1,...,\z_n\in\R^{108}$, the similarity weights $\uu_1,...,\uu_n\in\R_+^{108}$, and the MEX offsets $b_{r l p_h p_w}\in\R$ with $r=1,...,10$, $l=1,...,n$ and $p_w,p_h=1,2$. We used statistical estimation as described in sec.~\ref{sec:init} to initialize the templates $\z_l$ and the similarity weights $\uu_l$ (initialization was based on training images, without making use of their labels). The network was then trained by minimizing a softmax loss with stochastic gradient descent (SGD) that includes momentum and acceleration~(\cite{sutskever2013importance}). For SGD, we used a batch size of $64$, a momentum of $0.9$, and a learning rate of $0.01$ decreased by a factor of $10$ after $50$ epochs, running $100$ epochs in total. The weight decay for the templates was set to zero, and those for the similarity weights and offsets were set equal to each other, their value chosen via cross-validation.
We compared the SimNet to instances of two learning architectures. The first is a ConvNet with a single convolutional layer followed by a pooling layer followed by an output layer (see illustration in fig.~\ref{fig:exp_nets}(b)). The purpose of this comparison is to evaluate the SimNet against an analogous ConvNet, measuring the network sizes (number of learned parameters) required to reach given accuracies. A successful outcome here would be if the SimNet reached the same (or higher) level of performance as the ConvNet, with considerably smaller network size. The second comparison we held was against the ``single-layer'' network studied by Coates et al.~(\cite{coates2011analysis}), which has the same depth as the evaluated SimNet, and whose performance on CIFAR-10 is one of the best reported for networks of such depth (absolute state of the art in 2011). In~\cite{coates2011analysis}, a number of unsupervised learning methods were devised for ``coding'' the input image. The coding methods were based on ``templates'', such that each template corresponded to a single feature map. The feature maps were passed on to a sum-pooling operator, and from there a linear SVM was learned using supervised data. Many coding methods were experimented on, and the one that produced the best results, referred to as ``triangle'' coding, was a ``soft'' Euclidean measure applied to templates learned via k-means. This coding method, along with the other architectural settings that produced the best results, are illustrated in fig.~\ref{fig:exp_nets}(c). Finally, we question the importance of the unsupervised initialization scheme described in sec.~\ref{sec:init}, by training the evaluated SimNet with random initialization (as customary with ConvNets), and examining the effect on the cross-validation accuracies.
The results reported below show that the evaluated SimNet achieves performance comparable to that of the ConvNet and the network of Coates et al., with only a fraction of the number of learned parameters. The unsupervised initialization scheme indeed boosts performance, and can be viewed as one of the drivers behind the SimNet's superiority. We are currently working on the optimization of our code (including GPU acceleration), to enable evaluation of larger and deeper SimNets on more meaningful benchmarks, comparing against deep state of the art ConvNets.
\subsection{Benchmarking against the ConvNet}
The ConvNet was implemented using Caffe toolbox~(\cite{Jia13caffe}), with random initialization and SGD training. We used a batch size of $100$, momentum of $0.95$, and learning rate of $10^{-4}$ decreased by a factor of $10$ every $45$ epochs, running $150$ epochs in total. The global weight decay and the dense layer's DropOut rate were chosen via cross-validation. The ConvNet's input was an RGB image ($32\times32\times3$ array) normalized for brightness and contrast. For the SimNet we also added patch ``whitening'', in accordance with the suggestion of~\cite{coates2011analysis}. The positive effect of whitening for the SimNet (which has $l_1/l_2$ similarities) was verified experimentally, whereas for the ConvNet, we observed that whitening does not have a positive effect (complying with the observations of~\cite{coates2011analysis}).
Fig.~\ref{fig:benchmark_results_templates} shows the cross-validation accuracies of the evaluated networks as a function of the number of templates ($n$). Fig.~\ref{fig:benchmark_results_params} plots the same results against the number of learned parameters in the networks. For the SimNet, we experimented with up to $400$ templates (we believe that more templates would only give marginal improvements in accuracy, and thus did not continue further), and reached accuracies of $76.8\%$ and $77.1\%$ with weighted $l_1$ and $l_2$ similarities respectively. We ran the ConvNet with up to $6,400$ templates (beyond that Caffe had GPU memory management issues), with the highest accuracy standing at $76.2\%$. In comparison, taking into account that with weighted similarities each template carries with it a weight vector, the size (number of learned parameters) of the $400$-template SimNet with weighted similarities was less than $1/9$ the size of the $6,400$-template ConvNet, while achieving slightly superior accuracy. The performance of the SimNet with unweighted similarities on the other hand, is very similar to that of the ConvNet, thus highlighting the importance of the weights in the similarity layer. The weights double the number of parameters in the layer, but the increase in performance scales up super-linearly with the number of added parameters. In other words, weights provide a gain in accuracy which is much higher than what would be obtained by simply adding more templates until reaching the same network size. For example, the accuracies with weights at $100$ templates are considerably higher than the accuracies without weights at $200$ templates, despite the fact that in the latter case, the overall network size is higher.
It is worth noting that the performance of the SimNet with unweighted $l_1$ and $l_2$ similarities is comparable to that of the ConvNet. This confirms what we observed formally in subsec.~\ref{subsec:mlp} -- the hypothesis space (analyzed through the shapes of decision regions) corresponding to unweighted $l_2$-similarity is essentially the same as that which corresponds to linear similarity (convolutional operator). The hypothesis space corresponding to unweighted $l_1$-similarity is different, but apparently does not provide a higher degree of abstraction (further study of this is deferred to future work).
Although the SimNet accuracies achieved here are not state of the art for this dataset, the results demonstrate the potential of SimNets for modeling learning problems with significant reduction in network sizes compared to ConvNets.
\subsection{Benchmarking against the ``single-layer'' network of Coates et al.}
The ``single-layer'' network studied by Coates et al.~(\cite{coates2011analysis}) is of interest on several accounts. First, with GMM coding, the network is equivalent to the SimNet variant presented in eqn.~\ref{eqn:simnet_patch_class_2}. Second, their best result with ``triangle'' coding is one of the highest accuracies on CIFAR-10 reported for networks of this depth (absolute state of the art in 2011). Third, their observations with respect to the effect of whitening are relevant to the SimNet architecture, and indeed, we found that for the evaluated SimNet with $l_1$ and $l_2$ similarities, whitening makes a difference.
In~\cite{coates2011analysis}, the network ``templates'' (i.e. the parameters of the selected coding method) were set using unlabeled data, and were not modified in the supervised training phase. To facilitate a fair comparison against our SimNet (where templates are modified in supervised training), we added an additional supervised training phase, which applied to both the templates and the SVM coefficients. More specifically, we used SGD to jointly modify the network templates and SVM coefficients produced by~\cite{coates2011analysis}, in an attempt to reach higher accuracy levels than those reported by the authors. As it turned out, with the triangle coding they proposed, the supervised update of the templates did not improve accuracy any further than the original k-means clustering. Deep inspection of this phenomena revealed that the k-means clustering (along with the SVM that follows) provides a strong local minima for the learning problem, so even the training accuracy was not improved. This leads us to believe that the triangle coding is so successful precisely because it creates a representation for which k-means finds optimal templates, that cannot be improved even in the presence of labeled data. In~\cite{coates2011analysis} results are reported for up to $1,600$ templates. We used their code to reproduce these results, while running up to $6,400$ templates. The accuracy curve we obtained is displayed in fig.~\ref{fig:benchmark_results_templates}, with $77.3\%$ for $1,600$ templates, $78.3\%$ for $3,200$ templates, and $77.8\%$ for $6,400$ templates. The peak accuracy was achieved for $3,200$ templates, and was slightly higher than the SimNet peak accuracy, which stood at $77.1\%$ for weighted $l_2$-similarity and $400$ templates. The SimNet on the other hand was almost $1/5$ in size (see fig.~\ref{fig:benchmark_results_params}).
\subsection{The importance of unsupervised initialization}
To assess the importance of the SimNets' unsupervised initialization scheme presented in sec.~\ref{sec:init}, we trained the evaluated SimNet (fig.~\ref{fig:exp_nets}(a)) with weighted $l_1$ similarity, using no data for initialization. In particular, we initialized the templates $\z_1,...,\z_n$ randomly with a zero-mean unit-variance Gaussian distribution (in accordance with the fact that the input patches are whitened to have zero-mean and unit variance), and the weights $\uu_1,...,\uu_n$ with constant ones. Besides the difference in initialization, the SimNet was trained exactly as described above. Running the experiment with $200$ templates, cross-validation accuracy dropped from $76\%$ to $74.1\%$. With $400$ templates, accuracy declined from $76.8\%$ to $74.4\%$. With $50$ and $100$ templates, the learning algorithm did not converge. We conclude that the SimNet unsupervised initialization scheme indeed has significant impact on performance. The impact is especially acute for networks of small size. This complies with conventional wisdom, according to which training small networks poses more difficult optimization problems. The SimNet initialization scheme may provide an alternative to the common practice of over-specifying networks (constructing networks larger than necessary in order to ease the optimization task).
\section{Discussion}
We presented a deep layered architecture called SimNets, with similar ingredients as classical ConvNets. The architecture is driven by two operators: \emph{(i)} the similarity operator, which is a generalization of the convolutional operator in ConvNets, and \emph{(ii)} the MEX operator, which can realize classical operators found in ConvNets like ReLU and max pooling, but has additional capabilities that make SimNets a powerful generalization of ConvNets. One of the interesting properties of the SimNet architecture is that applying its two operators in succession -- similarity followed by MEX, results in what can be viewed as an artificial neuron in a high-dimensional feature space. Moreover, the multilayer perceptron construction of input to hidden layer to output, as well as the fundamental building block incorporating locality, sharing and pooling, are both generalizations of kernel machines.
We described two possible similarity measures: the $l_p$ similarity, which in its unweighted version gives rise to the Generalized Gaussian kernel, and the linear similarity, which is the operator found in ConvNets, and gives rise to the Exponential kernel. We also showed that the full specification of the $l_p$ similarity operator, which includes weights, goes beyond a kernel machine and carries with it a higher abstraction level than what a convolutional layer can express. Another interesting property of the SimNet architecture is that statistical estimation methods for Generalized Gaussian mixture distributions can be used for unsupervised initialization of network parameters. These initializations arise naturally from standard statistical assumptions, having the potential of employing unsupervised learning in an effective manner as part of deep learning.
Implementing deep SimNets with state of the art optimization techniques (including GPU acceleration) is an ongoing effort, but we were able to implement a basic SimNet and conduct benchmarks comparing it against two networks of the same depth -- an analogous ConvNet and the ``single-layer'' network of~\cite{coates2011analysis}. The results demonstrate that a SimNet can achieve comparable and/or better accuracy, while requiring a significantly smaller network (in terms of the number of learned parameters) -- around $1/9$ the size of the ConvNet and $1/5$ the size of the network in~\cite{coates2011analysis}.
The SimNet architecture departs from classical ConvNets in three main respects. First, the similarity layer can incorporate entry-wise weights when the $l_p$ similarity is used. With linear similarity (which is essentially an inner-product between an input patch and a convolutional kernel) incorporating weights is meaningless, as they blend into the convolutional kernels. We saw that the unweighted $l_p$ and linear similarities give rise to a kernel-SVM building block with the Generalized Gaussian and Exponential kernels, respectively. The weighted $l_p$ similarity on the other hand, cannot be realized in the kernel-SVM framework (thm.~\ref{thm:l_p_sim_wgt_no_K}), thereby offering a potentially stronger building block (whose effect is described in more detail in subsec.~\ref{subsec:mlp}). The experiments we carried out highlight the differences between weighted and unweighted $l_p$ similarities:
\begin{itemize}
\item Without weights, $l_p$ similarity and the linear similarity (convolutional operator) give rise to comparable performance. This suggests that without weights, SimNets do not exhibit superiority over ConvNets.
\item When weights are included, $l_p$ similarity displays a significant increase in performance, which scales up super-linearly with the number of parameters. That is to say, the increase in accuracy cannot be explained merely by the fact that the number of parameters in the similarity layer has been doubled (weights on top of templates).
\end{itemize}
These findings suggest that the strength of having the basic building block go beyond the hypothesis space of a kernel-SVM, has significant appeal in practice.
The second respect in which SimNets depart from ConvNets has to do with the ability of the MEX layer to incorporate offsets. When the MEX layer serves as the final layer of the network, these offsets play the role of classification coefficients. However, when the MEX layer is inserted as a pooling layer, the offsets can be interpreted as providing locality-based biases to the templates generated in a previous similarity layer. This is something that classical ConvNets cannot express. Evaluating the practical significance of the MEX offsets requires experimentation with deep layered SimNets, which is an ongoing effort.
The third departure (or distinction) from ConvNets, is that the SimNet architecture is endowed with a natural initialization based on unlabeled data. In the case of ConvNets, existing unsupervised initialization schemes have little to no advantage over random initializations. For the SimNets, we reported experimental results that demonstrate the superiority of the unsupervised initialization scheme over random initializations, showing that the effect is more acute when the networks are small. Besides its aid in training, the unsupervised scheme proposed also has the potential of determining the number of channels for a similarity layer based on variance analysis of patterns generated from previous layers. This implies that the structure of SimNets can potentially be determined automatically from (unlabeled) training data.
Future work is focused on further implementation, with the purpose of creating an open programming environment for the research community, that will enable wider scale experimentation of SimNets. Further theoretical studies are ongoing as well, with the intent to capture the sample complexity of SimNets, and to gain a better understanding of the typical network structure and size required under different conditions.
\subsubsection*{Acknowledgments}
The authors would like to thank Nitzan Guberman and Or Sharir for their dedicated contribution to the experiments carried out in this work. The work is partly funded by Intel grant ICRI-CI no. 9-2012-6133 and by ISF Center grant 1790/12.
\subsubsection*{References}
{
\bibliographystyle{plainnat}
|
\section{Introduction}
Spatial indexing of geographic data has always been an important component of database systems. Since the wide-spread adoption of social media and social networks, the size of data to be indexed has grown by multiple orders of magnitude, making even more demand on the efficiency of indexing algorithms and index structures. Certain fields of natural sciences also face similar problems: Astronomers, for example, have to perform spatial searches in databases of billions of observations where spatial criteria can be arbitrarily complex spherical regions. Inspired by astronomical problems, Szalay et al. \cite{htm, htm2} came up with a solution to index spatial data stored in Microsoft SQL Server years before native support of geographic indexing appeared in the product. Their indexing scheme is called Hierarchical Triangular Mesh (HTM) and uses an iteratively refined triangulation of the surface of the sphere to build a quad-tree index. In the present case study, we demonstrate how we applied HTM to index real GPS coordinates collected from the open data streams of Twitter, and performed a huge spatial join on the data to classify the coordinates by the administrative regions of the world. Our results show that pre-filtering capabilites of HTM are significantly better than that of the built-in spatial index of SQL Server and that HTM renders spatial joins originally thought impossible to be able to be computed in a reasonable time. Source code and full queries are available at the following url: \url{http://www.vo.elte.hu/htmpaper}.
\section{Data and indexing}
Our data set consists of short messages (tweets) collected over a period of two years from the publicly available ``sprinkler'' data stream of Twitter. More than half of the tweets, over a billion, are geo-tagged. We built a Microsoft SQL Server database of the tweets \cite{twitterdb} and wanted to classify the geo-tagged messages by political administrative regions to investigate the geographic embedding of the social network of Twitter users.
Microsoft SQL Server supports spatial indexing since version~2008 via a library implemented on top of the integrated .Net framework runtime (SQL CLR). The library works by projecting the hemisphere onto a quadrilateral pyramid, then projecting the pyramid onto a square, and tessellating the square using four levels of fix-sized rectangular grids to construct a spatial index. The number of grid cells can be set between $4 \times 4$ and $16 \times 16$ providing a maximal index depth of 32~bits. The index structure itself is materialized as a hidden, but otherwise normal database table containing one row for each cell touched by the geography object being indexed. The table uses 11~bytes for the spatial index which is complemented by the primary key of the indexed object, in our case an additional 10~bytes. The final index size can be controlled by limiting the number of cells stored for each geography object resulting in less effective pre-filtering of spatial matches when the index size is kept small. The hard limit on the number of index entries per geography object is 8192.
Similarly to the built-in spatial index of SQL~Server, HTM indexing is also implemented in .Net. By default, HTM calculates a 40~bit deep hash, the so called HTM ID, from the coordinates. The 40~bit hash length corresponds to an almost uniform resolution of 10~meters on the surface of the Earth. HTM tessellates two dimensional shapes with small spherical triangles, called trixels. Trixels are represented by integer intervals; all coordinates with an HTM ID falling into the interval are guaranteed to be covered by the trixel. As HTM is a custom library, we had full control over the index structures, therefore we simply stored the 8~byte~HTM identifiers in the same table where the tweets were, and built an auxiliary index on the table ordered by the HTM identifier. Together with the primary key, the auxiliary index size was 18~bytes per row, exactly one row per coordinate pair.
We classified tweets using the maps from \url{gadm.org}, an open database of global administrative regions of every country in the world. The maps were loaded into the database as geography objects using the built-in geography type. For indexing, however, we decided to use HTM. This raised a problem because the HTM library was built for astronomical applications where regions on the sphere are better represented as unions of convex shapes contoured by great \textit{or} small circles than by vertices of polygons connected by great circles. This union-of-convexes representation is not appropriate for highly detailed complex maps as shapes would need to be decomposed into convexes first, a process that significantly increases the size of the data structures.
\begin{figure}
\centering
\includegraphics[width=\figurewidthA]{ct72}
\caption{HTM cover of the U.S. state of California with level~9 trixels.}
\label{fig:cali}
\end{figure}
Unfortunately, no code exists to directly compute the HTM tessellation of maps in the polygon representation, so we had to use another approach. By combining the HTM library and the built-in geographic library of SQL~Server, we determined the approximate HTM tessellation of regions up to a given precision by iteratively refining HTM trixels at the boundaries. Our solution, see Algorithm~\ref{lst:alg}, goes as follows. We construct a coarse tessellation of the region based on the bounding circle, then we intersect each trixel with the region using the built-in functions of SQL Server. If a trixel is completely inside the region it is added to the result set. Similarly, a completely disjoint trixel is discarded. Trixels intersecting with the boundary are refined into four sub-trixels and the algorithm is called recursively. Passing only the intersection of the original map with the trixel to the recursive call reduces the total runtime of the tessellation significantly. The algorithm uses the maximum depth of HTM trixels as a stop condition to limit the resolution of the tessellation but can be easily modified to use an upper limit on the number of trixels instead. Also, instead of trying to keep the index tables small, we store every trixel of the tessellation. Trixels on the deepest level which intersect with the boundary of the geography object are flagged as ``partial''. Figure~\ref{fig:cali} illustrates the results of the level~9 HTM covering of California with partial trixels in green.
\begin{algorithm}
\begin{algorithmic}
\Function{EvalTrixels}{region, trixellist, maxlevel}
\State retlist $\leftarrow \varnothing$
\ForAll{t \textbf{in} trixellist}
\If{region.STContains(t)} \Comment{Full trixel}
\State t.Partial $\leftarrow$ false \Comment{Flag as full}
\State retlist.Add(t)
\Else \Comment{Partial or disjoint trixel}
\State region2 $\leftarrow$ region.STIntersection(t)
\If{region2 $\neq \varnothing$} \Comment{Partial trixel}
\If{t.Level $\geq$ maxlevel}
\State t.Partial $\leftarrow$ true \Comment{Flag as partial}
\State retlist.Add(t)
\Else
\State tlist2 $\leftarrow$ t.Refine(t.Level+1)
\State retlist.AddRange(\Call{EvalTrixels}{region2, tlist2, maxlevel})
\EndIf
\EndIf
\EndIf
\EndFor
\State \Return retlist
\EndFunction
\end{algorithmic}
\caption{The function used for creating the HTM tessellation of a region. The function parameters are the region to be tessellated (region), the list of covering trixels to be refined (trixellist), and the maximum depth (maxlevel). The \texttt{STContains} and \texttt{STIntersection} functions are provided by the SQL~Server geography library. Trixels added to the result set are flagged as either full or partial.}
\label{lst:alg}
\end{algorithm}
\section{Spatial joins}
\begin{query}
\begin{lstlisting}
CREATE TABLE tweet (
ID bigint PRIMARY KEY,
HTMID bigint,
coord geography )
CREATE INDEX IX_tweet_htm
ON tweet ( HTMID )
CREATE SPATIAL INDEX IX_tweet_spatial
ON tweet ( coord )
CREATE TABLE region (
ID int PRIMARY KEY,
geo geography )
CREATE SPATIAL INDEX IX_region_spatial
ON region ( geo )
CREATE TABLE regionHTM (
ID int, --foreign key to region.ID
start bigint,
end bigint,
partial bit )
\end{lstlisting}
\caption{Simplified schema of our database used for benchmarking HTM.}
\label{query:schema}
\end{query}
In order to explain the internals of HTM indexing, we create the schema of our database with Query~\ref{query:schema}. The HTM-based pre-filtering of a spatial join between a table containing GPS coordinates and another containing the tessellations of complex regions requires an inner join with a \texttt{BETWEEN} operator in the join constraint. Query~\ref{query:htmfilter} is a simplified example of such pre-filtering query. We will refer to these types of queries as \textit{range joins}. As range joins are highly optimized in the database engine, we expect excellent pre-filtering performance. The \texttt{LOOP JOIN} hint is added to suggest a query plan that consist of scan of the \texttt{tweet} table, while doing index seeks on the much smaller \texttt{regionHTM} table. This is optimal as long as the HTM index of the regions can be kept in memory. Using the built-in spatial index of SQL Server, pre-filtering of a spatial join can be done with Query~\ref{query:sqlfilter}. It translates into a rather complex execution plan that uses the spatial index for table \texttt{tweet} only, while calculating the tessellation of the geography objects in table \texttt{region} during query execution, or vice versa. By specifying query hints, one can tell the server which spatial index to use, but it seems impossible to use the spatial indices on \textit{both} tables at the same time. This behavior has a tremendous impact on the performance of spatial joins when using the built-in indices.
In case of the SQL Server geography index, exact containment testing can be done by simply replacing the function call to \texttt{Filter} with \texttt{STContains}, as in Query~\ref{query:sqlcontains}. When using the HTM index, points in full trixels are already accurately classified with Query~\ref{query:htmfilter}, only points in partial trixels need further processing to filter out false positive matches. This is done in Query~\ref{query:htmcontains} which again relies on the spatial functions of SQL Server. Also note, that Query~\ref{query:htmcontains}, by referencing the column \texttt{region.geo}, uses the entire region for testing point containment. In case of computing the spatial join of billions of coordinates with a limited number of complex regions, it is well worth to pre-compute the intersections of partial trixels and regions first and use them for containment testing instead of the whole regions. Due to publicational constraints we omit the query but all performance metrics quoted in the paper are measured using the pre-computed intersections of the regions and partial trixels for exact containment testing.
\section{Performance evaluation}
We measured the index generation time for the 50 continental states (plus Washington D.C.) of the United States using two different depths ($8 \times 8$ and $16 \times 16$ grids) of the SQL Server geography index and three different depths (level 12, 14 and 16) of HTM. For comparison, the $8 \times 8$ resolution of the SQL Server index roughly corresponds to a level 12 HTM index and the $16 \times 16$ grid resolution corresponds to a level 16 HTM index. The benchmarks were run on a 16-core database server with 96~GB memory. As our dataset fits into memory, queries are basically CPU-limited. Index generation times are summarized in Table~\ref{tab:idxgen}. Note, that SQL Server executed the geography index generation on two threads, while HTM ranges were generated on a single thread only. While we have no control over the internals of geography indices, the iterative refining of HTM tessellation could be replaced with a smarter, multi-threaded one. Also, the size of the geography indices is internally limited to 8192 entries per region, while the HTM indices were calculated without pruning, ultimately resulting in much larger index sizes.
\begin{query}
\begin{lstlisting}
SELECT t.ID, h.ID, h.partial
FROM tweet t
INNER LOOP JOIN regionHTM h
ON t.htmID BETWEEN h.start AND h.end
\end{lstlisting}
\caption{Pre-filtering of a spatial join with HTM index.}
\label{query:htmfilter}
\end{query}
\begin{query}
\begin{lstlisting}
SELECT t.ID, r.ID
FROM tweet t INNER JOIN region r
ON r.geo.Filter(t.coord) = 1
\end{lstlisting}
\label{query:sqlfilter}
\caption{Pre-filtering of a spatial join with the SQL Server geography index.}
\end{query}
\begin{query}
\begin{lstlisting}
SELECT tweet.ID, region.ID
FROM tweet t INNER JOIN region r
ON r.geo.STContains(t.coord) = 1
\end{lstlisting}
\caption{Exact containment testing with the SQL Server geography index.}
\label{query:sqlcontains}
\end{query}
\begin{query}[h!]
\begin{lstlisting}
SELECT tweet.ID, regionHTM.ID
FROM tweet t INNER LOOP JOIN regionHTM h
ON t.HTMID BETWEEN h.start AND h.end
INNER JOIN region r ON r.ID = h.ID
WHERE h.partial = 0 OR
r.geo.STContains(t.coord) = 1
\end{lstlisting}
\caption{Classifying points from partial trixels.}
\label{query:htmcontains}
\end{query}
It is also rather instructive to compare the two indexing schemes by the false positive rates of pre-filtering. The results are listed in Table~\ref{tab:fp}. False positives rates of the HTM index are significantly lower in all cases, especially for higher index depths. In the case of the SQL Server spatial index, increasing the resolution does not help, but the opposite: it just makes things worse, as the number of index rows (and the resolution of the tessellation) is limited to a maximum of 8192 cells, insufficiently small for complex maps. Strictly limiting index size only helps when the number of shapes to be indexed is large and the shapes are relatively small and simple. When spatial indices fit into the memory, or at least can be read quickly from the disk, pre-filtering using range joins is expected to be significantly faster, even for indices with millions of rows, rather than exact containment testing against complex shapes.
To test the performance of point classification, we prepared three samples having approximately 300~thousand, 1~million, 5~million points in each, uniformly sampled from the original database. Some tests were also run using the entire data set of more than one billion tweets. The coordinates covered the entire world but the majority of them were within the continental United States. The geographical distribution of the samples is realistic and follows the population density weighted by the local Twitter usage rate. To evaluate index performance, we computed a spatial join between a the samples of GPS coordinates with different cardinality and the 51 regions, first with pre-filtering only, then with exact containment test. Results of pre-filtering are listed in Table~\ref{tab:prefilter}, while exact containment metrics are shown in Table~\ref{tab:total}. All queries were executed using cold buffers, thus include I/O and CPU times.
The spatial join performance of the HTM index turned out to be about a hundred times better than the performance of the built-in geography index. Pre-filtering itself is about a thousand times faster than the built-in index, and usually could be done in a few seconds for the smaller samples. Such short query times are hard to be measured correctly, and values show a significant scatter when the queries are repeated. The two main reasons behind the significantly better performance of HTM are: 1) HTM-based pre-filtering could benefit from the spatial index on both tables, whereas SQL Server's geography library only used the index for one of the tables and calculated the tessellation for the other table on the fly. 2) The extensive pruning of index entries resulted in a very high rate of false positives in case of SQL~Server's geography index. Because of the pruning, increasing the index resolution could not actually increase the resolution of the tessellation in case of the rather complex maps. By using a significantly larger, but still manageable index table, and by intersecting the trixels of the tessellations with the regions to reduce the complexity of exact containment testing, HTM indexing could reduce the cost of spatial joins to a minimum. Based on these results, it is clear that running the point classification using only the built-in geography index of SQL Server index is not a viable solution for any task similar to ours, namely, when the number of points is in the billions range.
\begin{table}[t]
\centering
\begin{tabular}{l|r|r}
index type & time [s] & index rows \\ \hline
geography $8 \times 8$ & 13,352 & 412,055 \\
geography $16 \times 16$ & 6,215 & 410,040 \\
HTM level 12 & 4,366 & 267,763 \\
HTM level 14 & 5,151 & 1,331,632 \\
HTM level 16 & 9,952 & 6,354,932
\end{tabular}
\caption{Index generation time and number of index rows of the regions.}
\label{tab:idxgen}
\end{table}
\begin{table}
\centering
\begin{tabular}{l|r|r|r|r}
index type & CO & IL & MD & WA \\ \hline
geography $8 \times 8$ & <0.01\% & 0.16\% & 3.62\% & 1.11\% \\
geography $16 \times 16$ & <0.01\% & 4.66\% & 22.43\% & 3.14\% \\
HTM level 12 & 0.01\% & 1.71\% & 4.82\% & 1.30\% \\
HTM level 14 & <0.01\% & 0.18\% & 1.84\% & 0.47\% \\
HTM level 16 & <0.01\% & 0.04\% & 0.53\% & 0.23\%
\end{tabular}
\caption{False positive rate of spatial join pre-filtering for the U.S. states Colorado, Illinois, Maryland and Washington. Note, that false positive rates depend on the actual distribution of tweets and not only on the geometry of the states.}
\label{tab:fp}
\end{table}
\begin{table}
\centering
\begin{tabular}{l|r|r|r|r}
index type & 300k [s] & 1M [s] & 5M [s] & 1G [s] \\ \hline
geography $8 \times 8$ & 223 & 780 & 5009 & - \\
geography $16 \times 16$ & 223 & 883 & 4053 & - \\
HTM level 12 & 1 & 1 & 2 & 194 \\
HTM level 14 & 2 & 2 & 5 & 266 \\
HTM level 16 & 7 & 4 & 5 & 232
\end{tabular}
\caption{Pre-filtering time for the spatial join query}
\label{tab:prefilter}
\end{table}
\begin{table}
\centering
\begin{tabular}{l|r|r|r|r}
index type & 300k [s] & 1M [s] & 5M [s] & 1G [s] \\ \hline
geography $8 \times 8$ & 295 & 915 & 4276 & - \\
geography $16 \times 16$ & 301 & 773 & 4273 & - \\
HTM level 12 & 12 & 24 & 139 & 2370 \\
HTM level 14 & 7 & 13 & 58 & 1299 \\
HTM level 16 & 8 & 10 & 42 & 1032
\end{tabular}
\caption{Total time for the spatial join query}
\label{tab:total}
\end{table}
\section{Conclusions}
In this paper, we investigated the feasibility of efficient classification of GPS coordinates of Twitter messages by geographic regions using a relational database management system, Microsoft SQL Server~2012. We evaluated the performance of the built-in spatial indexing technology side by side with a customized solution based on Hierarchical Triangular Mesh (HTM) indexing. The built-in spatial index was found to be inadequate to perform spatial joins between large sets of GPS coordinates (on the scale of billions) and complex geographic regions. We showed that our solution, a heuristic combination of existing techniques for handling spatial data in a relational database environment, can easily be a hundred times faster and makes the computation of the aforementioned spatial join available in reasonable time. We pointed out that the strength of HTM indexing is the great control the database programmer has on both the index structure and query plans (via hints). We also demonstrated that aggressive pruning of spatial indices is not a good idea when indexing of very complex regions is a requirement, as range-join-based pre-filtering is significantly faster than exact containment testing, even in case of millions of index entries. To make exact containment testing even faster, we pre-computed the intersections of the complex geographic regions and partial HTM trixels and use these much smaller shapes to filter out false positives.
\section{Acknowledgments}
The authors thank the partial support of the European Union and the European Social Fund through project FuturICT.hu
(grant no.: TAMOP-4.2.2.C-11/1/KONV-2012-0013),
and the OTKA 103244 grants.
EITKIC 12-1-2012-0001 project was partially supported by the Hungarian Government,
managed by the National Development Agency, and financed by the Research and
Technology Innovation Fund and the MAKOG Foundation.
|
\section{Conclusion}
This paper presented the general Volterra series and discussed some typical simplifying assumptions. Limitations of the finite $V_{(N,M)}$ class Volterra Series were discussed with respect to physical systems. These limitations can often be beneficial as they represent behaviors not commonly seen in physical systems. Perhaps the key limitation to the Volterra series is the number of parameters required to describe a model, Laguerre Polynomials were introduced as a tool, by which to estimate the Volterra kernels and reduce the total number of model parameters required for a given system. A novel algorithm is presented to generalize identification of discrete VL systems. The proposed algorithm allows flexibility of parameterization to fit any class of system that can be described by a $V_{(N,M)}, l_{r}^{a_{n,i}}$ Volterra-Laguerre model. An example using the proposed algorithm on experimental data showed that using variable parameters has a higher probability of fitting the data set with less error.
\newpage
\section{Added generalization for Ease of Practical Application}
In order to simplify practical application, generalizations were made to make the algorithm more easily scalable. Each of the generalizations will be discussed separately and are listed below:
\begin{enumerate}
\item Allow different nonlinear degree $N$ for each system input $i=1 \ldots I$
\item Allow different Laguerre series $l_{r}$ for each Volterra term $n=1 \ldots N \text{ and input } i=1 \ldots I$
\item Allow different Laguerre time scale $a_{n,i}$ for each Laguerre series $l_{r}^{a_{n,i}}$
\end{enumerate}
\subsection{Separate Volterra order for each system input}
The Volterra series can be used to model a large class of systems.\Cref{eq:GenVoltConst} (shown below for reference) is the general $N^{th}$ order Volterra series.
\begin{align}
y\left ( t \right )=&\int_{-\infty}^{\infty }h_1(\sigma_1)u(t-\sigma_1)d\sigma_1 \nonumber \\
+&\int_{-\infty}^{\infty }\int_{-\infty}^{\infty }h_2(\sigma_1,\sigma_2)u(t-\sigma_1)u(t-\sigma_2)d\sigma_1d\sigma_2 + \cdots \nonumber\\
+&\int_{-\infty}^{\infty }\cdots \int_{-\infty}^{\infty }h_N(\sigma_1,\cdots,\sigma_i)u(t-\sigma_1)\cdots \nonumber\\
&\qquad \qquad \qquad u(t-\sigma_N)d\sigma1\cdots d\sigma_N\nonumber
\end{align}
$N=1$ is an example of the first order Volterra series and can describe systems with a linear relationship between the inputs and output. Using $N=2$ allows the Volterra series to describe second order relationships between the inputs and output. It cannot be assumed that, in a general system all inputs will have the same relationship to the output. In fact this will rarely be the case.
\subsection{Separate Laguerre series for each Volterra term}
Considering a single input being mapped to an output using an $N^{th}$ order Volterra series. There will be $N$ different Laguerre polynomials to approximate each kernel of the Volterra series. Depending on the complexity of the relationship between one term of the Volterra series and the output; a different order $R$ of the Laguerre polynomial could be used for each separate term of the Volterra series. This option gives one the ability to use more or less Laguerre coefficients to aproximate a certain term of the Volterra series if needed. A general example of an input with an uncomplicated first order relationship to the output and a more complicated second order relationship to the output could take advantage of fewer laguerre polynomials to approximate the first order Volterra kernel and more laguerre polynomials to approximate the second order Volterra kernel. Myriad other scenarios exist where this flexibility would be useful/necessary especially when considering scenarios involving multiple inputs.
\subsection{Separate Laguerre time scale for each Laguerre series}
Earlier in this document the Laguerre time scale was discussed as an important parameter in the Laguerre polynomial.
The Laguerre polynomial acts as a filter with time constant $1/a$. The time scale should be chosen to be the time constant of the response that is being modeled. Since a separate time scale can be chosen for each Volterra term this also allows adaptations to differences in responses within the same input. Again similar to the example above consider an input with a low frequency first order relationship to the output and a high frequency second order relationship to the output. A situation that would probably occur more frequently would be two inputs with significantly different dynamics.
\subsection{A Generalized Algorithm}
With these three modifications the equations given in the previous section need to be modified. Assume that there are $D$ I/O points being considered.
\begin{align}
\tilde{y}(k) &= [\mathbf{U}_{k}^{[1]},\mathbf{U}_{k}^{[2]},\ldots,\mathbf{U}_k^{[N]}]\mathbf{\Theta} \label{eq:MISOMatrixAPCONLI}
\end{align}
Where:
\begin{align}
\mathbf{U}^{[n]} &= [\mathbf{U^{n}B^{n}}]^{[n]}, n=1,\ldots,N \\
\mathbf{U}^{n} &= [\mathbf{u^1},\mathbf{u^2},\ldots,\mathbf{u^I}]
\end{align}
\begin{align}
\mathbf{u^{i}}&=
\begin{bmatrix}
u_i(k) &u_i(k-1) &u_i(k-2) &\cdots &u_i(k-M) \\
u_i(k+1) &u_i(k) &u_i(k-1) &\ddots &\ddots \\
u_i(k+2) &u_i(k+1) &u_i(k) &\ddots &\ddots \\
\vdots &\ddots &\ddots &\ddots &\ddots \\
u_i(k+D) &u_i(k+D-1) &\cdots &\ddots &u_i(k+D-M)
\end{bmatrix}
, i=1,\ldots,I
\end{align}
\begin{align}
\mathbf{B^{n}} &=
\begin{bmatrix}
\mathbf{B_{1}^{n}} &\mathbf{0} &\cdots &\mathbf{0} \\
\mathbf{0} &\mathbf{B_{2}^{n}} &\ddots &\vdots \\
\vdots &\ddots &\ddots &\mathbf{0} \\
\mathbf{0} &\cdots &\mathbf{0} &\mathbf{B_{I}^{n}}
\end{bmatrix}
\end{align}
\begin{align}
\mathbf{B_{i}^{n}} &=
\begin{bmatrix}
l_1^{a_{n,i}}(0) & l_2^{a_{n,i}}(0) &\cdots &l_{R_{n,i}}^{a_{n,i}}(0) \\
l_1^{a_{n,i}}(1) & l_2^{a_{n,i}}(1) &\cdots &l_{R_{n,i}}^{a_{n,i}}(1) \\
\vdots &\vdots &\ddots &\vdots \\
l_1^{a_{n,i}}(M) & l_2^{a_{n,i}}(M) &\cdots &l_{R_{n,i}}^{a_{n,i}}(M) \\
\end{bmatrix}
\end{align}
Where: $i=1,\ldots,I$ and $n =1,\ldots,N$. Finally:
\begin{align}
\mathbf{\Theta} = [\theta_{1,1,1},&\theta_{1,1,2},\cdots,\theta_{1,1,R_{n,i}}, \nonumber\\
&\theta_{1,2,1},\cdots,\theta_{1,I,R_{n,i}},\cdots,\theta_{N,I,R_{N,i}}]^T
\end{align}
The explanations and definitions of the following symbols should be noted. Recall that $a^{[n]}$ is the $n^{th}$ reduced Kronecker product. $N$ is the maximum Volterra order of the inputs $N=max(N_i)$. $I$ is the number of inputs. $k$ denotes the time step at which identification will begin on the data set, while this can be anywhere in the data set such that $k-M > 1$ usually $k=M+1$. $\mathbf{B^n}$ is a block diagonal matrix, the boldface zeros represent zero matrices with appropriate dimensions. $a_{n,i}$ is the Laguerre time scale that pertains to Volterra term $n$ and input $i$. These values can be specified or can be calculated by optimization of an initial guess, recall that this is a global optimization and that local minima may be a problem. $R_{n,i}$ is the order of the Laguerre polynomial that will be used to fit each polynomial, these values are specified by the user. Finally $\theta_{n,i,r}$ is the $\theta$ (or Laguerre coefficient) corresponding to Volterra term $n$, input $i$, and Laguerre polynomial $r$.
\subsection{Effect of Laguerre Reduction}\label{sec:VoltLagParameterization}
The parameterization of the Volterra series was discussed earlier in this document. It was identified to be a significant impediment to the practical application of Volterra model identification because of the number of I/O that are required to confidently identify such a large number of model parameters.
The replacement of the Volterra kernel with the Laguerre polynomials allows a reduction of the required parameters. This occurs because the Laguerre functions approximate each Volterra kernel with $R$ parameters instead of $M$ parameters (assuming R is the same for each term). The total number of coefficients in the reduced Volterra model $V_{(N,R)}$ can be derived similarly to \cref{eq:VoltParam} and is shown below in \cref{eq:VoltLagParam}.
\begin{align}
C_{(N,R)}=R^N \label{eq:VoltLagParam}
\end{align}
\autoref{tab:VoltLagParameterization} below shows the amount of parameters for a Volterra-Laguerre model with $N$ Volterra terms and an $R^{th}$ order Laguerre polynomial.
\begin{table}[h]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
& N=1 & N=2 & N=3 & N=4 \\ \hline
R=1 & 1 & 1 & 1 & 1 \\ \hline
R=2 & 2 & 4 & 8 & 16 \\ \hline
R=3 & 3 & 9 & 27 & 81 \\ \hline
R=4 & 4 & 16 & 64 & 256\\ \hline
\end{tabular}
\caption{Number of Volterra-Laguerre parameters based on N and R}
\label{tab:VoltLagParameterization}
\end{table}
The number of required parameters for a Volterra-Laguerre model is much less than that of a Volterra model. For example, consider a Volterra system with $M=50$ and $N=3$. \autoref{tab:VolterraParameterization} shows that this would require approximately $8000$ parameters to describe. Fitting the same Volterra model with $N=3$ and Laguerre polynomials of order $R=3$. The number of parameters is reduced to $27$ which is less than $0.5$ percent of the number required for the Volterra series.
It is important to remember that reduction of the number of required model parameters is desirable because of the amount of required I/O data to identify them. For a fixed amount of I/O data the identification of the parameters will have some statistical confidence inversely proportional to the number of parameters. If the number of required parameters are decreased then the statistical confidence will go up. For industrial applications this means that less time can be spent collecting data to achieve the same (or better) statistical confidence in the model parameters. The other way of looking at it is that a better model can be made with the same amount of I/O data. Either way use of the Volterra-Laguerre series can be extremely beneficial.
\subsubsection{Example - Variable Parameters}
A short demonstration highlights the value of having an algorithm that allows variation to $N$, $l_{r}$, and $a_{n,i}$. Data for this example was borrowed from \citet{bachlin2010wearable}. The dataset is a multivariate time-series for freezing of gait in patients with Parkinson's disease. 'Trunk acceleration - horizontal forward acceleration [mg]' and 'Upper leg (thigh) acceleration - horizontal forward acceleration [mg]' were used $u^{1}(t)$ and $u^{2}(t)$ respectively. And, 'Ankle (shank) acceleration - horizontal forward acceleration [mg]' was $y(t)$ . Hundreds of simulations were run using samples drawn uniformly from the following domains: $N$, the nonlinear degree or Volterra order was taken from integers $[1,5]$ for each input, $l_{r}$ was drawn from integers $[2,4]$ for each volterra kernel, and $a_{n,i}$ was taken from the set of real numbers on $[0.005, 100]$ for each laguerre series. Hundreds more simulations were run using equal $N$, $l_{r}$, and $a_{n,i}$ as well.
Compiling the results from these simulations and plotting the corresponding distributions of the sum of the squared error (SSE) indicates the difference in the expected error from a scenario with $N$, $l_{r}$, and $a_{n,i}$ fixed (according to practice in current literature) as opposed to the case where they can be different.
\begin{figure}
\centering
\includegraphics[width=5in]{SimulationDistribution.png}
\caption{Distribution density of baseline simulations vs. variable parameter distributions.}
\label{fig:SimulationDistribution}
\end{figure}
Normalizing the means to the lowest value of SSE the difference between the baseline simulations and the variable parameter simulations is approximately $5.5$ percent. It should be noted that the simulations with variable parameters have a tighter distribution. It should also be noted that lower SSE is not necessarily desirable because of over-fitting. This example illustrates that, in general a VL model fit with non-equal parameters will have less error than a model with equal parameters.
\section{Introduction}
\label{sec:VolterraSeriesBackground}
The Volterra Series were first studied by Vito Volterra and were named after him. The first application of the Volterra series to the study of nonlinear systems was done by Norbert Wiener \citep[see][p.517]{Schetzen1980}. The time invariant series can be represented by \eqref{eq:GenVoltConst} below.
\begin{align}
y\left ( t \right )=&\int_{-\infty}^{\infty }h_1(\sigma_1)u(t-\sigma_1)d\sigma_1 \nonumber \\
+&\int_{-\infty}^{\infty }\int_{-\infty}^{\infty }h_2(\sigma_1,\sigma_2)u(t-\sigma_1)u(t-\sigma_2)d\sigma_1d\sigma_2 + \cdots \nonumber\\
+&\int_{-\infty}^{\infty }\cdots \int_{-\infty}^{\infty }h_N(\sigma_1,\cdots,\sigma_i)u(t-\sigma_1)\cdots \nonumber\\
&\qquad \qquad \qquad u(t-\sigma_N)d\sigma1\cdots d\sigma_N \label{eq:GenVoltConst}
\end{align}
Here $u$ represents the input to the system, $y$ is the system output and $h_n$ is known as the $n^{th}$ Volterra kernel, it can also be called the $n^{th}$ order impulse response. This terminology comes because the first term of \cref{eq:GenVoltConst} is the same as the convolution integral which relates the output of a system to the input and the $1^{st}$ order impulse response of the system\footnote{Note that \cref{eq:GenVoltConst}, and consequently most of the other equations in this paper represent\ SISO systems for compactness. MIMO systems can be accounted for by solving several MISO sets, and expanding the expression with $u_{i}(t) \text{, }i=1 \ldots I$ where $I$ is the total number of inputs. For illustrative purposes it is sufficient to look at SISO systems for now.}. Higher order terms of the Volterra series can be seen as higher order impulse responses. The terms \emph{Volterra kernel} and \emph{impulse response} will be used interchangeably in this document.
Several modifications can and should be made to \eqref{eq:GenVoltConst} before performing an identification. These modifications include: discretization of the integrals, modification of the limits of integration based on known characteristics of physical systems, and reduction of the Volterra kernel using the Laguerre polynomials to reduce the number of model parameters.
\section{The Laguerre Polynomials}\label{sec:LaguerrePolynomialsBackground}
Laguerre Polynomials are named after Edmond Laguerre. The Laguerre Polynomials are a series of orthogonal polynomials that can be used to reduce the number of coefficients required to describe a Volterra kernel. The application of orthogonal functions to identification and control is not new, see \citep{Schetzen1980}, \citep{dumont1991}, \citep{zheng1994}, \citep{makila1990approximation}, \citep{clement1982laguerre}, \citep{Mahmoodi2007}, \citep{Zheng1995},and \citep{CommunicationLee1960}. A mathematical review of orthogonal and orthonormal fucntions can be found in Appendix \ref{sec:Orthogonality}.
\subsection{Making the Laguerre Functions}
\subsubsection{Forming the General Laguerre Representation}
The Laguerre functions can be obtained by forming an orthonormal set from the linearly independent set of functions in \eqref{eq:RootLag}. Note that since the functions are only non-zero for $t\geq 0$ that the Laguerre functions will be orthonormal on the domain $[0,\infty)$.
\begin{align}
v_n =
\begin{cases}
0 &\text{for }t<0\\
(at)^ne^{-at} &\text{for }t\geq 0; n=0,1,2,\ldots \label{eq:RootLag}
\end{cases}
\end{align}
The Laguerre function can then be represented as:
\begin{equation}
l_n(t)=\sum_{m=0}^{\infty}c_{mn}v_m(t) \label{eq:LagDef}
\end{equation}
It must adhere to the orthonormal condition given in Equations \eqref{eq:OrthonormCond},\eqref{eq:OrthonormCond_a} and \eqref{eq:OrthonormCond_b} (see Appendix \ref{sec:Orthogonality}). This can be done by choosing the coefficients $c_{mn}$ in order to satisfy the given equations. Examples of this procedure for the time and frequency domain can be found in \citet[Sec. 16.1-16.2]{Schetzen1980}. It can be shown that the general expression for the Laguerre functions in the time domain is \cref{eq:GeneralLaguerre} below.
\begin{align}
l_n(t) = \sqrt{2a}\sum_{k=0}^{n}\frac{(-1)^{k}n!2^{n-k}}{k![(n-k)!]^2}(2at)^{n-k}e^{-at}
\label{eq:GeneralLaguerre}
\end{align}
\subsubsection{Laguerre Time Scale Factor -- $a$}\label{sec:LaguerreTimeScale}
There is some interesting discussion that should take place concerning the value $a$ in \cref{eq:GeneralLaguerre}. It is a factor by which the time scale of the Laguerre functions can be lengthened or shortened. It is often referred to as the \emph{Laguerre time scale} because of this. Examination of the frequency domain representation of the Laguerre polynomials shows that the $a$ term defines the time constant of a filter. This leads to more discussion on the use of Laguerre polynomials as filters. It is sufficient to know for the purposes of this paper that the parameter $a$ or rather $1/a$ is the time constant of the ``Laguerre Filter" and that because of this filter the Laguerre functions can reject noise in experimental data if it is chosen properly.
\Cref{eq:GenLagFreq,,eq:GenLagFreqFilt} below are frequency domain representations of the Laguerre polynomials and make the filter more obvious. Observe that $\alpha$ is a low pass filter and $\beta$ is an all-pass filter making the net effect of the Laguerre Functions a filter with time constant $1/a$. For more detail on the frequency domain representation of the Laguerre Functions see \citet[Sec. 16.2]{Schetzen1980}. For more information on the Laguerre filter see \citet{Silva1995} and \cite{king1977}.
\begin{align}
L_n(s)&=\sqrt{2a}\frac{(a-s)^n}{(a+s)^{n+1}}; \sigma > -a \label{eq:GenLagFreq}\\
&\text{or} \nonumber \\
L_n(s)&=\underbrace{\left[\frac{\sqrt{2a}}{a+s}\right]}_{\alpha}\underbrace{\left[\frac{(a-s)}{(a+s)}\right]^n}_{\beta}; \sigma > -a \label{eq:GenLagFreqFilt}
\end{align}
Research has been performed to identify the optimal time scale factor $a$ for a given identification problem. It has been suggested that the factor should be placed near the dominant pole of the system \citep{Zheng1995}, however if the system has delay the factor $a$ will be greatly affected \citep{wang1994optimal}. Methods exist for calculating the optimal time scale value for linear systems, or require previous knowledge of the system and thus are not generally applicable to nonlinear system identification \citep{Clowes1965}, \citep{Fu1993}, and \citep{parks1971}. Current general practice is to perform a nonlinear optimization to calculate the value for $a$ that will yield the minimum error.
If the value for $a$ is optimal the coefficients of the higher order terms of the Laguerre polynomial will go to zero. This is valuable because the main purpose of using the Laguerre functions is to reduce the number of parameters that need to be identified for a $V_{(N,M)}$ model. If the value of $a$ is not optimal then the Laguerre functions can still be used but a higher order Laguerre polynomial will be required. Some further discussion of properties can be found in Appendix \ref{sec:LaguerreProps}.
\section{Simplifying the Volterra Series for Practical Application}
If a system is \emph{causal}, which means that the output at some time $t$ depends only on past inputs($u(t-\sigma)$ for $\sigma>0$) and not on future inputs ($u(t-\sigma)$ for $\sigma<0$). The Volterra series can be written as shown in \cref{eq:GenVoltCausal} below. Note that \eqref{eq:GenVoltCausal} only includes the first term of the series for simplicity. It should also be noted that all known physical systems appear to be \emph{causal}. A more detailed discussion of \emph{causality} can be found in \citet[p.21,89]{Schetzen1980}.
\begin{align}
y\left ( t \right )=&\int_{0}^{\infty }h_1(\sigma_1)u(t-\sigma_1)d\sigma_1
\label{eq:GenVoltCausal}
\end{align}
The Volterra series may also be discretized by using the convolution sum instead of the convolution integral, yielding \cref{eq:GenVoltDisc}. Again, only the first term of the series is shown for simplicity.
\begin{align}
y\left ( t \right )=\sum_{i_1=0}^{\infty }h_1(i_1)u(t-i_1)
\label{eq:GenVoltDisc}
\end{align}
Finally, if the system is assumed to have \emph{finite memory} or \emph{fading memory} and finite order another simplification can be made. \emph{Fading Memory} means that the there is some time $M$ in the past before which inputs will no longer have affect on the output of the system. \Cref{eq:GenVoltDiscFiniteMem} is referred to as the discrete, finite memory, $N^{th}$ order Volterra Series.
\begin{align}
y\left ( t \right )=&\sum_{n=1}^{N} \nu_{M}^{n}(t) \label{eq:GenVoltDiscFiniteMem}\\
\nu_{M}^{n}(t) =&\sum_{i_1=0}^{M}\cdots \sum_{i_n=0}^{M}h_n(i,\ldots,i_n)u(t-i_1)\cdots u(t-i_n) \nonumber
\end{align}
\subsection{Volterra Model Limitations}
This class of finite Volterra models is defined as the class of $V_{(N,M)}$ models by \citet{DoyleOgunnaikePearson2001}. Where $N$ is the nonlinear degree and $M$ is the dynamic order. In other words $N$ is the number of Volterra terms and $M$ is the memory length of the system. Using this notation it is easy to describe different Volterra Models by examining the behaviors of the limiting cases. These are: $V_{(\infty,M)}$,$V_{(N,\infty)}$, and $V_{(\infty,\infty)}$ \citep[Ch. 2]{DoyleOgunnaikePearson2001}, \citep[Sec.4.2]{Pearson1999}.
It is important to consider the limitations of the $V_{(N,M)}$ class. Some of these limitations include not being able to exhibit \emph{output multiplicity} \citep{boyd1985fading}. This can be described intuitively by saying that if a system can exhibit the same output by different local inputs (i.e. different steady state responses to the same steady state input), it must have had paths that differed initially. This leads to a similar conclusion which says conditionally stable impulse responses cannot be described by a \emph{fading memory} Volterra model. Volterra models also cannot produce persistent oscillations or chaos in response to asymptotically constant input sequences.
The limitations of the Volterra series can be seen as beneficial or detrimental depending on the desired output of the model. If a model for a system with persistent oscillations is desired then VL models should not be used. However, it is useful to have a model that implicitly rejects these types of characteristics if the physical system does not exhibit them. More information on this subject can be found in \citet{DoyleOgunnaikePearson2001,Pearson1999}
\subsection{Volterra Model Parameterization}\label{sec:VoltParameterization}
Another important practical limitation, that isn't dependent on the application, is \emph{Volterra Model Parameterization}. In other words how many parameters are required to define a $V_{(N,M)}$ model. The total number of parameters can be represented as $C_{(N,M)}$ \citep{DoyleOgunnaikePearson2001}. The following equations describe the calculation of $C_{(N,M)}$.
\begin{align}
C_{(N,M)} &= \sum_{n=0}^{M} C_n(M) \\
C_n(M) &= (M+1)^n \nonumber\\
C_{(N,M)} &= \sum_{n=0}^{M} C_n(M) = \frac{(M+1)^{N+1}-1}{M}\label{eq:VoltParam} \\
&\simeq M^N \nonumber
\end{align}
Here $C_n(M)$ is the total number of coefficients in $h_n$(The $n^{th}$ Volterra Kernel) of the Volterra model $V_{(N,M)}$. Although \citet[p.19]{DoyleOgunnaikePearson2001} discusses methods for reducing the number of coefficients, the relationship is still exponential and therefore remains a barrier for practical application. \Cref{tab:VolterraParameterization} below demonstrates how quickly the number of required parameters can grow, especially considering that $M$ is regularly between $50$ and $250$ in many industrial processes. The number of model parameters required makes any Voterra model with $N > 2$ impratical.
\begin{table}[htb]
\centering
\begin{tabular}{|r|r|r|r|r|}
\hline
& N=1 & N=2 & N=3 & N=4 \\ \hline
M=1 & 1 & 1 & 1 & 1 \\ \hline
M=2 & 2 & 4 & 8 & 16 \\ \hline
M=3 & 3 & 9 & 27 & 81 \\ \hline
M=4 & 4 & 16 & 64 & 256\\ \hline
M=10& 10 & 100& 1000& 10\,000\\ \hline
M=20& 20 & 400& 8000& 16\,000\\ \hline
M=50& 50 & 2500& 125\,000& 6\,250\,000\\ \hline
\end{tabular}
\caption{Number of Volterra Parameters based on N and M}
\label{tab:VolterraParameterization}
\end{table}
\section{Laguerre Estimation of the Volterra Kernel}
Recall the first order discrete Volterra kernel given in \cref{eq:GenVoltDisc} (shown below for reference).
\begin{align}
y\left ( t \right )=\sum_{i_1=0}^{\infty }h_1(i_1)u(t-i_1) \nonumber
\end{align}
The only unknown is $h_1(i_1)$, the $1^{st}$ order impulse response, since for system identification both $y(t)$ and $u(t)$ are recorded I/O data. $h_1(i_1)$ can be approximated by linear combination of the Laguerre functions.
In the case of the first order Volterra-Laguerre series. The first order Volterra kernel ($h_1(i_1)$) is approximated by linear combination of an $r^{th}$ order Laguerre polynomial. $h_1(i_1)$ meets the requirement of its square being finite over the interval through which the Laguerre functions are orthonormal (see \eqref{eq:ISErestrict}).The formulation is shown below:
\begin{align}
h_1(i_1)\approx \sum_{r=1}^{R} \theta_rl_r(t)
\label{eq:VoltLag1OrderKernel}
\end{align}
Here, $l_r(t)$ is given by \eqref{eq:GeneralLaguerre} and is shown below for reference.
\begin{align}
l_r(t) = \sqrt{2a}\sum_{k=0}^{r}\frac{(-1)^{k}r!2^{r-k}}{k![(r-k)!]^2}(2at)^{r-k}e^{-at} \nonumber
\end{align}
Substituting \cref{eq:VoltLag1OrderKernel} into \cref{eq:OrthogApprox} and truncating both $h_1(i_1)$ and $l_r(t)$ to a memory length of $M$ yields:
\begin{align}
y(t)\approx \sum_{i_1=0}^{M}\sum_{r=1}^{R}\theta_rl_r(t)u(t-i_1)
\end{align}
Now, defining the following:
\begin{align}
\mathbf{\Theta} &= [\theta_1,\theta_2,\ldots,\theta_R]^T \\
\mathbf{B} &=
\begin{bmatrix}
l_1(0) & l_2(0) &\cdots &l_R(0) \\
l_1(1) & l_2(1) &\cdots &l_R(1) \\
\vdots & \vdots &\ddots &\vdots \\
l_1(M) & l_2(M) &\cdots &l_R(M) \\
\end{bmatrix} \\
\mathbf{U_k}&=[u(k),u(k-1),\ldots,u(k-M)]
\end{align}
Then the Volterra system can be approximated by:
\begin{align}
\tilde{y}(k) = \mathbf{U}_k\mathbf{B\Theta}
\end{align}
In order to extend the representation to higher order Volterra series for a MIMO system it is first useful to define the reduced Kronecker product as \eqref{eq:RKronecker} \citep[p.100]{Rugh1981} :
\begin{align}
a^{[2]}=a\otimes a=&[a_1,a_2,\ldots,a_n]^{[2]}= \nonumber \\
&[a_1a_1,a_1a_2,\ldots,a_2a_2,a_2a_3,\ldots,a_na_n] \label{eq:RKronecker}
\end{align}
Using the reduced Kronecker product notation above a general MISO Volterra-Laguerre series with $I$ inputs can be approximated by \eqref{eq:MISOMatrix} below. For a MIMO system the separate MISO solutions can be combined.
\begin{align}
\tilde{y}(k) &= [\mathbf{U}_k,\mathbf{U}_k^{[2]},\ldots,\mathbf{U}_k^{[N]}]\mathbf{\Theta} \label{eq:MISOMatrix}\\
\text{where} \nonumber \\
\mathbf{U}_k^i &= [u_i(k),\ldots,u_i(k-m)], i=1,\ldots,I \\
\mathbf{U}_k &= [\mathbf{U}_k^1\mathbf{B},\ldots,\mathbf{U}_k^I\mathbf{B}]
\end{align}
This notation was originally derived in \citet{zheng2004volterra}.
|
\section{Introduction}
In this paper we compute the curvature of the determinant line bundle associated to a family of Dirac operators on the noncommutative two torus. Following Quillen's pioneering work \cite{Quillen1985}, and using zeta regularized determinants, one can endow the determinant line bundle over the space of Dirac operators on the noncommutative two torus with a natural Hermitian metric. Our result computes the curvature of the associated Chern connection on this holomorphic line bundle. In the noncommutative case the method of proof applied in \cite{Quillen1985} does not work and we had to use a different strategy. To this end we found it very useful to extend the canonical trace of Kontsevich-Vishik \cite{Kontsevich-Vishik1995} to the algebra of pseudodifferential operators on the noncommutative two torus.
This paper is organized as follows. In Section 2 we review some standard facts about Quillen's determinant line bundle on the space of Fredholm operators from \cite{Quillen1985}, and
about noncommutative two torus that we need in this paper.
In Section 3 we develop the tools that are needed in our computation of the curvature of the determinant line bundle in the noncommutative case.
We recall Connes' pseudodifferential calculus and define an analogue of the Kontsevich-Vishik trace for classical pseudodifferential symbols on the noncommutative torus.
A similar construction of the canonical trace can be found in \cite{Paycha-Levy2014}, where one works with the algebra of toroidal symbols.
Section 4 is devoted to Cauchy-Riemann operators on $\mathcal{A}_{\theta}$ with a fixed complex structure. This is the family of elliptic operators that we want to study its determinant line bundle.
In Section 5 using the calculus of symbols and the canonical trace we compute the curvature of determinant line bundle. Calculus of symbols and the canonical trace allow us to bypass local calculations involving Green functions in \cite{Quillen1985}, which is not applicable in our noncommutative case.
The study of the conformal and complex geometry of the noncommutative two torus started with the seminal work \cite{Connes-Tred2009} (cf. also \cite{Connes-Cohen1992} for a preliminary version) where a Gauss-Bonnet theorem is proved for a noncommutative two torus equipped with a conformally perturbed metric.
This result was refined and extended in \cite{Masoud-Farzad2012} where the Gauss-Bonnet theorem was proved for metrics in all translation invariant conformal structures.
The problem of computing the scalar curvature of the curved noncommutative two
torus was fully settled in \cite{Connes-Moscovici2014}, and, independently, in \cite{Farzad-Masoud2013}, and in \cite{Farzad-Masoud2014} for the four dimensional case. Other related works include
\cite{Marcolli-Tanvir2012, Ali-Masoud2014, dabsit1, dabsit2, Ros, Lesch2}.
\section{Preliminaries}
In this section we recall the definition of Quillen's determinant line bundle over the space of Fredholm operators. We also recall some basic notions about noncommutative torus that we need in this paper.
\subsection{The determinant line bundle}
Unless otherwise stated, in this paper by a Hilbert space we mean a separable infinite dimensional Hilbert space over the field of complex numbers. Let $ \mathcal{F} = {\rm Fred} (\mathcal{H}_0, \mathcal{H}_1)$ denote the set of Fredholm operators between Hilbert spaces $\mathcal{H}_0$ and $ \mathcal{H}_1$. It is an open subset, with respect to norm topology, in the complex Banach space of all bounded linear operators between $\mathcal{H}_0$ and $ \mathcal{H}_1$. The index map $ index : \mathcal{F} \to \mathbb{Z}$ is a homotopy invariant and in fact defines a bijection between connected components of
$\mathcal{F}$ and the set of integers $\mathbb{Z}$.
It is well known that $\mathcal{F}$ is a classifying space for $K$-theory: for any compact space $X$ we have a natural ring isomorphism
$$ K^0 (X) = [ X, \mathcal{F}]$$
between the $K$-theory of $X$ and the set of homotopy classes of continuous maps from $X$ to
$\mathcal{F}.$ In other words, homotopy classes of continuous families of Fredholm operators parametrized by $X$ determine the $K$-theory of $X$.
It thus follows that $\mathcal{F}$ is homotopy equivalent to $\mathbb{Z} \times BU$, the latter being also a classifying space for $K$-theory. Let $\mathcal{F}_0$ denote the set of Fredholm operators with index zero. By
Bott periodicity, $\pi_{2j}(\mathcal{F})\cong \mathbb{Z}$ and $\pi_{2j+1}(\mathcal{F}) = \{0\}$ for $j\ge 0$.
So by Hurewicz's theorem, $H^2(\mathcal{F}_0, Z) \cong \mathbb{Z}.$ Now the determinant line bundle ${\rm DET}$ defined below has the property that its first Chern class,
$c_1({\rm DET}),$ is a generator of $H^2(\mathcal{F}_0, \mathbb{Z}) \cong \mathbb{Z}$. We refer to \cite{Scott2010, Chakraborty-Mathai2009} and references therein for details.
In \cite{Quillen1985} Quillen defines a line bundle $\text{DET} \to \mathcal{F}$ such that for any $T \in \mathcal{F}$
$${\rm DET}_T = \Lambda^{max}({\rm ker}(T))^* \otimes \Lambda^{max}({\rm coker}(T)).$$
This is remarkable if we notice that ${\rm ker}(T)$ and ${\rm coker}(T)$ are not vector bundles due to
discontinuities in their dimensions as $T$ varies within $\mathcal{F}$. Let us briefly recall the construction of
this determinant line bundle DET. For each finite dimensional subspace $F$ of $ \mathcal{H}_1$ let
$U_F = \{ T\in \mathcal{F}_1: {\rm Im}(T) + F = \mathcal{H}_1\}$ denote the set of Fredholm operators whose range is transversal to $F$. It is an open subset of $\mathcal{F}$ and we have an open cover
$ \mathcal{F} = \bigcup U_F$.
For $T\in U_F$, the exact
sequence
\begin{equation}\label{eqn:index1}
0\to {\rm ker}(T) \to T^{-1}F \stackrel{T}{\to} F \to {\rm coker}(T)\to 0
\end{equation}
shows that the rank of
$T^{-1}F$ is constant when $T$ varies within a continuous family in $ U_F$.
Thus we can define a vector bundle $\mathcal{E}^F\to U_F$
by setting $\mathcal{E}^F_T = T^{-1}F.$ We can then define a line bundle ${\rm DET}^F \to U_F$ by setting
$$ {\rm DET}^F_T= \Lambda^{max} (T^{-1}F)^* \otimes \Lambda^{max} F.$$
We can use the inner products on $\mathcal{H}_0$ and $\mathcal{H}_1$ to split the above exact sequence \eqref{eqn:index1} canonically and get a canonical isomorphism
$ {\rm ker}(T) \oplus F \cong T^{-1}F \oplus {\rm coker}(T)$. Therefore
$$
\Lambda^{max}({\rm ker}(T))^* \otimes \Lambda^{max}({\rm coker}(T))\cong
\Lambda^{max} (T^{-1}F)^* \otimes \Lambda^{max} F.
$$
Now over each member of the cover $U_F$ a line bundle
${\rm DET}^F \to U_F$ is defined. Next one shows that over intersections $U_{F_1} \cap U_{F_2}$ there is an isomorphism ${\rm DET}^{F_1} \to {\rm DET}^{F_2} $ and moreover the isomorphisms satisfy a
cocycle condition over triple intersections $U_{F_1} \cap U_{F_2} \cap U_{F_3}.$ This shows that the line bundles ${\rm DET}^F \to U_F$ glue together to define a line bundle over $\mathcal{F}$. It is further shown in \cite{Quillen1985} that this line bundle is holomorphic as a bundle over an open subset of a complex Banach space.
It is tempting to think that since $c_1 (\text{DET})$ is the generator of $H^2(\mathcal{F}_0, \mathbb{Z}) \cong \mathbb{Z},$ there might exits a natural Hermitian metric on DET whose curvature 2-form would be a representative of this generator. One problem is that the induced metric from ${\rm ker}(T)$ and ${\rm ker}(T^*)$ on DET is not even continuous.
In \cite{Quillen1985} Quillen shows that for families of Cauchy-Riemann operators on a Riemann surface one can correct the Hilbert space metric by multiplying it by zeta regularized determinant and in this way one obtains a smooth Hermitian metric on the induced determinant line bundle. In Section 5 we describe a similar construction for noncommutative two torus.
\subsection{Noncommutative two torus}
For $\theta \in \mathbb{R}, $ the noncommutative two torus $A_{\theta}$ is by definition
the universal unital $C^*$-algebra generated by two unitaries $U, V$ satisfying
\[VU=e^{2 \pi i \theta} UV.\]
There is a continuous action of $\mathbb{T}^2$, $\mathbb{T}= \mathbb{R}/2\pi \mathbb{Z}$, on $A_{\theta}$ by $C^*$-algebra
automorphisms $\{ \alpha_s\}$, $s\in \mathbb{R}^2$, defined by
\[\alpha_s(U^mV^n)=e^{is.(m,n)}U^mV^n.\]
The space of smooth elements for this action will be denoted by
$A_{\theta}^{\infty}$. It is a dense subalgebra of $A_{\theta}$ which can be alternatively
described as the algebra of elements in $A_{\theta}$
whose (noncommutative) Fourier expansion has rapidly decreasing coefficients:
\[A_{\theta}^{\infty}=\left\{\sum_{m,n\in \mathbb{Z}}a_{m,n}U^mV^n: a_{m,n}\in\mathcal{S}(\mathbb{Z}^2)\right\}.\]
There is a normalized, faithful and positive, trace $\varphi_0$ on $A_{\theta}$ whose restriction on smooth elements is given by
\[\varphi_0(\sum_{m,n\in \mathbb{Z}}a_{m,n}U^mV^n)=a_{0,0}.\]
The algebra $A_{\theta}^{\infty}$ is equipped with the
derivations $\delta_1, \, \delta_2: A_{\theta}^{\infty} \to A_{\theta}^{\infty}$, uniquely defined by the relations
\[\delta_1(U)=U, \,\, \delta_1(V)=0, \quad \delta_2(U)=0, \,\, \delta_2(V)=V.\]
We have $\delta_j(a^*)= -\delta_j(a)^* $ for $j=1, 2$ and all $a\in A_{\theta}^{\infty}$.
Moreover, the
analogue of the
integration by parts formula in this setting is given by:
\[ \varphi_0(a\delta_j(b)) = -\varphi_0(\delta_j(a)b), \,\,\, \forall a,b \in A_{\theta}^{\infty}. \]
We apply the GNS construction to $A_{\theta}$. The state $\varphi_0$ defines an inner product
\[ \langle a, b \rangle = \varphi_0(b^*a), \,\,\, a,b \in A_{\theta}, \nonumber \]
and a pre-Hilbert structure on $A_{\theta}$. After completion we obtain a Hilbert space denoted
$\mathcal{H}_0$. The derivations $\delta_1, \delta_2$, as densely defined unbounded
operators on $\mathcal{H}_0$, are formally selfadjoint and have unique extensions to selfadjoint operators.
We introduce a complex structure associated with a complex number $\tau = \tau_1+i\tau_2, \, \tau_2 >0,$
by defining
\[ \bar{\partial} = \delta_1 + \tau \delta_2, \,\,\, \bar{\partial}^*= \delta_1 + \overline{\tau} \delta_2. \]
Note that $\bar{\partial}$ is an unbounded operator on $\mathcal{H}_0$ and $\bar{\partial}^*$ is
its formal adjoint. The analogue of the space of anti-holomorphic 1-forms on the ordinary two torus
is defined to be $$\Omega^{0,1}_{\theta}=\left\{\sum a \bar{\partial} b\;, a,b
\in A_{\theta}^{\infty}\right\}.$$
Using the induced inner product from $\psi $, one can turn $\Omega^{0,1}_{\theta}$ into a Hilbert space which we denote by
$\mathcal{H}^{0,1}$.
\section{The canonical trace and noncommutative residue}
In this section we define an analogue of the canonical trace of Kontsevich and Vishik \cite{Kontsevich-Vishik1995} for the noncommutative torus. Let us first recall the algebra of pseudodifferential symbols on the noncommutative torus \cite{Connes1980, Connes-Tred2009}.
\subsection{Pseudodifferential calculus on $\mathcal{A}_{\theta}$}
Using operator valued symbols, one can define an algebra of pseudodifferential operators on $A_{\theta}^{\infty}$. We shall use the notation
$\partial^{\alpha}=\frac{\partial^{\alpha_1}}{\partial \xi_1^{\alpha_1}} \frac{\partial^{\alpha_2}}{\partial \xi_2^{\alpha_2}} $, and $\delta^{\alpha}= \delta_1^{\alpha_1}\delta_2^{\alpha_2}, $ for a multi-index $\alpha= (\alpha_1, \alpha_2).$
\begin{definition}
For a real number $m$, a smooth map $\sigma: \mathbb{R}^2 \to A_{\theta}^{\infty}$ is said
to be a symbol of order $m$, if for all non-negative integers $i_1, i_2, j_1,
j_2,$
\[ ||\delta^{(i_1, i_2)} \partial^{(j_1, j_2)} \sigma(\xi) ||
\leq c (1+|\xi|)^{m-j_1-j_2},\]
where $c$ is a constant, and if there exists a smooth map $k: \mathbb{R}^2 \to
A_{\theta}^{\infty}$ such that
\[\lim_{\lambda \to \infty} \lambda^{-m} \sigma(\lambda\xi_1, \lambda\xi_2) = k (\xi_1, \xi_2).\]
The space of symbols of order $m$ is denoted by ${\mathcal S}^m(\mathcal{A}_{\theta})$.
\end{definition}
\begin{definition}\label{pseudodef}
To a symbol $\sigma$ of order $m$, one can associate an operator on $A_{\theta}^{\infty}$,
denoted by $P_{\sigma}$, given by
$$ P_{\sigma}(a) = \int \int e^{-is \cdot \xi} \sigma(\xi) \alpha_s(a) \,ds \,
d\xi.$$
Here, $d\xi=(2\pi)^{-2}d_L\xi$ where $d_L\xi$ is the Lebesgue measure on $\mathbb{R}^2$.
The operator $P_{\sigma}$ is said to be a pseudodifferential operator of order $m$.
\end{definition}
For
example, the differential operator $\sum_{j_1+j_2 \leq m } a_{j_1,j_2}
\delta^{(j_1, j_2)} $ is associated with the symbol $\sum_{j_1+j_2 \leq m } a_{j_1,j_2}
\xi_1^{j_1} \xi_2^{j_2}$ via the above formula.
Two symbols $\sigma$, $\sigma'\in {\mathcal S}^m(\mathcal{A}_{\theta})$ are said to be equivalent if and only if $\sigma-\sigma'\in {\mathcal S}^n(\mathcal{A}_{\theta})$ for all integers $n$. The equivalence of the symbols will be denoted by $\sigma \sim \sigma'$.
Let $P$ and $Q$ be pseudodifferential operators with the symbols
$\sigma$ and $\sigma'$ respectively. Then the adjoint $P^*$ and
the product $PQ$ are pseudodifferential operators with the following
symbols
\[
\sigma(P^*) \sim \sum_{ \ell = (\ell_1, \ell_2) \geq 0} \frac{1}{\ell ! }
\partial^{\ell} \delta^{\ell}
(\sigma(\xi))^*,
\]
\[
\sigma (P Q) \sim \sum_{\ell = (\ell_1, \ell_2) \geq 0} \frac{1}{\ell !}
\partial^{\ell} (\sigma (\xi))
\delta^{\ell}(\sigma'(\xi)).
\]
\begin{definition}
A symbol $\sigma\in\mathcal{S}^m(\mathcal{A}_{\theta})$ is called elliptic if $\sigma(\xi)$ is invertible for $\xi\neq0,$ and for some $c$
$$||\sigma(\xi)^{-1}||\leq c(1+|\xi|)^{-m},$$ for large enough $|\xi|$.
\end{definition}
A smooth map $\sigma:\mathbb{R}^2\to \mathcal{A}_{\theta}$ is called a classical symbol of order $\alpha \in \mathbb{C}$
if for any $N$ and each $0\leq j\leq N$ there exist $\sigma_{\alpha-j}:\mathbb{R}^2\backslash\{0\}\to\mathcal{A}_{\theta}$ positive homogeneous of degree $\alpha-j$, and a symbol $\sigma^N\in\mathcal{S}^{\Re(\alpha)-N-1}(\mathcal{A}_{\theta})$, such that
\begin{equation}
\label{sigma}
\sigma (\xi)=\sum_{j=0}^{N}\chi(\xi)\sigma_{\alpha-j}(\xi)+\sigma^N(\xi)\quad\xi\in\mathbb{R}^2.
\end{equation}
Here $\chi$ is a smooth cut off function on $\mathbb{R}^2$ which is equal to zero on a small ball around the origin, and is equal to one outside the unit ball.
It can be shown that the homogeneous terms in the expansion are uniquely determined by $\sigma$.
We denote the set of classical symbols of order $\alpha$ by $\mathcal{S}^{\alpha}_{cl}(\mathcal{A}_{\theta})$ and the associated classical pseudodifferential operators by $\Psi_{cl}^{\alpha}(\mathcal{A}_{\theta})$.
The space of classical symbols $\mathcal{S}_{cl}(\mathcal{A}_{\theta})$ is equipped with a Fr\'{e}chet topology induced by the semi-norms
\begin{equation}\label{frechet}
p_{\alpha,\beta}(\sigma)=\sup_{\xi\in\mathbb{R}^2}(1+|\xi|)^{-m+|\beta|}||\delta^{\alpha}\partial^{\beta}\sigma(\xi)||.
\end{equation}
The analogue of the Wodzicki residue for classical pseudodifferential operators on the noncommutative torus is defined in \cite{Farzad-Wong2011}.
\begin{definition}
The Wodzicki residue of a classical pseudodifferential operator $P_{\sigma}$ is defined as
$${\rm Res}(P_{\sigma})=\varphi_0\left({\rm res} (P_\sigma)\right),$$
where ${\rm res}(P_\sigma):=\int_{|\xi|=1}\sigma_{-2}(\xi)d\xi$.
\end{definition}
It is evident from its definition that Wodzicki residue vanishes on differential operators and on non-integer order classical pseudodifferential operators.
\subsection{The canonical trace}
In what follows, we define the analogue of Kontsevich-Vishik trace \cite{Kontsevich-Vishik1995} on non-integer order pseudodifferential operators on the noncommutative torus. For an alternative approach based on toroidal noncommutative symbols see \cite{Paycha-Levy2014}. For a thorough review of the theory in the classical case we refer to \cite{Paycha2012,Paycha-Scott2007}. First we show the existence of the so called cut-off integral for classical symbols.
\begin{proposition}\label{expansioninR}
Let $\sigma\in\mathcal{S}_{cl}^{\alpha}(\mathcal{A}_{\theta})$ and $B(R)$ be the ball of radius $R$ around the origin. One has the following asymptotic expansion
$$\int_{B(R)}\sigma(\xi)d\xi\sim_{R\rightarrow\infty}\sum_{j=0,\alpha-j+2\neq0}^{\infty}\alpha_j(\sigma)R^{\alpha-j+2}+\beta(\sigma)\log R+c(\sigma),$$
where $\beta(\sigma)=\int_{|\xi|=1}\sigma_{-2}(\xi)d\xi$
and the constant term in the expansion, $c(\sigma)$, is given by
\begin{equation}\label{cutoffinexpansion}
\int_{\mathbb{R}^n}\sigma^N+\sum_{j=0}^{N}\int_{B(1)}\chi(\xi)\sigma_{\alpha-j}(\xi)d\xi-\sum_{j=0,\alpha-j+2\neq0}^N \frac{1}{\alpha-j+2}\int_{|\xi|=1}\sigma_{\alpha-j}(\omega)d\omega.
\end{equation}
Here we have used the notation of (\ref{sigma}).
\end{proposition}
\begin{proof}
First, we write $\sigma (\xi)=\sum_{j=0}^{N}\chi(\xi)\sigma_{\alpha-j}(\xi)+\sigma^N(\xi)$
with large enough $N$, so that $\sigma^N$ is integrable. Then we have,
\begin{equation}\label{constant1}
\int_{B(R)}\sigma(\xi)d\xi=\sum_{j=0}^{N}\int_{B(R)}\chi(\xi)\sigma_{\alpha-j}(\xi)d\xi+\int_{B(R)}\sigma^N(\xi)d\xi.
\end{equation}
For $N>\alpha+1$, $\sigma^N\in\mathcal{L}^1(\mathbb{R}^2,\mathcal{A}_{\theta})$, so
\begin{equation}
\int_{B(R)}\sigma^N(\xi)d\xi\to\int_{\mathbb{R}^2}\sigma^N(\xi)d\xi,\quad R\to\infty.\nonumber
\end{equation}
Now for each $j\leq N$ we have
$$\int_{B(R)}\chi(\xi)\sigma_{\alpha-j}(\xi)d\xi=\int_{B(1)}\chi(\xi)\sigma_{\alpha-j}(\xi)d\xi+\int_{B(R)\backslash B(1)}\chi(\xi)\sigma_{\alpha-j}(\xi)d\xi.$$
Obviously $\int_{B(1)}\chi(\xi)\sigma_{\alpha-j}(\xi)d\xi<\infty$ and
by using polar coordinates $\xi=r\omega $, and homogeneity of $\sigma_{\alpha-j}$, we have
\begin{equation}\label{constant2}
\int_{B(R)\backslash B(1)}\chi(\xi)\sigma_{\alpha-j}(\xi)d\xi=\int_{1} ^R r^{\alpha-j+2-1}dr\int_{|\xi|=1}\sigma_{\alpha-j}(\xi)d\xi.
\end{equation}
Note that the cut-off function is equal to one on the set $\mathbb{R}^2 \backslash B(1)$.
For the term with $\alpha-j=-2$ one has
\begin{equation}
\int_{B(R)\backslash B(1)}\chi(\xi)\sigma_{\alpha-j}(\xi)d\xi= \log R \int_{|\xi|=1}\sigma_{\alpha-j}(\xi)d\xi.\nonumber
\end{equation}
The terms with $\alpha-j\neq -2$ will give us the following:
\begin{eqnarray}\label{constant3}
\int_{ B(R)\backslash B(1)}\chi(\xi)\sigma_{\alpha-j}(\xi)d\xi&=& \\
\frac{R^{\alpha-j+2}}{m-j+2}\int_{|\xi|=1}\sigma_{\alpha-j}(\xi)d\xi&-&\frac{1}{\alpha-j+2}\int_{|\xi|=1}\sigma_{\alpha-j}(\xi)d\xi.\nonumber
\end{eqnarray}
Adding all the constant terms in \eqref{constant1}-\eqref{constant3}, we get the constant term given in \eqref{cutoffinexpansion}.
\end{proof}
\begin{definition}
The cut-off integral of a symbol $\sigma\in\mathcal{S}_{cl}^{\alpha}(\mathcal{A}_{\theta})$ is defined to be the constant term in the above asymptotic expansion, and we denote it by $\int\hspace{-0.35cm}{-} \, \sigma(\xi)d\xi$.
\end{definition}
\begin{remark} Two remarks are in order here. First note that the cut-off integral of a symbol is independent of the choice of $N$. Second, it is also independent of the choice of the cut-off function
$\chi$.
\end{remark}
We now give the definition of the canonical trace for classical pseudodifferential operators on $\mathcal{A}_{\theta}$.
\begin{definition}
The canonical trace of a classical pseudodifferential operator $P\in\Psi^{\alpha}_{cl}(\mathcal{A}_{\theta})$ of non-integral order $\alpha$ is defined as
\begin{equation*}
{\rm TR}(P):=\varphi_0\left(\int\hspace{-0.35cm}{-} \, \sigma_P(\xi)d\xi\right).
\end{equation*}
\end{definition}
In the following, we establish the relation between the TR-functional and the usual trace on trace-class pseudodifferential operators.
Note that any pseudodifferential operator $P$ of order less that $-2$, is a trace-class operator on $\mathcal{H}_0$ and its trace is given by
$${\rm Tr}(P)=\varphi_0\left(\int_{\mathbb{R}^2}\sigma_{P}(\xi)d\xi\right).$$
On the other hand, for such operator the symbol is integrable and we have
\begin{equation}\label{TRTr}
\int\hspace{-0.35cm}{-} \,\sigma_P(\xi)=\int_{\mathbb{R}^2}\sigma_P(\xi)d\xi.
\end{equation}
Therefore, the ${\rm TR}$-functional and operator trace coincide on classical pseudodifferential operators of order less than $-2$.
Next, we show that the ${\rm TR}$-functional is in fact an analytic continuation of the operator trace and using this fact we can prove that it is actually a trace.
\begin{definition}
A family of symbols $\sigma(z)\in\mathcal{S}_{cl}^{\alpha(z)}(\mathcal{A}_{\theta})$, parametrized by $z\in W\subset \mathbb{C}$, is called a holomorphic family if
\begin{itemize}
\item[i)] The map $z\mapsto \alpha(z)$ is holomorphic.
\item[ii)] The map $z\mapsto \sigma(z)\in \mathcal{S}_{cl}^{\alpha(z)}(\mathcal{A}_{\theta})$ is a holomorphic map from $W$ to the Fr\'{e}chet space $\mathcal{S}_{cl}(\mathcal{A}^n_{\theta}).$
\item[iii)] The map $z\mapsto \sigma(z)_{\alpha(z)-j}$ is holomorphic for any $j$, where
\begin{equation}\label{sigma2}
\sigma(z)(\xi)\sim\sum_j\chi(\xi)\sigma(z)_{\alpha(z)-j}(\xi)\in\mathcal{S}_{cl}^{\alpha(z)}(\mathcal{A}_{\theta}).
\end{equation}
\item[iv)] The bounds of the asymptotic expansion of $\sigma(z)$ are locally uniform with respect to $z$, i.e, for any $N\geq1$ and compact subset $K\subset W$, there exists a constant $C_{N,K,\alpha,\beta}$ such that for all multi-indices $\alpha,\beta$ we have
$$\left|\left|\delta^{\alpha}\partial^{\beta}\left(\sigma(z)-\sum_{j<N}\chi\sigma(z)_{\alpha(z)-j}\right)(\xi)\right|\right|<C_{N,K,\alpha,\beta}|\xi|^{\Re(\alpha(z))-N-|\beta|}.$$
\end{itemize}
A family $\left\{P_z\right\}\in\Psi_{cl}(\mathcal{A}_{\theta})$ is called holomorphic if $P_z=P_{\sigma(z)}$ for a holomorphic family of symbols $\left\{\sigma(z)\right\}$.
\end{definition}
The following Proposition is an analogue of a result of Kontsevich and Vishik\cite{Kontsevich-Vishik1995}, for pseudodifferential calculus on noncommutative tori.
\begin{proposition}\label{laurentofholo}
Given a holomorphic family $\sigma(z)\in\mathcal{S}_{cl}^{\alpha(z)}(\mathcal{A}_{\theta})$, $z\in W \subset\mathbb{C}$, the map
$$z\mapsto \int\hspace{-0.35cm}{-} \,\sigma(z)(\xi)d\xi,$$
is meromorphic with at most simple poles located in
$$P=\left\{z_0\in W;~\alpha(z_0)\in\mathbb{Z}\cap[-2,+\infty]\right\}.$$
The residues at poles are given by
$${\rm Res}_{z=z_0} \int\hspace{-0.35cm}{-} \,\sigma(z)(\xi)d\xi=-\frac{1}{\alpha'(z_0)}\int_{|\xi|=1}\sigma(z_0)_{-2}d\xi.$$
\end{proposition}
\begin{proof}
By definition, one can write $\sigma(z)=\sum_{j=0}^N\chi(\xi)\sigma(z)_{\alpha(z)-i}(\xi)+\sigma(z)^N(\xi)$, and by Proposition \ref{expansioninR} we have,
\begin{align*}
\int\hspace{-0.35cm}{-} \,\sigma(z)(\xi)d\xi
&=\int_{\mathbb{R}^2}\sigma(z)^N(\xi)d\xi
+\sum_{j=0}^{N}\int_{B(1)}\chi(\xi)\sigma(z)_{\alpha(z)-j}(\xi)\\
&-\sum_{j=0}^{N}\frac{1}{\alpha(z)+2-j}\int_{|\xi|=1}\sigma(z)_{\alpha(z)-j}(\xi)d\xi.
\end{align*}
Now suppose $\alpha(z_0)+2-j_0=0$. By holomorphicity of $\sigma(z)$, we have $\alpha(z)-\alpha(z_0)=\alpha'(z_0)(z-z_0)+o(z-z_0)$. Hence
$${\rm Res}_{z=z_0} \int\hspace{-0.35cm}{-} \,\sigma(z)=\frac{-1}{\alpha'(z_0)}\int_{|\xi|=1}\sigma(z_0)_{-2}(\xi)d\xi.$$
\end{proof}
\begin{corollary}\label{analyticcontin}
The functional ${\rm TR}$ is the analytic continuation of the ordinary trace on trace-class pseudodifferential operators.
\end{corollary}
\begin{proof}
First observe that, by the above result, for a non-integer order holomorphic family of symbols $\sigma(z)$, the map $z\mapsto\int\hspace{-0.35cm}{-} \,\sigma(z)(\xi)d\xi$ is holomorphic. Hence, the map $\sigma\mapsto\int\hspace{-0.35cm}{-} \,\sigma(\xi)d\xi$ is the unique analytic continuation of the map $\sigma\mapsto\int_{\mathbb{R}^2}\sigma(\xi)d\xi$ from $\mathcal{S}_{cl}^{<-2}(\mathcal{A}_{\theta})$ to $\mathcal{S}^{\notin\mathbb{Z}}_{cl}(\mathcal{A}_{\theta})$. By \eqref{TRTr} we have the result.
\end{proof}
Let $Q\in\Psi^q_{cl}(\mathcal{A}_{\theta})$ be a positive elliptic pseudodifferential operator of order $q>0$.
The complex power of such an operator, $Q_\phi^z$, for $\Re(z)<0$ can be defined by the following Cauchy integral formula.
\begin{equation}\label{Qz}
Q_\phi^z=\frac{i}{2\pi}\int_{C_\phi}\lambda_\phi^z (Q-\lambda)^{-1} d\lambda.
\end{equation}
Here $\lambda^z_\phi$ is the complex power with branch cut $L_\phi =\{re^{i\phi}, r\geq 0\}$ and $C_\phi$ is a contour around the spectrum of $Q$ such that
$$C_\phi\cap {\rm spec}(Q)\backslash\{0\}=\emptyset,\qquad L_\phi\cap C_\phi=\emptyset,$$
$$ C_\phi\cap\{ {\rm spec}(\sigma(Q)^L(\xi)),\,\xi\neq 0\}=\emptyset.$$
In general an operator for which one can find a ray $L_\phi$ with the above property, is called an admissible operator with the spectral cut $L_\phi$. Positive elliptic operators are admissible and we take the ray $L_\pi$ as the spectral cut, and in this case we drop the index $\phi$ and write $Q^z$.
To extend \eqref{Qz} to $\Re(z)>0$ we choose a positive integer such that $\Re (z) <k$ and define
$$Q_\phi^z:=Q^kQ_\phi^{z-k}.$$
It can be proved that this definition is independent of the choice of $k$.
\begin{corollary}
Let $A\in\Psi ^{\alpha}_{cl}(\mathcal{A}_{\theta})$ be of order $\alpha\in\mathbb{Z}$ and let $Q$ be a positive elliptic classical pseudodifferential operator of positive order $q$. We have
$${\rm Res}_{z=0}{\rm TR}(AQ^{-z})= \frac{1}{q}{\rm Res}(A).$$
\end{corollary}
\begin{proof}
For the holomorphic family $\sigma(z)=\sigma(AQ^{-z})$, $z=0$ is a pole for the map $z\mapsto\int\hspace{-0.35cm}{-} \,\sigma(z)(\xi)d\xi$ whose residue is given by
$${\rm Res}_{z=0}\left(z\mapsto \int\hspace{-0.35cm}{-} \,\sigma(z)(\xi)d\xi\right)=-\frac{1}{\alpha'(0)}\int_{|\xi|=1}\sigma_{-2}(0)d\xi=-\frac{1}{\alpha'(0)}{\rm res}(A).$$
Taking trace on both sides gives the result.
\end{proof}
Now we can prove the trace property of ${\rm TR}$-functional.
\begin{proposition}
We have
${\rm TR}(AB)={\rm TR}(BA)$ for any $A,B\in\Psi_{cl}(\mathcal{A}_{\theta})$, provided that $ord(A)+ord(B)\notin\mathbb{Z}$.
\end{proposition}
\begin{proof}
Consider the families $A_z=AQ^z$ and $B_z=BQ^z$ where $Q$ is an injective positive elliptic classical operator of order $q>0$.
For $\Re(z)<<0$, the two families are trace class and ${\rm Tr}(A_zB_z)={\rm Tr}(B_zA_z)$. By the uniqueness of the analytic continuation, we have
$${\rm TR}(A_zB_z)={\rm TR}(B_zA_z),$$ for those $z$ for which $2qz+{\rm ord}(A)+{\rm ord}(B)\not \in \mathbb{Z}.$
At $z=0$, we obtain ${\rm Tr}(AB)={\rm TR}(BA).$
\end{proof}
\subsection{Log-polyhomogeneous symbols}
Proposition \ref{laurentofholo} can be extended and one can explicitly write down the Laurent expansion of the cut-off integral around each of the poles. The terms of the Laurent expansion involve residue densities of $z$-derivatives of the holomorphic family. In general, $z$-derivatives of a classical holomorphic family of symbols is not classical anymore and therefore we introduce log-polyhomogeneous symbols which include the $z$-derivatives of the symbols of the holomorphic family $\sigma(AQ^{-z})$.
\begin{definition}
A symbol $\sigma$ is called a log-polyhomogeneous symbol if it has the following form
\begin{equation}\label{logpolysym}
\sigma(\xi) \sim \sum_{j\geq 0}\sum_{l=0}^\infty\sigma_{\alpha-j,l}(\xi)\log^l|\xi|\quad |\xi|>0,
\end{equation}
with $ \sigma_{\alpha-j,l}$ positively homogeneous in $\xi$ of degree $\alpha - j$.
\end{definition}
An important example of an operator with such a symbol is $\log Q$ where $Q\in\Psi^q_{cl}(\mathcal{A}_{\theta})$ is a positive elliptic pseudodifferential operator of order $q>0$.
The logarithm of $Q$ can be defined by
$$\log Q
=Q\left.\frac{d}{dz}\right|_{z=0}Q^{z-1}
=Q\left.\frac{d}{dz}\right|_{z=0}\frac{i}{2\pi}\int_C\lambda^{z-1}(Q-\lambda)^{-1}d\lambda.$$
It is a pseudodifferential operator with symbol
\begin{equation}\label{symboflog}\sigma(\log Q)
\sim \sigma(Q)\star \sigma\Big(\left.\frac{d}{dz}\right|_{z=0} Q^{z-1}\Big),
\end{equation}
where $\star$ denotes the products of the pseudodifferential symbols.
Using symbol calculus and homogeneity properties, we can show that \eqref{symboflog} is a log-homogeneous symbol of the form
$$\sigma(\log Q)(\xi)=2\log|\xi| I+ \sigma_{cl}(\log Q)(\xi),$$
where $\sigma_{cl}(\log Q)$ is a classical symbol of order zero. This symbol can be computed using the homogeneous parts of the classical symbol $\sigma(Q^z)=\sum_{j=0}^\infty b(z)_{2z-j}(\xi)$ and it is given by the following formula (see e.g. \cite{Paycha2012}).
\begin{align}\label{cllog} \sigma_{cl}(\log Q)(\xi) &= \\
\sum_{k=0}^\infty \sum_{i+j+|\alpha|=k}\frac{1}{\alpha !}\partial^\alpha \sigma_{2-i}(Q)\delta^\alpha
& \left[|\xi|^{-2-j}\left.\frac{d}{dz}\right|_{z=0} b(z-1)_{2z-2-j}\left(\xi/|\xi|\right)\right]. \nonumber
\end{align}
The Wodzicki residue can also be extended to this class of pseudodifferential operators \cite{Lesch1999}. For an operator $A$ with log-polyhomogeneous symbol as \eqref{logpolysym} it can be defined by
$${\rm res}(A)=\int_{|\xi|=1}\sigma_{-2,0}(\xi)d\xi.$$
By adapting the proof of Theorem 1.13 in \cite{Paycha-Scott2007} to the noncommutative case, we have the following theorem which is written only for the families of the form $\sigma(AQ^{-z})$ which we will use in Section \ref{sec:computation}.
\begin{proposition}\label{Laurentat0}
Let $A\in\Psi_{cl}^\alpha(\mathcal{A}_{\theta})$ and $Q$ be a positive , in general an admissible, elliptic pseudodifferential operator of positive order $q$. If $\alpha\in P$ then $0$ is a possible simple pole for the function $z\mapsto {\rm TR}(AQ^{-z})$ with the following Laurent expansion around zero.
\begin{align*}
{\rm TR}(AQ^{-z})&=\frac{1}{q}{\rm Res}(A)\frac{1}{z}\\
&+\varphi_0\left(\int\hspace{-0.35cm}{-} \,\sigma(A)- \frac{1}{q}{\rm res}(A\log Q)\right)-{\rm Tr}(A\Pi_Q)\\
& +\sum_{k=1}^K(-1)^k\frac{(z)^k}{k!} \\
&\times \left(\varphi_0\left( \int\hspace{-0.35cm}{-} \,\sigma(A(\log Q)^k)d\xi-\frac{1}{q(k+1)}{\rm res}(A(\log Q)^{k+1})\right)-{\rm Tr}(A\log^k Q\Pi_Q)\right)\\
&+o(z^{K}).
\end{align*}
Where $\Pi_Q$ is the projection on the kernel of $Q$.
\end{proposition}
For operators $A$ and $Q$ as in the previous Proposition,
we define a zeta function by
\begin{equation}\label{zetafunction}
\zeta(A,Q,z)={\rm TR}(AQ^{-z}).
\end{equation}
By Corollary \ref{analyticcontin}, it is obvious that $\zeta(A,Q,z)$ is the analytic continuation of the zeta function ${\rm Tr}(AQ^{-z})$ defined by the regular trace only for $\Re(z)>>0$.
\begin{remark}\label{holomorphicityatzero}
If $A$ is a differential operator, the zeta function \eqref{zetafunction} is holomorphic at $z=0$ with the value equal to
$$\varphi_0\left(\int\hspace{-0.35cm}{-} \,\sigma(A)- \frac{1}{q}{\rm res}(A\log Q)\right)-{\rm Tr}(A\Pi_Q).$$
\end{remark}
\section{ Cauchy-Riemann operators on noncommutative tori}
In \cite{Quillen1985}, Quillen studies the geometry of the determinant line bundle on the space of all Cauchy-Riemann operators on a smooth vector bundle on a closed Riemann surface. To investigate the same notion on noncommutative tori, we first briefly recall some basic facts in the classical case on how Cauchy-Riemann operators are related to Dirac operators and spectral triples. Then by analogy we define our Cauchy-Riemann operator on $\mathcal{A}_{\theta}$, and consider the spectral triples defined by them.
Let $M$ be a compact complex manifold and $V$ be a smooth complex vector bundle on $M$.
Let $\Omega^{p,q}(M,V)$ denote the space of $(p,q)$ forms on $M$ with coefficients in $V$.
A $\bar\partial$-flat connection
on $V$ is a $\mathbb{C}$-linear map $D: \Omega^{0,0}(M,V)\to \Omega^{0,1}(M,V),$ such that for any $f\in C^\infty(M)$ and $u\in \Omega^{0,0}(M,V)$,
\begin{equation}
D(fu)=(\bar\partial f)\otimes u+ f Du,
\end{equation}
and $D^2=0$. Here to define $D^2$, note that
any $\bar\partial$-connection as above has a unique extension
to an operator $D:\Omega^{p,q}(M,V)\to\Omega^{p,q+1}(M,V)$, defined by
$$D(\alpha\otimes \beta)=\bar{\partial}\alpha \otimes u+(-1)^{p+q}\alpha\wedge Du, \quad \alpha\in \Omega^{p,q}(M),\,\, u\in C^\infty(V).$$
We refer to $\bar\partial$-flat connections as Cauchy-Riemann operators.
A holomorphic vector bundle $V$ has a canonical Cauchy-Riemann operator $\bar{\partial}_V:\Omega^{0}(M,V)\to \Omega^{0,1}(M,V)$, whose extension to $\Omega^{0,*}(M,V)$ forms the Dolbeault complex of $M$ with coefficients in $V$. In fact there is a one-one correspondence between Cauchy-Riemann operators on $V$ up to (gauge) equivalence, and holomorphic structures on $V$. We denote by $\mathcal{A}$ the set of all Cauchy-Riemann operators on $V$.
Any holomorphic structure on a Hermitian vector bundle $V$ determines a unique Hermitian connection, called the Chern connection, whose projection on $(0,1)$-forms, $\nabla^{0,1}(M, V)$, is the Cauchy-Riemann operator coming from the holomorphic structure.
Now, if $M$ is a K\"ahler manifold, the tensor product of the Levi-Civita connection for $M$ with the Chern connection on $V$ defines a Clifford connection on the Clifford module $(\Lambda^{0,+}\oplus \Lambda^{0,-})\otimes V$ and the operator $D_0=\sqrt{2}(\bar\partial_V+\bar\partial_V^*)$ is the associated Dirac operator (see e.g. \cite{Gilkey1984}).
Any other Dirac operator on the Clifford module $(\Lambda^{0,+}\oplus \Lambda^{0,-})\otimes V$ is of the form $D_0+A$ where $A$ is the connection one form of a Hermitian connection.
This connection need not be a Chern connection. However, on a Riemann surface (with a Riemannian metric compatible with its complex structure) any Hermitian connection on a smooth Hermitian vector bundle is the Chern connection of a holomorphic structure on $V$. Therefore, the positive part of any Dirac operator on $(\Lambda^{0,0}\oplus \Lambda^{0,1})\otimes V$ is a Cauchy-Riemann operator, and this gives a one to one correspondence between all Dirac operators and the set of all Cauchy-Riemann operators.
Next we define the analogue of Cauchy-Riemann operators for the noncommutative torus.
First, following \cite{Connes-Tred2009, Masoud-Farzad2012}, we fix a complex structure on $\mathcal{A}_{\theta}$ by a complex number $\tau$ in the upper half plane and construct the spectral triple
\begin{equation}\label{firstspec}
(\mathcal{A}_{\theta}, \mathcal{H}_{0}\oplus\mathcal{H}^{0,1},D_0=\left(\begin{array}{ll} 0 & \bar{\partial}^*\\ \bar{\partial}&0\end{array}\right)),
\end{equation}
where $\bar{\partial}:\mathcal{A}_{\theta}\to \mathcal{A}_{\theta}$ is given by $\bar\partial=\delta_1+\tau\delta_2$. The Hilbert space $\mathcal{H}_0$ is obtained by GNS construction from $\mathcal{A}_{\theta}$ using the trace $\varphi_0$ and $\bar{\partial}^*$ is the adjoint of the operator $\bar{\partial}$.
As in the classical case, we define our Cauchy-Riemann operators on $\mathcal{A}_{\theta}$ as the positive part of twisted Dirac operators. All such operators define spectral triples of the form
$$(\mathcal{A}_{\theta}, \mathcal{H}_{0}\oplus\mathcal{H}^{0,1},D_A=\left(\begin{array}{ll} 0 & \bar{\partial}^*+\alpha^*\\ \bar{\partial}+\alpha&0\end{array}\right)),$$
where $\alpha\in \mathcal{A}_{\theta}$ is the positive part of a selfadjoint element
$$A=\left(\begin{array}{ll} 0 & \alpha^*\\ \alpha&0\end{array}\right)\in\Omega^1_{D_0}(\mathcal{A}_{\theta}).$$
We recall that $\Omega^1_{D_0}(\mathcal{A}_{\theta})$ is the space of quantized one forms
consisting of the elements $\sum a_i[D_0, b_i]$ where $a_i,b_i\in\mathcal{A}_{\theta}$ \cite{Connesbook1994}.
Note that the in this case, the space $\mathcal{A}$ of Cauchy-Riemann operators is the space of $(0,1)$-forms on $\mathcal{A}_{\theta}$.
We should mention that in the noncommutative case, in the work of Chakraborty and Mathai \cite{Chakraborty-Mathai2009} a general family of spectral triples is considered and, under suitable regularity conditions, a determinant line bundle is defined for such families.
The curvature of the determinant line bundle however is not computed and that is the main object of study in the present paper, as well as in \cite{Quillen1985}.
\section{The curvature of the determinant line bundle for $\mathcal{A}_{\theta}$}\label{sec:computation}
For any $\alpha\in\mathcal{A}$, the Cauchy-Riemann operator $$\bar{\partial}_{\alpha}=\bar{\partial}+\alpha:\mathcal{H}_0\to\mathcal{H}^{0,1}$$ is a Fredholm operator.
We pull back the determinant line bundle DET on the space of Fredholm operators ${\rm Fred}(\mathcal{H}_0,\mathcal{H}^{0,1}),$ to get a line bundle $\mathcal{L}$ on $\mathcal{A}$. Following Quillen \cite{Quillen1985}, we define a Hermitian metric on $\mathcal{L}$ and compute its curvature in this section. Let us define a metric on the fiber
$$\mathcal{L}_{\alpha} = \Lambda^{max} ({\rm ker} \,\bar{\partial}_{\alpha})^*\otimes \Lambda^{max}({\rm ker} \,\bar{\partial}_{\alpha}^*).$$
as the product of the induced metrics on $\Lambda^{max} ({\rm ker} \,\bar{\partial}_{\alpha})^* $, $\Lambda^{max}({\rm ker} \,\bar{\partial}_{\alpha}^*)$,
with the zeta regularized determinant $e^{-\zeta'_{\Delta_{\alpha}}(0)}$.
Here we define the Laplacian as $\Delta_{\alpha}=\bar{\partial}_{\alpha}^*\bar{\partial}_{\alpha}:\mathcal{H}_0\to \mathcal{H}_0$, and its zeta function by
\begin{align*}
&\zeta(z)={\rm TR}(\Delta_{\alpha}^{-z}).
\end{align*}
It is a meromorphic function and by Remark \ref{holomorphicityatzero} it is regular at $z=0$ .
Similar proof as in \cite{Quillen1985} shows that this defines a smooth Hermitian metric on $\mathcal{L}$.
On the open set of invertible operators each fiber of $\mathcal{L}$ is canonically isomorphic to $\mathbb{C}$ and the nonzero holomorphic section $\sigma=1$ gives a trivialization. Also, according to the definition of the Hermitian metric, the norm of this section is given by
\begin{equation}
\|\sigma\|^2=e^{-\zeta'_{\Delta_\alpha}(0)}.
\end{equation}
\subsection{Variations of LogDet and curvature form}
We begin by explaining the motivation behind the computations of Quillen in \cite{Quillen1985}. Recall that a holomorphic line bundle equipped with a Hermitian inner product has a canonical connection compatible with the two structures. This is also known as the Chern connection. The curvature form of this connection is computed by $\bar{\partial} \partial \log\|\sigma\|^2,$
where $\sigma$ is any non-zero local holomorphic section.
In our case we will proceed by analogy and compute the second variation
$ \bar{\partial} \partial \log \|\sigma\|^2$
on the open set of invertible index zero Cauchy-Riemann operators. Let us consider a holomorphic family of invertible index zero Cauchy-Riemann operators $D_w=\bar{\partial}+\alpha_w$, where $\alpha_w$ depends holomorphically on the complex variable $w$ and compute
$$ \delta_{\bar{w}} \delta_w \zeta'_\Delta(0).$$
One has the following first variational formula,
\begin{equation*}
\delta_w\zeta(z)=\delta_w{\rm TR}(\Delta^{-z})
={\rm TR}(\delta_w\Delta^{-z})
=-z{\rm TR}(\delta_w\Delta\Delta^{-z-1}),
\end{equation*}
where in the second equality we were able to change the order of $\delta_w$ and ${\rm TR}$ because of the uniformity condition in the definition of holomorphic families (cf. \cite{Paycha-Rosenberg2006}).
Note that, although ${\rm TR}(\Delta^{-z})$ is regular at $z=0$,
${\rm TR}(\delta_w\Delta\Delta^{-z-1})$ might have a pole at $z=0$ since $\delta_w\Delta\Delta^{-z-1}|_{z=0}=\delta_w\Delta\Delta^{-1}$ is not a differential operator any more and may have non-zero residue.
Around $z=0$ one has the following Laurent expansion:
$$-z{\rm TR}(\delta_w\Delta\Delta^{-z-1})=-z(\frac{a_{-1}}{z}+a_0+a_1 z+\cdots).$$
Hence,
$$\left.\delta_w\zeta(z)\right|_{z=0}=-a_{-1},\qquad \left.\frac{d}{d z}\delta_w\zeta(z)\right|_{z=0}=-a_{0}.$$
Using Proposition \ref{Laurentat0} we have
$$\delta_w\zeta'(0)=\left.\frac{d}{d z}\delta_w\zeta(z)\right|_{z=0}=-\varphi_0 \left(\int\hspace{-0.35cm}{-} \,\sigma(\delta_w\Delta\Delta^{-1})-\frac{1}{2}{\rm res}_x(\delta_w\Delta\Delta^{-1}\log\Delta)\right).$$
To compute the right hand side of the above equality,
we need to note that since $D_w$ depends holomorphically on $w$, $\delta_wD^*=0$ and hence
$$\delta_w\Delta=\delta_w D^*D+ D^*\delta_w D=D^*\delta_w D.$$
Since $\delta_wD$ is a zero order differential operator, we have
\begin{align*}
\delta_w\zeta'(0)&=-\varphi_0\left(\int\hspace{-0.35cm}{-} \,\sigma(D^*\delta_w D\Delta^{-1})-\frac{1}{2}{\rm res}(D^*\delta_w D\Delta^{-1}\log\Delta)\right)\\
&=-\varphi_0\left(\int\hspace{-0.35cm}{-} \,\sigma(\delta_w D\Delta^{-1}D^*)-\frac{1}{2}{\rm res}(\delta_w D\log\Delta\,\Delta^{-1}D^*)\right)\\
&=-\varphi_0\left(\delta_w D\left(\int\hspace{-0.35cm}{-} \,\sigma(D^{-1})-\frac{1}{2}{\rm res}(\log\Delta\,D^{-1})\right)\right)\\
&=-\varphi_0\left(\delta_w D\, J\right),
\end{align*}
where
$$J=\int\hspace{-0.35cm}{-} \,\sigma(D^{-1})-\frac{1}{2}{\rm res}(\log(\Delta)D^{-1}).$$
The reader can compare this to the term $J$ in Quillen's computations \cite{Quillen1985}.
Now we compute the second variation
$\delta_{\bar w}\delta_{w}\zeta'(0).$
Since $D_w$ is holomorphic we have
\begin{align*}
\delta_{\bar w}\delta_{w}\zeta'(0)&=-\varphi_0\left(\delta_w D \delta_{\bar w}J\right).
\end{align*}
Next we compute the variation $\delta_{\bar{w}}J$. Note that since $D_w$ is invertible, $D_w^{-1}$ is also holomorphic and hence $\delta_{\bar{w}}\int\hspace{-0.35cm}{-} \,\sigma(D^{-1})=0$. Therefore
\begin{align*}
\delta_{\bar w}J&=\delta_{\bar{w}}\left( \int\hspace{-0.35cm}{-} \,\sigma(D^{-1})-\frac{1}{2}{\rm res}(\log\Delta\, D^{-1})\right)
=-\frac{1}{2}\delta_{\bar{w}}{\rm res}(\log\Delta\, D^{-1}).
\end{align*}
Thus, we have shown that
\begin{lemma}\label{secondvarnotsim}
For the holomorphic family of Cauchy-Riemann operators $D_w$, the second variation of $\zeta'(0)$ reads:
$$\delta_{\bar w}\delta_{w}\zeta'(0)=\frac{1}{2}\varphi_0\left(\delta_w D\delta_{\bar{w}}{\rm res}(\log\Delta\,D^{-1})\right).$$
\qed
\end{lemma}
Our next goal is to compute $\delta_{\bar{w}}{{\rm res}}(\log\Delta\,D^{-1})$.
This combined with the above lemma shows that the curvature form of the determinant line bundle equals the K\"ahler form on the space of connections.
\begin{lemma}\label{secondvarsim}
With above definitions and notations, we have
\begin{align*}
\sigma_{-2,0}(\log\Delta\, D^{-1})&=\frac{(\alpha+\alpha^*)\xi_1+(\bar\tau\alpha+\tau\alpha^*)\xi_2}{(\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2)(\xi_1+\tau\xi_2) }\\ \\
&-\log \left( \frac{\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2}{|\xi|^2}\right) \frac{\alpha}{\xi_1+\tau\xi_2},
\end{align*}
and
$$\delta_{\bar{w}}{\rm res}(\log(\Delta) D^{-1})=\frac{1}{2\pi\Im(\tau)}(\delta_{{w}}D)^*.$$
\end{lemma}
\begin{proof}
By writing down the homogeneous terms in the expansion of $\sigma_{\bullet,0}(\log\Delta)$ and $\sigma(D^{-1})$ and using the product formula of the symbols we see that
$$\sigma_{-2,0}(\log\Delta D^{-1})\sim\sigma_{-1,0}(\log\Delta)\sigma_{-1}(D^{-1})+\sigma_{0,0}(\log\Delta)\sigma_{-2}(D^{-1}).$$
Starting with the symbol of $\Delta$, we have
$$\sigma(\Delta)=\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2+(\alpha+\alpha^*)\xi_1+(\bar\tau\alpha+\tau\alpha^*)\xi_2+\bar{\partial}^*(\alpha).$$
Then, the homogeneous parts of $\sigma((\lambda-\Delta)^{-1})=\sum_{j} b_{-2-j}$ is given by the following recursive formula
\begin{align*}
b_{-2}&=(\lambda-\sigma_{2}(\Delta))^{-1},\\
b_{-2-j}&=-b_{-2}\sum_{k+l+|\gamma|=j, \, l<j}\partial^\gamma\sigma_{2-k}(\Delta)\delta^\gamma b_{-2-l}/\gamma!,
\end{align*}
which gives us
$$b_{-2}=\frac{1}{\lambda-(\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2)},$$
and
$$b_{-3}=\frac{1}{(\lambda-(\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2))^2}\left((\alpha+\alpha^*)\xi_1+(\bar\tau\alpha+\tau\alpha^*)\xi_2\right).$$
Also, $\Delta^z$ is a classical operator defined by
$$\Delta^z=\frac{1}{2\pi i}\int_C \lambda^z(\lambda-\Delta)^{-1}d\lambda,$$
with the homogeneous parts of the symbol given by
$$b(z)_{2z-j}:=\sigma_{2z-j}(\Delta^z)=\frac{1}{2\pi i}\int_C \lambda^zb_{-2-j}d\lambda.$$
Hence we have
\begin{align*}
b{(z)}_{2z}&=\frac{1}{2\pi i}\int_C \lambda^z\frac{1}{\lambda-(\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2)}d\lambda\\ \\
&=(\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2)^{z}
\end{align*}
\begin{align*}
b{(z)}_{2z-1}&=\frac{1}{2\pi i}\int_C \lambda^z\frac{\left((\alpha+\alpha^*)\xi_1+(\bar\tau\alpha+\tau\alpha^*)\xi_2\right)}{(\lambda-(\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2))^2}d\lambda \\ \\
&=z(\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2))^{z-1} \left((\alpha+\alpha^*)\xi_1+(\bar\tau\alpha+\tau\alpha^*)\xi_2\right).
\end{align*}
Using \eqref{cllog} and what we have computed up to here, it is clear that
\begin{align*}
\sigma_{0,0}(\log\Delta)(\xi)&= \sigma_{2}(\Delta)|\xi|^{-2}\left.\frac{d}{d z}\right|_{z=0} b{(z-1)}_{2z-2}\left(\xi/|\xi|\right)\\ \\
&=\sigma_{2}(\Delta)|\xi|^{-2}\left.\frac{d}{d z}\right|_{z=0}( (\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2)/|\xi|^2)^{z-1}\\ \\
&=\log ( (\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2)/|\xi|^2).
\end{align*}
Note that the above term is homogeneous of order zero in $\xi$.
\begin{align*}
& \sigma_{-1,0}(\log\Delta)(\xi)\\ \\
&= \sum_{i+j+|\alpha|=1}\frac{1}{\alpha !}\partial^\alpha \sigma_{2-i}(\Delta)\delta^\alpha|\xi|^{-2-j}\left.\frac{d}{d z}\right|_{z=0} b{(z-1)}_{2z-2-j}\left(\xi/|\xi| \right)\\ \\
&= \sigma_{2}(\Delta)|\xi|^{-3}\left.\frac{d}{d z}\right|_{z=0} b{(z-1)}_{2z-3}\left(\xi/|\xi|\right)\\
&+ \sigma_{1}(\Delta)|\xi|^{-2}\left.\frac{d}{d z}\right|_{z=0} b{(z-1)}_{2z-2}\left(\xi/|\xi|\right)\\ \\
&=\frac{1-\log (\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2)/|\xi|^2)}{ (\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2)} \left[(\alpha+\alpha^*)\xi_1+(\bar\tau\alpha+\tau\alpha^*)\xi_2\right]\\
& +\frac{ \log (\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2)/|\xi|^2)}{\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2} \left[(\alpha+\alpha^*)\xi_1+(\bar\tau\alpha+\tau\alpha^*)\xi_2\right]
\\ \\
&= (\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2)^{-1} \left[(\alpha+\alpha^*)\xi_1+(\bar\tau\alpha+\tau\alpha^*)\xi_2\right].
\end{align*}
Next we compute the symbol of $D^{-1}$. The symbol of $D$ reads
$$\sigma(D)={\xi_1+\tau \xi_2}+\alpha.$$
We need to compute the homogeneous parts of order -1 and -2 of $D^{-1}$. By using recursive formula for the symbol of the inverse we get:
\begin{align*}
\sigma_{-1}(D^{-1})&=\sigma_1(D)^{-1}=(\xi_1+\tau\xi_2)^{-1}\\
\sigma_{-2}(D^{-1})&=- \sigma_{-1}(D^{-1})\sum_{k+|\gamma|=1}\partial^\gamma\sigma_{1-k}(D)\delta^\gamma \sigma_{-1}(D^{-1})/\gamma!\\
&=- \sigma_{-1}(D^{-1})^2\sigma_{0}(D)\\
&=- (\xi_1+\tau\xi_2)^{-2}\alpha.
\end{align*}
Finally, we have
\begin{align*}
\sigma_{-2,0}(\log\Delta\, D^{-1})&=\sigma_{-1,0}(\log\Delta)\sigma_{-1}(D^{-1})+\sigma_{0,0}(\log\Delta)\sigma_{-2}(D^{-1})\\
&=(\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2)^{-1}(\xi_1+\tau\xi_2)^{-1} \left[(\alpha+\alpha^*)\xi_1+(\bar\tau\alpha+\tau\alpha^*)\xi_2\right]\\
&-\log ( (\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2)/|\xi|^2) (\xi_1+\tau\xi_2)^{-2}\alpha.
\end{align*}
Therefore, we compute the variation:
\begin{align}\label{variation}
\delta_{\bar{w}} \sigma_{-2,0}(\log\Delta\, D^{-1})&=(\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2)^{-1} \left[(\delta_{\bar w}\alpha^*)\xi_1+(\tau\delta_{\bar w}\alpha^*)\xi_2\right](\xi_1+\tau\xi_2)^{-1}\nonumber\\
&=(\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2)^{-1}(\delta_{\bar w}\alpha^*)\nonumber\\
&=(\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2)^{-1}(\delta_{{w}}D)^*.
\end{align}
In order to compute the variation of the residue density, we need to integrate (\ref{variation}) with respect to $\xi$ variable:
$$\delta_{\bar{w}}{\rm res}(\log(\Delta) D^{-1})=\int_{|\xi|=1}(\xi_1^2+2\Re(\tau) \xi_1\xi_2+ |\tau|^2\xi_2^2)^{-1}(\delta_{{w}}D)^*d\xi=\frac{1}{2\pi\Im(\tau)}(\delta_{{w}}D)^*.$$
Note that we have used the normalized Lebesgue measure in the last integral (see \eqref{pseudodef}).
\end{proof}
We record the main result of this paper in the following theorem. It computes the curvature of the determinant line bundle in terms of the natural K\"{a}hler form on the space of connections.
\begin{theorem}
The curvature of the determinant line bundle for the noncommutative two torus is given by
\begin{equation}
\delta_{\bar w}\delta_{w}\zeta'(0)=\frac{1}{4\pi\Im(\tau)}\varphi_0\left(\delta_w D(\delta_w D)^*\right).
\end{equation}\qed
\end{theorem}
\begin{remark}
In order to recover the classical result of Quillen for $\theta=0$, we have to take into account the change of the volume form due to a change of the metric. This means we have to multiply the above result by $\Im(\tau)$.
\end{remark}
\def\polhk#1{\setbox0=\hbox{#1}{\ooalign{\hidewidth
\lower1.5ex\hbox{`}\hidewidth\crcr\unhbox0}}} \def$'${$'$}
|
\section{Introduction}
\label{sec:intro}
Real world data often contain some degrees of freedom that might be redundant. Matrix
decomposition~\cite{Golub:1996,DemmelBook:1997,TrefethenBook:1997} is an
important tool in machine learning and data mining to normalize data.
A prominent example of data normalization by matrix decomposition is principal
component analysis (PCA). When the given point cloud is represented as a matrix
with each row being coordinates of points, PCA removes the degree of freedom in
translation and rotation of the point cloud with the help of singular value decomposition (SVD) on the matrix.
The selection of particular matrix decomposition corresponds to which
degrees of freedom we would like to remove. In the PCA example, SVD extracts an
orthonormal basis that makes the normalized data invariant to rotation.
There are cases when other degrees of freedom exist in data. For example, planar
objects like digits, characters or iconic symbols, often look distorted in
photos because the camera sensor plane may not be parallel to the plane carrying
the objects.
Therefore in this case, the degrees of freedom we would like to eliminate from
data are homography transforms \cite{hartley2003multiple}, which can be
approximated as combination of translation, rotation, shearing and squeezing when the planar objects are
sufficient far away relative to their size.
However, PCA is not applicable to eliminate these degrees of freedom, because
the normalized form found with PCA is not invariant under shearing and
squeezing. In general, based on the property of data, we would need new data
normalization methods that can uncover invariant structures depending on the
degrees of freedom we would like to remove.
\begin{figure}
\begin{center}
\includegraphics[height=32mm, width=100mm]{sparse_linear.pdf}
\end{center}
\caption{Normalization by optimization over orbit generated by special
linear group $\mathfrak{SL}(2)$ for 2D point clouds. The first row contains point
clouds before normalization; the second row consists of corresponding point clouds after
normalization for each entry in the first row.
It can be observed that point clouds in the second row are approximately the
same, modulo four orientations (rotated clockwise by angle of 0,
$\frac{\pi}2$, $\pi$, $\frac{3\pi}2$).}
\label{fig:special_linear}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[height=32mm, width=100mm]{bunnies.pdf}
\end{center}
\caption{Normalization by optimization over orbit generated by special
linear group $\mathfrak{SL}(3)$ for 3D point clouds. The first row contains point
clouds before normalization.
In particular, the ``rabbits'' are of different shapes and sizes.
The second row consists of corresponding point clouds after normalization
for each entry in the first row.
It can be observed that point clouds in the second row are approximately
the same, modulo different orientations of the same shape.}
\label{fig:special_linear_3d}
\end{figure}
In this paper we study the cases when degrees of freedom to be removed have a
group structure ${\mathfrak G}$ when combined. Under such a condition, a data matrix ${\bf X}$
can be mapped to its quotient set ${\bf X}/\sim$ by the equivalence relation $\sim$
defined as \[x_1\sim x_2\iff \exists g\in{\mathfrak G}, x_1 = g x_2 \text{.}\] We call the
elements of quotient set $\hat{{\bf X}} \in {\bf X}/\sim$ canonical forms of data, as they
are invariant with respect to (w.r.t.) group actions $g\in{\mathfrak G}$. An important
example of using the quotient set is the shape space method
\cite{dryden1998statistical}, which works in the quotient space of rotation
matrix and is closely related to PCA and SVD.
Here and later, we restrict ourselves to the case when ${\mathfrak G}$
is a matrix group and when the group acts by simple matrix product. The
quotient mapping ${\bf X}\rightarrow \hat{{\bf X}}$ can then be represented in the form of
matrix decomposition:
\[{\bf X} = {\bf G}\hat{{\bf X}}\text{, } {\bf G}\in{\mathfrak G}\text{.}\]
Instead of constructing separate algorithms for different ${\mathfrak G}$, we use an
optimization process to induce corresponding matrix decomposition techniques.
In particular, given a data matrix ${\bf M}$, we consider a group orbit optimization
(GOO) problem as follows:
\begin{align}
\label{eq:goo}
\inf_{{\bf G} \in {\mathfrak G}} \; \phi({\bf G} {\bf M}),
\end{align}
where $\phi:
{\mathbb F}^{n_1\times n_2}\rightarrow {\mathbb R}$ is a cost function and ${\mathbb F}$ is some number field.
In Section~\ref{sec:prelim} we present
several special classes of cost functions, which are used to construct new
formulations for several matrix decompositions including SVD, Schur, LU,
Cholesky and QR in Section~\ref{sec:main}. As an application, in Section~\ref{sec:normal} we illustrate how to use GOO to normalize low dimensional point cloud data over a special linear group.
Experiment results for two-dimensional and three-dimensional point cloud are
given in Figure~\ref{fig:special_linear} and Figure~\ref{fig:special_linear_3d}.
It can be observed that the effect of rotation, shearing and squeezing in
data has been mostly eliminated in the normalized point clouds. The detail of
this normalization is explained in Section~\ref{sec:normal}.
The GOO formulation also allows us to construct
generalizations of some matrix decompositions to tensor. Real world
data have tensor structure when some value depends on multiple factors. For
example, in an electronic-commerce site, user preferences in different brands
form a matrix. As such preferences change over time, the time-dependent
preferences form a \nth{3} order tensor. As in the matrix case, tensor
decomposition techniques~\cite{Kolda:2009:TDA:1655228.1655230,
krishnamurthy2013low} aim to eliminate degrees of freedom in data while respecting the tensor structure of data. In
Section~\ref{sec:tensor}, we use GOO to induce tensor decompositions that can be
used for normalizing tensor. In the unified framework of GOO, the GOO inducing
tensor decomposition when applied to a \nth{2} order tensor, is exactly the same
as the GOO inducing matrix decomposition, when the same group and cost function
is used for both GOO problems.
The remainder of paper is organized as follows. Section~\ref{sec:notation} gives
notation used in this paper. Section~\ref{sec:prelim} defines several
properties for describing the cost function used in defining GOO to induce
matrix and tensor decompositions. Section~\ref{sec:main} studies GOO formulations
that can induce SVD, Schur, LU, Cholesky, QR, etc. Section~\ref{sec:tensor}
demonstrates how to use GOO to induce tensor decompositions and prove a few inequalities relating
a few forms of GOO. Section~\ref{sec:normal} demonstrates how to normalize
point cloud data distorted by rotation, shearing and squeezing with GOO over
the special linear group.
Section~\ref{sec:exps} presents numerical algorithms and examples of matrix
decomposition, point cloud normalization and tensor decomposition.
Finally, we conclude the work in Section~\ref{sec:conclusion}.
\section{Notation}
\label{sec:notation}
\subsection{Matrix operation notation}
In this paper,
we let ${\bf I}_r$ denote the $r\times r$ identity matrix. Given an $n{\times}m$
matrix ${\bf X}=[x_{ij}]$, we denote $|{\bf X}|=[|x_{ij}|]$ and $\operatorname{vec}({\bf X})=[x_{11},
\ldots, x_{n1}, x_{12}, \ldots, x_{nm}]^\top$. The $\ell_p$-norm of ${\bf X}$ is
defined by \[\|{\bf X}\|_p\overset{\mathrm{def}}{=\joinrel=} (\sum_{ij} |x_{ij}|^p)^{\frac1p}\] for $p\ge 0$. Note that
we abuse the notation a little bit as $\|{\bf X}\|_p$ is not a norm when $p<1$.
When $p=2$, it is also called the Frobenius norm and usually denoted by
$\|{\bf X}\|_F$.
When applied to vector ${\bf x}$, $\|{\bf x}\|_2$ is the $\ell_2$-norm and it is
shortened as $\|{\bf x}\|$. The dual norm of the $p$-norm where $p\ge 1$ is
equivalent to the $q$-norm, where $\frac{1}{p}+\frac{1}{q}=1$.
We let $\|{\bf X}\|_{*p}$ denote the Schatten $p$-norm; that is, it is the $\ell_p$ norm of the vector of
the singular values of ${\bf X}$.
Assume that $\mathbb{F}$ is some number field.
Let $\conj{{\bf X}}$ be the complex conjugate
of ${\bf X}$, and ${\bf X}^*$ be the complex conjugate transpose of ${\bf X}$.
Let $\operatorname{dg}({\bf M})$ be a vector consisting of the diagonal entries of ${\bf M}$, and
$\operatorname{diag}({\bf v})$ be a matrix with ${\bf v}$ as its diagonals.
Given two matrices ${\bf A}$ and ${\bf B}$, ${\bf A}\odot {\bf B}$ is
their Hadamard product and
${\bf A} \otimes {\bf B}$ is the Kronecker product. Similarly, ${\bf x}\otimes {\bf y}$ is the Kronecker product of vectors ${\bf x}$ and
${\bf y}$. For groups ${\mathfrak G}_1$ and ${\mathfrak G}_2$, we denote group
$\{{\bf G}_1\otimes{\bf G}_2:\, {\bf G}_1\in{\mathfrak G}_1, {\bf G}_2\in{\mathfrak G}_2\}$ as ${\mathfrak G}_1\otimes{\mathfrak G}_2$. The
Kronecker sum for two square matrices ${\bf A}\in{\mathbb F}^{m\times m}, {\bf B}\in{\mathbb F}^{n\times
n}$ is defined as \[{\bf A}\oplus{\bf B} = {\bf A}\otimes {\bf I}_n + {\bf I}_m \otimes {\bf B} \text{.}\]
\begin{definition}
A matrix ${\bf A} \in \mathbb{F}^{m\times n}$ is said to be pseudo-diagonal if
there exist permutation matrices ${\bf P}$ and ${\bf Q}$ such that ${\bf P} {\bf A} {\bf Q}^\top$
is diagonal.
\end{definition}
\begin{remark}
Note that a diagonal matrix is also pseudo-diagonal.
\end{remark}
\begin{lemma}
Given a pseudo-diagonal matrix ${\bf A}$, we have that
\begin{enumerate}
\item[\emph{(i)}] ${\bf A}^*{\bf A}$, ${\bf A} {\bf A}^*$, ${\bf A}^\top{\bf A}$ and ${\bf A} {\bf A}^\top$
are diagonal.
\item[\emph{(ii)}] There exists a row permutation matrix ${\bf P}$ such that ${\bf P} {\bf A}$ is
diagonal.
\item[\emph{(iii)}] There exists a row permutation matrix ${\bf P}$ such that ${\bf A} {\bf P}^\top$
is diagonal.
\end{enumerate}
\end{lemma}
We let $\operatorname{Poly}{({\bf M})}$ be the polyhedral formed by points with coordinates being
rows of ${\bf M}$, and $\mu(\operatorname{Poly}{({\bf M})})$ be the Lebesgue measure of $\operatorname{Poly}{({\bf M})}$.
We let $\operatorname{Rasterize}(\operatorname{Poly}{({\bf M})})$ be a matrix ${\bf Z}$ where $z_{ij}$ is the
image pixel value at coordinate $(i, j)$ of image rasterized from polyhedral
$\operatorname{Poly}{({\bf M})}$ with unit grid.
\subsection{Tensor operation notation}
The notation of tensor operations used in this paper mostly follows that of
\cite{Kolda:2009:TDA:1655228.1655230}.
Given an order-$k$ tensor ${\mathcal X} \in \mathbb{F}^{n_1\times n_2\times \ldots \times n_k}$
and $k$ matrices $\{{\bf U}_i\}_{i=1}^k$ where $ {\bf U}_i \in
\mathbb{F}^{m_i\times n_i}$, we define $\times_k$ to be the inner product over the
$k$-th mode. That is, if ${\mathcal Y} = {\mathcal X} \times_a {\bf U}_a\in \mathbb{F}^{n_1\times
n_2\times \ldots \times n_{a-1} \times m_a \times n_{a+1} \times \ldots \times
n_k}$, then \[y_{i_1\cdots i_{a-1}j i_{a+1}\cdots i_k} = \sum_{i_a=1}^{n_a} x_{i_1i_2\cdots i_k} u_{j
i_a}.
\]
For shorthand,
we denote \[\prod_i {\mathcal X} {\bf U}_i \overset{\mathrm{def}}{=\joinrel=} {\mathcal X} \times_1 {\bf U}_1
\times_2 {\bf U}_2 \cdots \times_k {\bf U}_k \textrm{.}\]
Here ${\mathcal Y} = \prod_i {\mathcal X} {\bf U}_i$ when $\forall i, m_i = n_i$ is also known as
the Tucker decomposition in the literature \cite{tucker1966some}. With this notation,
the SVD of a real matrix ${\bf M}={\bf U}_1\mbox{\boldmath$\Sigma$\unboldmath} {\bf U}_2^\top$ can
be written as \[{\bf M}= \mbox{\boldmath$\Sigma$\unboldmath} \times_1 {\bf U}_1\times_2 {\bf U}_2=\prod_{i=1}^2
\mbox{\boldmath$\Sigma$\unboldmath} {\bf U}_i\text{.}\]
Using the vectorization operation for tensor, we have
\[
\operatorname{vec}{(\prod_i {\mathcal X} {\bf U}_i)} = [{\bf U}_n \otimes{\bf U}_{n-1}\otimes\dotsb\otimes{\bf U}_{1}] \operatorname{vec}{({\mathcal X})}
\overset{\mathrm{def}}{=\joinrel=} \otimes^\downarrow_i{\bf U}_i \operatorname{vec}{({\mathcal X})},
\]
where we denote $\otimes^\downarrow_i{\bf U}_i$ as shorthand for
${\bf U}_n\otimes{\bf U}_{n-1}\otimes\dotsb\otimes{\bf U}_{1}$.
We let $\operatorname{index}_{n_1, n_2, \ldots, n_k}(I)$ be a map from a sequence of indices $I = {i_1, i_2, \ldots,
i_k}$ to an integer such that
\begin{align}
[\operatorname{vec}{{\mathcal X}}]_{\operatorname{index}_{n_1, n_2, \ldots, n_k}({i_1, i_2, \ldots, i_k})} =
{\mathcal X}_{i_1, i_2, \ldots, i_k}
\textrm{.}
\end{align}
We note that $\operatorname{index}^{-1}_N(n)$ is well-defined.
The unfold operation maps a tensor to a tensor of lower order and is defined by
\[\operatorname{fold}^{-1}_J: {\mathbb F}^{n_1\times n_2\times \ldots \times n_k}\mapsto
\mathbb{F}^{m_1\times m_2\times \ldots \times m_l}\;,\] where $J$ is an index
set grouping of the indices $I=\{1, 2, \ldots, k\}$ into sets $J=\{J_1, J_2, \ldots,
J_l\}$, $m_o=\prod_{t\in J_o} t$, and satisfies:
\[
\operatorname{vec}[\operatorname{fold}^{-1}_J({\mathcal A})] = \operatorname{vec}({\mathcal A}) \text{.}
\]
When unfolding a single index, i.e.,
$J=\{\{j\}, I-\{j\}\}$, we also denote $\operatorname{fold}^{-1}_J$ as $\operatorname{fold}^{-1}_j$.
The $\ell_p$-norm of tensor ${\mathcal A}$ is defined as \[\|{\mathcal A}\|_p \overset{\mathrm{def}}{=\joinrel=}
\|\operatorname{fold}^{-1}_i \mathcal A\|_p\] for an arbitrary mode $i$.
For tensors ${\mathcal A}, {\mathcal B}\in {\mathbb F}^{n_1\times n_2\times \ldots \times n_k}$,
$\langle {\mathcal A}, {\mathcal B} \rangle$ is their Frobenius inner product
defined as:
\[
\langle {\mathcal A}, {\mathcal B} \rangle\overset{\mathrm{def}}{=\joinrel=}\langle \operatorname{vec}({\mathcal A}), \operatorname{vec}({\mathcal B}) \rangle \text{.}
\]
Finally, given $f: {\mathbb F}\rightarrow{\mathbb F}$ and ${\mathcal T}\in{\mathbb F}^{n_1\times n_2\times \ldots
\times n_k}$, $f({\mathcal T})$ is defined as a tensor-valued function with $f$ applied to
each entry of ${\mathcal T}$. Therefore, $f({\mathcal T})\in{\mathbb F}^{n_1\times n_2\times \ldots \times
n_k}$. When $f(x) = |x|$, we denote $f({\mathcal T})$ as $|{\mathcal T}|$.
\subsection{Group notation}
$\mathfrak{O}$ is the orthogonal group over real field ${\mathbb R}$. $\mathfrak{SO}$ is the special
orthogonal group over ${\mathbb R}$.
$\mathfrak{U}$ is the unitary group over complex field. We let
$\mathfrak{UUT}(n)$ denote the upper-unit-triangular group and $\mathfrak{LUT}(n)$ denote
the lower-unit-triangular group, both of which have all entries along the
diagonals being $1$.
$\mathfrak{H}$ is the group formed by (calibrated) homography transform below:
\[
{\bf H}_{2w} = {\bf R}_{2w}({\bf I}_3+ {\bf p}_2 \frac{{\bf n}^\top}{d}) \text{,}
\]
where ${\bf R}_{2w}\in\mathfrak{SO}$ is attitude of the camera; ${\bf p}_2$
is position of the camera, and ${\bf n}^\top {\bf x} = d$ is equation of the object
plane.
\section{Preliminaries}
\label{sec:prelim}
In this paper we would like to show that matrix and tensor decompositions
techniques can be induced from formulations of the group orbit optimization.
As we have seen in formula (\ref{eq:goo}), a GOO problem includes two key
ingredients:
a cost function $\phi$ and a group structure ${\mathfrak G}$.
Thus, we present preliminaries, including \emph{sparsifying function}
and a \emph{unit matrix group}. The sparsifying functions will be used to define
cost functions for some matrix decompositions in Table~\ref{tab:matrix-decomp}
that have diagonal matrices in decomposed formulations.
It should be noted that other classes of functions can be used together with
some unit matrix groups to induce interesting matrix and tensor decompositions.
Confer Schur decomposition in Table~\ref{tab:matrix-decomp} for an example.
\subsection{Sparsifying functions}
For two functions $f$ and $g$, we here and later denote their composition as $f\circ g$ s.t.\
$f\circ g(x)\overset{\mathrm{def}}{=\joinrel=} f(g(x))$.
We first prove several utility lemmas used for characterizing sparsifying
functions.
\begin{lemma}[Subadditive properties] \label{lem:41}
If $f(\sqrt{x})$ is subadditive, then
\begin{enumerate}
\item[\emph{(1)}] $\sum_{i=1}^n f(|x_i|)\geq f(\|{\bf x}\|)$ where
${\bf x}\in{\mathbb F}^n$.
\item[\emph{(2)}] $f(0)\geq 0$.
\end{enumerate}
\end{lemma}
\begin{proof} First we have that
$\sum_{i=1}^n f(|x_i|) = \sum_{i=1}^n f(\sqrt{x_i^2}) \geq
f(\sqrt{\sum_{i=1}^n x_i^2}) = f(\|{\bf x}\|)$.
By the subadditivity of $f(\sqrt{x})$ we further have $f(\sqrt{0}) +
f(\sqrt{0}) \ge f(\sqrt{0 + 0}) = f(\sqrt 0)$, hence $f(0) = f(\sqrt 0) \ge 0$.
\end{proof}
\begin{lemma} \label{lem:42}
If $f(e^x)$ is convex for any $x \in {\mathbb F}$, then when $\forall i, x_i\neq 0$, we
have: \[ \sum_{i=1}^n f(|x_i|) \geq n f\big((\prod_{i=1}^n |x_i|)^{\frac 1
n}\big)\quad .
\]
\end{lemma}
\begin{proof}
Since $f(e^x)$ is convex, we have
\[
\sum_{i=1}^n f(|x_i|) = \sum_{i=1}^n f(e^{\ln |x_i|})\geq n f(e^{\frac{1}{n} \sum_{i=1}^n \ln |x_i|})
= n f((\prod_{i=1}^n |x_i|)^{\frac{1}{n}}).
\]
\end{proof}
\begin{lemma}\label{lem:43} If $f$ is strictly concave and $f(0)\ge 0$, then
$f(tx)\geq t f(x)$ where $0\leq t \leq 1$, with equality only when
$t=0,1$ or $x=0$.
\end{lemma}
\begin{proof} We have $f(tx) = f(tx +(1-t)0)\geq t f(x)+(1-t)f(0)\geq t f(x)$.
Obviously, the first equality holds only when $x=0$ or $t=0,1$.
\end{proof}
\begin{lemma}
\label{lem:44}
Assume $f(x) = f(|x|)$. Then $f$ is
concave and $f(0)\geq 0$ iff $f$ is concave and subadditive.
\end{lemma}
\begin{proof} Because $f(x) = f(|x|)$, w.l.o.g.\ we assume $x \ge 0$.
We first prove ``$\Rightarrow$ part". When $a=0$ and $b=0$, we trivially have
$f(a)+f(b)\ge f(a+b)$. Otherwise, we have \[f(tx) = f(tx
+(1-t)0)\geq t f(x)+(1-t)f(0)\geq t f(x)\text{.}\] Thus, when $a\neq 0$ or $b
\neq 0$, \[f(a)+f(b)=f(\frac{(a+b)
a}{a+b})+f(\frac{(a+b) b}{a+b})\ge \frac{a}{a+b} f(a+b)+\frac{b}{a+b}
f(a+b)=f(a+b)\text{.}\]
As for ``$\Leftarrow$ part", we have $f(0) + f(0)\geq f(0 + 0)$.
Hence $f(0)\geq 0$.
\end{proof}
Now we are ready to define the sparsifying function.
\begin{definition}[sparsifying function] \label{def:42}
A function f is sparsifying if
\begin{enumerate}
\item[\emph{(a)}] $f$ is symmetric about the origin; i.e., $f(x)=f(|x|)$;
\item[\emph{(b)}] $f(\sum_i |x_i|)=\sum_i f(|x_i|) \Longrightarrow$ there is
at most one $i$ with $x_i\neq 0$.
\end{enumerate}
\end{definition}
The following theorem gives a sufficient condition for function $f$ to be sparsifying.
\begin{theorem}[sufficient condition for sparsifying]\label{thm:sparsifying}
If $f(x) = f(|x|)$ and $f$ is strictly concave and subadditive, then $f$ is
sparsifying.
\end{theorem}
\begin{proof}
Because $f(x) = f(|x|)$, w.l.o.g.\ we assume $x\ge 0$. By Lemma~\ref{lem:44}, $f$ is strictly concave and $f(0)\ge 0$.
When $\sum_i x_i= 0$, there is no $i$ with $x_i=0$. Otherwise, it follows from
Lemma~\ref{lem:43} that
\[
\sum_i f(x_i) = \sum_i f \big(\frac{x_i}{\sum_{j}
x_j} \sum_j x_j \big) \geq \sum_i \frac{x_i}{\sum_{j} x_j} f(\sum_j
x_j)=f(\sum_j x_j)\text{.}
\]
Also by Lemma~\ref{lem:43}, the equality holds iff $\frac{x_i}{\sum_j x_j}=0
\text{ or } 1$. Because $\sum_i \frac{x_i}{\sum_j x_j} = 1$, there is only
one $i$ with $x_i\neq 0$.
In both cases, there is at most one $i$ with $x_i\neq 0$.
\end{proof}
\begin{corollary} Conical combination of sparsifying functions. In particular,
if $f$ and $g$ are sparsifying, then so is $\alpha f + \beta g$ where $\alpha$ and $\beta$ are two nonnegative constants.
\end{corollary}
\begin{proof}
As strict concavity is preserved by conical combination, we only need
prove subadditivity is preserved by conical combination, which holds because:
\begin{align*}
(\alpha f + \beta g)(x + y) &= \alpha f(x + y) + \beta g(x + y) \\
&\le \alpha f(x) + \alpha f(y) + \beta g(x) + \beta g(y) \\
&= (\alpha f +
\beta g)(x) + (\alpha f + \beta g)(y).
\end{align*}
\end{proof}
It can be directly checked that the following functions are sparsifying.
\begin{example}
Following functions are sparsifying:
\begin{enumerate}
\item[\emph{(1)}] Power function: $f(x)=|x|^p$ for $0<p<1$;
\item[\emph{(2)}] Capped power function: $f(x)=\min(|x|^p,1)$ for $0<p<1$;
\item[\emph{(3)}] $f(x)=-|x|^p$ for $p>1$;
\item[\emph{(4)}] $f(x)= \log(1+|x|)$;
\item[\emph{(5)}] Shannon Entropy: $f(x)=-|x| \log |x|$ when $0\leq x$;
\item[\emph{(6)}] Squared entropy: $f(x)=-x^2 \log x^2$ when $0\leq x$;
\item[\emph{(7)}] $f(x)=a-(a+|x|^p)^{\frac1p}$ for $p>1$ and $a\geq 0$;
\item[\emph{(8)}] $f(x)=-a+(a+|x|^p)^{\frac1p}$ for $p<1$ and $a\geq 0$;
\end{enumerate}
\end{example}
\begin{remark}
We note that $\log|x|$ is not subadditive because $f(0)=-\infty<0$.
Although $f(x)=|x|^p$ for $p<0$ is subadditive, $|x|^{p}$ is not
concave. Thus, these two functions are not sparsifying.
\end{remark}
\subsection{Unit Matrix Groups}
\begin{definition}[unit group] \label{def:unit-group}
A matrix group ${\mathfrak G}$ is a \emph{unit group} if $|\det({\bf G})|=1,\, \forall {\bf G} \in
{\mathfrak G}$.
\end{definition}
Clearly, unitary, orthogonal, and unit-triangular matrix groups are unit
groups.
We now present some properties of the unit groups.
\begin{lemma}\label{lem:unit-group} Unit group has the following properties.
\begin{enumerate}
\item[\emph{(i)}] Unit group is well-defined, i.e., closed under multiplication
and inverse, and has an identity element which happens to be ${\bf I}$.
\item[\emph{(ii)}] The Kronecker product of unit groups is also a unit group. In particular, if
${\mathfrak G}_1$ and ${\mathfrak G}_2$ are unit groups, then ${\mathfrak G}_1\otimes {\mathfrak G}_2 =\{{\bf M}_1\otimes {\bf M}_2: {\bf M}_1 \in
{\mathfrak G}_1 \mbox{ } {\bf M}_2\in {\mathfrak G}_2 \}$ is also a unit group.
\item[\emph{(iii)}] $\{{\bf P}\otimes {\bf P}^{-\top}: {\bf P} \in \mathfrak{GL}(n)\}$ is a unit
group.
\item[\emph{(iv)}] $\{{\bf A}\otimes \conj{{\bf A}}: {\bf A} \in \mathfrak{GL}(n)\}$ is a
group, and is a unit group iff $\mathfrak{GL}(n)$ is a unit group.
\item[\emph{(v)}] $\{{\bf I}_n\otimes {\bf A}, {\bf A}\in{\mathfrak G}\}$ is a unit group iff ${\mathfrak G}$ is
a unit group. $\{{\bf A}\otimes{\bf I}_n, {\bf A}\in{\mathfrak G}\}$ is a unit group iff ${\mathfrak G}$ is
a unit group.
\end{enumerate}
\end{lemma}
\begin{proof}
\begin{enumerate}
\item [{(i)}] Let
${\bf G}_1,{\bf G}_2\in{\mathfrak G}$. Then $|\det({\bf G}_1^{-1})| = 1$ and \[ |\det( {\bf G}_1 {\bf G}_2 )| = |\det( {\bf G}_1
)| |\det({\bf G}_2)| = 1.
\]Hence ${\bf G}_1^{-1}\in{\mathfrak G}$ and ${\bf G}_1{\bf G}_2\in{\mathfrak G}$.
\item [{(ii)}] We first check ${\mathfrak G}_1\otimes {\mathfrak G}_2$ is a group. This
can be done by noting that $({\bf G}_1\otimes{\bf G}_2)^{-1} = {\bf G}_1^{-1}\otimes{\bf G}_2^{-1}\in{\mathfrak G}_1\otimes {\mathfrak G}_2$
when ${\bf G}_1\in{\mathfrak G}_1$, ${\bf G}_2\in{\mathfrak G}_2$; and \[({\bf G}_1\otimes{\bf G}_2)
({\bf G}_3\otimes{\bf G}_4) = ({\bf G}_1{\bf G}_3)\otimes({\bf G}_2{\bf G}_4)\text{.}\] Also ${\bf I}\in
{\mathfrak G}_1\otimes \GM2$. Moreover, since
$|\det({\bf G}_1\otimes {\bf G}_2)| = |\det({\bf G}_1)|^m |\det({\bf G}_2)|^n = 1$ for any ${\bf G}_1
\in {\mathfrak G}_1$ and ${\bf G}_2 \in {\mathfrak G}_2$, ${\mathfrak G}_1\otimes {\mathfrak G}_2$ is a unit group.
\item [{(iii)}] Closedness under multiplication and inverse can be
proved by noting
\[({\bf P}\otimes {\bf P}^{-\top})({\bf Q}\otimes {\bf Q}^{-\top})=({\bf P} {\bf Q})\otimes ({\bf P}^{-\top}
{\bf Q}^{-\top}) =({\bf P} {\bf Q})\otimes ({\bf P} {\bf Q})^{-\top}\text{.}\] Also we have \[({\bf P}
\otimes {\bf P}^{-\top})^{-1}={\bf P}^{-1}\otimes {\bf P}^{\top}\text{.}\] Thus ${\bf P}
\otimes {\bf P}^{-\top}$ forms a group with ${\bf I}$ as the identity. It is also a unit
group as $|\det({\bf P}\otimes {\bf P}^{-\top})|=|\det({\bf P})^n \det({\bf P}^{-\top})^n|=1$.
\item[{(iv)}] Closedness under multiplication and inverse can be
proved based on \[({\bf L}\otimes \conj{{\bf L}})({\bf R}\otimes \conj{{\bf R}})=({\bf L} {\bf R}) \otimes
(\conj{{\bf L}} \conj{{\bf R}})=({\bf L} {\bf R}) \otimes \conj{{\bf L} {\bf R}}\text{,}\] and $({\bf L}\otimes
\conj{{\bf L}})^{-1}={\bf L}^{-1} \otimes \conj{({\bf L}^{-1})}$. Thus ${\bf L}\otimes \conj{{\bf L}}$
forms a group with ${\bf I}$ as the identity.
Moreover $|\det({\bf L}\otimes \conj{{\bf L}})|=|\det({\bf L})^n
\det(\conj{{\bf L}})^n|=|\det({\bf L})|^{2n}$, i.e., ${\bf L}\otimes \conj{{\bf L}}$ forms a unit group iff ${\bf L}$ is from a unit group.
\item[{(v)}] Note $\{{\bf I}_n\}$ is a unit group with single element. By property
{(ii)} we can prove this property.
\end{enumerate}
\end{proof}
It is worth pointing out that ${\bf P}\otimes {\bf P}^{-1}$ does not form a group in
general because $({\bf P}\otimes {\bf P}^{-1})({\bf Q}\otimes {\bf Q}^{-1}) = ({\bf P}{\bf Q})\otimes
({\bf Q}{\bf P})^{-1} \not\equiv ({\bf P}{\bf Q})\otimes
({\bf P}{\bf Q})^{-1}$.
Finally, in Table~\ref{tab:matrix-decomp} we list matrix decompositions of ${\bf X}$ used in
this paper. When referring to the Cholesky decomposition, ${\bf X}$ should be positive definite.
\begin{table}[!ht] \centering \small
\caption{Matrix decompositions}
\begin{center}
\begin{tabular}{p{2cm} p{2.5cm} p{6cm}}
\toprule
Name & Decomposition & Constraint \\
\midrule real SVD & ${\bf X} = {\bf U} {\bf D} {\bf V}^\top$ & ${\bf U}, {\bf V}\in \mathfrak{O}(n)$, ${\bf D}$ is
diagonal\\
\midrule complex SVD & ${\bf X} = {\bf U} {\bf D} {\bf V}^*$ & ${\bf U},{\bf V} \in \mathfrak{U}(n)$, ${\bf D}$ is
diagonal\\
\midrule QR & ${\bf X} = {\bf Q} {\bf D} {\bf R}$ & ${\bf Q}\in\mathfrak{U}(n), {\bf R}\in\mathfrak{UUT}(n)$, ${\bf D}$ is
diagonal\\
\midrule LU & ${\bf X} = {\bf L} {\bf D} {\bf U}$ & ${\bf L}\in\mathfrak{LUT}(n), {\bf U}\in\mathfrak{UUT}(n)$, ${\bf D}$ is
diagonal \\
\midrule Cholesky & ${\bf X} = {\bf L} {\bf D} {\bf L}^\top$ & ${\bf L}\in\mathfrak{LUT}(n)$,
${\bf D}$ is diagonal \\
\midrule Schur & ${\bf X} = {\bf Q} {\bf U}{\bf Q}^*$ & ${\bf Q}\in\mathfrak{U}(n)$, ${\bf U}$ is upper triangular
\\
\hline
\end{tabular}
\end{center} \label{tab:matrix-decomp}
\end{table}
\section{Group Orbit Optimization}
\label{sec:main}
\subsection{Matrix Decomposition Induced from Group Orbit Optimization}
\label{subsec:induced-matrix-decomp}
\subsubsection{GOO formulation}
We now illustrate how matrix decomposition can be induced from GOO.
Given two groups $\GM1, \GM2$ and a data matrix
${\bf M}$, we consider the following optimization problem
\begin{align}
\inf_{{\bf G}_1 \in {\mathfrak G}_1, {\bf G}_2 \in {\mathfrak G}_2} \; \phi({\bf G}_2 {\bf M}{\bf G}_1^\top).
\label{eq:matrix-induce}
\end{align}
Assume that $\hat{{\bf G}_1}$ and $\hat{{\bf G}_2}$ are minimizers of the above
GOO and ${\bf D} = \hat{{\bf G}}_2 {\bf M}\hat{{\bf G}}_1^\top$,
then we refer to
\[{\bf M} = \hat{{\bf G}}_2^{-1} {\bf D} \hat{{\bf G}}_1^{-\top}\text{,}\]
as a matrix decomposition of ${\bf M}$ which is
induced from Formula~(\ref{eq:matrix-induce}).
When $\phi = \varphi \circ \operatorname{vec}$, an equivalent formulation of
Formula~(\ref{eq:matrix-induce}) is:
\[
\inf_{{\bf G}_1 \in {\mathfrak G}_1, {\bf G}_2 \in {\mathfrak G}_2} \; \phi({\bf G}_2
{\bf M}{\bf G}_1^\top)\equiv\inf_{{\bf G}\in{\mathfrak G}} \varphi({\bf G} \operatorname{vec}({\bf M}))\text{,}
\]
where ${\bf G}={\bf G}_1\otimes {\bf G}_2\in{\mathfrak G}$ and ${\mathfrak G}\overset{\mathrm{def}}{=\joinrel=}{\mathfrak G}_1\otimes{\mathfrak G}_2$.
\subsubsection{GOO over unit group}
For a general matrix group ${\mathfrak G}$, ${\bf G}\in{\mathfrak G}$ implies that $|\det({\bf G})| > 0$.
However, group structure may not be sufficient to induce non-trivial matrix
decomposition, as with some groups and cost functions the infimum will be
trivially zero. For example, with general linear group $\mathfrak{GL}$ and for any matrix
${\bf M}$, we have
\[
\inf_{{\bf G}\in\mathfrak{GL}} \|{\bf G} {\bf M}\|_p = 0,
\]
because $s{\bf I} \in \mathfrak{GL}$ and
\[\lim_{s\to 0} \inf_{s\in{\mathbb R}} \|s {\bf I} {\bf M}\|_p = \lim_{s\to0}s\|{\bf M}\|_p =0\;.
\]
Nevertheless, if we require ${\mathfrak G}$ to be a unit group, we have $|\det({\bf G})| = 1$.
Consequently, we can prevent the infimum from vanishing trivially for any
$\ell_p$-norm.
Thus, we mainly consider the case where ${\mathfrak G}$ is a unit group in this paper.
The following theorem shows that many matrix decompositions can be induced from the group orbit optimization.
\begin{theorem}
\label{theorem:decompostion-as-optimization}
SVD, LU, QR, Schur and
Cholesky decompositions of matrix ${\bf M} \in
\mathbb{F}^{m\times n}$ can be induced from GOO of the form
\[\inf_{{\bf G}_1 \in {\mathfrak G}_1, {\bf G}_2 \in {\mathfrak G}_2} \; \phi({\bf G}_2 {\bf M}{\bf G}_1^\top)\text{,}
\]
by using the corresponding unit group ${\bf G}$ and
cost function $\phi$, which are given in Table~\ref{tab:group}.
\end{theorem}
Clearly, the matrix groups in Table~\ref{tab:group} are unit groups by
Lemma~\ref{lem:unit-group}.
We will prove the rest of theorem in Section~\ref{subsec:diag-goo} and
Section~\ref{subsec:triang-goo}.
\begin{remark}
The cost function
for SVD, QR and Matrix Equivalence can be $\phi({\bf X}) = \|{\bf X}\|_p, \, 1\le p <2$. And
the cost function for LU, Schur and Cholesky can be $\phi({\bf X}) =
\sum_{ij}\|x_{ij} {\mathbb I}_{\{i<j\}}\|_p, \, 1\le p <2$.
\end{remark}
\begin{table}[!ht] \centering \small
\caption{Matrix decompositions induced from optimizations}
\begin{center}
\begin{tabular}{p{3.8cm} p{3.8cm} p{4cm}}
\toprule
Decomposition & Unit group ${\mathfrak G}={\mathfrak G}_1\otimes{\mathfrak G}_2$ & Objective function
$\phi({\bf X})$
\\
\midrule real SVD: ${\bf U} {\bf D} {\bf V}^\top$ & $\{{\bf V}\otimes {\bf U}: {\bf U}, {\bf V}\in \mathfrak{O}(n)\}$
& $\sum_{ij} f(|x_{ij}|)$ where $f(\sqrt{x})$ is strictly
concave, $f(0)\ge0$\\
\midrule complex SVD: ${\bf U} {\bf D} {\bf V}^*$ & $\{\conj{{\bf V}}\otimes {\bf U}: {\bf U},{\bf V}
\in \mathfrak{U}(n)\}$ & $\sum_{ij} f(|x_{ij}|)$ where $f(\sqrt{x})$ is strictly
concave, $f(0)\ge0$\\
\midrule QR: ${\bf Q} {\bf D} {\bf R}$ & $\{{\bf R}^\top\otimes {\bf Q}: {\bf Q}\in\mathfrak{U}(n),
{\bf R}\in\mathfrak{UUT}(n)\}$ & $\sum_{ij} f(|x_{ij}|)$ where $f(\sqrt{x})$ is strictly
concave and increasing, $f(0)\ge0$\\
\midrule Matrix Equivalence: ${\bf P} {\bf D} {\bf Q}$ & $\{{\bf Q}^\top\otimes {\bf P}:
{\bf Q}, {\bf P}\in\mathfrak{SL}(n)\}$ & $\sum_{ij} f(|x_{ij}|)$ where
$f(0)\ge0$, $f(\sqrt{x})$ is strictly concave and increasing;
$f(\sqrt{e^x})$ is convex\\
\midrule LU: ${\bf L} {\bf D} {\bf U}$ & $\{{\bf I}\otimes {\bf L}:
{\bf L}\in\mathfrak{LUT}(n)\}$ & $\sum_{ij} f(|x_{ij} {\mathbb I}_{\{i<j\}} |)$ where
$f(x)\not\equiv 0$, $f(x)=0\Rightarrow x=0$, $f(x)\ge 0$\\
\midrule Cholesky: ${\bf L} {\bf D} {\bf L}^\top$ & $\{{\bf I}\otimes {\bf L}:
{\bf L}\in\mathfrak{LUT}(n)\}$ & $\sum_{ij} f(|x_{ij} {\mathbb I}_{\{i<j\}} |)$ where
$f(x)\not\equiv 0$, $f(x)=0\Rightarrow x=0$, $f(x)\ge 0$\\
\midrule Schur: ${\bf Q} {\bf U}{\bf Q}^*$ & $\{\conj{\bf Q}\otimes {\bf Q}:
{\bf Q}\in\mathfrak{U}(n)\}$ & $\sum_{ij} f(|x_{ij} {\mathbb I}_{\{i<j\}} |)$ where
$f(x)\not\equiv 0$, $f(x)=0\Rightarrow x=0$, $f(x)\ge 0$\\
\end{tabular}
\end{center} \label{tab:group}
\end{table}
\begin{remark}
The formulation of QR decomposition exploits
the fact that ${\bf M}={\bf Q} {\bf R}$ is
equivalent to ${\bf M}={\bf Q}({\bf D} \tilde{{\bf R}})$ where ${\bf Q}\in\mathfrak{U}(n)$, ${\bf R}$ is
upper-triangular, $\tilde{{\bf R}}\in\mathfrak{UUT}(n)$, and ${\bf D}$ is diagonal.
\end{remark}
\begin{remark}
``Matrix Equivalence" in Table~\ref{tab:group} finds
a diagonal matrix equivalent to an invertible matrix ${\bf M}$ as defined in
Section~\ref{subsub:equiv}.
\end{remark}
\begin{remark}
However, there are matrix decompositions whose formulation cannot be
expressed as GOO in the same way as Table~\ref{tab:group}. For example, Polar
decomposition ${\bf M}={\bf U} {\bf L} {\bf D} {\bf L}^*$ where ${\bf U}\in\mathfrak{U}(n)$ and ${\bf L}\in\mathfrak{LUT}(n)$, though
derivable from SVD, cannot be induced from a GOO formulation of
diagonalization.
This is because $\conj{{\bf L}}\otimes {\bf U} {\bf L}$ does not form a group as it is not
closed under multiplication.
For another example, consider a formulation of decomposition ${\bf M} ={\bf L} {\bf D}
{\bf L}^{-\top}$ where ${\bf L}\in\mathfrak{LUT}(n)$ and ${\bf D}$ is diagonal. As we stated earlier, ${\bf L}\otimes
{\bf L}^{-1}$ is not a group in general, so ${\bf S} ={\bf L} {\bf D} {\bf L}^{-\top}$ cannot be induced
from a GOO formulation of diagonalization.
\end{remark}
\begin{remark}
For matrix decomposition of the form ${\bf M} ={\bf A} {\bf D} {\bf B}^\top$, where ${\bf A} \in \mathbb{F}^{m\times r}$ and ${\bf B} \in
\mathbb{F}^{n\times r}$ with $r\leq \min(m, n)$.
In this case, we can zero-pad ${\bf D}$ to $\tilde{{\bf D}}\in\mathbb{F}^{m\times n}$, and
extend ${\bf A}$ and ${\bf B}$ to $\tilde{{\bf A}}\in\mathbb{F}^{m\times m}$ and
$\tilde{{\bf B}}\in\mathbb{F}^{n\times n}$ which are square matrices. Accordingly, we formulate a decomposition ${\bf M} =\tilde{{\bf A}}\tilde{{\bf D}}\tilde{{\bf B}}^\top$ which may be induced from GOO.
\end{remark}
We next prove a lemma that characterizes the optimum.
\begin{lemma}[Criteria for infimum] \label{lem:criteria-infimum}
If $\phi({\bf G} {\bf D}) \geq \phi({\bf D})$ for any ${\bf G}\in{\mathfrak G}$ and there exists ${\bf A} \in
{\mathfrak G}$ s.t.\ ${\bf M}= {\bf A} {\bf D}$, then
\[
\inf_{{\bf G}\in{\mathfrak G}} \phi({\bf G} {\bf M}) = \phi({\bf D}).
\]
\end{lemma}
\begin{proof}
We note that $\inf_{{\bf G} \in {\mathfrak G}} \phi({\bf G} {\bf M}) = \inf_{{\bf G} \in {\mathfrak G}} \phi({\bf G} {\bf A}
{\bf D})$. By the group structure, the coset $\{{\bf G} {\bf A}: {\bf G}\in{\mathfrak G}\} =
{\mathfrak G}$. Hence we have
\[
\inf_{{\bf G}\in{\mathfrak G}} \phi({\bf G} {\bf M}) = \inf_{{\bf G} \in {\mathfrak G}} \phi( {\bf G} {\bf A} {\bf D}) =
\inf_{{\bf G}\in{\mathfrak G}} \phi({\bf G} {\bf D}).
\]
Using the condition $\forall {\bf G}\in{\mathfrak G}, \phi({\bf G}{\bf D}) \geq \phi({\bf D})$, we
have \[ \inf_{{\bf G}\in{\mathfrak G}} \phi({\bf G} {\bf D}) \geq
\inf_{{\bf G}\in{\mathfrak G}} \phi({\bf D})=\phi({\bf D}).
\]
On the other hand, as ${\bf I}\in{\mathfrak G}$ we have $\phi({\bf D})=\phi({\bf I}{\bf D})\geq
\inf_{{\bf G}\in{\mathfrak G}} \phi({\bf G} {\bf D})$.
Hence \[\phi({\bf D}) = \inf_{{\bf G}\in{\mathfrak G}} \phi({\bf G} {\bf D})= \inf_{{\bf G}\in{\mathfrak G}} \phi({\bf G}
{\bf M})\text{.}\]
\end{proof}
By virtue of Lemma~\ref{lem:criteria-infimum}, if we want to prove that matrix decomposition
$\operatorname{vec}{({\bf M})} =\tilde{{\bf G}}\operatorname{vec}{({\bf D})}$ is induced by a GOO {w.r.t.\ } $\phi$ and ${\mathfrak G}$, we only need prove
that there exists a $\tilde{\bf G}\in{\mathfrak G}$ s.t.\ $\operatorname{vec}{({\bf M})} =\tilde{{\bf G}}\operatorname{vec}{({\bf D})}$, and
$\phi({\bf G} \operatorname{vec}({\bf D})) \geq \phi(\operatorname{vec}({\bf D}))$ $\forall {\bf G}\in{\mathfrak G}$.
The equality condition will determine the uniqueness of the
optimum of the optimization problem.
\subsection{Matrix Diagonalization as GOO}\label{subsec:diag-goo}
Next we demonstrate how matrix diagonalization can be induced from GOO with
proper choice of cost function and unit group.
\subsubsection{Singular Value Decomposition}
First we discuss SVD of a complex matrix and of a real matrix.
\begin{lemma}[Cost function and group for SVD]\label{lem:svd-cost} Let
${\bf D}=[d_{ij}]$ be pseudo-diagonal, and ${\bf U},{\bf V} \in\mathfrak{U}(n)$.
Given a function $f$ such that $f(x) = f(|x|)$ and $f(\sqrt{|x|})$ is
strictly concave and subadditive, and $\phi({\bf X}) = \sum_{ij} f(x_{ij})$ we have \[
\phi({\bf U}{\bf D}{\bf V}^* )\geq \phi({\bf D}), \]
with equality iff there exists a row
permutation matrix ${\bf P}$ such that $|{\bf U}{\bf D}{\bf V}^*| = |{\bf P}{\bf D}|$.
Furthermore, if ${\bf U}, {\bf V} \in\mathfrak{O}(n)$, we have
\begin{align}\label{ineq:svd}
\phi({\bf U}{\bf D}{\bf V}^\top)\geq
\phi({\bf D}),
\end{align}
with equality iff there exists a row permutation
matrix ${\bf P}$ such that $|{\bf U}{\bf D}{\bf V}^\top| = |{\bf P}{\bf D}|$.
\end{lemma}
\begin{proof}
First we prove the inequality.
We write $g(x)=f(\sqrt{x})$ and ${\bf A} =
{\bf U}{\bf D}{\bf V}^*$. We let $g({\bf X}) = [g(x_{ij})]$ be a matrix-valued function of ${\bf X}=[x_{ij}]$. As $g$ is
concave and subadditive, by Lemma~\ref{lem:41} for a vector ${\bf v}=(v_1, \ldots,
v_n)^\top$, we have $\sum_{i=1}^n f(v_i) \ge f(\norm{\bf v})= g({\bf v}^*{\bf v})$. Applying
this to each column of ${\bf A}$, we have
\begin{align}\label{eq:subadditive}
\phi({\bf A})\geq \operatorname{tr}[ g ({\bf A}^*{\bf A})] = \operatorname{tr}[g({\bf V}{\bf D}^*{\bf D}{\bf V}^*)].
\end{align}
Alternatively, we can also apply the inequality to each row of ${\bf A}$ and have
\begin{align}\label{eq:subadditive-2}
\phi({\bf A})\geq \operatorname{tr}[ g ({\bf A} {\bf A}^*)] = \operatorname{tr}[g({\bf U}{\bf D}\D^*{\bf U}^*)].
\end{align}
As ${\bf D}$ is pseudo-diagonal, ${\bf D}^*{\bf D}$ is diagonal. Because $g$ is concave and ${\bf V} {\bf V}^*={\bf I}$, we can apply Jensen's
inequality, obtaining \[
\operatorname{tr}( g({\bf V}{\bf D}^*{\bf D} {\bf V}^*))\geq \operatorname{tr} ({\bf V} (g ({\bf D}^*{\bf D})){\bf V}^*).
\]
Hence altogether we have:
\[
\phi({\bf A})\ge\operatorname{tr} (g ({\bf V}{\bf D}^*{\bf D} {\bf V}^*))\geq \operatorname{tr}
({\bf V} (g ({\bf D}^*{\bf D})){\bf V}^*) = \operatorname{tr} (g({\bf D}^*{\bf D})) = \sum_{ij} f(d_{ij}) =\phi({\bf D}).
\]
Next we check the equality condition.
By Theorem~\ref{thm:sparsifying}, $g$ is sparsifying. For the
equality condition in inequality~(\ref{eq:subadditive}) to hold, ${\bf A}$ can have
at most one nonzero in each column. By the symmetry between (\ref{eq:subadditive}) and
(\ref{eq:subadditive-2}), and noting $\phi({\bf A}^\top) = \phi({\bf A})$ and $
\phi({\bf D}) = \phi({\bf D}^\top)$, ${\bf A}$ can also have at most one nonzero in each row for
$\phi({\bf A}) = \phi({\bf D})$ to hold.
Hence when the equality holds, ${\bf A}$ is pseudo-diagonal. Then there exists a
permutation matrix ${\bf P}$ such that ${\bf Z} = {\bf P}^{-1}|{\bf A}| = {\bf P}^{-1} {\bf Q} {\bf A}$ is a
diagonal matrix with elements on diagonal in descending order and are all non-negative,
where ${\bf Q}$ is a diagonal matrix s.t.\ $|{\bf Q}| = {\bf I}$.
By the uniqueness of singular values of a matrix, we have ${\bf Z} = |{\bf D}|$. Hence
equality in inequality\ref{ineq:svd} holds when $|{\bf A}| = {\bf P}|{\bf D}| = |{\bf P}{\bf D}|$.
The proof for ${\bf U}, {\bf V} \in\mathfrak{O}(n)$ is similar.
\end{proof}
Note that ${\bf D}$, modulo sign and permutation, is the global
minimizer for a large class of functions $f$.
After applying Lemma~\ref{lem:criteria-infimum}, we have the following theorem.
\begin{theorem}[SVD induced from optimization]
\label{thm:svd-opt}
We are given a function $f$ such that
$f(x) = f(|x|)$ and $f(\sqrt{|x|})$ is strictly concave and subadditive, and
$\phi({\bf X}) = \sum_{ij} f(x_{ij})$.
Let $\hat{{\bf U}}$ and $\hat{{\bf V}}$ be an optimal solution of the following optimization:
\[
\inf_{{\bf U}\in\mathfrak{U}(n),{\bf V}\in\mathfrak{U}(n)} \phi({\bf U}^*{\bf M}{\bf V}).
\]
Then if SVD of ${\bf M}$ is ${\bf M}={\bf U}{\bf S}{\bf V}^*$, there exist a permutation matrix
${\bf P}$ and a diagonal matrix ${\bf Z}$ such that ${\bf M} = \hat{{\bf U}}{\bf P}
{\bf Z}\hat{{\bf V}}^*$ and $|{\bf Z}|={\bf S}$.
\end{theorem}
\begin{corollary}
With $\phi$ as in Theorem~\ref{thm:svd-opt}, eignedecomposition of a Hermitian
matrix ${\bf M}$ can be induced from
\[
\inf_{{\bf U}\in\mathfrak{U}(n)} \phi({\bf U}^*{\bf M}{\bf U}).
\]
Similarly, eignedecomposition of a real symmetric matrix ${\bf M}$ can be induced from
\[
\inf_{{\bf U}\in\mathfrak{O}(n)} \phi({\bf U}^\top{\bf M}{\bf U}).
\]
\end{corollary}
From the above optimization, we can derive several inequalities.
\begin{corollary}[The Schatten $p$-norm and $\ell_p$-norm inequality]
\label{cor:schatten-ineq}
The
$\ell_p$-norm of matrix ${\bf A}$ is larger (smaller) than the Schatten $p$-norm of
${\bf A}$ when $0\leq p< 2$ $(>2)$.
In particular, we have
\[
\|{\bf M}\|_p \ge \inf_{{\bf U},{\bf V}\in\mathfrak{U}} \|{\bf U}{\bf M}{\bf V}^*\|_p = \|{\bf M}\|_{*p} \quad \text{ when }
0<p< 2,
\]
and
\[
\|{\bf M}\|_p \le \sup_{{\bf U},{\bf V}\in\mathfrak{U}} \|{\bf U}{\bf M}{\bf V}^*\|_p = \|{\bf M}\|_{*p}\quad \text{ when }
p> 2.
\]
\end{corollary}
\begin{proof} $f(x)=|x|^p$ satisfies $f(0)\geq 0$, and $f(\sqrt{|x|})$ is
strictly concave when $0\leq p <2$. On the other hand, $f(x)=-|x|^p$ satisfies
$f(0) \geq 0$ and $f(\sqrt{|x|})$ is strictly concave when $p>2$. By
Theorem~\ref{thm:svd-opt} we have
\[
\inf_{{\bf U},{\bf V}\in\mathfrak{U}} \|{\bf U}{\bf M}{\bf V}^*\|_p = \|{\bf M}\|_{*p} \quad \text{ when }
0<p< 2,
\]
and
\[
\inf_{{\bf U},{\bf V}\in\mathfrak{U}} -\|{\bf U}{\bf M}{\bf V}^*\|_p = -\|{\bf M}\|_{*p}\quad \text{ when }
p> 2.
\]
\end{proof}
\begin{corollary} [Duality gap]
Given SVD of ${\bf M}$ as ${\bf M}={\bf U}{\bf D}{\bf V}^*$, we have
\[ \|{\bf M}\|_p \geq \|{\bf D}\|_p\geq \|{\bf D} \|_q\geq \|{\bf M}\|_q \]
where $0<p<2<q$.
\end{corollary}
\begin{proof}
Due to the non-increasing property of the $\ell_p$-norm {w.r.t.\ } $p$, we have
$\|{\bf D}\|_p\ge \|{\bf D}\|_q$. Also by Theorem~\ref{thm:svd-opt} we have
$\|{\bf D}\|_p = \|{\bf M}\|_{*p}$ and $\|{\bf D}\|_q
= \|{\bf M}\|_{*q}$. Applying Corollary~\ref{cor:schatten-ineq} completes the proof.
\end{proof}
\begin{corollary}
The von Neumann entropy of density matrix ${\bf M}$ is smaller than the sum of the Shannon
entropies of rows (columns) of ${\bf M}$.
\end{corollary}
\begin{proof}
By noting that $f(x) = -x\log{x}$ is strictly concave and $f(0)\geq 0$ when
$x\geq 0$, and that the von Neumann entropy is entropy of diagonal matrix in SVD
of ${\bf M}$, the inequality holds.
\end{proof}
\subsubsection{QR Decomposition}
To derive GOO for QR decomposition, we first note that QR decomposition of a
matrix ${\bf M}$ can be rewritten as ${\bf M}= {\bf Q} {\bf D} {\bf R}$, where ${\bf Q}\in\mathfrak{U}(n)$,
${\bf D}$ is diagonal, and ${\bf R}\in\mathfrak{UUT}(m)$.
\begin{lemma}[Cost function and group for QR]
\label{lem:qr-cost}
Let $f$ satisfy that $f(x) =
f(|x|)$, $f(\sqrt{x})$ is concave and increasing; $f(0)\geq 0$. Let $\phi({\bf X}) = \sum_{ij} f(x_{ij})$.
If ${\bf Q}\in\mathfrak{U}(n)$, ${\bf R}\in\mathfrak{UUT}(n)$, and ${\bf D}$ is diagonal, then we have
\[
\phi({\bf Q} {\bf D} {\bf R}) \geq \phi({\bf D}),
\]
with equality when \ $|{\bf Q} {\bf D} {\bf R}| = |{\bf P}{\bf D}|$, where ${\bf P}$
is a row permutation matrix.
\end{lemma}
\begin{proof}
Let $g(x)=f(\sqrt{|x|})$ and ${\bf A}={\bf Q}{\bf D}{\bf R}$, and let $g({\bf X})=[g(x_{ij})]$.
First we prove the inequality.
As $g(x)$ is sparsifying and increasing, we have
\[
\phi({\bf A})\geq \operatorname{tr} [g ({\bf A}^*{\bf A})] = \operatorname{tr} [g ({\bf R}^* {\bf D}^*{\bf D}
{\bf R})] \geq \operatorname{tr} [g ({\bf D}^*{\bf D})] = \phi({\bf D}).
\]
Next we check the equality condition.
For the equality to hold in inequality $\operatorname{tr}[ g ({\bf R}^* {\bf D}^*{\bf D}
{\bf R})] \geq \operatorname{tr}[g ({\bf D}^*{\bf D})]$, as $g$ is increasing, ${\bf R}$ needs to be diagonal. Hence, ${\bf R}
= {\bf I}$.
Now for the equality to hold in $\phi({\bf A})\ge \operatorname{tr} [g ({\bf A}^*{\bf A})]$, ${\bf A}$ needs to be
pseudo-diagonal.
Thus, the equality holds only when $|{\bf A}| = {\bf P}|{\bf D}| = |{\bf P} {\bf D}|$ where ${\bf P}$ is a
row permutation matrix.
\end{proof}
Similarly, we derive the optimization inducing QR decomposition.
\begin{theorem}[QR induced from optimization]
Assume that the conditions are satisfied in Lemma~\ref{lem:qr-cost}.
Let $(\hat{{\bf Q}}, \hat{{\bf R}})$ be optimal solution of optimization as follows
\[
\inf_{{\bf Q}\in\mathfrak{U}(n),{\bf R}\in\mathfrak{UUT}(n)} \phi({\bf Q}^{-1}{\bf M}{\bf R}^{-1})\textrm{.}
\]
Then ${\bf M} =
\hat{{\bf Q}} {\bf Z}$ is the QR decomposition of ${\bf M}$, where
${\bf Z} = \hat{{\bf Q}}^{-1}{\bf M}$.
\end{theorem}
\subsubsection{Matrix Equivalence by the Special Linear
Group}\label{subsub:equiv}
An interesting question is whether we can extend the following optimization form
to more general groups ${\mathfrak G}_1$ and ${\mathfrak G}_2$:
\[
\inf_{{\bf U}\in{\mathfrak G}_1, \; {\bf V}\in{\mathfrak G}_2} \; \phi({\bf U}{\bf M}{\bf V}).
\]
It turns out that we can use the special linear group to construct a unit group, and
hence, induce matrix equivalence from an optimization.
\begin{lemma}[Matrix equivalence decomposition]
An invertible matrix ${\bf M} \in \mathbb{F}^{n\times n}$ can be decomposed as
${\bf M}={\bf A} {\bf D} {\bf B}$, where $\det({\bf A})=\det({\bf B})=1$, and ${\bf D}=\det({\bf M})^{1/n} {\bf I}$.
\end{lemma}
\begin{proof}
Just let
\begin{align}
{\bf A}=\det({\bf M})^{-1/n} {\bf M}, \quad {\bf B}={\bf I}_n, \quad {\bf D} =\det({\bf M})^{1/n} {\bf I},
\end{align}
which gives an existence proof.
\end{proof}
\begin{lemma}
\label{lem:matrix-equivalence-cost}
If $f(\sqrt{x})$ is strictly concave and increasing, and $f(\sqrt{e^x})$ is
convex, $f(0)\geq 0$, ${\bf A},{\bf B}\in\mathfrak{SL}(n)$ and ${\bf D} = \lambda{\bf I}$, then $\phi({\bf A} {\bf D}
{\bf B})\geq \phi({\bf D})$, with equality iff there exists a permutation matrix ${\bf P}$
such that $|{\bf A} {\bf D} {\bf B}| = |{\bf P} {\bf D}|$.
\end{lemma}
\begin{proof}
We write $g(x)=f(\sqrt{x})$ for shorthand.
First prove the inequality.
As ${\bf D} = \lambda{\bf I}_n$, we have
\[ \phi({\bf A} {\bf D} {\bf B}) = \phi({\bf A}{\bf B} {\bf D}).
\]
As $f(\sqrt{x})$ is strictly concave and $f(0)\geq 0$, we have
\[\phi({\bf A}{\bf B} {\bf D}) \ge \operatorname{tr}[ g ({\bf A}{\bf B} {\bf D}\D^*{\bf B}^*{\bf A}^*)] = \operatorname{tr}[
g (\lambda^2{\bf A}{\bf B}\B^*{\bf A}^*)].
\]
As ${\bf A}{\bf B}\B^*{\bf A}^*$ is Hermitian, we let its LDL decomposition be ${\bf L}{\bf Z}{\bf L}^*$ where
${\bf L}\in\mathfrak{LUT}(n)$. Because $g$ is increasing, we have
\[
\operatorname{tr}[ g (\lambda^2 {\bf L}{\bf Z}{\bf L}^*)] \ge \operatorname{tr}[ g (\lambda^2 {\bf Z})].
\]
Because $g(e^x)$ is convex and
\[
\det({\bf Z}) = \det({\bf L}{\bf Z}{\bf L}^*) = \det({\bf A}{\bf B}\B^*{\bf A}^*) =1\text{,}
\] we have
\[
\operatorname{tr}[ g (\lambda^2 {\bf Z})] \ge \operatorname{tr}[ g(\lambda^2 {\bf I})].
\]
In summary, we have
\[
\phi({\bf A}{\bf D} {\bf B})\ge \operatorname{tr}[ g(\lambda^2 {\bf I})] = \operatorname{tr}[ g({\bf D}\D^*)] = \phi({\bf D})\textrm{.}
\]
Next we check the equality condition. The equality holds in \[\operatorname{tr}[ g(\lambda^2
{\bf L}{\bf Z}{\bf L}^*)] \ge \operatorname{tr}[ g(\lambda^2 {\bf I})]\] iff ${\bf L}{\bf Z}{\bf L}^* = {\bf I}$ and ${\bf A}{\bf B}{\bf D}$ is
pseudo-diagonal. Hence, the equality holds iff $|{\bf A} {\bf D} {\bf B}| = {\bf P} |{\bf D}| = |{\bf P}
{\bf D}|$.
\end{proof}
\begin{theorem}[Matrix equivalence induced from optimization]
Let $f$ satisfy that $f(x) =
f(|x|)$, $f(\sqrt{x})$ is concave, $f(\sqrt{e^x})$ is convex and increasing, and
$f(0)\geq 0$.
Let $(\hat{{\bf A}}, \hat{{\bf B}})$ be an optimal solution of the following optimization problem
\[
\inf_{{\bf A}\in\mathfrak{SL}(n),{\bf B}\in\mathfrak{SL}(n)} \phi({\bf A}^{-1}{\bf M}{\bf B}^{-1})\textrm{.}
\]
Then there exists a row permutation matrix ${\bf P}$ such that ${\bf M} =
\hat{{\bf A}} ({\bf P} {\bf Z}) \hat{{\bf B}}$ is the matrix equivalence decomposition of ${\bf M}$ with
${\bf P} {\bf Z} = \lambda{\bf I}$, where ${\bf Z} = (\hat{{\bf A}}{\bf P})^{-1}{\bf M}\hat{{\bf B}}^{-1}$.
\end{theorem}
\subsection{Matrix Triangularization as GOO}\label{subsec:triang-goo}
Next we demonstrate how matrix triangularization can be induced from GOO with
proper choice of cost function and unit group. In fact, we can prove that any
triangularization can be induced from optimization {w.r.t.\ } a masked norm.
\begin{lemma}
A matrix decomposition ${\bf M} = {\bf G}_2 {\bf U} {\bf G}_1^\top, {\bf G}_1\in{\mathfrak G}_1, {\bf G}_2\in{\mathfrak G}_2$,
where ${\bf U}$ is upper triangular and ${\mathfrak G}_1 \otimes {\mathfrak G}_2$ is a unit group, can
be induced from the following optimization:
\begin{align}\label{form:triang}
\inf_{{\bf G}_1\in{\mathfrak G}_1, {\bf G}_2\in{\mathfrak G}_2} \phi({\bf G}_2{\bf M}{\bf G}_1^\top)
\textrm{.}
\end{align}
Here $\phi({\bf X}) = \sum_{ij} f(|x_{ij} {\mathbb I}_{\{i<j\}} |)$ where
$f(x)=0\Rightarrow x=0$, $f(x)\not\equiv 0$, $f(x)\ge 0$.
\end{lemma}
\begin{proof}
We trivially have
\[\forall {\bf G}, \; \phi({\bf G}_2 {\bf U} {\bf G}_1^\top) \ge 0 = \phi({\bf U})\text{.}
\]
By Lemma~\ref{lem:criteria-infimum} and
$\phi({\bf X}) = 0\iff {\bf X}={\bf 0}$, the decomposition can be induced from optimization.
\end{proof}
For example, the Schur decomposition can be induced from
Formula~\ref{form:triang} with $f(x) = |x|$.
We will give a numerical example in
Section~\ref{sec:exps}.
\section{Group Orbit Optimization on Tensor Data}
\label{sec:tensor}
The GOO problem on tensor ${\mathcal T}$ is defined as
\[\inf_{{\bf G}_i \in {\mathfrak G}_i} \phi(\prod_i {\mathcal T} {\bf G}_i).
\]
When there exists function $\varphi$ s.t.\ $\phi({\mathcal T}) = \varphi(\operatorname{vec}({\mathcal T}))$, we
get a form that bears resemblance to the matrix version:
\[\inf_{{\bf G}_i \in {\mathfrak G}_i} \varphi(\operatorname{vec}(\prod_i {\mathcal T} {\bf G}_i) )= \inf_{{\bf G}_i \in {\mathfrak G}_i} \varphi(\otimes^\downarrow_i{\bf G}_i\operatorname{vec}({\mathcal T})).
\]
Similar to the matrix case in Section~\ref{subsec:induced-matrix-decomp}, we now
illustrate how the Tucker decomposition can be induced from an optimization formulation.
Given a group ${\mathfrak G}=\otimes^\downarrow_i{\mathfrak G}_i$ and ${\bf G}=\otimes^\downarrow_i{\bf G}_i$ and a tensor ${\mathcal T}$,
we define the following optimization problem
\[\inf_{{\bf G} \in {\mathfrak G}} \; \varphi({\bf G}
\operatorname{vec}({\mathcal T})) = \inf_{{\bf G}_i \in {\mathfrak G}_i} \; \varphi((\otimes^\downarrow_i{\bf G}_i) \operatorname{vec}({\mathcal T}))
=\inf_{{\bf G}_i \in {\mathfrak G}_i} \; \varphi(\operatorname{vec}(\prod_i{\mathcal T} {\bf G}_i))).
\]
If we assume that $\hat{{\bf G}} =\otimes^\downarrow_i{\hat{{\bf G}}_i}= \arg \inf_{{\bf G} \in
{\mathfrak G}} \; \varphi({\bf G} \operatorname{vec}({\mathcal T}))$ and ${\mathcal Z} = \prod_i {\mathcal T} \hat{{\bf G}}_i $, then
${\mathcal M} = \prod_i {\mathcal Z} \hat{{\bf G}}_i^{-1}$ can be regarded as a tensor decomposition induced from
the optimization problem.
In this section, we particularly generalize the results in
Sections~\ref{sec:prelim} and \ref{sec:main} to tensors. In
Lemma~\ref{lem:subgroup}, we prove that we can use the subgroup relation to
induce a partial order of the infima of GOO.
We also show that
GOO {w.r.t.\ } the special linear group finds the ``sparsest" Tucker-like
decomposition of a tensor, and prove that GOO on tensor ${\mathcal T}$ is ``denser" than
GOO on any matrix unfolded from ${\mathcal T}$.
We also prove Theorem~\ref{thm:unfolded-diagonal}, which says that if a tensor
can be decomposed into a core tensor with certain shape, then it is optimal. As
a consequence, we prove that not all tensors have superdiagonal form under a GOO
{w.r.t.\ } any matrix group.
\subsection{Subgroup Hierarchy}
First we observe the following partial order of infima of GOO induced from a
subgroup relation.
\begin{lemma}[Infima partial order from subgroup relation]\label{lem:subgroup}
If ${\mathfrak G}_1$ is a subgroup of ${\mathfrak G}_2$, then for any $\phi:{\mathbb F}^{n_1\times
n_2\times\cdots n_k} \rightarrow {\mathbb R}$:
\[\inf_{{\bf G} \in {\mathfrak G}_1} \phi({\bf G} {\bf M}) \ge \inf_{{\bf G} \in {\mathfrak G}_2} \phi({\bf G}
{\bf M})\textrm{.}\]
\end{lemma}
\begin{proof}
As ${\mathfrak G}_1$ is a subgroup of ${\mathfrak G}_2$, the set of optimization variables of the left-hand side is
a subset of those of the right-hand side. Hence, the inequality holds.
\end{proof}
\begin{corollary}
For a matrix ${\bf M}$, we can construct an upper bound of the Schatten $p$-norm via
\[
\inf_{{\bf G}\in\mathfrak{U}}\|{\bf G}{\bf M}\|_p \ge \|{\bf M}\|_{*p}.
\]
\end{corollary}
\begin{proof}
\[\inf_{{\bf G}\in\mathfrak{U}}\|{\bf G}{\bf M}\|_p = \inf_{{\bf G}\in\mathfrak{U}}\|({\bf I}\otimes{\bf G})\operatorname{vec}({\bf M})\|_p
\ge \inf_{{\bf G}_1, {\bf G}_2\in\mathfrak{U}}\|({\bf G}_2\otimes{\bf G}_1)\operatorname{vec} ({\bf M}) \|_p =
\|{\bf M}\|_{*p}.\]
\end{proof}
\begin{lemma}[GOO {w.r.t.\ } special linear group]\label{lem:special_linear}
The infimum of GOO {w.r.t.\ } the special linear group is the smallest among all
GOO {w.r.t.\ } a unit matrix group ${\mathfrak G}$ and the same $\phi$ for a tensor, that is,
\begin{align}
\inf_{\otimes^\downarrow_i{\bf G}_i\in{\mathfrak G}} \phi(\prod_i {\mathcal T} {\bf G}_i)
\ge \inf_{{\bf G}_i \in \mathfrak{SL}(n_i)} \phi(\prod_i {\mathcal T} {\bf G}_i) \textrm{.}
\end{align}
\end{lemma}
\begin{proof}
First we note for ${\bf A}\in \mathbb{F}^{m\times m}$, ${\bf B}\in \mathbb{F}^{n\times
n}$, \[ \det({\bf A}\otimes
{\bf B})^{\frac{1}{mn}}=\det({\bf A})^{\frac1m}\det({\bf A})^{\frac1n} \textrm{.}\]
Let ${\bf Z}_i\in\mathfrak{GL}(n_i)$ and $\det(\otimes^\downarrow_i{\bf Z}_i)=1$. We can find
${\bf N}_i\in\mathfrak{SL}(n_i)$ such that \[{\bf N}_i=(\det({\bf Z}_i))^{-\frac1{n_i}}{\bf Z}_i\;.\]
Now as $\det(\otimes^\downarrow_i{\bf Z}_i)=\det(\otimes^\downarrow_i{\bf N}_i)=1$, we
have $\otimes^\downarrow_i{\bf N}_i=\otimes^\downarrow_i{\bf Z}_i$.
By Lemma~\ref{lem:subgroup}, we have
\[
\inf_{\otimes^\downarrow_i{\bf G}_i \in{\mathfrak G}} \phi(\prod_i {\mathcal T} {\bf G}_i)\ge \inf_{{\bf Z}_i\in\mathfrak{GL}(n_i),
\det(\otimes^\downarrow_i{\bf Z}_i)=1} \phi(\prod_i {\mathcal T} {\bf Z}_i) = \inf_{{\bf G}_i
\in \mathfrak{SL}(n_i)} \phi(\prod_i {\mathcal T} {\bf G}_i)\textrm{.}
\]
\end{proof}
Next we show that GOO gives a unified framework for matrix decomposition and
tensor decomposition. We can rewrite decomposition in Table~\ref{tab:group} in
tensor notation as \[{\bf M} =\prod_i {\bf D} {\bf G}_i \text{,}\]
which is induced {w.r.t.\ } a cost function $\phi$ by optimization
\[\inf_{{\bf G}\in{\mathfrak G}_i} \phi(\prod_i{\bf M} {\bf G}_i)\text{.}
\]
Now if there exists $\varphi$ s.t.\ $\phi = \varphi \circ \operatorname{vec}$,
we can generalize $\phi$ to tensor as
\[\tilde{\phi}({\mathcal T}) = <{\bf 1},\, f({\mathcal T})>\text{,}
\]
where ${\bf 1}$ is a tensor with all entries being $1$ and of the same dimension as
${\mathcal T}$.
This inspires us to define a tensor version of the above optimization and
decomposition as below:
\[
{\mathcal T}=\prod_i {\mathcal Z}{\bf G}_i \text{,}\]
and \[
\inf_{{\bf G}\in{\mathfrak G}_i} \tilde{\phi}(\prod_i{\mathcal T}{\bf G}_i)\text{.}
\]
In particular, if a matrix decomposition can be induced by entry-wise cost
function $\phi({\bf M}) = \sum_{ij} f(m_{ij})$ {w.r.t.\ } some unit group, we can consistently
generalize the matrix decompositions to tensors using cost function
$\tilde{\phi}({\mathcal T}) = <{\bf 1}, f({\mathcal T})>$. In this case, there is an inequality relation
that follows from the Lemma~\ref{lem:subgroup}.
\begin{lemma}[Lifting lemma]
\label{lem:lifting}
We are given a tensor ${\mathcal T}$ and its arbitrary unfolding
$\operatorname{fold}^{-1}_I({\mathcal T})$ {w.r.t.\ } an index set grouping $I$.
Let $\{m_i\}$ and $\{n_j\}$ be the sizes of the square matrices before and
after grouping. Then we have
\[
\inf_{{\bf M}_i\in{\mathfrak G}(m_i)} \phi(\prod_i {\mathcal T} {\bf M}_i) \ge
\inf_{{\bf N}_j\in{\mathfrak G}(n_j)} \phi(\prod_j \operatorname{fold}^{-1}_I({\mathcal T}) {\bf N}_j).
\]
\end{lemma}
\begin{proof}
We note ${\mathfrak G}(a)\otimes{\mathfrak G}(b)$ is a subgroup of ${\mathfrak G}(a+b)$. Hence
\begin{align*}
\inf_{{\bf M}_i\in{\mathfrak G}(m_i)} \phi(\prod_i {\mathcal T} {\bf M}_i)
&= \inf_{{\bf M}_i\in{\mathfrak G}(m_i)} \phi((\otimes^\downarrow_i {\bf M}_i)\operatorname{vec} ({\mathcal T})) \\
&\ge
\inf_{{\bf N}_j\in{\mathfrak G}(n_j)} \phi((\otimes^\downarrow_j{\bf N}_j) \operatorname{vec} ({\mathcal T}) )\\
&= \inf_{{\bf N}_j\in{\mathfrak G}(n_j)} \phi(\prod_j \operatorname{fold}^{-1}_I({\mathcal T}) {\bf N}_j).
\end{align*}
\end{proof}
\subsection{An Upper Bound for Some Tensor Norms}
In the literature, there are multiple generalizations of the Schatten $p$-norm to
tensors.
For example, the tensor unfolding
trace norm \cite{Liu:2013:TCE:2412386.2412938, tomioka2010extension} is
defined as a weighted sum of the trace norm of single index unfoldings of the tensor; namely,
\begin{align}\label{eq:tensor-unfold-norm}
\sum_{i=1}^k \alpha_i \|\operatorname{fold}^{-1}_i {\mathcal T}\|_{*} \; ,
\end{align}
where $\alpha_i\ge 0$ and $\sum_i\alpha_i = 1$.
Another generalization as given in \cite{signoretto2010nuclear} is defined by
\begin{align}\label{eq:tensor-unfold-norm-2}
(\sum_{i=1}^k \frac1k \|\operatorname{fold}^{-1}_i {\mathcal T}\|_{*q}^p)^\frac{1}{p} \; .
\end{align}
These tensor norms are interesting as they correspond to the Schatten norm
of matrix. We next study use of the Lemma~\ref{lem:lifting} to construct an upper
bound for the two norms.
Formula~\ref{eq:tensor-unfold-norm} and Formula~\ref{eq:tensor-unfold-norm-2}
try to capture the tensor structure by considering all single-index unfoldings of the tensor. However, there are many unfoldings that
are not single index. In general, for a $k$th-order tensor, there are $2^{k-1}-2$
possible unfoldings, as in the following example.
\begin{example}
A \nth{3} order tensor has 3 unfoldings {w.r.t.\ } the following index set grouping:
$\{\{1\},\{2,3\}\},\{\{2\},\{1,3\}\},\{\{3\},\{1,2\}\}$.
Here $\{\{1\},\{2,3\}\}$ means putting index 1 of the tensor in the first dimension
of the unfolded matrix, and index 2 and 3 of the tensor in the second dimension of the unfolded
matrix.
Additionally, a \nth{4} th-order tensor has 6 unfoldings.
\end{example}
It turns out the following GOO that respects the tensor structure produces an
upper bound for the Schatten $p$-norm of the matrices unfolded from a tensor.
\begin{lemma}[Infimum of GOO {w.r.t.\ } the unitary group]
\label{lem:unitary-goo-infimum}
For $0\le p<2$, and for any index set grouping $J$, we have:
\[\inf_{{\bf G}_i\in\mathfrak{U}_i} \|\prod_i {\mathcal T} {\bf G}_i\|_p \ge \inf_{{\bf G}_j\in\mathfrak{U}_j} \|\prod_j
(\operatorname{fold}^{-1}_J {\mathcal T}) {\bf G}_j\|_p\; . \]
Similarly for $p>2$, we have:
\[\sup_{{\bf G}_i\in\mathfrak{U}_i} \|\prod_i {\mathcal T} {\bf G}_i\|_p \le \sup_{{\bf G}_j\in\mathfrak{U}_j} \|\prod_j
(\operatorname{fold}^{-1}_J {\mathcal T}) {\bf G}_j\|_p\; . \]
\end{lemma}
\begin{proof}
The inequalities is obtained by applying Lemma~\ref{lem:lifting} to GOO {w.r.t.\ }
unitary group and $f(x)=\|x\|_p$ when $0\le p<2$ and $f(x)=-\|x\|_p$ when $p>2$.
\end{proof}
\begin{corollary}
For $0\le p<2$, we have
\[
\inf_{{\bf G}_i\in\mathfrak{U}_i} \|\prod_i {\mathcal T} {\bf G}_i\|_p \ge \sum_{i=1}^k \alpha_i
\|\operatorname{fold}^{-1}_i {\mathcal T}\|_{*} \; ,
\]
where $\alpha_i\ge 0$ and $\sum_i\alpha_i = 1$.
We also have:
\[
\inf_{{\bf G}_i\in\mathfrak{U}_i} \|\prod_i {\mathcal T} {\bf G}_i\|_p \ge
(\sum_{i=1}^k \frac1k \|\operatorname{fold}^{-1}_i {\mathcal T}\|_{*p}^q)^\frac{1}{q}.
\]
\end{corollary}
\begin{proof}
By Lemma~\ref{lem:unitary-goo-infimum}, we have
\[
\inf_{{\bf G}_i\in\mathfrak{U}_i} \|\prod_i {\mathcal T} {\bf G}_i\|_p \ge \max_{1\le i\le k}
\inf_{{\bf G}_j\in\mathfrak{U}_j} \|\prod_j (\operatorname{fold}^{-1}_i {\mathcal T}) {\bf G}_j\|_p = \max_{1\le i\le k}
\|\operatorname{fold}^{-1}_i {\mathcal T}\|_{*p}\;.
\]
We prove the inequalities as the following relations hold:
\[
\max_{1\le i\le k} \|\operatorname{fold}^{-1}_i {\mathcal T}\|_{*p} \ge \sum_{i=1}^k \alpha_i \|\operatorname{fold}^{-1}_i
{\mathcal T}\|_{*} \;,
\]
and
\[
\max_{1\le i\le k} \|\operatorname{fold}^{-1}_i {\mathcal T}\|_{*p} =
(\sum_{i=1}^k\frac1k\max_{1\le i\le k} \|\operatorname{fold}^{-1}_i
{\mathcal T}\|_{*p}^q)^\frac{1}{q} \ge
(\sum_{i=1}^k \frac1k \|\operatorname{fold}^{-1}_i
{\mathcal T}\|_{*p}^q)^\frac{1}{q}\;.
\]
\end{proof}
Due to the above inequality, an optimization that tries to minimize one of
tensor norms defined as in
Formula~\ref{eq:tensor-unfold-norm} and Formula~\ref{eq:tensor-unfold-norm-2} can have the tensor norm replaced by
$\inf_{{\bf G}_i\in\mathfrak{U}_i} \|\prod_i {\mathcal T} {\bf G}_i\|_p$, as minimizing upper bound of a
function $f$ can always minimize $f$.
\subsection{Sparse Structure in Tensor}
The tensor rank is defined as the minimum number of non-zero rank-1 tensors required to
sum up to ${\mathcal T}$, which is a generalization of the matrix rank. In the Tucker
decomposition ${\mathcal T}=\prod_i {\mathcal X}{\bf U}_i$, one can define the Tucker rank {w.r.t.\ } the
different constraints on the ${\bf U}_i$. For example, for a real tensor,
when the ${\bf U}_i$ are required to be orthogonal, the number of nonzeros in ${\mathcal X}$
is defined as a tensor strong orthogonal rank of ${\mathcal T}$ \cite{Kolda:2001:OTD:587708.587830}.
Trivially, the tensor rank is a lower bound of all the Tucker ranks.
Tensor decomposition has a large body of the literature
\cite{harshman1970foundations}
\cite{Zhang:2001:RAH:587704.587760}
\cite{Lathauwer:2000:BRR:354353.354405}
\cite{Lathauwer:2000:MSV:354353.354398}
\cite{Lathauwer:2005:CCD:1039893.1039944}
\cite{DeLathauwer:2008:DHT:1461964.1461967} \cite{ishteva2008dimensionality}, the
interested reader may refer to \cite{Kolda:2009:TDA:1655228.1655230} and the references therein.
We find that the strong orthogonal rank of tensor ${\mathcal T}$
is exactly the infimum of the following GOO problem
\[
\operatorname{rank}_{so}{\mathcal T} = \inf_{{\bf G}_i\in\mathfrak{U}_i} \nnz{\prod_i {\mathcal T} {\bf G}_i}.
\]
\begin{corollary}
\label{cor:special-linear-best} We have
\[
\inf_{\det(\otimes^\downarrow_i{\bf G}_i)=1} \nnz{\prod_i {\mathcal T} {\bf G}_i}
\ge \inf_{{\bf G}_i \in \mathfrak{SL}(n_i)} \nnz{\prod_i {\mathcal T} {\bf G}_i} \ge \operatorname{rank}({\mathcal T}) \textrm{.}
\]
\end{corollary}
\begin{proof}
The first inequality directly follows from Lemma~\ref{lem:special_linear} by
choosing $\phi=\nnz{\cdot}$.
For the second inequality, as the problem $\inf_{{\bf G}_i\in{\mathfrak G}_i} \nnz{\prod_i {\mathcal T}
{\bf G}_i}$ induces a tensor decomposition into $\nnz{\prod_i {\mathcal T} \hat{{\bf G}}_i}$
number of rank-1 tensors, by definition of the tensor rank we have the following inequality:
\[
\inf_{{\bf G}_i\in{\mathfrak G}_i} \nnz{\prod_i {\mathcal T} {\bf G}_i} \ge \operatorname{rank}({\mathcal T}).
\]
\end{proof}
When $\phi=\nnz{\cdot}$, Lemma~\ref{lem:lifting} provides a link between
the rank of the tensor ${\mathcal T}$ and the ranks of the matrices or the vectors unfolded from ${\mathcal T}$.
For example, when ${\mathfrak G}=\mathfrak{U}$ and ${\mathcal T}\neq{\bf 0}$, we have
\[
\inf_{{\bf G}\in\mathfrak{U}} \nnz{{\bf G} \operatorname{vec} ({\mathcal T})} = 1.
\]
However, for the same ${\mathcal T}$ unfolded to a matrix $\operatorname{fold}^{-1}_I ({\mathcal T})$, we have
\[
\inf_{{\bf G}_1\in\mathfrak{U}_1, {\bf G}_2\in\mathfrak{U}_2} \nnz{{\bf G}_1 \operatorname{fold}^{-1}_I({\mathcal T}) {\bf G}_2} =
\rank(\operatorname{fold}^{-1}_I({\mathcal T})) \ge 1 = \inf_{{\bf G}\in\mathfrak{U}} \nnz{{\bf G} \operatorname{vec} ({\mathcal T})}.
\]
Also, for the strong orthogonal rank of ${\mathcal T}$, we have
\[
\operatorname{rank}_{so}{\mathcal T} = \inf_{{\bf G}_i\in\mathfrak{U}_i} \nnz{\prod_i {\mathcal T}{\bf G}_i} \ge
\rank(\operatorname{fold}^{-1}_I({\mathcal T})).
\]
Hence, intuitively, for a higher order tensor ${\mathcal T}$, we can only hope to find
decomposition with progressively ``denser" core than the matrices and the vectors
unfolded from ${\mathcal T}$. This can be describe more formally in the following lemma.
\begin{lemma}[Optimal core when unfoldable to optimal diagonal]
\label{lem:optimal-core}
If a tensor ${\mathcal T}$ admits a decomposition ${\mathcal T}=\prod_i {\mathcal Z} {\bf G}_i$, where ${\bf G}_i\in
{\mathfrak G}_i$,
and there exists an index set grouping $J$ such that $\otimes^\downarrow_i{\mathfrak G}_i$ is
a sugroup of $\otimes^\downarrow_j\tilde{{\mathfrak G}}_j$, and
\[
\inf_{\otimes^\downarrow_j\tilde{{\bf G}}_j\in\otimes^\downarrow_j\tilde{{\mathfrak G}}_j}\varphi((\otimes^\downarrow_j
\tilde{{\bf G}}_j)\operatorname{vec}(\operatorname{fold}^{-1}_J({\mathcal Z}))) = \varphi(\operatorname{vec}(\operatorname{fold}^{-1}_J({\mathcal Z}))),
\]
then ${\bf Z}$ is the optimal sparse core in the following sense:
\[
\varphi(\operatorname{vec}({\mathcal Z})) = \inf_{{\bf G}_i\in{\mathfrak G}_i}\varphi(\operatorname{vec}(\prod_i {\mathcal T} {\bf G}_i)) \; .
\]
\end{lemma}
\begin{proof}
We have
\begin{align*}
\varphi(\operatorname{vec}({\mathcal Z})) \ge \inf_{{\bf G}_i\in{\mathfrak G}_i} \varphi((\otimes^\downarrow_i
{\bf G}_i)\operatorname{vec}({\mathcal Z})) &\ge \inf_{\tilde{{\bf G}}_j\in\tilde{{\mathfrak G}}_j} \phi((\otimes^\downarrow_j
\tilde{{\bf G}}_j)\operatorname{vec}(\operatorname{fold}^{-1}_J({\mathcal Z})))\\ &= \varphi(\operatorname{vec}(\operatorname{fold}^{-1}_J({\mathcal Z}))) =
\varphi(\operatorname{vec}({\mathcal Z})).
\end{align*}
Hence,
\begin{align*}
\varphi(\operatorname{vec}({\mathcal Z})) &=\inf_{{\bf G}_i\in{\mathfrak G}_i} \varphi((\otimes^\downarrow_i
{\bf G}_i)\operatorname{vec}({\mathcal Z})) \\ & = \inf_{{\bf G}_i\in{\mathfrak G}_i} \varphi(\operatorname{vec}(\prod_i {\mathcal Z} {\bf G}_i)) =
\inf_{{\bf G}_i\in{\mathfrak G}_i} \varphi(\operatorname{vec}(\prod_i {\mathcal T} {\bf G}_i)).
\end{align*}
\end{proof}
Applying Lemmas~\ref{lem:svd-cost},
and~\ref{lem:matrix-equivalence-cost} to Lemma~\ref{lem:optimal-core}, we immediately have the following theorem.
\begin{theorem}
\label{thm:unfolded-diagonal}
If a tensor ${\mathcal T}$ admits a decomposition ${\mathcal T}=\prod_i {\mathcal Z} {\bf G}_i$, and there exists
an index set grouping $J$ such that $\operatorname{fold}^{-1}_J ({\mathcal Z})$ is of optimal shape
{w.r.t.\ } $f$ and ${\mathfrak G}$ in Table~\ref{tab:optimal-core}, then ${\mathcal Z}$ is optimal
in the sense that
\[
\phi({\mathcal Z}) = \inf_{{\bf G}_i\in{\mathfrak G}_i} \phi(\prod_i {\mathcal T}{\bf G}_i).
\]
\end{theorem}
\begin{table}[!ht] \centering \small
\caption{Tensor decomposition generalized from matrix decomposition in the
GOO framework}
\begin{center}
\begin{tabular}{p{2.2cm} p{2.8cm} p{1.8cm} p{4.2cm}}
\toprule
Decomposition & Optimal core shape & Unit Group & Objective function
\\
\midrule Tensor SVD & unfoldable to some pseudo-diagonal matrix & $\mathfrak{U}$
& $\sum_{ij} f(|x_{ij}|)$ where $f(\sqrt{x})$ is strictly
concave, $f(0)\ge 0$\\
\midrule Tensor Equivalence & unfoldable to $\lambda {\bf I}_n$ & $\mathfrak{SL}$
& $\sum_{ij} f(|x_{ij}|)$ where $f(\sqrt{x})$ is
strictly concave and increasing; $f(\sqrt{e^x})$ is convex; $f(0)\ge 0$\\
\hline
\end{tabular}
\end{center} \label{tab:optimal-core}
\end{table}
If a tensor ${\mathcal T}$ admits a decomposition ${\mathcal T}=\prod_i {\mathcal Z} {\bf G}_i$, where ${\mathcal Z}$ is
superdiagonal, i.e., $z_{i_1, i_2, \cdots, i_n} \neq 0 \Rightarrow i_1 = i_2
=\cdots=i_n$, then by Theorem~\ref{thm:unfolded-diagonal}, ${\mathcal Z}$ is optimal
under Tensor SVD as $\operatorname{fold}^{-1}_J ({\mathcal Z})$ is diagonal.
However, Theorem~\ref{thm:unfolded-diagonal} covers more cases than the
superdiagonal case, like the example below:
For example, consider the following \nth{4} order
tensor.
\begin{example}[Non-superdiagonalizable optimal tensor]
\label{example:non-superdiagonalizable} We consider \[{\mathcal T}\in{\mathbb F}^{2\times 2\times
2\times 2},\; \operatorname{vec}{({\mathcal T})}=[{{{{1,0}, {0,1}}, {{0,0}, {0,0}}}, {{{0,0},
{0,0}}, {{1,0}, {0,1}}}}]^\top.\] We can unfold ${\mathcal T}$ to ${\bf I}_4$ with index set grouping $\{\{1,2\},\{3,4\}\}$.
Hence, ${\mathcal T}$ cannot be further ``sparsified" by GOO {w.r.t.\ } any matrix group,
even though it is not in superdiagonal form.
\end{example}
\begin{corollary}
There exist tensors that do not have a superdiagonal core under
any Tucker decomposition induced by GOO.
\end{corollary}
\begin{proof}
The tensor ${\mathcal T}$ in Example~\ref{example:non-superdiagonalizable} can be unfolded to a
scaled identity matrix. Hence, by Corollary~\ref{cor:special-linear-best} we have
\[
\inf_{\det(\otimes^\downarrow_i{\bf G}_i)=1} \nnz{\prod_i {\mathcal T} {\bf G}_i}
\ge \inf_{{\bf G}_i \in \mathfrak{SL}(n_i)} \nnz{\prod_i {\mathcal T} {\bf G}_i} = 4.
\]
This means that a Tucker decomposition of ${\mathcal T}$ induced by a GOO will have at least
four non-zero elements in the core matrix. However, the superdiagonal core can
only have at most two non-zero elements. Hence, ${\mathcal T}$ does not have a superdiagonal core under any
Tucker decomposition induced by GOO.
\end{proof}
It is known that the minimal rank tensor decomposition in the Tucker model is not
unique. For example, for tensor ${\mathcal T}\in{\mathbb F}^{2\times 2\times 2}$, we have
\begin{align}
\operatorname{vec}({\mathcal T})=[{{{1, 0}, {0, 3}},{{0, 0}, {0, 2}}}]^\top \textrm{,}
\end{align}
and there exist unitary matrices ${\bf U}_i\in\mathfrak{U}$ so that
\begin{align}
\operatorname{vec}(\prod_i {\mathcal T} {\bf U}_i)\approx
[{{{3.6055, 0},{0, 0.8320}},{{0, 0},{0, -0.5547}}}]^\top \textrm{.}
\end{align}
This means that GOO by $\mathfrak{U}$ {w.r.t.\ } different $f(x)=\|x\|_p$ may lead to
different optimum values. Hence, the class of entry-wise cost functions $f(x)=\|x\|_p,
0\le p<2$, may not be used to induce sparsity when the optimal core tensor cannot be
unfolded to a diagonal matrix. In other words, in the matrix case, any $p$,
$\|x\|_p, 0\le p<2$ can be used to find the sparest core; however, for a tensor
with order larger than 3, only $f(x)=\nnz{x}$ can be used for finding the
``sparsest" core of ${\mathcal T}$ under GOO.
In practice, this may be done by the following asymptotic formulation:
\[
\inf_{{\bf G}_i\in\mathfrak{U}} \nnz{\prod_i{\mathcal T}{\bf G}_i} = \lim_{p\to 0} \inf_{{\bf G}_i\in\mathfrak{U}}
\|\prod_i{\mathcal T}{\bf G}_i\|_p.
\]
In Section~\ref{sec:exps} we will present several concrete examples.
\section{Data Normalization}
\label{sec:normal}
Data normalization seeks to eliminate some arbitrary degrees of freedom in data.
For example, when we are concerned with shape of an object, its attitude and
position in space will become irrelevant. Given a point clouds, which is a
sequence of coordinates of points, we demonstrate a method to
obtain a representation of point clouds that does not depend on its attitude and
position in this section. We craft the method as a special case of GOO
with some particular choice of group and cost function.
\subsection{Shape Analysis: Matching vs. Normalization}
Point cloud data arise when interest points are extracted from images. If there
are $k$ points, each with a $d$-tuple coordinate, a $k\times d$ matrix can be
formed to describe the object.
Given point cloud data describing an object, shape matching tries to find an
object of the closest shape within a candidate set of shapes under some measure.
The shape space method for shape matching works by matching two objects with
known point-to-point correspondence over given group orbits.
For example, if an object described by ${\bf A}$ is known to be a rotated version of
another known object ${\bf B}$, we can find out parameters describing the rotation by
the following optimization formulation:
\[ g({\bf A}, {\bf B}) \overset{\mathrm{def}}{=\joinrel=} \inf_{{\bf R}\in\mathfrak{U}(d)}\|{\bf A}{\bf R} - {\bf B} \| \textrm{.} \] If there
are $n$ candidates $\{{\bf B}_i\}_{i=1}^n$, then the best matching object can be
found by $\arg\min_i g({\bf A}, {\bf B}_i)$.
However, to make the above method work, a point correspondence procedure must
be established in the first place, which means that the same row of ${\bf A}$ and
${\bf B}$ should refer to the same point. This meets difficulties in real world data
applications because
\begin{enumerate}
\item[(1)] ${\bf A}$ and ${\bf B}$ may have different numbers of rows;
\item[(2)] ${\bf A}$ and ${\bf B}$ may have many rows, leading to exponential number of
possible correspondence.
\end{enumerate}
Here we present a method to match point cloud by using normalization to simplify
matching. The first step of the method is normalizing each of objects
$\{{\bf A}_i\}_{i=1}^m$ and $\{{\bf B}_i\}_{i=1}^n$ with following optimization:
\[\hat{{\bf M}} = \mathop{\rm arginf}_{{\bf R}\in\mathfrak{U}(d)} \phi({\bf M} {\bf R}) \textrm{.} \] As in
Section~\ref{subsec:induced-matrix-decomp}, the above optimization leads to
following decompositions:
\[ {\bf A}_i = \hat{{\bf A}}_i {\bf R}_i\text{ and }{\bf B}_i = \hat{{\bf B}}_i {\bf U}_i\text{,} \] where
${\bf R}_i,{\bf U}_i\in{\mathfrak G}_i$ for some group ${\mathfrak G}_i$.
The second step of the method carries out matching of objects $\{{\bf A}_i\}_{i=1}^m$
against $\{{\bf B}_i\}_{i=1}^n$ by using the normalized forms
$\{\hat{{\bf A}}_i\}_{i=1}^m$ and $\{\hat{{\bf B}}_i\}_{i=1}^n$. Matching between
$\{\hat{{\bf A}}_i\}_{i=1}^n$ and $\{\hat{{\bf B}}_i\}_{i=1}^n$ is expected to be simpler
because less degrees of freedom remain after normalization.
A well-known data normalization method is Principal Component Analysis (PCA),
which eliminates the following degrees of freedom: translation, scaling and
rotation.
As any rigid body movement can be expressed as combination of translation and
rotation, PCA provides a method to standardize data {w.r.t.\ } the rigid body
movement.
However, there may be other distortions of data. Thus, we discuss using general
group for normalizing point cloud data to eliminate the effect of non-rigid body
transforms. An illustrative example has been shown in
Figure~\ref{fig:special_linear} and Figure~\ref{fig:special_linear_3d} in
Section~1, where we see that normalized point clouds can be
matched by enumerating a small number of orientation.
\subsection{Normalization of Point Cloud Data by the Special Linear Group}
In this section we use group ${\bf P}^\top\otimes{\bf I}_m, {\bf P}\in\mathfrak{SL}(n)$ for
normalization of point cloud data.
Here we assume the distortions to the point
cloud data are of a few categories of degrees of freedom: including
mirroring, rotation, shearing and squeezing, which we seek to eliminate using
the special linear group orbit.
\begin{lemma} The
special linear group can represent any combination of mirroring, rotation,
shearing and squeezing operations for point cloud data.
\end{lemma}
\begin{proof}
Every special linear matrix ${\bf M}$
can be QR decomposed as ${\bf M} = {\bf Q} {\bf R}$, and ${\bf R}$ can be decomposed into ${\bf R} = {\bf D}
{\bf U}$ where ${\bf D}$ is diagonal and ${\bf U}\in\mathfrak{UUT}$. Accordingly, we have a decomposition ${\bf M} =
{\bf Q} {\bf D} {\bf U}$,
where ${\bf U}$ models the shearing operation and ${\bf Q}$ models the
rotation. As $\det ({\bf M}) = 1$, $|\det {\bf D}| =
|\frac{\det {\bf M}}{\det {\bf Q} \det {\bf U}}| = 1$.
Hence, the diagonal matrix ${\bf D}$ is the squeezing operation (optionally with the
mirror operation). Hence the action of ${\bf G}\in\mathfrak{SL}$ applied to ${\bf M}$ is equivalent
to the sequential application of rotation, squeezing, mirroring, and shearing. Now as
the special linear group is a group, arbitrary composition of these operations
can still be represented as some ${\bf G}'\in\mathfrak{SL}$.
\end{proof}
We show that for some point clouds, the normalized form are exactly the
axis-aligned hypercubes.
\begin{lemma}
Given a matrix ${\bf M}\in{\mathbb R}^{n\times d}$, if $\operatorname{Poly}{({\bf M})}$ is an axis-aligned hypercube and
$\det{{\bf G}}=1$, then
\[
\|{\bf M}{\bf G}\|_\infty \ge \|{\bf M}\|_\infty.
\]
\end{lemma}
\begin{proof}
First note that as $\det{({\bf G})}=1$, given a Lebesgue measure $\mu$, we
have:
\[
\mu(\operatorname{Poly}{({\bf M}{\bf G})}) = \mu(\operatorname{Poly}{({\bf M})})\textrm{.}
\]
We can construct a bounding box $\operatorname{Poly}{({\bf Z})}$ for $\operatorname{Poly}{({\bf M}{\bf G})}$ with center at the
origin and edge length $2\|{\bf M}{\bf G}\|_\infty$. Note that $\operatorname{Poly}{({\bf Z})}$ is also
a hypercube and $\|{\bf Z}\|_\infty=\|{\bf M}{\bf G}\|_\infty$. We thus have
\[
\mu(\operatorname{Poly}{({\bf Z})}) \ge \mu(\operatorname{Poly}{({\bf M}{\bf G})})=\mu(\operatorname{Poly}{({\bf M})})\textrm{.}
\]
Because for any axis-aligned hypercube $\operatorname{Poly}{({\bf Y})}$ we have
$\mu(\operatorname{Poly}{{\bf Y}})=2^d\|{\bf Y}\|_\infty^d$, the following holds:
\[
2^d\|{\bf M}{\bf G}\|_\infty^d=2^d\|{\bf Z}\|_\infty^d \ge 2^d\|{\bf M}\|_\infty^d \textrm{.}
\]
\end{proof}
By Lemma~\ref{lem:criteria-infimum}, we can prove the following corollary.
\begin{corollary}\label{thm:special-linear-normalize}
Let ${\bf M}\in{\mathbb R}^{n\times k}$ be a matrix such that $\operatorname{Poly}({\bf M})$ can be transformed
by ${\bf G}\in\mathfrak{SL}(k)$ into a hypercube.
Then the following optimization problem attains its optimum when $\operatorname{Poly}({\bf M} \hat{{\bf G}})$
is a hypercube:
\[
\inf_{{\bf G}\in\mathfrak{SL}(n)} \|{\bf M} {\bf G}\|_\infty
\]
\end{corollary}
\begin{remark}
The optimization problem in Theorem~\ref{thm:special-linear-normalize} is not convex.
For example, in ${\mathbb R}^2$, a square can be rotated by 90 degrees, 180 degrees,
and 270 degrees while still being a square. Nevertheless, the degree of freedom associated with rotation,
squeezing and shearing described by the special linear group is reduced to
only one of four configurations. The three other optimal $\hat{{\bf G}}$ can be
enumerated when one optimal ${\bf G}$ is known.
\end{remark}
\begin{remark}
Optimality of $\inf_{{\bf G}\in\mathfrak{SL}(2)} \|{\bf M} {\bf G}\|_\infty$ depends
on whether $\operatorname{Poly}({\bf M} {\bf G})$ is a parallelogram. In practice, we
find the above method works well in normalizing general point data, especially
for those arise in shape recognition. In
Section~\ref{sec:exps} we will present several concrete examples.
\end{remark}
\section{Numerical Algorithm and Examples}
\label{sec:exps}
\subsection{Algorithm}\label{subsec:algorithm}
Most of optimizations involved in this paper are constrained optimization
problem of the following form:
\begin{align}\label{opt:constrained-group}
\inf_{{\bf G}\in{\mathfrak G}} \; \varphi({\bf G} \operatorname{vec}({\bf M})) \text{,}\end{align}
where ${\mathfrak G}$ is a unit group. When ${\mathfrak G}$ is a Lie group, alternatively we
can turn the above optimization to another constrained optimization:
\[ \inf_{{\bf G} \in {\mathfrak g}} \; \varphi(\exp({\bf G}) \operatorname{vec}({\bf M}))\text{,} \] where $\exp({\bf G})$
is matrix exponential of matrix ${\bf G}$ and ${\mathfrak g}$ is the Lie algebra associated with
Lie group ${\mathfrak G}$. This formulation may have constraints that are easier to encode
in numeric software. For example, in \[ \inf_{{\bf G}_1\in\mathfrak{SO}(m), {\bf G}_2\in\mathfrak{SO}(n)}
\varphi(\operatorname{vec}({\bf G}_2 {\bf M} {\bf G}_1^\top)) \text{,}\] we have
${\mathfrak G}=\mathfrak{SO}(m)\otimes\mathfrak{SO}(n)$, ${\mathfrak g} = \{{\bf Z}_2\oplus{\bf Z}_1:
{\bf Z}_1\in{\mathbb F}^{n\times n}, {\bf Z}_1 + {\bf Z}_1^\top=0, {\bf Z}_2\in{\mathbb F}^{m\times m},
{\bf Z}_2 + {\bf Z}_2^\top=0\}$.
Hence, we can turn the optimization into a
constrained optimization over ${\mathfrak g}$ as
\[\inf_{{\bf Z}\in{\mathfrak g}}\varphi(\exp({\bf Z})\operatorname{vec}({\bf M}))\text{.}\]
Moreover, in this particular case, we can turn the above optimization into an
unconstrained optimization:\[ \inf_{{\bf G}_1\in{\mathbb F}^{n\times n}, {\bf G}_2\in{\mathbb F}^{m\times m}}
\varphi(\operatorname{vec}(\exp({\bf G}_2-{\bf G}_2^\top) {\bf M}\exp({\bf G}_1-{\bf G}_1^\top)^\top)) \text{.}\]
In Table~\ref{tab:lie-group-encoding} we list a few more cases when the
constrained optimization of Formula~\ref{opt:constrained-group} can be turned
into an unconstrained optimization.
\begin{table}[!ht] \centering \small
\caption{Encoding of constraints for Lie groups}
\begin{center}
\begin{tabular}{p{2cm} p{4cm} p{4cm}}
\toprule
Lie group & Lie algebra & Encoding of Constraint \\
\midrule $\mathfrak{SO}$ & $\{{\bf Z}: {\bf Z}+{\bf Z}^\top=0\}$ & ${\bf Z} = {\bf X} - {\bf X}^\top$\\
\midrule $\mathfrak{LUT}$ & $\{{\bf Z}: \operatorname{dg}{\bf Z}={\bf 0}, {\bf Z}\in\mathfrak{LUT}\}$ & ${\bf Z} = {\bf X} \odot
[{\mathbb I}_{i > j}]$\\
\midrule $\mathfrak{UUT}$ & $\{{\bf Z}: \operatorname{dg}{\bf Z}={\bf 0}, {\bf Z}\in\mathfrak{UUT}\}$ & ${\bf Z} = {\bf X} \odot
[{\mathbb I}_{i < j}]$\\
\midrule $\mathfrak{SL}$ & $\{{\bf Z}: \operatorname{tr}{\bf Z}=0\}$ & ${\bf Z} = {\bf X} - (\operatorname{tr} {\bf X}) {\bf v} {\bf v}^\top$,
where ${\bf v}=[1, 0, 0, \ldots, 0]^\top$\\
\midrule ${\mathfrak G}_1\otimes {\mathfrak G}_2$ & ${\mathfrak g}_1\oplus{\mathfrak g}_2$ & \\
\hline
\end{tabular}
\end{center} \label{tab:lie-group-encoding}
\end{table}
The exponential mapping used for
optimization over Lie groups is related to other optimization on manifold
methods \cite{udriste1994convex} \cite{edelman1998geometry} \cite{absil2009optimization}.
In this section all numerical optimizations are solved by Nelder-Meld heuristic
global optimization algorithm \cite{nelder1965simplex} implemented in Mathematica\texttrademark\
9.0.0, unless noted.
\subsection{GOO Inducing Matrix Decomposition}
We empirically illustrate several examples of GOO inducing matrix decomposition.
Due to the large amount of computation required by Nelder-Meld algorithm, here
we only give a few examples involving small matrices.
\begin{example}[Compute SVD of a $3\times 3$ real matrix] \label{exp:svd}
Given a matrix ${\bf M}$:
\[
{\bf M}\approx\left[
\begin{array}{ccc}
0.17658 & 0.517888 & 0.448587 \\
0.214066 & 0.718154 & 0.849892 \\
0.796042 & 0.197801 & 0.233489
\end{array}
\right],
\]
the SVD of ${\bf M} = {\bf U} \mbox{\boldmath$\Lambda$\unboldmath} {\bf V}^\top$ is given as
\[
{\bf U} \approx
\left[
\begin{array}{ccc}
-0.483076 & -0.175226 & -0.857865 \\
-0.768129 & -0.385453 & 0.511276 \\
-0.420256 & 0.905937 & 0.0516068
\end{array}
\right],
\]
\[
\mbox{\boldmath$\Lambda$\unboldmath}\approx \left[
\begin{array}{ccc}
1.43557 & 0. & 0. \\
0. & 0.66535 & 0. \\
0. & 0. & 0.0910448
\end{array}
\right],
\]
\[
{\bf V}\approx \left[
\begin{array}{ccc}
-0.406999 & 0.913369 & -0.0104708 \\
-0.616441 & -0.28311 & -0.734745 \\
-0.674057 & -0.292586 & 0.678263
\end{array}
\right].
\]
We now use the optimization problem:
\[\inf_{{\bf U}, {\bf V}\in\mathfrak{SO}(3)} \|{\bf U}^\top{\bf M}{\bf V}\|_1\]
to find the
SVD of ${\bf M}$. A numerical solution, produced by the heuristic global
optimization, is given as
\[
\hat{{\bf U}}\approx\left[
\begin{array}{ccc}
0.483076 & -0.857865 & -0.175227 \\
0.768129 & 0.511277 & -0.385453 \\
0.420256 & 0.0516062 & 0.905937
\end{array}
\right],
\]
\[
\hat{{\bf U}}^\top{\bf M}\hat{{\bf V}} \approx \left[
\begin{array}{ccc}
9.16351\times 10^{-10} & -8.52198\times 10^{-9} & {\bf 1.43557} \\
-4.58791\times 10^{-7} & {\bf -9.10448\times 10^{-2}} & 4.36886\times 10^{-9} \\
{\bf 6.6535\times 10^{-1}} & 8.13893\times 10^{-8} & 2.67077\times 10^{-10} \\
\end{array}
\right],
\]
\[
\hat{{\bf V}}\approx\left[
\begin{array}{ccc}
0.913369 & 0.010471 & 0.406999 \\
-0.28311 & 0.734745 & 0.616441 \\
-0.292585 & -0.678263 & 0.674057 \\
\end{array}
\right].
\]
Note that $\hat{{\bf U}}$, $\hat{{\bf V}}$, $\hat{{\bf U}}^\top{\bf M}\hat{{\bf V}}$ are permuted
approximations of ${\bf U}$, ${\bf V}$, $\mbox{\boldmath$\Lambda$\unboldmath}$ modulo sign, respectively.
\end{example}
\begin{example}[Compute QR of a $3\times 3$ matrix] We use the same ${\bf M}$ as in
Example~\ref{exp:svd}. QR decomposition of ${\bf M}$ is given by ${\bf M} = {\bf Q}{\bf D}{\bf R}$ where
\[{\bf Q}\approx
\left[
\begin{array}{ccc}
-0.20946 & -0.541716 & -0.814046 \\
-0.253927 & -0.773817 & 0.580283 \\
-0.944271 & 0.328254 & 0.0245274 \\
\end{array}
\right],
\]
\[{\bf D}\approx\left[
\begin{array}{ccc}
-0.843023 & 0. & 0. \\
0. & -0.771339 & 0. \\
0. & 0. & 0.133735 \\
\end{array}
\right],\]
\[{\bf R}\approx\left[
\begin{array}{ccc}
1. & 0.566548 & 0.628984 \\
0. & 1. & 1.0683 \\
0. & 0. & 1. \\
\end{array}
\right].\]
We use optimization of the form
\[\inf_{{\bf Q}\in\mathfrak{SO}(3),{\bf R}\in\mathfrak{UUT}(3)}
\|{\bf Q}^\top{\bf M}{\bf R}^{-1}\|_1\]
to find QR of ${\bf M}$. A numerical solution, produced
by the heuristic global optimization, is given as
\[\hat{{\bf Q}}\approx
\left[
\begin{array}{ccc}
0.814046 & 0.541716 & 0.20946 \\
-0.580283 & 0.773817 & 0.253927 \\
-0.0245274 & -0.328254 & 0.944271 \\
\end{array}
\right],
\]
\[\hat{{\bf Q}}^\top{\bf M}\hat{{\bf R}}^{-1}\approx
\left[
\begin{array}{ccc}
4.65594\times 10^{-9} & -7.03107\times 10^{-9} & \bf -1.33735\times 10^{-1} \\
2.32509\times 10^{-8} & \bf 7.71339\times 10^{-1} & 6.05164\times 10^{-9} \\
\bf 8.43023\times 10^{-1} & 5.62326\times 10^{-3} & 3.42317\times 10^{-9} \\
\end{array}
\right],
\]
\[\hat{{\bf R}}\approx
\left[
\begin{array}{ccc}
1 & 0.559878 & 0.621858 \\
0 & 1 & 1.0683 \\
0 & 0 & 1 \\
\end{array}
\right].
\]
Note that $\hat{{\bf Q}}$, $\hat{{\bf Q}}^\top{\bf M}\hat{{\bf R}}^{-1}$, and $\hat{{\bf R}}$ are
permuted approximations of ${\bf Q}$, ${\bf D}$, ${\bf R}$ modulo sign,
respectively.
Note although there is a significant difference between ${\bf R}$ and $\hat{{\bf R}}$ as
$\|{\bf R}-\hat{{\bf R}}\|_F\approx0.00976074$, the decomposition is still good
approximation as we have \[\|\hat{{\bf Q}}\hat{{\bf D}}\hat{{\bf R}} -
{\bf M}\|_F\approx3.01503\times 10^{-16}\text{.}\]
\end{example}
\begin{example}[Compute matrix equivalence decomposition of a $3\times 3$
matrix] We use the same ${\bf M}$ as in Example~\ref{exp:svd}. Matrix equivalence
decomposition of ${\bf M}$ is not unique. Anyway the optimal core modulo sign
and permutation would be \[{\bf D}\approx
\left[
\begin{array}{ccc}
0.44304 & 0. & 0. \\
0. & 0.44304 & 0. \\
0. & 0. & 0.44304
\end{array}
\right].
\]
We use the optimization \[\inf_{{\bf A}\in\mathfrak{SL}(3),{\bf B}\in\mathfrak{SL}(3)}
\|{\bf A}^{-1}{\bf M}{\bf B}^{-1}\|_1\] to find the matrix equivalence decomposition of ${\bf M}$. A
numerical solution produced by the heuristic global optimization is given as
\[\hat{{\bf A}}^{-1}{\bf M} \hat{{\bf B}}^{-1}\approx
\left[
\begin{array}{ccc}
2.00925\times 10^{-9} & {\bf 4.42891\times 10^{-1}} & 2.77141\times 10^{-9} \\
1.18223\times 10^{-9} & 2.10486\times 10^{-8} & {\bf 4.43117\times 10^{-1}} \\
{\bf 4.43112\times 10^{-1}} & -7.24341\times 10^{-9} & 1.79685\times 10^{-10} \\
\end{array}
\right].
\]
\end{example}
\begin{example}[Compute LU of a $3\times 3$ matrix] We use the same ${\bf M}$ as in
Example~\ref{exp:svd}. The LU decomposition of ${\bf M}$ without pivoting is given by
\[{\bf L}\approx
\left[
\begin{array}{ccc}
1 & 0 & 0 \\
1.21229 & 1 & 0 \\
4.50812 & -2.36585\times 10^1 & 1 \\
\end{array}
\right],
\]
\[{\bf U}\approx
\left[
\begin{array}{ccc}
1.7658\times 10^{-1} & 5.17888\times 10^{-1} & 4.48587\times 10^{-1} \\
2.77556\times 10^{-17} & 9.03226\times 10^{-2} & 3.06073\times 10^{-1} \\
1.11022\times 10^{-16} & 0. & 5.45245 \\
\end{array}
\right]
\textrm{.}\]
We use the optimization \[\inf_{{\bf L}\in\mathfrak{LUT}(3)}
\|\Delta \odot ({\bf L}^{-1}{\bf M})\|_1\] to find the LU of ${\bf M}$, where
$\Delta_{ij}={\mathbb I}_{i>j}$. A numerical solution produced
by the heuristic global optimization is given as
\[\hat{{\bf L}}\approx
\left[
\begin{array}{ccc}
1 & 0 & 0 \\
1.21229 & 1 & 0 \\
4.50812 & -2.36585\times 10^1 & 1 \\
\end{array}
\right]
\textrm{,}\]
\[\hat{{\bf U}}\approx
\left[
\begin{array}{ccc}
1.7658\times 10^{-1} & 5.17888\times 10^{-1} & 4.48587\times 10^{-1} \\
-1.64336\times 10^{-17} & 9.03226\times 10^{-2} & 3.06073\times 10^{-1} \\
2.22045\times 10^{-16} & 4.44089\times 10^{-16} & 5.45245 \\
\end{array}
\right]
\textrm{.}\]
\end{example}
\begin{example}[Compute Cholesky decomposition of a $3\times 3$ matrix] We use
the ${\bf M}^\top{\bf M}$ as input with ${\bf M}$ from Example~\ref{exp:svd}. The Cholesky
decomposition of ${\bf M}^\top{\bf M}$ is given by ${\bf M}^\top{\bf M} = {\bf U}^*{\bf U}$ where
\[{\bf U}^*\approx
\left[
\begin{array}{ccc}
0.843023 & 0. & 0. \\
0.477613 & 0.771339 & 0. \\
0.530248 & 0.824024 & 0.133735 \\
\end{array}
\right]
\textrm{.}
\]
We use the optimization \[\inf_{{\bf L}\in\mathfrak{LUT}(3)}
\|\Delta \odot ({\bf L}^{-1}{\bf M})\|_1\] to find the LU of ${\bf M}$, where
$\Delta_{ij}={\mathbb I}_{i>j}$. A numerical solution
is given as
\[\hat{{\bf L}}\boldsymbol\Lambda\approx
\left[
\begin{array}{ccc}
0.843023 & 0. & 0. \\
0.477613 & 0.771339 & 0. \\
0.530248 & 0.824024 & 0.133735 \\
\end{array}
\right].
\]
Here $\boldsymbol\Lambda$ is a diagonal matrix with square root of
diagonals of $\hat{{\bf L}}^{-1}{\bf M}$ as its diagonals.
\end{example}
\begin{example}[Compute Schur decomposition of a $3\times 3$ matrix] We use the
same ${\bf M}$ as in Example~\ref{exp:svd}. The Schur decomposition of ${\bf M} = {\bf Q}{\bf U}{\bf Q}^{-1}$
is given by
\[\Re{\bf Q}\approx
\left[
\begin{array}{ccc}
-4.4809\times 10^{-1} & 2.13496\times 10^{-2} & -3.11518\times 10^{-1} \\
-6.81147\times 10^{-1} & -5.00762\times 10^{-1} & 1.85125\times 10^{-3} \\
-4.25133\times 10^{-1} & 7.79816\times 10^{-1} & 3.25373\times 10^{-1} \\
\end{array}
\right]
\textrm{,}\]
\[\Im{\bf Q}\approx
\left[
\begin{array}{ccc}
-1.91558\times 10^{-1} & 2.76329\times 10^{-1} & -7.67245\times 10^{-1} \\
-2.91189\times 10^{-1} & -2.4227\times 10^{-2} & 4.47096\times 10^{-1} \\
-1.81744\times 10^{-1} & -2.52434\times 10^{-1} & 9.23392\times 10^{-2} \\
\end{array}
\right]
\textrm{,}\]
\[\Re{\bf U}\approx
\left[
\begin{array}{ccc}
1.38943 & -2.5768\times 10^{-1} & -1.15903\times 10^{-1} \\
0. & -1.30605\times 10^{-1} & -7.89372\times 10^{-2} \\
0. & 0. & -1.30605\times 10^{-1}
\end{array}
\right]
\textrm{,}\]
\[\Im{\bf U}\approx
\left[
\begin{array}{ccc}
-1.38778\times 10^{-16} & 2.11604\times 10^{-1} & 3.88987\times 10^{-2} \\
0. & 2.13379\times 10^{-1} & -5.69018\times 10^{-1} \\
0. & 0. & -2.13379\times 10^{-1} \\
\end{array}
\right]
\textrm{,}\]
We can use the optimization \[\inf_{{\bf Q}\in\mathfrak{U}(3)}
\|\Delta\odot({\bf Q}^{-1}{\bf M}{\bf Q})\|_1\] to find the Schur decomposition of ${\bf M}$, where
$\Delta_{ij}={\mathbb I}_{i>j}$. A numerical solution is given as
\[\Re\hat{{\bf Q}}\approx
\left[
\begin{array}{ccc}
-0.159122 & 0.456745 & -0.924103 \\
-0.697698 & 0.596808 & 0.46622 \\
0.777452 & 0.665151 & 0.227028
\end{array}
\right]
\textrm{,}\]
\[\Im\hat{{\bf Q}}\approx
\left[
\begin{array}{ccc}
-0.288006 & -0.0686255 & 0.0156732 \\
0.161731 & 0.0500731 & 0.177932 \\
0.0861939 & 0.00219548 & -0.301602
\end{array}
\right]
\textrm{,}\]
\[\Re\hat{{\bf U}}\approx
\left[
\begin{array}{ccc}
\bf -1.30605\times 10^{-1} & -3.71633\times 10^{-1} & -7.07291\times 10^{-1} \\
-2.09852\times 10^{-9} & \bf 1.38943 & -1.09975\times 10^{-1} \\
-1.39859\times 10^{-8} & 9.97677\times 10^{-9} & \bf -1.30604\times 10^{-1}
\end{array}
\right]
\textrm{,}\]
\[\Im\hat{{\bf U}}\approx
\left[
\begin{array}{ccc}
\bf -2.13379\times 10^{-1} & -1.95892\times 10^{-2} & 2.66849\times 10^{-2} \\
2.82526\times 10^{-9} & \bf 2.10889\times 10^{-9} & -1.05424\times 10^{-1} \\
9.09511\times 10^{-9} & 8.28696\times 10^{-9} & \bf 2.13379\times 10^{-1} \\
\end{array}
\right]
\textrm{.}\]
We note that $\hat{{\bf U}}$ is permuted approximation of ${\bf U}$ modulo sign.
\end{example}
\subsection{GOO Inducing Tensor Decomposition}
We empirically illustrate several examples of GOO inducing tensor decomposition.
Due to the large amount of computation required by the Nelder-Meld algorithm, here
we only give examples involving small-size tensors.
\begin{example}[Non-uniqueness of strong-orthogonal decomposition]
\label{ex:non-unique-strong-orthogonal-decomp}
Consider tensor ${\mathcal A}$ as in Example 3.3 of \cite{Kolda:2001:OTD:587708.587830},
which we reproduce below:
\begin{align}\label{eq:kolda-example-3.3}
{\mathcal A}=\sigma_1 {\bf a}\otimes {\bf b}\otimes {\bf b}+\sigma_2 {\bf b}\otimes {\bf b}\otimes {\bf b}+\sigma_3
{\bf a}\otimes {\bf a}\otimes {\bf a}\text{,}
\end{align}
where $\sigma_1>\sigma_2>\sigma_3, {\bf a}\perp {\bf b}, \|{\bf a}\|=\|{\bf b}\|=1$.
We note that Formula~(\ref{eq:kolda-example-3.3}) is already a strong orthogonal
decomposition of ${\mathcal A}$. Nevertheless, an alternative strong orthogonal
decomposition is given therein as
\begin{align}\label{eq:kolda-example-3.3-alt}
{\mathcal A}=\hat{\sigma}_1
\hat{{\bf a}}\otimes{\bf b}\otimes{\bf b}+\hat{\sigma}_2\hat{{\bf a}}\otimes{\bf a}\otimes{\bf a}+\hat{\sigma}_3\hat{{\bf b}}\otimes{\bf a}\otimes{\bf a}
\text{,}\end{align}
where
\[
\hat{\sigma}_1 = \sqrt{\sigma_1^2+\sigma_2^2},\qquad
\hat{\sigma}_2=\frac{\sigma_1\sigma_3}{\hat{\sigma}_2},\qquad\hat{\sigma}_3=\frac{\sigma_2\sigma_3}{\hat{\sigma}_1}\text{,}
\]
\[
\hat{{\bf a}} = \frac{\sigma_1 {\bf a} + \sigma_2 {\bf b}}{\hat{\sigma}_1},\text{ and }\;
\hat{{\bf b}} = \frac{\sigma_2 {\bf a} - \sigma_1 {\bf b}}{\hat{\sigma}_1}\text{.}
\]
Without loss of generality, we let $\sigma_1 =3, \sigma_2 = 2, \sigma_3 = 1,
{\bf a}=[1,0]^\top$, and ${\bf b}=[0,1]^\top$. Then
\[
\operatorname{vec}{({\mathcal A})} = [1,0,0,0,0,0,3,2]^\top\text{,}
\]
\[
\hat{\sigma}_1\approx3.60555,\qquad\hat{\sigma}_2\approx0.832050,\qquad\hat{\sigma}_3\approx
0.5547002.
\]
In framework of GOO, we can induce a strong orthogonal decomposition of tensor
${\mathcal A}$ by the following optimization:
\[
\inf_{{\bf G}_1\in \mathfrak{SO}(2), {\bf G}_2\in \mathfrak{SO}(2), {\bf G}_3\in \mathfrak{SO}(2)} \|{\mathcal A} \times_1
{\bf G}_1\times_2 {\bf G}_2\times_3 {\bf G}_3\|_1 \text{.}
\]
One numerical solution of core tensor is:
\[\operatorname{vec}({\mathcal A}\times_1\hat{{\bf G}}_1\times_2\hat{{\bf G}}_2\times_3\hat{{\bf G}}_3) \approx
\left[
\begin{array}{c}
\bf 3.60555 \\
1.24142\times 10^{-8} \\
1.92366\times 10^{-10} \\
-2.91798\times 10^{-9} \\
-1.80012\times 10^{-8} \\
-6.43786\times 10^{-10} \\
\bf 8.3205\times 10^{-1} \\
\bf -5.547\times 10^{-1} \\
\end{array}
\right]
\text{.}
\]
Note that the large nonzero values (in bold) are approximations of $\hat{\sigma}_1$,
$\hat{\sigma}_2$, and $\hat{\sigma}_3$, modulo sign.
\end{example}
\begin{example}[The Special Linear Group finds Sparser Core in Tensor
Decomposition] ${\mathcal A}$ is given as in
Example~\ref{ex:non-unique-strong-orthogonal-decomp}. We
can induce a ``sparser'' decomposition of tensor with the following GOO:
\[
\inf_{{\bf G}_1\in \mathfrak{SL}(2), {\bf G}_2\in \mathfrak{SL}(2), {\bf G}_3\in \mathfrak{SL}(2)} \|{\mathcal A} \times_1
{\bf G}_1\times_2 {\bf G}_2\times_3 {\bf G}_3\|_1 \text{.}
\]
One numerical solution of core tensor is:
\[\operatorname{vec}({\mathcal A}\times_1\hat{{\bf G}}_1\times_2\hat{{\bf G}}_2\times_3\hat{{\bf G}}_3) \approx
\left[
\begin{array}{c}
\bf 1.41421 \\
-5.7745\times 10^{-9} \\
-9.00468\times 10^{-9} \\
1.49267\times 10^{-9} \\
-4.79358\times 10^{-9} \\
-8.18472\times 10^{-9} \\
-2.80551\times 10^{-9} \\
\bf 1.41421 \\
\end{array}
\right]
\text{.}
\]
Note that there are only two significant nonzero values (in bold), in contrast to
three in the strong orthogonal decomposition. Since
${\mathcal A}\times_1\hat{{\bf G}}_1\times_2\hat{{\bf G}}_2\times_3\hat{{\bf G}}_3$ is superdiagonal, it
is the ``sparsest'' core tensor under any Tucker decompositions.
\end{example}
\begin{example}[A tensor that does not have Superdiagonal Form but is also of
Lowest Rank under any Tucker
Decomposition]
We give a numerical solution to Example~\ref{example:non-superdiagonalizable}
where
\[{\mathcal T}\in{\mathbb F}^{2\times 2\times
2\times 2},\; \operatorname{vec}{({\mathcal T})}=[{{{{1,0}, {0,1}}, {{0,0}, {0,0}}}, {{{0,0},
{0,0}}, {{1,0}, {0,1}}}}]^\top.\]
The solution to
\[
\inf_{{\bf G}_1\in \mathfrak{SL}(2), {\bf G}_2\in \mathfrak{SL}(2), {\bf G}_3\in \mathfrak{SL}(2),{\bf G}_4\in\mathfrak{SL}(2)} \|{\mathcal A}
\times_1 {\bf G}_1\times_2 {\bf G}_2\times_3 {\bf G}_3\times_4{\bf G}_4\|_1
\]
is
\[\operatorname{vec}({\mathcal T}\times_1\hat{{\bf G}}_1\times_2\hat{{\bf G}}_2\times_3\hat{{\bf G}}_3\times_4\hat{{\bf G}}_4)
\approx \left[
\begin{array}{c}
\bf 1. \\
5.85196\times 10^{-9} \\
2.40735\times 10^{-10} \\
\bf -1. \\
-2.39741\times 10^{-9} \\
-1.40295\times 10^{-17} \\
-5.77138\times 10^{-19} \\
2.3974\times 10^{-9} \\
4.74159\times 10^{-9} \\
2.77475\times 10^{-17} \\
1.14146\times 10^{-18} \\
-4.74158\times 10^{-9} \\
\bf 9.99999\times 10^{-1} \\
5.85194\times 10^{-9} \\
2.40734\times 10^{-10} \\
\bf -9.99998\times 10^{-1} \\
\end{array}
\right]\text{.}
\]
Hence there are four significant nonzero values even under GOO {w.r.t.\ } the special
linear group.
\end{example}
\subsection{Normalization of point cloud {w.r.t.\ } special linear group}
Here we apply the GOO defined in Section~\ref{sec:normal} to a publicly
available set of 2D point cloud data
\href{http://vision.lems.brown.edu/content/available-software-and-databases/#Datasets-Shape}{here}.
As the optimization variable only consists of a small matrix
${\bf M}\in{\mathbb F}^{2\times 2}$, we are able to deal with large point clouds consisting
of more than thousands of points.
The detailed steps are as follows:
\begin{algorithm}\label{alg:special-linear}
\begin{enumerate}
\item[Step 1] Normalize the point cloud corresponding to ${\bf M}$ {w.r.t.\ } special linear
group as \[\hat{{\bf M}} = \mathop{\rm arginf}_{{\bf G}\in\mathfrak{SL}(n)} \|{\bf M} {\bf G}\|_\infty\text{.}\]
\item[Step 2] (Optional) Let $\hat{{\bf M}}_x$ and $\hat{{\bf M}}_y$ be two columns of
$\hat{{\bf M}}$. We can use a simple criterion to select one from four possible forms of normalized
point clouds: $[\hat{{\bf M}}_x, \hat{{\bf M}}_y]$, $[-\hat{{\bf M}}_y, \hat{{\bf M}}_x]$,
$[-\hat{{\bf M}}_x, -\hat{{\bf M}}_y]$, and $[\hat{{\bf M}}_y, -\hat{{\bf M}}_x]$ to further
eliminate ambiguity.
An example is to pick the matrix $\hat{\hat{{\bf M}}}$ that minimizes $\phi({\bf X}) = \|g({\bf X})\|_F$
where $g(x) = \max(0, x)$. $\hat{\hat{{\bf M}}}$ is called a canonical form of ${\bf M}$
in this section.
\end{enumerate}
\end{algorithm}
\begin{remark}
Step~2 in Algorithm~\ref{alg:special-linear} is found to be useful in
eliminating the ambiguity in orientation in some circumstances. However, even if
Step~2 fails or is skipped, one can still use $\hat{{\bf M}}$ as ``canonical''
form and enumerate the few number of possible orientations. The result of
Algorithm~\ref{alg:special-linear} without Step~2 is shown in
Figure~\ref{fig:special_linear} and Figure~\ref{fig:special_linear_3d}.
\end{remark}
In Figure~\ref{fig:normal_form} we perform a side-by-side comparison of results
of several normalization techniques.
The point clouds in the row marked with ``Distorted'' are produced by applying
random shearing, mirroring, squeezing and rotation to the same point cloud. The point
clouds in the row marked with ``PCA'' are results of applying PCA to the
matrices corresponding to the distorted point clouds in the ``Distorted'' row. It
can be seen that PCA can remove the degree of freedom corresponding to rotation in the input data, but fails to
remove effect of squeezing and shearing. The row marked with ``GOO\_SO'' is
produced by using GOO with orthogonal group:
\[
\inf_{{\bf G}\in\mathfrak{O}(n)} \|{\bf M} {\bf G}\|_\infty\text{.}\]
We can see that effect of
rotation is removed but effects of squeezing and shearing remain. The row
marked with ``GOO\_SL'' is the canonical forms of matrices corresponding
to point clouds derived by Algorithm~\ref{alg:special-linear}. We can see that
the normalized point clouds are approximately the same, and effects of rotation, squeezing and shearing are
almost completely eliminated.
\begin{figure}
\subfigtopskip = 0pt
\begin{center}
\centering
\subfigure{\includegraphics[height=64mm,
width=130mm]{normalizing_1.pdf}}
\subfigure{\includegraphics[height=64mm,
width=130mm]{normalizing_5.pdf}}
\subfigure{\includegraphics[height=64mm,
width=130mm]{normalizing_6.pdf}}
\end{center}
\caption{This figures show the results of normalizing distorted point clouds
by different methods.
The rows marked with ``Distorted'' consist of distorted point clouds used as
input to various normalization methods.
The rows marked with ``PCA'' contain results of normalization by principal
component analysis. It can be seen that the effects of
rotational distortion have been partially eliminated, but results of shearing
and squeezing remain.
The rows marked with ``GOO\_SO'' contain results of normalization using
special orthogonal group in GOO. It can be seen that the effects of
rotational distortion have been partially eliminated, but results of shearing
and squeezing remain.
The rows marked with ``GOO\_SL'' contain results of
normalization using Algorithm~\ref{alg:special-linear}, where it can be seen that the algorithm
can produce approximately the same point clouds after eliminating distortions
like shearing, squeezing and rotation.}
\label{fig:normal_form}
\end{figure}
In Figure~\ref{fig:sampling_rate} we study the impact of number of points on
the canonical form found by the GOO.
We can see that though the number of points in the canonical form vary between
180 and 260, the canonical form is nearly the same, module different
orientations. In this case although Step~2 in
Algorithm~\ref{alg:special-linear} cannot completely eliminate the ambiguity of
four possible orientations of point clouds, we can simply remove this ambiguity by
enumerating all four possible orientations when doing comparison.
This property means that when comparing shape of two point clouds, it is not necessary to
require two point clouds to have exactly the same number of points when we are
comparing based on the canonical forms.
\begin{figure}
\subfigtopskip = 0pt
\begin{center}
\centering
\subfigure{\includegraphics[height=48mm,
width=130mm]{sampling_rate_1.pdf}}
\subfigure{\includegraphics[height=48mm,
width=130mm]{sampling_rate_5.pdf}}
\subfigure{\includegraphics[height=48mm,
width=130mm]{sampling_rate_6.pdf}}
\end{center}
\caption{The point clouds in rows marked with ``Distorted'' are the results
of applying distortion generated by random special linear matrix to original
point clouds.
The point clouds in rows marked with ``GOO\_SL'' are after normalization by
special linear group. From left to right, the sparser point clouds are
generated by sampling from the rightmost densest point cloud respectively.
It can be seen that for all three examples, though density varies, the
shape of the normalized point clouds remains stable, modulo four possible
orientations.}
\label{fig:sampling_rate}
\end{figure}
\section{Related work}
In this section we discuss the related work not yet covered in the previous sections.
An early example of GOO is a so-called quadratic assignment problem \cite{sahni1976p} where the
following optimization problem is studied:
\[
\inf_{{\bf X}\in\Pi_n} \operatorname{tr}({\bf W} {\bf X} {\bf D} {\bf X}^\top) \text{,}
\]
where $\Pi_n$ is the permutation matrix group. Due to the combinatorial nature
of $\Pi_n$, QAP is NP-hard. In contrast, we mainly work on
non-combinatorial matrix groups in this paper.
In \cite{zhang2012tilt}, a non-linear GOO is used to find texture
invariant to rotation for 2D point cloud ${\bf M}\in{\mathbb F}^{n\times 2}$:
\[
\inf_{{\bf G}\in\mathfrak{O}} \|\operatorname{Rasterize}(\operatorname{Poly}({\bf M} {\bf G}))\|_* \text{.}
\]
As $\mathfrak{O}$ is a unit group, the optimization is well defined and the
induced matrix decomposition is found to be useful as a rotation-invariant
representation for texture. The same paper also considers finding
homography-invariant representation for texture for 2D point cloud
${\bf M}\in{\mathbb F}^{n\times 2}$:
\[
\inf_{{\bf G}\in\mathfrak{H},\; \mu(\operatorname{Poly}(\lambda {\bf M} {\bf G}))=\const} \|\operatorname{Rasterize}[\operatorname{Poly}(\lambda
{\bf M} {\bf G})]\|_*
\text{.}
\]
Note that here a coefficient $\lambda$ is intentionally
added to ensure $\mu$ measure of the point cloud be preserved {w.r.t.\ } the action
of ${\bf G}$.
In \cite{hu2013fast} the following formulation is used to get the Ky-Fan $k$-norm
\cite{horn1991topics} of a matrix ${\bf M}\in{\mathbb F}^{m\times n}$ when $m\ge k$ and
$n\ge k$:
\[
\sup_{{\bf G}_1\in{\mathbb F}^{m\times k}, {\bf G}_1^\top{\bf G}_1 = {\bf I}_k,{\bf G}_2\in{\mathbb F}^{n\times k},
{\bf G}_2^\top{\bf G}_2 = {\bf I}_k} \operatorname{tr}({\bf G}_1^\top {\bf M} {\bf G}_2)\text{.}
\]
Note that the above optimization is not a GOO when $k^2\neq m n$ as in that case
${\bf G}_2^\top \otimes {\bf G}_1^\top \in {\mathbb F}^{k^2\times m n}$ does not form a group.
\section{Conclusion}
\label{sec:conclusion}
In this paper, we have studied an optimization problem over the group orbit
generated by action of group ${\mathfrak G}$ and referred to it as the \emph{Group Orbit
Optimization} (GOO).
We have shown that SVD/QR/LU/Cholesky decomposition can be reformulated under
the GOO framework as in Theorem~\ref{theorem:decompostion-as-optimization}.
Moreover, we have used GOO to induce tensor decomposition in
Theorem~\ref{thm:unfolded-diagonal}. The unified framework of GOO for matrix
decomposition and tensor decomposition allows us to bridge them. In
particular, we have presented Lemma~\ref{lem:lifting}, which relates the infimum of
the tensor-based GOO with the infimum of GOO of the matrix unfolded to the
tensor. Finally, we have applied GOO to point cloud data to demonstrate the use of data
normalization in shape matching when objects are represented as point clouds.
Our work has demonstrated that the unified framework of GOO for
data normalization is both of theoretical interests in providing a new perspective on
matrix and tensor decompositions, and of practical interests in modeling and
elimination of distortions present in real world data.
|
\section{Introduction}
Quantum computation promises to efficiently implement algorithms intractable on classical computers \cite{bib:NielsenChuang00}. Many physical realizations have been proposed, using different physical systems to represent qubits and their evolution. Linear optics quantum computing (LOQC) \cite{bib:KokLovett}, originally proposed by Knill, Laflamme \& Milburn (KLM) \cite{bib:KLM01}, is one of the most promising proposals, owing to the inherently long decoherence times of photons, and the relative ease with which to prepare, evolve, and measure them.
Boson-sampling, proposed by Aaronson \& Arkhipov \cite{bib:AaronsonArkhipov10}, is a simple, but non-universal approach to implementing a specific quantum algorithm using quantum optics, which is strongly believed to be intractable on a classical computer. Whilst not universal for quantum computing, it has attracted much attention as it is significantly simpler than KLM, requiring only single-photon state preparation, passive linear optics, and photo-detection. Several elementary proof-of-principle experimental demonstrations have been performed \cite{bib:Broome20122012, bib:Crespi3, bib:Tillmann4, bib:Spring2}. For an elementary introduction to boson-sampling, see Gard \emph{et al.} \cite{bib:GardBSintro}.
Recently, Motes, Gilchrist, Dowling \& Rohde (MGDR) \cite{bib:MotesLoop} presented a scheme for efficient, universal boson-sampling using a fiber-loop architecture with time-bin encoding, which has frugal experimental requirements, and is highly scalable. However, the scheme was only shown to be universal for boson-sampling, and does not allow universal quantum computation.
Here we show that the same architecture can be made universal for quantum computation with only a straightforward modification to the classical processing and control. The scheme requires only two embedded fiber-loops, three dynamically controlled beamsplitters, a single push-button photon source, and a single time-resolved photo-detector. These experimental requirements are fixed, irrespective of the size of the computation, limited only by fiber lengths (which scales polynomially with the number of qubits) and their respective loss rates. The scheme has only a single point of interference, making alignment of optical elements significantly more favorable than other proposals, such as bulk optics or waveguide implementations, which require simultaneously aligning $\mathrm{poly}(n)$ points of interference, $n$ being the number of optical modes.
The viability of time-bin encoding using a loop architecture has been demonstrated with recent quantum walk experiments by Schreiber \emph{et al.} \cite{bib:Schreiber10, bib:Schreiber12}. Humphreys \emph{et al.} \cite{bib:HumphreysSingleSpatialMode} presented a similar scheme for LOQC in a single spatial mode, based on using the polarization degree of freedom to `address' individual time-bins, and implement arbitrary LOQC transformations, which they showed was sufficient for universal LOQC. Additionally, they performed an elementary experimental demonstration of a controlled-phase (CZ) gate (a maximally entangling two-qubit gate, which is universal when combined with single-qubit operations) using their approach. Their scheme relied on dynamically controlled polarization rotations to perform time-bin addressing, and polarization-dependent temporal displacement operations (implemented via birefringent crystals) to interfere time-bins. The MGDR scheme has different experimental requirements, requiring only dynamically controlled beamsplitters to implement arbitrary linear optics interferometers.
\section{Boson-sampling}
The full fiber-loop architecture for boson-sampling, as originally presented by MGDR, is shown in Fig. \ref{fig:full_arch}. A push-button single-photon source prepares a pulse-train of single-photon and vacuum states, where each time-bin is separated by time $\tau$. $\tau$ must be sufficiently large compared to the photons' wavepackets that they remain temporally orthogonal. The pulse-train enters a system of two embedded fiber-loops, controlled by three (classically) dynamically controlled beamsplitters. MGDR showed that this configuration, with appropriate control over the dynamically controlled beamsplitters, can implement an arbitrary, passive linear optics network, implementing a transformation on the mode operators of the form,
\begin{equation} \label{eq:creation_U}
\hat{a}_i^\dag \to \sum_{j=1}^n U_{i,j} \hat{a}_j^\dag,
\end{equation}
where $\hat{a}^\dag_i$ is the photon creation operator on the $i$th mode (time-bin), for any \mbox{$n\times n$} unitary $U$.
\begin{figure}[!htb]
\includegraphics[width=0.9\columnwidth]{full_architecture}
\caption{The MGDR architecture for implementing arbitrary passive linear optics transformations on an $n$ time-bin-encoded pulse-train. The push-button source prepares a pulse train of single-photon and vacuum states, separated by time $\tau$, across $n$ time-bins. The pulse-train enters the outer loop, classically controlled via two on/off switches. The pulse-train then passes through the inner loop some number of times, as determined by the control of the central dynamically controlled beamsplitter. With \mbox{$\mathrm{poly}(n)$} round-trips of the outer loop, an arbitrary $n$-mode linear optics transformation may be implemented. The round-trip time of the inner loop is $\tau$, and the round-trip time of the outer loop is \mbox{$>n\tau$}. The first and last dynamically controlled beamsplitters need only be on/off (completely reflective or completely transmissive), whereas the central dynamically controlled beamsplitter must be able to implement arbitrary $SU(2)$ transformations.} \label{fig:full_arch}
\end{figure}
We begin by reviewing how this experimental configuration allows arbitrary transformations of the form of Eq. \ref{eq:creation_U} to be implemented. In Fig. \ref{fig:single_loop} we show the expansion for the inner loop as its equivalent spatially-encoded beamsplitter network with $n$ modes. In Fig. \ref{fig:two_loops} we show that with two passes through the inner loop, arbitrary pairwise beamsplitter transformations may be implemented in the case of \mbox{$n=3$} modes. These two passes through the inner loop are implemented via two round-trips through the outer loop.
Having established that arbitrary pairwise beamsplitter operations are possible, it follows from the Reck \emph{et al.} \cite{bib:Reck94} decomposition that an arbitrary 3-mode linear optics network may be implemented with multiple applications of pairwise interactions. With arbitrary 3-mode linear optics networks at our disposal in the MGDR architecture, in Fig. \ref{fig:induction} we show via an inductive argument that with \mbox{$\mathrm{poly}(n)$} passes through the outer loop, arbitrary pairwise transformations between the $n$ modes are possible, and thus arbitrary $n$-mode linear optics transformations may be implemented using the Reck \emph{et al.} decomposition with $\mathrm{poly}(n)$ round-trips of the outer loop.
\begin{figure}[!htb]
\includegraphics[width=0.9\columnwidth]{single_loop}
\caption{Mapping between the inner fiber-loop (a) and its equivalent spatially-encoded transformation (b). $n$ passes through the inner loop is equivalent to a diagonal array of spatially-encoded beamsplitters. On its own, this configuration is not sufficient for arbitrary linear optics transformations as per Eq. \ref{eq:creation_U}.} \label{fig:single_loop}
\end{figure}
\begin{figure}[!htb]
\includegraphics[width=0.8\columnwidth]{universality}
\caption{(Color online) For \mbox{$n=3$} modes, with two consecutive applications of the inner loop (i.e two round-trips of the outer loop) and appropriate choices of beamsplitter reflectivities, arbitrary pairwise beamsplitter operations may be implemented: (a) between modes 1 and 2; (b) between modes 1 and 3; (c) between modes 2 and 3. It follows from the Reck \emph{et al.} decomposition that \mbox{$O(n^2)$} pairwise beamsplitter operations enables arbitrary 3-mode linear optics transformations.} \label{fig:two_loops}
\end{figure}
\begin{figure}[!htb]
\includegraphics[width=0.7\columnwidth]{induction}
\caption{Inductive argument that the MGDR architecture enables arbitrary $n$-mode linear optics transformations. We have established in Fig. \ref{fig:two_loops} that arbitrary 3-mode transformations are possible, which includes all permutations. To enable arbitrary $n$-mode transformations, we must additionally allow arbitrary beamsplitter operations between any of the first $n$ modes and the \mbox{$(n+1)$}th mode. We let $P$ be a permutation that maps any of the first $n$ modes to the $n$th mode, apply the desired beamsplitter interaction between the $n$th mode and the \mbox{$(n+1)$}th mode, and then apply the inverse permutation. Thus, if arbitrary $n$-mode transformations are possible, it follows inductively that arbitrary \mbox{($n+1$)}-mode transformations are possible, thus enabling universal linear optics transformations for arbitrary $n$.} \label{fig:induction}
\end{figure}
In the original proposal for boson-sampling presented by Aaronson \& Arkhipov, we require an input state comprising a \emph{specific} configuration of single-photon and vacuum states. With non-deterministic photon sources -- as is the case with most present-day technologies -- this might be achieved by multiplexing many heralded sources, as was considered by Motes \emph{et al.} \cite{bib:MotesSPDC}. However, recently Lund \emph{et al.} \cite{bib:RandomBS} showed that when using spontaneous parametric down-conversion (SPDC) as heralded sources, multiplexing is not necessary to implement a computational problem equivalent to boson-sampling. Rather, as long as the heralding signature is known, the exact heralding configuration does not affect the computational complexity. This is referred to as `randomized' boson-sampling. The MGDR architecture is perfectly suited to randomized boson-sampling, where we employ an SPDC with high repetition rate as the photon source preparing the pulse-train.
\section{Universal quantum computing}
We have established that the fiber-loop architecture may implement arbitrary passive linear optics transformations of the form of Eq. \ref{eq:creation_U}. Whilst this is sufficient for arbitrary boson-sampling on $n$-modes, which only requires passive linear optics, it is not sufficient for universal LOQC, which requires ancillary states, post-selection, and feed-forward. To demonstrate the universality of this scheme for LOQC, we will show that it can be made equivalent to various universal approaches for LOQC, with appropriate adjustments.
We will first show how the fiber-loop architecture can be mapped to the original KLM scheme for universal LOQC. Then we will show mappings to two subsequent variations employing cluster states -- by Nielsen \cite{bib:Nielsen04}, and Browne \& Rudolph \cite{bib:BrowneRudolph05} -- that substantially reduce experimental resource requirements.
In all of these variations, the necessary modification to the MGDR architecture is to introduce `ancilla injection' and `ancilla extraction', whereby we dynamically introduce ancillary photons into the photon pulse-train, or extract a subset of the photons in the pulse-train for measurement. Injecting and measuring these ancillary photons is the crucial ingredient necessary to make the scheme universal for LOQC, and does not require any experimental overhead compared to the original MGDR architecture, since the required switching to perform ancilla injection and extraction is already a part of the architecture. The only overhead compared to the original MGDR architecture is more complicated classical control of the dynamic switches, which requires feed-forward from the measurement results obtained during the ancilla extraction stage.
\subsection{Knill, Laflamme \& Milburn} \label{sec:KLM}
It was long believed that universal optical quantum computation would only be possible with strong Kerr non-linearities. In a seminal result, KLM showed that just linear optics transformations, photo-detection, and feed-forward, is sufficient for efficient universal quantum computation. The key observation was that, in conjunction with ancillary states (vacuum and single-photon Fock states), post-selection enables an effective non-linearity to be implemented. Specifically, a non-deterministic non-linear sign-shift (NS) gate may be implemented, which simulates a strong Kerr non-linearity. This gate subsequently allows for the construction of a non-deterministic CZ gate. Whilst non-deterministic, KLM presented an encoding scheme, based on gate teleportation \cite{bib:GottesmanChuang99}, which overcomes gate non-determinism, with only \mbox{$\mathrm{poly}(n)$} overhead, where $n$ is the number of logical qubits. Since the advent of the original KLM scheme, numerous proposals have been presented, which significantly improve on KLM's original scaling requirements, most promisingly using cluster state approaches \cite{bib:Raussendorf01, bib:Raussendorf03, bib:Nielsen04, bib:Nielsen06, bib:BrowneRudolph05, bib:Varnava05}.
The KLM scheme is efficient, requiring \mbox{$\mathrm{poly}(n)$} ancillary states and \mbox{$\mathrm{poly}(n)$} time-steps. Single-qubit operations are trivial using dual-rail encoding (whereby a qubit is encoded as a superposition of a single photon across two optical modes, \mbox{$\ket{\psi}=\alpha\ket{1,0}+\beta\ket{0,1}$}), requiring only arbitrary \mbox{$SU(2)$} beamsplitter operations. However, two-qubit entangling gates, such as the CZ gate, which are required to construct a universal gate set, require ancillary states, post-selection, and feed-forward.
The KLM scheme can be represented in the form shown in Fig. \ref{fig:KLM}, comprising logical state preparation, $\ket{\psi_L}$, and \mbox{$\mathrm{poly}(n)$} iterations of: ancillary state preparation, $\ket{\psi_A^{(i)}}$; photo-detection; and, passive linear optics, classically controlled by the previous measurement outcomes, $m_i$. This scheme is sufficient for universal LOQC, using \mbox{$\mathrm{poly}(n)$} circuit size. We will now show that the MGDR architecture can be made equivalent to the KLM scheme.
\begin{figure}[!htb]
\includegraphics[width=0.5\columnwidth]{KLM}
\caption{A representation of the KLM scheme for universal LOQC. We prepare a logical state $\ket{\psi_L}$, comprising $n_L$ modes, and apply some \mbox{$\mathrm{poly}(n)$} number of iterations of: $n_A$-mode ancillary state preparation, $\ket{\psi_A^{(i)}}$, at the $i$th step; passive linear optics, $\hat{U}_i$; measurement of the ancillary state, yielding measurement signature $m_i$; and, feed-forward of $m_i$ to control the next passive unitary, $\hat{U}_{i+1}$. From KLM, this scheme enables efficient implementation of a universal gate set, enabling universal quantum computation.} \label{fig:KLM}
\end{figure}
We have already determined that the MGDR architecture is sufficient for implementing all the $\hat{U}_i$ in Fig. \ref{fig:KLM}. The remaining requirement to demonstrate equivalence with KLM is that we are able to add arbitrary ancillary states prior to each $\hat{U}_i$, and measure them following $\hat{U}_i$, prior to the application of $\hat{U}_{i+1}$. We will consider these two stages separately, which we will refer to as `ancilla injection' and `ancilla extraction' respectively.
The protocol for ancilla injection is shown in Fig. \ref{fig:ancilla_injection}. This takes place at the first of the on/off dynamic switches in conjunction with the push-button single-photon source. The on/off switch is set to be completely reflective for duration \mbox{$n_L\tau$}, which completely couples all the logical time-bins into the the outer loop. Immediately following this, the switch is set to become completely transmissive for duration \mbox{$n_A\tau$}, thereby completely coupling the ancillary state $\ket{\psi_A^{(i)}}$, which is prepared via the push-button source, into the outer loop. Then the state immediately after the first switch is the logical state (across $n_L$ time-bins) followed by the ancillary state (across $n_A$ time-bins), yielding a combined pulse-train over \mbox{$n_L+n_A$} temporal modes. Following this, the MGDR protocol is applied to implement the required passive linear optics network $\hat{U}_i$.
\begin{figure}[!htb]
\includegraphics[width=\columnwidth]{ancilla_injection}
\caption{The shown beamsplitter refers to the first beamsplitter in Fig. \ref{fig:full_arch}. Injection of ancillary state $\ket{\psi_A}$ into the outer loop is achieved by first setting the dynamic beamsplitter to be completely reflective for duration \mbox{$n_L\tau$}, which completely couples the logical state $\ket{\psi_L}$ into the outer loop. Then the beamsplitter is set to be completely transmissive, which completely couples the ancillary state $\ket{\psi_A}$ (prepared via the push-button source) into the outer loop. Immediately following the beamsplitter, we have a pulse-train of \mbox{$n_L+n_A$} time-bins, which combined represent the logical and ancillary states.} \label{fig:ancilla_injection}
\end{figure}
Next we must measure the ancillary modes (but not the logical modes) via ancilla extraction. This is the same as the ancilla injection protocol in reverse, as shown in Fig. \ref{fig:ancilla_extraction}. We set the last dynamic switch to be completely reflective for duration $n_L\tau$, completely coupling $\ket{\psi_L}$ into the outer loop. Following this, we set the switch to become completely transmissive, which couples the last $n_A$ time-bins, representing $\ket{\psi_A}$, out of the loop and into the photo-detector, yielding measurement signature $m_i$. Now only the logical modes remain inside the outer loop, until ancilla injection at the end of the round-trip. The measurement signature $m_i$ of the $n_A$ ancilla time-bins is used to classically control the passive linear optics transformation $\hat{U}_{i+1}$ at the subsequent step, which simply corresponds to a classical programming of the sequence of beamsplitter reflectivities in the central beamsplitter of Fig. \ref{fig:full_arch}.
\begin{figure}[!htb]
\includegraphics[width=0.85\columnwidth]{ancilla_extraction}
\caption{The shown beamsplitter refers to the last beamsplitter in Fig. \ref{fig:full_arch}. For the first $n_L\tau$ time-steps, the beamsplitter is set to be completely reflective, which completely couples the logical state $\ket{\psi_L}$ into the outer loop. Following this, the beamsplitter is set to be completely transmissive, which completely couples the ancillary time-bins, $\ket{\psi_A}$, out of the outer loop and into the photo-detector. The measurement signature, $m_i$, is used to classically control the subsequent unitary $\hat{U}_{i+1}$.} \label{fig:ancilla_extraction}
\end{figure}
Clearly, the protocol of: (1) initial logical state preparation, $\ket{\psi_L}$; (2) ancilla state preparation, $\ket{\psi_A^{(i)}}$, at every time-step $i$; (3) dynamically controlled passive linear optics, $\hat{U}_i$; (3) measurement of the ancillary modes, yielding measurement signatures $m_i$; and, (4) dynamic control of $\hat{U}_{i+1}$ at the subsequent step, is sufficient to enable the decomposition shown in Fig. \ref{fig:KLM}, enabling full KLM LOQC using \mbox{$\mathrm{poly}(n)$} physical resources. From KLM, using dual-rail encoding, the number of modes required to represent the logical state $\ket{\psi_L}$, scales as \mbox{$n_L=2n_Q$}, where $n_Q$ is the number of logical qubits. The number of ancillary states scales as \mbox{$n_A=\mathrm{poly}(n_L)$}. And thus the total number of photons and temporal modes in the system scales as $\mathrm{poly}(n_Q)$. Thus, the scheme is efficient.
\subsection{Cluster states}
`Cluster state' quantum computing \cite{bib:Raussendorf01, bib:Raussendorf03, bib:Nielsen06}, also known as `graph state quantum computing, `measurement-based quantum computing' and `one-way quantum computing', is an approach to universal quantum computing that differs from the usual and more familiar circuit-based approach. Rather than preparing an input state, applying some combination of single- and two-qubit gates belonging to a universal gate set, and measuring the output state, we instead prepare a `cluster state', which is a highly entangled multi-qubit state.
A cluster state may be represented as a graph, in which vertices represent qubits prepared in the \mbox{$\ket{+}=(\ket{0}+\ket{1})/\sqrt{2}$} state, and edges represent the application of CZ gates between the respective vertices. A simple example is shown in Fig. \ref{fig:cluster_state}. Having prepared this state, we have a `substrate' from which any quantum computation may be implemented using only single-qubit measurements, where the measurement bases are determined via classical feed-forward from previous measurement outcomes.
\begin{figure}[!htb]
\includegraphics[width=0.35\columnwidth]{cluster_state}
\caption{A simple example of a four-qubit cluster state. Vertices represent qubits initialized in the $\ket{+}$ state, and edges represent the application of CZ gates. CZ gates commute, so the time ordering of the operations is irrelevant.} \label{fig:cluster_state}
\end{figure}
However, whilst a cluster state may be considered in terms of CZ gates acting on the edges of a graph, CZ gates are not the only way to prepare cluster states. For example, Bell pairs, which may be produced via a variety of means, such as SPDC, are locally equivalent to two-qubit cluster states.
The attractive feature of cluster states is that, having prepared the substrate state, an arbitrary quantum computation requires only single-qubit measurements. In a photonic context this is extremely attractive, as single-qubit measurements are trivial.
Cluster states have many other attractive properties. Gross \emph{et al.} \cite{bib:Gross06}, and Rohde \& Barrett \cite{bib:RohdeBarrett07} showed that cluster states may be prepared efficiently using non-deterministic gates, gate non-determinism being the bane of LOQC and the reason why KLM has such high experimental resource overhead. Additionally, cluster states were shown by Varnava \emph{et al.} \cite{bib:Varnava05} to be highly robust against qubit loss, and cluster states are directly suited to various topological codes, to achieve fault-tolerance with extremely favorable fault-tolerance thresholds.
Nielsen \cite{bib:Nielsen04} first demonstrated that by combining a non-deterministic KLM CZ gate with the cluster state formalism, physical resource requirements -- in terms of the number of required qubits and CZ gates -- could be reduced by orders of magnitude compared to KLM. Subsequently, Browne \& Rudolph \cite{bib:BrowneRudolph05} demonstrated that by replacing the KLM CZ gate with parity measurements, physical resource requirements could be further reduced substantially.
Next we will show that the fiber-loop architecture, combined with ancilla injection and extraction, can be mapped to both the Nielsen, and Browne \& Rudolph schemes, making the architecture suitable for highly efficient LOQC.
\subsubsection{Nielsen cluster states}
The Nielsen \cite{bib:Nielsen04} approach to optical cluster state quantum computing directly combines a non-deterministic KLM CZ gate with the cluster state formalism.
To illustrate the concept, let us consider the preparation of a two qubit cluster state (i.e a Bell pair). We prepare two $\ket{+}$ states, and apply the CZ gate, as described by KLM. KLM actually describe a hierarchy of CZ gates, with increasing success probability, but also with (polynomially) increasing resource overhead. Let us choose a particular such gate and let the gate success probability be $p_\mathrm{gate}$. Then, on average we will have to repeat the protocol \mbox{$1/p_\mathrm{gate}$} times to prepare a two-qubit cluster state.
The basic building block of the Nielsen protocol is to bond $\ket{+}$ states together to form star-shaped `micro-cluster' states, with $k$ vertices emanating from a central vertex. We perform this in advance, preparing a number of star clusters offline. The central vertex in each star cluster will ultimately belong to our final cluster state. The emanating branches facilitate joining multiple star clusters together. At the first instance, we take two stars and attempt to bond them together via their branches using the KLM CZ gate. When the CZ gate succeeds, we have prepared two `fused' stars. When it fails, the branches to which the CZ gate was applied are simply removed, leaving us with two star clusters with \mbox{$k-1$} branches, which may be reused for the next attempt at bonding the micro-clusters. However, with $k$ branches, we may have $k$ attempts to perform the joining of the stars. Thus, with
\begin{equation} \label{eq:nielsen_bond_suc}
k=\frac{\mathrm{log}(1-p_\mathrm{bond})}{\mathrm{log}(1-p_\mathrm{gate})}
\end{equation}
branches, we can successfully join two stars with success probability $p_\mathrm{bond}$. The bonding of two micro-clusters is shown in Fig. \ref{fig:micro_cluster}.
\begin{figure}[!htb]
\includegraphics[width=0.75\columnwidth]{micro_cluster}
\caption{An example of the Nielsen protocol for bonding two micro-clusters. We non-deterministically prepare two small clusters with central vertices $A$ and $B$, each with a series of $k$ branches emanating from the central vertex. To join the microclusters we perform non-deterministic CZ gates between the branches of the two micro-clusters in parallel. If a CZ gate fails it simply removes the respective branch qubits from the micro-clusters, and if it succeeds it creates an edge between the respective qubits. For a given gate success probability ($p_\mathrm{gate}$) and desired bonding success probability ($p_\mathrm{bond}$), the required number of branches $k$ is given by Eq. \ref{eq:nielsen_bond_suc}. Having successfully joined the two micro-clusters, all qubits other than $A$ and $B$ are measured in the $Y$ basis, which simply removes all those qubits, whilst creating edges between their neighboring qubits. Thus, after doing this to all qubits other than $A$ and $B$, we are left with just $A$ and $B$, with an edge between them. This idea generalizes to arbitrary graph topologies, whilst remaining efficient.} \label{fig:micro_cluster}
\end{figure}
The protocol continues as one would expect, successively bonding larger and larger smaller clusters to form larger clusters, until we have a cluster of the size required for our computation. Using this protocol, arbitrarily large cluster states may be prepared efficiently, with resource requirements that Nielsen showed were substantially more favorable than the original KLM protocol.
In Sec. \ref{sec:KLM} we already established that the fiber-loop architecture is sufficient for universal KLM LOQC. Thus, the CZ gate required for Nielsen-style cluster state preparation is available to us. Additionally, we must manipulate the topology of the graph, depending on which CZ gates succeed or fail. Manipulating the topology of the graph is trivial, since it only requires permuting time-bins, and of course permutations on optical modes are unitary operations, and thus may be implemented using the MGDR protocol. Permutations only require completely transmissive or reflective beamsplitter operations in the MGDR architecture, and thus do not constitute major mode-matching problems.
Thus, by combining the MGDR protocol with ancilla injection and extraction, we are able to efficiently implement the cluster state approach of Nielsen, yielding a far more efficient approach to universal LOQC than simply using the fiber-loop architecture to implement KLM directly.
\subsubsection{Browne \& Rudolph cluster states}
Whilst the Nielsen scheme is far more favorable than KLM, it nonetheless relies on the non-deterministic KLM CZ gate as its basic building block. Implementing this gate is challenging, as it is based on a Mach-Zehnder interferometer, which requires stabilization on the order of the photons' wavelength, and also requires a large number of optical elements. A far more favorable approach would be to employ only Hong-Ou-Mandel \cite{bib:HOM87}-type interference, which only requires stabilization on the order of the coherence length, which, from an experimental point of view, is extremely advantageous.
To this end, Browne \& Rudolph \cite{bib:BrowneRudolph05} proposed a scheme not requiring any Mach-Zehnder interference. Instead, their scheme relies on polarization encoded qubits (\mbox{$\ket{\psi}=\alpha_0\ket{H}+\alpha_1\ket{V}$}), where bonding smaller clusters into larger clusters is implemented via a joint parity measurement, which is easily implemented via a polarizing beamsplitter (PBS).
Browne \& Rudolph's scheme employs three components: a resource of polarization-encoded Bell pairs; and, type-I and type-II `fusion' gates. The resource of Bell pairs may be prepared offline using a KLM gate (note that entangled pairs prepared via SPDC are \emph{not} suitable for their scheme, owing to the higher order photon-number terms). The type-I and type-II fusion gates are variations on parity measurement, implemented via a combination of waveplates and a PBS, shown in Fig. \ref{fig:fusion}. With these primitives at our disposal, we may prepare arbitrary cluster states with very high efficiency. Refer to Ref. \cite{bib:BrowneRudolph05} for full details of the scheme and its scaling characteristics. A detailed analysis of mode-matching effects in this scheme was presented by Rohde \& Ralph \cite{bib:RohdeRalph06}.
\begin{figure}[!htb]
\includegraphics[width=0.65\columnwidth]{fusion}
\caption{The type-I and type-II fusion gates, comprising two polarization-encoded qubits, a PBS, waveplates, and polarization-resolved photo-detectors. The type-I fusion gate destroys a single qubit, whilst the type-II fusion gate destroys both qubits. Both gates require only Hong-Ou-Mandel-type interference, and thus are more favorable from a mode-matching perspective than the KLM CZ gate, which requires Mach-Zehnder interference.} \label{fig:fusion}
\end{figure}
We will assume that a resource of Bell pairs is available (e.g prepared using KLM gates), and proceed to show that the remaining two requirements -- PBSs and waveplates -- may be mapped to the MGDR architecture, where polarization encoding is mapped to time-bin encoding.
In polarization encoding, a qubit takes the form \mbox{$\ket\psi = \alpha_0\ket{H} + \alpha_1\ket{V}$}, whereas in time-bin encoding the same state takes the form \mbox{$\ket{\psi} = \alpha_0\ket{1_t,0_{t+\tau}} + \alpha_1\ket{0_t,1_{t+\tau}}$}, where the qubit arrives at time $t$ and the time-bin separation is $\tau$.
First, let us consider the action of a waveplate in polarization encoding. A waveplate implements a polarization rotation such that $\vec\alpha' = \hat{U}\vec\alpha$, where \mbox{$\hat{U}\in SU(2)$}. Now we wish to find the equivalent operation in time-bin encoding. Time-bin encoding can be thought of as dual-rail encoding, where the two rails are temporally rather than spatially separated. In dual-rail encoding, the operation equivalent to a waveplate in polarization encoding is the beamsplitter operation. Thus, to implement a single-qubit rotation, we must implement an arbitrary beamsplitter between the temporal modes at $t$ and \mbox{$t+\tau$}. The MGDR protocol describes how to implement arbitrary pairwise beamsplitter operations between temporal modes.
Next, let us consider the PBS operation. In the two-mode polarization basis, the PBS operation implements the transformation,
\begin{eqnarray}
\hat{a}^\dag_{H_1} &\to& \hat{a}^\dag_{H_2}, \nonumber \\
\hat{a}^\dag_{V_1} &\to& \hat{a}^\dag_{V_1}, \nonumber \\
\hat{a}^\dag_{H_2} &\to& \hat{a}^\dag_{H_1}, \nonumber \\
\hat{a}^\dag_{V_2} &\to& \hat{a}^\dag_{V_2},
\end{eqnarray}
on the photonic creation operators. That is, horizontally polarized photons are transmitted, whilst vertically polarized photons are reflected. The full transformation on all four basis states is illustrated in Fig. \ref{fig:PBS}. This transformation may be represented in matrix form as,
\begin{equation} \label{eq:PBS_mapping}
\hat{U}_\mathrm{PBS} = \left(\begin{array}{cccc}
0 & 0 & 1 & 0 \\
0 & 1 & 0 & 0 \\
1 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 \end{array}\right),
\end{equation}
in the basis $\{H_1,V_1,H_2,V_2\}$. Note that this is simply a permutation matrix that swaps the first and third time-bins, which of course is unitary, and thus may be implemented directly using the MGDR protocol. As discussed earlier, permutation operations do not require arbitrary beamsplitter operations -- we need only switch between completely transmissive and completely reflective at the central dynamically controlled beamsplitter -- thereby minimizing mode-matching effects.
\begin{figure}[!htb]
\includegraphics[width=0.65\columnwidth]{PBS}
\caption{(top) The action of a PBS on all four basis states of two polarization-encoded qubits. (bottom) The equivalent transformation in time-bin encoding based on the transformation of Eq. \ref{eq:PBS_mapping}, which is simply a swap of the 1st and 3rd time-bins.} \label{fig:PBS}
\end{figure}
To complete the type-I and type-II fusion gates we must measure some of the modes, which is accomplished via the ancilla extraction protocol.
Having demonstrated the ability of the fiber-loop architecture to implement type-I and type-II fusion gates, it follows from Browne \& Rudolph that arbitrary cluster states may be efficiently prepared, and their preparation strategies and scaling results apply directly.
As an interesting aside, using polarization encoding waveplates implement a polarization rotation, which does not generate entanglement, whereas the PBS is the entangling operation. On the other hand, in time-bin encoding the equivalent of the waveplate is the time-bin beamsplitter operation, which generates temporal entanglement, whereas the equivalent of the PBS -- a swap operation on time-bins -- does not. Nonetheless, despite the reversed role of entanglement generation, the two encodings are isomorphic.
\section{Conclusion}
We have presented a simple experimental architecture for universal LOQC using constant experimental resources. The scheme requires a single push-button photon source or Bell pair source, three dynamically controlled beamsplitters, two nested fiber-loops, and a single time-resolved photo-detector. The experimental requirements for the architecture are fixed, irrespective of the size of the computation, yet can efficiently implement full, universal LOQC, equivalent to the KLM protocol or subsequent and more efficient cluster state protocols. There is a single point of interference in the architecture, which may make optical alignment in this scheme far more favorable than existing bulk optics or waveguide approaches. The size of the computation that can be implemented is limited only by the length of the outer fiber-loop, which must have round-trip time of at least \mbox{$n\tau$}, where $n$ is the number of time-bins, which scales polynomially with the number of logical qubits in the computation. By removing the ancilla injection and ancilla extraction stages of the protocol, this scheme reduces to the MGDR architecture for universal boson-sampling. The required technologies to implement this scheme are, for the larger part, available today. Thus, elementary demonstrations of this protocol might be viable in the near future.
\begin{acknowledgments}
We thank Keith Motes, Alexei Gilchrist, Timothy Ralph, Geoff Pryde, Andrew White, and Jonathan Dowling for helpful discussions. We thank Yoram Zekri for helping to motivate this work. This research was conducted by the Australian Research Council Centre of Excellence for Engineered Quantum Systems (Project number CE110001013).
\end{acknowledgments}
|
\section{Introduction}\label{sec:introduction}
Metal borohydrides are amongst the most promising hydrogen storage
materials, \cite{Graetz_2009:new_approaches,
Ronnebro_2011:development_group, Li_2011:recent_progress,
Rude_2011:tailoring_properties, Jain_2010:novel_hydrogen} as they have
some of the highest storage densities. Unfortunately, the hydrogen
desorption temperature for the most attractive borohydrides is too high
for on-board storage, where it should be below
85$^\circ$C.\cite{DOE_Targets_Onboard_2009, Yang_2010:high_capacity}
Thus, borohydrides have been studied intensely in order to lower their
desorption temperature. Destabilizing via reactions with other hydrides
has been suggested, \cite{Vajo_2005:reversible_storage,
Alapati_2007:using_first, Alapati_2008:large-scale_screening,
Li_2008:dehydriding_rehydriding} as well as simple
doping,\cite{Nickels_2008:tuning_decomposition,
Hoang_2009:hydrogen-related_defects} cation
substitution,\cite{Setten_2007:ab_initio} anion
substitution,\cite{Brinks_2008:adjustment_stability} or adding
catalysts\cite{Li_2007:effects_ball}---but unfortunately, the results
still fall short of the required DOE
targets.\cite{DOE_Targets_Onboard_2009}
The borohydride Mg(BH$_4$)$_2$\ is of particular interest; it has substantial
storage density of 14.9~mass\% and its decomposition has been shown to
be fully reversible under certain
conditions,\cite{Severa_2010:direct_hydrogenation} but it completes its
first major intermediate step around 570~K and doesn't fully desorb
until it reaches temperatures around
820~K.\cite{Li_2008:dehydriding_rehydriding,
Soloveichik_2009:magnesium_borohydride} If we can lower its desorption
temperature, it might be one of the few materials to satisfy the DOE
targets. The desorption temperature is determined by the thermodynamics
and kinetics of the desorption reaction. Although it is generally
believed that kinetics is the ``culprit'' in the case of
Mg(BH$_4$)$_2$,\cite{Ozolins_2008:first-principles_prediction} it is still
desirable to first optimize the thermodynamics before addressing the
kinetic barrier.\cite{Alapati_2008:large-scale_screening} In the present
manuscript, we investigate the thermodynamics of the desorption of Mg(BH$_4$)$_2$\
and the dramatic effect Zn alloying has on it.
We have recently become aware of some very nice work also studying Mg/Zn
borohydride solid solutions, and will use this opportunity to compare
our results with theirs and point out similarities and
differences.\cite{Albanese_2013:theoretical_experimental} Further
simulations have been performed studying the desorption of
Mg(BH$_4$)$_2$,\cite{Zhang_2012:theoretical_prediction,
Setten_2008:density_functional} but without accounting for van der Waals
contributions, known to be necessary to achieve the correct energetic
ordering among polymorphs of
Mg(BH$_4$)$_2$;\cite{Huan_2013:thermodynamic_stability, Bil_2011:van_waals} we will
thus compare our results of the desorption pathway with other works to
ascertain the effect of van der Waals interactions. We also argue that
given the minimum practical delivery pressure from storage system of 3
bar, the overall desorption reaction is, in fact, outside the ideal
thermodynamic window of $-40$$^\circ$C to
+85$^\circ$C,\cite{DOE_Targets_Onboard_2009} but can be brought there by
alloying with Zn. We also find a linear relationship between the overall
hydrogen desorption enthalpy and Zn concentration in Mg(BH$_4$)$_2$.
Borohydrides are complex hydrides, in which tetrahedral anion units,
such as [BH$_4$]$^-$ and [AlH$_4$]$^-$, are bound to more
electropositive cations, such as Li, Na, K, Mg, and Ca. The ionic
bonding and charge transfer between the cations and the anion units is a
key feature of the stability of these
hydrides.\cite{Nakamori_2006:correlation_between} Targeting this
feature, the desorption temperature of Mg$_2$NiH$_4$ was successfully
lowered by destabilizing it through cation
substitution.\cite{Setten_2007:ab_initio} The same can also be achieved
by altering the anion complex, as demonstrated with fluorine
substitution in the hydrogen sublattice of
Na$_3$AlH$_6$.\cite{Brinks_2008:adjustment_stability,
Graetz_2009:new_approaches} More generally, an extensive experimental
study revealed an almost linear correlation between the desorption
temperature $T_d$ in K and the Pauling electronegativity $\chi_P$ as
well as the hydrogen desorption reaction enthalpy $\Delta H_r$ in
kJ/mol~H$_2$:\cite{T_d_fit}
\begin{eqnarray}
T_d &=& 1234 - 517.3\;\chi_P\;,\label{equ:fit_chi}\\
T_d &=& 423.4 + 8.34\;\Delta H_r\;\label{equ:fit_H}.
\end{eqnarray}
The study included data for $\mathcal{M}$(BH$_4$)$_n$ with $\mathcal{M}$
= Li, Na, K, Mg, Zn, Sc, Zr, and
Hf,\cite{Nakamori_2006:correlation_between} and was later extended by
Ca, Ti, V, Cr, and Mn.\cite{Nakamori_2007:thermodynamical_stabilities}
Zn(BH$_4$)$_2$ stands out in that it has the lowest temperature for full
decomposition of only 410~K.\cite{Jeon_2006:mechanochemical_synthesis}
Zn(BH$_4$)$_2$ itself is thermodynamically unstable at room temperature
and produces diborane gases in its decomposition reaction---it is thus
not directly interesting for hydrogen storage. But, as we will argue
below, Zn is an ideal alloyant for the Mg(BH$_4$)$_2$\ system, forming
Mg$_{1-x}$Zn$_x$(BH$_4$)$_2$, with remarkable effects. Zn alloying of
Mg(BH$_4$)$_2$\ is supported by the following facts: (i) Zn alloying has been found
experimentally to significantly lower the decomposition
temperature;\cite{Albanese_2013:theoretical_experimental,
Kalantzopoulos_2014:hydrogen_storage} (ii) Zn alloying lowers the
decomposition temperature of other
borohydrides;\cite{Ravnsbaek_2009:series_mixed-metal} (iii) Zn is known
to form a borohydride with the same stoichiometry as Mg(BH$_4$)$_2$\ and their
structures are essentially the same
\cite{Nakamori_2006:correlation_between} (the ionic radii of Zn and Mg
of 0.88 and 0.86 \AA\ are virtually identical); (iv) alloying in other
borohydrides, such as Li$_{1-x}$Cu$_x$BH$_4$, shows a desorption
temperature between the two
constituents;\cite{Miwa_2005:first-principles_study,
Nickels_2008:tuning_decomposition} (v) Zn is known to not form hydrides
or borides like other
borohydrides,\cite{Nakamori_2006:correlation_between} simplifying the
hydrogen desorption reaction significantly; and finally (vi) Zn is an
abundant major industrial metal.
\section{Computational Details}
In order to study the thermodynamics of Mg(BH$_4$)$_2$\ and the effect of Zn
alloying, we need the enthalpies and entropies for all materials
suspected in the decomposition of Mg(BH$_4$)$_2$. To this end, we performed {\it ab
initio} simulations at the DFT level, as
implemented in \textsc{Vasp},\cite{Kresse_1996:efficient_iterative,
Kresse_1999:ultrasoft_pseudopotentials} to calculate the ground-state
energies and phonon densities of states. We used PAW pseudopotentials
with a 650 eV kinetic energy cutoff and an energy convergence criterion
of $10^{-7}$~eV. We used $k$-point meshes giving convergence to within
1~meV/atom, with the exception of the metallic compounds and elemental
boron, which were converged to within 3~meV/atom. This yielded a
$k$-point mesh of e.g.\ $5\times5\times4$ for small unit cells like
MgB$_2$ and a mesh of $1\times1\times1$ for large unit cells like Mg(BH$_4$)$_2$.
Structures were relaxed with respect to unit-cell parameters and atom
positions until all forces were less than 0.1~meV/\AA. This way, we
found the lowest-energy structure of boron to be the 106 atom
$\beta$-rhombohedral structure suggested by van Setten et
al.\cite{Setten_2007:thermodynamic_stability} Phonons were calculated with the
symmetry-reduced finite-displacement method with displacements of
0.015~\AA. Supercells were created such that they were the same
dimensions as the $k$-point mesh used for the original unit cell. The
exceptions were the metals Mg and Zn, whose phonons were calculated with
a $5\times5\times2$ supercell and a $3\times3\times4$ $k$-point mesh.
While the ground-state structure of Mg(BH$_4$)$_2$\ is experimentally
known to be of P6$_1$22 symmetry with 30 formula units per unit cell,
theoretical studies with a variety of exchange-correlation (XC)
functionals find numerous other structures, all disagreeing with
experiment.\cite{Bil_2011:van_waals} The reason for this is that
Mg(BH$_4$)$_2$---similar to other borohydrides
\cite{Lodziana_2010:multivalent_metal}---exhibits a small but important
contribution from van der Waals interactions, on the order of 0.1~eV per
BH$_4$ unit. Including those contributions via the XC functional
vdW-DF\cite{Thonhauser_2007:van_waals,
Langreth_2009:density_functional}---with the (semi)local XC as
originally defined in Ref.~[\onlinecite{Dion_2004:van_waals}]---we were
the first to find the correct ground-state structure in agreement with
experiment.\cite{Bil_2011:van_waals} We further found that the closely
related P3$_1$12 structure is less than 1~meV per atom higher in energy.
This structure has only 9 formula units per unit cell and the local
coordination is almost identical to the P6$_1$22 structure, resulting in
almost identical phonon densities of states. Thus, in the present study we are
using vdW-DF and the numerically feasible P3$_1$12 structure for all our
simulations. This is further justified by comparing one of our reaction
enthalpies to the results of Albanese et al.\ in
Ref.~[\onlinecite{Albanese_2013:theoretical_experimental}], who in fact
used the large P6$_1$22 structure and obtained almost identical
results---see details below. As our structure has 9 formula units per
unit cell, we report results for Zn concentrations in steps of 1/9.
Alloying was done by randomly replacing Mg atoms with Zn, with
variations---due to which atoms are being replaced---being on the order
of only fractions of a kJ/mol at the most. Our results for the enthalpy
of mixing compare well with Albanese et al., who found similarly small
values.\cite{Albanese_2013:theoretical_experimental}
\section{Results}
\begin{table*}
\caption{\label{tab:bader_analysis}Bader charges (in units of $e$) for
the metal site $\mathcal{M}$ and the BH$_4$ units as a function of
concentration $x$ in Mg$_{1-x}$Zn$_x$(BH$_4$)$_2$. Given are also the
Pauling electronegativity $\chi_P$ and the estimated desorption
temperature $T_d$ (in K) from Eq.~(\ref{equ:fit_chi}). The reported
charges and electronegativities are values averaged over all 9 formula
units in the unit cell.}
\begin{tabular*}{\textwidth}{@{}l@{\extracolsep{\fill}}rrrrrrrrrr@{}}\hline\hline
$x$ & 0 & 1/9 & 2/9 & 3/9 & 4/9 & 5/9 & 6/9 & 7/9 & 8/9 & 1 \\\hline
$\mathcal{M}$ & +1.72 & +1.65 & +1.59 & +1.53 & +1.46 & +1.40 & +1.33 & +1.26 & +1.20 & +1.13\\
BH$_4$ & $-$0.86 & $-$0.83 & $-$0.79 & $-$0.76 & $-$0.73 & $-$0.70 & $-$0.67 & $-$0.63 & $-$0.60 & $-$0.56\\\hline
$\chi_P$ & 1.20 & 1.24 & 1.29 & 1.33 & 1.38 & 1.42 & 1.47 & 1.51 & 1.56 & 1.60\\
$T_d$ & 613 & 593 & 567 & 546 & 520 & 499 & 474 & 453 & 427 & 406 \\\hline\hline
\end{tabular*}
\end{table*}
\begin{figure}
\includegraphics[width=\columnwidth]{compound}
\caption{\label{fig:storage_density}H$_2$ gravimetric storage density in
mass\% (to be read off the left axis) and H$_2$ volumetric storage
density in kg~H$_2$/m$^3$ (right axis) for different Zn concentrations.}
\end{figure}
\subsection{Storage Densities}
The gravimetric storage density of Mg(BH$_4$)$_2$\ will decrease with increasing Zn
content, as can be seen in Fig.~\ref{fig:storage_density}. But, as we
shall see below, only modest alloying is necessary to achieve the
required effect, and the loss in gravimetric density is reasonable. On
the other hand, the volumetric density increases with increasing Zn
content. The ``bump'' or deviation from linear behavior, observed in
Fig.~\ref{fig:storage_density} around 50\% Zn concentration, is also
seen in the work of Albanese et
al.,\cite{Albanese_2013:theoretical_experimental} and may be indicative
of the formation of a superstructure.
\subsection{Structural Stability and Ionic Character}
We begin by analyzing the structural stability in terms of the ionic
character as a function of Zn concentration. As mentioned above, the
ionic bonding is a key feature of the stability of borohydrides. In
Table~\ref{tab:bader_analysis} we quantify this picture and present a
Bader charge analysis as a function of Zn concentration, using the fast
implementation proposed by Henkelman et
al.\cite{Henkelman_2006:fast_robust} Note that the Bader analysis,
similarly to the Mulliken analysis, is an intuitive (but not unique) way
of partitioning the electron charge. Interestingly, even for the pure
Mg(BH$_4$)$_2$\ structure, the ionic character significantly deviates from the
nominal values of +2 and $-1$. As expected, the ionic character
diminishes as a function of Zn concentration due to the higher
electronegativity of Zn, approximately 0.06~$e$ per 10\% Zn. A more
detailed analysis of the charge density reveals the following: While the
direct effect of Zn is localized, hydrogens further away from Zn get a
small compensating charge if they are within a tetrahedra next to a Zn.
Through this mechanism, even low levels of alloying influence almost the
entire structure. In Table~\ref{tab:bader_analysis} we also show the
Pauling electronegativity $\chi_P$ as a function of Zn concentration and
the corresponding estimated desorption temperatures $T_d$ according to
Eq.~(\ref{equ:fit_chi}). Purely based on this experimentally found
connection for borohydrides, we can already estimate that modest Zn
alloying might reduce the desorption temperature on the order of 100~K.
In the remainder of this paper, we will quantify this empirical
estimate.
\subsection{Thermodynamics of the Hydrogen Desorption at 1 Bar Hydrogen
Pressure}
We now move to the main point of this paper, i.e.\ the thermodynamics of
the hydrogen desorption and the effect of Zn alloying. To this end, we
calculate the temperature dependent vibrational contribution to the
enthalpy and entropy as
\begingroup\small
\begin{eqnarray}
H_{\rm vib}&=&\int_0^\infty d\omega
\bigg(\frac{1}{2}+\frac{1}{\exp[\hbar\omega/kT]-1}\bigg)
g(\omega)\hbar\omega\;,\label{eq:enthalpy_phonon}\\[1ex]
S_{\rm vib} &=& \int_0^\infty d\omega
\bigg(\frac{\hbar\omega}{2T}\coth\frac{\hbar\omega}{2kT}-k\ln\Big[2\sinh\frac{\hbar\omega}{2kT}\Big]\bigg)
g(\omega)\;,\mspace{30mu}\label{eq:entropy_phonon}
\end{eqnarray}
\endgroup
where $\omega$ is the vibrational frequency, $g(\omega)$ is the phonon
density of states, $T$ is the temperature, and $k$ is Boltzmann's
constant. The enthalpy then is the sum of the DFT ground-state energy
and this vibrational contribution. Formation enthalpies, in turn, are
calculated as differences in enthalpies of the material and its
constituent elements (thus, elements in their natural state have formation enthalpies of
0). From this, we calculate enthalpies and entropies of reaction, using
formation enthalpies and absolute entropies of all materials
(tabulated in the appendix). For reactions involving Zn
alloying (Reactions~2, 3, and 13 in Table~\ref{tab:react_values}), the
entropy of mixing was calculated according to
\begin{equation}
S_{\rm mix}= -k_B\; [c\ln{c}+(1-c)\ln{(1-c)}]\;,
\end{equation}
where $c$ is the concentration of Zn; this entropy of mixing was added
to the entropy of reaction calculated from vibrational frequencies,
resulting in e.g.\ a decrease in reaction entropy of
$\sim5$~J/K/mol~H$_2$ and a corresponding increase in the critical
temperature of $\sim10$~K. Note that all structures we found are true
local minima and none of the density of states show imaginary
frequencies.
Following the approach of van Setten et
al.,\cite{Setten_2008:density_functional} the temperature-dependent
thermodynamic values of H$_2$ were obtained using experimental
data.\cite{Hemmes_1986:thermodynamic_properties} In particular, we used
$H_{\rm{H_2\,gas}}(T)=E_{\rm{H_2}}+E^{\rm{ZPE}}_{\rm{H_2}}+H^{\rm{exp}}_{\rm{H_2\,gas}}(T)$,
where the electronic energy $E_{\rm{H_2}}$ and zero-point energy
$E^{\rm{ZPE}}_{\rm{H_2}}$ were calculated using DFT and the last term
was taken from experiment.\cite{Hemmes_1986:thermodynamic_properties}
Specifically, data was taken from H$_2$ at 1~bar for increments of 100~K
with values in between being linearly interpolated. Note that the
enthalpy of H$_2$ changes very little over moderate pressure changes
(e.g. the change in enthalpy going from 1 to 100~bar at 300~K is only
0.114~kJ/mol), meaning our formation enthalpies should be useful even
when looking at high pressures. The entropy of hydrogen gas used to
calculate the entropies of reaction in Table~\ref{tab:react_values}
and the reactions at 3~bar discussed later on were likewise taken from
the same experimental data.\cite{Hemmes_1986:thermodynamic_properties}
For reference, the entropies of H$_2$ at 300~K for 1 and 3~bar are
130.77 and 121.63~J/K/mol~H$_2$, respectively. Note that the entropy
of hydrogen gas for other pressures can be accurately estimated by the
equation $S_{\rm{H_2}}=-R\ln{p}+S_0$, where $p$ is the pressure in bar
and $S_0$ is the entropy at 1~bar.
\begin{table*}
\caption{\label{tab:react_values}Reaction enthalpies $\Delta H_r$ in
kJ/mol~H$_2$ and entropies $\Delta S_r$ in J/K/mol~H$_2$ at 300 K for
several Mg(BH$_4$)$_2$\ desorption reactions. The critical temperature $T_c$
predicted from the van't Hoff equation $\ln{p}=-\Delta H/RT+\Delta S/R$
for 1 bar H$_2$ pressure is given in K. We give a rough estimate of the
kinetic barrier in K as the difference $T_{\textrm{barrier}}=T_d-T_c$,
where $T_d$ is the approximate experimental desorption temperature.}
\begin{tabular*}{\textwidth}{@{}lr@{\quad$\longrightarrow$\quad}l@{\extracolsep{\fill}}rrrrr@{}}\hline\hline
No.& Reactants & Products & $\Delta H_r^{\rm 300K}$ & $\Delta S_r^{\rm 300K}$ & $T_c$ & $T_d$ & $T_{\textrm{barrier}}$\\\hline
1 & Mg(BH$_4$)$_2$\ & $\textrm{MgB}_2+4\textrm{H}_2$ & 40.3 & 112.15 & 360 & 820$^a$ & 460\\
2 & Mg$_{2/3}$Zn$_{1/3}$(BH$_4$)$_2$ & $\frac{2}{3}\textrm{MgB}_2+\frac{1}{3}\textrm{Zn}+\frac{2}{3}\textrm{B}+4\textrm{H}_2$ & 25.3 & 106.72 & 237 \\
3 & Mg$_{1/3}$Zn$_{2/3}$(BH$_4$)$_2$ & $\frac{1}{3}\textrm{MgB}_2+\frac{2}{3}\textrm{Zn}+\frac{4}{3}\textrm{B}+4\textrm{H}_2$ & 10.4 & 106.57 & 98 \\
4 & Zn(BH$_4$)$_2$ & $\textrm{Zn}+2\textrm{B}+4\textrm{H}_2$ & $-3.92$ & 112.57 & $-35$ & 410$^b$ & 445\\
\hline\noalign{\vskip 0.3mm}
5 & Mg(BH$_4$)$_2$\ & $\frac{1}{6}\textrm{MgB$_{12}$H$_{12}$}+\frac{5}{6}\textrm{MgH}_2+\frac{13}{6}\textrm{H}_2$ & 24.8 & 104.71 & 237 & 570$^c$ & 333 \\
6 & MgH$_2$ & $\textrm{Mg}+\textrm{H}_2$ & 78.7 & 132.99 & 592 & 640$^d$ & 48 \\
7 & $\frac{1}{6}\textrm{MgB$_{12}$H$_{12}$}$ & $\frac{1}{6}\textrm{Mg}+2\textrm{B}+\textrm{H}_2$ & 85.4 & 119.30 & 716 & 730$^d$ & 14 \\
\noalign{\vskip 0.3mm}\hline\noalign{\vskip 0.3mm}
8 & Mg(BH$_4$)$_2$\ & $\frac{1}{6}\textrm{MgB$_{12}$H$_{12}$}+\frac{5}{6}\textrm{Mg}+3\textrm{H}_2$ & 39.8 & 112.56 & 353 \\
9 & $\frac{1}{6}\textrm{MgB$_{12}$H$_{12}$}+\frac{5}{6}\textrm{Mg}$ & $\textrm{MgB}_2+\textrm{H}_2$ & 42.0 & 110.91 & 379 \\
10 & $\frac{1}{6}\textrm{MgB$_{12}$H$_{12}$}$ & $\frac{1}{6}\textrm{MgH}_2+2\textrm{B}+\frac{5}{6}\textrm{H}_2$ & 86.8 & 116.56 & 744 \\
11 & $\frac{1}{6}\textrm{MgB$_{12}$H$_{12}$}+\frac{5}{6}\textrm{MgH}_2$ & $\frac{1}{2}\textrm{MgB}_4+\frac{1}{2}\textrm{MgH}_2+\frac{4}{3}\textrm{H}_2$ & 63.5 & 121.13 & 525 \\
\noalign{\vskip 0.3mm}\hline\noalign{\vskip 0.3mm}
12 & Mg(BH$_4$)$_2$\ & $\textrm{MgH}_2+2\textrm{B}+3\textrm{H}_2$ & 42.0 & 108.00 & 389 \\
13 & Mg$_{4/5}$Zn$_{1/5}$(BH$_4$)$_2$ & $\frac{4}{5}\textrm{MgH}_2+\frac{1}{5}\textrm{Zn}+2\textrm{B}+\frac{16}{5}\textrm{H}_2$& 30.5 & 104.98 & 291 \\\noalign{\vskip 0.3mm}\hline\hline
\end{tabular*}
\raggedright $^a$Ref.~[\onlinecite{Soloveichik_2009:magnesium_borohydride}]; $^b$Ref.~[\onlinecite{Jeon_2006:mechanochemical_synthesis}]; $^c$Ref.~[\onlinecite{Yan_2008:differential_scanning}]; $^d$Ref.~[\onlinecite{Li_2008:dehydriding_rehydriding}].
\end{table*}
The overall energy required for the hydrogen decomposition reaction to
start can be split into two pieces, i.e.\ the \emph{reaction enthalpy}
and an additional \emph{kinetic barrier}. We argue that the latter is
similar for the initial hydrogen release of many decomposition reactions
of borohydrides and focus first on the reaction enthalpies.
Table~\ref{tab:react_values} contains the most pertinent results from
this study; it lists reaction enthalpies for a number of possible
reactions, calculated from formation enthalpies for a range of
materials. Where a direct comparison with experiment is possible, we
generally find very good agreement. For example, experimental formation
enthalpies for MgH$_2$ and MgB$_2$ of $-75.3$ and $-41.2$~kJ/mol are in
excellent agreement with our calculated values of $-78.7$ and
$-43.4$~kJ/mol (see appendix A).\cite{Vajo_2004:altering_hydrogen,
Balducci_2005:thermodynamics_intermediate} Furthermore, our calculated
reaction enthalpies for Reactions 1 and 8 of
40.3~kJ/mol~H$_2$ and 39.8~kJ/mol H$_2$ are also in good agreement with
the value of $40\pm2$~kJ/mol~H$_2$ found by Yan et al.\ in
Ref.~[\onlinecite{Yan_2008:differential_scanning}].\footnote{The authors
of Ref.~[\onlinecite{Yan_2008:differential_scanning}] propose here the
decomposition of MgB$_{12}$H$_{12}$\ before MgH$_2$, which we find to be unlikely as it
has a critical temperature of 716~K. We suspect the actual reaction
studied was the one we proposed.} Previous theoretical studies have
calculated an overall reaction enthalpy of 38--39~kJ/mol~H$_2$ at room
temperature,\cite{Setten_2008:density_functional,
Zhang_2012:theoretical_prediction,
Ozolins_2008:first-principles_prediction,
Ozolins_2009:first-principles_prediction} which closely agrees with our
value of 40.3~kJ/mol~H$_2$. We attribute the difference to vdW-DF, which
stabilizes Mg(BH$_4$)$_2$\ with respect to its reaction products by accounting for
the long-range van der Waals forces. Comparison of our results for other
decomposition reactions with previous studies shows similar agreement,
with differences on the order of several kJ/mol~H$_2$ due to vdW-DF.
Although accounting for long-range van der Waals forces has been found
to be critical to finding the correct energetic ordering among
polymorphs in borohydrides,\cite{Huan_2013:thermodynamic_stability,
Bil_2011:van_waals} it seems to affect total reaction enthalpies by only
several kJ/mol~H$_2$ at the most.
Reactions 1 -- 4 of Table~\ref{tab:react_values} show the effects of Zn
alloying on the overall reaction enthalpy. We find the effect to be
linear and every 10\% Zn results in a further lowering of the desorption
enthalpy by approximately 4.5~kJ/mol~H$_2$, as can be seen in
Fig.~\ref{fig:reaction_enthalpies}. Zn
alloying thus provides a convenient way of tuning the desorption
enthalpy over a wide range. Note that the desorption enthalpy of
Reactions 2 and 5 are approximately the same, but the important
difference is that the alloyed phase releases the entire hydrogen, while
the latter only releases half of the available hydrogen. Furthermore,
note that Reactions 1 -- 4 are different from the decomposition modeled
in Ref.~[\onlinecite{Albanese_2013:theoretical_experimental}] in that we
model the entire hydrogen release, i.e.\ the outcome is MgB$_2$ +
4H$_2$, whereas they assume the intermediate MgH$_2$. For comparison, we
have modeled their reaction also (Reaction 13 in
Table~\ref{tab:react_values}), obtaining 30.5~kJ/mol~H$_2$ in excellent
agreement to their value of 30~kJ/mol~H$_2$---in fact, justifying
approximations that both our groups have made. It is important to also
compare the two corresponding ``pure'' reactions, i.e.\ Reaction 1 and
12. The reaction enthalpy for the entire hydrogen release in Reaction 1
is, in fact, a little lower compared to Reaction 12, which led us to
pursue the pure reaction pathway.
\begin{figure}
\includegraphics[width=\columnwidth]{mgb2_enthalpy}
\caption{\label{fig:reaction_enthalpies}Reactions enthalpies
$\Delta H_r$ can readily be calculated from the
tabulated data in the appendices. As an example, we plot here the
reaction enthalpy as a function of temperature and alloying for the
reaction $\textrm{Mg$_{1-x}$Zn$_x$(BH$_4$)$_2$} \rightarrow
\textrm{Mg$_{1-x}$B$_{2(1-x)}$} + \textrm{Zn$_x$} + 2\textrm{B$_x$} +
4\textrm{H$_2$}$. Note that the effect of Zn concentration on the
reaction enthalpy is almost perfectly linear.}
\end{figure}
Note that we did not list the effects of Zn alloying on intermediate
reactions because it was unknown whether Zn would remain in MgB$_{12}$H$_{12}$\ or
phase separate, and in either case there was no experimental data on the
corresponding structures. Reactions 5 -- 7 show suspected intermediate
steps, while Reactions 8 -- 13 show alternative reactions discussed
later; data for other possible reactions can be readily calculated using
extensive thermodynamic data tabulated in the appendix.
Of course, it is also well known that there are more intermediates in
the decomposition of Mg(BH$_4$)$_2$\ involving Mg(B$_x$H$_y$)$_z$
complexes,\cite{Ozolins_2008:first-principles_prediction,
Newhouse_2010:reversibility_improved} especially in the formation and
decomposition of MgB$_{12}$H$_{12}$. While we did not study all of these
possible intermediates, our calculated thermodynamic values in the
appendix can give insight into some other favorable
pathways. Within the framework of pathways considered, our results show
that the most favorable pathway for hydrogen desorption that agrees with
experiment is:
\begin{align}
\textrm{Mg(BH$_4$)$_2$} &\rightarrow \frac{1}{6}\textrm{MgB$_{12}$H$_{12}$}\label{eq:reaction_mgb12h12}
+\frac{5}{6}\textrm{MgH$_{2}$}
+ \frac{13}{6}\textrm{H}_2\\
&\rightarrow \frac{1}{6}\textrm{MgB$_{12}$H$_{12}$}
\label{eq:reaction_mg}
+\frac{5}{6}\textrm{Mg}
+ 3\textrm{H}_{2}\\
&\rightarrow \textrm{Mg}
\label{eq:reaction_elements}
+2\textrm{B}
+ 4\textrm{H}_{2}\\
&\rightarrow \textrm{MgB$_{2}$} +
4\textrm{H$_2$}\label{eq:reaction_mgb2}
\end{align}
Thermodynamically, as seen in Reaction 9, the most favorable pathway
after the formation of MgB$_{12}$H$_{12}$\ is the formation of MgB$_2$, but from
experimental differential scanning calorimetry and chemical intuition,
it is likely that there is at least one intermediate step from MgB$_{12}$H$_{12}$\ to
MgB$_2$. The theoretical total hydrogen loss for steps
\eqref{eq:reaction_mgb12h12}, \eqref{eq:reaction_mg}, and
\eqref{eq:reaction_mgb2} is 8.1\%, 11.2\%, and 14.9\%, respectively.
The formation of MgB$_{12}$H$_{12}$, MgH$_2$, Mg, and MgB$_2$ are all well-known steps
confirmed by X-ray diffraction and solid state
NMR.\cite{Li_2008:dehydriding_rehydriding, Yan_2011:formation_process,
Soloveichik_2009:magnesium_borohydride, Yang_2011:decomposition_pathway}
They are also supported as thermodynamically favorable by a number of
theoretical studies.\cite{Zhang_2012:theoretical_prediction,
Zhang_2010:theoretical_prediction} Furthermore, the temperature
evolution of the reaction products also supports this reaction pathway:
MgH$_2$ appears first, then Mg without MgH$_2$, and finally MgB$_2$.
Equation~\eqref{eq:reaction_mgb12h12} also corresponds to the mass\% of
hydrogen desorbed for certain temperatures held for long periods of
time. Yan et al.\ found that just over 8 mass\% of H$_2$ desorb after
1000~min at 573~K; this corresponds almost exactly to the expected value
of 8.1\%\ for
Eq.~\eqref{eq:reaction_mgb12h12}.\cite{Yan_2011:formation_process} It
is also worth noting that it took over an hour before even 4 mass\% of
H$_2$ was desorbed, and the curves for lower temperatures never
equilibrated after 1000~min. Due to the sluggish kinetics of this step
of the reaction it is difficult to see the individual reaction steps in
a typical thermogravimetry curve; by the time only a fraction of Mg(BH$_4$)$_2$\
has undergone the first step of the reaction, the temperature has
increased so that the second step of the reaction has already begun.
However, we can see further evidence for the proposed reaction pathway
by the results of Matsunaga et al., who found that Mg(BH$_4$)$_2$\ could be
rehydrogenated from 11\%\ to 8\%\ at 623~K, which corresponds to the
rehydrogenation of Mg to MgH$_2$ (see Fig.~6 of
\nocite{Matsunaga_2008:hydrogen_storage}
Ref.~[\onlinecite{Matsunaga_2008:hydrogen_storage}]).
We also find that the previously suggested decomposition of MgB$_{12}$H$_{12}$\ to
MgH$_2$ and elements shown in Reaction~10 to be unfavorable as a second
step of the reaction, with a critical temperature of 744~K, and suggest
the decomposition of MgH$_2$ ($T_c=592$~K) as occurring before MgB$_{12}$H$_{12}$.
After MgH$_2$ decomposes, we suggest the decomposition of MgB$_{12}$H$_{12}$\, as our
critical temperature of 716~K coincides closely with the final large dip
observed in differential scanning calorimetry plots near 1 bar H$_2$
pressure.\cite{Yan_2008:differential_scanning} Finally, we find the
formation of MgB$_4$ to be plausible, as evidenced by the favorable
thermodynamics of Reaction~11.
Reaction~13 is the decomposition reaction modeled by Albanese et al.\ as
having the optimal decomposition
enthalpy;\cite{Albanese_2013:theoretical_experimental} our calculated
value of 30.5~kJ/mol~H$_2$ is in exact agreement with their result of
30~kJ/mol~H$_2$. Note that our enthalpy and entropy used for 20\%\
Zn-alloyed Mg(BH$_4$)$_2$\ were calculated from a linear interpolation of our
thermodynamic data; this is valid because the change in enthalpy and
entropy values with Zn concentration is nearly exactly linear (see the
appendix). It is interesting to see that, when we
interpolate our results for Reactions 1 -- 4 to a Zn concentration of
1/5 to match the concentration in Reaction 13, we find a reaction
enthalpy and entropy of 31.46~kJ/mol~H$_2$ and 112~J/K/mol~H$_2$,
leading to a critical temperature $T_c=281$~K, which is essentially the
same as for Reaction 13 in Table~\ref{tab:react_values}. Although our
results for Reaction~13 agree with the work done by Albanese et al., our
work can be seen as an extension of the theoretical part of that study,
as we have actually calculated the full thermodynamic data---including
reaction entropies---instead of interpolating them. We have done so for
their proposed partial decomposition (Reaction 13) and in addition for a
large number of other reactions (Reactions 1 -- 11), including the full
hydrogen decomposition as well as others that can easily be deduced from
the tabulated thermodynamic data in our appendix.
\subsection{Thermodynamics of the Hydrogen Desorption at 3 Bar Hydrogen
Pressure}
From the results above it follows that kinetics is not the only problem
with Mg(BH$_4$)$_2$, as the reaction from Mg(BH$_4$)$_2$\ to MgB$_2$ has a calculated critical
temperature of 360~K (86$^{\circ}$C). Note that this critical
temperature was calculated at 1~bar~H$_2$ pressure, while the minimum
DOE target for delivery pressure is around 3~bar; the critical
temperature at 3 bar is 430~K (157$^{\circ}$C), well outside the optimal
thermodynamic window. Furthermore, while Zuttel et al.\ have argued that
the entropic contribution to metal hydride reactions is approximately
130~J/K/mol~H$_2$ for most simple metal-hydrogen
systems,\cite{Zuttel_2003:libh4_new} lower values can occur for complex
metal hydrides, such as 97~J/K/mol~H$_2$ for
$\textrm{LiBH}_4\rightarrow\textrm{LiH}+\textrm{B}+3/2\textrm{H}_2$.\cite{Alapati_2007:using_first}
It seems likely then that Mg(BH$_4$)$_2$\ also has a lower than normal entropy of
reaction, in agreement with our calculated values in
Table~\ref{tab:react_values}, given its similarity to LiBH$_4$. Because
of this comparatively low value for the reaction entropy, even the
commonly cited value of 40~kJ/mol H$_2$ is too high for a practical PEM
fuel cell. In fact, using our calculated entropy value, for a minimum
delivery pressure of 3 bar and temperature of 80$^\circ$C, the maximum
desired enthalpy is 33~kJ/mol~H$_2$. Also note that while the estimated
desired reaction enthalpy for a hydrogen storage material is
20--50~kJ/mol~H$_2$ (for reactions with low to high entropic
contributions), the ideal range for efficiency is
20--30~kJ/mol~H$_2$.\cite{Yang_2010:high_capacity} All of this points
towards the conclusion that Mg(BH$_4$)$_2$\ needs to be destabilized with respect
to its reaction products in order to achieve the ideal thermodynamic
reaction window. E.g.\ in the case of 1/3 Zn alloying, we find a
reaction enthalpy of 25.3~kJ/mol~H$_2$ and a critical temperature of
270~K ($-3^{\circ}$C) at 3 bar, both in the ideal range for a PEM fuel
cell. This is based on the suspected decomposition reaction
$\textrm{Mg$_{1-x}$Zn$_x$(BH$_4$)$_2$}\rightarrow\textrm{Mg$_{1-x}$B$_{2(1-x)}$}
+ \textrm{Zn$_x$}+2\textrm{B$_x$}+ 4\textrm{H$_2$}$. We find this to be
the most likely reaction as Zn(BH$_4$)$_2$ ``decomposes directly to
elemental Zn due to instabilities of Zn hydride and
boride,''\cite{Nakamori_2006:correlation_between} and we find Zn-alloyed
MgB$_2$ to be unstable and expect it to phase separate to MgB$_2$, Zn,
and B based on preliminary calculations. Of course, for high
concentrations of Zn, we expect the formation of diborane, as seen
experimentally by Albanese et al.\ and Kalantzopoulos et
al.,\cite{Albanese_2013:theoretical_experimental,
Kalantzopoulos_2014:hydrogen_storage} but for low concentrations of Zn
it is known that diborane is not formed.
\subsection{Kinetics of the Hydrogen Desorption and overall Desorption
Temperature}
We now switch to a discussion of the kinetic barrier. By comparing our
calculated critical temperatures $T_c$ with the known experimental
temperatures of hydrogen desorption
$T_d$,\cite{Li_2008:dehydriding_rehydriding, Yan_2011:formation_process,
Soloveichik_2009:magnesium_borohydride,
Jeon_2006:mechanochemical_synthesis} we can make a rough estimate of the
kinetic barrier, given in Table~\ref{tab:react_values}. Interestingly,
from Eq.~(\ref{equ:fit_H}) we see that---by definition---borohydrides
with $\Delta H_r = 0$ have a kinetic barrier of 423.4~K; using the same
argument but a slightly different fit (see footnote \citenum{T_d_fit}),
we find a kinetic barrier of 439.4~K. Borohydrides with $\Delta H_r
\approx 0$, such as Zn(BH$_4$)$_2$, thus must have a kinetic barrier of
the same order of magnitude, in very good agreement to our calculated
value of 445~K for Zn.
From Table~\ref{tab:react_values} we see that the kinetic barrier for
the full hydrogen release reaction is almost the same for Mg(BH$_4$)$_2$\ and
Zn(BH$_4$)$_2$, which in turn means that differences in calculated
critical temperatures in Table~\ref{tab:react_values} approximately
result in the same differences for the overall desorption temperature.
We thus estimate a decrease of desorption temperature of Mg(BH$_4$)$_2$\ when
alloying with Zn for Reaction 1 of approximately 123~K for 1/3 Zn
concentration and 262~K for 2/3 Zn. Note that we obtain almost identical
numbers using a different argument: Simply evaluating the experimental
relationship in Eq.~(\ref{equ:fit_H}), purely based on our calculated
changes in enthalpy, gives a lowering of the desorption temperature of
125~K for 1/3 Zn concentration and 250~K for 2/3 Zn.
From our results we see that the formation of the intermediate
MgB$_{12}$H$_{12}$ is mainly responsible for the kinetic barrier, and
methods to improve the kinetics of borohydrides should focus on reaction
pathways avoiding the formation of this intermediate, as suggested by
several other groups.\cite{Yan_2011:formation_process,
Severa_2010:direct_hydrogenation,
Soloveichik_2009:magnesium_borohydride} Because the kinetic barrier is
roughly constant, at least for the full hydrogen release reaction, we do
not expect alloying to have a direct impact on the height of the kinetic
barrier. But, it is really the overall energy required to start the
hydrogen desorption that is of interest---and that linearly decreases
with Zn concentration. Note that the reaction kinetics can, at least in
principle, be accelerated by using catalysts or by controlling the
particle size of reactants.\cite{Alapati_2007:using_first} Furthermore,
it is also conceivable that the kinetics of the dehydrogenation reaction
can be influenced by the inclusion of impurities, as studied by van de
Walle's group.\cite{Hoang_2009:hydrogen-related_defects}
\section{Conclusions}
In summary, \emph{ab initio} calculations were performed to accurately
determine formation enthalpies for many different materials suspected in
the decomposition of Mg(BH$_4$)$_2$, to find reaction enthalpies for the most
likely reaction pathways, and to study the effect of Zn concentration on
the overall desorption reaction. We find that the overall
thermodynamics of the desorption reaction can be optimized by alloying
Mg(BH$_4$)$_2$\ with around 33\% Zn. We estimate that in this case the temperature
for the complete decomposition reaction is lowered by 123~K. Verifying
the kinetics explicitly through \emph{ab initio} calculations is
difficult and is the subject of ongoing research. Supported by
encouraging experimental
results,\cite{Albanese_2013:theoretical_experimental,
Kalantzopoulos_2014:hydrogen_storage} we conclude that alloying Mg(BH$_4$)$_2$\
with Zn is an ideal option for fine-tuning the desorption reaction
without sacrificing too much gravimetric storage density.
\begin{acknowledgements}
This work was supported in full by NSF Grant No.\ DMR-1145968.
\end{acknowledgements}
|
\section{Introduction}
Existence of new hadrons has been investigated for a long time,
and a few hundred hadrons have been discovered \cite{pdg-2014}.
Almost all the hadrons are understood
by the internal configurations of $q\bar q$ and $qqq$ as proposed
in the original quark model. Exotic hadrons, such as tetraquark
($qq\bar q \bar q$), pentaquark ($qqqq\bar q$), and glueball ($gg$),
have been investigated since the quark-model proposal in 1964.
For example, $f_0 (980)$ and $a_0 (980)$ are known as exotic hadron
candidates. In a native quark model, scalar mesons in the 1 GeV region
could be interpreted as
$\sigma = f_0 (600) = (u\bar u+d\bar d)/\sqrt{2}$,
$f_0 (980) = s \bar s$,
$a_0 (980) = u \bar d, \, (u\bar u-d\bar d)/\sqrt{2}, \,
d \bar u$.
Since the strange quark is heavier than up and down quarks,
these quark configurations suggest the mass hierarchy
$m (\sigma) \sim m (a_0) < m (f_0)$, which is in
contradiction to the experimental one
$m (\sigma) < m (a_0) \sim m (f_0)$, where $f_0 (980)$
is simply denoted as $f_0$. Furthermore, the strong-decay width
$f_0 \to 2\pi$ in the quark model largely overestimates
the experimental data \cite{f0-decay}.
Radiative and two-photon decay widths
also indicate multiquark structure \cite{gamma-f0-a0}.
Therefore, $f_0$ and $a_0$ could be considered
as exotic tetraquark hadrons or $K\bar K$ molecules.
The $a_0$-$f_0$ mixing intensity also provides
a clue for their $K \bar K$ compositeness \cite{sk-2014}.
Furthermore, $\Lambda (1405)$ has been considered as an exotic hadron
because the $\Lambda (1405)$ mass is anomalously light.
Although the ground-state $\Lambda$ is heavier than the nucleon
due to the heavier strange-quark mass, their lowest excitation states
with $(1/2)^-$ show the reversed mass relation
$m_{\Lambda(1405)} < m_{N(1535)}$, which is difficult
to be understood within the quark model.
It is likely to be a $\bar KN$ state.
It was pointed out in Ref. \cite{sk-2014} that a future
measurement of the $\Lambda (1405)$ radiative decay width
should constrain the $\bar KN$ compositeness.
These hadrons have been investigated in global observables such as
spins, parities, masses, and decay widths. For finding clear evidences
of their exotic signatures, we consider to use high-energy
reactions. Because quark and gluon degrees of freedom are relevant
at high energies, internal configurations are expected to become apparent.
Parton distribution functions are not directly measured
because unstable exotic hadrons cannot be used as fixed targets.
Then, it was studied that fragmentation functions could be
used for such a purpose by considering the difference between
favored and disfavored functions \cite{framentaion}.
Here, we propose to use high-energy exclusive reactions
for probing the internal structure of exotic hadron candidates.
First, we explain that the constituent counting rule can be used
for exotic hadron candidates in Sec.\,\ref{counting} by the scaling
behavior of exclusive cross sections \cite{kks-2013}.
Second, we show in Sec.\,\ref{gpd-gda}
that the generalized parton distributions (GPDs)
and generalized distribution amplitudes (GDAs) can be used
for tomography of exotic hadron candidates \cite{kk-2014}.
\section{Constituent-counting rule for hard exclusive production of
an exotic hadron}
\label{counting}
\begin{figure}[b]
\vspace{-0.2cm}
\begin{minipage}{0.48\textwidth}
\begin{center}
\includegraphics[width=3.5cm]{elastic-ab-cd.eps}
\end{center}
\vspace{-0.2cm}
\caption{Exclusive process $a+b \to c+d$ \cite{kks-2013}.}
\label{fig:exclusive-ab-cd}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}{0.48\textwidth}
\begin{center}
\includegraphics[width=4.5cm]{gluon-ex-exclusive.eps}
\end{center}
\vspace{-0.2cm}
\caption{Hard gluon exchange process for exclusive reaction
\cite{kks-2013}.}
\label{fig:hard-glun-exchange}
\end{minipage}
\end{figure}
The cross section of a large-angle exclusive scattering $a+b \to c+d$
is given by
$ d\sigma_{ab \to cd} / dt \simeq
\overline{\sum}_{pol} \, | M_{ab \to cd} |^2 / (16 \pi s^2) $,
where $s$ and $t$ are Mandelstam variables defined by
the momenta $p_h \ (h=a,\,b,\,c,\,d)$ as
$s = (p_a + p_b)^2$ and $t = (p_a - p_c)^2$.
The matrix element is expressed as
\cite{exclusive-theory}
\begin{align}
M_{ab \to cd} = & \int [dx_a] \, [dx_b] \, [dx_c] \, [dx_d] \,
\phi_c ([x_c]) \, \phi_d ([x_d])
\nonumber \\[-0.1cm]
& \ \ \
\times
H_{ab \to cd} ([x_a],[x_b],[x_c],[x_d],Q^2) \,
\phi_a ([x_a]) \, \phi_b ([x_b]) ,
\label{eqn:mab-cd}
\end{align}
where $H_{ab \to cd}$ is the partonic scattering amplitude,
and $\phi_h$ is the light-cone distribution amplitude of
the hadron $h$ as illustrated in Fig. \ref{fig:exclusive-ab-cd}.
A set of the light-cone momentum fractions,
$x_i=p_i^+/p_h^+$ with $i$-th parton momentum
$p_i$, is denoted $[x_h]$ for partons in a hadron $h$.
The high-energy behavior of the cross section is described by
the constituent counting rule in perturbative QCD.
As shown in Fig. \ref{fig:hard-glun-exchange},
quarks should share large momenta, by exchanging hard gluons,
so that they should stick together to form a hadron
in a large-angle-exclusive reaction.
Assigning hard momentum factors for the internal quarks,
gluons, and external quarks, we obtain the constituent-counting rule
for the cross section in perturbative QCD:
\begin{align}
\frac{d\sigma_{ab \to cd}}{dt} = \frac{1}{s^{\, n-2}} \, f_{ab \to cd}(t/s).
\label{eqn:cross-counting}
\end{align}
Here, the factor $n$ is the number of constituents defined by
$n = n_a+n_b+n_c+n_d$, and $f(t/s)$ is a scattering-angle
dependent function.
This expression indicates that the cross section
is proportional to $1/s^{\, n-2}$ with the number
of constituents. This scaling behavior
is called the constituent-counting rule.
The counting rule has been experimentally investigated
\cite{counting-exp}.
In various hadron two-body reactions, BNL measurements
support the scaling predicted by the counting rule.
Furthermore, it is also indicated in lepton-facility
measurements. In Fig. \ref{fig:gamma-p},
the cross section of $\gamma + p \to \pi^+ + n$ is shown
at $\theta_{cm}=90^\circ$ as a function of $\sqrt{s}$.
Here, the cross section is multiplied by the factor
$s^7$ because the number is $n-2=7$.
The figure shows typical resonance structure at low energies,
whereas the cross section is almost constant at high energies.
The data indicate that the transition from hadron degrees
of freedom to the quark and gluon ones occurs at
$\sqrt{s}=2.5$ GeV.
The scaling of the cross section could be used for
probing internal structure of an exotic hadron because
the slope is controlled by the number of constituents ($n$)
\cite{kks-2013}.
As an example, we estimate the cross section of $\Lambda (1405)$
production, $\pi^- + p \to K^0 +\Lambda (1405)$.
For this purpose, we first investigate the ground-$\Lambda$ production
$\pi^- + p \to K^0 +\Lambda$, for which there are many available data
at $\theta_{cm}=90^\circ$. The cross section data of
$\pi^- + p \to K^0 +\Lambda$ show the scaling with
$n=10.1 \pm 0.6$ at high energies, and it is consistent
with the number of constituents $n=2+3+2+3=10$.
Therefore, it is promising to use the counting rule
also for $\Lambda (1405)$ for finding whether it
is an ordinary $qqq$ state of five-quark one.
There is only one experimental data available for
the $\Lambda (1405)$ cross section at $\theta_{cm}=90^\circ$.
We use this data together with the counting rule
to show the scaling of the cross section at high energies
in Fig. \ref{fig:lambda-1405-scaling}
by assuming that the perturbative region starts at $\sqrt{s}=2$ GeV.
Two curves are shown by assuming
three- or five-quark state for $\Lambda (1405)$.
There are large differences between the two curves as
the energy becomes larger, which should be used for judging
whether $\Lambda (1405)$ is an exotic five-quark state.
Such an experiment is possible, for example, at J-PARC.
\begin{figure}[h!]
\vspace{-0.4cm}
\begin{minipage}{0.48\textwidth}
\begin{center}
\includegraphics[width=6.0cm]{s7dsdt_gamma.eps}
\end{center}
\vspace{-0.3cm}
\caption{Cross section of $\gamma + p \to \pi^+ + n$
and scaling at large energies \cite{kks-2013}.}
\label{fig:gamma-p}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}{0.48\textwidth}
\begin{center}
\includegraphics[width=5.85cm]{dsdOmega-1405.eps}
\end{center}
\vspace{-0.3cm}
\caption{Scaling of
$\pi^- + p \to K^0 + \Lambda \, (1405)$ cross section
at high energies \cite{kks-2013}.}
\label{fig:lambda-1405-scaling}
\end{minipage}
\end{figure}
\vspace{-0.8cm}
\section{Tomography of exotic hadrons by generalized parton distributions
and generalized distribution amplitudes}
\label{gpd-gda}
\begin{wrapfigure}[9]{r}{0.45\textwidth}
\vspace{-0.80cm}
\begin{center}
\includegraphics[width=5.5cm]{gpd-1.eps}
\end{center}
\vspace{-0.1cm}
\caption{Virtual Compton scattering for GPD \cite{kk-2014}.}
\label{fig:gpd-1}
\vspace{-0.5cm}
\end{wrapfigure}
We proposed that the internal quark configuration could be
found by looking at the scaling of an exclusive cross section
at high energies. For finding another independent evidence
at high energies and for probing much details of internal
structure in the three-dimensional form, we propose to
use hadron tomography by using generalized parton distributions
(GPDs) and generalized distribution amplitudes (GDAs) \cite{kk-2014}.
Recently, three dimensional structure of the nucleon has been
investigated by using the GPDs and TMDs
(transverse-momentum-dependent parton distributions)
\cite{gpd-gda-summary},
particularly for understanding the origin of nucleon spin
due to orbital angular momenta of partons.
The GPDs are measured by the virtual Compton process in Fig. \ref{fig:gpd-1}.
We define kinematical variable for expressing the GPDs.
First, the momenta $\bar P$, $\bar q$, and $\Delta$ are defined by
the nucleon and photon momenta as
$\bar P = (p+p')/2$,
$\bar q = (q+q')/2$,
$\Delta = p'-p = q-q'$.
Then, the momentum-transfer-squared quantities are given by
$Q^2 = -q^2$, $\bar Q^2 = - \bar q^2$, and $t = \Delta^2$.
The generalized scaling variable $x$ and a skewdness parameter $\xi$
are defined by
$x = Q^2/(2p \cdot q)$ and
$\xi = \bar Q^2 /(2 \bar P \cdot \bar q)$.
The variable $x$ indicates the lightcone momentum fraction
carried by a quark in the nucleon.
The skewdness parameter $\xi$ or the momentum $\Delta$ indicates
the momentum transfer from the initial nucleon to the final one
or the one between the quarks.
\begin{figure}[b]
\vspace{-0.2cm}
\begin{minipage}{0.48\textwidth}
\begin{center}
\includegraphics[width=5.0cm]{exotic-pdf2.eps}
\end{center}
\vspace{-0.1cm}
\caption{PDFs of exotic hadrons in comparison
with parametrizations of pion and proton PDFs.
The solid curves are calculated by using a simple
function suggested by the constituent counting rule \cite{kk-2014}.
}
\label{fig:exotic-pdfs}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}{0.48\textwidth}
\begin{center}
\includegraphics[width=5.0cm]{2d-form-q2t.eps}
\end{center}
\vspace{-0.0cm}
\caption{Transverse form factors for $x=0.2$, 0.4 and
$\Lambda=0.5$, 1.0 GeV \cite{kk-2014}.}
\label{fig:transverse-form}
\end{minipage}
\end{figure}
The GPDs for the nucleon are defined by off-forward matrix elements
of quark (and gluon) operators with a lightcone separation
between nucleonic states:
\begin{align}
& \! \! \! \! \! \! \! \!
\int\frac{d y^-}{4\pi}e^{i x \bar P^+ y^-}
\left< p' \left|
\overline{\psi}(-y/2) \gamma^+ \psi(y/2)
\right| p \right> \Big |_{y^+ = \vec y_\perp =0}
\nonumber \\
& \! \! \! \! \! \! \! \! \! \! \! \!
= \frac{1}{2 \bar P^+} \, \overline{u} (p')
\left [ H_q (x,\xi,t) \gamma^+
+ E_q (x,\xi,t) \frac{i \sigma^{+ \alpha} \Delta_\alpha}{2 \, M}
\right ] u (p) .
\label{eqn:gpd-n}
\end{align}
Here, $\sigma^{\alpha\beta}$ is
$\sigma^{\alpha\beta}=(i/2)[\gamma^\alpha, \gamma^\beta]$,
and the unpolarized GPDs of the nucleon are
$H_q (x,\xi,t)$ and $E_q (x,\xi,t)$.
The quark field is denoted as $\psi(y/2)$, and
$u (p)$ is the Dirac spinor.
The GPDs contain information on both longitudinal
momentum distributions of partons and transverse
form factors. It can be seen first by taking
the forward limit ($\Delta,\, \xi,\, t \rightarrow 0$):
$H_q (x, 0, 0) = q(x)$,
where $q(x)$ is an unpolarized parton distribution function (PDF)
in the nucleon.
Next, their first moments are the form factors of the nucleon:
$ \int_{-1}^{1} dx H_q(x,\xi,t) = F_1 (t)$ and
$ \int_{-1}^{1} dx E_q(x,\xi,t) = F_2 (t)$.
Because of these two ingredients, a useful functional form of
the GPDs
\begin{align}
H_q^h (x,\xi=0,t)= f_n (x) \, F_n^h (t, x) ,
\label{eqn:gpd-paramet1}
\end{align}
is often used. Here, $f_n (x)$ is a longitudinal PDF,
and $F_n^h (t, x)$ is a transverse form factor at $x$.
Because there is no stable exotic hadron, the GPDs cannot be
directly measured except for transition GPDs, for example,
for $p \to \Lambda (1405)$. However, we consider
a {\it gedankenexperiment} by which the PDFs or the GPDs
of the exotic hadrons can be obtained
by assuming a stable target.
A simple form of the PDFs is given by
$ f_n (x) = C_n \, x^{\alpha_n} \, (1-x)^{\beta_n} $,
where the parameters are constrained by
the valence-quark number
$ \int_0^1 dx \, f_n (x) = n$
and the quark momentum
$ \int_0^1 dx \, x \, f_n (x) = \langle x \rangle_q$.
The parameter $\beta_n$ could be determined by
the constituent counting rule
as $\beta_n = 2n -3+2\Delta S_z$ with
the spin factor $\Delta S_z=|S_z^q-S_z^h|$
because the factor
$(1-x)^{\beta_n}$ controls the behavior in the elastic
limit $x \to 1$.
Using this $\beta_n$, the number of constituents $n$,
and the momentum fraction $\langle x \rangle_q$,
we obtain $C_n$ and $\alpha_n$.
The calculated PDFs are shown in Fig. \ref{fig:exotic-pdfs}
for $n=2$ (meson), 3 (baryon), 4 (tetraquark), and 5 (pentaquark).
In comparison, typical PDFs determined from experimental data
are shown by the dashed curves for the pion and nucleon
at $Q^2=2$ GeV$^2$.
Although differences between the dashed and solid curves
vary depending on the $Q^2$ value, we obtain a reasonable
agreement. It means that the PDFs obtained by the counting rule
are reasonable magnitude estimates.
As the valence-quark number becomes larger ($n=4$, 5),
the distribution shifts toward the small-$x$ region.
In Fig. \ref{fig:transverse-form}, transverse form factors
are shown by assuming the exponential form
$ F_n^h (t,x) = e^{(1-x) t/(x \Lambda^2)} $
with the cutoff parameter constrained by the transverse size as
$ \langle r_\perp^2 \rangle = 4(1-x) / (x \Lambda^2) $.
The $q_\perp^2$ dependence changes significantly whether
the hadron has compact $q\bar q$ ($qqq$)-like structure
or diffuse tetraquark (pentaquark, hadron molecule) one.
\begin{wrapfigure}[12]{r}{0.43\textwidth}
\vspace{-0.9cm}
\begin{center}
\includegraphics[width=5.5cm]{gda-1.eps}
\end{center}
\vspace{-0.3cm}
\caption{$\gamma^* \gamma \to h \bar h$ process for GDA \cite{kk-2014}.}
\label{fig:gda-1}
\vspace{-0.5cm}
\end{wrapfigure}
The GDAs are defined in the same way with the GPDs
in the $s$-$t$ crossed channel as shown in Fig. \ref{fig:gda-1}.
They describe the production of a hadron pair $h\bar h$,
$\gamma^* \gamma \to h \bar h$. The final-hadron momenta are
denoted $p$ and $p'$, the initial photon momenta are
$q$ ($Q^2=-q^2$) and $q'$ (${q'}^2=0$).
$P$ is the total momentum $P=p+p'$, and $k$ is the quark momentum.
The center-of-mass (c.m.) energy squared $s$ is equal to
the invariant-mass squared $W^2$ of the final hadron pair,
$s= (q+q')^2 = (p+p')^2 = W^2$.
Then, the variable $\zeta$ is defined by
$ \zeta = p \cdot q' / (P \cdot q')
= p^+ / P^+ = (1+\beta \cos\theta)/2 $,
where $\beta$ is the velocity of a hadron,
$ \beta = |\vec p \, | / p^0 = ( 1-4 m_h^2 / W^2 )^{1/2} $,
with the final hadron mass $m_h$, and
$\theta$ is the scattering angle in the c.m. frame.
The GDAs are expressed
by these three variables ($z$, $\zeta$, $s=W^2$).
The GDAs are defined by the same lightcone operators
between the vacuum and the hadron pair $h\bar h$:
\begin{align}
\Phi_q^{h\bar h} (z,\zeta,s)
= \int \frac{d y^-}{2\pi}\, e^{i (2z-1)\, P^+ y^-}
\langle \, h(p) \, \bar h(p') \, | \,
\overline{\psi}(-y/2) \gamma^+ \psi(y/2)
\, | \, 0 \rangle \Big |_{y^+=\vec y_\perp =0} \, ,
\end{align}
for a quark. The gluon GDA is defined in the similar way.
The GDAs are related to the GPDs by the $s$-$t$ crossing
if the factorizations can be applied as shown
in Figs. \ref{fig:gpd-1} and \ref{fig:gda-1}.
The crossing means to move the final state $\bar h$ ($p'$)
to the initial $h$ ($p$). It indicates that the momenta ($p$, $p'$)
of the GDAs should be replaced by ($p'$, $-p$) in the GPDs.
Then, the relations between the variables are given by
$ z \leftrightarrow (1-x/\xi)/2$,
$ \zeta \leftrightarrow (1-1/\xi)/2$, and
$ W^2 \leftrightarrow t$,
which indicates that the GDAs are related to the GPDs by
\begin{align}
\Phi_q^{h\bar h} (z,\zeta,W^2)
\longleftrightarrow
H_q^h \left ( x=\frac{1-2z}{1-2\zeta},
\xi=\frac{1}{1-2\zeta}, t=W^2 \right ) .
\label{eqn:gda-gpd-relation}
\end{align}
Because there are many studies on the GPDs, this relation seems to
be useful for estimating the GDAs. However, we find that the relevant
kinematical regions are
$ 0 \le |x| < \infty$, $ 0 \le |\xi| < \infty$,
$ |x| \le |\xi|$, $ t \ge 0 $,
which are not necessarily the usual physical regions of the GDAs,
so that the information of the GPDs is not used directly
for the GDAs.
The cross section for $e\gamma \to e h \bar h$ is given by
\cite{diehl-2000}
\begin{align}
\frac{{d\sigma }}{{d{Q^2}d{W^2}d\cos \theta}}
= \frac{\beta \, \alpha ^3}{8 s_{e\gamma}^2 \, Q^2 (1 - \varepsilon ) }
{\left| {{A_{+ +}}(\zeta ,{W^2})} \right|^2} , \ \
A_{++} = \sum\limits_q {\frac{{e_q^2}}{2}} \int_0^1 {dz}
\frac{{2z - 1}}{{z(1 - z)}}\Phi _q^{h \bar h}(z,\zeta ,{W^2}) ,
\label{eqn:corss-2}
\end{align}
in the leading order of $\alpha_s$ by neglecting gluon GDAs.
Here, $A_{+ +}$ is the helicity amplitude
$ A_{i j}= \varepsilon _\mu ^{( i )}(q) \,
\varepsilon _\nu ^{( j )}(q') \, {T^{\mu \nu }} /e^2 $
for the hadron tensor $T^{\mu \nu }$ of $\gamma^* \gamma \to h \bar h$, and
$\Phi _q^{h \bar h}$ is a quark GDA.
In oder to find possible signatures of exotic hadrons, we
use a simple function
\begin{align}
\Phi_q^{h\bar h (I=0)} (z,\zeta,W^2)
= N_{h(q)} \, z^\alpha (1-z)^\beta (2z-1) \, \zeta (1-\zeta) \, F_{h(q)} (W^2) ,
\label{eqn:gda-paramet}
\end{align}
which satisfies the sum rules for the quark GDAs for
the isospin $I=0$ two-meson final states:
\begin{align}
\int_0^1 dz \, \Phi_q^{h\bar h (I=0)} (z,\zeta,W^2) = 0, \ \
\int_0^1 dz \, (2z -1) \, \Phi_q^{h\bar h (I=0)} (z,\zeta,W^2)
= - 2 M_{2(q)}^h \zeta (1-\zeta) F_{h(q)} (W^2) ,
\label{eqn:gda-sum-I=0}
\end{align}
\begin{wrapfigure}[12]{r}{0.42\textwidth}
\vspace{-0.80cm}
\begin{center}
\includegraphics[width=5.6cm]{cross-form.eps}
\end{center}
\vspace{-0.35cm}
\caption{Form-factor effects on the cross section
as a function of $W^2$ at $Q^2=$10 GeV \cite{kk-2014}.}
\label{fig:cross-form}
\vspace{-0.5cm}
\end{wrapfigure}
\noindent
where $M_{2(q)}^h$ is the momentum fraction carried by quarks.
The function $F_{h(q)} (W^2)$ is a form factor of the quark part
of the energy-momentum tensor, and it may be taken in the form
suggested by the counting rule:
$ F_{h(q)} (W^2)
= 1 / [ 1 + (W^2-4 m_h^2)/\Lambda^2 ]^{n-1} $,
where $n=2$ for ordinary $q\bar q$ mesons and $n=4$
for tetraquark hadrons.
Using the form factor together with the GDA expression in
Eq. (\ref{eqn:gda-paramet}), we obtain the cross section
for $e+\gamma \to e + h + \bar h$ as the function of $W^2$
in Fig. \ref{fig:cross-form} by considering $h=f_0 (980)$ or $a_0 (980)$
\cite{kk-2014}.
Depending on the size and the constituent number, the cross section
varies much as the function of $W^2$. Therefore,
measurements of the $e\gamma \to e' h\bar h$ cross section
should be valuable for determining internal structure of
exotic hadrons through the GDAs.
\section{Summary}
\vspace{-0.2cm}
Exotic hadron candidates have been investigated in hadron spectroscopy and their
decays; however, the high-energy processes can probe their internal structure.
In this work, we studied exclusive reactions by using the idea of
constituent counting rule and by the hadron tomography
with the GPDs and GDAs. Because the quark and gluon degrees of freedom
are relevant at high energies, it should be more appropriate to use
the high-energy exclusive reactions for probing the internal configurations
of exotic hadron candidates.
\vspace{-0.0cm}
\section*{Acknowledgements}
\vspace{-0.2cm}
This work was supported by the MEXT KAKENHI Grant Number 25105010.
|
\section{Introduction}
Although previously considered a system in equilibrium, observations of the Hydra I cluster core, dominated by NGC 3311, have shown instead a complex, ongoing process of assembly of matter. Recently, \citet{2012A&A...545A..37A} reported the existence of an faint surface brightness offset envelope. In part I of our work (see Hilker at al. in this proceedings), we have shown that the envelope substructure have line-of-sight velocity distributions distinct from the surrounding regions, which is also related to a group of infalling galaxies.
\citet{2011A&A...533A.138C} has shown that the stellar content in the region of the offset envelope is different from the other regions, thus attesting that stellar populations can be used to trace and constrain the processes involved in galaxy disruption and in the build up of the massive cD galaxies. However, a spatially resolved study to observe the shape and extent of the stellar population features was still missing. Therefore, here we aim to further demonstrate that the Hydra I core is indeed actively forming using stellar populations as tracers of the past evolution of the system.
\section{Data and Methods}
We have used our novel FORS2/VLT spectroscopic data set (programme 88.B-448; PI: Richtler), i.e. short slits placed in an onion shell-like pattern onto NGC 3311 to mimic a coarse `IFU'. The data set consists of high resolution optical spectra for 130 positions out to $\sim$30 kpc ($\sim$3 effective radii). We modelled seven absorption
line features in the Lick/IDS system with the alpha-enhanced models of \citet{2011MNRAS.412.2183T} using a Monte Carlo Markov Chains method in order to obtain uncertainties and to access whether we can break the well known metallicity-age degeneracy.
\section{Results and Conclusion}
We present results for the luminosity-weighted ages, metallicities, and alpha element abundances in Fig. 1. The offset envelope exhibits populations which are slightly younger, more metal-rich and have lower alpha abundances. These results are consistent with the "two stages" formation scenario of central galaxies: the central parts form first "in situ", possibly in a quasi monolithic collapse with violent starbursts. The outer regions then grow by accreting less massive systems with extended periods of star formation.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{barbosa_fig1a.eps}
\includegraphics[width=\linewidth]{barbosa_fig1b.eps}
\caption{Luminosity-weighted stellar populations of the Hydra I cluster core: ages (left), metallicity (centre) and alpha-elements content (right). Upper and lower panels display contours for the V-band image and the offset envelope respectively, as in \citet{2012A&A...545A..37A}.}\label{fig:fig1}
\end{figure}
\section*{Acknowledgements}
\noindent
Based on observations made with ESO Telescopes at the La Silla Paranal Observatory. CEB and CMdO are grateful to the S\~{a}o Paulo Research Foundation (FAPESP)
funding (Procs. 2006/56213-9, 2011/21325-0 and 2012/22676-3). TR acknowledges support from FONDECYT project Nr.\,1100620, the BASAL Centro de Astrof\'isica y Tecnolog\'ias Afines (CATA) PFB-06/2007, and a visitorship at ESO/Garching.
|
\section{Introduction}
The description of CP violation in the Standard Model demands an
accurate knowledge of the Cabibbo-Kobayashi-Maskawa (CKM) quark mixing
matrix. Being the b-sector the one known with lesser precision,
a precise quantitative study of the weak decays of $B$ and $B_s$
mesons is needed. In this contribution we summarise our studies of the
semileptonic decays $B_s$ into $K$ and $D_s$ estates, and some of
the nonleptonic decays studied in
Refs.\cite{Albertus:2014gba, Albertus:2014bfa}, evaluated
in the context of a constituent quark model.
\section{Semileptonic $B_s\to K$ decay}
The hadronic matrix element for this decay can be parametrised in
terms of the form factors $f_+$ and $f_0$. If we neglect the mass of
the leptons, only $f_+$ contributes to the differential decay width
\begin{eqnarray}
\frac{d\Gamma}{dq^2}=\frac{G_F^2}{192\pi^3}\,|V_{ub}|^2\,
\frac{\lambda^{3/2}(q^2,M_{B_s}^2,M_K^2)}{M_{B_s}^3}\,f_+^2(q^2)
\label{eq:dgdq2}
\end{eqnarray}
with $G_F$ being the Fermi constant, $|V_{ub}|$ the modulus of the
corresponding CKM matrix element and $\lambda(a,b,c)=a^2
+b^2+c^2-2ab-2ac-2bc$.
We evaluate the valence quark contribution to the form factor that we supplement
with a $B^*$-pole one to improve its behaviour at high $q^2$ values
\cite{Albertus:2014gba}. To
extend the above predictions beyond its region of applicability (near
$q^2_{\rm max}$), we adopt a multiply-subtracted Omnes dispersion
relation \cite{Albertus:2006gx, Flynn:2007ii, Flynn:2007qd, Flynn:2006vr},
and we take
\begin{equation}
f^+(q^2)\approx\frac{1}{M^2_{B^*}-q^2}\prod_{j=0}^n\Big[f_+(q_j^2)
\Big(M^2_{B^*}-q^2_j\Big)\Big]^{\alpha_j(q^2)}, \alpha_j(q^2)=\prod_{j\ne k=0}^n\frac{q^2-q_k^2}{q_j^2-q_k^2}
\label{eq:omnes}
\end{equation}
for $q^2 < s_{th}=(m_{B_s}+m_K)^2$ and where the different $q_j^2 \in
]-\infty,s_{th}[$ are the different subtraction points. The values of
the $f_+$ form factor at the subtraction points are taken as free
parameters that we fit to our quark model results (valence plus $B^*$-pole)
at high $q^2$ and previous light cone sum rules (LCSR) results in the low
$q^2$ region \cite{Duplancic:2008tk}. We take the subtraction points at $q^2=0,
q^2=\frac{q^2_{\rm max}}{3},q^2=\frac{2q^2_{\rm max}}{3}$ and
$q^2=q^2_{\rm max}$.
\begin{figure}[ht]
\centering
\includegraphics[width=4cm,clip]{conrado0.eps}
\includegraphics[width=4.4cm,clip]{conrado1.eps}
\caption{Left panel: Comparison of the different approaches to the
$f_+$ form factor. Right panel: Differential decay width calculated
with a 68\% confidence level band obtained with our fitting
procedure}
\label{fig1}
\end{figure}
In the left panel of Fig.~\ref{fig1} we compare the $f_+$ form factor obtained
in our combined approach with the ones
calculated using LCSR techniques \cite{Duplancic:2008tk} and lattice
QCD~\cite{Bouchard:2014ypa}.
The latter are an extrapolation of the lattice
points obtained in Ref.~\cite{Bouchard:2013zda} which are also shown.
In the right panel, we plot the differential
decay width. We compare our results with the ones in Ref.~\cite{Li:2001yv},
obtained in a LCSR+$B^*$-pole calculation, and in
Ref.~\cite{Bouchard:2014ypa}, obtained in lattice QCD.
In both panels we also include a 68\%
confidence level band for our predictions.
The total decay width that we get is $
\Gamma(B_s\to Kl^+\nu_l)=(5.47^{+0.54}_{-0.46})|V_{ub}|^2\times 10^{-9}\,{\rm
MeV}$.
This is to be compared with the results
$(4.63^{+0.97}_{-0.88})|V_{ub}|^2\times
10^{-9}\,{\rm MeV}$~\cite{Li:2001yv}, where we have propagated a 10\% uncertainty
in the
form factor, and $(5.1 \pm 1.0)|V_{ub}|^2\times
10^{-9}\,{\rm MeV}$~\cite{Bouchard:2014ypa}. The three estimates are compatible
within uncertainties. More details are given in Ref.~\cite{Albertus:2014gba}.
\section{Semileptonic $B_s\to D^{(*)}_{s}$ decays}
We have considered the semileptonic decays of $B_s$ meson into $D^{(*)}_{s}$
states with $J^\pi$ quantum numbers $0^-$, $0^+$, $1^-$, $1^+$, $2^-$
and $2^+$. The form factor decomposition required for each channel can
be found in Ref.~\cite{Albertus:2014bfa}. We have adopted the helicity formalism of
Ref.~\cite{Ivanov:2005fd} to compute the contraction of the leptonic
and hadronic tensors. Expressions for the helicity amplitudes can be
found in Ref.~\cite{Albertus:2014bfa}
\begin{table}
\centering
\caption{Branching fractions for the indicated decay channels, in percentage.\label{tab:bra1}}
\begin{tabular}{c|cc}
\hline
$M'$ & $l=e,\mu$ & $l=\tau$ \\\hline
$D_s^+ $ & 2.32 & 0.67 \\
$D_{s0}^{*+} $ & 0.39 & 0.04 \\
$D_s^{*+} $ & 6.26 & 1.53 \\
$D_{s1}^+(2460)$ & 0.47 & 0.04 \\
$D_{s1}^+(2536)$ & 0.32 & 0.03 \\
$c\bar s(2^-) $ & 9.2\ten{3} & 2.0\ten{4} \\
$D_{s2}^{*+} $ & 0.44 & 0.03 \\\hline
\end{tabular}
\end{table}
In Table~\ref{tab:bra1} we show our results for the branching ratios.
As shown in
Table V of Ref.~\cite{Albertus:2014bfa} the results of
this work are in good agreement with those from Ref.~\cite{Faustov:2012mt},
obtained in a relativistic quark model approach. The agreement is also good
with the quark model calculation of Ref.~\cite{Zhao:2006at}. Our results
for decays into orbitally excited final $D_s^*$ mesons agree with our
previous results from Ref.~\cite{segovia:2011dg}, though in that work the
potential model that has been used was much more sophisticated. Our results also compare well to the
sum-rules calculation of Refs.~\cite{Azizi:2008xy, Azizi:2008vt}, while the
result of Ref.~\cite{Blasi:1993fi} is lower by a factor of two. The same
happens if we compare with the results of Ref.~\cite{Chen:2012fi} or
Ref.~\cite{Li:2009wq}. In Ref.~\cite{Albertus:2014bfa} we also check our
results against Heavy Quark Symmetry predictions.
\section{$\bar B_s \to c\bar s M_F$}
We have also calculated the decay width for
two-meson nonleptonic reactions $\bar B_s \to c\bar s M_F$ where $M_F$
is a light pseudoscalar or
vector meson. These decays correspond to a $b \to c$ transition at the
quark level. These transitions are governed, neglecting penguin
operators, by the effective Hamiltonian of Eq.~(53) of
Ref~\cite{Albertus:2014bfa}. See
Refs.~\cite{Ebert:2006nz,Beneke:1999br} for further details. We shall
work in the factorisation approximation, i. e., the hadron matrix
elements of the effective Hamiltonian are evaluated as a product of
quark-current matrix elements. One of these is the matrix element of
the $B_s$ transition to one of the final mesons, while the other
is determined by the decay constant of the other meson. In
Table~\ref{tab:brcomp2} we compare our calculations with previous
results and experimental data when available. More results are
shown in Ref.~\cite{Albertus:2014bfa}.
\begin{table}
\begin{tabular}{lccccccc}
\hline\hline
& This work & \cite{Faustov:2012mt} & \cite{Blasi:1993fi} & \cite{Chen:2012fi} &\cite{Li:2009wq}
& Experiment \cite{PhysRevD.86.010001}\\\hline
$\bar B_s \to D_s^+ \pi^-$ & 0.53 & 0.35 & 0.5 & $0.27_{-0.03}^{+0.07}$ & $0.17_{-0.06}^{+0.07}$ & $0.32\pm0.4$ \\
$\bar B_s \to D_s^+ \rho^-$ & 1.26 & 0.94 & 1.3 & $0.64_{-0.11}^{+0.17}$ & $0.42_{-1.4}^{+1.7}$ & $0.74\pm0.17$\\
$\bar B_s \to D_s^+ K^-$ & 0.04 & 0.028 & 0.04 & $0.021_{-0.002}^{+0.002}$ & $0.013_{-0.004}^{+0.005}$&& \\
$\bar B_s \to D_s^+ K^{*-}$ & 0.08 & 0.047 & 0.06 & $0.038_{-0.005}^{+0.005}$ && \\
$\bar B_s \to D_{s0}^{*+} \pi^-$ & 0.10 &0.09 & $0.052_{-0.021}^{+0.25}$ &&&\\
$\bar B_s \to D_{s0}^{*+} \rho^-$ & 0.27 &0.22 & $0.013_{-0.05}^{+0.06}$ &&&\\
$\bar B_s \to D_{s0}^{*+} K^-$ & 0.009 &0.007 & $0.004_{-0.002}^{+0.002}$ &&&\\
$\bar B_s \to D_{s0}^{*+} K^{*-}$ & 0.16 &0.012 & $0.008_{-0.003}^{+0.004}$ &&&\\
$\bar B_s \to D_s^{*+} \pi^-$ & 0.45 & 0.27 & 0.2 & $0.31_{-0.02}^{+0.03}$ & & $0.21\pm0.06$\\
$\bar B_s \to D_s^{*+} \rho^-$ & 1.35 & 0.87 & 1.3 & $0.9_{-1.5}^{+1.5}$ & & $1.03\pm2.6$ \\
$\bar B_s \to D_s^{*+} K^-$ & 0.04 & 0.021 & 0.02 & $0.024_{-0.002}^{+0.002}$ & \\
$\bar B_s \to D_s^{*+} K^{*-}$ & 0.08 & 0.048 & 0.06 & $0.056_{-0.007}^{+0.006}$ & \\
$\bar B_s \to D_{s1}^+(2460) \pi^-$ &0.15 &0.19 & & &&\\
$\bar B_s \to D_{s1}^+(2460) \rho^-$ &0.36 &0.49 & & &&\\
$\bar B_s \to D_{s1}^+(2460) K^-$ &0.012 &0.014 & & &&\\
$\bar B_s \to D_{s1}^+(2460) K^{*-}$ &0.020 &0.026 & & &&\\
$\bar B_s \to D_{s1}^+(2536) \pi^-$ &0.07 &0.029 & & &&\\
$\bar B_s \to D_{s1}^+(2536) \rho^-$ &0.19 &0.083 & & &&\\
$\bar B_s \to D_{s1}^+(2536) K^-$ &0.0054 &0.0021 & & &&\\
$\bar B_s \to D_{s1}^+(2536) K^{*-}$ &0.01 &0.0044 & & &&\\
$\bar B_s \to (2^-)^+ \pi^-$ &7.1\ten{5} &&&&&\\
$\bar B_s \to (2^-)^+ \rho^-$ &0.0047 &&&&&\\
$\bar B_s \to (2^-)^+ K^-$ &5.2\ten{6} &&&&&\\
$\bar B_s \to (2^-)^+ K^{*-}$ &2.2\ten{8} &&&&&\\
$\bar B_s \to D_{s2}^{*+} \pi^-$ &0.1 &0.16 & & &&\\
$\bar B_s \to D_{s2}^{*+} \rho^-$ &0.27 &0.42 & & &&\\
$\bar B_s \to D_{s2}^{*+} K^-$ &0.008 &0.012 & & &&\\
$\bar B_s \to D_{s2}^{*+} K^{*-}$ &0.016 &0.022 & & &&\\
\hline
\end{tabular}
\caption{\label{tab:brcomp2} Branching ratios for the decays above.}
\end{table}
\begin{acknowledgement}
This research was supported by the Spanish Ministerio de Econom\'\i a
y Competitividad and European FEDER funds under Contracts
Nos. FPA2010-21750-C 02-02, FIS2011-28853-C02-02, and CS D2007-00042, by
Generalitat Valenciana under Contract No. PROMETEO/20090090, by Junta
de Andaluc\'\i a under Contract No. FQM-225, by the EU HadronPhysics3
project, Grant Agreement No. 283286, and by the University of Granada
start-up Project for Young Researchers contract No. PYR-2014-1. C.A.
wishes to acknowledge a CPAN postdoctoral contract and C.H.
-D. thanks the support of the JAE-CSIC Program.
\end{acknowledgement}
|
\section{Introduction}
I wish to congratulate Professor Imbens on a lucid and erudite review of
the instrumental variable literature. The paper contrasts an econometric
view of instrumental variable models, where treatment confounding is due to
agents rationally choosing an optimal treatment for their situation, and
the statistical view, where treatment confounding arises due to noncompliance,
unobserved baseline differences between individuals, or other such issues.
While the paper does an admirable job describing the statistics view of
the instrumental variables based on the potential outcome model of Neyman
and Rubin, it does not much discuss the growing statistics literature on causal
graphical models, except to mention that causal graphs are a useful tool for
displaying the exclusion restriction assumption crucial for the use of
instrumental variables.
I would like to give a brief and hopefully complementary account of how
causal graphical models serve to clarify
and help address the issues of confounding (what Heckman calls the selection
problem) that make causal inference from
observational data such a challenging endeavor.
\section{Graphs as a General Method for Dealing with Confounding}
Causal inference in statistics has been greatly influenced by Neyman's idea of
explicitly representing interventions or forced treatment assignments on the
outcome (\cite{neyman23app}), and by Rubin's idea of using the stable unit
treatment value assumption (SUTVA) and ignorability assumptions to
equate potential outcome parameters with functionals of the observed data
(\cite{rubin74potential}).
Professor Imbens discusses these ideas at length in the paper. The essence of
Rubin's method is that assumptions on potential outcome random variables allow
one to properly adjust for the presence of confounding. Unfortunately, in
complex, possibly longitudinal settings it is not easy to see what assumptions
are needed, or whether it is even possible to identify parameters of interest
as functionals of observed data. For this task, graphical causal models,
first used by Wright in the context of animal genetics
(\cite{wright21correlation}), and expanded into a general methodology for
causal inference by Spirtes, Glymour and Scheines (\citeyear{spirtes93causation}), Pearl
(\citeyear{pearl00causality}), Robins
(\citeyear{robins86new, robins97estimation}), and others
have proven to be invaluable.
\begin{figure*}
\begin{center}
\begin{tikzpicture}[>=stealth, node distance=1.2cm]
\tikzstyle{format} = [draw, thick, circle, minimum size=5.0mm,
inner sep=0pt]
\begin{scope}
\path[->, thick]
node[format] (a) {$A$}
node[format, above right of=a, gray] (c) {$C$}
node[format, below right of=c] (y) {$Y$}
(c) edge (a)
(c) edge (y)
(a) edge (y)
node[below of=c, yshift=-0.2cm] (l) {\footnotesize{(a)}}
;
\end{scope}
\begin{scope}[xshift=3.0cm]
\path[->, thick]
node[format] (a) {$A$}
node[format, above right of=a] (c) {$C$}
node[format, below right of=c] (y) {$Y$}
(c) edge (a)
(c) edge (y)
(a) edge (y)
node[below of=c, yshift=-0.2cm] (l) {\footnotesize{(b)}}
;
\end{scope}
\begin{scope}[xshift=6.0cm]
\path[->, thick]
node[format] (z) {$Z$}
node[format, right of=z] (a) {$A$}
node[format, above right of=a, gray] (c) {$C$}
node[format, below right of=c] (y) {$Y$}
(z) edge (a)
(c) edge (a)
(c) edge (y)
(a) edge (y)
node[below of=c, yshift=-0.2cm] (l) {\footnotesize{(c)}}
;
\end{scope}
\begin{scope}[xshift=10.2cm]
\path[->, thick]
node[format] (a) {$A$}
node[format, above right of=a, gray] (c) {$C$}
node[format, below right of=c] (y) {$Y$}
node[format, right of=a, xshift=-0.35cm] (w) {$W$}
(c) edge (a)
(c) edge (y)
(a) edge (w)
(w) edge (y)
node[below of=c, yshift=-0.2cm] (l) {\footnotesize{(d)}}
;
\end{scope}
\end{tikzpicture}
\end{center}
\caption{\textup{(a)} The standard problem of causal inference---an unobserved
confounder $C$, and possible approaches to the problem using observational
data. \textup{(b)} Observing
the confounder and adjustment/stratification methods.
\textup{(c)} $Z$ as an instrumental variable. \textup{(d)} A strong independent
mediator as an ``instrument'' for identification.}\label{fig:confounding}
\end{figure*}
Consider Figure~\ref{fig:confounding}(a), where vertices represent random
variables of interest: a treatment $A$, an outcome $Y$, and a source of
unobserved confounding $C$
(lightly shaded in the graph to represented unobservability).
Following Neyman, we quantify the causal effect of $A$ on $Y$ by means of a
function of the distribution of the potential outcome $Y(a)$ ($Y$ after we
force $A$ to a value $a$). For instance, we may use the average causal
effect (ACE): $E[Y(a)] - E[Y(a')]$, where $a$ is the active treatment value,
and $a'$ is the baseline treatment value. We are interested in using
observed data to make inferences about such effects, which entails
dealing with confounding in some way. The assumptions underlying this
graph which we will use can be expressed in terms of potential outcomes if
desired. For example, the \emph{finest fully randomized causally interpretable
structured tree graph} (FFRCISTG) model of Robins (\citeyear{robins86new})
corresponding to this graph states that for all value assignments
$a$ and $c$ to $A$ and $C$, random variables $C$, $A(c)$ and $Y(a,c)$
are mutually independent, while the
\emph{nonparametric structural equation model with independent errors}
(NPSEM-IE) of Pearl (\citeyear{pearl00causality}) corresponding to this graph
states that for all value assignments $a,c$ and $c'$ to $A$ and $C$,
random variables $C$, $A(c)$ and $Y(c',a)$ are mutually independent.
Note that the former set of assumptions can be viewed
as a kind of mutual ignorability assumption derived from the graph, while
the latter set can be viewed as a mutual version of what Imai
called the \emph{sequential ignorability} assumption (\cite{imai10id}).
A general method for associating an arbitrary graph with sets of assumptions
on potential outcomes can be found, for instance, in the paper by \citet{thomas13swig}.
Thus, graphs are merely a visual representation of familiar potential outcome
models. A~purely visual view, however, can prove quite helpful.
A common approach within the Rubin framework is to assume
that a conditional ignorability assumption $(Y(a) \ci A \mid C)$ holds.
Here, $\ci$ is the conditional independence symbol.
This assumption logically follows from the assumptions
defining the FFRCISTG model
of Figure~\ref{fig:confounding}(a).\footnote{The NPSEM-IE always makes at least
as many assumptions as the FFRCISTG model, and in many cases more. Thus, any
assumption entailed by the FFRCISTG model of a graph is also entailed by the
NPSEM-IE for the same graph.}
Moreover, if $C$ is observed (represented graphically by
Figure~\ref{fig:confounding}(b), where $C$ is now normally shaded), this
assumption in turn entails that
the distribution of $Y(a)$ can be expressed as a functional of
the observed data via the adjustment formula
\[
p\bigl(Y(a)\bigr) = \sum_{c} p(Y \mid a,c) p(c)
\]
which in turn can be estimated by a variety of methods, including propensity
score methods (\cite{rosenbaum83propensity}),
inverse weighting methods (\cite{horvitz52weights}),
the parametric g-formula (\cite{robins87graphical})
or doubly robust methods (\cite{robins94estimation}).
If conditional ignorability is not a sensible assumption, or we cannot make use
of it due to strong sources of confounding that cannot be measured,
as is often the case in econometric applications, we may instead try to find
an instrument, which is shown graphically in Figure~\ref{fig:confounding}(c).
Here, the missing arrow from $Z$ to $Y$ represents the exclusion restriction
assumption namely that $Y(a,z) = Y(a,z')$ for any $a,z$ and $z'$,
and the lack of arrows from $C$ to $Z$ represents the assumption that
the instrument behaves as if randomly assigned: $Z \ci \{ A(z), Y(a,z) \}$
for any $a,z$. These assumptions also logically follow from the assumptions
defining the FFRCISTG model for the graph in Figure~\ref{fig:confounding}(c).
As Professor Imbens discusses, given
these assumptions, one could obtain bounds on the effect, or using further
assumptions, obtain point identification.
It may be that we cannot make sure of the conditional ignorability assumption
(that is, large parts of $C$ are not observable),
and it is the case that a variable for which
the above instrumental assumptions hold cannot be found.
In this case, it is possible to
use the systematic representation of restrictions on potential outcomes given
by a graph to derive additional methods of attack on the confounding
problem which can complement the instrumental variable and conditional
ignorability approaches.
For example, it may be possible that a strong \emph{mediating} variable
for the effect of $A$ on $Y$ exists and is observable, and moreover,
the mechanism by which this mediation happens is independent of the source
of the confounding between the treatment $A$ and outcome $Y$, given the
treatment $A$. This situation is shown in Figure~\ref{fig:confounding}(d).
In terms of assumptions on potential outcomes, such a strong mediator between
$A$ and $Y$ is represented as stating that $Y(w,a) = Y(w,a')$ for all $w,a,a'$.
In words, $A$ has no direct effect on $Y$ once we fix $W$ to any value $w$.
The unconfoundedness of the mediator may be represented as stating that
$\{ Y(w), A \} \ci W(a)$. It is easy
to see that the former assumption is a kind of exclusion restriction
corresponding to the absence of a directed arc from $A$ to $Y$, and the latter
is a kind of ignorability assumption, which corresponds to the absence of
an arc from $C$ to $W$, and the absence of other sources of confounding
between $W$ and $A$ and $Y$. Professor Imbens discusses these kinds of assumptions
in the context of instrumental variables in Sections~5.1 and 5.2.
Pearl has shown a result that is equivalent to stating that given a version of
SUTVA, and above assumptions, the following \emph{front-door formula} holds:
\begin{eqnarray}\label{eqn:front-door}
\hspace*{24pt}p\bigl(Y(a)\bigr) = \sum_{w} p(w \mid a) \sum
_{a'} p\bigl(Y \mid w, a'\bigr) p
\bigl(a'\bigr).
\end{eqnarray}
As before, the above assumptions logically follow from the assumptions defining
the FFRCISTG model for Figure~\ref{fig:confounding}(d).
A (slightly contrived) example of a situation represented in Figure~\ref{fig:confounding}(d) is as follows. We may suspect that consumption of
skyr (a kind of Icelandic dairy product) is protective against stomach cancer by
means of a particular type of flora found in skyr. However, we suspect
those with Icelandic citizenship may both consume more skyr than average, and
have different rates of stomach cancer than the general population. If we
cannot observe citizenship status, but we can do a simple test for the presence
of the protective flora, and moreover, we suspect the protective causal
mechanism is not influenced by the confounding variable directly but only
through skyr consumption, and moreover, this mechanism mediates
\emph{all} of the effect of skyr consumption, then we can find a way to express
the causal effect of skyr consumption on incidence of stomach cancer using
observational data via (\ref{eqn:front-door}).
Since graphs are merely a systematic \emph{visual} way of arranging information
on potential outcomes, they have proven extremely helpful for generating
solutions to problems posed by confounding in very general settings.
For instance, \citet{tyler13on}
used graphs to show that many informal
definitions for what a ``confounder'' is in the literature are ``incorrect''
in the sense of not agreeing with intuition in given examples, and not obeying
certain natural properties we expect a confounder to obey, while a definition
that is ``correct'' in this sense is fairly subtle.
Furthermore, a mediator-based approach resulting in (\ref{eqn:front-door})
has been generalized using causal graphs to a fully general method of
``deconfounding'' which can successfully be applied in complex longitudinal
settings even when no standard ignorability assumptions can be used,
and no good instruments can be found.
Consider a hypothetical longitudinal study represented by the causal graph
shown in Figure~\ref{fig:jamie}, where bidirected arrows represent the presence
of some hidden common cause. For instance, a bidirected arrow from $A$ to $C$
means there is a hidden common cause of $A$ and $C$. In this study,
$B$ and $D$ are administered treatments, $Y$ is the outcome, $A$ is
an observed baseline confounder, and $C$ is an intermediate health measure.
These variables are confounded, but in a very particular way displayed
by bidirected arrows in the graph. For instance, there
is no hidden common cause of $B$ and any other variable, and $D$ has
a hidden cause in common only with~$A$.
\begin{figure}
\begin{center}
\begin{tikzpicture}[>=stealth, node distance=1.2cm]
\tikzstyle{format} = [draw, thick, circle, minimum size=5.0mm,
inner sep=0pt]
\begin{scope}
\path[->, thick]
node[format] (a) {$A$}
node[format, right of=a] (b) {$B$}
node[format, right of=b] (c) {$C$}
node[format, right of=c] (d) {$D$}
node[format, right of=d] (e) {$Y$}
(a) edge (b)
(b) edge (c)
(b) edge[bend right] (d)
(c) edge (d)
(d) edge (e)
(c) edge[bend right] (e)
(b) edge[bend left=45] (e)
(a) edge[<->, thick, bend left] (c)
(c) edge[<->, thick, bend left] (e)
(a) edge[<->, thick, bend right=45] (e)
(a) edge[<->, thick, bend left=45] (d)
;
\end{scope}
\end{tikzpicture}
\end{center}
\caption{A longitudinal study with exposures $B,D$, outcome $Y$ and multiple
sources of unobserved confounding.}
\label{fig:jamie}
\end{figure}
We are interested in the total effect of the drug on the outcome, which can
be obtained from the distribution of $Y(b,d)$, in order to better determine the
optimal treatment assignment. Robins (\citeyear{robins86new})
gave a general method for expressing $Y(b,d)$ as a functional
of the observed data given that an assumption called sequential ignorability
(different from one in \cite{imai10id}) holds. Sequential ignorability on the observable variables happens
not to be implied by either the FFRCISTG model, or the NPSEM-IE of any underlying hidden variable DAG consistent with
Figure~\ref{fig:jamie}, but it can be shown that assumptions implied by these
models can be used to derive the following identity:
\begin{eqnarray}\label{eqn:complex}
p\bigl(Y(b,d) = y\bigr) &=& \sum_c \biggl( \sum
_{a} p(y,d,c \mid b,a) p(a) \biggr)\nonumber\\[-8pt]\\[-8pt]
&&{} \cdot
\frac{\sum_{a} p(c \mid b,a) p(a) }{
\sum_{a} p(d,c \mid b,a) p(a) }. \nonumber
\end{eqnarray}
Just as in the previous cases, it is possible to derive in a systematic
way a list of restrictions on potential outcomes over observable variables
implied by the graph in Figure~\ref{fig:jamie}, and use this list to
derive (\ref{eqn:complex}) without referring to the graph at all.
However, without the help of the graph
it is not so easy to see what this list of restrictions might look like, or how
the derivation based on this list might proceed. In more complex longitudinal
settings, with many more than 5 variables the problem becomes even more severe.
A general method for identifying potential outcome distributions as functionals
of observed data for causal graphs was given in the paper by \citet{tian02on}, and was proven
complete (in the sense of only failing on nonidentifiable distributions)
by \citeauthor{shpitser06id} (\citeyear{shpitser06idc}, \citeyear{shpitser06id}, \citeyear{shpitser07hierarchy}), \citet{huang06do}.
While it is possible to rederive these results
in terms of assumptions on potential outcomes, graph theory provided
mathematical terminology and intuitions that proved crucial in practice
for deriving these general ``deconfounding'' results.
\section{Conclusion}
Professor Imbens' excellent review shows that in the context
of economics, confounding arises due to the agent's decision algorithm, while
in the context of statistics confounding arises for other reasons (for
instance, due to lack of compatibility among patients in an observational
study). Instrumental variables are an important technique for dealing with
confounding in causal inference, though alternative methods involving
stratification were developed under the potential outcome model of Rubin and
Neyman.
More general lines of attack on the problem of confounding
were developed using graphical causal models, first used by Wright in
genetics, and extended into a general model by Spirtes, Glymour and Scheines
(\citeyear{spirtes93causation}), Pearl (\citeyear{pearl00causality}), Robins
(\citeyear{robins86new}, \citeyear{robins97estimation})
and others. These methods allow ``deconfounding'' of the problem even in
cases where confounders cannot be observed, and good instrumental variables
do not exist.
|
\section{Introduction}
Semiconductor waveguides offer the potential for highly integrated and low-power optical devices for applications such as optical communications~\cite{Kuo:06,Lee:09}, ultrafast measurements~\cite{Foster2008} and quantum communications~\cite{Xiong:11}. Silicon in particular combines very high nonlinearity~\cite{Leuthold2010} and well-developed micro-fabrication techniques~\cite{Lipson:2005}. In addition to displaying nonlinear phenomena common to all Kerr materials such as four-wave mixing~\cite{Fukuda:05} and soliton compression~\cite{ColmanP.2010}, silicon and other semiconductors exhibit multi-photon absorption and free-carrier effects. This gives rise to additional phenomena such as nonlinear losses~\cite{Yin:07}, the free-carrier induced blueshift~\cite{Monat:09} and pulse acceleration~\cite{Blanco-Redondo2014}. Given the large body of experimental work in silicon photonics, a deep theoretical understanding of multiphoton and free-carrier effects on optical pulse propagation is important to evaluate their impact on silicon and semiconductor photonics devices. This can also elucidate novel nonlinear wave phenomena.
Several theoretical approaches have been applied to silicon photonics devices. In general, the nonlinear equations for wave propagation can be solved numerically~\cite{Yin:07}. Alternatively, analytical methods provide a more fundamental understanding by exposing the phenomena underlying the observed effects. In some cases, the equations can be solved directly under some assumptions. For example, an integral solution for pulse propagation in silicon was obtained in the case of small variation of the temporal pulse~\cite{Rukhlenko:2010}. Such methods provide information on intrapulse temporal dynamics, but basic physical interactions are often embedded in implicit equations making it challenging to derive intuitive understanding. Alternatively, to elucidate the physical scalings of pulse propagation dynamics explicitly, we can derive a set of evolution equations for discrete parameters such as energy and pulse duration. These latter methods allow trends to be isolated and quickly evaluated over a broad range of parameters. A well-known example is the calculation of the Raman-induced soliton self-frequency shift in optical fibers~\cite{Gordon:86ssfs}. Recently, perturbation theory has been applied to soliton interaction with free-carriers effects~\cite{Saleh:2011,Roy:13}. This technique can directly estimate the soliton blueshift, but pulse parameters are limited to those of a fundamental soliton.
A more general analytic treatment of pulse propagation in silicon photonics should allow unconstrained parameters. The method of moments provides general trends by allowing arbitrary pulse shapes and parameter sets. The root mean square (RMS) quantities of interest for the pulse are defined and differential evolution equations are extracted~\cite{Perez:07}. The method of moments has been applied to glass media such as optical fibers to estimate jitter in optical communications~\cite{Gordon:86noise}, as well as model intrapulse Raman scattering~\cite{Santhanam2003413,Chen:10} and the nonlinear dynamics of short pulses~\cite{Tsoy:06,Burgoyne:07}. However, to our knowledge, the moment method has not been applied to semiconductor effects such as nonlinear losses and free-carrier interactions.
In this article, we apply the method of moments to pulse propagation in silicon photonics. We derive general equations for pulse evolution taking into account two-photon absorption (TPA) and free-carrier dispersion (FCD) and absorption (FCA) in addition to dispersion and self-phase modulation (SPM). Analytical expressions for the free-carrier induced blueshift and acceleration are obtained. These are found to depend explicitly on the pulse peak power only, due to the accumulated nature of free-carrier effects. Surprisingly, they are found to be independent of the pulse temporal duration. The moment equations are then solved numerically to study the influence of free-carrier effects in silicon photonic crystal and nanowire waveguides. Finally, the moment trends are compared with full numerical simulations and experiments, and good agreement is found over a broad range of input powers. The model developed here provides a concise and general picture of the complex pulse shaping effects observed in silicon waveguides directly in terms of the basic physical mechanisms underpinning them.
\section{The moment method in silicon waveguides}
Pulse propagation in semiconductor waveguides can be modelled using a generalized nonlinear Schr\"odinger equation (GNLSE)~\cite{Yin:07}. For semiconductor waveguides such as silicon where two-photon absorption is the dominant nonlinear absorption mechanism, the equation is:
\begin{multline}
\frac{\partial A}{\partial z} = -\frac{\alpha}{2}A -i\frac{\beta_2}{2}\frac{\partial^2 A}{\partial t^2} +\frac{\beta_3}{6}\frac{\partial^3 A}{\partial t^3} +i\gamma|A|^2A\\ - \frac{\gamma_{TPA}}{2}|A|^2A+\bigg(in_{FC}k_0-\frac{\sigma}{2}\bigg)N_cA.
\label{eqn:nlse}
\end{multline}
Here $A(z,t)$ is the amplitude envelope (the power $P=|A|^2$), $N_c(z,t)$ is the free-carrier density, $\alpha$ is the linear propagation loss and $\beta_2$ and $\beta_3$ are the group-velocity dispersion (GVD) and third-order dispersion (TOD) coefficients, respectively. The Kerr nonlinear parameter causing self-phase modulation is $\gamma=n_2k_0/A_{eff}$ and $\gamma_{TPA}=\alpha_{TPA}/A_{eff}$ is the two-photon absorption parameter. The free-carrier absorption parameter is $\sigma$ and the carrier density-dependent refractive index coefficient $n_{FC}$ causes free-carrier dispersion. Note that $n_{FC}$ is generally negative since in a free-carrier plasma higher density lowers the refractive index~\cite{soref:87}. The bulk nonlinear parameters are the intensity dependent refractive index $n_2$ and absorption $\alpha_{TPA}$. The effective mode area is $A_{eff}$ and the vacuum wavevector is $k_0$.
For the free-carrier density in silicon, we consider that two-photon absorption is the dominant generation mechanism. This is the case when doping and carrier injection are absent. The free-carrier density is then governed by the equation:
\begin{equation}
\frac{\partial N_c}{\partial t} = \rho_{FC}|A|^4-\frac{N_c}{\tau_c}.
\label{eqn:Nc}
\end{equation}
We have defined the free-carrier generation rate $\rho_{FC}=\alpha_{TPA}/(2\hbar\omega_0{A_{eff}}^2)$ with $\hbar\omega_0$ the photon energy, and $\tau_c$ is the free-carrier lifetime. From here on we will assume that the free-carrier recombination term is negligible. This requires $\tau_c$ to be much longer than an individual pulse but much shorter that the pulse train period. For silicon waveguides we have $\tau_c\sim1-10$~ns~\cite{Boyraz:04,Corcoran:10}, so this assumption is valid for femtosecond and picosecond pulses with megahertz repetition rates.
The method of moments reduces the dynamics of pulse propagation down to a set of discrete parameters. We start by defining the moments of energy $E$, acceleration $T_c$, spectral shift $\Omega$, RMS temporal width $\sigma_t$ and RMS chirp $\tilde{C}$~\cite{Santhanam2003413}:
\begin{align}
&E = \int_{-\infty}^{\infty}|A|^2dt, \label{eqn:E}\\
&T_c = \frac{1}{E}\int_{-\infty}^{\infty}t|A|^2dt, \\
&{\sigma_t}^2 = \frac{1}{E}\int_{-\infty}^{\infty}(t-T_c)^2|A|^2dt, \\
&\Omega = \frac{i}{2E}\int_{-\infty}^{\infty}(A^*\partial_tA-A\partial_tA^*)dt,\\
&\tilde{C} = \frac{i}{2E}\int_{-\infty}^{\infty}(t-T_c)(A^*\partial_tA-A\partial_tA^*)dt. \label{eqn:C}
\end{align}
Propagation equations for each moment can be obtained by taking the derivative of Eqs.~\ref{eqn:E}-\ref{eqn:C}, combining with the GNLSE in Eq.\ref{eqn:nlse} and using integration by parts, assuming the pulse amplitude vanishes at infinity~\cite{Perez:07}. To simplify the equations, we define the nonlinear loss coefficient:
\begin{equation}
\Gamma_{NL} = \gamma_{TPA}|A|^2 + \sigma N_c.
\end{equation}
We will now show the full, general moment equations. In practice not all terms contribute significantly given a set of waveguide and input pulse parameters, but for now we present the most general case. The full moment evolution equations are:
\begin{widetext}
\begin{align}
&\frac{dE}{dz} = -\alpha E - \int_{-\infty}^{\infty}\Gamma_{NL}|A|^2dt, \label{eqn:dEdz}\\
&\frac{dT_c}{dz} = \beta_2\Omega + \frac{\beta_3}{2E}\int_{-\infty}^{\infty}|\partial_tA|^2dt
-\frac{1}{E}\int_{-\infty}^{\infty}(t-T_c)\Gamma_{NL}|A|^2dt, \\
&\frac{d\sigma_t}{dz} = \frac{\beta_2\tilde{C}}{\sigma_t} + \frac{\beta_3}{2\sigma_t E}\int_{-\infty}^{\infty}(t-T_c)|\partial_tA|^2dt
-\frac{1}{2E\sigma_t}\int_{-\infty}^{\infty}\big((t-T_c)^2-{\sigma_t}^2\big)\Gamma_{NL}|A|^2dt, \\
&\frac{d\Omega}{dz} = \frac{-n_{FC}k_0}{E}\int_{-\infty}^{\infty}|A|^2\partial_tN_cdt
+\frac{1}{E}\int_{-\infty}^{\infty}\bigg(\frac{A^*\partial_tA-A\partial_tA^*}{2i}+\Omega|A|^2\bigg)\Gamma_{NL}dt, \\
&\frac{d\tilde{C}}{dz} = \frac{\beta_2}{E}\int_{-\infty}^{\infty}|\partial_t A|^2dt + i\frac{\beta_3}{4E}\int_{-\infty}^{\infty}(\partial_t^2A\partial_tA^*-\partial_t^2A^*\partial_tA)dt - \Omega\frac{dT_c}{dz}
+ \frac{\gamma}{2E}\int_{-\infty}^{\infty}|A|^4dt \nonumber\\
&\hspace{3cm} +\frac{-n_{FC}k_0}{E}\int_{-\infty}^{\infty}(t-T_c)|A|^2\partial_tN_cdt + \frac{1}{E}\int_{-\infty}^{\infty}\bigg((t-T_c)\frac{A^*\partial_tA-A\partial_tA^*}{2i}+\tilde{C}|A|^2\bigg)\Gamma_{NL}dt. \label{eqn:dCdz}
\end{align}
\end{widetext}
\section{Moments of a hyperbolic secant pulse}
The moment Eqs.~\ref{eqn:dEdz}-\ref{eqn:dCdz} contain the full GNLSE dynamics without assumptions, but it is difficult to extract trends in this form. They can be reduced to ordinary coupled differential equations by specifying a pulse shape. The general shape is assumed to be maintained through propagation, while its parameters are allowed to vary continuously. Here we will use a chirped hyperbolic secant pulse, which often occurs in soliton-like propagation:
\begin{multline}
A(z,t) = \sqrt{\frac{E}{2\tau}}\text{sech}\bigg( \frac{t-T_c}{\tau} \bigg)\times\\
\exp\bigg(-i\Omega(t-T_c) -i\frac{C}{2{\tau}^2}(t-T_c)^2\bigg)
\label{eqn:sech}
\end{multline}
The pulse duration $\tau$ and chirp $C$ in Eq.~\ref{eqn:sech} are linked to the RMS values by the constant $K=12/\pi^2\approx1.22$ as $\tau^2=K{\sigma_t}^2$ and $C=K\tilde{C}$~\cite{Santhanam2003413}. By combining Eqs.~\ref{eqn:dEdz}-\ref{eqn:dCdz} with Eq.~\ref{eqn:sech}, we obtain the set of moment equations:
\begin{align}
\frac{dE}{dz} &= -\alpha E - \frac{\gamma_{TPA}}{3}\frac{E^2}{\tau} - \frac{\sigma\rho_{FC}}{6}\frac{E^3}{\tau} \label{eqn:dEdzsech}, \\
\frac{dT_c}{dz} &= \beta_2\Omega + \frac{\beta_3}{2}\bigg( \Omega^2 + \frac{1+\frac{\pi^2}{4}C^2}{3\tau^2} \bigg) -\frac{7\sigma\rho_{FC}}{72}E^2, \label{eqn:dTdzsech}\\
\frac{d\tau}{dz} &= (\beta_2+\beta_3\Omega)\frac{C}{\tau} + \frac{\gamma_{TPA}}{\pi^2}E, \label{eqn:dtaudzsech}\\
\frac{d\Omega}{dz} &= -\frac{2}{15}n_{FC}k_0\rho_{FC}\frac{E^2}{\tau^2} -\frac{7}{72}\sigma\rho_{FC}\frac{C E^2}{\tau^2}, \label{eqn:dWdzsech}\\
\frac{dC}{dz} &= (\beta_2+\beta_3\Omega) \frac{\frac{4}{\pi^2} + C^2}{\tau^2} + \frac{2}{\pi^2}(\gamma+\gamma_{TPA}C)\frac{E}{\tau}. \label{eqn:dCdzsech}
\end{align}
Notice that the pulse duration Eq.~\ref{eqn:dtaudzsech} and chirp Eq.~\ref{eqn:dCdzsech} depend on the effective GVD $\beta_2' = \beta_2+\beta_3\Omega$ experienced by a frequency-shifted pulse~\cite{Tsoy:06}. The current equations do not account for pure TOD broadening, usually a small effect for pulses longer than 100 fs. Including this would require a pulse ansatz allowing asymmetry and higher order phase, which would complicate the equations.
These results are consistent with those from soliton perturbation theory including TPA and FCD~\cite{Roy:13}, noting that for a secant pulse the peak power $P = E/(2\tau)$. The moment method employed in this work is more general as it allows pulse duration and chirp to vary freely. Consequently, the moment method has a broader range of applicability across multiple pulse shapes beyond the assumptions underlying perturbation theory.
In addition to the well-known dispersion and SPM effects~\cite{Burgoyne:07}, the system of moments in Eqs.~\ref{eqn:dEdzsech}-\ref{eqn:dCdzsech} capture several important multiphoton and free-carrier effects unique to semiconductor waveguides. These are summarised in Table~\ref{tab:effects}.
\begin{table}[h!]
\caption{Multi-photon and free-carrier effects in silicon.}
\label{tab:effects}
\begin{center}
\begin{tabular}{lc}
\hline
Two-photon absorption & $\gamma_{TPA}E^2/\tau$ \\
Free-carrier absorption & $\sigma\rho_{FC}E^3/\tau$ \\
FCD acceleration & $\beta_2\Omega_{FCD}$\\
FCA pulse trail suppression & $\sigma\rho_{FC}E^2$ \\
TPA effective broadening & $\gamma_{TPA}E$ \\
FCD blueshift & $|n_{FC}|k_0\rho_{FC}E^2/\tau^2$ \\
FCA chirped frequency shift & $-\sigma\rho_{FC}C E^2/\tau^2$ \\
TPA chirp redistribution & $\gamma_{TPA}CE/\tau$ \\
\hline
\end{tabular}
\end{center}
\end{table}
The energy Eq.~\ref{eqn:dEdzsech} accounts for nonlinear loss due to two-photon and free-carrier absorption. It is also clear that the FCA effect is second-order to the dominant TPA effect from their respective cubic and square energy dependence. The FCA term in the temporal shift Eq.~\ref{eqn:dTdzsech} is caused by accumulated free-carriers suppressing the trailing edge of the pulse, shifting its center of mass toward the leading edge. This is usually weak in PhC waveguides compared to the dispersive effects, which shift the entire pulse~\cite{Blanco:optica14}. The TPA term in the duration Eq.~\ref{eqn:dtaudzsech} represents effective pulse broadening due to TPA suppressing the pulse peak more strongly and is also second-order to the dispersive effect in PhC. However, the TPA and FCA temporal effects can be dominant in nanowire waveguides due to their smaller dispersion~\cite{Liao:13}. The FCD blueshift can be seen in Eq.\ref{eqn:dWdzsech} (remember that $n_{FC}<0$). When coupled with the dispersion term in Eq.~\ref{eqn:dTdzsech}, the blueshift causes pulse acceleration or deceleration, depending on the sign of $\beta_2$.
The FCA frequency shift in Eq.\ref{eqn:dWdzsech} is caused by asymmetric absorption of a chirped pulse and is usually small. For example, using the parameters in section~\ref{sec:propsilicon} below, when $C<|n_{FC}|k_0/\sigma=15$, the FCA shift is small compared to the FCD blueshift. Similarly, the TPA term in the chirp Eq.~\ref{eqn:dCdzsech} is due to power-dependent suppression of a chirped pulse and is also often small. For the parameters below, the SPM chirp is dominant for $C < \gamma/\gamma_{TPA} = 2.4$. The FCD contribution to chirp in Eq.~\ref{eqn:dCdz} cancels out for symmetric pulses, similar to Raman scattering~\cite{Santhanam2003413}.
Another common pulse shape is the chirped gaussian. The resulting moment equations are similar to Eqs.~\ref{eqn:dEdzsech}-\ref{eqn:dCdzsech} with different numerical factors and are given in Appendix~\ref{app:gauss}. The essential physics remain the same.
\section{Free-carrier induced blueshift and acceleration}
\label{sec:blueshift}
We now use the moment equations to derive new physical insight about the free-carrier induced frequency and temporal shifts. The first term in Eq.~\ref{eqn:dWdzsech} represents the free-carrier induced blueshift. This general feature of free-carrier dispersion in plasmas has been predicted and observed in gases~\cite{Wood:1991}, silicon waveguides~\cite{Monat:09,Roy:13,Blanco-Redondo2014} and gas-filled hollow-core fibers~\cite{Fedotov:2007,Saleh:2011}. As discussed above, the second term is due to FCA on chirped pulses, which is usually smaller than the FCD term for small chirps and therefore negligible.
The $E^2/\tau^2$ dependence of the free-carrier blueshift corresponds to a $P^2$ dependance. The pulse duration dependence of the FCD phase cancels out because free-carriers accumulate over the pulse ($N_c$ is the integral of Eq.~\ref{eqn:Nc}). Thus, at constant peak power the free-carrier blueshift is not affected by pulse duration. This is a key conclusion of this work. Note that a $P^3$ dependence is expected of a three-photon limited semiconductor system~\cite{ColmanP.2010}, while for near-threshold tunnelling ionization in gases this would scale as $P$~\cite{Saleh:2011}. As an experimental consideration, we note other effects which suppress peak power, such as dispersion, become weaker for longer pulses. Consequently, the observed blueshift may be larger for longer pulses in these systems.
The pulse duration independence of the FCD blueshift contrasts with the well studied Raman self-frequency shift (SFS). In general, Raman SFS scales like $E/\tau^3$~\cite{Santhanam2003413}, corresponding to a $P/\tau^2$ dependence. This is expected since Raman scattering depends on the local pulse steepness (derivative in \cite[Eq. 1]{Santhanam2003413}) which increases for shorter pulses.
We now estimate the blueshift accumulated during propagation. Defining $P(z)= E/(2\tau) =P_0\eta(z)$, with the input peak power $P_0$ and peak power decay factor $\eta \le 1$, and assuming $n_{FC}<0$, the blueshift becomes:
\begin{align}
\Omega(z) &= \frac{8}{15}|n_{FC}|k_0\rho_{FC}\int_0^zP(z')^2dz', \nonumber\\
&= \frac{8}{15}|n_{FC}|k_0\rho_{FC}{P_0}^2z_{eff}.
\label{eqn:blueshift}
\end{align}
We have defined the FCD effective length:
\begin{equation}
z_{eff} = \int_0^z\eta(z')^2dz'.
\end{equation}
The effective length $z_{eff}$ capture the effects of linear and nonlinear losses and other mechanisms reducing peak power, such as dispersion. For long propagation, $z_{eff}$ converges to a characteristic length $L_{max}$ discussed below. We can also show that for short propagation $z\ll L_{max}$ we have $z_{eff}\approx z$. Thus the blueshift is initially linear in $z$ before converging to a maximum value.
We now estimate the free-carrier induced temporal shift. To first order, consider the GVD term in Eq.~\ref{eqn:dTdzsech}:
\begin{align}
T_c(z) &= \frac{8}{15}\beta_2|n_{FC}|k_0\rho_{FC}{P_0}^2\int_0^zz_{eff}'dz'.
\label{eqn:dtzeff}
\end{align}
For short propagation the temporal shift $T_c \propto z^2$, corresponding to a constant free-carrier induced pulse acceleration. When the blueshift saturates we have $T_c \propto z$, a fixed velocity from the accumulated free-carrier blueshift.
It is often possible to obtain a closed expression for the effective length $z_{eff}$. Moreover, this effective length converges to a saturated non-linear length $L_{max}$ after sufficient propagation. Table~\ref{tab:zeff} summarizes the effective lengths for the main peak power decay mechanisms in silicon photonics. The decay factors for TPA and linear loss can obtained from the NLSE Eq.~\ref{eqn:nlse}. The peak power reduction from dispersive broadening is obtained by solving Eq.~\ref{eqn:dtaudzsech} and Eq.~\ref{eqn:dCdzsech} together with GVD only and zero input chirp. The dispersion length is $L_D = \tau(0)^2/|\beta_2|$.
\begin{table}[h!]
\caption{Effective lengths for FCD effects.}
\label{tab:zeff}
\begin{center}
\begin{tabular}{|l|c|c|c|}
\hline
& $\eta(z)$ & $z_{eff}$ & $L_{max}$ \\
\hline
TPA & $(1+\gamma_{TPA}P_0z)^{-1}$ & $z/(1+\gamma_{TPA}P_0z)$ & $1/(\gamma_{TPA}P_0)$ \\ \hline
Loss & $e^{-\alpha z}$ & $(1-e^{-2\alpha z})/2\alpha$& $1/(2\alpha)$ \\ \hline
GVD & $1/\sqrt{1+\frac{4}{\pi^2}\frac{z^2}{{L_D}^2}}$ & $\frac{\pi}{2}L_D\tan^{-1}{\bigg( \frac{2}{\pi}\frac{z}{L_D} \bigg)}$ & $(\pi/2)^2L_D$\\
\hline
\end{tabular}
\end{center}
\end{table}
\section{Pulse propagation in silicon PhC waveguides}
\label{sec:propsilicon}
In this section we study the propagation of pulses in silicon waveguides and identify the dominant contributions to the nonlinear pulse dynamics. The system of moment Eqs.~\ref{eqn:dEdzsech}-\ref{eqn:dCdzsech} can be solved numerically in the general case. Solving these simple equations is much faster than full numerical solutions of the NLSE and can quickly yield trends for pulse propagation in semiconductors, as long as the pulse shape does not deviate strongly from the input. The computational advantage is particularly strong here since full simulations require the NLSE and free-carrier equations to be solved together.
The parameters used correspond to a slow-light silicon photonic crystal (PhC) waveguide described elsewhere~\cite{Blanco-Redondo2014,Blanco:optica14}. At a wavelength $\lambda_0$ of 1543~nm, the effective modelling parameters are given in Table~\ref{tab:param}. The slow-light scaling of the bulk parameters used to obtain these is discussed in Appendix~\ref{app:slowlight}.
\begin{table}[h!]
\caption{Physical parameters for a silicon PhC waveguide at 1543~nm}
\label{tab:param}
\begin{center}
\begin{tabular}{lclc}
\hline
loss & 5.0 dB/mm & $L$ & 0.396 mm \\
$\beta_2$ & -10.4 ps$^2$/mm & $\beta_3$ & -0.53 ps$^3$/mm\\
$\gamma$ & 2948 W$^{-1}$m$^{-1}$ & $\gamma_{TPA}$ & 1207 W$^{-1}$m$^{-1}$\\
$n_{FC}$ & -3.4$\cdot 10^{-26}$ m$^2$ & $\sigma$ & 9.0$\cdot10^{-21}$ m$^2$ \\
$\rho_{FC}$ & $1.46\cdot10^{34}$ m$^{-3}$W$^{-2}$s$^{-1}$ &\\
\hline
\end{tabular}
\end{center}
\end{table}
The input pulse parameters are $E(0) = 10.4$~pJ, $\tau(0) = 0.74$~ps and $C(0) = 0$, corresponding to an unchirped hyperbolic secant with a full-width half-maximum (FWHM) duration of 1.3~ps and 4~W peak power. The coupled moment Eqs.~\ref{eqn:dEdzsech}-\ref{eqn:dCdzsech} are solved using a 4th-order Runge-Kutta integrator. We present results with different effects turned off to isolate their influence. The calculations without TPA correspond to dropping the power-dependent loss from Eq.~\ref{eqn:nlse} while keeping free-carrier generation active. Although this is not possible in reality, it allows the influence of multiphoton loss to be isolated from that of free-carrier effects.
The free-carrier effects are most pronounced when looking at the frequency and temporal shifts. The blueshift $\Omega$ is caused by free-carrier dispersion, and free-carrier absorption has a minor effect here. The frequency shift $\Delta f = \Omega/(2\pi)$ is shown in Fig.~\ref{fig:momentdf}. The blueshift accumulates mostly over one effective length. The lengths scales as defined in section~\ref{sec:blueshift} are $L_{max}^{D} = 0.13$~mm, $L_{max}^{TPA} = 0.21$~mm and $L_{max}^{lin} = 0.44$~mm. In this case, the blueshift is limited mostly by dispersion and two-photon absorption, while linear loss plays a minor role. With TPA loss removed a larger blueshift is accumulated until dispersion limits the FCD again. For comparison, the blueshift predicted by the approximate integral Eq.~\ref{eqn:blueshift} is also plotted. To account for the contribution of different effects to $z_{eff}$, we apply the effective lengths in cascade, from the strongest to the weakest effect. Thus, $z_{eff}^{D}$ is first computed and used as the input for $z_{eff}^{TPA}$. There is close agreement with the full moments solution, so Eq.~\ref{eqn:blueshift} can provide a good estimate of the blueshift without having to calculate all the moments.
\begin{figure}[htbp]
\centerline{\includegraphics[width=3in]{moments_df.eps}}
\caption{Frequency shift $\Delta f = \Omega/(2\pi)$ of a hyperbolic secant pulse along a silicon PhC waveguide from the full moment equations with different effects suppressed (lines). Also shown is the analytic result from Eq.\ref{eqn:blueshift} including FCD blueshift only (circles).}
\label{fig:momentdf}
\end{figure}
The temporal shift $T_c$ is shown in Fig.~\ref{fig:momentTc}. When the free-carrier blueshift is combined with anomalous dispersion, the pulse advances in time compared to the linear case~\cite{Blanco-Redondo2014}. At normal dispersion the pulse would lag instead. The relatively small negative TOD adds a minor contribution here. In contrast to the frequency shift, the temporal shift occurs throughout the whole waveguide since it depends on the accumulated blueshift, not the local peak power. The temporal shift $T_c$ increases quadratically in $z$ as long as the blueshift is increasing linearly in $z$, as described by Eq.~\ref{eqn:dtzeff}. However, as the peak power decreases during propagation due to dispersion and loss during propagation, the FCD blueshift saturates and $z$ approaches a constant $L_{max}$ and $T_c$ increases linearly, as expected. For comparison, results from the effective length treatment from Eq.~\ref{eqn:dtzeff} are shown, where the $z_{eff}$ used in the blueshift calculation was integrated numerically. The trend is well accounted for, but the temporal shift is slightly underestimated since TOD and FCA are not included in the approximation.
\begin{figure}[htbp]
\centerline{\includegraphics[width=3in]{moments_accel.eps}}
\caption{Temporal shift $T_c$ of a hyperbolic secant pulse along a silicon PhC waveguide from the full moment equation with different effects suppressed (lines). Also shown is the analytic result from Eq.\ref{eqn:dtzeff} with GVD only (circles).}
\label{fig:momentTc}
\end{figure}
The evolution of the pulse duration $\tau$ is shown in Fig.~\ref{fig:momenttau}. This is dominated by dispersive broadening, while the free-carrier effects have little incidence at this input energy. Pulse broadening is slightly reduced with TPA suppressed. TPA slightly broaden the pulse as in Eq.~\ref{eqn:dtaudzsech}, and this is enhanced by the nonlinear loss which reduces the SPM pulse compression for anomalous dispersion. FCD contributes only slightly to temporal broadening. Removing TOD (not shown) has the exact same effect. This is because the blueshift increases the effective GVD $\beta_2' = \beta_2+\beta_3\Omega$ at the pulse center frequency when $\beta_2$ and $\beta_3$ have the same sign, so this is a combined FCD-TOD effect~\cite{Blanco:optica14}. The trend for the chirp moment is similar to the one for pulse duration and is not shown here.
\begin{figure}[htbp]
\centerline{\includegraphics[width=3in]{moments_tau.eps}}
\caption{Pulse duration $\tau$ of a hyperbolic secant pulse along a silicon PhC waveguide with different effects suppressed.}
\label{fig:momenttau}
\end{figure}
The evolution of the pulse energy $E$ is shown in Fig.~\ref{fig:momentE}. Two-photon absorption is the dominant loss mechanism over the characteristic length $L_{max}^{TPA}$, and linear loss also contributes over the whole length. Free-carrier absorption plays a minor role here. Removing dispersion enhances nonlinear losses slightly since higher peak powers are maintained without dispersive broadening.
\begin{figure}[htbp]
\centerline{\includegraphics[width=3in]{moments_E.eps}}
\caption{Energy $E$ of a hyperbolic secant pulse in a silicon PhC waveguide with different effects suppressed.}
\label{fig:momentE}
\end{figure}
Overall, the FCD blueshift in Eq.~\ref{eqn:dWdzsech} and TPA loss in Eq.~\ref{eqn:dEdzsech} are clearly the dominant terms, as mentioned before. Thus, in many cases it is sufficient to consider only these two when solving Eq.~\ref{eqn:dEdzsech}-\ref{eqn:dCdzsech} to obtain the main trends.
\section{Comparison with simulations and experiments in silicon PhC waveguides}
We now compare the predictions of the moment method with experiments and full numerical simulations of the GNLSE given by Eq. \ref{eqn:nlse}. The experiments used a high sensitivity frequency-resolved electrical gating (FREG) system to probe the temporal intensity and phase of the output pulses~\cite{Dorrer:02}. The waveguide is a silicon PhC with parameters given in Table~\ref{tab:param}, and the details of the experiments are given in~\cite{Blanco-Redondo2014,Blanco:optica14}. The input pulses had a fixed input duration of about 1.3~ps FWHM ($\tau=0.74$~ps RMS) and the input peak power inside the waveguide was varied over several watts. The numerical simulations used the split-step Fourier method to solve Eqs.~\ref{eqn:nlse}-\ref{eqn:Nc} with the same device parameters and input pulse as the moment method.
We first consider the free-carrier blueshift $\Omega$ in Fig.~\ref{fig:expdf}(a). The blueshift increases quadratically with input peak power over most of the range, as expected from Eq.~\ref{eqn:blueshift}. The moment method agrees well with NLSE simulations and experiments, while slightly overestimating at high powers. This is because the moment method underestimates pulse broadening, which will be discussed below in relation to the pulse duration. The trends for the temporal shift $T_c$ are shown in Fig.~\ref{fig:expdf}(b). Free-carrier acceleration is clearly observed, and agreement is fairly consistent as it was for the blueshift.
\begin{figure}[htbp]
\centerline{\includegraphics[width=3.25in]{exp_df_accel_2col.eps}}
\caption{(a) Frequency shift $\Delta f = \Omega/(2\pi)$ and (b) temporal shift $T_c$ from the moment equations, NLSE simulations and experimental measurements for a silicon PhC waveguide at increasing input peak power.}
\label{fig:expdf}
\end{figure}
The trends for pulse duration are shown in Fig.~\ref{fig:exptau}(a). To better compare with experiments, we use the FWHM duration which is less sensitive to experimental noise. The effective FWHM for the moment method and NLSE are obtained by scaling the RMS duration by the shape factor ($t_{FWHM}=2\cosh^{-1}(\sqrt{2})\sqrt{12/\pi^2}\sigma_t$ for hyperbolic secant). The agreement is quite good up to 3~W. Dispersive broadening dominates while SPM partially cancels dispersion as power increases. For high powers the moment method and simulations diverge considerably. As we discuss in detail in a recent submission~\cite{Blanco:optica14}, this is caused by free-carrier dispersion induced spectral broadening, which induces temporal broadening when coupled with dispersion in the NLSE simulations. The moment method assumes a fixed pulse shape, so it instead predicts further SPM compression. As discussed before, the trend for chirp is similar to that for pulse duration and is not shown.
\begin{figure}[htbp]
\centerline{\includegraphics[width=3.25in]{exp_tau_en_2col.eps}}
\caption{(a) Effective FWHM duration from the moment equations ($1.763\tau$) and NLSE simulations ($1.763\cdot1.1\cdot\sigma_t$), and experimental FWHM duration for a silicon PhC waveguide at increasing input peak power. The dashed line indicates the input pulse duration. (b) Pulse energy $E$ from the moment equations, NLSE simulations and experimental measurements.}
\label{fig:exptau}
\end{figure}
However, as shown above, predictions for the other moments remain accurate at high powers. The moment method's inability to predict FCD temporal broadening does not affect other predictions as much, since the nonlinear effects like blueshift and TPA accumulate mostly in the first part of propagation before dispersion and losses suppress them.
In contrast, most of the dispersive broadening occurs after significant nonlinear effects and losses have accumulated, explaining why discrepancies are stronger for pulse duration predictions.
Finally, results for pulse energy $E$ are shown in Fig.~\ref{fig:exptau}(b). The general trends agree well, with TPA causing significant nonlinear loss. There is some discrepancy at high powers which could be due to uncertainties in the effective area and nonlinear coefficients.
\section{Temporal dynamics in silicon nanowires}
We now apply the method of moments to a silicon nanowire waveguide. The waveguide parameters provided by these devices will allows us to highlight effects that were not readily observable in the photonic crystal waveguides discussed so far. We will focus on effective pulse broadening caused by peak power suppression by two-photon absorption and temporal shifting due to the asymmetry of free-carrier absorption.
We use waveguide parameters similar to previous reports~\cite{Liao:13}. The input pulses have a FWHM duration of 1.3~ps ($\tau=0.74$~ps) and a centre wavelength of 1545~nm. The nanowire has a length $L=4$~mm, linear loss of 3~dB/cm and dispersion $\beta_2=-1.5$~ps$^2$/m. The effective non-linearity is $\gamma = 349$~W$^{-1}$m$^{-1}$ and the two-photon absorption is $\gamma_{TPA} = 143$~W$^{-1}$m$^{-1}$. The carrier generation rate is $\rho_{FC}=7.9\cdot10^{33}$ m$^{-3}$W$^{-2}$s$^{-1}$. The free-carrier dispersion coefficient is $n_{FC}= 1.35\cdot10^{-27}$~m$^3$ and the free-carrier absorption is $\sigma = 1.45\cdot10^{-21}$ m$^2$~\cite{Yin:07}.
The output pulse duration from the moments method for increasing input peak power is shown in Fig.~\ref{fig:tau_nw} and compared to full NLSE simulations. Dispersive broadening is negligible in the nanowire ($L_D=363$~mm). Instead, two-photon absorption is the dominant broadening mechanism, as described by the energy-dependent term in Eqn.~\ref{eqn:dtaudzsech}. The TPA suppresses the peak of the pulse more strongly, resulting in an effective broadening. With FCA turned off, other losses are reduced and thus the TPA effects are strengthened. The quantitative deviation between moments and the NLSE are probably due to the pulse evolving away from the input shape.
\begin{figure}[htbp]
\centerline{\includegraphics[width=3in]{tau_nw.eps}}
\caption{Pulse duration $\tau$ in a silicon nanowire at increasing input peak power from the moments method with different effects suppressed and compared to NLSE simulations.}
\label{fig:tau_nw}
\end{figure}
The output temporal shift for different input powers in the nanowire is shown in Fig.~\ref{fig:accell_nw}. This is dominated by free-carrier absorption, as described by the energy-dependent term in Eqn.~\ref{eqn:dTdzsech}. The accumulation of free-carriers towards the trailing edge of the pulse causes asymmetric absorption, causing the pulse temporal centroid to shift to earlier times. Removing TPA reduces losses and thus increases the free-carrier density and asymmetric absorption. In contrast, in the photonic crystal waveguides the temporal shift is caused by the free-carrier dispersion coupled with GVD.
\begin{figure}[htbp]
\centerline{\includegraphics[width=3in]{accel_nw.eps}}
\caption{Temporal shift $T_c$ in a silicon nanowire at increasing input peak power from the moments method with different effects suppressed and compared to NLSE simulations.}
\label{fig:accell_nw}
\end{figure}
Overall, the trends modelled for nanowires are similar to previous reports~\cite{Liao:13}. The silicon nanowires represent a distinct regime where dispersion is small and nonlinear absorption from multi-photon and free-carrier effects dominate temporal dynamics. In contrast, PhC devices are dominated by linear and free-carrier dispersions which redistribute energy rather than selectively suppress it.
\section{Conclusion}
We have developed a moment method approach to free-carrier effects in semiconductors including two-photon absorption, free-carrier dispersion and free-carrier absorption. General expressions for the evolution of energy, temporal acceleration, pulse duration, frequency shift and chirp were obtained. The system of evolution equations for a hyperbolic secant pulse was obtained. From this we derived the free-carrier blueshift and temporal acceleration based on an effective length method accounting for linear and nonlinear peak power losses. We showed that the blueshift and acceleration depend explicitly on the pulse peak power only and are independent of the pulse duration, due to the buildup of free-carriers across the pulse.
We modelled pulse evolution in a silicon photonic crystal and nanowire waveguides and found that the free-carrier blueshift is limited mainly by dispersion and nonlinear absorption. The free-carrier absorption was found to play a negligible role in photonic crystals. In contrast, free-carrier and two-photon absorption are dominant in nanowires. Finally, comparing the moment method with full numerical simulations and experiments revealed good agreement for free-carrier phenomena such as multiphoton absorption, blueshifting and temporal acceleration. The temporal pulse width and chirp can be accurately predicted as long as free-carrier spectral broadening coupled with dispersion does not significantly distort the temporal pulse.
Overall, the moment method is a powerful tool for isolating the basic scaling laws behind complex multiphoton and free-carrier pulse shaping effects in silicon photonics. It also allows rapid calculation of pulse evolution trends in actual silicon waveguides to understand and optimise their design. The results and methods developed here apply to any two-photon absorption semiconductor waveguides and can be straightforwardly extended to n-photon absorption systems.
|
\section*{Keywords}
{\sffamily Proto Transitive Relations, PRAX, PRAS, Generalized Transitivity, Rough Dependence, Rough Objects, Granulation, Algebraic Semantics, Approximate Relations, Approximate Semantics, Kleene Algebras, Axiomatic Theory of Granules, Geometry of Knowledge, Contamination Problem.
}
\newpage
\pagestyle{empty}
\section*{About the Author}
{\sffamily A. Mani is an active researcher in algebra, logic, rough sets, vagueness, philosophy and
foundations of Mathematics. She has published extensively on the subjects in a number of
international peer-reviewed journals for more than a decade. Her current affiliations
include the University of Calcutta in Kolkata and Division-R of STVROM. She is active in various academic groups
like ISRS, IRSS, ASL and FOM. She is also a teacher, free software activist, feminist, consultant in statistical and soft computing and service provider.}
\tableofcontents
\mainmatter
\chapter[Introduction]{Introduction}
\automark[section]{chapter}
Proto-transitivity is one of the infinite number of possible generalizations of transitivity. These types of generalized relations happen often in application contexts. Failure to recognize them causes mathematical models to be inadequate or underspecified and tends to unduly complicate algorithms and approximate methods. From among the many possible alternatives that fall under \emph{generalized transitivity}, we chose \emph{proto-transitivity} because of application contexts, its simple set theoretic definition, connections with factor relations and consequent generative value among such relations. It has a special role in modelling knowledge as well.
Proto-transitive approximation spaces \textsf{PRAX} have been introduced by the present author in \cite{AM270} and the structure of definite objects has been characterized in it to a degree. It is relatively a harder structure from a semantic perspective as the representation of rough objects is involved \cite{AM270}. Aspects of knowledge interpretation in \textsf{PRAX} contexts have been considered in \cite{AM270} and in \cite{AM3690} the relation of approximations resulting from approximation of relations to the approximations from the original relation are studied in the context of \textsf{PRAX}. These are used for defining an approximate semantics for \textsf{PRAX} and their limitations are explored by the present author in the same paper. All of these are expanded upon in this monograph.
Rough objects as explained in \cite{AM99,AM240} are collections of objects in a classical domain (Meta-C) that appear to be indistinguishable among themselves in another rough semantic domain (Meta-R). But their representation in most \textsf{RST}s in purely order theoretic terms is not known. For \textsf{PRAX}, this is solved in \cite{AM270}. Rough objects in a \textsf{PRAX} need not correspond to intervals of the form $]a, b[$ with the definite object $b$ covering (in the ordered set of definite objects) the definite object $a$.
If $R$ is a relation on a set $S$, then $R$ can be approximated by a wide variety of partial or quasi-order relations in both classical and rough set perspective \cite{RJ2011}. Though the methods are essentially equivalent for binary relations, the latter method is more general. When the relation $R$ satisfies proto-transitivity, then many new properties emerge. This aspect is developed further in the present monograph and most of \cite{AM3690} is included.
When $R$ is a quasi-order relation, then a semantics for the set of ordered pairs of lower and upper approximations $\{(A^{l}, A^{u}) ; \, A\subseteq S\}$ has recently been developed in \cite{SJ,JPR}. Though such a set of ordered pairs of lower and upper approximations are not rough objects in the \textsf{PRAX} context, we can use the approximations for an additional semantic approach to it. We prove that differences of consequent lower and upper approximations suggest partial structures for \emph{measuring} structured deviation. The developed method should also be useful for studying correspondences between the different semantics \cite{AM1800,AM3600}. Because of this we devote some space to the nature of transformation of granules by the relational approximation process.
In this research monograph, we also investigate the nature of possible concepts of \emph{rough dependence} first. Though the concept of independence is well studied in probability theory, the concept of dependence is rarely explored in any useful way. It has been shown to be very powerful in classical probability theory \cite{BD2010} - the formalism is valid over probability spaces, but its axiomatic potential is left unexplored. Connections between rough sets and probability theory have been explored from rough measure and information entropy viewpoint in a number of papers \cite{PZ2002,SL2006,GPS04,YY2003,BCC2007}. The nature of rough independence is also explored in \cite{AM3930} by the present author and there is some overlap with the present work. Apart from problems relating to contamination, we show that the comparison by way of corresponding concepts of dependence fails in a very essential way.
Further, using the introduced concepts of rough dependence we internalize the approximate semantics instead of depending on correspondences. This allows for richer variants of the earlier semantics of rough objects.
This monograph is reasonably self-contained and is organized as follows: In the rest of this chapter we introduce the basics of proto-transitivity, recall relevant information of Nelson algebras, granules and granulations. In the following chapter, we define relevant approximations in \textsf{PRAX} and study their basic properties and those of definite elements. In the third chapter, we propose an abstract and three other extended examples justifying our study. In the following chapter, we describe the algebraic structures that can be associated with the semantic properties of definite objects in a \textsf{PRAX}. The representation of rough objects is done from an interesting perspective in the fifth chapter. In the sixth chapter, we define new derived operators in a \textsf{PRAX} and consider their connection with non monotonic reasoning. These are of relevance in representation again. In the following chapter, atoms in the partially ordered set of rough object are described. This is followed by an algebraic semantics that relies on multiple types of aggregation and commonality operations. In the ninth chapter, a partial semantics similar to the increasing Nelson algebraic semantics is formulated. This semantics is completed in three different ways in the fourteenth chapter after internalization of dependency. In the tenth and eleventh chapters approximate relations and approximate semantics are considered - the material in these chapters includes expansions of the results in \cite{AM3690}. In the following two chapters, we define concepts of rough dependence, compare them with those of probabilistic dependence and demonstrate their stark differences - the material in these chapters are expansions of \cite{AM3930}. The knowledge interpretation of \textsf{PRAX} is revisited in the fifteenth chapter.
\section[Basics]{Basic Concepts, Terminology}
\begin{definition}
A binary relation $R$ on a set $S$ is said to be \emph{weakly-transitive, transitive or
proto-transitive} respectively on $S$ \ifof
$S$ satisfies
\begin{mitemize}
\item { If whenever $Rxy,\,Ryz$ and $x\,\neq\,y\,\neq\,z$ holds, then $Rxz$.
(i.e. $(R\circ R)\setminus \Delta_{S}\,\subseteq R$ (where $\circ$ is relation composition)
, or}
\item {whenever $Rxy \,\&\,Ryz$ holds then $Rxz$ (i.e. $(R\circ
R) \subseteq R$), or }
\item {Whenever $Rxy,\,Ryz , \,Ryx ,\, Rzy$ and $x\,\neq\,y\,\neq\,z$ holds, then $Rxz$ follows, respectively. Proto-transitivity of
$R$ is equivalent to $R\cap R^{-1} \,=\,\tau(R)$ being weakly transitive.}
\end{mitemize}
\end{definition}
We will use the following simpler example to illustrate many of the concepts and situations in the monograph. For detailed motivations see \textsf{Ch.}\ref{moex} on motivation and examples.
\begin{examp}\label{agre}
A simple real-life example of a proto-transitive, non transitive relation would be the relation $\mathbb{P}$, defined by
\[\mathbb{P}xy \;\mathsf{if\;and\; only \;if \;} x \mathrm{\;thinks\; that\;} y \mathrm{\;thinks\; that\; color\; of \;object\;} O \mathrm{\;is \;a \;maroon}.\]
But we will use the following simple example from databases as a persistent one (especially in the chapters on approximation of relations) to illustrate a number of concepts. It has other attributes apart from the main one for illustrating more involved aspects.
Let $\mathcal{I}$ be survey data in table form with column names being for sex, gender, sexual orientations, other personal data and opinions on sexist contexts with each row corresponding to a person. We write
\[Rab\;\mathsf{if\; and\; only\; if\;} \mathrm{person \;} a \mathsf{\;agrees\; with\;} b's \mathrm{\;opinions}.\]
The predicate \textsf{agrees with} can be constructed empirically or from the data by a suitable heuristic. Often $R$ is a proto-transitive, reflexive relation and this condition can be imposed to complete partial data as well (as a rationality condition). If $a$ agrees with the opinions of $b$, then we will say that $a$ is an \emph{ally} of $b$ - if $b$ is also an ally of $a$, then they are \emph{comrades}. Finding optimal subsets of allies can be an interesting problem in many contexts especially given the fact that responses may have some vagueness in them.
\end{examp}
\begin{definition}
A binary relation $R$ on a set $S$ is said to be \emph{semi-transitive} on $S$ \ifof
$S$ satisfies
\begin{mitemize}
\item {Whenever $\tau(R)a b \& R b c $ holds then $R a c$ follows and }
\item {Whenever $\tau(R)a b \& R c a $ holds then $R c b$ follows.}
\end{mitemize}
\end{definition}
Henceforth we will use $Rxy$ for $(x, y)\in R$ uniformly. $Ref (S),$ $Sym(S),$ $Tol(S),$
$r\tau (S),$ $w\tau(S),\, p\tau(S),\, s\tau(S),\, EQ(S)$ will respectively denote the set
of reflexive, symmetric, tolerance, transitive, weakly transitive, pseudo transitive,
semi-transitive and equivalence relations on the set $S$ respectively.
The following proposition has steep ontological commitments.
\begin{proposition}
For a relation $R$ on a set $S$, the following are satisfied:
\begin{mitemize}
\item {$R$ is \emph{weakly transitive} \ifof $(R\cap R^{-1})\setminus \Delta_{S}
\subseteq R$.}
\item {$R$ is \emph{transitive} \ifof $(R\cap R^{-1}) \subseteq R$.}
\end{mitemize}
\end{proposition}
By a \emph{pseudo order}, we will mean an antisymmetric, reflexive relation. A
\emph{quasi-order} is a reflexive, transitive relation, while a partial order a
reflexive, antisymmetric and transitive relation.
Let $\alpha \subseteq \rho$ be two binary relations on $S$, then $\rho | \alpha$ will be
the relation on $S|\rho$ defined via $(x, y)\in \rho | \alpha$ \ifof $(\exists b \in x ,
c\in y ) (b, c) \in \rho$. The relation $Q|\tau(Q)$ for a relation $Q$ will be denoted by
$\sigma (Q)$.
The following are known:
\begin{proposition}
If $Q$ is a quasi-order on $S$, then $Q|\tau(Q)$ is a partial order on $S|\tau(Q)$.
\end{proposition}
\begin{proposition}
If $R\in Ref(S)$, then $R\in p\tau(S)$ \ifof $\tau(R)\in EQ(S)$.
\end{proposition}
\begin{proposition}
In general, \[w\tau(S) \,\subseteq\, s\tau (S) \,\subseteq\, p\tau(S) .\]
\end{proposition}
\begin{proposition}
If $R\in p\tau(S) \cap Ref(S)$, then the following are equivalent:
\begin{description}
\item [A1]{$([a], [b])\in R|\tau (R)$ \ifof $(a, b) \in R$.}
\item [A2]{$R$ is semi-transitive.}
\end{description}
\end{proposition}
In \cite{ICH}, it is proved that
\begin{theorem}
If $R\in Ref(S)$, then the following are equivalent:
\begin{description}
\item [A3]{$R|\tau (R)$ is a pseudo order on $S|\tau(R)$ and A1 holds.}
\item [A2]{$R$ is semi-transitive.}
\end{description}
\end{theorem}
Note that \emph{Weak transitivity} of \cite{ICH} is \emph{proto-transitivity} here. $Ref (S)$, $r\tau (S),$ $w\tau(S)$, $p\tau(S)$, $EQ(S)$ will respectively denote the set of reflexive, transitive, weakly transitive, proto transitive, and equivalence relations on the set $S$ respectively. Clearly, $w\tau(S) \,\subseteq\, p\tau(S)$.
\begin{proposition}
$\forall {R\in Ref(S)}(R\in p\tau(S) \leftrightarrow \tau(R)\in EQ(S))$.
\end{proposition}
\begin{definition}
A \emph{Partial Algebra} $P$ is a tuple of the form \[\left\langle\underline{P},\,f_{1},\,f_{2},\,\ldots ,\, f_{n}, (r_{1},\,\ldots ,\,r_{n} )\right\rangle\] with $\underline{P}$ being a set, $f_{i}$'s being partial function symbols of arity $r_{i}$. The interpretation of $f_{i}$ on the set $\underline{P}$ should be denoted by $f_{i}^{\underline{P}}$, but the superscript will be dropped in this monograph as the application contexts are simple enough. If predicate symbols enter into the signature, then $P$ is termed a \emph{Partial Algebraic System}. (see \cite{BU,LJ} for the basic theory)
\end{definition}
In a partial algebra, for term functions $p,\,q$, \[p\stackrel{\omega}{=}q\; \mathrm{iff} \;(\forall x\,\in\,dom(p)\,\cap\,dom(q)) p(x)=q(x).\] The \emph{weak strong equality} is defined via, \[p\stackrel{\omega^{*}}{=}q\; \mathrm{iff} \;(\forall x\,\in\,dom(p)\,=\,dom(q)) p(x)=q(x).\] For two terms $s,\,t$, $s\,\stackrel{\omega}{=}\,t$ shall mean, if both sides are defined then the two terms are equal (the quantification is implicit). $s\,\stackrel{\omega ^*}{=}\,t$ shall mean if either side is defined, then the other is and the two sides are equal (the quantification is implicit).
\section{Nelson Algebras}
By a \emph{De Morgan lattice} $\Delta ML$ we will mean an algebra of the form $L\,=\,\left\langle \underline{L},\, \vee,\, \wedge , \, c , \, 0,\, 1 \right\rangle $ with $\vee,\,\wedge$ being distributive lattice operations and $c$ satisfying
\begin{mitemize}
\item {$x^{cc}\,=\, x\;;\;(x\vee y)^{c}\,=\,x^{c}\,\wedge\, y^{c}\, ;$}
\item {$(x\,\leq\, y\,\leftrightarrow\, y^{c}\, \leq\, x^{c})\;;\; (x\,\wedge\, y)^{c}\,=\,x^{c}\,\vee\, y^{c}\, ;$}
\end{mitemize}
It is possible to define a partial unary operation $\star$, via $x^{\star}\,=\, \bigwedge \{x\, :\, x\,\leq\,x^{c} \}$ on any $\Delta ML$. If it is total, then the $\Delta ML$ is said to be \emph{complete}. In a complete $\Delta ML$ $L$, we have
\begin{mitemize}
\item {$x^{\star}\,\nleq\, x^{c}\;;\;x^{\star \star}\,=\,x \, ;$}
\item {$(x\,\leq\, y\,\longrightarrow\, y^{\star}\, \leq\, x^{\star})$.}
\item {$x^{c}\,=\, \bigvee \{y\, :\, x^{\star}\,\nleq\, y\}$.}
\end{mitemize}
A $\Delta ML$ is said to be a \emph{Kleene algebra} if it satisfies $x\,\wedge\,x^{c}\, \leq\, y\,\vee\, y^{c} $. If
$L^{+}\,=\, \{x\,\vee\, x^{c}\, :\, x\in L\}$ and $L^{-}\,=\, \{x\,\wedge\, x^{c}\, :\, x\in L\}$, then in a Kleene algebra we have
\begin{mitemize}
\item {$(L^{-})^{c}\,=\, L^{+}$ is a filter and $(L^{+})^{c}\,=\, L^{-}$ is an ideal. }
\item {$(\forall a, b\in L^{-})\,a\,\leq\, b^{c} $; $(\forall a, b\in L^{+})\,a^{c}\,\leq\, b$.}
\item {$x\in L^{-}$ if and only if $x \,\leq\, x^{c}$.}
\end{mitemize}
A \emph{Heyting algebra} $K$, is a relatively pseudo-complemented lattice, that is $(\forall a, b)\, a\,\Rightarrow b\,=\, \bigvee \{x\,;\, a\wedge x\,\leq \, b\}\,\in\, K$.
A \emph{Quasi-Nelson} algebra $Q$ is a Kleene algebra that satisfies $(\forall a, b)\, a\Rightarrow (a^{c}\vee b )\, \in \,Q$. $a\Rightarrow (a^{c}\vee b )$ is abbreviated by $a\rightarrow b$ below. Such an algebra satisfies all of the sentences N1--N4:
\begin{align*}
x\rightarrow x\,=\, 1\tag{N1}\\
(x^{c}\vee y)\,\wedge\, (x\rightarrow y)\,=\,x^{c}\vee y\tag{N2}\\
x\,\wedge\, (x\rightarrow y)\,=\, x\, \wedge\, (x^{c}\vee y)\tag{N3}\\
x\rightarrow (y\wedge z)\,=\, (x\rightarrow y)\,\wedge\, (x\rightarrow z)\tag{N4}\\
(x\wedge y)\rightarrow z\,=\, x\rightarrow (y\rightarrow z).\tag{N5}
\end{align*}
A Nelson algebra is a quasi-Nelson algebra satisfying \textsf{N5}. A Nelson algebra can also be defined directly as an algebra of the form $\left\langle A, \vee , \wedge, \rightarrow , c , 0, 1 \right\rangle $ with $\left\langle A, \vee , \wedge, c , 0, 1 \right\rangle $ being a Kleene algebra with the binary operation $\rightarrow$ satisfying \textsf{N1--N5}.
\section{Granules and Granular Computing Paradigms}
The idea of granular computing is as old as human evolution. Even in the available information on earliest human habitations and dwellings, it is possible to identify a primitive granular computing process (\textsf{PGCP}) at work. This can for example be seen from the stone houses, dating to 3500 BCE, used in what is present-day Scotland. The main features of this and other primitive versions of the paradigm may be seen to be
\begin{mitemize}
\item {Problem requirements are not rigid.}
\item {Concept of granules may be vague.}
\item {Little effort on formalization right up to approximately the middle of the previous century.}
\item {Scope of abstraction is very limited.}
\item {Concept of granules may be concrete or abstract (relative all materialist viewpoints).}
\end{mitemize}
The precision based granular computing paradigm, traceable to Moore and Shannon's paper \cite{Sha56}, will be referred to as the \emph{classical granular computing paradigm} \textsf{CGCP} is usually understood as the granular computing paradigm (The reader may note that the idea is vaguely present in \cite{Sha48}). The distinct terminology would be useful to keep track of the differences with other paradigms. CGCP has since been adapted to fuzzy and rough set theories in different ways.
Granules may be assumed to subsume the concept of information granules -- information at some level of precision. In granular approaches to both rough and fuzzy sets, we are usually concerned with such types of granules. Some of the fragments involved in applying CGCP may be:
\begin{mitemize}
\item {Paradigm Fragment-1: Granules can exist at different levels of precision.}
\item {Paradigm Fragment-2: Among the many precision levels, choose a precision level at which the problem at hand is solved. }
\item {Paradigm Fragment-3: Granulations (granules at specific levels or processes) form a hierarchy (later development).}
\item {Paradigm Fragment-4: It is possible to easily switch between precision levels.}
\item {Paradigm Fragment-5: The problem under investigation may be represented by the hierarchy of multiple levels of granulations.}
\end{mitemize}
The different stages of development of granular computing paradigms are as in the following:
\begin{mitemize}
\item {Classical Primitive Paradigm till middle of previous century.}
\item {CGCP: Since Shannon's information theory}
\item {CGCP in fuzzy set theory. It is natural for most real-valued types of fuzzy sets, but even in such domains unsatisfactory results are normal. Type-2 fuzzy sets have an advantage over type-1 fuzzy sets in handling data relating to emotion words, for example, but still far from satisfactory. For one thing linguistic hedges have little to do with numbers. A useful reference would be \cite{LZ9}.}
\item {For a long period (up to 2008 or so), the adaptation of CGCP for RST has been based solely on precision and related philosophical aspects. The adaptation is described for example in \cite{Ya01}. In the same paper the hierarchical structure of granulations is also stressed. This and many later papers on CGCP (like \cite{TYL}) in rough sets speak of structure of granulations. }
\item {Some Papers with explicit reference to multiple types of granules from a semantic viewpoint include \cite{AM69,AM99,SW3,SW,AM105}.}
\item {The axiomatic approach to granularity initiated in \cite{AM99} has been developed by the present author in the direction of contamination reduction in \cite{AM240}. From the order-theoretic/algebraic point of view, the deviation is in a very new direction relative the precision-based paradigm. The paradigm shift includes a new approach to measures. }
\end{mitemize}
There are other adaptations of CGCP to soft computing like \cite{KCM} that we will not consider.
Unless the underlying language is restricted, granulations can bear upon the theory with unlimited diversity. Thus for example in classical \textsf{RST}, we can take any of the following as granulations: collection of equivalence classes, complements of equivalence classes, other partitions on the universal set $S$, other partition in $S$, set of finite subsets of $\mathcal{S}$ and set of finite subsets of $\mathcal{S}$ of cardinality greater than 2. This is also among the many motivations for the axiomatic approach.
A formal simplified version of the the axiomatic approach to granules is in \cite{AM1800}. The axiomatic
theory is capable of handling most contexts and is intended to permit relaxation of
set-theoretic axioms at a later stage. The axioms are considered in the framework of Rough
Y-Systems (\textsf{RYS}) that maybe seen as a generalized form of \emph{abstract
approximation spaces} \cite{CC5} and approximation framework
\cite{CD3}. It includes relation-based RST, cover-based RST and more. These structures are
provided with enough structure so that a classical semantic domain (Meta-C) and at least
one rough semantic domain (called Meta-R) of roughly equivalent objects along with
admissible operations and predicates are associable. But the exact way of association is
not something absolute as there is no real end to recursive approximation processes of
objects.
In the present monograph we will stick to successor, predecessor and related granules generated by elements and will avoid the precision based paradigm.
\chapter{Approximations and Definite Elements in PRAX}
\begin{definition}
By a \emph{Proto Approximation Space} $S$ (\textsf{PRAS for short}), we will mean a pair
of the form $\left\langle \underline{S},\, R \right\rangle $ with $\underline{S}$ being a
set and $R$ being a proto-transitive relation on it. If $R$ is also reflexive, then it
will be called a \emph{Reflexive Proto Approximation Space} (\textsf{PRAX}) for short).
$\underline{S}$ may be infinite.
\end{definition}
If $S$ is a PRAX or a PRAS, then we will respectively denote \emph{successor neighborhoods},
\emph{inverted successor or predecessor neighborhoods} and \emph{symmetrized
successor neighborhoods} generated by an element $x \in S$ as follows:
\[[x]\,=\, \{y ;\, Ryx\}.\]
\[[x]_{i}\,=\, \{y;\, Rxy\} .\]
\[[x]_{o}\,=\, \{y ;\, Ryx\, \&\, Rxy\}.\]
Taking these as granules, the associated granulations will be denoted by
$\mathcal{G} \,=\,\{[x]:\, x\in S \}$, $\mathcal{G}_{i}$ and $\mathcal{G}_{o}$ respectively. In all that
follows $S$ will be a \textsf{PRAX} unless indicated otherwise.
\begin{definition}
Definable approximations on $S$ include ($A\subseteq S$):
\begin{align*}
\tag{Upper Proto} A^{u}\,=\, \bigcup _{[x]\cap A \neq \emptyset} {[x]}.\\
\tag{Lower Proto} A^{l}\,=\, \bigcup_{[x]\subseteq A} {[x]}.\\
\tag{Symmetrized Upper Proto} A^{u_o}\,=\, \bigcup _{[x]_{o}\cap A \neq \emptyset}
{[x]_{o}}.\\
\tag{Symmetrized Lower Proto} A^{l_o}\,=\, \bigcup_{[x]_{o}\subseteq A} {[x]_{o}}.\\
\tag{Point-wise Upper} A^{u+}\,=\, \{ x \, :\, [x]\cap A \neq \emptyset \}.\\
\tag{Point-wise Lower} A^{l+}\,=\, \{x \,:\,[x]\subseteq A \}\,.
\end{align*}
\end{definition}
\begin{examp}
In the context of our example \ref{agre}, $[x]$ is the \emph{set of allies} $x$, while $[x]_{o}$ is the set of comrades of $x$. $A^{l}$ is the \emph{union of the set of all allies of at least one of of the members of} $A$ if they are all in $A$. $A^{u}$ is the union of the set of all allies of persons having at least one ally in $A$. $A^{l_{+}}$ is the set of all those persons in $A$ all of whose allies are within $A$. $A^{u_{+}}$ is the set of all those persons having allies in $A$.
\end{examp}
\begin{definition}
If $A\subseteq S$ is an arbitrary subset of a \textsf{PRAX} or a
\textsf{PRAS} $S$, then
\begin{gather}
A^{ux}\,=\, \bigcup _{[x]_{o}\cap A \neq \emptyset} {[x]}.\\
A^{lx}\,=\, \bigcup_{[x]_{o}\subseteq A} {[x]}. \\
A^{u*}= \bigcup \{[x] : [x]\cap A \neq \emptyset \& (\exists y)
([x],[y])\in \sigma (R), \, (x, y)\in R, \, x\neq y , [y]\subseteq A \}.\\
A^{l*}= \bigcup\{[x] : [x]\subseteq A \,\& (\exists y) (([x],[y])\in
\sigma(R),\, x\neq y, \, [y]\subseteq A ) \}.
\end{gather}
\end{definition}
The following inverted approximations are also of relevance as they provide Galois
connections in case of point-wise approximations (see \cite{JJ}) under particular
assumptions. Our main approximations of interest will be $l, u, l_{o}, u_{o}$.
\begin{definition}
In the context of the above definition, the following will be referred to as inverted
approximations:
\begin{gather*}
A^{ui}\,=\, \bigcup _{[x]_{i}\cap A \neq \emptyset} {[x]_{i}}\,\\
A^{li}\,=\, \bigcup _{[x]_{i}\subseteq A} {[x]_{i}}\,\\
A^{\vartriangle}\,=\, \{ x \, :\, [x]_{i}\cap A \neq \emptyset \}\\
A^{\triangledown}\,=\, \{x \,:\,[x]_{i}\subseteq A \}\,\\
\end{gather*}
\end{definition}
\begin{proposition}
In a PRAX $S$ and for a subset $A\subseteq S$, all of the following hold:
\begin{mitemize}
\item {$(\forall x) \,[x]_{o}\subseteq [x]$ }
\item {It is possible that $A^{l}\,\neq \, A^{l+}$ and in general, $A^{l}\,\parallel \,
A^{lo}$.}
\end{mitemize}
\end{proposition}
\begin{proof}
The proof of the first two parts are easy. For the third, we chase the argument up to a
trivial counter example (see the following chapter).
\[\bigcup_{[x]\subseteq A} {[x]}\,\subseteq \bigcup_{[x]_{o}\subseteq A} {[x]}\, \supseteq
\,\bigcup_{[x]_{o}\subseteq A} {[x]_{o}} \]
\[\bigcup_{[x]_{o}\subseteq A} {[x]_{o}} \supseteq \bigcup_{[x]\subseteq A} {[x]_{o}}\,
\subseteq \,\bigcup_{[x]\subseteq A} {[x]}. \]
\end{proof}
\begin{proposition}
For any subset $A$ of $S$, \[A^{u_{o}}\,\subseteq\, A^{u}.\]
\end{proposition}
\begin{proof}
Since $[x]_{o}\cap A\neq \emptyset$, therefore
\[A^{u_{o}}\,=\, \bigcup _{[x]_{o}\cap A \neq \emptyset} {[x]_{o}} \,\subseteq \, \bigcup
_{[x] \cap A \neq \emptyset} {[x]_{o}}\, \subseteq\,A^{u_{o}}\,=\, \bigcup _{[x]\cap A \neq
\emptyset} {[x]}\,=\, A^{u}.\]
\end{proof}
\begin{definition}
If $X$ is an approximation operator, then by a $X$-\emph{definite element},
we will mean a subset $A$ satisfying $A^{X}\,=\, A$. The set of all $X$-definite elements
will be denoted by $\delta_{X}(S)$, while the set of $X$ and $Y$-definite elements ($Y$
being another approximation operator) will be denoted by $\delta_{XY}(S)$. In particular,
we will speak of \emph{lower proto-definite}, \emph{upper proto definite} and
\emph{proto-definite} elements (those that are both lower and upper proto-definite).
\end{definition}
\begin{theorem}
In a \textsf{PRAX} $S$, the following hold:
\begin{mitemize}
\item {$\delta_{u}(S)\,\subseteq\,\delta_{u_{o}}(S)$, but $\delta_{l_{o}}(S)\,=\,
\delta_{u_{o}}(S)$ and $\delta_{u}(S)$ is a complete sublattice of $\wp (S)$ with respect
to inclusion. }
\item {$\delta_{l}(S)\,\parallel\,\delta_{l_{o}}(S)$ in general. ($\parallel$ means \emph{is not comparable}.)}
\item {It is possible that $\delta_{u}\,\nsubseteq \delta_{u_{o}}$.}
\end{mitemize}
\end{theorem}
\begin{proof}
\begin{mitemize}
\item {As $R$ is reflexive, if $A,\, B $ are upper proto definite, then $A\cup B$ and $A\cap B$
are both upper proto definite. So $\delta_{u}(S)$ is a complete sublattice of $\wp (S)$.}
\item {If $A\,\in\,\delta_{u} $, then $(\forall x \in A ) [x]\subseteq A$ and
$(\forall x \in A^{c}) [x]\cap A = \emptyset$.}
\item {So $(\forall x \in A^{c})\, [x]_{o}\cap A = \emptyset $.
But as $A\,\subseteq\,A^{u_{o}}$ is necessary, we must have $A\in \delta_{u_{o}}$.}
\end{mitemize}
\qed
\end{proof}
$A^{u+},\, A^{l+}$ have relatively been more commonly used in the literature and
have also been the only kind of approximation studied in \cite{JJ} for example (the
inverse relation is also considered from the same perspective).
\begin{definition}
A subset $B\,\subseteq\,A^{l+} $ will be said to be \emph{skeleton} of $A$ if and only if \[\bigcup_{x\in B} [x] \,=\, A^{l} ,\] and the set skeletons of $A$ will be denoted by $\mathbf{sk}(A)$.
\end{definition}
The skeleton of a set $A$ is important because it relates all three classes of approximations.
\begin{theorem}
In the context of the above definition, we have
\begin{mitemize}
\item {$\mathbf{sk}(A)$ is partially ordered by inclusion with greatest element $A^{l+}$.}
\item {$\mathbf{sk}(A)$ has a set of minimal elements $\mathbf{sk}_{m}(S)$.}
\item {$\mathbf{sk}(A)\,=\, \mathbf{sk}(A^{l})$}
\item {$\mathbf{sk}(A)\,=\, \mathbf{sk}(B)\,\leftrightarrow\, A^{l}=B^{l}\,\&\, A^{l+}= B^{l+}$.}
\item {If $B \in \mathbf{sk}(A)$, then $A^{l}\,\subseteq\, B^{u}$.}
\item {If $\cap \mathbf{sk}(A)\,=\, B$, then $A^{l_o}\,\cap\, \bigcup_{x\in B} [x]\,=\, \emptyset$.}
\end{mitemize}
\end{theorem}
\begin{proof}
Much of the proof is implicit in other results proved earlier in this chapter.
\begin{mitemize}
\item {If $x\in A^{l}\setminus A^{l+}$, then $[x]\nsubseteq A^{l}$ and many subsets $B$ of $A^{l+}$ are in $\mathbf{sk}(A)$. If $B\subset K\subset A^{l+}$ and $B \in \mathbf{sk}(A)$, then $K\in \mathbf{sk}(A)$. Further we have a minimal elements in the inclusion order (even if $A$ is infinite) by the induced properties of inclusion in $\wp (S)$.}
\item {has been proved above.}
\item {More generally, if we have $A^{l}\,\subseteq\, B\,\subseteq\, A$, then $B^{l}\,=\,A^{l} $. So $\mathbf{sk}(A)\,=\, \mathbf{sk}(A^{l})$. }
\item {Follows from definition.}
\item {If $B \in \mathbf{sk}(A)$, then $A^{l}\,=\,B^{l}\, \subseteq\, B^{u}$. }
\end{mitemize}
\qed
\end{proof}
\begin{theorem}
All of the following hold in PRAX:
\begin{mitemize}
\item {$(\forall A)\, A^{cl+} = A^{u+c},\,\,A^{cu+} = A^{l+c} $ - that is $l+$ and $u+$ are
mutually dual}
\item {$u+$ ($l+$ resp.) is a monotone $\vee$- (complete $\wedge$- resp.) morphism.}
\item {$\partial (A) = \partial(A^{c})$, where $\partial$ stands for the boundary operator.}
\item {$\Im (u+)$ (the image of $u+$) is an interior system while $\Im (l+)$ is a closure system.}
\item {$\Im (u+)$ and $\Im (l+)$ are dually isomorphic lattices.}
\end{mitemize}
\end{theorem}
\begin{theorem}
In a \textsf{PRAX}, $(\forall A\in \wp (S))\, A^{l+} \subseteq A^{l},\:\:A^{u+} \subseteq A^{u}$ and all of the following hold.
\begin{align*}
\tag{Bi}(\forall A\in \wp(S))\, A^{ll}\,=\, A^{l}\, \& \, A^{u} \subseteq\, A^{uu} .\\
\tag{l-Cup}(\forall A, B \in \wp(S))\, A^{l}\cup B^{l}\,\subseteq (A\cup B)^{l}.\\
\tag{l-Cap}(\forall A, B \in \wp(S))\, (A\cap B)^{l}\,\subseteq\, A^{l}\cap B^{l}.\\
\tag{u-Cup}(\forall A, B \in \wp(S))\,(A\cup B)^{u}\,=\, A^{u}\cup B^{u}.\\
\tag{u-Cap}(\forall A, B \in \wp(S))\,(A\cap B)^{u} \,\subseteq\, A^{u}\cap B^{u} .\\
\tag{Dual}(\forall A\in \wp (S))\, A^{lc}\,\subseteq\, A^{cu}.
\end{align*}
\end{theorem}
\begin{proof}
\begin{description}
\item[l-Cup]{For any $A, B\in \wp{S}$, $x\in (A\cup B)^{l}$
\begin{impemize}
\item {$(\exists y\in (A\cup B) )\, x\in [y]\, \subseteq \, A\cup B $.}
\item {$(\exists y\in A )\, x\in [y]\, \subseteq \, A\cup B $ or $(\exists y\in B
)\, x\in [y]\, \subseteq \, A\cup B $.}
\item {$(\exists y\in A) \, x\in [y]\, \subseteq \, A $ or $(\exists y\in A) \, x\in [y]\,
\subseteq \, B $ or $(\exists y\in B) \, x\in [y]\, \subseteq \, A $ or $(\exists y\in B)
\, x\in [y]\, \subseteq \, B $ - this is implied by $x\in A^{l} \cup B^{l}$.}
\end{impemize}}
\item[l-Cap]{For any $A, B\in \wp{S}$, $x\in (A\cap B)^{l}$
\begin{impemize}
\item {$ x\in \, A\cap B $}
\item {$(\exists y\in A\cap B )\, x\in [y]\, \subseteq \, A\cap B $ and $x\,\in\, A,\;
x\,\in\, B$}
\item {$(\exists y\in A) \, x\in [y]\, \subseteq \, A $ and $(\exists y\in B) \, x\in
[y]\,\subseteq \,B $ - Clearly this statement implies $x\in A^{l} \& x\in B^{l}$,
but the converse is not true in general.}
\end{impemize}}
\item[u-Cup]{$x\,\in\, (A\cup B)^{u}$
\begin{impemize}
\item {$x\,\in\, \bigcup_{[y]\cap (A\cup B)\neq \emptyset } [y] $}
\item {$x\,\in\, \bigcup_{([y]\cap A)\cup ([y]\cap B)\neq \emptyset} $}
\item {$x\,\in\,\bigcup_{[y]\cap A\neq \emptyset} [y] $ or $x\,\in\,\bigcup_{[y]\cap
B\neq \emptyset} [y] $}
\item {$x\in A^{u}\cup B^{u}$.}
\end{impemize}}
\item[u-Cap] {By monotonicity, $(A\cap B)\,\subseteq\, A^{u}$ and $(A\cap B)\,\subseteq\,
B^{u}$, so $(A\cap B)^{u}\,\subseteq\, A^{u}\cap B^{u} $.}
\item[Dual]{If $z\in A^{lc}$, then $ z\in [x]^{c}$ for all $[x]\subseteq A$ and either,
$z\in A\setminus A^{l}$ or $z\in A^{c}$. If $z\in A^{c}$ then $z\in A^{cu}$.
If $z\in A\setminus A^{l}$ and $z\neq A^{cu \setminus A^{c}}$ then $[z]\cap A^{c}\,=\,
\emptyset$. But this contradicts $z\notin A^{cu} \setminus A^{c}$.
So $(\forall A\in \wp (S))\, A^{lc}\,\subseteq\, A^{cu} .$}
\end{description}
\qed
\end{proof}
\begin{theorem}
In a \textsf{PRAX} $S$, all of the following hold:
\begin{align}
{(\forall A, B\in \wp (S))\, (A\cap B)^{l+}\,=\, A^{l+}\cap B^{l+}}. \\
{(\forall A, B\in \wp (S))\, A^{l+}\cup B^{l+}\, \subseteq (A\cup B)^{l+}}. \\
{(\forall A \in \wp (S))\, (A^{l+})^{c} = (A^{c})^{u+} \,\& \,A^{l+}\subseteq A^{l_{o}}}.
\end{align}
\end{theorem}
\begin{proof}
\begin{enumerate}
\item {$x\in (A\cap B)^{l+}$
\begin{impemize}
\item {$[x]\subseteq A\cap B $}
\item {$[x]\subseteq A $ and $[x]\subseteq B $}
\item {$x\in xA^{l+} $ and $x\in B^{l+}$.}
\end{impemize}}
\item {$x \in A^{l+}\cup B^{l+} $
\begin{impemize}
\item {$[x]\subseteq A^{l+} $ or $[x]\subseteq B^{l+}$}
\item {$[x]\subseteq A $ or $[x]\subseteq B$}
\end{impemize}
$\Rightarrow [x]\subseteq A\cup B\, \Leftrightarrow\, x\in (A\cup B)^{l+} $.}
\item {$z\in A^{l+c}$
\begin{impemize}
\item {$z\,\notin\, A^{l+}$}
\item {$[z]\,\nsubseteq A $}
\item {$z\cap A^{c}\,\neq \,\emptyset $}
\end{impemize}
}
\end{enumerate}
\qed
\end{proof}
\begin{theorem}
If $u+,\, l+ $ are treated as self maps on the power-set $\wp (S)$, $S$ being a PRAX or a
\textsf{PRAS} then all of the following hold:
\begin{mitemize}
\item {$(\forall x)\, x^{cl+}=x^{u+c},\,\,x^{cu+}=x^{l+c} $ - that is $l+$ and $u+$ are
mutually dual}
\item {$l+,\, u+$ are monotone.}
\item {$l+$ is a complete $\wedge$-morphism, while $u+$ is a $\vee$-morphism.}
\item {$\partial (x) = \partial(x^{c})$, where partial stands for the boundary operator.}
\item {$\Im (u+)$ is an interior system while $\Im (l+)$ is a closure system.}
\item {$\Im (u+)$ and $\Im (l+)$ are dually isomorphic lattices.}
\end{mitemize}
\end{theorem}
\begin{theorem}
\[\mathrm{In\; a\;} \mathsf{PRAX}\; S, (\forall A \subseteq S)\, A^{l+} \subseteq A^{l},\:\:A^{u+} \subseteq A^{u}.\]
\end{theorem}
\begin{proof}
\begin{mitemize}
\item {If $x\in A^{l+}$, then $[x]\subseteq A$ and so $[x]\subseteq A^{l},\, x\in A^{l}$.}
\item {If $x\in A^{l}$, then $(\exists y \in A) [y]\subseteq A,\, Rxy$. But it is possible
that $[x] \nsubseteq A$, therefore it is possible that $x\notin A^{l+}$ and
$A^{l}\nsubseteq A^{l+}$.}
\item {If $x\in A^{u+}$, then $[x]\cap A \neq \emptyset$, so $x\in A^{u}$.}
\item {So $A^{u+}\subseteq A^{u}$.}
\item {Note that $x\in A^{u}$, \ifof $(\exists z\in S) \,x\in [z],\, [z]\cap A\neq
\emptyset $, but this does not imply $x\in A^{u+}$. }
\end{mitemize}
\qed
\end{proof}
\begin{theorem}
In a \textsf{PRAX} $S$, all of the following hold:
\begin{align}
{(\forall A\in \wp (S))\, A^{l+}\,\subseteq\, A^{l_{o}} }. \\
{(\forall A\in \wp (S))\, A^{u_{o}}\,\subseteq\, A^{u+} }. \\
{(\forall A\in \wp (S))\, A^{lc}\,\subseteq\, A^{cu}}.
\end{align}
\end{theorem}
\begin{proof}
\begin{enumerate}
\item {
\begin{mitemize}
\item {If $x\in A^{l+}$, then $[x]\subseteq A$.}
\item {But as $[x]_{o}\subseteq [x]$, $A^{l+}\subseteq A^{l_{o}}$.}
\end{mitemize}}
\item {This follows easily from definitions.}
\item {
\begin{mitemize}
\item {If $z\in A^{lc}$, then $ z\in [x]^{c}$ for all $[x]\subseteq A$ and either, $z\in
A\setminus A^{l}$ or $z\in A^{c}$.}
\item {If $z\in A^{c}$ then $z\in A^{cu}$.}
\item {If $z\in A\setminus A^{l}$ and $z\neq A^{cu \setminus A^{c}}$ then $[z]\cap A^{c}\,=\,
\emptyset$.}
\item {But this contradicts $z\notin A^{cu} \setminus A^{c}$.}
\item {So $(\forall A\in \wp (S))\, A^{lc}\,\subseteq\, A^{cu} .$}
\end{mitemize}}
\end{enumerate}
\qed
\end{proof}
From the above, we have the following relation between approximations in general ($A^{u_+}
\longrightarrow A^{u}$ should be read as $A^{u_+}$ \emph{is included in} $A^u$):
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[node distance=2cm, auto]
\node (Alp) {$A^{l_{+}}$};
\node (A0) [right of=Alp] {};
\node (A) [right of=A0] {$A$};
\node (Al) [above of=A0] {$A^{l}$};
\node (Alo) [below of=A0] {$A^{l_{o}}$};
\node (Auo) [right of=A] {$A^{u_{o}}$};
\node (Aup) [right of=Auo] {$A^{u_{+}} $};
\node (Au) [right of=Aup] {$A^{u}$};
\draw[->] (Alp) to node {}(Al);
\draw[->] (Alp) to node {}(Alo);
\draw[->] (Al) to node {}(A);
\draw[->] (Alo) to node {}(A);
\draw[->] (A) to node {}(Auo);
\draw[->] (Auo) to node {}(Aup);
\draw[->] (Aup) to node {}(Au);
\end{tikzpicture}
\end{center}
\caption{Relationship Between Approximations}
\end{figure}
If a relation $R$ is purely reflexive and not proto-transitive on a set $S$, then the
relation $\tau(R)\,=\, R\cap R^{-1}$ will not be an equivalence and for a $A\subset S$,
it is possible that $A^{u_{o} l}\subseteq A $ or $A^{u_{o} l}\parallel A$ or $A\subseteq
A^{u_{o} l}$.
\chapter{Motivation and Examples}\label{moex}
Generalized transitive relations occur frequently in general information systems,
but are often not recognized as such and there is hope for improved semantics and
\textsf{KI} relative the situation for purely reflexive relation based \textsf{RST}. Not
all of the definable approximations have been investigated in even closely related
structures of general \textsf{RST}. Contamination-free semantics \cite{AM240} for the
contexts are also not known. Finally these relate to \textsf{RYS} and variants. A proper
characterization of roughly equal (requal) objects is also motivated by \cite{AM240}.
\section{Abstract Example}
Let $\S\,=\, \{a, b, c, e, f, g, h, l, n\}$ and let $R$ be a binary relation on it
defined via
\begin{align*}
R\,=\,& \{(a,\,a),\,
(l,\,l),\,(n,\,n),\,(n,\,h),\,(h,\,n),\,(l,\,n),\,(g,\,c),\,(c,\,g)\\
& (g,\,l),\,(b,\,
g), \,(g,\,b ), \, (h, \,g ), \,(a,\,b),\,(b,\,c),\,(h,\,a),\,(a,\,c)\}.
\end{align*}
Then $\left\langle S,\, R\right\rangle $ is a \textsf{PRAS}.
If $P$ is the reflexive closure of $R$ (that is
$P\,=\, R\cup \Delta_{S}$), then $\left\langle S,\, P\right\rangle $ is a \textsf{PRAX}.
The successor neighborhoods associated with different elements of $S$ are as follows
(\textbf{E} is a variable taking values in $S$):
\begin{table}[h]
\caption{Successor Neighborhoods}
\begin{small}
\begin{tabular}[h]{|c|c|c|c|c|c|c|c|c|c|}
\hline
\textbf{E} & $a\,$ & $ b\,$ & $c\,$ & $g\,$ & $e\,$ & $f \,$ & $h\,$ & $l\,$ & $n\,$\\
\hline
\textbf{$[E]$} & $\{a, h \}$ & $\{b, c, g \}$ & $\{b, c, g \}$ & $\{b, c,
g, h \}$ &$\{e \}$ & $\{ f\}$ & $\{h,n \}$ & $\{l,g \}$ & $\{n,l,g,h\}$\\
\hline
\textbf{$[E]_{o}$} & $\{ a\}$ & $\{b, c, g \}$ & $\{ b, c, g\}$ &$\{b,
c, g \}$ & $\{e\}$ & $\{ f\}$ & $\{h, n \}$ & $\{ l\}$ &$\{n,h \}$\\
\hline
\end{tabular}
\end{small}
\end{table}
\begin{gather*}
\mathrm{If}\; A\,=\,\{a,\, h,\, f\},\\
\mathrm{then}\; A^{l}\,=\,\{a,\, h,\, f\},\\
A^{l_o}\,=\,\{a,\, f\}\; \mathrm{and}\;A^{l_o}\,\subset \, A^{l}. \\
\mathrm{If}\; F\,=\,\{l\},\\
\mathrm{then}\; F^{l}\,=\,\emptyset,\; F^{l_{o}}\,=\,F \\
\mathrm{and}\;F^{l}\,\subset \, F^{l_{o}}.
\end{gather*}
Now let $Z\,=\, N \cup S \cup X$, where $N$ is the set of naturals, $X$ is the set of
elements of the infinite sequences $\{x_{i}\},\,\{y_j\} $. Let $Q$ be a relation on $Z$
such that
\begin{gather}
Q\cap S^{2} = P,\\
Q\cap N^{2} \;\mathrm{is\; some\; equivalence},\\
(\forall i\in N) (i, x_{3i +1}),\, (x_{2i}, i), \, (x_{i}, x_{i+1}),\,(y_{i}, y_{i+1})\in Q.
\end{gather}
$Q$ is then a proto-transitive relation. For any $i\in N$, let $P_{i}\,=\,\{y_{k}:\,k\neq 2j \& k < i \}\,\cup\,\{x_{2j}:\, 2j <
i\}$ - this will be used in later chapters. The extension of the example to involve nets and densely ordered subsets is standard.
\section{Caste Hierarchies and Interaction}
The caste system and religion are among the deep-seated evils of Indian society that
often cut across socio-economic classes and level of education. For the formulation of
strategies aimed at large groups of people towards the elimination of such evils it would
be fruitful to study interaction of people belonging to different castes and religions on
different social fronts.
Most of these castes would have multiple subcaste hierarchies in addition.
Social interactions are necessarily constrained by their type and untouchability
perception. If $x, \, y$ are two castes, then with respect to a possible social
interaction $\alpha$, people belonging to $x$ will either regard people belonging to $y$
as untouchable or otherwise. As the universality is so total, it is possible to write
$\mathbb{U}_{\alpha}xy$ to mean that $y$ is untouchable for $x$ for the interaction
$\alpha$. Usually this is a asymmetric relation and $y$ would be perceived as a
\emph{lower caste} by members of $x$ and many others.
Other predicates will of course be involved in deciding on the possibility of the social
interaction, but if $\mathbb{U}_{\alpha} xy$ then the interaction is forbidden relative
$x$. If $\alpha$ is "context of possible marriage", then the complementary relation
($\mathbb{C}_{\alpha}$ say) is a reflexive proto-transitive relation. For various other
modes of interaction similar relations may be found.
In devising remedial educational programmes targeted at mixed groups, it would be important
to understand approximate perceptions of the group and the semantics of PRAX
would be very relevant.
\section{Compatibility Prediction Models}
When we want to predict compatibility among individuals or objects, then the following model can be used. Specific examples include situations involving data from dating sites like OK-Cupid.
Let one woman be defined by a sequence of sets of features $a_1 ,\, \ldots,\, a_n$ at different temporal instants and another woman by
$b_1 ,\, \ldots,\, b_n$. Let $\omega (a_{i},\, b_{i}) $ be the set of features that are desired by $a_{i}$, but missing in $b_{i}$. Let $\rho$ be an equivalence relation on a subset $K$ of $S$ -- the set of all features, that determines the classical rough approximations $l_{\rho}, u_{\rho}$ on $\wp (K)$.
Let $(a, b) \in R$ if and only if $(\omega (a_{n},\, b_{n})^{l_{\rho}}$ is \emph{small} (for example, that can mean being an atom of $\wp (K)$). The predicate $R$ is intended to convey \emph{may like to be related}. In dating sites, this is understood in terms of profile matches: if a woman's profile matches another woman's and conversely and similarly with another woman's, then the other two woman are assumed to be mutually compatible.
\begin{proposition}
$R$ is a proto-transitive relation and $\left \langle \underline{S},\, R \right \rangle $ is a \textsf{PRAS}.
\end{proposition}
\begin{proof}
Obviously $R$ need not be reflexive or symmetric in general.
If $(a, b),\, (b, c),\, (b, a),\, (c, b) \,\in \, R$, then $(a, c),\, (c, a) \, \in\, R$ is a reasonable rule.
\end{proof}
So we have a concrete example of a \textsf{PRAS} that is suggestive of many more practical contexts.
\section{Indeterminate Information System Perspective}
It is easy to derive \textsf{PRAX} from population census, medical, gender studies and other databases and these correspond to information systems. We make the connection clearer through this example.
If our problem is to classify a specific population $O$, for a purpose based on scientific data on sex, gender continuum, sexual orientation and other factors, then our data base would be an indeterminate information system of the form \[\mathcal{I}\,=\, \left\langle O,\, At,\, \{V_{a} :\, a\in At\},\, \{\varphi _{a} :\, a\in At\} \right\rangle ,\] where $At$ is a set of attributes, $V_{a}$ a set of possible values corresponding to the attribute $a$ and $\varphi_{a}: O\, \longmapsto\, \wp(V_{a}) $ the valuation function. Sex is determined by many attributes corresponding to hormones, brain structure, karyotypes, brain configuration, anatomy, clinical sex etc. We can associate free/bound values of over six hormones, the values of which vary widely over populations. Suppose we are interested in a subset of attributes for which the inclusion/ordering of values (corresponding to any one of the
attributes in the subset) of an object in another is relevant. We may, for example, be interested in patterns in sexual compatibility/relationships corresponding to such inclusions. This relation is proto-transitive. Formally for a $B\,\subseteq \, At$, if we let $(x,\,y)\,\in\, \rho_{B}$ if and only if $(\exists a\in B) \varphi_{a} x \,\subseteq \, \varphi_{a} y$, then $\rho_{B}$ is often proto-transitive via another predicate on $B$.
\chapter{Algebras of Rough Definite Elements}
In this chapter we prove key results on the fine structure of definite elements.
\begin{theorem}
On the set of proto definite elements $\delta_{lu}(S)$ of a \textsf{PRAX} $S$, we can
define the following:
\begin{align}
{x\wedge y\,\stackrel{\Delta}{=}\, x\cap y .}\\
{x\vee y\,\stackrel{\Delta}{=}\, x\cup y .}\\
{0\,\stackrel{\Delta}{=}\, \emptyset .}\\
{1\,\stackrel{\Delta}{=}\,S .}\\
{x^{c}\,\stackrel{\Delta}{=}\,S \setminus x .}
\end{align}
\end{theorem}
\begin{proof}
We need to show that the operations are well defined. Suppose $x,\,y$ are proto-definite
elements, then
\begin{enumerate}
\item {\[(x\cap y)^{u}\subseteq x^{u}\cap y^{u} \,=\, x\cap y.\]
\[(x\cap y)^{l}\,=\, (x^{u}\cap y^{u})^{l} \,=\, (x \cap y)^{ul} = (x\cap y)^{u}
\,=\, x\cap y .\] Since $a^{ul} = a^{u}$ for any $a$.}
\item {\[(x\cup y)^{u}\,=\, x\cup y\,=\, x^{l}\cup y^{l}\subseteq (x\cup y)^{l}.\]}
\item {$0\,\stackrel{\Delta}{=}\, \emptyset$ is obviously well defined.}
\item {Obvious.}
\item {Suppose $A\in \delta_{lu}(S)$, then $(\forall z\in A^{c}) \,[z]\cap A \,=\,
\emptyset$ is essential, else $[z]$ would be in $A^{u}$. This means $[z]\subseteq A^{c}$
and so $A^{c}\,=\, A^{cl}$. If there exists a $a\in A$ such that $[a]\cap A^{c}\neq
\emptyset$, then $[a]\subseteq A^{u}\,=\, A$. So $A^{c}\in \delta_{lu}(S)$. }
\end{enumerate}
\qed
\end{proof}
\begin{theorem}
The algebra $\delta_{proto}(S)\,=\, \left\langle \delta_{lu}(S), \vee,\wedge, c , 0, 1
\right\rangle $ is a Boolean lattice.
\end{theorem}
\begin{proof}
Follows from the previous theorem. The lattice order can be defined via,
$x\leq y$ \ifof \, $\,x\cup y = y$ and $x\cap y = x$.
\end{proof}
\chapter{The Representation of Roughly Objects}
The representation of roughly equal elements in terms of definite elements are well known in case of classical rough set theory. In case of more general spaces including tolerance spaces \cite{AM240}, most authors have been concerned with describing the interaction of rough approximations of different types and not of the interaction of roughly equal
objects. Higher order approaches, developed by the present author as in \cite{AM99} for
bitten approximation spaces, permit constructs over sets of roughly equal objects. In the
light of the contamination problem \cite{AM99,AM240}, it would be an improvement to
describe without higher order constructs. In this chapter a new method of representing
roughly equal elements based on expanding concepts of definite elements is developed.
\begin{definition}\label{lesspre}
On $\wp (S)$, we can define the following relations:
\begin{gather*}
A\,\preceq \,B\,\mathrm{if \; and \; only \; if} \, A^{l}\subseteq B^{l}\,\& \,A^{u}\subseteq B^{u}.\tag{Rough Inclusion}\\
A\,\approx\, B\,\mathrm{if \; and \; only \; if}\, A\,\preceq\,B \,\& \, B\,\preceq\, A. \tag{Rough Equality}
\end{gather*}
\end{definition}
\begin{proposition}
The relation $\preceq$ defined on $\wp (S)$ is a bounded partial order and $\approx$ is an equivalence. The quotient $\wp(S)|\approx$ will be said to be the set of \emph{roughly equivalent objects}.
\end{proposition}
\begin{definition}
A subset $A$ of $\wp (S)$ will be said to a set of \emph{roughly equal} elements \ifof
\[(\forall x, y\in A)\, x^{l}\,=\, y^{l}\,\& \, x^{u}\,=\,y^{u}.\]
It will be said to be \emph{full} if no other subset properly including $A$ has the
property.
\end{definition}
Relative the situation for a general \textsf{RYS}, we have
\begin{theorem}[Meta-Theorem]
In a \textsf{PRAX} $S$, full set of roughly equal elements is necessarily a union of
intervals in $\wp(S)$.
\end{theorem}
\begin{definition}
A non-empty set of non singleton subsets $\alpha\,=\, \{x\,:\,x\subseteq \wp(S)\}$ will be
said to be a \emph{upper broom} \ifof all of the following hold:
\begin{gather*}
{(\forall x, y \in \alpha)\, x^{u}\,=\, y^{u} . }\\
{(\forall x, y \in \alpha)\, x\,\parallel \, y. }\\
{\mathrm{If\;} \alpha \subset \beta, \mathrm{\;then\;} \beta \mathrm{\;fails\; to\; satisfy\; at\; least\;
one \;of \;the\; above\; two\; conditions.}}
\end{gather*}
The set of upper brooms of $S$ will be denoted by $\pitchfork (S)$.
\end{definition}
\begin{definition}
A non-empty set of non singleton subsets $\alpha\,=\, \{x\,:\,x\subseteq \wp(S)\}$ will be
said to be a
\emph{lower broom} \ifof all of the following hold:
\begin{gather}
{(\forall x, y \in \alpha)\, x^{l}\,=\, y^{l}\neq x .}\\
{(\forall x, y \in \alpha)\, x\,\parallel \, y .}\\
{\mathrm{\;If\;} \beta \subset \alpha \& Card(\beta)\geq 2, \mathrm{\;then\;} \beta \mathrm{\;fails\; to\; satisfy\; condition\;}(1) \;\mathrm{or}\;(2). }
\end{gather}
The set of lower brooms of $S$ will be denoted by $\psi (S)$.
\end{definition}
\begin{proposition}
If $x\in \delta_{lu}(S)$ then $\{x\}\notin \pitchfork (S)$ and $\{x\}\notin \psi (S)$.
\end{proposition}
In the next definition, the concept of union of intervals in a partially ordered set is
modified in a way for use with specific types of objects.
\begin{definition}
By a \emph{bruinval}, we will mean a subset of $\wp (S)$ of one of the following forms:
\begin{mitemize}
\item {Bruinval-0: Intervals of the form $(x, y),\,[x, y),\, [x, x],\, (x, y]$ for
$x,y\in \wp(S)$.}
\item {Open Bruinvals: Sets of the form $[x, \alpha)\,=\, \{z\, : \, x\, \leq\, z\,
< \, b\,\&\, b\in \alpha\}$, $(x, \alpha]\,=\, \{z\, : \, x\, < \, z\,
\leq\, b\,\&\, b\in \alpha\}$ and $(x, \alpha)\,=\, \{z\, : \, x\, < \, z\,
< \, b\,, b\in \alpha\}$ for $\alpha \in \wp(\wp (S))$.}
\item {Closed Bruinvals: Sets of the form $[x, \alpha]\,=\, \{z\, : \, x\, \leq\, z\,
\leq\, b\,\&\, b\in \alpha\}$ for $\alpha \in \wp(\wp (S))$.}
\item {Closed Set Bruinvals: Sets of the form $[\alpha, \beta]\,=\,\{z\, : \, x\, \leq\,
z\,\leq\, y\,\&\, x\in \alpha \& y\in \beta\}$ for $\alpha , \beta \in \wp(\wp (S))$ }
\item {Open Set Bruinvals: Sets of the form $(\alpha, \beta)\,=\,\{z\, : \, x\, < \,
z\,< \, y\,, x\in \alpha \& y\in \beta\}$ for $\alpha , \beta \in \wp(\wp (S))$.}
\item {Semi-Closed Set Bruinvals: Sets of the form $[[\alpha, \beta]]$ defined as
follows: $\alpha = \alpha_{1}\cup\alpha_{2} $, $\beta = \beta_{1}\cup\beta_{2}$ and
$[[\alpha,\beta]]\,=\,(\alpha_{1},\beta_{1})\cup
[\alpha_{2},\beta_{2}]\cup(\alpha_{1},\beta_{2}]\cup [\alpha_{2},\beta_{1}) $ for $\alpha
, \beta \in \wp(\wp (S))$.}
\end{mitemize}
\end{definition}
In the example of the second chapter, the representation of the rough object
$(P_{i}^{l}, P_{i}^{u})$ requires set bruinvals.
\begin{proposition}
If $S$ is a \textsf{PRAX}, then a set of the form $[x, y]$ with $x, y\in \delta_{lu} (S)$
will be a set of roughly equal subsets of $S$ \ifof $x\,=\, y$.
\end{proposition}
\begin{proposition}
A bruinval-0 of the form $(x, y)$ is a full set of roughly equal
elements if
\begin{center}
\begin{mitemize}
\item {$x, y \in \delta_{lu} (S)$,}
\item {$x$ is covered by $y$ in the order on $\delta_{lu} (S)$.}
\end{mitemize}
\end{center}
\end{proposition}
\begin{proposition}
If $x, y\in \delta_{lu} (S)$ then sets of the form $[x, y),\, (x,
y]$ cannot be a non-empty set of roughly equal elements, while those of the form $[x, y]$
can be \ifof $x = y$.
\end{proposition}
\begin{proposition}
A bruinval-0 of the form $[x, y)$ is a full set of roughly equal
elements if
\begin{center}
\begin{mitemize}
\item {$x^{l}, y^{u}\in \delta_{lu} (S)$, $x^{l} = y^{l}$ and $x^{u} = y^{u}$,}
\item {$x^{l}$ is covered by $y^{u}$ in $\delta_{lu} (S)$ and}
\item {$x\setminus (x^{l})$ and $y^{u}\setminus y$ are singletons}
\end{mitemize}
\end{center}
\end{proposition}
\begin{remark}
In the above proposition the condition $x^{l}, y^{u}\in \delta_{lu} (S)$, is not
necessary.
\end{remark}
\begin{theorem}
If a bruinval-0 of the form $[x, y]$ satisfies
\begin{gather*}
{x^{l} = y^{l} = x\, \&\, x^{u} = y^{u} .}\\
{Card(y^{u}\setminus y) = 1.}
\end{gather*}
then $[x, y]$ is a full set of roughly equal objects.
\end{theorem}
\begin{proof}
Under the conditions, if $[x, y]$ is not a full set of roughly equal objects, then there
must exist at least one set $h$ such that $h^{l} = x$ and $h^{u} = y^{u}$ and $h\notin
[x, y]$. But this contradicts the order constraint $x^{l}\, \leq \, h \, y^{u}$. Note
that $y^{u}\notin [x, y]$ under the conditions.
\qed
\end{proof}
\begin{theorem}
If a bruinval-0 of the form $(x, y]$ satisfies
\begin{gather*}
{x^{l} = y^{l} = x \;\&\; (\forall z\in (x, y])\,z^{u} = y^{u},}\\
{Card(y^{u}\setminus y)=1.}
\end{gather*}
then $(x, y]$ is a full set of roughly equal objects, that does not intersect the full set
$[x, x^{u}]$.
\end{theorem}
\begin{proof}
By monotonicity it follows that $(x, y]$ is a full set
of roughly equal objects. then there
must exist at least one set $h$ such that $h^{l} = x$ and $h^{u} = y^{u}$ and $h\notin
[x, y]$. But this contradicts the order constraint $x^{l}\, \leq \, h \, y^{u}$. Note
that $y^{u}\notin [x, y]$ under the conditions.
\qed
\end{proof}
\begin{theorem}
A bruinval-0 of the form $(x^{l}, x^{u})$ is not always a set of roughly equal
elements, but will be so when $x^{uu}= x^{u}$. In the latter situation it will be full if
$[x^{l},x^{u})$ is not full.
\end{theorem}
The above theorems essentially show that the description of rough objects depends on too
many types of sets and the order as well. Most of the considerations extend to other
types of bruinvals as is shown below and remain amenable.
\begin{theorem}
An open bruinval of the form $(x, \alpha)$ is a full set of roughly equal elements \ifof
\begin{gather*}
{\alpha \,\in\,\pitchfork (S).}\\
{(\forall y\in \alpha)\, x^{l}\,=\, y^{l}, x^{u}\,=\, y^{u}}\\
{(\forall z) (x^{l}\subseteq z \subset x \longrightarrow z^{u}\subset x^{u}).}
\end{gather*}
\end{theorem}
\begin{proof}
It is clear that for any $y\in \alpha$, $(x,y)$ is a convex interval and all elements in
it have same upper and lower approximations. The third condition ensures that $[z,\alpha)$
is not a full set for any $z\in [ x^{l}, x)$.
\qed
\end{proof}
\begin{definition}
An element $x\in \wp (S)$ will be said to be a \emph{weak upper critical element relative}
$z\subset x$ \ifof $(\forall y\in \wp (S))\, (z\,=\,y^{l} \,\&\, x \subset y\,
\longrightarrow x^{u}\subset y^{u} )$.
An element $x\in \wp (S)$ will be said to be an \emph{upper critical element relative}
$z\subset x$ \ifof
$(\forall v,\,y\in \wp (S))\, (z\,=\,y^{l}\,=\,v^{l} \,\&\,v \subset x \subset y\,
\longrightarrow \,v^{u} = x^{u}\subset
y^{u} )$. Note that the inclusion is strict.
An element $a$ will be said to be \emph{bi-critical relative} $b$ \ifof $(\forall x,\,
y\in \wp (S)) (a\subset x \subseteq y \subset b\,\longrightarrow\, x^{u} = y^{u}\,\&\,
x^{l} = y^{l}\,\&\, x^{u} \subset b^{u}\,\&\, a^{l}\subset x^{l})$.
\end{definition}
If $x$ is an upper critical point relative $z$, then $[z,x)$ or $(z,x)$ is a set of
roughly equivalent elements.
\begin{definition}
An element $x\in \wp (S)$ will be said to be an \emph{weak lower critical element
relative} $z\supset x$ \ifof
$(\forall y\in \wp (S))\, (z\,=\,y^{u}\,\&\,y \subset x\, \longrightarrow \,y^{l}\subset
x^{l} )$.
An element $x\in \wp (S)$ will be said to be an \emph{lower critical element
relative} $z\supset x$ \ifof
$(\forall y, v \in \wp (S))\, (z\,=\,y^{u}\,=\, v^{u}\&\,y \subset x \subset v\,
\longrightarrow \,y^{l}\subset
x^{l}= v^{l} )$.
An element $x\in \wp (S)$ will be said to be an \emph{lower critical element} \ifof
$(\forall y\in \wp (S))\, (y \subset x\, \longrightarrow \,y^{l}\subset x^{l} )$
An element that is both lower and upper critical will be said to be \emph{critical}.
The set of upper critical, lower critical and critical elements respectively will be
denoted by $UC(S)$, $LC(S)$ and $CR(S)$.
\end{definition}
\begin{proposition}
In a \textsf{PRAX}, every upper definite subset is also upper critical, but the converse
need not hold.
\end{proposition}
The most important thing about the different lower and upper critical points is that they
help in determining full sets of roughly equal elements by determining the boundaries of
intervals in bruinvals of different types.
\section{Types of Associated Sets}
Because of reflexivity it might appear that lower approximations in \textsf{PRAX} and classical \textsf{RST} are too similar at least in the perspective of lower definite objects. It is necessary to classify subsets of a \textsf{PRAX} $S$, to see the differences relative the behavior of lower approximations in classical \textsf{RST}. We will make use of this in some of the semantics as well.
\begin{definition}
For each element $x\in \wp (S)$ we can associate the following sets:
\begin{gather}
F_{0}(x)\,=\, \{y \,:\,(\exists a\in x^{c}) \, Rya\, \& y\in x \} \tag{Forward Looking}\\
F_{1}(x)\,=\, \{y \,:\,(\exists a\in x^{c}) \, Rya\, \& R zy \,\& z\in x \} \tag{1-Forward Looking}\\
\pi _{0} (x)\,=\, \{y \,:\, y\in x \,\&\, (\exists a\in x^{c})\,Ray \} \tag{Progressive}\\
St (x)\,=\, \{y\,:\, [y]\,\subseteq\, x\, \&\, \neg (y\in F_{0}(x)) \} \tag{Stable}\\
Sym (x)\,=\, \{y\,:\, y\in x \,\&\, (\forall z\in x)(Ryz\, \leftrightarrow\, Rzy) \} \tag{Relsym}
\end{gather}
\end{definition}
\emph{Forward looking} set associated with a set $x$ includes those elements not in $x$ whose successor neighborhoods intersect $x$. Elements of the set may be said to be relatively forward looking. \emph{Progressive} set of $x$ includes those elements of $x$ whose successor neighborhoods are not included in $x$. It is obvious that progressive elements are all elements of $x\setminus x^{l} $. Stable elements are those that are strongly within $x$ and are not directly reachable in any sense from outside. $Sym(x)$ includes those elements in $x$ which are symmetrically related to all other elements within $x$.
Even though all these are important we cannot easily represent them in the rough domain. Their approximations have the following properties:
\begin{proposition}
In the above context, we have
\begin{gather*}
(\pi_{0}(x))^{l}\,=\, \emptyset \, \& \, (\pi_{0}(x))^{u}\,\subseteq\,x^{u}\setminus x^{l} \\
(F_{0}(x))^{u}\,\subseteq\, x^{u} \\
St(x)^{l}\,\subseteq\, x^{l} \,\& \, F_{0}(x)\,=\,\emptyset\, \longrightarrow\,St(x)\,=\, x^{l+} \\
{Sym(x)}^{u}\,\subseteq \,x^{u} \,\&\, (Sym(x))^{l}\,\subseteq\, x^{l}.
\end{gather*}
\end{proposition}
\begin{proof}
Proof is fairly direct.
\qed
\end{proof}
\chapter{More on Representation of Rough Objects}
We have already shown in the previous chapter that the representation of rough objects by definite objects is not possible in a \textsf{PRAX}. So it is important to look at possibilities based on other types of derived approximations. We do this and solve the problem right up to representation theorems for the derived operators in this chapter.
\begin{definition}
If $x\in \wp (S)$, then
\begin{mitemize}
\item {Let $\Pi _{\heartsuit}^{o}(x)\,=\, \{y\,;\, x\,\subseteq\, y \,\&\, x^{l}\,=\,y^{l} \&\, y^{u}\,\subseteq\, x^{uu}\}$.}
\item {Form the set of maximal elements $\Pi _{\heartsuit}(x)$ of $\Pi _{\heartsuit}^{o}(x)$ with respect to the inclusion order.}
\item {Select a unique element $\chi (\Pi _{\heartsuit}(x)) $ through a fixed choice function $\chi$.}
\item {Form $(\chi (\Pi _{\heartsuit}(x)))^{u}$.}
\item {$x^{\heartsuit \chi}\,=\, (\chi (\Pi _{\heartsuit}(x)))^{u}$ will be said to be the \emph{almost upper approximation} of $x$ relative $\chi$. }
\item {$x^{\heartsuit \chi}$ will be abbreviated by $x^{\heartsuit}$ when we work with fixed $\chi$.}
\end{mitemize}
The choice function will be said to be \emph{regular} if and only if $(\forall x, y)\,(x\,\subseteq y \,\&\, x^{l}\,=\, y^{l}\,\longrightarrow \,\chi (\Pi _{\heartsuit}(x))\,=\, \chi (\Pi _{\heartsuit}(y)) )$. We will assume regularity unless specified otherwise in what follows.
\end{definition}
\begin{definition}
If $x\in \wp (S)$, then
\begin{mitemize}
\item {Let $\Pi _{\diamondsuit}^{o}(x)\,=\, \{y\,;\, x\,\subseteq\, y \,\&\, x^{l}\,=\,y^{l}\}$.}
\item {Form the set of maximal elements $\Pi _{\diamondsuit}(x)$ of $\Pi _{\diamondsuit}^{o}(x)$ with respect to the inclusion order.}
\item {Select a unique element $\chi (\Pi _{\diamondsuit}(x)) $ through a fixed choice function $\chi$.}
\item {$x^{\diamondsuit \chi}\,=\, \chi (\Pi _{\diamondsuit}(x))$ will be said to be the \emph{lower limiter} of $x$ relative $\chi$. }
\item {$x^{\diamondsuit \chi}$ will be abbreviated by $x^{\diamondsuit}$ when we work with fixed $\chi$.}
\end{mitemize}
\end{definition}
\begin{definition}
If $x\in \wp (S)$, then
\begin{mitemize}
\item {Let $\Pi _{\flat}^{o}(x)\,=\, \{y\,;\, y\,\subseteq\, x \,\&\, x^{u}\,=\,y^{u}\}$.}
\item {Form the set of maximal elements $\Pi _{\flat}(x)$ of $\Pi _{\flat}^{o}(x)$ with respect to the inclusion order.}
\item {Select a unique element $\xi (\Pi _{\flat}(x)) $ through a fixed choice function $\xi$.}
\item {$x^{\flat \xi}\,=\, \xi (\Pi _{\flat}(x))$ will be said to be the \emph{ upper limiter} of $x$ relative $\chi$. }
\item {$x^{\flat \xi}$ will be abbreviated by $x^{\flat}$ when we work with fixed $\xi$.}
\end{mitemize}
\end{definition}
\begin{proposition}\label{non1}
In the context of the above definition, the almost upper approximation satisfies all of the following:
\begin{gather}
(\forall {x})\, x\,\subseteq \, x^{\heartsuit} \tag{Inclusion}\\
(\forall {x})\, x^{\heartsuit}\,\subseteq\, x^{\heartsuit \heartsuit} \tag{Non-Idempotence}\\
(\forall x\, y)\,(x\,\subseteq \, y\,\subseteq\, x^{\heartsuit}\,\longrightarrow\, x^{\heartsuit}\,\subseteq \, y^{\heartsuit}) \tag{Cautious Monotony}\\
(\forall {x})\, x^{u}\,\subseteq \, x^{\heartsuit} \tag{Supra Pseudo Classicality}\\
S ^{\heartsuit}\,=\, S \tag{Top.}
\end{gather}
\end{proposition}
\begin{proof}
\begin{mitemize}
\item {Inclusion: Follows from the construction. If we have one element granules or successor neighborhoods included in $x$, then these must be in the lower approximation. If a granule $y$ is not included in $x$, but intersects it in $f$, then it is possible to include $f$ in each of $\Pi_{\heartsuit}(x)$. So inclusion follows.}
\item {Non-Idempotence: The reverse inclusion does not happen as $x^{u}\,\subseteq\, x^{uu}$. }
\item {Cautious monotony: It is clear that monotony can fail in general because of the choice aspect, but if we have $x\,\subseteq \, y\,\subseteq\, x^{\heartsuit}$, then $x^{l}\subseteq y^{l}$ and $y^{\heartsuit}$ has to be equal to $x^{\heartsuit}$ or include more granules because of regularity of the choice function.}
\item {Supra Pseudo Classicality: We use the adjective \emph{pseudo} because $u$ is not a classical consequence operator. In the construction of $x^{\heartsuit}$, we select from super-sets of $x^{l}$ that can generate maximal upper approximations and take the upper approximation of the selected. So that includes $x^{u}$ in general.}
\end{mitemize}
\qed
\end{proof}
We have used the names of conditions in relation to the standard terminology used in non-monotonic reasoning. The upper approximation operator $u$ is similar to classical consequence operator, but lacks idempotence. So the fourth property has been termed as \emph{supra} \emph{pseudo} \emph{classicality} as opposed to \emph{supra classicality}. This means we are in a more general domain of reasoning relative the domains of \cite{DM94}.
\begin{theorem}
In the context of \ref{non1}, we have the following additional properties:
\begin{gather*}
(\forall x)\, x^{\heartsuit} \subseteq x^{u \heartsuit} \tag{Sub Left Absorption}\\
(\forall x)\, x^{\heartsuit} \subseteq x^{\heartsuit u} \tag{Sub Right Absorption}\\
\boxdot(\forall x, y)\,(x^{u}\,=\, y^{u}\,\nrightarrow\,x^{\heartsuit}\,=\, y^{\heartsuit} ) \tag{No Left Logical Equivalence}\\
\boxdot (\forall x, y)\,(x^{\heartsuit}\,=\, y^{\heartsuit}\,\nrightarrow\,x^{l}\,=\, y^{l}) \tag{No Jump Equivalence}\\
\boxdot(\forall x, y, z)\,(x\subseteq y^{\heartsuit}\,\&\, z\subseteq x^{u}\,\nrightarrow\,z\subseteq y^{\heartsuit} ) \tag{No Weakening}\\
\boxdot(\forall x, y)\,(x\,\subseteq\, y\,\subseteq\, x^{u}\,\nrightarrow\,x^{\heartsuit}\,=\, y^{\heartsuit} ) \tag{No subclassical cumulativity}\\
(\forall x, y)\,x^{\heartsuit}\,\cap\, y^{\heartsuit}\,\subseteq\,(x^{u}\,\cap\, y^{u})^{\heartsuit} \tag{Distributivity}\\
(\forall x, y, z)\,(x\cup z)^{\heartsuit}\,\cap\, (y\cup z)^{\heartsuit}\,\subseteq\,(z \cup (x^{u}\,\cap\, y^{u}))^{\heartsuit} \tag{Weak Distributivity}\\
(\forall x, y, z)\,(x\,\cup\, y)^{\heartsuit}\,\cap\, (x\,\cup\, z)^{\heartsuit}\,\subseteq (x\,\cup\, (y\,\oplus\,z))^{\heartsuit}
\tag{Disjunction in Antecedent}\\
(\forall x, y)\,(x\cup y)^{\heartsuit}\, \cap\, (x\cup y^{c})^{\heartsuit}\,\subseteq\, x^{\heartsuit} \tag{Proof by Cases}\\
\mathrm{If}\; y \subseteq (x\cup z)^{\heartsuit},\; \mathrm{then}\; x\,\implies y\subseteq z^{\heartsuit} \tag{Conditionalization.}
\end{gather*}
\end{theorem}
\begin{proof}
\begin{description}
\item [Sub Left Absorption]{For any $x$, $x^{\heartsuit}$ is the upper approximation of a maximal subset $y$ containing $x$ such that $x^{l}\,=\, y^{l}$ and $x^{u\heartsuit}$ is the upper approximation of a maximal subset $z$ containing $x^u$ such that $x^{ul}\,=\,x^{u}\,=\, z^{l}$. Since, $x^{l}\,\subseteq\, x^{ul}$ and $x\,\subseteq\, x^{u}$, so $x^{\heartsuit} \subseteq x^{u \heartsuit}$ follows. }
\item [Sub Right Absorption]{Follows from the properties of $u$.}
\item [No Left Logical Equivalence]{Two subsets $x,\, y$ can have unequal lower approximations and equal upper approximations and so the implication does not hold in general. $\boxdot$ should be treated as an abbreviation for \emph{in general}.}
\item [No Jump Equivalence]{The reason is similar to that of the previous negative result.}
\item [No weakening]{In general if $x\subseteq y^{\heartsuit}\,\&\, z\subseteq x^{u}$, then it is possible that $x^{u} \subseteq y^{\heartsuit}$ or $y^{\heartsuit}\,\subseteq \, x^{u}$. So we cannot be sure about $z\subseteq y^{\heartsuit}$.}
\item [No Subclassical Cumulativity]{If $x\,\subseteq\, y\,\subseteq\, x^{u}$, then $x^{l}\,\subseteq \, y^{l} $ in general and so elements of $\Pi_{\heartsuit}(x)$ may be included in $\Pi_{\heartsuit}(y)$, the two may be unequal and we may not be able to use a uniform choice function on them. So we need not have $x^{\heartsuit}\,=\, y^{\heartsuit}$. }
\item [Distributivity]{If $ z \in x^{\heartsuit}\,\cap\, y^{\heartsuit}$, then $z \in (\chi(\Pi_{\heartsuit}(x)))^{u} $ and $z \in (\chi(\Pi_{\heartsuit}(y)))^{u} $. So if $z\in x^{l}$ and $z\in y^{l}$, then $z\in (x^{u}\,\cap\, y^{u})^{\heartsuit} $.
Since in general, $(a\cap b)^{u}\,\subseteq\, a^u \cap b^{u}$ and $(a^{u}\cap b^u )^l = (a^{u}\cap b^u )$, we have the required inclusion. }
\end{description}
\begin{gather*}
(\forall x, y, z)\,(x\cup z)^{\heartsuit}\,\cap\, (y\cup z)^{\heartsuit}\,\subseteq\,(z \cup (x^{u}\,\cap\, y^{u}))^{\heartsuit} \tag{Weak Distributivity}\\
(\forall x, y, z)\,(x\,\cup\, y)^{\heartsuit}\,\cap\, (x\,\cup\, z)^{\heartsuit}\,\subseteq (x\,\cup\, (y\,\oplus\,z))^{\heartsuit}
\tag{Disjunction in Antecedent}\\
(\forall x, y)\,(x\cup y)^{\heartsuit}\, \cap\, (x\cup y^{c})^{\heartsuit}\,\subseteq\, x^{\heartsuit} \tag{Proof by Cases}\\
\mathrm{If}\; y \subseteq (x\cup z)^{\heartsuit},\; \mathrm{then}\; x\,\implies y\subseteq z^{\heartsuit} \tag{Conditionalization.}
\end{gather*}
\qed
\end{proof}
\begin{proposition}
\begin{gather*}
(\forall x, y)(x^{\diamondsuit}\,=\, y^{\diamondsuit}\,\longrightarrow\, x^{l}\,=\, y^{l;})\\
(\forall x, y)(x^{\flat}\,=\, y^{\flat}\,\longrightarrow\, x^{u}\,=\, y^{u}.)
\end{gather*}
\end{proposition}
\begin{flushleft}
\textbf{Discussion}:
\end{flushleft}
In non monotonic reasoning, if $C$ is any consequence operator $:\wp(S)\,\longmapsto\, \wp(S)$, then the following named properties of crucial importance in semantics (in whatever sense, \cite{DM94,DM2003}):
\begin{gather}
A\subseteq B\subseteq C(A) \longrightarrow C(B) \subseteq C(A) \tag{Cut}\\
A\subseteq B\subseteq C(A) \longrightarrow C(B) = C(A) \tag{Cumulativity}\\
x\subseteq y\subseteq x^{u} \longrightarrow \, x^{\heartsuit}\,=\, y^{\heartsuit} \tag{subclassical subcumulativity}
\end{gather}
\begin{proposition}
In the context of the above definition, the lower limiter satisfies all of the following:
\begin{gather}
(\forall {x})\, x\,\subseteq \, x^{\diamondsuit} \tag{Inclusion}\\
(\forall {x})\, x^{\diamondsuit\diamondsuit}\,=\, x^{ \diamondsuit} \tag{Idempotence}\\
(\forall x\, y)\,(x\,\subseteq \, y\,\subseteq\, x^{\diamondsuit}\,\longrightarrow\, x^{\diamondsuit}\,= \, y^{\diamondsuit}) \tag{Cumulativity}\\
(\forall {x})\, x^{u}\,\subseteq \, x^{\diamondsuit} \tag{Upper Inclusion}\\
S ^{\diamondsuit}\,=\, S \tag{Top}\\
(\forall x\, y)\,(x\,\subseteq \, y\, \subseteq\, x^{\diamondsuit}\,\longrightarrow\, x^{\diamondsuit}\,= \, y^{\diamondsuit}) \tag{Cumulativity}
\end{gather}
\end{proposition}
\emph{The above proposition means that the upper limiter corresponds to ways of reasoning in a stable way in the sense that the aggregation of conclusions does not affect inferential power or cut-like amplification}.
We prove a limited concrete representation theorem for operators like $\heartsuit$ in special cases and $\diamondsuit$. The representation theorem is valid for similar operators in non-monotonic reasoning. The representation theorem permits us to identify cover based formulations of \textsf{PRAX}.
\begin{definition}
A collection of sets $\mathcal{S}$ will be said to be a \emph{closure system} of a type as per the following conditions:
\begin{gather*}
\tag{Closure System} (\forall \mathcal{H}\subseteq \mathcal{S})\, \cap \mathcal{H}\in \mathcal{S}. \\
\tag{U-Closure System} (\forall \mathcal{H}\subseteq \mathcal{S})\, (\cap \mathcal{H})^{u}\in \mathcal{S}. \\
\tag{L-Closure System} (\forall \mathcal{H}\subseteq \mathcal{S})\, (\cap \mathcal{H})^{l}\in \mathcal{S}. \\
\tag{LU-Closure System} (\forall \mathcal{H}\subseteq \mathcal{S})\, (\cap \mathcal{H})^{l},(\cap \mathcal{H})^{u}\, \in \mathcal{S}. \\
\tag{Bounded} (\exists 0 , \top \in \mathcal{S})(\forall X\in \mathcal{S})\, 0 \subseteq X \subseteq \top\, .
\end{gather*}
\end{definition}
\begin{proposition}
In a \textsf{PRAX} $S$, the set $\mathcal{U}(S)\,=\,\{x^{u}; \, x\in \wp(S) \} $ is not a bounded U-closure system.
\end{proposition}
\begin{proposition}
\[(\forall {x}) \, x^{\heartsuit u }\,\subseteq \, x^{u \heartsuit } .\]
\end{proposition}
\begin{proof}
Because $x^{l}\,\subseteq x^{u}$, an evaluation of possible granules involved in the construction of $x^{\heartsuit u }$ and $x^{u \heartsuit }$ proves the result.
\qed
\end{proof}
\begin{theorem}
In a \textsf{PRAX} $S$, the set $\mathcal{\heartsuit}(S)\,=\,\{x^{\heartsuit}; \, x\in \wp(S) \} $ is a bounded LU-closure system if the choice operation is regular.
\end{theorem}
\begin{proof}
\begin{mitemize}
\item {$x^{\heartsuit}$ is the upper approximation of a specific $y$ containing $x$ that is maximal subject to $x^{l}\,=\, y^{l}$.}
\item {$x^{\heartsuit u}$ is the upper approximation of the upper approximation of a specific $y$ containing $x$ that is maximal subject to $x^{l} = y^{l}$ and its upper approximation.}
\item {Clearly, \[(\chi(\Pi_{\heartsuit} (x))\,\cap\, \chi (\Pi_{\heartsuit} (y)))^{u}\,\subseteq\, (\chi(\Pi_{\heartsuit} (x)))^{u}\,\cap\, (\chi(\Pi_{\heartsuit} (y)))^{u}. \] }
\item {The expression on the right of the inclusion is obviously a union of granules in the \textsf{PRAX}. }
\item {From a constructive bottom-up perspective, let $p_{1},\, p_{2},\, \ldots \, p_{s}$ be a collection of subsets of $x\setminus x^{l}$ such that }
\end{mitemize}
\begin{gather*}
\cup p_{i}\subseteq x\setminus x^{l}\\
(\exists z)\, p_{i}^{u}\,=\, [z] \\
\cup_{i\neq j}(p_i \cap p_j ) \; \mathrm{is \; minimal \; on\; all \; such\; collections.}
\end{gather*}
\begin{mitemize}
\item {Now we add subsets $k(p_{i})$ of $x^{uu}\setminus x^{c}$ to $x$ to form the required maximal subset.}
\item {For the lower approximation part, we simple use the preservation of $l$ by $cap$.}
\end{mitemize}
\qed
\end{proof}
\begin{proposition}
For each $x\in \wp(S)$ let $x^{\curlyvee}\,=\, (x^{\heartsuit})^{u} $, then we have the following properties:
\begin{gather*}
(\forall x ) \, x \,\subseteq x^{\curlyvee} .\\
(\forall x )\, x^{\curlyvee \curlyvee}\,=\, x^{\curlyvee}.
\end{gather*}
\end{proposition}
\begin{proof}
$x^{\heartsuit u \heartsuit u}\,=\, x^{\heartsuit u}$. Because if we could add a part of a class that retains the equality of lower approximations, then that should be adjoinable in the construction of $x^{\heartsuit}$ as well.
\qed
\end{proof}
The following limited representation theorem can be useful for connections with covers.
\begin{definition}
Let $X$ be a set and $C\,:\, \wp (X)\,\longmapsto \,\wp(X)$ a map satisfying all the following conditions:
\begin{gather*}
(\forall A\in \wp(S))\, A\subseteq C(A) \tag{Inclusion}\\
(\forall A\in \wp(S))\, C (C(A))\,=\,C(A) \tag{Idempotence}\\
(\forall A, B \in \wp(S))\,(A\subseteq B \subseteq C(A)\,\longrightarrow\, C(A)\,\subseteq\, C(B)) \tag{Cautious Monotony,}
\end{gather*}
then $C$ will said to be a \emph{cautious closure operator} (\textsf{CCO}) on $X$.
\end{definition}
\begin{definition}\label{relevant}
Let $H\,=\, \left\langle\underline{H},\, \preceq\right\rangle$, be a partially ordered set over a set $\underline{H}$. A subset $\mathcal{K}$ of the set of order ideals $\mathcal{F}(H)$ of $H$ will be said to be \emph{relevant} for a subset $B\,\subseteq\, H$ (in symbols $\rho (\mathcal{K},\, H)$) if and only if the following hold:
\begin{gather*}
(\exists G\in \mathcal{K})(\forall P \in \mathcal{K})\, {P\,\subseteq\, G }. \\
(\forall P \in \mathcal{K})\, P\,\subseteq B .\\
\mathrm{For \; any}\; \mathcal{L}\subseteq \mathcal{F}(H), \; \mathrm{if}\; \mathcal{K}\subseteq \mathcal{L}, \; \mathrm{then}\\ (\exists \top \in \mathcal{L})(\forall Y \in \mathcal{L})\, Y\subseteq \top \neq H \,\&\, \cap \mathcal{L}\,=\, \cap \mathcal{K}.
\end{gather*}
\end{definition}
\begin{definition}
In the context of \textsf{Def.}\ref{relevant}, a map $\jmath : \wp(L) \,\longmapsto \, \wp (L)$ defined as below will be said to be \emph{safe}
\[\jmath(Z)\, =\,\left\{
\begin{array}{ll}
\cap \mathcal{K}, \:\mathrm{if \: all\; relevant \; collections\; for\; Z \; have\; same \; intersection.} \\
\cap\{ \alpha : Z\subseteq \alpha \in \mathcal{F}(H) \},\; \mathrm{else}.
\end{array}
\right. \]
\end{definition}
\begin{proposition}
A safe map $\jmath$ is a cautious closure operator.
\end{proposition}
\begin{proof}
The verification of idempotence and inclusion is direct.
\begin{mitemize}
\item {For $A, B\in \wp(L)$, if we have $A\subseteq B \subseteq \jmath (A) $,}
\item {then either $A\subseteq B \subseteq \jmath(B) \subseteq \jmath (A) $ or $A\subseteq B \subseteq \jmath (A)\subseteq \jmath (B) $ must be true. }
\item {If the former inclusions hold, then it is necessary that $\jmath (A) \,=\, \jmath (B)$.}
\item {If $\jmath (B)$ is defined as the the intersection of order ideals and $\jmath (A)$ as that of relevant subcollections, then it is necessary that $\jmath (A)\subseteq \jmath (B)$. So cautious monotony holds. It can also be checked that monotonicity fails in this kind of situation.}
\end{mitemize}
\qed
\end{proof}
\begin{theorem}
On every Boolean ordered unary algebra of the form \[\mathcal{H}\,=\,\left\langle \wp (H), \subseteq , C \right \rangle, \] there exists a partial order $\leq$ on $K$ such that $\left\langle \wp (K), \subseteq, \jmath \right \rangle $
is isomorphic to $\mathcal{H}$.
\end{theorem}
\chapter{Atoms in the POSET of Rough Objects}
\begin{definition}\label{less}
For any two elements $x, y\in \wp(S)|\approx$, let
\[x\,\leq\, y \,\ifsf \, (\forall a\in x)(\forall b\in y) a^{l}\subseteq
b^{l}\,\&\,a^{u}\subseteq b^{u}. \]
$\wp(S)|\approx$ will be denoted by $H$ in what follows.
\end{definition}
\begin{proposition}
The relation $\leq$ defined on $H$ is a bounded and directed partial order. The least
element will be denoted by $0$ ($0\,=\, \{\emptyset\}$) and the greatest by $1$
($1\,=\,\{S\}$).
\end{proposition}
\begin{definition}
For any $a, b\in H$, let $UB(a, b)\,=\, \{x\, :\,a\leq x \,\&\,b\leq x \}$ and
$LB(a, b)\,=\, \{x\, :\,x\leq a \,\&\,x\leq b \}$. By a \emph{s-ideal} (strong ideal) of
$H$, we will mean a subset $K$ that satisfies all of
\begin{gather*}
{(\forall x\in H)(\forall a\in K)(x\leq a\,\longrightarrow\, x\in K),}\\
{(\forall a, b\in K)\, UB(a, b)\cap K\neq \emptyset .}
\end{gather*}
An \emph{atom} of $H$ is any element that covers $0$. The set of atoms of $H$ will be
denoted by $At(H)$.
\end{definition}
\begin{theorem}
Atoms of $H$ will be of one of the following types:
\begin{description}
\item [Type-0]{Elements of the form $(\emptyset, [x])$, that intersect no other set of
roughly equivalent sets. }
\item [Type-1]{Bruinvals of the form $(\emptyset, \alpha)$, that do not contain full sets
of roughly equivalent sets.}
\item [Type-2]{Bruinvals of the form $(\alpha, \beta)$, that do not contain full sets of
roughly equivalent sets and are such that $(\forall x) x^{l}\,=\, \emptyset$.}
\end{description}
\end{theorem}
\begin{proof}
It is obvious that a bruinval of the form $(\alpha, \beta)$ can be an atom only if
$\alpha$ is the $\emptyset$. If not, then each element $x$ of the bruinval $(\emptyset,
\alpha)$ will satisfy $x^{l}=\emptyset \,\subset \, x^{u}$, thereby contradicting the
assumption that $(\alpha, \beta)$ is an atom.
If $[x]$ intersects no other successor neighborhood, then \[(\forall y\in (\emptyset,
[x])) y^{l}\,=\, \emptyset\,\&\,x^{u}\,=\, [x]\] and it will be a minimal set of
roughly equal elements containing $0$.
The other part can be verified based on the representation of possible sets of roughly
equivalent elements.
\qed
\end{proof}
\begin{theorem}
The partially ordered set $H$ is atomic.
\end{theorem}
\begin{proof}
We need to prove that any element $x$ greater than $0$ is either an atom or
there exists an atom $a$ such that $a \leq x$, that is
\[(\forall x)(\exists a\in At(H))(0 < x\,\longrightarrow\, a\leq x ) .\]
Suppose the bruinval $(\alpha,\beta)$ represents a non-atom, then it is necessary that
\[(\forall x\in \alpha)\, x^{l} \neq \emptyset \,\&\, x^{u}\subseteq S . \]
Suppose the neighborhoods included in $x^{u}$ are $\{[y]\,:\, y\in B\subseteq S\}$.
If all combinations of bruinvals of the form $(\emptyset,\gamma)$ formed from these
neighborhoods are not atoms, then it is necessary that the upper approximation of every
singleton subset of a set in $\gamma$ properly contains another non-trivial
upper approximation. This is impossible.
So $H$ is atomic.
\qed
\end{proof}
\chapter{Algebraic Semantics-1}
An algebraic semantics is a complete description of reasoning about rough objects
involved in the context of \textsf{PRAX} or \textsf{PRAS} or any particular
instances thereof. In the present author's view the objects of interest should be roughly equal elements in some sense and the semantics should avoid objects of other kinds (from other semantic domains) thereby contaminating the semantics. But in any perspective, semantics relative any semantic domain is of interest. When it comes to the question of defining sensible operations over rough objects, given the ontological constraints, there is scope for much variation.
If $A,\, B\,\in\, \wp (S)$ and $A\approx B$ then $A^{u}\approx B^{u}$ and $A^{l}\approx B^{l}$, but $\neg (A\approx A^{u}) $ in general. We have already seen that $\leq$ is a partial order relation on $\wp(S)|\approx$. In this chapter we will be working mostly on $\wp (S)|\approx$ and will use lower case Greek alphabets for elements in it.
\begin{theorem}
The following operations can be defined on $\wp (S)|\approx$ ($A, \, B\in \wp (S)$ and $[A],\,[B]$ are corresponding classes):
\begin{gather}
L[A]\,\stackrel{\Delta}{=}\,[A^{l}] \\
[A]\odot [B]\,\stackrel{\Delta}{=}\, [\bigcup _{X\in [A],\, Y\in [B] } (X\cap
Y)]\\
[A]\oplus [B]\,\stackrel{\Delta}{=}\, [\bigcup _{X\in [A],\, Y\in [B] } (X\cup
Y)]\\
U[A]\,\stackrel{\Delta}{=}\,[A^{u}]\\
[A]\cdot [B]\,\stackrel{\Delta}{=}\,\lambda (LB([A], [B]))\\
[A]\circledast [B]\,\stackrel{\Delta}{=}\,\lambda (UB([A], [B]))\\
[A] + [B]\,\stackrel{\Delta}{=}\,\{X \,:\, X^{l}\,=\, (A^{l}\cap B^{l})^{l}\,\&\,X^{u}\,=\, A^{u}\cup B^{u} \} \\
[A] \times [B]\,\stackrel{\Delta}{=}\,\{X \,:\,X^{l}\,=\, A^{l}\cup B^{l}\,\&\, X^{u}\,=\, A^{l}\cup B^{l}\,\cup (A^{u}\cap B^{u}) \\
[A]\otimes [B] \,\stackrel{\Delta}{=}\,\{X\,:\, X^{l}\,=\,A^{l}\cup B^{l}\,\&\, X^{u}\,=\,A^{u}\cup B^{u}\}.
\end{gather}
\end{theorem}
\begin{proof}
If $A\approx B$ then $A^{u}\approx B^{u}$ and $A^{l}\approx B^{l}$, but $\neg (A\approx A^{u}) $ in general.
\begin{enumerate}
\item {If $B\in [A]$, then $B^{l}\,=\, A^{l}$, $B^{u}\,=\, A^{u}$ and $L[A] = L[B] \, =\, [A^{l}]$.}
\item {$[A]\odot [B]\,\stackrel{\Delta}{=}\, [\bigcup _{X\in [A],\, Y\in [B] } (X\cap
Y)]$ is obviously well defined as sets of the form $[A]$ are elements of partitions}
\item {Similar to the above.}
\item {If $B\in [A]$, then $B^{u}\,=\, A^{u}$ and so $[B^{u}]\,=\, [A^{u}]$.}
\item {$[A]\cdot [B]\,\stackrel{\Delta}{=}\,\lambda (LB([A], [B]))$. }
\item {$[A]\circledast [B]\,\stackrel{\Delta}{=}\,\lambda (UB([A], [B]))$.}
\item {$ [A] + [B]\,\stackrel{\Delta}{=}\,\{X \,:\, X^{l}\,=\, A^{l}\cap B^{l}\,\&\,X^{u}\,=\, A^{u}\cup B^{u} \} $. As the definitions is in terms of $A^{l},\, B^{l}, A^{u},\, B^{u} $, so there is no issue. }
\item {Similar to above.}
\item {Similar to above.}
\end{enumerate}
\end{proof}
$+$, $\times$ and $\otimes$ will be referred to as \emph{pragmatic} aggregation, commonality and commonality operations as they are less ontologically committed to the classical domain and more dependent on the main rough domain of interest. $+$ and the other pragmatic operations cannot be compared by the $\leq$ relation and so do not confirm to intuitive understanding of the concepts of aggregation and commonality.
The following theorems summarize the essential properties of the defined operations:
\begin{theorem}
\begin{gather*}
\tag{L1}{LL(\alpha)\,=\, L(\alpha). }\\
\tag{L2}{(\alpha \, \leq \, \beta\, \longrightarrow\,L(\alpha)\,\leq \, L(\beta) ).}\\
\tag{L3}{(L(\alpha)\,=\, [\alpha]\,\longrightarrow\,\alpha\,=\, \{\alpha^{l}\} ).}\\
\tag{U1}{(U(\alpha)\,\cap \, UU(\alpha) \neq \emptyset \,\longrightarrow\, U(\alpha)\,=\, UU(\alpha)).}\\
\tag{U2}{(UU(\alpha)\,=\, \emptyset \, \nrightarrow \, U(\alpha)\,=\, \emptyset).}\\
\tag{U3}{(\alpha\,\leq\, \beta\,\longrightarrow\, U(\alpha)\,\leq \, U(\beta).}\\
\tag{U4}{(U(\alpha)\,=\, \alpha\,\longrightarrow\,\alpha\,=\, \alpha^{l}\,=\, \alpha^{u} ).}\\
\tag{U5}{UL(\alpha)\,\leq \, U(\alpha).}\\
\tag{U6}{LU(\alpha)\,=\, U(\alpha).}
\end{gather*}
\end{theorem}
\begin{proof}
Let $\alpha\in \wp (S)|\approx$, then we can associate a pair of lower and upper approximations denoted by $\alpha_l$ and $\alpha_u$ respectively. By $\alpha^u$ and $\alpha^l$ we mean the global operations respectively on the set $\alpha$ (seen as an element of $\wp(S)$). These take singleton values and so we do not really need the approximations $\alpha_l$ and $\alpha_{u}$ and shall use the former.
\textsf{Proof of L1:}
\begin{gather*}
\alpha\in\wp(S)|\approx,\;\mathrm{so}\; \alpha = \{X\,;\, \alpha_l \,=\, X^l \, \&\, \alpha_u \,=\,X^{u},\&\, X\in\wp (S) \}.\\
\alpha^{l}\,=\, \{X^{l}; \, X\,\in\, \alpha \}\,=\, \{\alpha_l \} \\
\mathrm{So}\; [\alpha^{l}]\,=\, \{Y\,; \, Y^{l}\,=\, \alpha^{l}\,\&\, Y^{u}\,=\, \alpha^{lu} \}.\\
(L(\alpha))^{l}\,=\, \{Y^l \,; \,Y^{l}\,=\, \alpha^{l}\,\&\, Y^{u}\,=\, \alpha^{lu} \}\,=\,\{\alpha^{l}\}.\\
\mathrm{This\; yields}\; LL(\alpha)\,=\,L(\alpha). \tag{L1}
\end{gather*}
\textsf{Proof of U1:}
\begin{gather*}
\alpha^u = \{X^u \,;\,\alpha^{l} \,=\, X^{l} \, \&\, \alpha^u \,=\,X^{u} \}\,=\, \{\alpha^{u}\}.\\
U(\alpha)\,=\,[\alpha^{u}]\,=\, \{Y\,; \, Y^{l}\,=\, \alpha^{u}\,\&\, Y^{u}\,=\, \alpha^{uu} \}. \\
\mathrm{So}\; U(\alpha)^{u}\,=\, \{\alpha^{uu}\}.\\
UU(\alpha)\,=\, [U(\alpha)^{u}]\,=\, [\alpha^{uu}]\,=\,\{Y\,;\,Y^{l}\,=\,\alpha^{uu}\,\&\, Y^{u}\,=\,\alpha^{uuu}\}.\\
\mathrm{Since}\; \alpha\,\subseteq\,\alpha^{u}\,\subseteq\, \alpha^{uu}\subseteq\, \alpha^{uuu},\\
\mathrm{therefore}\; (U(\alpha)\,\cap \, UU(\alpha) \neq \emptyset \,\longrightarrow\, U(\alpha)\,=\, UU(\alpha). \tag{U1}
\end{gather*}
The other parts can be proved from the above considerations.
\qed
\end{proof}
\begin{theorem}
In the context of the above theorem, the following hold:
\begin{gather*}
\tag{CO1}{\alpha\odot \beta\,=\, \beta\odot \alpha)}\\
\tag{CO2}{\alpha\,\leq\, \alpha\odot \alpha }\\
\tag{CO3}{\alpha \leq \alpha\odot \top }\\
\tag{CO4}{\alpha\odot \alpha \,=\,\alpha\odot (\alpha\odot \alpha) \,=\, \alpha \odot \top}\\
\tag{AO1}{\alpha\oplus \beta\,=\, \beta\oplus \alpha)}\\
\tag{AO2}{\alpha \leq \alpha\oplus \beta}\\
\tag{AO3}{\alpha \leq \alpha\oplus \bot }\\
\tag{AO4}{(\alpha\oplus \alpha) \oplus \alpha \,=\, \alpha\oplus \alpha}\\
\tag{AC}{\mathrm{In\;general}, \alpha\oplus (\alpha\odot \beta)\neq \alpha .}
\end{gather*}
\end{theorem}
\begin{proof}
\begin{description}
\item [CO1]{The definition of $\odot$ does not depend on the order in which we take the arguments as set theoretic intersection and union are commutative. To be precise $\bigcup _{X\in [A],\, Y\in [B] } (X\cap Y)\,=\, \bigcup _{X\in [A],\, Y\in [B] } (Y\cap
X)$.}
\item [CO2]{$\bigcup _{X\in [A],\, Y\in [A] } (X\cap Y)\,=\, \bigcup_{X\in [A]} X $. But because $X^{l}\cup Y^{l}\subseteq (X\cup Y)^{l}$ in general, so we do not have equality.}
\item [CO3]{Follows from the last inequality.}
\item [CO4]{In $[\alpha\odot (\alpha\odot \alpha)]$, we cannot introduce any new elements that are not in $[\alpha\odot \alpha]$ as the inequality in [CO2] is due to the lower approximation and we have already included all possible subsets }
\item [AO1]{The definition of $\oplus$ does not depend on the order in which we take the arguments as set theoretic union is commutative.}
\item [AO2]{Even when $\beta=\alpha$, we can have the inequality for reasons mentioned earlier.}
\end{description}
Proof of [AO3, AO4, AC] are analogous or direct.
\qed
\end{proof}
The above result means that $\odot$ is an imperfect commonality relation. It is a proper commonality among a certain subset of elements of $H$.
\begin{theorem}
In the context of the above theorem, the following properties of $+,\, \times, \otimes$ are provable:
\begin{gather*}
\tag{+I}{\alpha + \alpha \,=\, \alpha},\\
\tag{+C}{\alpha + \beta \,=\, \beta + \alpha},\\
\tag{cI}{\alpha \times \alpha \,=\, \alpha},\\
\tag{cC}{\alpha \times \beta \,=\, \beta \times \alpha},\\
\tag{+Is}{\alpha \leq \beta \longrightarrow \alpha + \gamma \leq \beta + \gamma },\\
\tag{cIs}{\alpha \leq \beta \longrightarrow \alpha \times \gamma \leq \beta \times \gamma },\\
\tag{+In}{\alpha \leq \beta \longrightarrow \alpha \leq \alpha \times \beta \leq \beta},\\
\tag{R1}{\alpha + \beta \leq \alpha \oplus \beta},\\
\tag{Mix1}{\alpha \times \beta \leq (\alpha \times \beta) \oplus \alpha.}
\end{gather*}
\end{theorem}
\begin{proof}
Most of the proof is in Sec.\ref{appsem}, so we do not repeat them here.
\qed
\end{proof}
\begin{definition}
By a \emph{Concrete Pre-PRAX Algebraic System} (\textsf{CPPRAXA}), we will mean a system of the form
\[\mathfrak{H}\,=\,\left\langle H,\, \leq, L, U, \oplus , \odot , + , \times, \otimes, \bot , \top \right\rangle,\] with all of the operations being as defined in this chapter.
\end{definition}
Apparently we need to involve the algebraic properties of the rough objects of $l_{o},\, u_{o}$ to arrive at a representation theorem.
Further we can improve the operations defined to some extent by the related operations of the following chapter. Results concerning this will appear separately. Definable filters in general have reasonable properties.
\begin{definition}
Let $K$ be an arbitrary subset of a \textsf{CPPRAXA} $\mathfrak{H}$. Consider the following statements:
\begin{gather*}
\tag{F1} {(\forall x\in K)(\forall y\in \mathfrak{H})(x \leq y \Rightarrow
y\in K) .}\\
\tag{F2} {(\forall x, y\in K)\, x\oplus y, Lx \in K .}\\
\tag{F3} {(\forall a, b \in \mathfrak{H})(1\neq a\oplus b\in K\,\Rightarrow\, a\in K
\;\mathrm{or}\; b\in K) .}\\
\tag{F4} {(\forall a, b \in \mathfrak{H})(1\neq UB(a, b)\in K\,\Rightarrow\, a\in K
\;\mathrm{or}\; b\in K) .}\\
\tag{F5} {(\forall a, b\in K)\,LB(a, b)\cap K\neq \emptyset .}
\end{gather*}
\begin{mitemize}
\item {If $K$ satisfies \textbf{F1} then it will be said to be an \emph{order filter}. The
set of such filters on $\mathfrak{H}$ will be denoted by $\mathfrak{O}_{F}(\mathfrak{H})$.}
\item {If $K$ satisfies \textbf{F1, F2} then it will be said to be a \emph{filter}. The
set of such filters on $\mathfrak{H}$ will be denoted by $\mathcal{F}(\mathfrak{H})$.}
\item {If $K$ satisfies \textbf{F1, F2, F3} then it will be said to be a \emph{prime filter}.
The set of such filters on $\mathfrak{H}$ will be denoted by $\mathcal{F}_{P}(\mathfrak{H})$.}
\item {If $K$ satisfies \textbf{F1, F4} then it will be said to be a \emph{prime
order filter}. The set of such filters on $\mathfrak{H}$ will be denoted by $\mathfrak{O}_{PF}(\mathfrak{H})$.}
\item {If $K$ satisfies \textbf{F1, F5} then it will be said to be an \emph{strong order
filter}. The set of such filters on $\mathfrak{H}$ will be denoted by $\mathfrak{O}_{SF}(\mathfrak{H})$.}
\end{mitemize}
Dual concepts of ideals of different kinds can be defined.
\end{definition}
\begin{proposition}
Filters of different kinds have the following properties:
\begin{mitemize}
\item {Every set of filters of a kind is ordered by inclusion.}
\item {Every filter of a kind is contained in a maximal filter of the same kind.}
\item {$\mathfrak{O}_{SF}(\mathfrak{H})$ is an algebraic lattice, with its compact elements being
the finitely generated strong order filters in it.}
\end{mitemize}
\end{proposition}
\begin{definition}
For $F, P \in \mathcal{F} (\mathfrak{H})$, we can define the following operations:
\[F\wedge P\,\stackrel{\Delta}{=} F\cap P\]
\[F\vee P\,\stackrel{\Delta}{=} \left\langle F\cup P\right\rangle, \]
where $\left\langle F\cup P\right\rangle$ denotes the smallest filter containing $F\cup
P$.
\end{definition}
\begin{theorem}
$\left\langle \mathcal{F}(\mathfrak{H}),\, \vee , \, \wedge , \bot, \top \right\rangle$ is an
atomistic bounded lattice.
\end{theorem}
\chapter{Algebraic Semantics-2}\label{appsem}
We have seen that ordered pairs of the form $(A^{l},\, A^{u})$ do correspond to rough objects by definition. If we choose to ignore the representation and finer aspects of possible reasonable aggregation and commonality operations, then we still obtain an interesting order structure based fragment of semantic processes that is very useful in the approximation based semantics that we consider in subsequent chapters.
\begin{definition}
In a \textsf{PRAX} $S$, let \[\mathcal{R}(S)\,=\, \{(A^{l},\,A^{u})\, ; \, A\in \wp(S)\}.\]
Then we can define all of the following operations on $\mathcal{R}(S)$:
\begin{gather*}
(A^{l},\,A^{u})\,\vee\, (B^{l},\,B^{u})\,\stackrel{\Delta}{=}\, (A^{l}\cup B^{l},\,A^{u}\cup B^{u} ). \tag{Aggregation}\\
\mathrm{If}\; (A^{l}\cap B^{l},\,(A^{u}\cap B^{u}))\in \mathcal{R}(S)\;\mathrm{then}\;\\
(A^{l},\,A^{u})\,\wedge\, (B^{l},\,B^{u})\,\stackrel{\Delta}{=}\, (A^{l}\cap B^{l},\,(A^{u}\cap B^{u})). \tag{Commonality}\\
\mathrm{If}\; (A^{uc},\,A^{lc})\in \mathcal{R}(S)\;\mathrm{then}\;\\
\sim (A^{l},\,A^{u})\,\stackrel{\Delta}{=}\, (A^{uc},\,A^{lc}). \tag{Weak Complementation}\\
\bot \,\stackrel{\Delta}{=}\, (\emptyset ,\, \emptyset ). \;\; \top \,\stackrel{\Delta}{=}\, (S,\, S). \tag{Bottom, Top}\\
(A^{l},\,A^{u})\,\barwedge\, (B^{l},\,B^{u})\,\stackrel{\Delta}{=}\, ((A^{l}\cap B^{l})^{l},\,(A^{u}\cap B^{u})^{l}).\tag{Proper Commonality}
\end{gather*}
\end{definition}
\begin{definition}
In the context of the above definition, a partial algebra of the form $\mathfrak{R}(S)\,=\,\left\langle \mathcal{R}(S),\, \vee ,\, \wedge ,\, c ,\,\bot ,\,\top \right\rangle $ will be termed a \emph{proto-vague algebra} and $\mathfrak{R}_{f}(S)\,=\,\left\langle \mathcal{R}(S),\, \vee ,\, \wedge ,\,\barwedge.\, c ,\,\bot ,\,\top \right\rangle $ will be termed a \emph{full proto-vague algebra}.
More generally, if $L,\, U$ are arbitrary rough lower and upper approximation operators over the \textsf{PRAX}, and if we replace each occurrence of $l$ by $L$ and $u$ by $U$ in the above definition then we will term the resulting algebra of the above form a $LU$-\emph{proto-vague partial algebra}. Thus we will speak of $l_{o} u_{o}$-proto-vague algebras and such.
\end{definition}
\begin{theorem}
A full proto-vague partial algebra $\mathfrak{R}_{f}(S)$ satisfies all of the following:
\begin{enumerate}
\item {$\vee, \barwedge $ are total operations.}
\item {$\vee$ is a semi-lattice operation satisfying idempotency, commutativity and associativity.}
\item {$\wedge$ is a weak semi-lattice operation satisfying idempotency, weak strong commutativity and weak associativity. With $\vee$ it forms a weak distributive lattice.}
\item {$\sim$ is a weak strong idempotent partial operation; $\sim\sim\sim \alpha \,\stackrel{\omega^{*}}{=}\,\sim \alpha.$ }
\item {$\sim (\alpha \vee \beta)\,\stackrel{\omega}{=}\,\sim\alpha \wedge \sim\beta $ (Weak De Morgan condition) holds.}
\item {$\barwedge$ is an idempotent, commutative and associative operation that forms a lattice with $\vee$. }
\item {$\alpha \barwedge \bot \,=\, \alpha \wedge \bot \,=\, \bot$. $\alpha \vee \bot \,=\, \alpha$; $\alpha \barwedge \top \,=\, \alpha \wedge \top \,=\, \alpha$. $\alpha \vee \top \,=\, \top$.}
\item {$\sim (\alpha \wedge \beta)\,=\, (\sim \alpha \vee \sim \beta) \longrightarrow \sim (\alpha \barwedge \beta) \,=\,(\sim \alpha \vee \sim \beta).$ }
\item {$\alpha \vee (\beta \barwedge \gamma)\subseteq (\alpha \vee \beta) \barwedge (\alpha \vee \gamma) $, but distributivity fails.}
\end{enumerate}
\end{theorem}
\begin{proof}
Let $\alpha\,=\, (X^{l}, X^{u}),\; \beta\,=\, (Y^{l}, Y^{u})$ and $\gamma\,=\, (Z^{l}, Z^{u})$ for some $X,\, Y, \, Z\in \wp (S)$, then
\begin{enumerate}
\item {$\alpha \vee \beta\,=\, (X^{l}\cup Y^{l},X^{u}\cup Y^{u})$ belongs to $\mathfrak{R}(S)$ because the components are unions of successor neighborhoods and $X^{l}\cup Y^{l}\,\subseteq \,X^{u}\cup Y^{u}$. The proof for $\wedge$ is similar. }
\item {$\alpha \vee (\beta \vee \gamma)\,=\, (X^{l},X^{u}) \vee ((Y^{l}, Y^{u})\vee (Z^{l}, Z^{u}))\,=\, (X^{l},X^{u}) \vee (Y^{l}\cup Z^{l}, Y^{u}\cup Z^{u})\,=\,(X^{l}\cup Y^{l}\cup Z^{l}, X^{u}\cup Y^{u}\cup Z^{u})\,=\, (\alpha\vee\beta) \vee \gamma.$}
\item {We prove weak absorptivity and weak distributivity alone.
$(X^{l}\cap (X^{l}\cup Y^{l}))\,=\, X^{l}$ and $(X^{u}\cap (X^{u}\cup Y^{u}))\,=\, X^{l}$ hold in all situations. If $(X^{l}\cup (X^{l}\cap Y^{l}))$ is defined then it is equal to $X^{l}$ and if $(X^{u}\cup (X^{u}\cup Y^{u}))$ is defined, then it is equal to $X^{u}$. So \[\alpha \vee (\alpha \wedge \beta)\,\stackrel{\omega}{=}\, \alpha\,{=}\,\alpha \wedge (\alpha \vee \beta).\]
For distributivity ($\alpha \vee (\beta \wedge \gamma)\,\stackrel{\omega}{=}\,(\alpha\vee\beta)\wedge(\alpha\vee \gamma ) $ and $\alpha \wedge (\beta \vee \gamma)\,\stackrel{\omega}{=}\,(\alpha\wedge\beta)\vee(\alpha\wedge \gamma ) $) again it is a matter of definability working in coherence with set-theoretic distributivity.
}
\item {If $\sim \alpha$ is defined then $\sim \alpha\,=\,(X^{uc},X^{lc})$ and \[\sim\sim \alpha \,=\, \sim (X^{uc},X^{lc})\,=\, (X^{lcc},X^{ucc})\,=\,(X^{l},X^{u}),\] by definition. If $\sim \sim \alpha$ is defined, then $\sim \alpha$ is necessarily defined. So
\[\sim\sim\sim \alpha \,\stackrel{\omega^{*}}{=}\,\sim \alpha. \]}
\item {If $\sim (\alpha \vee \beta)$ and $\sim\alpha \wedge \sim\beta$ are defined then $\sim (\alpha \vee \beta)\,=\, \sim ((X^{l}\cup Y^{l}),(X^{u}\cup Y^{u}))\,=\, ((X^{uc}\cap Y^{uc}),(X^{lc}\cap Y^{lc}))\,\stackrel{\omega^{*}}{=}\,(X^{uc},X^{lc})\wedge (Y^{uc}, Y^{lc})\, =\,\sim\alpha \wedge \sim\beta $. So $\sim (\alpha \vee \beta)\,\stackrel{\omega^{*}}{=}\,\sim\alpha \wedge \sim\beta $.}
\item {$\alpha\barwedge \beta \,=\, \beta \barwedge \alpha \,\&\, \alpha \barwedge \alpha \,=\, \alpha$ are obvious.
$\alpha\barwedge(\beta \barwedge \gamma )\,=\,((X^{l}\cap (Y^{l}\cap Z^{l})^{l})^{l},\,(X^{u}\cap (Y^{u}\cap Z^{u})^{u})^{u} ) $
The components are basically the unions of common granules among the three. No granule in the final evaluation is eliminated by choice of order of operations. So $\alpha\barwedge(\beta \barwedge \gamma )\,=\, (\alpha\barwedge \beta) \barwedge \gamma $.
$\alpha \barwedge (\alpha \vee \beta) \,=\, ((X^{l}\cap (X^{l}\cup Y^{l}))^{l},\,(X^{u}\cap (X^{u}\cup Y^{u}))^{l} )\,=\, \alpha $.
Further, $\alpha \vee (\alpha \barwedge \beta) \,=\, ((X^{l}\cup (X^{l}\cap Y^{l})^{l}),\,(X^{u}\cup (X^{u}\cap Y^{u})^{l}) )\,=\,\alpha$. So $\vee , \barwedge$ are lattice operations.}
\item {
\begin{mitemize}
\item {Since $\bot = (\emptyset , \emptyset )$, $\alpha \barwedge \bot \,=\, \alpha \wedge \bot \,=\, \bot$ and $\alpha \vee \bot \,=\, \alpha$ follow directly.}
\item {Since $\top = (S, S)$, $\alpha \barwedge \top \,=\, \alpha \wedge \top \,=\, \alpha$ and $\alpha \vee \top \,=\, \top$ follow directly.}
\end{mitemize}}
\item {Follows from the previous proofs.}
\item {
\begin{mitemize}
\item {$\alpha \vee (\beta \barwedge \gamma) \,=\, ((X^{l}\cup (Y^{l}\cap Z^{l})^{l}),\,(X^{u}\cup (Y^{u}\cap Z^{u})^{l}) )$. If $a\in S$ and $[a]\subseteq X^{l}\cup (Y^{l}\cap Z^{l})^{l}$, and $[a]\subseteq (Y^{l}\cap Z^{l})^{l}$, then $[a]\subseteq Y^{l}$ and $[a]\subseteq Z^{l}$. So $[a]\subseteq X^{l}\cup Y^{l}$ and $[a]\subseteq X^{l}\cup Z^{l}$.}
\item {If $[a]\subseteq X^{l}\cup (Y^{l}\cap Z^{l})^{l}$ and if $[a]\,=\, P \cup Q$, with $P\subseteq X^{l}$, $Q\subseteq (Y^{l}\cap Z^{l})^{l}$ then $[a]\subseteq X^{l}\cup Y^{l}$ and $[a]\subseteq X^{l}\cup Z^{l}$. This proves $\alpha \vee (\beta \barwedge \gamma)\subseteq (\alpha \vee \beta) \barwedge (\alpha \vee \gamma) $.}
\item {If $[a]\subseteq ((X^{l}\cup Y^{l})\cap (X^{l}\cup Y^{l}))^{l}$ then $[a]\subseteq X^{l}\cup Y^{l}$ and $[a]\subseteq X^{l}\cup Z^{l}$. This means $[a]\,=\, P \cup Q$, with $P\subseteq X^{l}$, $Q\subseteq Y^{l}$ and $Q\subseteq Z^{l}$ and $Q$ is contained in union of some other granules. So $Q\subseteq Y^{l}\cap Z^{l}$, but we cannot ensure $Q\subseteq (Y^{l}\cap Z^{l})^{l}$ (required counterexamples are easy to construct). It follows that $((X^{l}\cup Y^{l})\cap (X^{l}\cup Y^{l}))^{l} \nsubseteq X^{l}\cup (Y^{l}\cap Z^{l})^{l}$.}
\end{mitemize}}
\end{enumerate}
\end{proof}
The following theorem provides us a condition for ensuring that $\sim \alpha$ is defined.
\begin{theorem}
If $X^{uu}\,=\,X^{u} $, then $\sim(X^{l},X^{u})\,=\, (X^{uc},X^{lc})$ but the converse is not necessarily true.
\end{theorem}
\begin{proof}
\begin{mitemize}
\item {$\sim(X^{l},X^{u})$ is defined if and only if $X^{uc}$ is a union of granules.}
\item {If $X^{uu}\,=\,X^{u} $ then $X^{uc}$ is a union of granules generated by \emph{some} of the elements in $X^{uc}$ , but the converse need not hold.}
\item {So we have the result.}
\end{mitemize}
\end{proof}
Let $W$ be any quasi-order relation that approximates $R$, and let the granules $[x]_w , \,[x]_{wi} $ and $l_w ,\, u_w$ be lower and upper approximations defined by analogy with the definitions of $l,\, u$. If $R\subset W$, then $(\forall x\in S)\,[x]\subseteq [x]_w$ and we have the following scenario ($A,\, B \in \wp(S)$. We write $A\parallel B$ for $A\nsubseteq B \,\&\,B\nsubseteq A$):
\begin{mitemize}
\item {If $A\subset B$ and $A^{u} = B^{u}$, then it is possible that $A^{u_w}\subset B^{u_w}$.}
\item {If $A\subset B$ and $A^{l} = B^{l}$, then it is possible that $A^{l_w}\subset B^{l_w}$.}
\item {If $A\subset B$ and $A^{u_w}= B^{u_w}$, then it is possible that $A^{u} \subset B^{u}$.}
\item {If $A\subset B$ and $A^{l_w}= B^{l_w}$, then it is possible that $A^{l} \subset B^{l}$.}
\item {If $A\parallel B$ and $A^{l}=B^{l}$, then it is possible that $A^{l_w} \parallel B^{l_w}$.}
\item {If $A\parallel B$ and $A^{l_w}=B^{l_w}$, then it is possible that $A^{l} \parallel B^{l}$.}
\item {If $A\parallel B$ and $A^{u}=B^{u}$, then it is possible that $A^{u_w} \parallel B^{u_w}$.}
\item {If $A\parallel B$ and $A^{u_w}=B^{u_w}$, then it is possible that $A^{u} \parallel B^{u}$.}
\item {If $A\subset B$, $A^{l} = B^{l}$ and $A^{u} = B^{u}$ , then it is possible that $A^{u_w}\subset B^{u_w}\,\&\,A^{l_w}\subset B^{l_w}$.}
\end{mitemize}
The above properties mean that meaningful correspondences between vague partial algebras and Nelson algebras may be quite complex.
Focusing on granular evolution alone, we can define
\begin{gather*}
(\forall x\in S) \, \varphi_{o}([x])\,=\, \bigcup_{z\in [x]} [z]_{w}.\\
(\forall A \in \wp (S))\, \varphi (A^{l})\,=\, \bigcup_{[x]\,\subseteq A^{l}} \varphi_{o}([x]).\\
(\forall A \in \wp (S))\, \varphi (A^{u})\,=\, \bigcup_{[x]\,\subseteq A^{u}} \varphi_{o}([x]).
\end{gather*}
$\varphi(A^{l}\cup B^{l})\,=\, \bigcup_{[x]\subseteq A^{l}\cup B^{l}}$.
If $[x]\subseteq A^{l}\cup B^{l}$
$\varphi$ can be naturally extended by components to a map $\tau$ as per \[\tau (A^{l},A^{u})\,=\, (\varphi(A^{l}),\, \varphi(A^{u})).\]
\begin{proposition}
If $R\subseteq R_{w}$ and $R_w$ is transitive, then
\begin{mitemize}
\item {If $z\in [x]$ and $x\in [z]$, then $\varphi ([z]) = \varphi([x])$.}
\item {If $z\in [x]$, then $\varphi ([z])\subseteq \varphi([x])$.}
\item {\[(\forall A \in \wp (S))\, \varphi(A^{l}) = \bigcup_{[x]\subseteq A^{l}}\varphi ([x]) = \bigcup_{[x]\subseteq A^{l}} [x]_{w} \]}
\end{mitemize}
\end{proposition}
\begin{proof}
\begin{mitemize}
\item {$z\in [x]$ yields $Rzx$. So if $Raz$, then $Rax$ and it is clear that $\varphi ([z]) \subseteq \varphi([x])$. $Rbx \& Rzx \& Rxz$ implies $R_{w}bz$ . }
\item {This is the first part of the above.}
\item {Follows from the above.}
\end{mitemize}
\end{proof}
\begin{definition}
We will use the following abbreviations for handling different types of subsets of $S$:
\begin{gather*}
\Gamma_{u}(S)\,=\, \{A^{u}; A\in \wp (S) \}. \tag{Uppers}\\
\Gamma_{uw}(S)\,=\, \{A^{u_{w}}; A\in \wp (S) \}. \tag{w-Uppers}\\
\Gamma (S)\,=\, \{B; \, (\exists A\in \wp (S))\, B=A^{l}\;\mathrm{or}\;B=A^{u}\}. \tag{lower definites}
\end{gather*}
Note that $\delta_{l}(S)$ is the same as $\Gamma (S)$ and similarly for $\delta_{lw}(S)$.
\end{definition}
$\tau$ has the following properties:
\begin{proposition}
If $R\subseteq R_{w}$ and $R_w$ is transitive, then
\begin{gather*}
\tau(\bot)\,=\, \bot_{w}.\\
\tau(\top )\,=\, \top_{w}.\\
(\forall \alpha, \beta \in \mathfrak{R}(S))\, \tau(\alpha\vee \beta)\,{=}\, \tau(\alpha)\vee \tau(\beta) .\\
(\forall \alpha, \beta \in \mathfrak{R}(S))\, \tau(\alpha\wedge \beta)\,\stackrel{\omega}{=}\, \tau(\alpha)\wedge \tau(\beta) .
\end{gather*}
\end{proposition}
\begin{proof}
\end{proof}
\begin{definition}
For each $\alpha\in \mathfrak{R}_{w}(S)$, the set of ordered pairs $\tau^{\dashv}(\alpha)$ will be termed as a \emph{co-rough object} of $S$, where \[\tau^{\dashv}(\alpha)\,=\, \{\beta \,; \, \beta\in \mathfrak{R}(S)\, \& \,\tau(\beta)\,=\, \alpha\}.\]
The collection of all co-rough objects will be denoted by $\mathfrak{CR}(S)$.
\end{definition}
This permits us to define a variety of closely related semantics of \textsf{PRAX} when $R\subseteq R_{w}$ and $R_w$ is transitive. These include:
\begin{itemize}
\item {The map $\tau : \mathfrak{R}_{f}(S)\, \longmapsto\, \mathfrak{R}_{w}(S)$. $\mathfrak{R}_{w}(S)$ being a Nelson algebra over an algebraic lattice.}
\item {$\mathfrak{R}_{f}(S)\,\cup\,\mathfrak{CR}(S)$ along with induced operations yields another semantics of \textsf{PRAX}.}
\item {$\mathfrak{R}(S) \,\cup\,\mathfrak{R}_{w}(S)$ enriched with algebraic and dependency operations described in \ref{dep}.}
\end{itemize}
\chapter{Approximate Relations}
If $R$ is a binary relation on a set $X$, then we let $R^{o} \, \stackrel{\partial}{=}\, R\,\cup \, \Delta_{X}$. The weak transitive closure of $R$ will be denoted by $R^{\#}$. If $R^{(i)}$ is the $i$-times composition $\stackrel{\underbrace{R\circ R \ldots \circ R}}{\textrm{i-times}}$, then $R^{\#}\,=\, \bigcup R^{(i)}$. $R$ is \emph{acyclic} if and only if $(\forall x) \, \neg R^{\#} xx$.
The relation $R^{\cdot}$ is defined by $R^{\cdot} ab$ if and only if $R a b \,\&\, \neg (R^{\#}a b \,\& \, R^{\#} b a)$.
\begin{definition}
If $R$ is a relation on a set $S$, then the relations $R^{\leftthreetimes},\, R^{cyc}$ and $R^{h} $ will be defined via
\begin{gather}
{R^{\lf} a b \mathrm{\; if \;and\; only\; if\;} [b]_{R^{o}}\subset [a]_{R^{o}} \,\&\, [a]_{i R^{o}}\subset [b]_{i R^{o}}}\\
{R^{cyc} a b \mathrm{\; if \;and\; only\; if\;} R^{\#} a b\, \& \,R^{\#} b a}\\
{R^{h} a b \mathrm{\; if \;and\; only\; if\;} R^{\lf} a b\, \& \,R^{\cdot} a b .}
\end{gather}
In case of PRAX, $R^{o}\, =\,R $, so the definition of $R^{\lf}$ would involve neighborhoods of the form $[a]$ and $[a]_{i}$ alone. $R^{\lf}\subset R$ and $R^{\lf}$ is a partial order.
\end{definition}
\begin{examp}
In our example \ref{agre}, $R^{\#}ab$ happens when $a$ is an ally of an ally of $b$. $R^{\lf}ab$ happens \textsf{iff} every ally of $b$ is an ally of $a$ and if $a$ is ally of $c$, then $b$ is an ally of $c$ - this can happen, for example, when $b$ is a Marxist feminist and $a$ is a socialist feminist. $R^{cyc}ab$ happens when \emph{$a$ is an ally of an ally of $b$ and $b$ is an ally of an ally of} $a$. $R^{\cdot}ab$ happens whenever $a$ is an ally of $b$, but $b$ is not an ally of anybody who is an ally of $a$.
\end{examp}
\begin{theorem}
$R^{h}\,=\, \emptyset .$
\end{theorem}
\begin{proof}
\begin{gather*}
{R^{h} a b \Leftrightarrow\, R^{\lf} a b\, \&\, R^{\cdot} a b}\\
{\Leftrightarrow \,\tau(R) a b\, \&\, (R\setminus \tau(R)) a b}\\
{\mathrm{But}\; \neg (\exists a ) (R\setminus \tau (R)) a a}.
\end{gather*}
{So $R^{h}\,=\, \emptyset$.}
\qed
\end{proof}
\begin{proposition}
All of the following hold in a PRAX $S$:
\begin{align}
{R^{\cdot}ab \,\leftrightarrow\,(R\setminus \tau(R)) ab }\\
{(\forall a, b) \neg (R^{\cdot} ab \,\&\, R^{\cdot} ba)}\\
{(\forall a, b, c)(R^{\cdot} a b \,\&\, R^{\cdot} b c \,\longrightarrow\, \neg R^{\cdot} a c).}
\end{align}
\end{proposition}
\begin{proof}
\begin{mitemize}
\item {$ R^{\cdot}ab \,\leftrightarrow\, R ab \, \&\, \neg (R^{\#}ab \, R^{\#} ba) $.}
\item {But $\neg (R^{\#}ab \, R^{\#} ba)$ is possible only when both $Rab$ and $Rba$ hold.}
\item {So $ R^{\cdot}ab \,\leftrightarrow\, R ab \, \&\, \neg (\tau (R) ab ) \,\leftrightarrow\, (R\setminus \tau (R)) ab$.}
\end{mitemize}
\qed
\end{proof}
\begin{theorem}
\begin{align}
{R^{\# \cdot}\,=\, R^{\#}\setminus \tau (R)}\\
{R^{\cdot \#}\,=\, (R\setminus \tau (R))^{\#}}\\
{(R\setminus \tau (R))^{\#} \,\subseteq \, R^{\#}\setminus \tau (R) .}
\end{align}
\end{theorem}
\begin{proof}
\begin{enumerate}
\item {\begin{align*}
R^{\# \cdot} a b \leftrightarrow & R^{\#} a b\,\&\, \neg (R^{\#\#} a b \,\& \, R^{\#\#} b a) \\
\leftrightarrow & R^{\#} a b \,\&\, \neg (R^{\#} a b \,\& \, R^{\#} b a) \\
\leftrightarrow & R^{\#} a b \,\&\, \neg \tau (R) a b \\
\leftrightarrow & (R^{\#}\setminus \tau (R)) a b.
\end{align*}}
\item {\begin{align*}
R^{\cdot \#} a b \leftrightarrow & (R^{\cdot})^{\#} a b \\
\leftrightarrow & (R\setminus \tau (R))^{\#} a b .
\end{align*}}
\item {Can be checked by a contradiction or a direct argument.}
\end{enumerate}
\qed
\end{proof}
We now look at possible properties that approximations of prototransitive relations may/should possess. If $<$ is a strict partial order on $S$ and $R$ is a relation, then consider the conditions :
\begin{gather}
\tag{PO1} (\forall a, b)(a < b\, \longrightarrow \, R^{\#} a b).\\
\tag{PO2} (\forall a, b)(a < b \, \longrightarrow \, \neg R^{\#} b a) .\\
\tag{PO3} (\forall a, b)(R^{\lf} a b \,\&\, R^{\cdot} a b \,\longrightarrow\, a < b.\\
\tag{PO4} \mathrm{If\;} a\equiv_{R} b, \mathrm{\;then\;} a\equiv_{<} b.\\
\tag{PO5} (\forall a, b)(a < b\, \longrightarrow \, R a b).
\end{gather}
As per \cite{RJ2011}, $<$ is said to be a \emph{partial order approximation} \textsf{POA} (resp. \emph{weak partial order approximation} \textsf{WPOA}) of $R$ if and only if \bbf{PO1, PO2, PO3, PO4} (resp. \bbf{PO1, PO3, PO4}) hold. A \textsf{POA} $<$ is \emph{inner approximation} \textsf{IPOA} of $R$ if and only if \bbf{PO5} holds. \bbf{PO4} has a role beyond that of approximation and depends on both successor and predecessor neighborhoods. $R^{h},\, R^{\cdot \lf}$ are \textsf{IPOA}, while $R^{\cdot \#},\, R^{\# \cdot}$ are \textsf{POA}s.
By a \emph{lean quasi order approximation} $<$ of $R$, we will mean a quasi order satisfying \bbf{PO1} and \bbf{PO2}.
The corresponding sets of such approximations of $R$ will be denoted by $POA(R),\, WPOA(R),\, IPOA(R), IWPOA(R) $ and $LQO(R)$
\begin{theorem}
For any $A,\, B\in LQO(R)$, we can define the operations $\& ,\vee , \top$:
\begin{gather*}
(\forall x, y) (A \& B) x y\; \mathrm{if\; and\; only\; if}\;(\forall x, y) A x y \,\&\, B x y. \\
(A \vee B) \,=\,(A \cup B)^{\#},\\
\top = R^{\#}.
\end{gather*}
\end{theorem}
\begin{proof}
\begin{mitemize}
\item {If $A a b$ then $R^{+} a b $ and if $B a b$ then $R^{+} a b $.}
\item {But if $(A\& B) a b$, then both $A a b$ and $B a b$.}
\item {So $R^{+} a b$.}
\end{mitemize}
Similarly it can be shown that $A\vee B \in LQO(R)$. It is always defined and contained within $R^{\#}$ as it is the transitive completion of $A\cup B$. $\top\,=\, R^{\#}$ as transitive closure is a closure operator.
\qed
\end{proof}
\begin{theorem}
In a \textsf{PRAX}, $R^{\cdot \#} \& R^{\# \cdot} x y \,\leftrightarrow \, (R\setminus \tau(R))^{\#} x y .$
\end{theorem}
\section{Granules of Derived Relations}
The behavior of approximations and rough objects corresponding to derived relations is investigated in this section.
\begin{definition}
The relation $R^{\#\cdot}$ will be termed the \emph{trans ortho-completion} of $R$. The following granules will be associated with each $x\in S$ :
\begin{gather}
[x]_{ot}\,=\, \{y\, ;\, R^{\#\cdot} y x \, \} \\
[x]_{ot}^{i}\,=\, \{y\, ;\, R^{\#\cdot} x y \, \} \\
[x]_{ot}^{o}\,=\, \{y\, ;\, R^{\#\cdot}y x \,\&\, R^{\#\cdot} x y \}. \end{gather}
Let the corresponding approximations be $l_{ot},\, u_{ot}$ and so on.
\end{definition}
\begin{theorem}
In a PRAX $S$, $(\forall x\in S)\, [x]_{ot}^{o}\,=\, \{x \}.$
\end{theorem}
\begin{proof}
$R^{\#\cdot} x y\,\&\,R^{\#\cdot} y x $ means that the pair $(x,\, y)$ is in the transitive completion of $R$ and not in $\tau(R)$.
So $y\in [x]_{ot}^{o}$ if and only if
\[(\exists a,\, b)\, R x a \,\&\, R a y \, \&\, (\neg R a x \vee \neg R y a )\,\& \,(R y b \,\&\, R b x)\, \&\, (\neg R b y \vee \neg R x b).\]
If we assume that $x\,\neq \, y$, then each of the possibilities leads to a contradiction as is shown below. In the context of the above statement:
\begin{flushleft}
\textbf{Case-1}
\end{flushleft}
\begin{mitemize}
\item {$R x a \,\&\,R a y \,\&\,\neg R a x \,\&\, R y a \,\&\, R y b \,\&\, R b x \,\&\, \neg R b y \,\&\, R x b $.}
\item {This yields $R^{\#} x a \,\&\, R^{\#} bb \,\&\, R^{\#} b a \,\&\, R^{\#} a b .$}
\item {So, $R^{\#} x b \,\&\, R^{\#} y a \,\&\, R^{\#} a x$ and we have contradicted our original assumption.}
\end{mitemize}
\begin{flushleft}
\textbf{Case-2}
\end{flushleft}
\begin{mitemize}
\item {$R x a \,\&\,R a y \,\&\, R a x \,\&\,\neg R y a \,\&\, R y b \,\&\, R b x \,\&\, R b y \,\&\,\neg R x b $.}
\item {This yields the contradiction $R^{\#} a b$. }
\end{mitemize}
\begin{flushleft}
\textbf{Case-3}
\end{flushleft}
\begin{mitemize}
\item {$R x a \,\&\,R a y \,\&\,\neg R a x \,\&\, R y a \,\&\, R y b \,\&\, R b x \,\&\, R b y \,\&\,\neg R x b.$}
\item {This yields $R^{\#} b a \,\&\, R^{\#} a b \,\&\,R^{\#} a a \,\&\, R^{\#} b b $ and $R^{\#} y y \& R^{\#} x y \& R^{\#} y x \& R y a \& R^{\#} x a$. }
\item {But such a $R^{\#}$ is not possible.}
\end{mitemize}
Somewhat similarly the other cases can be seen to lead to contradictions.
\qed
\end{proof}
By the \emph{symmetric center} of a relation $R$, we will mean the set $K_{R}\,=\, \bigcup e_{i} (\tau(R) \setminus \Delta_{S})$ - basically the union of elements in either component of $\tau (R)$ minus the diagonal relation on $S$.
\begin{proposition} $(\forall x)\, [x]\triangle [x]_{ot}\,\neq\, \emptyset$ as
\begin{align*}
x\,\notin\, K_{R}\, \longrightarrow\, [x]\subset [x]_{ot}\\
x\,\in\, K_{R}\, \longrightarrow\, [x]\nsubseteq [x]_{ot} \,\&\, \{x\}\subset [x]\cap [x]_{ot}.
\end{align*}
\end{proposition}
\begin{proof}
\begin{align*}
z\in [x]_{ot} \, & \leftrightarrow\, R^{\#\cdot} z x\\
\, & \leftrightarrow\, R^{\#} z x \,\&\, \neg \tau(R) z x\\
\, & \leftrightarrow\, (R z x \,\&\, \neg R x z ) \,\mathrm{or}\, (\neg R z x \,\& \, \neg R x z \,\&\, (R^{\#}\setminus R) z x).
\end{align*}
\qed
\end{proof}
$K_{R}$ can be used to partially categorize subsets of $S$ based on intersection.
\begin{proposition}
$(R\setminus \tau(R))^{\#}\cup \tau (R)$ is not necessarily a quasi order.
\end{proposition}
\begin{proof}
$(x, y)\in (R\setminus \tau(R))^{\#}\cup \tau (R) $ and $(x, y)\notin \tau(R)$ and $x\in K_{R}\,\&\, y \notin K_{R}$ and $\exists z\in K_{R}\, \&\, z\neq x \,\&\, Rzx$ do not disallow $Rz y$. So $(R\setminus \tau(R))^{\#}\cup \tau (R)$ is not necessarily a quasi-order. We leave the missing part to the reader.
\qed
\end{proof}
\begin{proposition}
$((R\setminus \tau(R))^{\#}\cup \tau (R))^{\#}\, =\, R^{\#}.$
\end{proposition}
\begin{proof}
Clearly $R\subseteq ((R\setminus \tau(R))^{\#}\cup \tau (R))^{\#}$ and it can be directly checked that if $a\in ((R\setminus \tau(R))^{\#}\cup \tau (R))^{\#} \setminus R$ then $a\in R^{\#}\setminus R$ and conversely.
\end{proof}
\chapter{Transitive Completion and Approximate Semantics}
The interaction of the rough approximations in a \textsf{PRAX} and the rough approximations in the transitive completion can be expected to follow some order. \emph{The definite or rough objects most closely related to the difference of lower approximations and those related to the difference of upper approximations can be expected to be related in a nice way}. We show that this \emph{nice way} is not really a \emph{rough way}. But the results proved remain relevant for the formulation of semantics that involves that of the transitive completion as in \cite{JPR,SJ}. A rough theoretical alternative is possible by simply starting from sets of the form $A^{*}\,=\,(A^{l}\setminus A^{l_{\#}})\cup (A^{u_{\#}} \setminus A^{u})$ and taking their lower ($l_{\#}$) and upper ($u_{\#}$) approximations - the resulting structure would be a partial algebra derived from a Nelson algebra over an algebraic lattice (\cite{AM3690}).
\begin{proposition}
For an arbitrary proto-transitive reflexive relation $R$ on a set $S$, (we use $\#$ subscripts for neighborhoods, approximation operators and rough equalities of the weak transitive completion) all of the following hold:
\begin{align*}
(\forall x \in S)\, [x]_{R}\,\subseteq [x]_{R^{\#}} \tag{Nbd} \\
(\forall A \subseteq S )\, A^{l}\,\subseteq A^{l_{\#}} \,\&\,A^{u}\,\subseteq A^{u_{\#}} \tag{App}\\
(\forall A \subseteq S )(\forall B \in [A]_{\approx})(\forall C \in [A]_{\approx_{\#}})\, B^{l}\,\subseteq C^{l_{\#}} \,\&\,B^{u}\,\subseteq C^{u_{\#}} \tag{REq}
\end{align*}
The reverse inclusions are false in general in the second assertion in a specific way.
Note that the last condition induces a more general partial order $\preceq$ over $\wp(\wp(S))$ via $A \preceq B$ if and only if $(\forall C\in A)(\forall E\in B)\,C^{l}\subseteq E^{l_{\#}}\, \&\,C^{u}\subseteq E^{u_{\#}}$.
\end{proposition}
\begin{proof}
The first of these is direct. For simplicity, we will denote the successor neighborhoods of $x$ by $[x]$ and $[x]_{\#}$ respectively. We look at the possibility tracking in the first part of the second assertion.
\begin{mitemize}
\item {If $z\in A^{l_{\#}}$ then $z\in A^{l}$ as $[x]_{\#}\subseteq A$ implies $[x]\subseteq A$.}
\item {If $z\in A^{l}$ then $(\exists x) \, z\in [x]\subseteq A^{l}$.}
\item {For this $x$, $z\in [x]_{\#}$, but it is possible that $[x]_{\#}\subseteq A$ or $[x]_{\#}\nsubseteq A$.}
\item {If $[x]_{\#}\nsubseteq A$, and $(\exists b\notin A )\,R_{\#} a x \,\&\, R a b \, \&\, R b x $ then we have a contradiction as $Rb x$ means $b\in [x]$.}
\item {If $[x]_{\#}\nsubseteq A$, and $(\exists b\in A )\,R_{\#} a x \,\&\, R a b \, \&\, R b x $ all we need is a $c\notin A \,\& R c b$ that is compatible with $R_{\#} c x$ and $A^{l}\nsubseteq A^{l_{\#}}$.}
\end{mitemize}
\qed
\end{proof}
\begin{definition}
By the \emph{l-scedastic approximation} $\hat{l}$ and the \emph{u-scedastic approximation} $\hat{u}$ of a subset $A\,\subseteq S$ we will mean the following approximations:
\[A^{\hat{l}}\,=\, (A^{l}\setminus A^{l_{\#}})^{l}, \;\; A^{\hat{u}}\,=\, ( A^{u_{\#}} \setminus A^{u})^{u_{\#}}. \]
The above cross difference approximation is the best possible from closeness to properties of rough approximations.
\end{definition}
\begin{theorem}
For an arbitrary subset $A\subseteq S$ of a \textsf{PRAX} $S$,the following statements and diagram of inclusion ($\rightarrow)$ hold:
\begin{mitemize}
\item {$A^{l_{\#}l} = A^{l_{\#}} = A^{ll_{\#}} = A^{l_{\#} l_{\#}}$}
\item {If $A^{u}\subset A^{u_{\#}}$ then $A^{u u_{\#}}\subseteq A^{u_{\#} u_{\#}}$. }
\end{mitemize}
\begin{center}
\begin{figure}[h]
\begin{tikzpicture}[node distance=2cm, auto]
\node (Alh) {$A^{l_{\#}}$};
\node (Al) [right of=Alh] {$A^{l}$};
\node (Alhu) [below of=Al] {$A^{l_{\#}u}$};
\node (Alu) [right of=Alhu] {$A^{lu}$};
\node (Alhuh) [below of=Alu] {$A^{l_{\#} u_{ \#}}$};
\node (A) [right of=Al] {$A$};
\node (Au) [right of=A] {$A^{u} $};
\node (Auh) [right of=Au] {$A_{u_{\#}}$};
\draw[->] (Alh) to node {}(Al);
\draw[->] (Al) to node {}(A);
\draw[->] (A) to node {}(Au);
\draw[->] (Au) to node {}(Auh);
\draw[->] (Alh) to node {}(Alhu);
\draw[->] (Alhu) to node {}(Alu);
\draw[->] (Alhu) to node {}(Alhuh);
\draw[->] (Alhuh) to node {}(Auh);
\draw[->] (Alu) to node {}(Au);
\end{tikzpicture}
\caption{Relation Between Approximate Approximations}
\end{figure}
\end{center}
\end{theorem}
\begin{proof}
It is clear that $A^{l}\subseteq A^{u}\subseteq A^{u_{\#}}$. So $A^{l}\,\nsubseteq \, A^{u_{\#}}\setminus A^{u}$.
\begin{align*}
x\in (A^{l} \setminus A^{l_{\#}})^{l} \, & \Rightarrow \, (\exists y)\, [y]_{\#}\nsubseteq A \,\&\, x \in [y] \subset A \,\&\, x\in [y]_{\#}\\
\, & \Rightarrow \, x\in A^{u_{\#}}\,\&\, x \in A^{u} \\
\, & \Rightarrow \, x\notin A^{u_{\#}}\setminus A^{u}.\\
\mathrm{But} \; [y]_{\#}\subset A^{u_{\#}} \, & \, (\exists z)\, z\in A^{u_{\#}}\,\&\, z\notin A^{u}\,\&\, z\in [y]_{\#}.\\
\mathrm{So} \; [y]_{\#}\,\subset\, (A^{u_{\#}}\setminus A^{u})^{u_{\#}}\, & \, \mathrm{and} \, \mathrm{it} \,\mathrm{is}\, \mathrm{possible} \,\mathrm{that}\, [y]_{\#}\,\nsubseteq\, (A^{u_{\#}}\setminus A^{u})^{u}.
\end{align*}
\qed
\end{proof}
\begin{theorem}
For an arbitrary subset $A\subseteq S$ of a \textsf{PRAX} $S$,
\begin{align*}
(A^{l}\setminus A^{l_{\#}})^{l}\,\nsubseteq \, (A^{u_{\#}}\setminus A^{u})^{u_{\#}}\,\longrightarrow \, A^{u_{\#}}\,=\, A^{u}.\\
A^{u_{\#}}\,\neq \, A^{u} \,\longrightarrow \, A^{l}\setminus A^{l_{\#}})^{l}\,\subseteq \, (A^{u_{\#}}\setminus A^{u})^{u_{\#}}.
\end{align*}
\end{theorem}
\begin{proof}
\begin{itemize}
\item {Let $ S\,=\, \{a, b, c, e, f\}$ and}
\item {let $R$ be the transitive completion satisfying $ R a b ,\, R b c , \, R e f$.}
\item {If $B\,=\, \{a,\, b\}$, $B^{\hat{l}}\,=\, B$,}
\item {but $B^{u_{\#}}\,=\,\{a,\, b,\, c\}\,=\, B^{u}$.}
\item {So $B^{\hat{u}} \,=\, \emptyset$.}
\item {The second part follows from the proof of the above proposition under the restriction in the premise.}
\end{itemize}
\qed
\end{proof}
\begin{theorem}
Key properties of the scedastic approximations follow:
\begin{enumerate}
\item {$(\forall B\in \wp (S))(B^{\hat{l}}\,=\, B\, \nrightarrow \, B^{\hat{u}}\,=\, B) $.}
\item {$(\forall B\in \wp (S))(B^{\hat{u}}\,=\, B\, \rightarrow \, B^{\hat{l}}\,=\, B) $.}
\item {$(\forall B\in \wp (S))\, B^{\hat{l} \hat{l}}\,=\, B^{\hat{l}}$.}
\item {$(\forall B\in \wp (S))\,B^{\hat{u} \hat{u}}\,\neq \, B^{\hat{u}}$. }
\item {It is possible that $(\exists B\in \wp (S)\,B^{\hat{u} \hat{u}}\,\subset \, B^{\hat{u}} )$.}
\end{enumerate}
\end{theorem}
\begin{proof}
\begin{enumerate}
\item {The counter example in the proof of the above theorem works for this statement.}
\item {$x\in B\,\leftrightarrow \, x\in (B^{u_{\#}}\setminus B^{u})^{u_{\#}}\,\leftrightarrow\, (\exists y\in B^{u_{\#}})(\exists z\in B^{u_{\#}}\setminus B^{u})\, x,\, z\in [y]_{\#} \,\&\, z\in B^{u_{\#}}\,\&\, z \notin B^{u} $. But this situation requires that elements of the form $z$ be related to $x$ and so we should have $B^{u_{\#}}\,=\, B^{u}$. }
\item {$B^{\hat{l} \hat{l}}\,=\, (B^{\hat{l} l}\setminus B^{\hat{l} l_{\#}})^{l}\,=\, ((B^{l}\setminus B^{l_{\#}})^{l}\setminus \emptyset )^{l} \,=\, B^{\hat{l}}$. The missing step is of proving $(B^{l}\setminus B^{l_{\#}})^{l l_{\#}}\,=\, \emptyset $.}
\item [4-5]{ We prove the last two assertions together. We provide a counterexample and also show the essential pattern of deviation.
\begin{itemize}
\item []{Let $S\,=\, \{a, b, c, e, f\}$ and $R$ be a reflexive relation s.t. $R a b ,\, R b c , \, R e f$.}
\item []{If $A\,=\, \{a,\, e\}$, then $A^{u_{\#}} \,=\, \{a,\, b,\,c,\, e\}$ and $A^{u}\,=\, \{a,\, b,\, e\}$.}
\item []{Therefore $A^{\hat{u}}\,=\, \{c \} \,\&\, A^{\hat{u}\hat{u}}\,=\, \emptyset \;\&\; A^{\hat{u} \hat{u}}\,\subset \, A^{\hat{u}}$.}
\item []{In general if $B$ is some subset, then $x\in B^{\hat{u}}\,=\, (A^{u_{\#}}\setminus A^{u} )^{u_{\#}}
\Rightarrow\, (\exists y \in A^{u_{\#}})(\exists z) \, y\in [z]_{\#}\, \&\, y\notin A^{u}\, \&\, y\notin A\, \&\,z\in A \,\&\, y\notin [z] \,\&\, y\in [x]_{\#}$.}
\end{itemize}}
\end{enumerate}
\qed
\end{proof}
An interesting problem can be given $A$ for which $A^{u_{\#}}\,\neq\, A^{u}$, when does there exist a $B$ such that
\[B^{l}\,=\,(A^{l}\setminus A^{l_{\#}})^{l}\,=\,A^{\hat{l}} \;\&\; B^{u}\,=\,(A^{u_{\#}}\setminus A^{u})^{u_{\#}}\,=\, A^{\hat{u}} ?\]
\chapter{Rough Dependence}\label{dep}
In this chapter, we introduce a concept of \emph{rough dependence} in general rough set theory. By the term \emph{rough dependence}, we intend to capture the relation between two objects (crisp or rough) that have some representable rough objects in common. There is no process for similarity with the concept \emph{mutual exclusivity} of probability theory and in rough sets we are actually handling evolution without regard to temporality. We would like to eventually analyze the extent to which ontology of not-necessarily-rough origin can be integrated in a seamless way. But in this monograph we will introduce basic concepts, compare them with probabilistic concepts and look at the semantic value of introduced functions and predicates.
Overall the following problems are basic and relevant for use in semantics:
\begin{mitemize}
\item {Which concepts of rough dependence provide for an adequate semantics of rough objects in the \textsf{PRAX} context? }
\item {More generally how does this relation vary over other \textsf{RST}s?}
\item {Characterize the connection between granularity and rough dependence?}
\end{mitemize}
By \emph{relation based RST} we mean rough theories originating from generalized approximation spaces of the form $U\,=\,\left\langle \underline{U},\,R\, \right\rangle$, with $\underline{U}$ being a set and $R$ being any binary relation on $\underline{U}$. If instead of a relation we start from a cover of the set, then we will refer to the rough theory as a \emph{cover based RST}.
\begin{definition}
The $\tau \nu$-\emph{infimal degree of dependence} $\beta_{i \tau \nu}$ of $A$ on $B$ will be defined as \[\beta_{i \tau \nu} (A,\, B)\,=\,\inf _{\nu (S) }\,\oplus \,\{C\,:\,C\in \tau(S) \, \&\,\pc C A \,\& \, \pc C B\}.\] Here the infimum means the largest $\nu(S)$ element contained in the aggregation.
The $\tau \nu$-\emph{supremal degree of dependence} $\beta_{s \tau \nu}$ of $A$ on $B$ will be defined as \[\beta_{s \tau \nu} (A,\, B)\,=\,\sup _{\nu (S) }\,\oplus \,\{C\,:\,C\in \tau(S) \,\&\, \pc C A \,\& \, \pc C B\}.\] Here the supremum means the least $\nu(S)$ element containing the sets.
The definition extends to \textsf{RYS} \cite{AM240} in a natural way.
\end{definition}
Note that all of the definitions do not use real-valued rough measures and the cardinality of sets in accord with one of the principles of avoiding contamination. The ideas of dependence are more closely related to certain semantic operations in classical \textsf{RST}. But these were never seen to be of much interest. The connections with probability theories has been part of a number of papers including \cite{ZPB,ZP5,PZ2002,SL2006,YY2008}, however neither dependence nor independence have received sufficient attention. This is the case with other papers on entropy. It should be noted that the idea of independence in statistics is seen in relation to probabilistic approaches, but dependence has largely not been given much importance in applications.
The positive region of a set $X$ is $X^{l}$, while the negative region is $X^{uc}$ -- this region is independent from $x$ in the sense of attributes being distinct, but not in the sense of derivability or inference by way of rules. When we talk of dependence or independence of a set relative another, then a basic question would be about possible balance between the two meta principles of independence in the rough theory and the relation to the granular concepts of independence.
\begin{definition}
Two elements $x,\, y$ in a RBRST or CBRST $S$ will be said to be \emph{PN-independent} $ I_{PN}(xy)$ if and only if \[x^{l}\,\subseteq \,y^{uc}\; \&\;y^{l}\,\subseteq \,x^{uc} . \]
Two elements $x,\, y$ in a RBRST or CBRST $S$ will be said to be \emph{PN-dependent} $\varsigma_{PN}(xy)$ if and only if \[x^{l}\,\nsubseteq \,y^{uc}\; \&\;y^{l}\,\nsubseteq \,x^{uc} . \]
\end{definition}
\begin{theorem}
Over the \textsf{RYS} corresponding to classical \textsf{RST}, we have the following properties of dependence degrees when $\tau(S)\,=\, \mathcal{G}(S)$ - the granulation of $S$ and $\nu(S)\,=\, \delta_{l}(S)$ - the set of lower definite elements. We omit the subscripts $\tau \nu$ and braces in $\beta_{i \tau \nu}(x, y)$ in the following:
\begin{enumerate}
\item {$\beta_{i}x y\,=\, x^{l}\, \cap\,y^ l \,=\,\beta_{s}x y $ (subscripts $i , \, s$ on $\beta$ can therefore be omitted).}
\item {$\beta x x \,=\, x^l $. }
\item {$\beta x y \,=\, \beta y x$. }
\item {$\beta (\beta x y) x \, =\, \beta x y$.}
\item {$\pc (\beta x y)(\beta x (y \oplus z))$.}
\item {$(\pc y^{l} z \longrightarrow \pc (\beta x y)(\beta x z))$.}
\item {$\beta x y \,=\, \beta x^{l} y^{l} \,=\, \beta x y^{l} $.}
\item {$\beta 0 x = 0 \,; \;\, \beta x 1 = x^{l}$.}
\item {$(\pc x y \longrightarrow \beta x y = x^{l})$.}
\end{enumerate}
\end{theorem}
We prove this in the next chapter.
\begin{theorem}
For classical \textsf{RST}, a semantics over the classical semantic domain can be formulated with no reference to lower and upper approximation operators using the operations $\cap,\, c ,\, \beta $ on the power-set of $S$, $S$ being an approximation space.
\end{theorem}
\begin{proof}
We have already shown that $l$ is representable in terms of $\beta$.
So the the result follows.
\end{proof}
\section{Dependence in PRAX}
When we set $\nu (S)\,= \delta_{l}(S) $ and $\tau(S)\,=\, \mathcal{G}(S) $ - the successor neighborhood granulation, then the situation in \textsf{PRAX} contexts is similar, but we cannot define $u$ from $l$ and complementation. However when we set $\nu (S)\,= \delta_{u}(S) $, then the situation is very different.
\begin{theorem}\label{vag}
Over the \textsf{RYS} corresponding to \textsf{PRAX} with $\pc \,=\, \subseteq $, $\oplus\,=\, \cup$ and $\odot \,=\, \cap$, we have the following properties of dependence degrees when $\tau(S)\,=\, \mathcal{S}$ - the granulation of $S$ and $\nu(S)\,=\, \delta_{l}(S)$ - the set of lower definite elements. In fact this holds in any reflexive \textsf{RBRST}. We omit the subscripts $\tau \nu$ and braces in $\beta_{i \tau \nu}(x, y)$ in the following:
\begin{enumerate}
\item {$\beta_{i}x y\,=\, x^{l}\, \cap\,y^ l \,=\,\beta_{s}x y $ (subscripts $i , \, s$ on $\beta$ can therefore be omitted).}
\item {$\beta x x \,=\, x^l $; $\beta x y \,=\, \beta y x$. }
\item {$(x\odot y\,=\, 0 \,\longrightarrow \, \beta_{i} x y \,=\, 0)$, but the converse is false. }
\item {$\beta (\beta x y) x \, =\, \beta x y$.}
\item {$\pc (\beta x y)(\beta x (y \oplus z))$.}
\item {$(\pc y^{l} z \longrightarrow \pc (\beta x y)(\beta x z))$.}
\item {$\beta x y \,=\, \beta x^{l} y^{l} \,=\, \beta x y^{l} $.}
\item {$\beta 0 x = 0 \,; \;\, \beta x 1 = x^{l}$.}
\item {$(\pc x y \longrightarrow \beta x y = x^{l})$.}
\end{enumerate}
\end{theorem}
\begin{proof}
\begin{enumerate}
\item {$\beta_{i}x y$ is the union of the collection of successor neighborhoods generated by elements $x$ and $y$ that are included in both of them. So $\beta_{i}x y \,=\, x^{l}\, \cap\,y^ l \,=\,\beta_{s}x y $.}
\item {$\beta x x \,=\, x^l $; $\beta x y \,=\, \beta y x$. is obvious}
\item {If $(x\odot y\,=\, 0$, then $x$ and $y$ have no elements in common and cannot have common successor neighborhoods. If $\beta_{i} x y \,=\, 0$, then $x,\,y$ have no common successor neighborhoods, but can still have common elements. So the statement follows.}
\item {$\beta x y \subseteq x^{l}\,\subseteq x$ by the first statement. So $\beta (\beta x y) x \, =\, \beta x y$.}
\item {$\pc (\beta x y)(\beta x (y \oplus z))$ follows by monotonicity.}
\item {If $\pc y^{l} z$ is the same thing as $y^{l}\subseteq z$. $\beta x y\, =\, x^{l}\cap y^{l}$ and $\beta x z \,=\, x^{l}\cap z^{l}$ by the first statement. So we have $(\pc y^{l} z \longrightarrow \pc (\beta x y)(\beta x z))$.}
\item {$\beta x y \,=\, \beta x^{l} y^{l} \,=\, \beta x y^{l} $ holds because $l$ is an idempotent operation in a \textsf{PRAX} \cite{AM270}.}
\item {Rest of the statements are obvious.}
\end{enumerate}
\qed
\end{proof}
Even though the properties are similar for reflexive \textsf{RBRST} when $\nu (S)\,= \delta_{l}(S) $ and $\tau(S)\,=\, \mathcal{G}(S) $, there are key differences that can be characterized in terms of special sets.
\begin{itemize}
\item {$\beta x y = z $ if and only if $(\forall a\in z)(\exists b \in z) \, a\in [z]\,\subseteq x\cap y$.}
\item {So we can select a minimal $K_{z}\,\subseteq z$ satisfying $(\forall a\in z )(\exists b\in K_{z})\,a\in \,[b]\,\subseteq x $ and $(\forall e\in K_{z})\,[e]\subseteq x\cap y $. Minimality being with respect to the inclusion order.}
\item {Let $\mathcal{P}_{z}$ be the collection of all such $K_{z}$ and let $\mathcal{B}_{z}$ be the subcollection of $\mathcal{P}_{z}$ satisfying the condition: if $K\,\in\,\mathcal{B}_{z} $ then $(\forall a\in K)(\forall b\in [a])(\exists J\in \mathcal{B}_{z})\, b\in J$. $\mathcal{P}_{z}$ will be called the local basis and $\mathcal{B}_{z}$, the local super basis of $z$. }
\end{itemize}
\begin{proposition}
For classical \textsf{RST} $(\forall z)\,\mathcal{B}_{z}\,=\, \mathcal{P}_{z}$ and conversely.
\end{proposition}
\begin{theorem}
In the context of \ref{vag}, if we set $\nu (S)\,= \delta_{u}(S) $ and $\tau(S)$ is as before, then we have (by $\beta xy $, we mean $\beta_{i} xy$)
\begin{enumerate}
\item {$\pc (\beta x y) (\beta_{i \delta_{l}(S)} x y) $, }
\item {$\pc(\beta x x) ( x^l) $; $\beta x y \,=\, \beta y x$. }
\item {$(x\odot y\,=\, 0 \,\longrightarrow \, \beta_{i} x y \,=\, 0)$, but the converse is false. }
\item {$\beta (\beta x y) x \, =\, \beta x y$.}
\item {$\pc (\beta x y)(\beta x (y \oplus z))$.}
\item {$(\pc y^{l} z \longrightarrow \pc (\beta x y)(\beta x z))$.}
\item {$\beta x y \,=\, \beta x^{l} y^{l} \,;\, \pc (\beta x y^{l})(\beta x^{u} y^{u})$.}
\item {$\beta 0 x = 0 \,; \;\, \pc (\beta x 1) (x^{l})$.}
\item {$(\pc x y \longrightarrow \pc (\beta z x)(\beta z y))$}
\item {$(\beta x y)^l\, =\,\beta xy $.}
\end{enumerate}
\end{theorem}
\begin{proof}
\begin{enumerate}
\item {By definition $\beta_{i \tau \nu} (A,\, B)\,=\,\inf _{\nu (S) }\,\oplus \,\{C\,:\,C\in \tau(S) \, \&\,\pc C A \,\& \, \pc C B\}$, so $\beta x y$ is the greatest upper definite set contained in the union of common successor neighborhoods included in $x$ and $y$. So it is necessarily a subset of $x^{l}\cap y^{l}$. In a \textsf{PRAX}, $u$ is not idempotent and in general $x^u \,\subseteq\, x^{uu}$ (\cite{AM270}). So $\pc (\beta x y) (\beta_{i \delta_{l}(S)} x y) $. }
\item {The statements $\pc(\beta x x) ( x^l) $ and $\beta x y \,=\, \beta y x$ follow from the above.}
\item {The proof is similar to that of third statement of \ref{vag}. }
\item {In constructing $\beta (\beta x y) x$ from $\beta x y$, we are not searching for upper definite subsets strictly contained in the latter. So the property follows.}
\item {$\pc (\beta x y)(\beta x (y \oplus z))$ follows by monotonicity.}
\item {Obvious from previous statements.}
\item {Note that $\beta x^{u} y^{u}$ is a subset of $x^{u}\cap y^{u}$ and in general contains $\beta x y$.}
\item {Is a special case of the first statement. $0$ is the empty set and $1$ is the top. }
\item {Follows by monotonicity.}
\item {Upper definite subsets are necessarily lower definite, so $(\beta x y)^l\, =\,\beta xy $.}
\end{enumerate}
\qed
\end{proof}
The main properties of PN-dependence is as below:
\begin{theorem}
In the context of \ref{vag}, all of the following hold (we drop the subscript 'PN' in $\varsigma_{PN}$):
\begin{enumerate}
\item {$\varsigma xx $. }
\item {$(\varsigma x y \,\leftrightarrow \varsigma y x)$.}
\item {In general, $\varsigma x y \,\&\, \varsigma z y $ does not imply $\varsigma x z$. But $\neg \varsigma x z$ is more likely if we assume a bit of frequentism.}
\item {In general, $\varsigma x y \,\nrightarrow \, \varsigma x^{u} y^{u} $ and $\varsigma x^{u} y^{u} \,\nrightarrow \, \varsigma x y $.}
\item {$(x\cdot y =0 \,\longrightarrow\, \neg \varsigma x y) $.}
\item {$(\pc x y \,\longrightarrow\, \varsigma x y) $.}
\end{enumerate}
\end{theorem}
\begin{theorem}
In the context of \ref{vag}, if $\beta x y \neq 0$ then $\varsigma x y$, but the converse need not hold. In classical \textsf{RST}, the converse holds as well.
\end{theorem}
\begin{proof}
If $\beta x y \neq 0$, then it follows that $x^{l}\,\cap\, y^{l}\neq \emptyset$ under the assumptions.
If we assume $x^l \subseteq y^{uc} \,\vee \, y^{l}\subseteq x^{uc}$, then in each of the three cases we have a contradiction.
So the first part of the result follows.
In the classical case, if $x^{l}\subseteq y^{uc}$ is not empty, then it should be a union of successor neighborhoods and similarly for
$y^{l}\subseteq x^{uc}$. These two parts should necessarily be common to $x^{l}$ and $y^{l}$. So the converse holds for classical \textsf{RST}. The proof does not work for \textsf{PRAX} and we know why it does not succeed.
\qed
\end{proof}
\chapter{Comparison with Dependence in Probabilistic Theories}
Probability measures may not exist in the first place over any given collection of sets, so even \textsf{CBRST} is necessarily more general and the idea of mutual exclusivity is not the correct concept corresponding to rough dependence. The basic idea of probabilistic dependence is oriented because occurrence of an event can be favorable or unfavorable for another event. In standard versions of rough set theory this has no corresponding concept. The concept of dependence in probability is rarely considered in the literature. The version in \cite{BD2010} uses a not-so intuitive valuation but is nevertheless useful. We abstract the subjective aspect of the valuation for comparison.
Among the different understandings of probabilistic causation, frequentism (\cite{AH2009}) and the tendency to omit necessary conditions are particularly problematic in various soft computing situations. In avoiding real-valued rough measures, we are committing to avoid the excesses of frequentism in rough sets.
If $(X,\,\mathcal{S} ,\, p)$ is a probability space with $X$ being a set, $\mathcal{S}$ being a $\sigma$-algebra over $X$ and $p$ being a probability function (we can use a collection of probability functions and handle more complex notions of dependence in 'probability structures', but these add little to the comparison), then the most natural dependence function $\delta :\, {\mathcal{S}}^{2}\,\longmapsto \, \Re $ is defined by
\[\partial (x, \, y)\,=\, p(x\cap y)\,-\, p(x)\,\cdot \, p(y)\]
This function satisfies a number of properties that can be used to characterize dependence. In the subjective probability domain where $p$ takes value in a bounded partially ordered partial semi-ring or your favorite partially ordered algebra, we will need to replace $\delta$ with a pair of predicates. So orientation of dependence seems to be fundamental in general forms of probability theory as well.
Two events $x,\, y \in X $ are \emph{mutually exclusive} if and only if $x\cap y\,\neq\, \emptyset$. This concept can be extended to countable sets of events in a natural way. Also it is worthwhile to modify the concept of mutual exclusivity as in following definition:
\begin{definition}
Two events $x,\, y$ will be said to be \emph{weakly mutually exclusive} (\textsf{WME}) if and only if \[x\,\cap\,y \,\neq\, z \, \& \,p(z)\,=\, 0 .\]
\end{definition}
Most results of probability theory involving mutual exclusivity continue to hold with the weaker assumption of \textsf{WME} and importantly is a better (though artificial) concept for comparison with the situation for rough sets.
\begin{definition}
In the above context, let
\begin{itemize}
\item {$\pi x y$ if and only if $p(x) \cdot p(y)\, <\, p(x\cap y) $}
\item {$\sigma x y $ if and only if $p(x\cap y)\,<\, p(x) \cdot p(y) $}
\end{itemize}
\end{definition}
\begin{proposition}
All of the following hold in a probability space:
\begin{itemize}
\item {$\pi x y^{c} \,\leftrightarrow \, \sigma y x $}
\item {$\pi x y \,\leftrightarrow \, \pi y x$}
\item {$(x\cap y \neq \emptyset \,\longrightarrow\, (\pi x a\,\&\, \pi y a \,\longrightarrow\, \pi (x\cup y) a ))$}
\item {$(x\cap y \neq \emptyset \,\longrightarrow\, (\sigma x a\,\&\, \sigma y a \,\longrightarrow\, \sigma (x\cup y) a ))$}
\item {$(\emptyset\,\neq\,x\subseteq y \, \longrightarrow\,\pi x y )$}
\item {$(x \cap y\,=\, \emptyset \,\longrightarrow\, \sigma x y)$}
\end{itemize}
\end{proposition}
Instead of using the the function $\partial(x, y)$, we can use the relations $\pi, \, \sigma$,as the former lacks a comparable contamination-free counterpart in rough set theory and also has peculiar properties like $\partial(x, x) \in [0,1/4]$.
\begin{proposition}
In the probability space above $0 \leq \partial(x, x)\leq 0.25$, $-0.25 \leq \partial (x,x^{c}) \leq 0$ and
$x,\,y$ are independent implies $\partial(x, y) = 0$, but not conversely.
\end{proposition}
\begin{proof}
The proof of the inequalities follow by a simple application of real analysis.
\qed
\end{proof}
So it follows that the interpretation of the function $\partial (x,y)$ as in \cite{BD2010} is actually incomplete. It combines certainty of the event with dependence.
Even though we can speak of positive, negative and neutral regions corresponding to an arbitrary subset $A$ of a \textsf{RBRST} or \textsf{CBRST} $S$, natural ideas of dependence do not correspond to the scenario in probability space. In fact,
\begin{theorem}
Predicates having properties identical with those of $\pi$ and $\sigma$ cannot be defined in the context of \ref{vag}.
\end{theorem}
Proof of this and more general results will appear separately.
\chapter{Dependency Semantics of PRAX}
We develop dependency based semantics in at least two ways. The \emph{internalization based semantics} is essentially about adjoining predicates to the Nelson algebra corresponding to $\mathfrak{R}_{w}(S)$. The \emph{cumulation based semantics} is essentially about cumulating both the semantics of $\mathfrak{R}(S)$, adjusting operations and adjoining predicates. We use broader dependency based predicates in this case, but the value of the method is in fusion of the methodologies.
The central blocks of development of the cumulation based dependency semantics are the following:
\begin{mitemize}
\item {Take $\mathfrak{R}(S) \,\cup\,\mathfrak{R}_{w}(S)$ as the universal set of the intended partial/total algebraic system. }
\item {Use a one point completion of $\tau$ to distinguish between elements of $\mathfrak{R}_{w}(S)\,\setminus\, \mathfrak{R}(S)$ and those in $\mathfrak{R}(S)$.}
\item {Extend the idea of operational dependency to pairs of sets.}
\item {Extend operations of aggregation, commonality and dual suitably.}
\item {Interpretation and meaning of semantic dependence?}
\end{mitemize}
The first step is obvious, but involves elimination of other potential sets arising from the properties of the map $\tau$.
\section*{One Point Completion}
Because we have $R\subseteq R_{w}$ and $R_{w}$ is transitive, so
\begin{proposition}
\[\alpha\in \mathfrak{R}(S)\cap \mathfrak{R}_{w}(S)\;\mathrm{ if\; and\; only\; if\;} \tau (\alpha) \,=\, \alpha.\]
\end{proposition}
We adjoin an element $0$ to $\mathfrak{R}(S) \,\cup\,\mathfrak{R}_{w}(S)$ to form $\mathfrak{R}^{*}(S)$ and extend $\tau$ (interpreted as a partial operation) to $\overline{\tau}$ as follows:
\[\overline{\tau} (\alpha)\,=\,\left\lbrace \begin{array}{ll}
\tau(\alpha) & \mathrm{if}\; \alpha\in \mathfrak{R}(S), \\
0 & \mathrm{if} \; \alpha\notin \mathfrak{R}(S).
\end{array}\right.
\]
Note that this operation suffices to distinguish between elements common to $\mathfrak{R}(S)$ and $\mathfrak{R}_{w}(S)$, and those exclusively in $\mathfrak{R}(S)$ and not in $\mathfrak{R}_{w}(S)$.
\section*{Dependency on Pairs}
We have the option of considering all dependency relative the Nelson algebra or $\mathfrak{R}(S)$. First we consider everything relative the former -so that we may be able to avoid the references to the latter.
\begin{definition}
By the \emph{paired infimal degree of dependence} $\beta^{+}_{i \tau_1 \tau_2 \nu_1 \nu_2}$ of $\alpha$ on $\beta$ will be defined as \[(\beta_{i \tau_1 \nu_1} (e_{1}\alpha,\, e_{1}\beta), \,\beta_{i \tau_2 \nu_2} (e_{2}\alpha,\, e_{2}\beta)) .\] Here the infimums involved are the largest $\nu_1(S)$ and $\nu_{2} (S)$ elements contained in the aggregation and the $e_{j}\alpha$ is the $j$-th component of $\alpha$.
\end{definition}
We will however be interested in the following well defined specialization with $\tau_{1}(S)\,=\,\tau_{2} (S)\, =\,\mathcal{G}_{w}(S)$, $\nu_{1}\,=\,\delta_{lw}(S) $ and $\nu_{2}\,=\,\Gamma_{uw}(S)$ in all that follows. When we need to specialize the dependencies between a element in $\mathfrak{R}(S)$ and its image in $\mathfrak{R}_{w}(S)$, we can define:
\begin{definition}
Under the above assumptions, by the \emph{relative semantic dependence} $\varrho (\alpha)$ of $\alpha\in \mathfrak{R}(S)$, we will mean
\[\varrho (\alpha) \,=\, \beta^{+}_{i}(\alpha, \tau(\alpha)).\]
\end{definition}
The idea of relative semantic dependence refers to elements in $\mathfrak{R}(S)$ and it can be reinterpreted as a relation on $\mathfrak{R}_{w}(S)$.
\section*{Internalization Based Semantics}
\begin{definition}
A relation $\Upsilon$ on $\mathfrak{R}_{w}(S)$ will be said to be a \emph{relsem-relation} if and only if
\[\Upsilon\tau(\alpha)\gamma \,\leftrightarrow \, (\exists \beta \in \tau^{\dashv}\tau(\alpha))\,\gamma\,=\,\varrho(\beta). \]
Note that, $\tau(\alpha)\,=\,\tau(\beta)$ by definition of $\tau^{\dashv}$.
\end{definition}
Through the above definitions we have arrived at the following internalized approximate definition:
\begin{definition}
By an \emph{Approximate Proto Vague Semantics} of a \textsf{PRAX} ${S}$ we will mean an algebraic system of the form \[\mathfrak{P}(S)\,=\,\left\langle \mathfrak{R}_{w}(S),\, \Upsilon \vee, \wedge, c, \bot, \top \right\rangle ,\] with $\left\langle \mathfrak{R}_{w}(S),\, \vee_{w}, \wedge_{w}, c, \bot, \top \right\rangle $ being a Nelson algebra over an algebraic lattice and $\Upsilon$ being as above.
\end{definition}
\begin{theorem}
$\Upsilon$ has the following properties:
\begin{align*}
\alpha\,=\,\tau(\alpha)\,\longrightarrow\, \Upsilon \alpha \alpha .& \\
\Upsilon \alpha \gamma \, \longrightarrow\, \gamma \wedge_{w} \alpha\,=\, \gamma .& \\
\Upsilon \alpha \gamma \,\&\, \Upsilon \gamma \alpha \,\longrightarrow\, \alpha\,=\, \gamma .& \\
\Upsilon \bot \bot \,\&\, \Upsilon \top \top .& \\
\Upsilon \alpha \gamma \,\&\,\Upsilon \beta \gamma \longrightarrow\,\Upsilon (\alpha\vee_{w} \beta) \gamma . &
\end{align*}
\end{theorem}
\begin{proof}
\begin{mitemize}
\item {If $\alpha\,=\, \tau(\alpha)$, then $\alpha\,=\,\varrho (\alpha)\,=\, \beta^{+}_{i}(\alpha, \tau(\alpha))$. So $\Upsilon \alpha \alpha$.}
\item {If $\Upsilon \alpha \gamma$, then it follows from the definition of $\beta^{+}_{i}$, that the components of gamma are respectively included in those of $\alpha$. So $\gamma \wedge \alpha\,=\, \gamma $.}
\item {Follows from the previous. }
\item {Proof is easy.}
\item {From the premise we have $(\exists \mu \in \tau^{\dashv}\tau(\alpha))\,\gamma\,=\,\varrho(\mu)$ and $(\exists \nu \in \tau^{\dashv}\tau(\beta))\,\gamma\,=\,\varrho(\nu)$. This yields $(\exists \lambda \in \tau^{\dashv}\tau(\alpha\vee_{w}\beta))\,\gamma\,=\,\varrho(\lambda)$ as can be checked from the components.}
\end{mitemize}
\qed
\end{proof}
$\Upsilon_{\tau(\alpha)}\,=\,\{\gamma \,; \,\Upsilon\tau(\alpha)\gamma \}$ is the approximate reflection of the set of $\tau$-equivalent elements in $\mathfrak{R}(S)$ identified by their dependence degree. In the approximate semantics we do not completely lose track of aggregation and commonality as the above theorem shows.
\begin{definition}
By the $\varrho /\sigma$-\emph{semantic dependences} $\varrho (\alpha)$, $\sigma(\alpha)$ of $\alpha\in \mathfrak{R}(S)$, we mean \[\varrho (\alpha) \,=\, \beta^{+}_{i}(\alpha, \tau(\alpha))\;\mathrm{and}\]
\[\sigma (\alpha) = \beta^{+}_{i}(\alpha, ((\varphi (e_{1}\alpha) \setminus e_{1}\alpha)^{l},(\varphi (e_{2}\alpha) \setminus e_{1}\alpha)^{u}))\] respectively. Such relations are optional in the internalization process.
\end{definition}
For a falls-down semantics, the natural candidates include the ones corresponding to largest equivalence or the largest semi-transitive contained in $R$. We reserve the latter for a separate paper. For the former, the general technique (using $\sigma(\alpha)$) extends to \textsf{PRAX} as follows:
\begin{definition}
\begin{mitemize}
\item {Define a map $\int$ from set of neighborhoods to $l$-definite elements via \[\int ([x]_o)\,=\,\cup_{y\in[x]_o} [y] \] and extend it to images of $l_o, u_o$ via, \[\oint (A^{l_o})= \cup_{[y]_{o}\subseteq A^{l_o}} \int ([y]_o).\]}
\item {Extend this to a map $\ltimes :\mathfrak{R}_{o}(S)\mapsto \mathfrak{R}(S)$ via \[\ltimes(\alpha) = (\oint (e_{1}\alpha),\oint (e_{2}\alpha)).\]}
\item {Define a predicate $\Pi$ on $\mathfrak{R}_{o}(S)$ as per \[\Pi \alpha \nu\,\leftrightarrow\,(\exists \gamma\in \ltimes^{\dashv}\ltimes(\alpha))\, \beta_{i}^{+}(\alpha, \gamma) = \nu .\] Let $\Pi_{\alpha} =\{\nu \,; \Pi \alpha \nu\}$.}
\item {By a \emph{Direct Falls Down} semantics of \textsf{PRAX}, we will mean an algebraic system of the form
\[\mathfrak{I}(S)\,=\,\left\langle \mathfrak{R}_{o}(S),\, \Pi , \vee, \wedge, c, \bot, \top \right\rangle,\] with $\left\langle \mathfrak{R}_{o}(S),\, \vee_{o}, \wedge_{o}, \rightarrow , c, \bot, \top \right\rangle $ being a semi-simple Nelson algebra \cite{PPM2}.}
\item {The falls down semantics determines a cover $\mathfrak{I}^{*}(S) = \{\Pi_{\alpha}\,;\, \alpha\in \mathfrak{R}_{o}(S) \}$}
\end{mitemize}
\end{definition}
\begin{theorem}
In the above context, all of the following hold:
\begin{mitemize}
\item {$\Pi \alpha \alpha$.}
\item {$(\Pi \alpha \mu\, \& \,\Pi \mu \alpha \longrightarrow \alpha = \mu)$. }
\item {$(\Pi \alpha \gamma \, \longrightarrow \, \gamma \subseteq \alpha)$. The converse is false.}
\item {$\alpha\neq \bot \,\&\, \Pi \alpha \gamma \,\&\, \Pi \alpha \mu \,\longrightarrow\, \beta_{i}^{+}(\gamma, \mu) \neq \bot$.}
\item {$\mu \in \Pi_{\alpha} \,\&\, \mu\subseteq \nu\,\subseteq \alpha\,\longrightarrow\, \nu\in \Pi_{\alpha}$.}
\end{mitemize}
\end{theorem}
\section*{Cumulation Based Semantics}
The idea of cumulation is correctly a way of enhancing our original semantics based on proto-vagueness algebras with the Nelson algebraic semantics and the operational dependence. We define it for a central problem relating to the underlying semantic domains and
\begin{definition}
By a \emph{cumulative proto-vague algebra} we will mean a partial algebra of the form
\[\mathfrak{C}(S)\,=\, \left\langle \mathfrak{R}^{*}(S),\overline{tau},\, \oplus, \odot, \otimes, \dagger , \bot , \top \right \rangle. \]
\end{definition}
\begin{flushleft}
\textbf{Problem:}
\end{flushleft}
When can the cumulation based semantics be deduced from (that is the extra operations can be defined from the original ones) within a full proto-vagueness algebra?
\chapter{Geometry of Granular Knowledge Interpretation}
In this chapter we provide a brief overview of knowledge interpretation in the \textsf{PRAX} contexts in the light of the results on representation of rough objects. For details of the knowledge interpretation of rough sets, the reader is referred to \cite{AM909,ZPB,CHP}. By an extension of those considerations any proto-transitive relation corresponds to knowledge. Here we will be concerned with representation of knowledge and that in turn depends on our choice of semantic domain - the most natural is the one corresponding to the rough objects. But we know that representation is involved.
\emph{Any knowledge, however involved, may be seen as a collection of
concepts with admissible operations of reasoning defined on them}. Knowledges associated
\textsf{PRAX} have various peculiarities corresponding to the
semantic evolution of rough objects in it. To start with, the semantic domains of representation
properly contain the semantic domains of interpretation. Not surprisingly, it is
because the rough objects corresponding to $l,\, u$ cannot be represented perfectly in
terms of objects from $\delta_{lu}(S)$ alone. \emph{In the nongranular perspective too,
this representation aspect should matter} - ''\emph{should}'', because it is matter of
choice during generalization from the classical case in the non granular approach.
The natural \rsds of $l,\, u$ is Meta-R, while that of $l_{o},\, u_{o}$ is
$\mathfrak{O}$ (say, corresponding rough objects of $\tau (R)$). These can be seen as
separate domains or as parts of a minimal containing domain that permits enough
expression. As we have seen knowledge is correctly representable in terms of \emph{atomic concepts of knowledge} at semantic
domains placed between Meta-C and Meta-R and not at the latter. So the
characterization of possible semantic domains and their mutual ordering - leading to their
geometry is of interest.
The following will be assumed to be \emph{part} of the interpretation:
\begin{mitemize}
\item {Two types of rough objects corresponding to Meta-R and $\mathfrak{O}$ and their
natural correspondence correspond to concepts or weakenings thereof. A concept relative
one semantic domain need not be one of the other.}
\item {A granule of the \rsd $\mathcal{O}$ is necessarily a concept of $\mathcal{O}$, but
a granule of Meta-R may not be a concept of $\mathcal{O}$ or Meta-R.}
\item {Critical points are not necessarily concepts of either semantic domain.}
\item {Critical points and the representation of rough objects require the \rsds to be
extended.}
\end{mitemize}
The above obviously assumes that a \textsf{PRAX} $S$ has at least two kinds of knowledge
associated (in relation to the Pawlak-sense interpretation). To make the interpretations
precise, we will indicate them by $\mathcal{I}_{1}(S)$ and $\mathcal{I}_{o}(S)$
respectively (corresponding to the approximations to $l,\, u$ and $l_{o},\, u_{o}$
respectively). The pair $(\mathcal{I}_{1}(S),\,\mathcal{I}_{o}(S))$ will also be
referred to as the \emph{generalized \textsf{KI}}.
\begin{definition}
Given two \textsf{PRAX} $S = \left\langle \underline{S},\,R \right\rangle$, $
V = \left\langle \underline{S},\,Q \right\rangle$, $S$ will be said to be \emph{o-coarser}
than $V$ \ifof $\mathcal{I}_{o}(S)$ is coarser than $\mathcal{I}_{o}(V)$ in Pawlak-sense
( that is $\tau(R)\subseteq \tau (Q)$). Conversely, $V$ will be said to be a
\emph{o-refinement} of $S$.
$S$ will be said to be \emph{p-coarser} than $V$ \ifof
$\mathcal{I}_{1}(S)$ is coarser than $\mathcal{I}_{1}(V)$ in the sense $R\subseteq Q$.
Conversely, $V$ will be said to be a \emph{p-refinement} of $S$.
\end{definition}
An extended concept of positive regions is defined next.
\begin{definition}
If $S_{1}\,=\, \left\langle \underline{S},Q \right\rangle $ and $S_{2}\,=\, \left\langle
\underline{S},P \right\rangle $ are two \textsf{PRAX} such that $Q\subset R$, then
by the
\emph{granular positive region} of $Q$ with respect to $R$ is given by $gPOS_{R}(Q)\,=\,
\{ [x]_{Q}^{l_{R}}\, :\, x\in \underline{S} \}$, where $[x]_{Q}^{l_{R}}$ is the lower
approximation (relative $R$) of the $Q$-related elements of $x$. Using this we can define
the \emph{granular extent of dependence} of knowledge encoded by $R$ on the knowledge
encoded by $Q$ by natural injections $:gPOS_{R}(Q) \longmapsto \mathcal{G}_{R} $.
\end{definition}
Lower critical points can be naturally interpreted as preconcepts that are
definitely included in the discourse, while upper critical points are preconcepts that
include most of the discourse. The problem with this interpretation is that it's
representation requires a semantic domain at which critical points of different kinds can
be found. A key requirement for such a domain would be the feasibility of rough counting
procedures like \textsf{IPC} \cite{AM240}. We will refer to a semantic domain that has
critical points of different types as basic objects as a \textsf{Meta-RC}.
The following possible axioms of granular knowledge that also figure in \cite{AM909} (due to the the present author), get into difficulties with the present approach and even when we restrict attention to $\mathcal{I}_{1}(S)$:
\begin{enumerate}
\item {Individual granules are atomic units of knowledge.}
\item {Maximal collections of granules subject to a concept of mutual
independence are admissible concepts of knowledge.}
\item {Parts common to subcollections of maximal collections of granules are also
knowledge.}
\end{enumerate}
The first axiom holds in weakened form as the granulation $\mathcal{G}$ for
$\mathcal{I}_{1}(S)$ is only lower definite and affects the other. The possibility of
other nice granulations being possible for the \textsf{PRAX} case appears to be possible
at the cost of other nice properties. So we can conclude that in proper
\textsf{KR} happens at semantic domains like \textsf{Meta-RC} where critical
points of different types are perceived. Further at Meta-R, rough objects may correspond
to knowledge or conjectures - if we require the concept of proof to be an ontological
concept or beliefs. The scenario can be made more complex with associations
of $\mathfrak{O}$ knowledges.
From a non-granular perspective, in Meta-R rough objects must
correspond to knowledge with some of them lacking a proper evolution - there is no problem
here. Even if we permit $\mathfrak{O}$ objects, then in the perspective we would be able
to speak of two kinds of closely associated knowledges.
The connections with non-monotonic reasoning and the approximate Nelson algebra semantics developed in this monograph suggest further enhancements to the above. These will be explored separately.
\chapter*{Further Directions and Remarks}
In this research we have developed the basic theory of rough sets over proto transitive relations, characterized the nature of rough objects and possible approximations, and have developed two different algebraic semantics for the same. Some of the work constitute a continuation of earlier work by the present author. Important connections between approximations and operators of generalized operators of non-monotonic reasoning have also been established in the monograph. This opens the door for many new kinds of semantic connections that we hope to consider in future. Various examples at the level of real-life applications are also outlined in the monograph. Concepts have been illustrated through a persistent example. Knowledge interpretation in \textsf{PRAX} contexts have also ben outlined.
In continuation of earlier work by the present author \cite{AM3690} on semantic consequences of the relation between protransitivity and its approximations is developed in detail. The theory of rough dependence from the knowledge perspective is also specialized to \textsf{PRAX} and extended for the purposes of the semantics in this monograph. Connections with probabilistic dependence is shown to be lacking any reasonable basis and we once again unsettle unbridled frequentism in rough set theory. The relation of the developed theory with entropy is strongly motivated by the knowledge interpretation \cite{AM909,AM270} and will be part of future work.
The first algebraic semantics was seen to be inadequate in not being particularly elegant and requiring additional predicates for a reasonable abstract representation theorem. This was one reason for restricting derivations involving rough objects of $\tau(R)$.
The internalization of a semantics of \textsf{PRAX} objects in Nelson algebras through ideas of rough dependence is shown to lead to a beautiful semantics. Further formulations of associated logics will be part of a future paper. The technique can be extended to define approximate semantics in various other rough set-theoretical contexts.
\bibliographystyle{splncs.bst}
|
\section*{References}}
\journal{Journal of Multivariate Analysis}
\begin{document}
\begin{frontmatter}
\title{Second order statistics of robust estimators of scatter. \\ Application to GLRT detection for elliptical signals\tnoteref{t1}}
\tnotetext[t1]{ Couillet's work is supported by the ERC MORE EC--120133. Pascal's work is supported by the DGA grant no. 2013.60.0011.00.470.75.01.}
\author[supelec]{Romain Couillet}
\ead{<EMAIL>}
\address[supelec]{Telecommunication department, Sup\'elec, Gif sur Yvette, France}
\author[kaust]{Abla Kammoun}
\ead{<EMAIL>}
\address[kaust]{King Abdullah's University of Science and Technology, Saudi Arabia}
\author[sondra]{Fr\'ed\'eric Pascal}
\ead{<EMAIL>}
\address[sondra]{SONDRA Laboratory, Sup\'elec, Gif sur Yvette, France}
\begin{abstract}
A central limit theorem for bilinear forms of the type $a^*\hat{C}_N(\rho)^{-1}b$, where $a,b\in\CC^N$ are unit norm deterministic vectors and $\hat{C}_N(\rho)$ a robust-shrinkage estimator of scatter parametrized by $\rho$ and built upon $n$ independent elliptical vector observations, is presented. The fluctuations of $a^*\hat{C}_N(\rho)^{-1}b$ are found to be of order $N^{-\frac12}$ and to be the same as those of $a^*\hat{S}_N(\rho)^{-1}b$ for $\hat{S}_N(\rho)$ a matrix of a theoretical tractable form. This result is exploited in a classical signal detection problem to provide an improved detector which is both robust to elliptical data observations (e.g., impulsive noise) and optimized across the shrinkage parameter $\rho$.
\end{abstract}
\begin{keyword}
random matrix theory \sep robust estimation \sep central limit theorem \sep GLRT.
\end{keyword}
\end{frontmatter}
\section{Introduction}
\label{sec:intro}
As an aftermath of the growing interest for large dimensional data analysis in machine learning, in a recent series of articles \citep{COU13,COU13b,COU14,ZHA14,KAR13}, several estimators from the field of robust statistics (dating back to the seventies) started to be explored under the assumption of commensurably large sample ($n$) and population ($N$) dimensions. Robust estimators were originally designed to turn classical estimators into outlier- and impulsive noise-resilient estimators, which is of considerable importance in the recent big data paradigm. Among these estimation methods, robust regression was studied in \citep{KAR13} which reveals that, in the large $N,n$ regime, the difference in norm between estimated and true regression vectors (of size $N$) tends almost surely to a positive constant which depends on the nature of the data and of the robust regressor. In parallel, and of more interest to the present work, \citep{COU13,COU13b,COU14,ZHA14} studied the limiting behavior of several classes of robust estimators $\hat{C}_N$ of scatter (or covariance) matrices $C_N$ based on independent zero-mean elliptical observations $x_1,\ldots,x_n\in\CC^N$. Precisely, \citep{COU13} shows that, letting $N/n<1$ and $\hat{C}_N$ be the (almost sure) unique solution to
\begin{align*}
\hat{C}_N &= \frac1n \sum_{i=1}^n u\left( \frac1Nx_i^*\hat{C}_N^{-1}x_i \right) x_ix_i^*
\end{align*}
under some appropriate conditions over the nonnegative function $u$ (corresponding to Maronna's M-estimator \citep{MAR76}), $\Vert\hat{C}_N-\hat{S}_N\Vert\asto 0$ in spectral norm as $N,n\to\infty$ with $N/n\to c\in(0,1)$, where $\hat{S}_N$ follows a standard random matrix model (such as studied in \citep{CHO95,HAC13}). In \citep{ZHA14}, the important scenario where $u(x)=1/x$ (referred to as Tyler's M-estimator) is treated. It is in particular shown for this model that for identity scatter matrices the spectrum of $\hat{C}_N$ converges weakly to the Mar\u{c}enko--Pastur law \citep{MAR67} in the large $N,n$ regime. Finally, for $N/n\to c\in(0,\infty)$, \citep{COU14} studied yet another robust estimation model defined, for each $\rho\in(\max\{0,1-n/N\},1]$, by $\hat{C}_N=\hat{C}_N(\rho)$, unique solution to
\begin{align}
\label{eq:hatCN}
\hat{C}_N(\rho) &= \frac1n \sum_{i=1}^n \frac{x_ix_i^*}{\frac1Nx_i^*\hat{C}_N^{-1}(\rho)x_i} + \rho I_N.
\end{align}
This estimator, proposed in \citep{PAS13}, corresponds to a hybrid robust-shrinkage estimator reminding Tyler's M-estimator of scale \citep{TYL87} and Ledoit--Wolf's shrinkage estimator \citep{LED04}. This estimator is particularly suited to scenarios where $N/n$ is not small, for which other estimators are badly conditioned if not undefined. For this model, it is shown in \citep{COU14} that $\sup_{\rho}\Vert\hat{C}_N(\rho)-\hat{S}_N(\rho)\Vert\asto 0$ where $\hat{S}_N(\rho)$ also follows a classical random matrix model.
The aforementioned approximations $\hat{S}_N$ of the estimators $\hat{C}_N$, the structure of which is well understood (as opposed to $\hat{C}_N$ which is only defined implicitly), allow for both a good apprehension of the limiting behavior of $\hat{C}_N$ and more importantly for a better usage of $\hat{C}_N$ as an appropriate substitute for sample covariance matrices in various estimation problems in the large $N,n$ regime. The convergence in norm $\Vert\hat{C}_N-\hat{S}_N\Vert\asto 0$ is indeed sufficient in many cases to produce new consistent estimation methods based on $\hat{C}_N$ by simply replacing $\hat{C}_N$ by $\hat{S}_N$ in the problem defining equations. For example, the results of \citep{COU13b} led to the introduction of novel consistent estimators based on functionals of $\hat{C}_N$ (of the Maronna type) for power and direction-of-arrival estimation in array processing in the presence of impulsive noise or rare outliers \citep{COU14c}. Similarly, in \citep{COU14}, empirical methods were designed to estimate the parameter $\rho$ which minimizes the expected Frobenius norm $\tr [(\hat{C}_N(\rho)-C_N)^2]$, of interest for various outlier-prone applications dealing with non-small ratios $N/n$.\footnote{Other metrics may also be considered as in e.g.\@ \citep{YAN14} with $\rho$ chosen to minimize the return variance in a portfolio optimization problem.}
Nonetheless, when replacing $\hat{C}_N$ for $\hat{S}_N$ in deriving consistent estimates, if the convergence $\Vert\hat{C}_N-\hat{S}_N\Vert\asto 0$ helps in producing novel consistent estimates, this convergence (which comes with no particular speed) is in general not sufficient to assess the performance of the estimator for large but finite $N,n$. Indeed, when second order results such as central limit theorems need be established, say at rate $N^{-\frac12}$, to proceed similarly to the replacement of $\hat{C}_N$ by $\hat{S}_N$ in the analysis, one would ideally demand that $\Vert\hat{C}_N-\hat{S}_N\Vert=o(N^{-\frac12})$; but such a result, we believe, unfortunately does not hold. This constitutes a severe limitation in the exploitation of robust estimators as their performance as well as optimal fine-tuning often rely on second order performance. Concretely, for practical purposes in the array processing application of \citep{COU14c}, one may naturally ask which choice of the $u$ function is optimal to minimize the variance of (consistent) power and angle estimates. This question remains unanswered to this point for lack of better theoretical results.
\bigskip
The main purpose of the article is twofold. From a technical aspect, taking the robust shrinkage estimator $\hat{C}_N(\rho)$ defined by \eqref{eq:hatCN} as an example, we first show that, although the convergence $\Vert\hat{C}_N(\rho)-\hat{S}_N(\rho)\Vert\asto 0$ (from \citep[Theorem~1]{COU14}) may not be extensible to a rate $O(N^{1-\varepsilon})$, one has the bilinear form convergence $N^{1-\varepsilon} a^*(\hat{C}_N^k(\rho)-\hat{S}_N^k(\rho))b\asto 0$ for each $\varepsilon>0$, each $a,b\in\CC^N$ of unit norm, and each $k\in\ZZ$. This result implies that, if $\sqrt{N}a^*\hat{S}_N^k(\rho)b$ satisfies a central limit theorem, then so does $\sqrt{N}a^*\hat{C}_N^k(\rho)b$ with the same limiting variance. This result is of fundamental importance to any statistical application based on such quadratic forms. Our second contribution is to exploit this result for the specific problem of signal detection in impulsive noise environments via the generalized likelihood-ratio test, particularly suited for radar signals detection under elliptical noise \citep{CON95,PAS13}. In this context, we determine the shrinkage parameter $\rho$ which minimizes the probability of false detections and provide an empirical consistent estimate for this parameter, thus improving significantly over traditional sample covariance matrix-based estimators.
The remainder of the article introduces our main results in Section~\ref{sec:results} which are proved in Section~\ref{sec:proof}. Technical elements of proof are provided in the appendix.
{\it Notations:} In the remainder of the article, we shall denote $\lambda_1(X),\ldots,\lambda_n(X)$ the real eigenvalues of $n\times n$ Hermitian matrices $X$. The norm notation $\Vert \cdot\Vert$ being considered is the spectral norm for matrices and Euclidean norm for vectors. The symbol $\imath$ is the complex $\sqrt{-1}$.
\section{Main Results}
\label{sec:results}
Let $N,n\in\NN$, $c_N\triangleq N/n$, and $\rho \in (\max\{0,1-c_N^{-1}\},1]$.
Let also $x_1,\ldots,x_n\in\CC^N$ be $n$ independent random vectors defined by the following assumptions.
\begin{assumption}[Data vectors]
\label{ass:x}
For $i\in\{1,\ldots,n\}$, $x_i=\sqrt{\tau_i}A_Nw_i=\sqrt{\tau_i}z_i$, where
\begin{itemize}
\item $w_i\in\CC^N$ is Gaussian with zero mean and covariance $I_N$, independent across $i$;
\item $A_NA_N^*\triangleq C_N\in\CC^{N\times N}$ is such that $\nu_N\triangleq \frac1N\sum_{i=1}^N{\bm\delta}_{\lambda_i(C_N)}\to \nu$ weakly, $\limsup_N\Vert C_N\Vert <\infty$, and $\frac1N\tr C_N=1$;
\item $\tau_i>0$ are random or deterministic scalars.
\end{itemize}
\end{assumption}
Under Assumption~\ref{ass:x}, letting $\tau_i=\tilde{\tau}_i/\Vert w_i\Vert$ for some $\tilde{\tau}_i$ independent of $w_i$, $x_i$ belongs to the class of elliptically distributed random vectors. Note that the normalization $\frac1N\tr C_N=1$ is not a restricting constraint since the scalars $\tau_i$ may absorb any other normalization.
It has been well-established by the robust estimation theory that, even if the $\tau_i$ are independent, independent of the $w_i$, and that $\lim_n \frac1n \sum_{i=1}^n\tau_i=1$ a.s., the sample covariance matrix $\frac1n\sum_{i=1}^n x_ix_i^*$ is in general a poor estimate for $C_N$. Robust estimators of scatter were designed for this purpose \citep{MAR76,TYL87}. In addition, if $N/n$ is non trivial, a linear shrinkage of these robust estimators against the identity matrix often helps in regularizing the estimator as established in e.g., \citep{PAS13,CHE11}. The robust estimator of scatter considered in this work, which we denote $\hat{C}_N(\rho)$, is defined (originally in \citep{PAS13}) as the unique solution to
\begin{align*}
\hat{C}_N(\rho) &= (1-\rho) \frac1n \sum_{i=1}^n \frac{x_ix_i^*}{\frac1Nx_i\hat{C}_N^{-1}(\rho)x_i} + \rho I_N.
\end{align*}
\subsection{Theoretical Results}
The asymptotic behavior of this estimator was studied recently in \citep{COU14} in the regime where $N,n\to\infty$ in such a way that $c_N\to c\in(0,\infty)$. We first recall the important results of this article, which shall lay down the main concepts and notations of the present work.
First define
\begin{align*}
\hat{S}_N(\rho) &= \frac1{\gamma_N(\rho)}\frac{1-\rho}{1-(1-\rho)c_N} \frac1n\sum_{i=1}^n z_iz_i^* + \rho I_N
\end{align*}
where $\gamma_N(\rho)$ is the unique solution to
\begin{align*}
1 &= \int \frac{t}{\gamma_N(\rho)\rho + (1-\rho)t}\nu_N(dt).
\end{align*}
For any $\kappa>0$ small, define $\mathcal R_\kappa\triangleq [\kappa+\max\{0,1-c^{-1}\},1]$. Then, from \cite[Theorem~1]{COU14}, as $N,n\to\infty$ with $c_N\to c\in(0,\infty)$,
\begin{align*}
\sup_{\rho \in \mathcal R_\kappa} \left\Vert \hat{C}_N(\rho) - \hat{S}_N(\rho) \right\Vert \asto 0.
\end{align*}
A careful analysis of the proof of \cite[Theorem~1]{COU14} (which is performed in Section~\ref{sec:proof}) reveals that the above convergence can be refined as
\begin{align}
\label{eq:cv12-eps}
\sup_{\rho \in \mathcal R_\kappa} N^{\frac12-\varepsilon} \left\Vert \hat{C}_N(\rho) - \hat{S}_N(\rho) \right\Vert \asto 0
\end{align}
for each $\varepsilon>0$. This suggests that (well-behaved) functionals of $\hat{C}_N(\rho)$ fluctuating at a slower speed than $N^{-\frac12+\varepsilon}$ for some $\varepsilon>0$ follow the same statistics as the same functionals with $\hat{S}_N(\rho)$ in place of $\hat{C}_N(\rho)$. However, this result is quite weak as most limiting theorems (starting with the classical central limit theorems for independent scalar variables) deal with fluctuations of order $N^{-\frac12}$ and sometimes in random matrix theory of order $N^{-1}$. In our opinion, the convergence speed \eqref{eq:cv12-eps} cannot be improved to a rate $N^{-\frac12}$. Nonetheless, thanks to an averaging effect documented in Section~\ref{sec:proof}, the fluctuation of special forms of functionals of $\hat{C}_N(\rho)$ can be proved to be much slower. Although among these functionals we could have considered linear functionals of the eigenvalue distribution of $\hat{C}_N(\rho)$, our present concern (driven by more obvious applications) is rather on bilinear forms of the type $a^*\hat{C}_N^k(\rho)b$ for some $a,b\in\CC^N$ with $\Vert a\Vert=\Vert b\Vert=1$, $k\in\ZZ$.
Our first main result is the following.
\begin{theorem}[Fluctuation of bilinear forms]
\label{th:bilin}
Let $a,b\in\CC^N$ with $\Vert a\Vert=\Vert b\Vert=1$. Then, as $N,n\to\infty$ with $c_N\to c\in(0,\infty)$, for any $\varepsilon>0$ and every $k\in\ZZ$,
\begin{align*}
\sup_{\rho\in\mathcal R_\kappa} N^{1-\varepsilon} \left| a^*\hat{C}_N^{k}(\rho)b - a^*\hat{S}_N^{k}(\rho)b \right| &\asto 0.
\end{align*}
\end{theorem}
Some comments and remarks are in order. First, we recall that central limit theorems involving bilinear forms of the type $a^*\hat{S}_N^k(\rho)b$ are classical objects in random matrix theory (see e.g.\@ \citep{KAM09,MES08} for $k=-1$), particularly common in signal processing and wireless communications. These central limit theorems in general show fluctuations at speed $N^{-\frac12}$. This indicates, taking $\varepsilon<\frac12$ in Theorem~\ref{th:bilin} and using the fact that almost sure convergence implies weak convergence, that $a^*\hat{C}_N^k(\rho)b$ exhibits the same fluctuations as $a^*\hat{S}_N^k(\rho)b$, the latter being classical and tractable while the former is quite intricate at the onset, due to the implicit nature of $\hat{C}_N(\rho)$.
Of practical interest to many applications in signal processing is the case where $k=-1$. In the next section, we present a classical generalized maximum likelihood signal detection in impulsive noise, for which we shall characterize the shrinkage parameter $\rho$ that meets minimum false alarm rates.
\subsection{Application to Signal Detection}
In this section, we consider the hypothesis testing scenario by which an $N$-sensor array receives a vector $y\in\CC^N$ according to the following hypotheses
\begin{align*}
y &= \left\{
\begin{array}{ll}
x &,~\mathcal H_0 \\
\alpha p + x&,~\mathcal H_1
\end{array}
\right.
\end{align*}
in which $\alpha>0$ is some unknown scaling factor constant while $p\in\CC^N$ is deterministic and known at the sensor array (which often corresponds to a steering vector arising from a specific known angle), and $x$ is an impulsive noise distributed as $x_1$ according to Assumption~\ref{ass:x}.
For convenience, we shall take $\Vert p\Vert=1$.
Under $\mathcal H_0$ (the null hypothesis), a noisy observation from an impulsive source is observed while under $\mathcal H_1$ both information and noise are collected at the array. The objective is to decide on $\mathcal H_1$ versus $\mathcal H_0$ upon the observation $y$ and prior pure-noise observations $x_1,\ldots,x_n$ distributed according to Assumption~\ref{ass:x}. When $\tau_1,\ldots,\tau_n$ and $C_N$ are unknown, the corresponding generalized likelihood ratio test, derived in \citep{CON95}, reads
\begin{align*}
T_N(\rho) &\overset{\mathcal H_1}{\underset{\mathcal H_0}{\gtrless}} \Gamma
\end{align*}
for some detection threshold $\Gamma$ where
\begin{align*}
T_N(\rho) \triangleq \frac{|y^*\hat{C}_N^{-1}(\rho)p|}{\sqrt{y^*\hat{C}_N^{-1}(\rho)y}\sqrt{p^*\hat{C}_N^{-1}(\rho)p}}.
\end{align*}
More precisely, \citep{CON95} derived the detector $T_N(0)$ only valid when $n\geq N$. The relaxed detector $T_N(\rho)$ allows for a better conditioning of the estimator, in particular for $n\simeq N$. In \citep{PAS13}, $T_N(\rho)$ is used explicitly in a space-time adaptive processing setting but only simulation results were provided. Alternative metrics for similar array processing problems involve the signal-to-noise ratio loss minimization rather than likelihood ratio tests; in \citep{ABR13a,ABR13b}, the authors exploit the estimators $\hat{C}_N(\rho)$ but restrict themselves to the less tractable finite dimensional analysis.
Our objective is to characterize the false alarm performance of the detector. That is, provided $\mathcal H_0$ is the actual scenario (i.e.\@ $y=x$), we shall evaluate $P(T_N(\rho)>\Gamma)$. Since it shall appear that, under $\mathcal H_0$, $T_N(\rho)\asto 0$ for every fixed $\Gamma>0$ and every $\rho$, by dominated convergence $P(T_N(\rho)>\Gamma)\to 0$ which does not say much about the actual test performance for large but finite $N,n$. To avoid such empty statements, we shall then consider the non-trivial case where $\Gamma=N^{-\frac12}\gamma$ for some fixed $\gamma>0$. In this case our objective is to characterize the false alarm probability
\begin{align*}
P\left( T_N(\rho) > \frac{\gamma}{\sqrt{N}} \right).
\end{align*}
Before providing this result, we need some further reminders from \citep{COU14}. First define
\begin{align*}
\underline{\hat{S}}_N(\rho) &\triangleq (1-\rho)\frac1n\sum_{i=1}^n z_iz_i^* + \rho I_N.
\end{align*}
Then, from \cite[Lemma~1]{COU14}, for each $\rho\in (\max\{0,1-c^{-1}\},1]$,
\begin{align*}
\frac{\hat{S}_N(\rho)}{\rho+\frac1{\gamma_N(\rho)}\frac{1-\rho}{1-(1-\rho)c}} = \underline{\hat{S}}_N(\underline\rho)
\end{align*}
where
\begin{align*}
\underline\rho \triangleq \frac{{\rho}}{\rho+\frac1{\gamma_N(\rho)}\frac{1-\rho}{1-(1-\rho)c}}.
\end{align*}
Moreover, the mapping $\rho\mapsto \underline\rho$ is continuously increasing from $(\max\{0,1-c^{-1}\},1]$ onto $(0,1]$.
From classical random matrix considerations (see e.g.\@ \citep{SIL95}), letting $Z=[z_1,\ldots,z_n]\in\CC^{N\times n}$, the empirical spectral distribution\footnote{That is the normalized counting measure of the eigenvalues.} of $(1-\rho)\frac1nZ^*Z$ almost surely admits a weak limit $\mu$. The Stieltjes transform $m(z)\triangleq \int (t-z)^{-1}\mu(dt)$ of $\mu$ at $z\in\CC\setminus {\rm Supp}(\mu)$ is the unique complex solution with positive (resp.\@ negative) imaginary part if $\Im[z]>0$ (resp.\@ $\Im[z]<0$) and unique real positive solution if $\Im[z]=0$ and $\Re[z]<0$ to
\begin{align*}
m(z) &= \left( -z + c\int \frac{(1-\rho)t}{1+(1-\rho)tm(z)} \nu(dt) \right)^{-1}.
\end{align*}
We denote $m'(z)$ the derivative of $m(z)$ with respect to $z$ (recall that the Stieltjes transform of a positively supported measure is analytic, hence continuously differentiable, away from the support of the measure).
With these definitions in place and with the help of Theorem~\ref{th:bilin}, we are now ready to introduce the main result of this section.
\begin{theorem}[Asymptotic detector performance]
\label{th:T}
Under hypothesis $\mathcal H_0$, as $N,n\to\infty$ with $c_N\to c\in(0,\infty)$,
\begin{align*}
\sup_{\rho\in\mathcal R_\kappa} \left| P\left( T_N(\rho) > \frac{\gamma}{\sqrt{N}} \right) - \exp \left( - \frac{\gamma^2}{2\sigma_N^2(\underline\rho)} \right) \right| &\to 0
\end{align*}
where $\rho\mapsto\underline\rho$ is the aforementioned mapping and
\begin{align*}
\sigma_N^2(\underline{\rho}) &\triangleq \frac12 \frac{ p^*C_NQ_N^2(\underline\rho)p}{ p^*Q_N(\underline\rho)p \cdot \frac1N\tr C_NQ_N(\underline\rho)\cdot \left(1-c (1-\rho)^2 m(-\underline{\rho})^2 \frac1N\tr C_N^2Q_N^2(\underline\rho)\right) }
\end{align*}
with $Q_N(\underline\rho)\triangleq (I_N + (1-\underline{\rho})m(-\underline{\rho})C_N)^{-1}$.
\end{theorem}
Otherwise stated, $\sqrt{N}T_N(\rho)$ is uniformly well approximated by a Rayleigh distributed random variable $R_N(\underline\rho)$ with parameter $\sigma_N(\underline\rho)$.
Simulation results are provided in Figure~\ref{fig:hist_detector_20} and Figure~\ref{fig:hist_detector_100} which corroborate the results of Theorem~\ref{th:T} for $N=20$ and $N=100$, respectively (for a single value of $\rho$ though). Comparatively, it is observed, as one would expect, that larger values for $N$ induce improved approximations in the tails of the approximating distribution.
\begin{figure}[h!]
\centering
\begin{tabular}{cc}
\begin{tikzpicture}[font=\footnotesize,scale=.7]
\renewcommand{\axisdefaulttryminticks}{4}
\tikzstyle{every major grid}+=[style=densely dashed]
\tikzstyle{every axis y label}+=[yshift=-10pt]
\tikzstyle{every axis x label}+=[yshift=5pt]
\tikzstyle{every axis legend}+=[cells={anchor=west},fill=white,
at={(0.98,0.98)}, anchor=north east, font=\scriptsize ]
\begin{axis}[
xmin=0,
ymin=0,
xmax=4,
bar width=3pt,
grid=major,
ymajorgrids=false,
scaled ticks=true,
ylabel={Density}
]
\addplot+[ybar,mark=none,color=black,fill=gray!40!white] coordinates{
(0.05,0.055)(0.15,0.138)(0.25,0.263)(0.35,0.324)(0.45,0.42)(0.55,0.516)(0.65,0.544)(0.75,0.626)(0.85,0.621)(0.95,0.631)(1.05,0.639)(1.15,0.618)(1.25,0.625)(1.35,0.588)(1.45,0.515)(1.55,0.512)(1.65,0.417)(1.75,0.395)(1.85,0.331)(1.95,0.3)(2.05,0.232)(2.15,0.186)(2.25,0.143)(2.35,0.093)(2.45,0.081)(2.55,0.065)(2.65,0.053)(2.75,0.031)(2.85,0.019)(2.95,0.01)(3.05,0.006)(3.15,0.002)(3.25,0.)(3.35,0.001)(3.45,0.)(3.55,0.)(3.65,0.)(3.75,0.)(3.85,0.)(3.95,0.)(4.05,0.)(4.15,0.)(4.25,0.)(4.35,0.)(4.45,0.)(4.55,0.)(4.65,0.)(4.75,0.)(4.85,0.)(4.95,0.)(5.05,0.)
};
\addplot[black,smooth,line width=1pt] plot coordinates{
(0.,0.)(0.1,0.109902)(0.2,0.21619)(0.3,0.315448)(0.4,0.40464)(0.5,0.481262)(0.6,0.543458)(0.7,0.590088)(0.8,0.620745)(0.9,0.635727)(1.,0.635965)(1.1,0.62292)(1.2,0.598449)(1.3,0.564673)(1.4,0.523829)(1.5,0.478146)(1.6,0.429733)(1.7,0.380485)(1.8,0.332025)(1.9,0.285669)(2.,0.24241)(2.1,0.202932)(2.2,0.167636)(2.3,0.136674)(2.4,0.109997)(2.5,0.087403)(2.6,0.068576)(2.7,0.053135)(2.8,0.040662)(2.9,0.030736)(3.,0.02295)(3.1,0.01693)(3.2,0.012338)(3.3,0.008885)(3.4,0.006321)(3.5,0.004445)(3.6,0.003088)(3.7,0.00212)(3.8,0.001439)(3.9,0.000965)(4.,0.00064)(4.1,0.000419)(4.2,0.000271)(4.3,0.000174)(4.4,0.00011)(4.5,0.000069)(4.6,0.000042)(4.7,0.000026)(4.8,0.000016)(4.9,0.000009)(5.,0.000006)
};
\legend{ {Empirical hist.\@ of $T_N(\rho)$},{Distribution of $R_N(\underline\rho)$} }
\end{axis}
\end{tikzpicture}
&
\begin{tikzpicture}[font=\footnotesize,scale=.7]
\renewcommand{\axisdefaulttryminticks}{4}
\tikzstyle{every major grid}+=[style=densely dashed]
\tikzstyle{every axis y label}+=[yshift=-10pt]
\tikzstyle{every axis x label}+=[yshift=5pt]
\tikzstyle{every axis legend}+=[cells={anchor=west},fill=white,
at={(0.98,0.02)}, anchor=south east, font=\scriptsize ]
\begin{axis}[
xmin=0,
ymin=0,
xmax=4,
ymax=1,
bar width=1.5pt,
grid=major,
ymajorgrids=false,
scaled ticks=true,
mark repeat=10,
ylabel={Cumulative distribution}
]
\addplot[black,mark=*] coordinates{
(0.,0.)(0.006969,0.01)(0.140717,0.02)(0.20486,0.03)(0.244764,0.04)(0.283584,0.05)(0.315349,0.06)(0.346699,0.07)(0.377842,0.08)(0.406186,0.09)(0.429656,0.1)(0.451618,0.11)(0.475542,0.12)(0.500291,0.13)(0.520232,0.14)(0.539641,0.15)(0.559926,0.16)(0.577158,0.17)(0.596035,0.18)(0.616435,0.19)(0.638308,0.2)(0.657572,0.21)(0.673424,0.22)(0.68921,0.23)(0.704637,0.24)(0.719824,0.25)(0.736868,0.26)(0.752971,0.27)(0.768967,0.28)(0.784846,0.29)(0.801679,0.3)(0.819511,0.31)(0.835377,0.32)(0.848875,0.33)(0.866909,0.34)(0.880432,0.35)(0.89928,0.36)(0.913679,0.37)(0.932125,0.38)(0.948459,0.39)(0.965796,0.4)(0.977908,0.41)(0.992774,0.42)(1.010217,0.43)(1.023342,0.44)(1.039632,0.45)(1.053273,0.46)(1.068361,0.47)(1.086139,0.48)(1.105446,0.49)(1.121136,0.5)(1.136039,0.51)(1.151077,0.52)(1.168546,0.53)(1.184198,0.54)(1.200874,0.55)(1.21731,0.56)(1.233584,0.57)(1.248894,0.58)(1.266999,0.59)(1.28289,0.6)(1.297469,0.61)(1.314827,0.62)(1.331183,0.63)(1.34982,0.64)(1.366348,0.65)(1.381957,0.66)(1.398737,0.67)(1.418269,0.68)(1.435907,0.69)(1.456034,0.7)(1.476974,0.71)(1.496225,0.72)(1.514232,0.73)(1.530731,0.74)(1.551749,0.75)(1.570442,0.76)(1.591612,0.77)(1.614273,0.78)(1.63925,0.79)(1.664769,0.8)(1.685157,0.81)(1.709925,0.82)(1.733259,0.83)(1.759881,0.84)(1.787953,0.85)(1.816399,0.86)(1.844413,0.87)(1.872439,0.88)(1.906016,0.89)(1.938662,0.9)(1.974716,0.91)(2.008711,0.92)(2.050453,0.93)(2.097168,0.94)(2.146087,0.95)(2.201662,0.96)(2.268444,0.97)(2.365899,0.98)(2.481182,0.99)(2.634444,1.)
};
\addplot[black,smooth] plot coordinates{
(0.,0.)(0.1,0.00551)(0.2,0.02186)(0.3,0.048514)(0.4,0.084613)(0.5,0.129022)(0.6,0.180384)(0.7,0.237194)(0.8,0.297868)(0.9,0.36082)(1.,0.424522)(1.1,0.487569)(1.2,0.548725)(1.3,0.606949)(1.4,0.661424)(1.5,0.711554)(1.6,0.756962)(1.7,0.797473)(1.8,0.833086)(1.9,0.863948)(2.,0.890323)(2.1,0.912556)(2.2,0.931049)(2.3,0.946228)(2.4,0.958527)(2.5,0.968364)(2.6,0.976133)(2.7,0.982192)(2.8,0.986859)(2.9,0.990409)(3.,0.993078)(3.1,0.995058)(3.2,0.996511)(3.3,0.997564)(3.4,0.998318)(3.5,0.998851)(3.6,0.999224)(3.7,0.999481)(3.8,0.999657)(3.9,0.999776)(4.,0.999855)(4.1,0.999908)(4.2,0.999942)(4.3,0.999963)(4.4,0.999977)(4.5,0.999986)(4.6,0.999992)(4.7,0.999995)(4.8,0.999997)(4.9,0.999998)(5.,0.999999)
};
\legend{ {Empirical dist.\@ of $T_N(\rho)$},{Distribution of $R_N(\underline\rho)$} }
\end{axis}
\end{tikzpicture}
\end{tabular}
\caption{Histogram distribution function of the $\sqrt{N}T_N(\rho)$ versus $R_N(\underline\rho)$, $N=20$, $p=N^{-\frac12}[1,\ldots,1]^\trans$, $[C_N]_{ij}=0.7^{|i-j|}$, $c_N=1/2$, $\rho=0.2$.}
\label{fig:hist_detector_20}
\end{figure}
\begin{figure}[h!]
\centering
\begin{tabular}{cc}
\begin{tikzpicture}[font=\footnotesize,scale=.7]
\renewcommand{\axisdefaulttryminticks}{4}
\tikzstyle{every major grid}+=[style=densely dashed]
\tikzstyle{every axis y label}+=[yshift=-10pt]
\tikzstyle{every axis x label}+=[yshift=5pt]
\tikzstyle{every axis legend}+=[cells={anchor=west},fill=white,
at={(0.98,0.98)}, anchor=north east, font=\scriptsize ]
\begin{axis}[
xmin=0,
ymin=0,
xmax=4,
bar width=3pt,
grid=major,
ymajorgrids=false,
scaled ticks=true,
ylabel={Density}
]
\addplot+[ybar,mark=none,color=black,fill=gray!40!white] coordinates{
(0.05,0.062)(0.15,0.152)(0.25,0.292)(0.35,0.332)(0.45,0.412)(0.55,0.515)(0.65,0.576)(0.75,0.592)(0.85,0.614)(0.95,0.669)(1.05,0.638)(1.15,0.617)(1.25,0.615)(1.35,0.546)(1.45,0.532)(1.55,0.426)(1.65,0.391)(1.75,0.378)(1.85,0.331)(1.95,0.252)(2.05,0.225)(2.15,0.185)(2.25,0.144)(2.35,0.125)(2.45,0.098)(2.55,0.063)(2.65,0.066)(2.75,0.052)(2.85,0.032)(2.95,0.021)(3.05,0.017)(3.15,0.007)(3.25,0.008)(3.35,0.006)(3.45,0.003)(3.55,0.)(3.65,0.003)(3.75,0.002)(3.85,0.)(3.95,0.001)(4.05,0.)(4.15,0.)(4.25,0.)(4.35,0.)(4.45,0.)(4.55,0.)(4.65,0.)(4.75,0.)(4.85,0.)(4.95,0.)(5.05,0.)
};
\addplot[black,smooth,line width=1pt] plot coordinates{
(0.,0.)(0.1,0.109131)(0.2,0.214698)(0.3,0.313332)(0.4,0.402036)(0.5,0.478332)(0.6,0.540381)(0.7,0.587044)(0.8,0.617904)(0.9,0.633237)(1.,0.633945)(1.1,0.621449)(1.2,0.597572)(1.3,0.564395)(1.4,0.524123)(1.5,0.478956)(1.6,0.430981)(1.7,0.382081)(1.8,0.333873)(1.9,0.287674)(2.,0.244483)(2.1,0.204995)(2.2,0.169624)(2.3,0.138538)(2.4,0.111702)(2.5,0.088927)(2.6,0.069911)(2.7,0.054281)(2.8,0.041628)(2.9,0.031536)(3.,0.023602)(3.1,0.017452)(3.2,0.01275)(3.3,0.009204)(3.4,0.006566)(3.5,0.004629)(3.6,0.003225)(3.7,0.002221)(3.8,0.001511)(3.9,0.001017)(4.,0.000676)(4.1,0.000444)(4.2,0.000289)(4.3,0.000185)(4.4,0.000118)(4.5,0.000074)(4.6,0.000046)(4.7,0.000028)(4.8,0.000017)(4.9,0.00001)(5.,0.000006)
};
\legend{ {Empirical hist.\@ of $T_N(\rho)$},{Density of $R_N(\underline\rho)$} }
\end{axis}
\end{tikzpicture}
&
\begin{tikzpicture}[font=\footnotesize,scale=.7]
\renewcommand{\axisdefaulttryminticks}{4}
\tikzstyle{every major grid}+=[style=densely dashed]
\tikzstyle{every axis y label}+=[yshift=-10pt]
\tikzstyle{every axis x label}+=[yshift=5pt]
\tikzstyle{every axis legend}+=[cells={anchor=west},fill=white,
at={(0.98,0.02)}, anchor=south east, font=\scriptsize ]
\begin{axis}[
xmin=0,
ymin=0,
xmax=4,
ymax=1,
bar width=1.5pt,
grid=major,
ymajorgrids=false,
scaled ticks=true,
mark repeat=10,
ylabel={Cumulative distribution}
]
\addplot[black,mark=*] coordinates{
(0.,0.)(0.004147,0.01)(0.134196,0.02)(0.194839,0.03)(0.231715,0.04)(0.268037,0.05)(0.29753,0.06)(0.326162,0.07)(0.361765,0.08)(0.387635,0.09)(0.415288,0.1)(0.441611,0.11)(0.464966,0.12)(0.485413,0.13)(0.509661,0.14)(0.529458,0.15)(0.548931,0.16)(0.567224,0.17)(0.585663,0.18)(0.606538,0.19)(0.622153,0.2)(0.639809,0.21)(0.657092,0.22)(0.674893,0.23)(0.691094,0.24)(0.709848,0.25)(0.723726,0.26)(0.742494,0.27)(0.759849,0.28)(0.776713,0.29)(0.792786,0.3)(0.809041,0.31)(0.826186,0.32)(0.843159,0.33)(0.860687,0.34)(0.876359,0.35)(0.891876,0.36)(0.908243,0.37)(0.921541,0.38)(0.935376,0.39)(0.950716,0.4)(0.966973,0.41)(0.982105,0.42)(0.996863,0.43)(1.010958,0.44)(1.031895,0.45)(1.04803,0.46)(1.059607,0.47)(1.074379,0.48)(1.090857,0.49)(1.105972,0.5)(1.12501,0.51)(1.14239,0.52)(1.157615,0.53)(1.171878,0.54)(1.188458,0.55)(1.204927,0.56)(1.222216,0.57)(1.237269,0.58)(1.251976,0.59)(1.271192,0.6)(1.287183,0.61)(1.301946,0.62)(1.319419,0.63)(1.338576,0.64)(1.355493,0.65)(1.377815,0.66)(1.394742,0.67)(1.415819,0.68)(1.433655,0.69)(1.451259,0.7)(1.468387,0.71)(1.486074,0.72)(1.509745,0.73)(1.534089,0.74)(1.555325,0.75)(1.579848,0.76)(1.601953,0.77)(1.626237,0.78)(1.650685,0.79)(1.678268,0.8)(1.704434,0.81)(1.728554,0.82)(1.757243,0.83)(1.782361,0.84)(1.810967,0.85)(1.840125,0.86)(1.869583,0.87)(1.904995,0.88)(1.944813,0.89)(1.984082,0.9)(2.025297,0.91)(2.068876,0.92)(2.11431,0.93)(2.168671,0.94)(2.231536,0.95)(2.303031,0.96)(2.383405,0.97)(2.477321,0.98)(2.621018,0.99)(2.800403,1.)
};
\addplot[black,smooth] plot coordinates{
(0.,0.)(0.1,0.005472)(0.2,0.021707)(0.3,0.04818)(0.4,0.084042)(0.5,0.128172)(0.6,0.179233)(0.7,0.235735)(0.8,0.296114)(0.9,0.358798)(1.,0.422273)(1.1,0.485146)(1.2,0.546184)(1.3,0.60435)(1.4,0.658826)(1.5,0.709012)(1.6,0.754524)(1.7,0.795178)(1.8,0.830964)(1.9,0.86202)(2.,0.888599)(2.1,0.91104)(2.2,0.929735)(2.3,0.945108)(2.4,0.957585)(2.5,0.967584)(2.6,0.975496)(2.7,0.981679)(2.8,0.986451)(2.9,0.99009)(3.,0.99283)(3.1,0.99487)(3.2,0.996369)(3.3,0.997458)(3.4,0.99824)(3.5,0.998795)(3.6,0.999184)(3.7,0.999453)(3.8,0.999638)(3.9,0.999762)(4.,0.999846)(4.1,0.999901)(4.2,0.999937)(4.3,0.999961)(4.4,0.999976)(4.5,0.999985)(4.6,0.999991)(4.7,0.999995)(4.8,0.999997)(4.9,0.999998)(5.,0.999999)
};
\legend{ {Empirical dist.\@ of $T_N(\rho)$},{Distribution of $R_N(\underline\rho)$} }
\end{axis}
\end{tikzpicture}
\end{tabular}
\caption{Histogram distribution function of the $\sqrt{N}T_N(\rho)$ versus $R_N(\underline\rho)$, $N=100$, $p=N^{-\frac12}[1,\ldots,1]^\trans$, $[C_N]_{ij}=0.7^{|i-j|}$, $c_N=1/2$, $\rho=0.2$.}
\label{fig:hist_detector_100}
\end{figure}
The result of Theorem~\ref{th:T} provides an analytical characterization of the performance of the GLRT for each $\rho$ which suggests in particular the existence of values for $\rho$ which minimize the false alarm probability for given $\gamma$. Note in passing that, independently of $\gamma$, minimizing the false alarm rate is asymptotically equivalent to minimizing $\sigma_N^2(\underline\rho)$ over $\rho$. However, the expression of $\sigma_N^2(\underline\rho)$ depends on the covariance matrix $C_N$ which is unknown to the array and therefore does not allow for an immediate online choice of an appropriate $\underline\rho$. To tackle this problem, the following proposition provides a consistent estimate for $\sigma_N^2(\underline\rho)$ based on $\hat{C}_N(\rho)$ and $p$.
\begin{proposition}[Empirical performance estimation]
\label{prop:1}
For $\rho\in(\max\{0,1-c_N^{-1}\},1)$ and $\underline\rho$ defined as above, let $\hat{\sigma}_N^2(\underline\rho)$ be given by
\begin{align*}
\hat{\sigma}_N^2(\underline\rho) &\triangleq \frac12 \frac{1-\underline\rho \cdot \frac{p^*\hat{C}_N^{-2}(\rho)p}{p^*\hat{C}_N^{-1}(\rho)p}\cdot \frac1N\tr \hat{C}_N(\rho)}{\left( 1-c + c\underline\rho \frac1N\tr\hat{C}_N^{-1}(\rho)\cdot \frac1N\tr \hat{C}_N(\rho) \right)\left( 1 - \underline\rho \frac1N\tr\hat{C}_N^{-1}(\rho)\cdot \frac1N\tr \hat{C}_N(\rho)\right)}.
\end{align*}
Also let $\hat{\sigma}_N^2(1)\triangleq \lim_{\underline\rho\uparrow 1}\hat{\sigma}_N^2(\underline\rho)$. Then we have
\begin{align*}
\sup_{\rho \in \mathcal R_\kappa} \left| \sigma_N^2(\underline\rho) - \hat{\sigma}_N^2(\underline\rho) \right| &\asto 0.
\end{align*}
\end{proposition}
Since both the estimation of $\sigma_N^2(\underline\rho)$ in Proposition~\ref{prop:1} and the convergence in Theorem~\ref{th:T} are uniform over $\rho\in\mathcal R_\kappa$, we have the following result.
\begin{corollary}[Empirical performance optimum]
\label{co:1}
Let $\hat{\sigma}_N^2(\underline\rho)$ be defined as in Proposition~\ref{prop:1} and define $\hat{\rho}_N^*$ as any value satisfying
\begin{align*}
\hat{\rho}_N^* &\in \argmin_{ \rho\in \mathcal R_\kappa } \left\{ \hat{\sigma}_N^2(\underline\rho) \right\}
\end{align*}
(this set being in general a singleton). Then, for every $\gamma>0$,
\begin{align*}
P\left( \sqrt{N}T_N(\hat{\rho}_N^*) > \gamma \right) - \inf_{\rho\in \mathcal R_\kappa} \left\{ P\left( \sqrt{N}T_N(\rho) > \gamma \right) \right\} &\to 0.
\end{align*}
\end{corollary}
This last result states that, for $N,n$ sufficiently large, it is increasingly close-to-optimal to use the detector $T_N(\hat{\rho}_N^*)$ in order to reach minimal false alarm probability. A practical graphical confirmation of this fact is provided in Figure~\ref{fig:FAR} where, in the same scenario as in Figures~\ref{fig:hist_detector_20}--\ref{fig:hist_detector_100}, the false alarm rates for various values of $\gamma$ are depicted. In this figure, the black dots correspond to the actual values taken by $P(\sqrt{N}T_N(\rho)>\gamma)$ empirically obtained out of $10^6$ Monte Carlo simulations. The plain curves are the approximating values $\exp(-\gamma^2/(2\sigma_N(\underline\rho)^2))$. Finally, the white dots with error bars correspond to the mean and standard deviations of $\exp(-\gamma^2/(2\hat\sigma_N(\underline\rho)^2))$ for each $\underline\rho$, respectively.
It is first interesting to note that the estimates $\hat\sigma_N(\underline\rho)$ are quite accurate, especially so for $N$ large, with standard deviations sufficiently small to provide good estimates, already for small $N$, of the false alarm minimizing $\rho$. However, similar to Figures~\ref{fig:hist_detector_20}--\ref{fig:hist_detector_100}, we observe a particularly weak approximation in the (small) $N=20$ setting for large values of $\gamma$, corresponding to tail events, while for $N=100$, these values are better recovered. This behavior is obviously explained by the fact that $\gamma=3$ is not small compared to $\sqrt{N}$ when $N=20$.
Nonetheless, from an error rate viewpoint, it is observed that errors of order $10^{-2}$ are rather well approximated for $N=100$. In Figure~\ref{fig:FAR2}, we consider this observation in depth by displaying $P(T_N(\hat{\rho}_N^*)>\Gamma)$ and its approximation $\exp(-N\Gamma^2/(2\hat\sigma_N^2(\underline\rho)))$ for $N=20$ and $N=100$, for various values of $\Gamma$. This figures shows that even errors of order $10^{-4}$ are well approximated for large $N$, while only errors of order $10^{-2}$ can be evaluated for small $N$.\footnote{Note that a comparison against alternative algorithms that would use no shrinkage (i.e., by setting $\rho=0$) or that would not implement a robust estimate is not provided here, being of little relevance. Indeed, a proper selection of $c_N$ to a large value or $C_N$ with condition number close to one would provide an arbitrarily large gain of shrinkage-based methods, while an arbitrarily heavy-tailed choice of the $\tau_i$ distribution would provide a huge performance gain for robust methods. It is therefore not possible to compare such methods on fair grounds.}
\begin{figure}[h!]
\centering
\begin{tabular}{cc}
\begin{tikzpicture}[font=\footnotesize,scale=.7]
\renewcommand{\axisdefaulttryminticks}{4}
\tikzstyle{every major grid}+=[style=densely dashed]
\tikzstyle{every axis y label}+=[yshift=-10pt]
\tikzstyle{every axis x label}+=[yshift=5pt]
\tikzstyle{every axis legend}+=[cells={anchor=west},fill=white,
at={(0.98,0.02)}, anchor=south east, font=\scriptsize ]
\begin{semilogyaxis}[
xmin=0,
xmax=1,
ymax=1,
grid=major,
ymajorgrids=false,
scaled ticks=true,
xlabel={$\rho$},
ylabel={$P(\sqrt{N}T_N(\rho)>\gamma)$}
]
\addplot[black] plot coordinates{
(0.010000,0.132470)(0.020000,0.129832)(0.030000,0.127406)(0.040000,0.125179)(0.050000,0.123139)(0.060000,0.121277)(0.070000,0.119582)(0.080000,0.118045)(0.090000,0.116660)(0.100000,0.115417)(0.110000,0.114311)(0.120000,0.113336)(0.130000,0.112486)(0.140000,0.111756)(0.150000,0.111141)(0.160000,0.110639)(0.170000,0.110244)(0.180000,0.109954)(0.190000,0.109766)(0.200000,0.109677)(0.210000,0.109685)(0.220000,0.109788)(0.230000,0.109983)(0.240000,0.110270)(0.250000,0.110648)(0.260000,0.111114)(0.270000,0.111669)(0.280000,0.112311)(0.290000,0.113040)(0.300000,0.113856)(0.310000,0.114758)(0.320000,0.115747)(0.330000,0.116822)(0.340000,0.117984)(0.350000,0.119233)(0.360000,0.120570)(0.370000,0.121996)(0.380000,0.123511)(0.390000,0.125116)(0.400000,0.126811)(0.410000,0.128599)(0.420000,0.130480)(0.430000,0.132456)(0.440000,0.134528)(0.450000,0.136696)(0.460000,0.138964)(0.470000,0.141332)(0.480000,0.143802)(0.490000,0.146376)(0.500000,0.149055)(0.510000,0.151841)(0.520000,0.154736)(0.530000,0.157743)(0.540000,0.160862)(0.550000,0.164095)(0.560000,0.167446)(0.570000,0.170915)(0.580000,0.174504)(0.590000,0.178216)(0.600000,0.182052)(0.610000,0.186014)(0.620000,0.190104)(0.630000,0.194324)(0.640000,0.198675)(0.650000,0.203159)(0.660000,0.207776)(0.670000,0.212530)(0.680000,0.217420)(0.690000,0.222448)(0.700000,0.227615)(0.710000,0.232922)(0.720000,0.238368)(0.730000,0.243955)(0.740000,0.249682)(0.750000,0.255549)(0.760000,0.261556)(0.770000,0.267703)(0.780000,0.273987)(0.790000,0.280408)(0.800000,0.286965)(0.810000,0.293655)(0.820000,0.300476)(0.830000,0.307425)(0.840000,0.314500)(0.850000,0.321698)(0.860000,0.329014)(0.870000,0.336445)(0.880000,0.343987)(0.890000,0.351635)(0.900000,0.359384)(0.910000,0.367230)(0.920000,0.375167)(0.930000,0.383190)(0.940000,0.391293)(0.950000,0.399471)(0.960000,0.407718)(0.970000,0.416027)(0.980000,0.424394)(0.990000,0.432813)(1.000000,0.441279)
};
\addplot[black,only marks,mark=o,mark options={scale=0.75},error bars/.cd,y dir=both,y explicit, error bar style={mark size=1.5pt}] plot coordinates{
(0.050000,0.122858)+-(0.004658,0.004658)(0.100000,0.115290)+-(0.007070,0.007070)(0.150000,0.111293)+-(0.008363,0.008363)(0.200000,0.110090)+-(0.009027,0.009027)(0.250000,0.111235)+-(0.009386,0.009386)(0.300000,0.114445)+-(0.009656,0.009656)(0.350000,0.119667)+-(0.009976,0.009976)(0.400000,0.126863)+-(0.010424,0.010424)(0.450000,0.136131)+-(0.011170,0.011170)(0.500000,0.147467)+-(0.012326,0.012326)(0.550000,0.161167)+-(0.013813,0.013813)(0.600000,0.177344)+-(0.015758,0.015758)(0.650000,0.196270)+-(0.018054,0.018054)(0.700000,0.218062)+-(0.020671,0.020671)(0.750000,0.242979)+-(0.023689,0.023689)(0.800000,0.270952)+-(0.026959,0.026959)(0.850000,0.301740)+-(0.030677,0.030677)(0.900000,0.334893)+-(0.034650,0.034650)(0.950000,0.369402)+-(0.038847,0.038847)
};
\addplot[black,only marks,mark=*,mark options={scale=0.75}] plot coordinates{
(0.050000,0.102554)(0.100000,0.095524)(0.150000,0.092681)(0.200000,0.091376)
(0.250000,0.093394)(0.300000,0.097212)(0.350000,0.103109)(0.400000,0.111273)(0.450000,0.121610)(0.500000,0.134944)(0.550000,0.152715)(0.600000,0.172006)(0.650000,0.197016)(0.700000,0.226062)(0.750000,0.258595)(0.800000,0.296503)(0.850000,0.339433)(0.900000,0.384739)(0.950000,0.432905)(1.000000,0.478739)
};
\addplot[black] plot coordinates{
(0.010000,0.010587)(0.020000,0.010118)(0.030000,0.009698)(0.040000,0.009321)(0.050000,0.008982)(0.060000,0.008680)(0.070000,0.008409)(0.080000,0.008168)(0.090000,0.007954)(0.100000,0.007764)(0.110000,0.007598)(0.120000,0.007453)(0.130000,0.007328)(0.140000,0.007221)(0.150000,0.007132)(0.160000,0.007060)(0.170000,0.007003)(0.180000,0.006962)(0.190000,0.006935)(0.200000,0.006922)(0.210000,0.006924)(0.220000,0.006938)(0.230000,0.006966)(0.240000,0.007007)(0.250000,0.007061)(0.260000,0.007128)(0.270000,0.007208)(0.280000,0.007302)(0.290000,0.007409)(0.300000,0.007530)(0.310000,0.007665)(0.320000,0.007814)(0.330000,0.007979)(0.340000,0.008158)(0.350000,0.008354)(0.360000,0.008566)(0.370000,0.008796)(0.380000,0.009043)(0.390000,0.009310)(0.400000,0.009596)(0.410000,0.009903)(0.420000,0.010232)(0.430000,0.010584)(0.440000,0.010960)(0.450000,0.011362)(0.460000,0.011790)(0.470000,0.012247)(0.480000,0.012734)(0.490000,0.013253)(0.500000,0.013805)(0.510000,0.014392)(0.520000,0.015017)(0.530000,0.015681)(0.540000,0.016388)(0.550000,0.017138)(0.560000,0.017936)(0.570000,0.018783)(0.580000,0.019682)(0.590000,0.020636)(0.600000,0.021649)(0.610000,0.022724)(0.620000,0.023863)(0.630000,0.025072)(0.640000,0.026352)(0.650000,0.027710)(0.660000,0.029147)(0.670000,0.030669)(0.680000,0.032279)(0.690000,0.033983)(0.700000,0.035785)(0.710000,0.037690)(0.720000,0.039702)(0.730000,0.041826)(0.740000,0.044068)(0.750000,0.046432)(0.760000,0.048924)(0.770000,0.051549)(0.780000,0.054312)(0.790000,0.057218)(0.800000,0.060272)(0.810000,0.063479)(0.820000,0.066845)(0.830000,0.070374)(0.840000,0.074071)(0.850000,0.077939)(0.860000,0.081985)(0.870000,0.086210)(0.880000,0.090619)(0.890000,0.095215)(0.900000,0.100002)(0.910000,0.104981)(0.920000,0.110156)(0.930000,0.115527)(0.940000,0.121096)(0.950000,0.126865)(0.960000,0.132834)(0.970000,0.139003)(0.980000,0.145372)(0.990000,0.151942)(1.000000,0.158710)
};
\addplot[black,only marks,mark=o,mark options={scale=0.75},error bars/.cd,y dir=both,y explicit, error bar style={mark size=1.5pt}] plot coordinates{
(0.050000,0.008954)+-(0.000743,0.000743)(0.100000,0.007786)+-(0.001034,0.001034)(0.150000,0.007211)+-(0.001170,0.001170)(0.200000,0.007047)+-(0.001250,0.001250)(0.250000,0.007217)+-(0.001323,0.001323)(0.300000,0.007694)+-(0.001418,0.001418)(0.350000,0.008505)+-(0.001558,0.001558)(0.400000,0.009696)+-(0.001762,0.001762)(0.450000,0.011363)+-(0.002074,0.002074)(0.500000,0.013608)+-(0.002542,0.002542)(0.550000,0.016628)+-(0.003196,0.003196)(0.600000,0.020636)+-(0.004120,0.004120)(0.650000,0.025945)+-(0.005360,0.005360)(0.700000,0.032905)+-(0.006995,0.006995)(0.750000,0.042004)+-(0.009156,0.009156)(0.800000,0.053704)+-(0.011903,0.011903)(0.850000,0.068460)+-(0.015448,0.015448)(0.900000,0.086600)+-(0.019812,0.019812)(0.950000,0.108034)+-(0.025015,0.025015)
};
\addplot[black,only marks,mark=*,mark options={scale=0.75}] plot coordinates{
(0.050000,0.001664)(0.100000,0.001371)(0.150000,0.001260)(0.200000,0.001244)(0.250000,0.001266)(0.300000,0.001392)(0.350000,0.001612)(0.400000,0.001974)(0.450000,0.002550)(0.500000,0.003313)(0.550000,0.004514)(0.600000,0.006323)(0.650000,0.009090)(0.700000,0.013042)(0.750000,0.019099)(0.800000,0.027746)(0.850000,0.040506)(0.900000,0.057496)(0.950000,0.080889)(1.000000,0.108458)
};
\node at (axis cs:0.4,0.3) {$\gamma=2$};
\draw (axis cs:0.4,0.12) ellipse [black,x radius=2,y radius=0.5];
\node at (axis cs:0.4,0.03) {$\gamma=3$};
\draw (axis cs:0.4,0.005) ellipse [black,x radius=2,y radius=1.5];
\legend{ {Limiting theory},{Empirical estimator},{Detector} }
\end{semilogyaxis}
\end{tikzpicture}
&
\begin{tikzpicture}[font=\footnotesize,scale=.7]
\renewcommand{\axisdefaulttryminticks}{4}
\tikzstyle{every major grid}+=[style=densely dashed]
\tikzstyle{every axis y label}+=[yshift=-10pt]
\tikzstyle{every axis x label}+=[yshift=5pt]
\tikzstyle{every axis legend}+=[cells={anchor=west},fill=white,
at={(0.98,0.02)}, anchor=south east, font=\scriptsize ]
\begin{semilogyaxis}[
xmin=0,
xmax=1,
ymax=1,
grid=major,
ymajorgrids=false,
scaled ticks=true,
xlabel={$\rho$},
ylabel={$P(\sqrt{N}T_N(\rho)>\gamma)$}
]
\addplot[black] plot coordinates{
(0.010000,0.132556)(0.020000,0.130002)(0.030000,0.127659)(0.040000,0.125515)(0.050000,0.123558)(0.060000,0.121777)(0.070000,0.120163)(0.080000,0.118708)(0.090000,0.117404)(0.100000,0.116244)(0.110000,0.115221)(0.120000,0.114330)(0.130000,0.113565)(0.140000,0.112922)(0.150000,0.112395)(0.160000,0.111983)(0.170000,0.111680)(0.180000,0.111483)(0.190000,0.111391)(0.200000,0.111401)(0.210000,0.111510)(0.220000,0.111717)(0.230000,0.112020)(0.240000,0.112418)(0.250000,0.112910)(0.260000,0.113495)(0.270000,0.114171)(0.280000,0.114940)(0.290000,0.115800)(0.300000,0.116752)(0.310000,0.117795)(0.320000,0.118930)(0.330000,0.120157)(0.340000,0.121477)(0.350000,0.122890)(0.360000,0.124398)(0.370000,0.126000)(0.380000,0.127699)(0.390000,0.129495)(0.400000,0.131391)(0.410000,0.133386)(0.420000,0.135483)(0.430000,0.137683)(0.440000,0.139989)(0.450000,0.142401)(0.460000,0.144923)(0.470000,0.147555)(0.480000,0.150300)(0.490000,0.153159)(0.500000,0.156136)(0.510000,0.159233)(0.520000,0.162451)(0.530000,0.165793)(0.540000,0.169262)(0.550000,0.172859)(0.560000,0.176587)(0.570000,0.180449)(0.580000,0.184447)(0.590000,0.188584)(0.600000,0.192861)(0.610000,0.197281)(0.620000,0.201846)(0.630000,0.206559)(0.640000,0.211421)(0.650000,0.216435)(0.660000,0.221602)(0.670000,0.226924)(0.680000,0.232402)(0.690000,0.238039)(0.700000,0.243834)(0.710000,0.249789)(0.720000,0.255905)(0.730000,0.262182)(0.740000,0.268620)(0.750000,0.275218)(0.760000,0.281977)(0.770000,0.288895)(0.780000,0.295972)(0.790000,0.303204)(0.800000,0.310590)(0.810000,0.318129)(0.820000,0.325815)(0.830000,0.333647)(0.840000,0.341621)(0.850000,0.349732)(0.860000,0.357975)(0.870000,0.366345)(0.880000,0.374838)(0.890000,0.383447)(0.900000,0.392165)(0.910000,0.400988)(0.920000,0.409907)(0.930000,0.418917)(0.940000,0.428010)(0.950000,0.437180)(0.960000,0.446419)(0.970000,0.455720)(0.980000,0.465078)(0.990000,0.474484)(1.000000,0.483934)
};
\addplot[black,only marks,mark=o,mark options={scale=0.75},error bars/.cd,y dir=both,y explicit, error bar style={mark size=1.5pt}] plot coordinates{
(0.050000,0.123520)+-(0.001961,0.001961)(0.100000,0.116258)+-(0.003004,0.003004)(0.150000,0.112500)+-(0.003619,0.003619)(0.200000,0.111525)+-(0.003995,0.003995)(0.250000,0.113046)+-(0.004201,0.004201)(0.300000,0.116926)+-(0.004334,0.004334)(0.350000,0.122988)+-(0.004459,0.004459)(0.400000,0.131404)+-(0.004595,0.004595)(0.450000,0.142259)+-(0.004695,0.004695)(0.500000,0.155536)+-(0.005049,0.005049)(0.550000,0.171997)+-(0.005367,0.005367)(0.600000,0.191571)+-(0.005818,0.005818)(0.650000,0.214513)+-(0.006541,0.006541)(0.700000,0.241244)+-(0.007450,0.007450)(0.750000,0.271917)+-(0.008836,0.008836)(0.800000,0.306142)+-(0.010523,0.010523)(0.850000,0.344681)+-(0.012411,0.012411)(0.900000,0.386030)+-(0.015349,0.015349)(0.950000,0.429722)+-(0.018687,0.018687)
};
\addplot[black,only marks,mark=*,mark options={scale=0.75}] plot coordinates{
(0.050000,0.123990)(0.100000,0.115930)(0.150000,0.111060)(0.200000,0.109690)(0.250000,0.111230)(0.300000,0.114650)(0.350000,0.120460)(0.400000,0.129600)(0.450000,0.141420)(0.500000,0.154690)(0.550000,0.174360)(0.600000,0.194010)(0.650000,0.218770)(0.700000,0.246950)(0.750000,0.279720)(0.800000,0.318310)(0.850000,0.355070)(0.900000,0.398220)(0.950000,0.447030)(1.000000,0.492040)
};
\addplot[black] plot coordinates{
(0.010000,0.010602)(0.020000,0.010148)(0.030000,0.009741)(0.040000,0.009377)(0.050000,0.009051)(0.060000,0.008760)(0.070000,0.008501)(0.080000,0.008271)(0.090000,0.008068)(0.100000,0.007890)(0.110000,0.007735)(0.120000,0.007601)(0.130000,0.007487)(0.140000,0.007392)(0.150000,0.007314)(0.160000,0.007254)(0.170000,0.007210)(0.180000,0.007182)(0.190000,0.007168)(0.200000,0.007170)(0.210000,0.007186)(0.220000,0.007216)(0.230000,0.007260)(0.240000,0.007318)(0.250000,0.007390)(0.260000,0.007476)(0.270000,0.007577)(0.280000,0.007692)(0.290000,0.007823)(0.300000,0.007968)(0.310000,0.008129)(0.320000,0.008306)(0.330000,0.008500)(0.340000,0.008712)(0.350000,0.008942)(0.360000,0.009190)(0.370000,0.009459)(0.380000,0.009748)(0.390000,0.010059)(0.400000,0.010394)(0.410000,0.010752)(0.420000,0.011136)(0.430000,0.011547)(0.440000,0.011987)(0.450000,0.012457)(0.460000,0.012959)(0.470000,0.013494)(0.480000,0.014065)(0.490000,0.014675)(0.500000,0.015324)(0.510000,0.016017)(0.520000,0.016754)(0.530000,0.017540)(0.540000,0.018376)(0.550000,0.019267)(0.560000,0.020214)(0.570000,0.021223)(0.580000,0.022295)(0.590000,0.023436)(0.600000,0.024649)(0.610000,0.025938)(0.620000,0.027308)(0.630000,0.028764)(0.640000,0.030310)(0.650000,0.031951)(0.660000,0.033693)(0.670000,0.035541)(0.680000,0.037501)(0.690000,0.039578)(0.700000,0.041779)(0.710000,0.044110)(0.720000,0.046578)(0.730000,0.049188)(0.740000,0.051947)(0.750000,0.054862)(0.760000,0.057940)(0.770000,0.061188)(0.780000,0.064612)(0.790000,0.068219)(0.800000,0.072015)(0.810000,0.076007)(0.820000,0.080202)(0.830000,0.084605)(0.840000,0.089223)(0.850000,0.094060)(0.860000,0.099121)(0.870000,0.104413)(0.880000,0.109938)(0.890000,0.115701)(0.900000,0.121704)(0.910000,0.127951)(0.920000,0.134445)(0.930000,0.141185)(0.940000,0.148174)(0.950000,0.155412)(0.960000,0.162900)(0.970000,0.170636)(0.980000,0.178621)(0.990000,0.186853)(1.000000,0.195330)
};
\addplot[black,only marks,mark=o,mark options={scale=0.75},error bars/.cd,y dir=both,y explicit, error bar style={mark size=1.5pt}] plot coordinates{
(0.050000,0.009048)+-(0.000322,0.000322)(0.100000,0.007900)+-(0.000456,0.000456)(0.150000,0.007341)+-(0.000528,0.000528)(0.200000,0.007201)+-(0.000576,0.000576)(0.250000,0.007425)+-(0.000617,0.000617)(0.300000,0.008010)+-(0.000663,0.000663)(0.350000,0.008974)+-(0.000728,0.000728)(0.400000,0.010414)+-(0.000816,0.000816)(0.450000,0.012448)+-(0.000921,0.000921)(0.500000,0.015215)+-(0.001107,0.001107)(0.550000,0.019077)+-(0.001335,0.001335)(0.600000,0.024311)+-(0.001655,0.001655)(0.650000,0.031357)+-(0.002145,0.002145)(0.700000,0.040842)+-(0.002831,0.002831)(0.750000,0.053472)+-(0.003899,0.003899)(0.800000,0.069831)+-(0.005378,0.005378)(0.850000,0.091197)+-(0.007359,0.007359)(0.900000,0.117723)+-(0.010475,0.010475)(0.950000,0.149908)+-(0.014589,0.014589)
};
\addplot[black,only marks,mark=*,mark options={scale=0.75}] plot coordinates{
(0.050000,0.008520)(0.100000,0.007550)(0.150000,0.006520)(0.200000,0.006560)(0.250000,0.006420)(0.300000,0.006850)(0.350000,0.008040)(0.400000,0.009110)(0.450000,0.011040)(0.500000,0.014580)(0.550000,0.017780)(0.600000,0.023820)(0.650000,0.030230)(0.700000,0.041990)(0.750000,0.054280)(0.800000,0.071530)(0.850000,0.092110)(0.900000,0.121480)(0.950000,0.157840)(1.000000,0.197780)
};
\node at (axis cs:0.4,0.3) {$\gamma=2$};
\draw (axis cs:0.4,0.12) ellipse [black,x radius=2,y radius=0.5];
\node at (axis cs:0.4,0.025) {$\gamma=3$};
\draw (axis cs:0.4,0.01) ellipse [black,x radius=2,y radius=0.5];
\legend{ {Limiting theory},{Empirical estimator},{Detector} }
\end{semilogyaxis}
\end{tikzpicture}
\end{tabular}
\caption{False alarm rate $P(\sqrt{N}T_N(\rho)>\gamma)$, $N=20$ (left), $N=100$ (right), $p=N^{-\frac12}[1,\ldots,1]^\trans$, $[C_N]_{ij}=0.7^{|i-j|}$, $c_N=1/2$.}
\label{fig:FAR}
\end{figure}
\begin{figure}[h!]
\centering
\begin{tikzpicture}[font=\footnotesize]
\renewcommand{\axisdefaulttryminticks}{4}
\tikzstyle{every major grid}+=[style=densely dashed]
\tikzstyle{every axis y label}+=[yshift=-10pt]
\tikzstyle{every axis x label}+=[yshift=5pt]
\tikzstyle{every axis legend}+=[cells={anchor=west},fill=white,
at={(0.98,0.98)}, anchor=north east, font=\scriptsize ]
\begin{semilogyaxis}[
xmin=0,
ymin=1e-4,
xmax=1,
ymax=1,
grid=major,
ymajorgrids=false,
scaled ticks=true,
xlabel={$\Gamma$},
mark repeat=2;
ylabel={$P(T_N(\rho)>\Gamma)$}
]
\addplot[black] plot coordinates{
(0.,1.)(0.01,0.998896)(0.02,0.995589)(0.03,0.990103)(0.04,0.982474)(0.05,0.97275)(0.06,0.960997)(0.07,0.94729)(0.08,0.931716)(0.09,0.914376)(0.1,0.895377)(0.11,0.874837)(0.12,0.852881)(0.13,0.82964)(0.14,0.805251)(0.15,0.779853)(0.16,0.753589)(0.17,0.726602)(0.18,0.699035)(0.19,0.671028)(0.2,0.642722)(0.21,0.614251)(0.22,0.585744)(0.23,0.557328)(0.24,0.529119)(0.25,0.501229)(0.26,0.473761)(0.27,0.446809)(0.28,0.420461)(0.29,0.394792)(0.3,0.369873)(0.31,0.345761)(0.32,0.322507)(0.33,0.300153)(0.34,0.278732)(0.35,0.258268)(0.36,0.238778)(0.37,0.220272)(0.38,0.202751)(0.39,0.186212)(0.4,0.170645)(0.41,0.156033)(0.42,0.142358)(0.43,0.129595)(0.44,0.117715)(0.45,0.106688)(0.46,0.096481)(0.47,0.087058)(0.48,0.078381)(0.49,0.070414)(0.5,0.063117)(0.51,0.056451)(0.52,0.050377)(0.53,0.044858)(0.54,0.039856)(0.55,0.035333)(0.56,0.031254)(0.57,0.027585)(0.58,0.024293)(0.59,0.021346)(0.6,0.018716)(0.61,0.016373)(0.62,0.014292)(0.63,0.012448)(0.64,0.010818)(0.65,0.009381)(0.66,0.008117)(0.67,0.007007)(0.68,0.006036)(0.69,0.005188)(0.7,0.004449)(0.71,0.003807)(0.72,0.003251)(0.73,0.002769)(0.74,0.002354)(0.75,0.001997)(0.76,0.00169)(0.77,0.001427)(0.78,0.001202)(0.79,0.001011)(0.8,0.000848)(0.81,0.00071)(0.82,0.000593)(0.83,0.000494)(0.84,0.000411)(0.85,0.000341)(0.86,0.000282)(0.87,0.000233)(0.88,0.000192)(0.89,0.000158)(0.9,0.00013)(0.91,0.000106)(0.92,0.000087)(0.93,0.000071)(0.94,0.000057)(0.95,0.000047)(0.96,0.000038)(0.97,0.00003)(0.98,0.000025)(0.99,0.00002)(1.,0.000016)
};
\addplot[black,only marks,mark=*,mark options={scale=0.75}] plot coordinates{
(0.,1.)(0.01,0.99896)(0.02,0.995809)(0.03,0.990564)(0.04,0.983483)(0.05,0.974368)(0.06,0.963289)(0.07,0.950217)(0.08,0.935453)(0.09,0.919055)(0.1,0.900819)(0.11,0.881349)(0.12,0.860069)(0.13,0.837701)(0.14,0.813929)(0.15,0.789225)(0.16,0.763256)(0.17,0.736593)(0.18,0.709161)(0.19,0.681532)(0.2,0.652937)(0.21,0.624299)(0.22,0.595142)(0.23,0.565768)(0.24,0.537037)(0.25,0.508517)(0.26,0.479995)(0.27,0.45221)(0.28,0.424418)(0.29,0.397055)(0.3,0.370642)(0.31,0.344963)(0.32,0.320174)(0.33,0.296156)(0.34,0.273264)(0.35,0.251097)(0.36,0.230089)(0.37,0.210095)(0.38,0.191335)(0.39,0.173287)(0.4,0.156354)(0.41,0.140724)(0.42,0.126194)(0.43,0.112588)(0.44,0.099988)(0.45,0.088543)(0.46,0.077981)(0.47,0.068355)(0.48,0.059736)(0.49,0.051963)(0.5,0.044954)(0.51,0.038702)(0.52,0.033099)(0.53,0.028047)(0.54,0.023645)(0.55,0.019754)(0.56,0.016417)(0.57,0.013608)(0.58,0.011183)(0.59,0.009115)(0.6,0.00741)(0.61,0.005953)(0.62,0.004697)(0.63,0.003682)(0.64,0.002875)(0.65,0.002209)(0.66,0.001709)(0.67,0.001276)(0.68,0.000938)(0.69,0.000692)(0.7,0.000487)(0.71,0.000338)(0.72,0.000227)(0.73,0.000163)(0.74,0.000123)(0.75,0.000076)(0.76,0.00005)(0.77,0.00003)(0.78,0.000014)(0.79,0.000006)(0.8,0.000002)(0.81,0.000001)(0.82,0.)(0.83,0.)(0.84,0.)(0.85,0.)(0.86,0.)(0.87,0.)(0.88,0.)(0.89,0.)(0.9,0.)(0.91,0.)(0.92,0.)(0.93,0.)(0.94,0.)(0.95,0.)(0.96,0.)(0.97,0.)(0.98,0.)(0.99,0.)(1.,0.)
};
\addplot[black] plot coordinates{
(0.,1.)(0.01,0.994528)(0.02,0.978292)(0.03,0.951819)(0.04,0.915955)(0.05,0.871823)(0.06,0.820761)(0.07,0.764257)(0.08,0.703876)(0.09,0.641191)(0.1,0.577714)(0.11,0.51484)(0.12,0.453802)(0.13,0.395635)(0.14,0.341159)(0.15,0.290974)(0.16,0.245462)(0.17,0.20481)(0.18,0.169025)(0.19,0.13797)(0.2,0.111391)(0.21,0.088952)(0.22,0.070257)(0.23,0.054886)(0.24,0.04241)(0.25,0.032412)(0.26,0.024501)(0.27,0.018318)(0.28,0.013547)(0.29,0.009908)(0.3,0.007168)(0.31,0.005129)(0.32,0.00363)(0.33,0.002541)(0.34,0.00176)(0.35,0.001205)(0.36,0.000816)(0.37,0.000547)(0.38,0.000362)(0.39,0.000237)(0.4,0.000154)(0.41,0.000099)(0.42,0.000063)(0.43,0.000039)(0.44,0.000024)(0.45,0.000015)(0.46,0.000009)(0.47,0.000005)(0.48,0.000003)(0.49,0.000002)(0.5,0.000001)(0.51,0.000001)(0.52,0.)(0.53,0.)(0.54,0.)(0.55,0.)(0.56,0.)(0.57,0.)(0.58,0.)(0.59,0.)(0.6,0.)(0.61,0.)(0.62,0.)(0.63,0.)(0.64,0.)(0.65,0.)(0.66,0.)(0.67,0.)(0.68,0.)(0.69,0.)(0.7,0.)(0.71,0.)(0.72,0.)(0.73,0.)(0.74,0.)(0.75,0.)(0.76,0.)(0.77,0.)(0.78,0.)(0.79,0.)(0.8,0.)(0.81,0.)(0.82,0.)(0.83,0.)(0.84,0.)(0.85,0.)(0.86,0.)(0.87,0.)(0.88,0.)(0.89,0.)(0.9,0.)(0.91,0.)(0.92,0.)(0.93,0.)(0.94,0.)(0.95,0.)(0.96,0.)(0.97,0.)(0.98,0.)(0.99,0.)(1.,0.)
};
\addplot[black,only marks,mark=*,mark options={scale=0.75}] plot coordinates{
(0.000000,1.000000)(0.010000,0.994368)(0.020000,0.978268)(0.030000,0.952258)(0.040000,0.917284)(0.050000,0.872770)(0.060000,0.821488)(0.070000,0.766126)(0.080000,0.706967)(0.090000,0.643454)(0.100000,0.580648)(0.110000,0.517506)(0.120000,0.455490)(0.130000,0.396165)(0.140000,0.341170)(0.150000,0.290833)(0.160000,0.244757)(0.170000,0.202599)(0.180000,0.166316)(0.190000,0.135318)(0.200000,0.108255)(0.210000,0.085529)(0.220000,0.066516)(0.230000,0.051258)(0.240000,0.039086)(0.250000,0.029266)(0.260000,0.021645)(0.270000,0.015920)(0.280000,0.011507)(0.290000,0.008071)(0.300000,0.005601)(0.310000,0.003907)(0.320000,0.002626)(0.330000,0.001749)(0.340000,0.001136)(0.350000,0.000720)(0.360000,0.000436)(0.370000,0.000256)(0.380000,0.000152)(0.390000,0.000100)(0.400000,0.000045)(0.410000,0.000024)(0.420000,0.000014)(0.430000,0.000010)(0.440000,0.000003)(0.450000,0.000003)(0.460000,0.000000)(0.470000,0.000000)(0.480000,0.000000)(0.490000,0.000000)(0.500000,0.000000)(0.510000,0.000000)(0.520000,0.000000)(0.530000,0.000000)(0.540000,0.000000)(0.550000,0.000000)(0.560000,0.000000)(0.570000,0.000000)(0.580000,0.000000)(0.590000,0.000000)(0.600000,0.000000)(0.610000,0.000000)(0.620000,0.000000)(0.630000,0.000000)(0.640000,0.000000)(0.650000,0.000000)(0.660000,0.000000)(0.670000,0.000000)(0.680000,0.000000)(0.690000,0.000000)(0.700000,0.000000)(0.710000,0.000000)(0.720000,0.000000)(0.730000,0.000000)(0.740000,0.000000)(0.750000,0.000000)(0.760000,0.000000)(0.770000,0.000000)(0.780000,0.000000)(0.790000,0.000000)(0.800000,0.000000)(0.810000,0.000000)(0.820000,0.000000)(0.830000,0.000000)(0.840000,0.000000)(0.850000,0.000000)(0.860000,0.000000)(0.870000,0.000000)(0.880000,0.000000)(0.890000,0.000000)(0.900000,0.000000)(0.910000,0.000000)(0.920000,0.000000)(0.930000,0.000000)(0.940000,0.000000)(0.950000,0.000000)(0.960000,0.000000)(0.970000,0.000000)(0.980000,0.000000)(0.990000,0.000000)(1.000000,0.000000)
};
\node at (axis cs:0.2,0.004) {$N=100$};
\draw (axis cs:0.32,0.004) ellipse [black,x radius=2,y radius=0.5];
\node at (axis cs:0.75,0.01) {$N=20$};
\draw (axis cs:0.65,0.004) ellipse [black,x radius=2,y radius=1.5];
\legend{ {Limiting theory},{Detector}}
\end{semilogyaxis}
\end{tikzpicture}
\caption{False alarm rate $P(T_N(\rho_N^*)>\Gamma)$ for $N=20$ and $N=100$, $p=N^{-\frac12}[1,\ldots,1]^\trans$, $[C_N]_{ij}=0.7^{|i-j|}$, $c_N=1/2$.}
\label{fig:FAR2}
\end{figure}
\section{Proof}
\label{sec:proof}
In this section, we shall successively prove Theorem~\ref{th:bilin}, Theorem~\ref{th:T}, Proposition~\ref{prop:1}, and Corollary~\ref{co:1}. Of utmost interest is the proof of Theorem~\ref{th:bilin} which shall be the concern of most of the section and of Appendix~\ref{app:key_lemma} for the proof of a key lemma.
Before delving into the core of the proofs, let us introduce a few notations that shall be used throughout the section.
First recall from \citep{COU14} that we can write, for each $\rho\in(\max\{0,1-c_N^{-1}\},1]$,
\begin{align*}
\hat{C}_N(\rho)=\frac{1-\rho}{1-(1-\rho)c_N} \frac1n\sum_{i=1}^n \frac{z_iz_i^*}{\frac1Nz_i^*\hat{C}_{(i)}^{-1}(\rho)z_i} + \rho I_N
\end{align*}
where $\hat{C}_{(i)}(\rho)=\hat{C}_N(\rho)-(1-\rho)\frac1n\frac{z_iz_i^*}{\frac1Nz_i^*\hat{C}_N^{-1}(\rho)z_i}$.
Now, we define
\begin{align*}
\alpha(\rho) &= \frac{1-\rho}{1-(1-\rho)c_N} \\
d_i(\rho) &= \frac1Nz_i^* \hat{C}_{(i)}^{-1}(\rho)z_i = \frac1Nz_i^* \left( \alpha(\rho) \frac1n\sum_{j\neq i} \frac{z_jz_j^*}{d_j(\rho)} + \rho I_N \right)^{-1}z_i \\
\tilde d_i(\rho) &= \frac1Nz_i^* \hat{S}_{(i)}^{-1}(\rho)z_i = \frac1Nz_i^* \left( \alpha(\rho) \frac1n\sum_{j\neq i}^n \frac{z_jz_j^*}{\gamma_N(\rho)} + \rho I_N \right)^{-1}z_
\end{align*}
Clearly by uniqueness of $\hat{C}_N$ and by the relation to $\hat{C}_{(i)}$ above, $d_1(\rho),\ldots,d_n(\rho)$ are uniquely defined by their $n$ implicit equations.
We shall also discard the parameter $\rho$ for readability whenever not needed.
\subsection{Bilinear form equivalence}
In this section, we prove Theorem~\ref{th:bilin}. As shall become clear, the proof unfolds similarly for each $k\in\ZZ\setminus\{0\}$ and we can therefore restrict ourselves to a single value for $k$. As Theorem~\ref{th:T} relies on $k=-1$, for consistency, we take $k=-1$ from now on. Thus, our objective is to prove that, for $a,b\in\CC^N$ with $\Vert a\Vert=\Vert b\Vert=1$, and for any $\varepsilon>0$,
\begin{align*}
\sup_{\rho\in \mathcal R_\kappa} N^{1-\varepsilon}\left| a^*\hat{C}_N^{-1}(\rho)b - a^*\hat{S}_N^{-1}(\rho)b \right| \asto 0.
\end{align*}
For this, forgetting for some time the index $\rho$, first write
\begin{align}
\label{eq:firsteq}
a^*\hat{C}_N^{-1}b - a^*\hat{S}_N^{-1}b &= a^*\hat{C}_N^{-1} \left( \frac{\alpha}n \sum_{i=1}^n \left[ \frac1{\gamma_N} - \frac1{d_i} \right] z_iz_i^* \right) \hat{S}_N^{-1}b \\
\label{eq:secondeq}
&= \frac{\alpha}n \sum_{i=1}^n a^*\hat{C}_N^{-1} z_i\frac{d_i-\gamma_N}{\gamma_Nd_i}z_i^*\hat{S}_N^{-1}b.
\end{align}
In \citep{COU14}, where it is shown that $\Vert \hat{C}_N-\hat{S}_N\Vert\asto 0$ (that is the spectral norm of the inner parenthesis in \eqref{eq:firsteq} vanishes), the core of the proof was to show that $\max_{1\leq i\leq n}|d_i-\gamma_N|\asto 0$ which, along with the convergence of $\gamma_N$ away from zero and the almost sure boundedness of $\Vert\frac1n\sum_{i=1}^nz_iz_i^*\Vert$ for all large $N$ (from e.g.\@ \citep{SIL98}), gives the result. A thorough inspection of the proof in \citep{COU14} reveals that $\max_{1\leq i\leq n}|d_i-\gamma_N|\asto 0$ may be improved into $\max_{1\leq i\leq n}N^{\frac12-\varepsilon}|d_i-\gamma_N|\asto 0$ for any $\varepsilon>0$ but that this speed cannot be further improved beyond $N^{\frac12}$. The latter statement is rather intuitive since $\gamma_N$ is essentially a sharp deterministic approximation for $\frac1N\tr \hat{C}_N^{-1}$ while $d_i$ is a quadratic form on $\hat{C}_{(i)}^{-1}$; classical random matrix results involving fluctuations of such quadratic forms, see e.g.\@ \citep{KAM09}, indeed show that these fluctuations are of order $N^{-\frac12}$. As a consequence, $\max_{1\leq i\leq n}N^{1-\varepsilon}|d_i-\gamma_N|$ and thus $N^{1-\varepsilon}\Vert \hat{C}_N-\hat{S}_N\Vert$ are not expected to vanish for small $\varepsilon$.
This being said, when it comes to bilinear forms, for which we shall naturally have $N^{\frac12-\varepsilon}|a^*\hat{C}_N^{-1}b - a^*\hat{S}_N^{-1}b|\asto 0$, seeing the difference in absolute values as the $n$-term average \eqref{eq:secondeq}, one may expect that the fluctuations of $d_i-\gamma_N$ are sufficiently loosely dependent across $i$ to further increase the speed of convergence from $N^{\frac12-\varepsilon}$ to $N^{1-\varepsilon}$ (which is the best one could expect from a law of large numbers aspect if the $d_i-\gamma_N$ were truly independent). It turns out that this intuition is correct.
Nonetheless, to proceed with the proof, it shall be quite involved to work directly with \eqref{eq:secondeq} which involves the rather intractable terms $d_i$ (as the random solutions to an implicit equation). As in \citep{COU14}, our approach will consist in first approximating $d_i$ by a much more tractable quantity. Letting $\gamma_N$ be this approximation is however not good enough this time since $\gamma_N-d_i$ is a non-obvious quantity of amplitude $O(N^{-\frac12})$ which, due to intractability, we shall not be able to average across $i$ into a $O(N^{-1})$ quantity. Thus, we need a refined approximation of $d_i$ which we shall take to be $\tilde{d}_i$ defined above. Intuitively, since $\tilde{d}_i$ is also a quadratic form closely related to $d_i$, we expect $d_i-\tilde{d}_i$ to be of order $O(N^{-1})$, which we shall indeed observe. With this approximation in place, $d_i$ can be replaced by $\tilde{d}_i$ in \eqref{eq:secondeq}, which now becomes a more tractable random variable (as it involves no implicit equation) that fluctuates around $\gamma_N$ at the expected $O(N^{-1})$ speed.
Let us then introduce the variable $\tilde{d}_i$ in \eqref{eq:firsteq} to obtain
\begin{align*}
a^*\hat{C}_N^{-1}b - a^*\hat{S}_N^{-1}b
&= a^*\hat{C}_N^{-1} \left( \frac{\alpha}n \sum_{i=1}^n \left[ \frac1{\gamma_N} - \frac1{\tilde{d}_i} \right] z_iz_i^* \right) \hat{S}_N^{-1}b \nonumber \\
&+ a^*\hat{C}_N^{-1} \left( \frac{\alpha}n \sum_{i=1}^n \left[ \frac1{\tilde{d}_i} - \frac1{d_i} \right] z_iz_i^* \right) \hat{S}_N^{-1}b \\
&\triangleq \xi_1+\xi_2.
\end{align*}
We will now show that $\xi_1=\xi_1(\rho)$ and $\xi_2=\xi_2(\rho)$ vanish at the appropriate speed and uniformly so on $\mathcal R_\kappa$.
Let us first progress in the derivation of $\xi_1(\rho)$ from which we wish to discard the explicit dependence on $\hat{C}_N$. We have
\begin{align*}
\xi_1 &= a^*\hat{C}_N^{-1} \left( \frac{\alpha}n \sum_{i=1}^n \left[ \frac1{\gamma_N} - \frac1{\tilde{d}_i} \right] z_iz_i^* \right) \hat{S}_N^{-1}b \\
&= a^*\hat{S}_N^{-1} \left( \frac{\alpha}n \sum_{i=1}^n \left[ \frac1{\gamma_N} - \frac1{\tilde{d}_i} \right] z_iz_i^* \right) \hat{S}_N^{-1}b \nonumber \\
&+ a^*(\hat{C}_N^{-1}-\hat{S}_N^{-1}) \left( \frac{\alpha}n \sum_{i=1}^n \left[ \frac1{\gamma_N} - \frac1{\tilde{d}_i} \right] z_iz_i^* \right) \hat{S}_N^{-1}b \\
&= a^*\hat{S}_N^{-1} \left( \frac{\alpha}n \sum_{i=1}^n \frac{\tilde{d}_i-\gamma_N}{\gamma_N^2} z_iz_i^* \right) \hat{S}_N^{-1}b \nonumber \\
&- a^*\hat{S}_N^{-1} \left( \frac{\alpha}n \sum_{i=1}^n \frac{(\tilde{d}_i-\gamma_N)^2}{\gamma_N^2\tilde{d}_i} z_iz_i^* \right) \hat{S}_N^{-1}b \nonumber \\
&+ a^*(\hat{C}_N^{-1}-\hat{S}_N^{-1}) \left( \frac{\alpha}n \sum_{i=1}^n \left[ \frac1{\gamma_N} - \frac1{\tilde{d}_i} \right] z_iz_i^* \right) \hat{S}_N^{-1}b \\
&\triangleq \xi_{11} + \xi_{12} + \xi_{13}.
\end{align*}
The terms $\xi_{12}$ and $\xi_{13}$ exhibit products of two terms that are expected to be of order $O(N^{-\frac12})$ and which are thus easily handled. As for $\xi_{11}$, it no longer depends on $\hat{C}_N$ and is therefore a standard random variable which, although involved, is technically tractable via standard random matrix methods. In order to show that $N^{1-\varepsilon}\max\{|\xi_{12}|,|\xi_{13}|\}\asto 0$ uniformly in $\rho$, we use the following lemma.
\begin{lemma}
\label{le:1}
For any $\varepsilon>0$,
\begin{align*}
\max_{1\leq i\leq n}\sup_{\rho\in\mathcal R_\kappa}N^{\frac12-\varepsilon}|\tilde{d}_i(\rho)-\gamma_N(\rho)| &\asto 0 \\
\max_{1\leq i\leq n}\sup_{\rho\in\mathcal R_\kappa}N^{\frac12-\varepsilon}|d_i(\rho)-\gamma_N(\rho)| &\asto 0.
\end{align*}
\end{lemma}
Note that, while the first result is a standard, easily established, random matrix result, the second result is the aforementioned refinement of the core result in the proof of \citep[Theorem~1]{COU14}.
\begin{proof}[Proof of Lemma~\ref{le:1}]
We start by proving the first identity.
From \cite[p.~17]{COU14} (taking $w=-\gamma_N\rho \alpha^{-1}$), we have, for each $p\geq 2$ and for each $1\leq k\leq n$,
\begin{align*}
\EE\left[ \left| \tilde{d}_k(\rho) - \gamma_N(\rho) \right|^p\right] &= O(N^{-\frac{p}2})
\end{align*}
where the bound does not depend on $\rho>\max\{0,1-1/c\}+\kappa$.
Let now $\max\{0,1-1/c\}+\kappa=\rho_0<\ldots<\rho_{\lceil\sqrt{n}\rceil}=1$ be a regular sampling of $\mathcal R_\kappa$ in $\lceil\sqrt{n}\rceil$ intervals. We then have, from Markov inequality and the union bound on $n(\lceil\sqrt{n}\rceil+1)$ events, for $C>0$ given,
\begin{align*}
P \left( \max_{1\leq k\leq n,0\leq i\leq \lceil\sqrt{n}\rceil} \left| N^{\frac12-\varepsilon}( \tilde{d}_k(\rho_i) - \gamma_N(\rho_i)) \right| > C \right) &\leq KN^{-p\varepsilon+\frac32}
\end{align*}
for some $K>0$ only dependent on $p$ and $C$. From the Borel Cantelli lemma, we then have $\max_{k,i} | N^{\frac12-\varepsilon}( \tilde{d}_k(\rho_i) - \gamma_N(\rho_i))|\asto 0$ as long as $-p\varepsilon+3/2<-1$, which is obtained for $p>5/(2\varepsilon)$. Using $|\gamma_N(\rho)-\gamma_N(\rho')|\leq K|\rho-\rho'|$ for some constant $K$ and each $\rho,\rho'\in\mathcal R_\kappa$ (see \citep[top of Section~5.1]{COU14}) and similarly $\max_{1\leq k\leq n}|\tilde{d}_k(\rho)-\tilde{d}_k(\rho')|\leq K|\rho-\rho'|$ for all large $n$ a.s.\@ (obtained by explicitly writing the difference and using the fact that $\Vert z_k\Vert^2/N$ is asymptotically bounded almost surely), we get
\begin{align*}
\max_{1\leq k\leq n}\sup_{\rho\in\mathcal R_\kappa}N^{\frac12-\varepsilon}|\tilde{d}_k(\rho)-\gamma_N(\rho)| &\leq \max_{k,i} N^{\frac12-\varepsilon} | \tilde{d}_k(\rho_i) - \gamma_N(\rho_i) | + KN^{-\varepsilon} \\ &\asto 0.
\end{align*}
The second result relies on revisiting the proof of \cite[Theorem~1]{COU14} incorporating the convergence speed on $\tilde{d}_k-\gamma_N$.
For convenience and compatibility with similar derivations that appear later in the proof, we slightly modify the original proof of \cite[Theorem~1]{COU14}.
We first define $f_i(\rho)=d_i(\rho)/\gamma_N(\rho)$ and relabel the $d_i(\rho)$ in such a way that $f_1(\rho)\leq \ldots\leq f_n(\rho)$ (the ordering may then depend on $\rho$). Then, we have by definition of $d_n(\rho)=\gamma_N(\rho) f_n(\rho)$
\begin{align*}
\gamma_N(\rho) f_n(\rho) &= \frac1Nz_n^*\left( \alpha(\rho) \frac1n\sum_{i<n} \frac{z_iz_i^*}{\gamma_N(\rho) f_i(\rho)} + \rho I_N \right)^{-1}z_n \\
&\leq \frac1Nz_n^*\left( \alpha(\rho) \frac1{f_n(\rho)} \frac1n\sum_{i<n} \frac{z_iz_i^*}{\gamma_N(\rho)} + \rho I_N \right)^{-1}z_n
\end{align*}
where we used $f_n(\rho)\geq f_i(\rho)$ for each $i$. The above is now equivalent to
\begin{align*}
\gamma_N(\rho) &\leq \frac1Nz_n^* \left( \alpha(\rho) \frac1n\sum_{i<n} \frac{z_iz_i^*}{\gamma_N(\rho)} + f_n(\rho) \rho I_N \right)^{-1}z_n.
\end{align*}
We now make the assumption that there exists $\eta>0$ and a sequence $\{\rho^{(n)}\}\in\mathcal R_\kappa$ such that $f_n(\rho^{(n)})>1+N^{\eta-\frac12}$ infinitely often, which is equivalent to saying $d_n(\rho^{(n)})>\gamma_N(\rho^{(n)})(1+N^{\eta-\frac12})$ infinitely often (i.o.). Then, from these assumptions and the above first convergence result
\begin{align}
\label{eq:gammaineq}
\gamma_N(\rho^{(n)}) &\leq \frac1Nz_n^*\left( \alpha(\rho^{(n)}) \frac1n\sum_{i<n} \frac{z_iz_i^*}{\gamma_N(\rho^{(n)})} + \rho^{(n)} (1+N^{\eta-\frac12}) I_N \right)^{-1}z_n \nonumber \\
&= \tilde{d}_n(\rho^{(n)}) - N^{\eta-\frac12} \frac1Nz_n^*\left( \frac1n\sum_{i<n} \frac{\alpha(\rho^{(n)}) z_iz_i^*}{\rho^{(n)}\gamma_N(\rho^{(n)})} + (1+N^{\eta-\frac12}) I_N \right)^{-1} \nonumber \\
&\times\left( \frac1n\sum_{i<n} \frac{\alpha(\rho^{(n)}) z_iz_i^*}{\gamma_N(\rho^{(n)})} + \rho^{(n)} I_N \right)^{-1}z_n.
\end{align}
Now, by the first result of the lemma, letting $0<\varepsilon<\eta$, we have
\begin{align*}
\left| \tilde{d}_n(\rho^{(n)}) - \gamma_N(\rho^{(n)}) \right| &\leq \max_{\rho\in\mathcal R_\kappa} \left| \tilde{d}_n(\rho) - \gamma_N(\rho) \right| \leq N^{\varepsilon-\frac12}
\end{align*}
for all large $n$ a.s., so that, for these large $n$, $\tilde{d}_n(\rho^{(n)}) \leq \gamma_N(\rho^{(n)}) + N^{\varepsilon-\frac12}$. Applying this inequality to the first right-end side term of \eqref{eq:gammaineq} and using the almost sure boundedness of the rightmost right-end side term entails
\begin{align*}
0 \leq N^{\varepsilon-\frac12} - KN^{\eta-\frac12}
\end{align*}
for some $K>0$ for all large $n$ a.s. But, $N^{\varepsilon/2-1/2} - KN^{\eta/2-1/2}<0$ for all large $N$, which contradicts the inequality. Thus, our initial assumption is wrong and therefore, for each $\eta>0$, we have for all large $n$ a.s., $d_n(\rho)<\gamma_N(\rho)+N^{\eta-\frac12}$ uniformly on $\rho\in\mathcal R_\kappa$. The same calculus can be performed for $d_1(\rho)$ by assuming that $f_1(\rho^{\prime(n)})<1-N^{\eta-\frac12}$ i.o.\@ over some sequence $\rho^{\prime(n)}$; by reverting all inequalities in the derivation above, we similarly conclude by contradiction that $d_1(\rho)>\gamma_N(\rho)-N^{\eta-\frac12}$ for all large $n$, uniformly so in $\mathcal R_\kappa$. Together, both results finally lead, for each $\varepsilon>0$, to
\begin{align*}
\max_{1\leq k\leq n} \sup_{\rho\in\mathcal R_\kappa} \left| N^{\frac12-\varepsilon} \left( d_k(\rho) - \gamma_N(\rho) \right) \right| &\asto 0
\end{align*}
obtained by fixing $\varepsilon$, taking $\eta$ such that $0<\eta<\varepsilon$, and using $\max_k \sup_{\rho} |d_k(\rho)-\gamma_N(\rho)|<N^{\eta-\frac12}$ for all large $n$ a.s.
\end{proof}
Thanks to Lemma~\ref{le:1}, expressing $\hat{C}_N^{-1}(\rho)-\hat{S}_N^{-1}(\rho)$ as a function of $d_i(\rho)-\gamma_N(\rho)$ and using the (almost sure) boundedness of the various terms involved, we finally get $N^{1-\varepsilon}\xi_{12}\asto 0$ and $N^{1-\varepsilon}\xi_{13}\asto 0$ uniformly on $\rho$.
\bigskip
It then remains to handle the more delicate term $\xi_{11}$, which can be further expressed as
\begin{align*}
\xi_{11} &= \frac{\alpha}{\gamma_N^2}a^*\hat{S}_N^{-1}\left( \frac1n \sum_{i=1}^n (\tilde{d}_i-\gamma_N) z_iz_i^* \right)\hat{S}_N^{-1}b \nonumber \\
&= \frac{\alpha}{\gamma_N^2} \frac1n\sum_{i=1}^n a^*\hat{S}_N^{-1}z_iz_i^*\hat{S}_N^{-1}b \left( \tilde{d}_i - \gamma_N \right).
\end{align*}
For that, we will resort to the following lemma, whose proof is postponed to Appendix~\ref{app:key_lemma}.
\begin{lemma}
Let $\first$ and $\second$ be random or deterministic vectors, independent of $z_1,\cdots,z_n$, such that $\max\left(\EE[\|\first\|^{k}], \EE[\|\second\|^k]\right) \leq K$ for some $K>0$ and all integer $k$. Then, for each integer $p$,
\begin{align*}
\EE\left[\left|\frac{1}{n}\sum_{i=1}^n \first^*\SN z_iz_i^* \SN \second \left(\frac{1}{N}z_i^* \Si z_i-\gammanrho\right) \right|^{2p}\right] =O\left(N^{-2p}\right)
\end{align*}
\label{le:keylemma1}
\end{lemma}
By the Markov inequality and the union bound, similar to the proof of Lemma~\ref{le:1}, we get from Lemma~\ref{le:keylemma1} (with $a=c$ and $d=b$) that, for each $\eta>0$ and for each integer $p\geq 1$,
\begin{align*}
P \left( \sup_{\rho\in \{\rho_0<\ldots<\rho_{\lceil\sqrt{n}\rceil}\}} N^{1-\varepsilon} |\xi_{11}| > \eta \right) &\leq K N^{-p\varepsilon+\frac12}
\end{align*}
with $K$ only function of $\eta$ and $\rho_0<\ldots<\rho_{\lceil\sqrt{n}\rceil}$ a regular sampling of $\mathcal R_\kappa$.
Taking $p>3/(2\varepsilon)$, we finally get from the Borel Cantelli lemma that
\begin{align*}
N^{1-\varepsilon} \xi_{11} &\asto 0
\end{align*}
uniformly on $\{\rho_0,\ldots,\rho_{\lceil\sqrt{n}\rceil}\}$ and finally, using Lipschitz arguments as in the proof of Lemma~\ref{le:1}, uniformly on $\mathcal R_\kappa$. Putting all results together, we finally have
\begin{align*}
\sup_{\rho\in\mathcal R_\kappa} N^{1-\varepsilon} |\xi_1(\rho)| &\asto 0
\end{align*}
which concludes the first part of the proof.
\bigskip
We now continue with $\xi_2(\rho)$. In order to prove $N^{1-\varepsilon}\xi_2(\rho)\asto 0$ uniformly on $\rho\in \mathcal R_\kappa$, it is sufficient (thanks to the boundedness of the various terms involved) to prove that
\begin{align*}
\max_{1\leq i\leq n}\sup_{\rho\in\mathcal R_\kappa} \left| N^{1-\varepsilon} \left(\tilde{d}_i(\rho) - d_i(\rho)\right)\right| &\asto 0.
\end{align*}
To obtain this result, we first need the following fundamental proposition.
\begin{proposition}
\label{prop:2}
For any $\varepsilon>0$,
\begin{align*}
\max_{1\leq k\leq n}\sup_{\rho \in\mathcal R_\kappa} \left| N^{1-\varepsilon} \left( \tilde{d}_k(\rho) - \frac1N z_k^*\left( \alpha(\rho) \frac1n \sum_{i\neq k} \frac{z_iz_i^*}{\tilde{d}_i(\rho)} + \rho I_N \right)^{-1}z_k \right) \right| \asto 0.
\end{align*}
\end{proposition}
\begin{proof}
By expanding the definition of $\tilde{d}_k$, first observe that
\begin{align*}
&\tilde{d}_k - \frac1N z_k^*\left( \alpha \frac1n \sum_{i\neq k} \frac{z_iz_i^*}{\tilde{d}_i} + \rho I_N \right)^{-1}z_k \\
&=\alpha \frac1n\sum_{i\neq k} \frac1N z_k^* \hat{S}_{(k)}^{-1} z_iz_i^* \frac{\gamma_N - \tilde{d}_i}{\gamma_N\tilde{d}_i} \left( \alpha \frac1n \sum_{j\neq k} \frac{z_jz_j^*}{\tilde{d}_j} + \rho I_N \right)^{-1}z_k.
\end{align*}
Similar to the derivation of $\xi_1$, we now proceed to approximating $\tilde{d}_i$ in the central denominator and each $\tilde{d}_j$ in the rightmost inverse matrix by the non-random $\gamma_N$. We obtain (from Lemma~\ref{le:1})
\begin{align*}
&\tilde{d}_k - \frac1N z_k^*\left( \alpha \frac1n \sum_{i\neq k} \frac{z_iz_i^*}{\tilde{d}_i} + \rho I_N \right)^{-1}z_k \\
&= \frac{\alpha}{\gamma_N^2} \frac1n\sum_{i\neq k} \frac1N z_k^* \hat{S}_{(k)}^{-1} z_iz_i^* (\gamma_N - \tilde{d}_i) \hat{S}_{(k)}^{-1} z_k + o(N^{\varepsilon-1})
\end{align*}
almost surely, for $\varepsilon>0$ and uniformly so on $\rho$.
The objective is then to show that the first right-hand side term is $o(N^{\varepsilon-1})$ almost surely and that this holds uniformly on $k$ and $\rho$. This is achieved by applying Lemma~\ref{le:keylemma1} with $c=d=z_k$. Indeed, Lemma~\ref{le:keylemma1} ensures that, for each integer $p$,\footnote{Note that Lemma~\ref{le:keylemma1} can strictly be applied here for $n-1$ instead of $n$; but since $1/n-1/(n-1)=O(n^{-2})$, this does not affect the result.}
\begin{align*}
\EE\left[\left|\frac{1}{n}\sum_{i\neq k}\frac{1}{N}z_k^*S_{(k)}^{-1}(\rho) z_iz_i^* S_{(k)}^{-1}(\rho)z_k\left(\frac{1}{N}z_i^*{S}_{(i,k)}^{-1}(\rho)z_i-\gamma_N(\rho)\right)\right|^p\right]=O(N^{-p})
\end{align*}
From this lemma, applying Markov's inequality, we have for each $k$,
\begin{align*}
P \left( N^{1-\varepsilon} \left| \frac1n\sum_{i\neq k} \frac1Nz_k^* \hat{S}_{(k)}^{-1}z_iz_i^*\hat{S}_{(k)}^{-1}z_k \left( \frac1Nz_i^*\hat{S}_{(i,k)}^{-1}z_i - \gamma_N \right) \right| > \eta \right) \leq K N^{-p\varepsilon}
\end{align*}
for some $K>0$ only dependent on $\eta>0$. Applying the union bound on the $n(n+1)$ events for $k=1,\ldots,n$ and for $\rho\in\{\rho_0,\ldots,\rho_n\}$, regular $n$-discretization of $\mathcal R_\kappa$, we then have
\begin{align*}
&P \left( \max_{k,j} N^{1-\varepsilon} \left| \frac1n\sum_{i\neq k} \frac1Nz_k^* \hat{S}_{(k)}^{-1}z_iz_i^*\hat{S}_{(k)}^{-1}z_k \left( \frac1Nz_i^*\hat{S}_{(i,k)}^{-1}z_i - \gamma_N(\rho_j) \right) \right| > \eta \right) \nonumber \\
&\leq K N^{-p\varepsilon+2}.
\end{align*}
Taking $p>3/\varepsilon$, by the Borel Cantelli lemma the above convergence holds almost surely, we finally get
\begin{align*}
\max_{k,j} \left| N^{1-\varepsilon} \left( \tilde{d}_k(\rho_j) - \frac1N z_k^*\left( \alpha(\rho_j) \frac1n \sum_{i\neq k} \frac{z_iz_i^*}{\tilde{d}_i(\rho_j)} + \rho_j I_N \right)^{-1}z_k \right) \right| \asto 0.
\end{align*}
Using the $\rho$-Lipschitz property (which holds almost surely so for all large $n$ a.s.) on both terms in the above difference concludes the proof of the proposition.
\end{proof}
The crux of the proof for the convergence of $\xi_2$ starts now.
In a similar manner as in the proof of Lemma~\ref{le:1}, we define $\tilde{f}_i(\rho)=d_i(\rho)/\tilde{d}_i(\rho)$ and reorder the indexes in such a way that $\tilde{f}_1(\rho)\leq \ldots\leq \tilde{f}_n(\rho)$ (this ordering depending on $\rho$). Then, by definition of $d_n(\rho)=\tilde{f}_i(\rho)\tilde{d}_i(\rho)$,
\begin{align*}
\tilde{d}_n(\rho) \tilde{f}_n(\rho) &= \frac1Nz_n^*\left( \alpha(\rho) \frac1n\sum_{i<n} \frac{z_iz_i^*}{\tilde{d}_i(\rho)\tilde{f}_i(\rho)} + \rho I_n \right)^{-1}z_n \\
&\leq \frac1Nz_n^*\left( \alpha(\rho) \frac1{\tilde{f}_n(\rho)} \frac1n\sum_{i<n} \frac{z_iz_i^*}{\tilde{d}_i(\rho)} + \rho I_n \right)^{-1}z_n
\end{align*}
where we used $\tilde{f}_n(\rho)\geq \tilde{f}_i(\rho)$ for each $i$. This inequality is equivalent to
\begin{align*}
\tilde{d}_n(\rho) &\leq \frac1Nz_n^* \left( \alpha(\rho) \frac1n\sum_{i<n} \frac{z_iz_i^*}{\tilde{d}_i(\rho)} + \tilde{f}_n(\rho) \rho I_n \right)^{-1}z_n.
\end{align*}
Assume now that, over some sequence $\{\rho^{(n)}\}\in\mathcal R_\kappa$, $\tilde{f}_n(\rho^{(n)})>1+N^{\eta-1}$ infinitely often for some $\eta>0$ (or equivalently, $d_n(\rho^{(n)})>\tilde{d}_n(\rho^{(n)})+N^{\eta-1}$ i.o.). Then we would have
\begin{align*}
\tilde{d}_n(\rho^{(n)}) &\leq \frac1Nz_n^*\left( \alpha(\rho^{(n)}) \frac1n\sum_{i<n} \frac{z_iz_i^*}{\tilde{d}_i(\rho^{(n)})} + \rho^{(n)} (1+N^{\eta-1}) I_N \right)^{-1}z_n \\
&=\tilde{d}_n(\rho^{(n)}) - N^{\eta-1} \frac1Nz_n^*\left( \frac1n\sum_{i<n} \frac{\alpha(\rho^{(n)}) z_iz_i^*}{\rho^{(n)}\tilde{d}_i(\rho^{(n)})} + (1+N^{\eta-1}) I_N \right)^{-1}\nonumber \\
&\times\left( \frac1n\sum_{i<n} \frac{\alpha(\rho^{(n)}) z_iz_i^*}{\tilde{d}_i(\rho^{(n)})} + \rho I_N \right)^{-1}z_n.
\end{align*}
But, by Proposition~\ref{prop:2}, letting $0<\varepsilon<\eta$, we have, for all large $n$ a.s.,
\begin{align*}
\frac1Nz_n^*\left( \alpha(\rho^{(n)}) \frac1n\sum_{i<n} \frac{z_iz_i^*}{\tilde{d}_i(\rho^{(n)})} + \rho^{(n)} I_n \right)^{-1}z_n \leq \tilde{d}_n(\rho^{(n)}) + N^{\varepsilon-1}
\end{align*}
which, along with the uniform boundedness of the $\tilde{d}_i$ away from zero, leads to
\begin{align*}
\tilde{d}_n(\rho^{(n)}) &\leq \tilde{d}_n(\rho^{(n)}) + N^{\varepsilon-1} - KN^{\eta-1}
\end{align*}
for some $K>0$. But, as $N^{\varepsilon-1} - KN^{\eta-1}<0$ for all large $N$, we obtain a contradiction. Hence, for each $\eta>0$, we have for all large $n$ a.s., $d_n(\rho)<\tilde{d}_n(\rho)+N^{\eta-1}$ uniformly on $\rho\in\mathcal R_\kappa$. Proceeding similarly with $d_1(\rho)$, and exploiting $\limsup_n \sup_\rho \max_i |\tilde{d}_i(\rho)|=O(1)$ a.s.\@, we finally have, for each $0<\varepsilon<\frac12$, that
\begin{align*}
\max_{1\leq k\leq n} \sup_{\rho\in\mathcal R_\kappa}\left| N^{1-\varepsilon} \left( d_k(\rho) - \tilde{d}_k(\rho) \right) \right| &\asto 0
\end{align*}
(for this, take an $\eta$ such that $0<\eta<\varepsilon$ and use $\max_k \sup_\rho |d_k(\rho)-\tilde{d}_k(\rho)|<N^{\eta-1}$ for all large $n$ a.s.).
Getting back to $\xi_2$, we now have
\begin{align*}
N^{1-\varepsilon} |\xi_2(\rho)| &= N^{1-\varepsilon} \left|a^*\hat{C}_N^{-1}(\rho) \left( \frac{\alpha(\rho)}n \sum_{i=1}^n \frac{d_i(\rho)-\tilde{d}_i(\rho)}{d_i(\rho)\tilde{d}_i(\rho)} z_iz_i^* \right)\hat{S}_N^{-1}(\rho)b\right|.
\end{align*}
But, from the above result,
\begin{align*}
N^{1-\varepsilon} \left\Vert \frac{\alpha(\rho)}n \sum_{i=1}^n \frac{d_i(\rho)-\tilde{d}_i(\rho)}{d_i(\rho)\tilde{d}_i(\rho)} z_iz_i^* \right\Vert &\leq N^{1-\varepsilon} \max_{1\leq k\leq n} \left| \frac{d_k(\rho)-\tilde{d}_k(\rho)}{d_k(\rho)\tilde{d}_k(\rho)} \right| \left\Vert \frac{\alpha(\rho)}n \sum_{i=1}^n z_iz_i^* \right\Vert \\
&\asto 0
\end{align*}
uniformly so on $\rho\in\mathcal R_\kappa$ which, along with the boundedness of $\Vert \hat{C}_N^{-1}\Vert$, $\Vert \hat{S}_N^{-1}\Vert$, $\Vert a\Vert$, and $\Vert b\Vert$, finally gives $N^{1-\varepsilon} \xi_2\asto 0$ uniformly on $\rho\in\mathcal R_\kappa$ as desired.
\bigskip
We have then proved that for each $\varepsilon>0$,
\begin{align*}
\sup_{\rho\in\mathcal R_\kappa}\left| N^{1-\varepsilon} \left( a^*\hat{C}_N^{-1}(\rho)b - a^*\hat{S}_N^{-1}(\rho)b \right)\right| \asto 0
\end{align*}
which proves Theorem~\ref{th:bilin} for $k=-1$. The generalization to arbitrary $k$ is rather immediate. Writing recursively $\hat{C}_N^k-\hat{S}_N^k=\hat{C}_N^{k-1}(\hat{C}_N-\hat{S}_N)+(\hat{C}_N^{k-1}-\hat{S}_N^{k-1})\hat{S}_N$, for positive $k$ or $\hat{C}_N^k-\hat{S}_N^k=\hat{C}_N^k(\hat{S}_N-\hat{C}_N)\hat{S}_N^{-1}+(\hat{C}_N^{k-1}-\hat{S}_N^{k-1})\hat{S}_N^{-1}$, \eqref{eq:firsteq} becomes a finite sum of terms that can be treated almost exactly as in the proof. This concludes the proof of Theorem~\ref{th:bilin}.
\subsection{Fluctuations of the GLRT detector}
This section is devoted to the proof of Theorem~\ref{th:T}, which shall fundamentally rely on Theorem~\ref{th:bilin}. The proof will be established in two steps. First, we shall prove the convergence for each $\rho\in\mathcal R_\kappa$, which we then generalize to the uniform statement of the theorem.
Let us then fix $\rho\in\mathcal R_\kappa$ for the moment.
In anticipation of the eventual replacement of $\hat{C}_N(\rho)$ by $\underline{\hat{S}}_N(\underline\rho)$, we start by studying the fluctuations of the bilinear forms involved in $T_N(\rho)$ but with $\hat{C}_N(\rho)$ replaced by $\underline{\hat{S}}_N(\underline\rho)$ (note that $T_N(\rho)$ remains constant when scaling $\hat{C}_N(\rho)$ by any constant, so that replacing $\hat{C}_N(\rho)$ by $\underline{\hat{S}}_N(\underline\rho)$ instead of by $\underline{\hat{S}}_N(\underline\rho)\cdot\frac1N\tr \hat{S}_N(\rho)$ as one would expect comes with no effect).
Our first goal is to show that the vector $\sqrt{N}(\Re[y^*\underline{\hat S}^{-1}_N(\underline\rho) p],\Im[y^*\underline{\hat S}^{-1}_N(\underline\rho) p])$ is asymptotically well approximated by a zero mean Gaussian vector with given covariance matrix. To this end, let us denote $A=[y~p]\in\CC^{N\times 2}$ and $Q_N=Q_N(\underline\rho)=(I_N+(1-\underline{\rho}) m(-\underline{\rho})C_N)^{-1}$. Then, from \cite[Lemma~5.3]{CHA12} (adapted to our current notations and normalizations), for any Hermitian $B\in\CC^{2\times 2}$ and for any $u\in\RR$,
\begin{align}
&\EE\left[\exp\left( \imath \sqrt{N} u \tr BA^*\left[\underline{\hat{S}}_N(\underline\rho)^{-1}-\frac1{\underline\rho}Q_N(\underline\rho) \right]A \right) ~\Big|~y\right] \nonumber \\
&= \exp\left(-\frac12 u^2 \Delta_N^2(B;y;p) \right) + O(N^{-\frac12}) \label{eq:characteristic_fun}
\end{align}
where we denote by $\EE[\cdot |y]$ the conditional expectation with respect to the random vector $y$ and where
\begin{align*}
\Delta_N^2(B;y;p) &\triangleq \frac{cm(-\underline{\rho})^2(1-\underline{\rho})^2\tr \left( A B A^* C_NQ_N^2(\underline\rho)\right)^2}{\underline{\rho}^2 \left(1-c m(-\underline{\rho})^2(1-\underline{\rho})^2 \frac1N\tr C_N^2Q_N^2(\underline\rho)\right) } .
\end{align*}
Also, we have from classical central limit results on Gaussian random variables
\begin{align*}
\EE\left[\exp\left( \imath \sqrt{N} u \tr B \left[A^*Q_N(\underline\rho)A - \Gamma_N \right]\right)\right] &= \exp \left( -\frac12u^2 \Delta_N^{\prime 2}(B;p) \right) + O(N^{-\frac12})
\end{align*}
where
\begin{align*}
\Gamma_N &\triangleq \frac1{\underline\rho}\begin{bmatrix} \frac1N\tr C_NQ_N(\underline\rho)& 0 \\ 0 & p^*Q_N(\underline\rho)p \end{bmatrix} \\
\Delta_N^{\prime 2}(B;p) &\triangleq \frac{B_{11}^2}{\underline\rho^2} \frac1N\tr C_N^2Q_N^2(\underline\rho) + \frac{2|B_{12}|^2}{\underline\rho^2} p^*C_NQ_N^2(\underline\rho)p.
\end{align*}
Besides, the $O(N^{-\frac12})$ terms in the right-hand side of \eqref{eq:characteristic_fun} remains $O(N^{-\frac12})$ under expectation over $y$ (for this, see the proof of \cite[Lemma~5.3]{CHA12}).
Altogether, we then have
\begin{align*}
& \EE\left[\exp\left( \imath \sqrt{N} u \tr B\left[ A^*\underline{\hat{S}}_N^{-1}(\underline\rho)A - \Gamma_N\right] \right)\right] \nonumber \\
&= \EE \left[\exp\left(-\frac12 u^2 \Delta_N^2(B;y;p)\right) \right] \exp\left(-\frac12 u^2 \Delta_N^{\prime 2}(B;p) \right) + O(N^{-\frac12}).
\end{align*}
Note now that
\begin{align*}
A^* C_NQ_N^2(\underline\rho)A - \Upsilon_N &\asto 0
\end{align*}
where
\begin{align*}
\Upsilon_N &\triangleq \begin{bmatrix} \frac1N\tr C_N^2 Q_N^2(\underline\rho) & 0 \\ 0 & p^*C_NQ_N^2(\underline\rho) p \end{bmatrix}
\end{align*}
so that, by dominated convergence, we obtain
\begin{align*}
&\EE\left[\exp\left( \imath \sqrt{N} u \tr B\left[ A^*\underline{\hat{S}}_N^{-1}(\underline\rho)A - \Gamma_N\right] \right)\right] \nonumber \\
&= \exp\left(-\frac12 u^2 \left[ \Delta_N^2(B;p) + \Delta_N^{\prime 2}(B;p) \right]\right) + o(1)
\end{align*}
where we defined
\begin{align*}
\Delta_N^2(B;p) &\triangleq \frac{cm(-\underline{\rho})^2(1-\underline{\rho})^2 \tr \left( B \Upsilon_N\right)^2}{\underline{\rho}^2 \left(1-c m(-\underline{\rho})^2(1-\underline{\rho})^2 \frac1N\tr C_N^2Q_N^2(\underline\rho)\right)}.
\end{align*}
By a generalized L\'evy's continuity theorem argument (see e.g.\@ \cite[Proposition~6]{HAC06}) and the Cramer-Wold device, we conclude that
\begin{align*}
\sqrt{N}\left(y^*\underline{\hat{S}}_N^{-1}(\underline{\rho})y,\Re[y^*\underline{\hat{S}}_N^{-1}(\underline{\rho})p],\Im[y^*\underline{\hat{S}}_N^{-1}(\underline{\rho})p]\right) - Z_N &= o_P(1)
\end{align*}
where $Z_N$ is a Gaussian random vector with mean and covariance matrix prescribed by the above approximation of $\sqrt{N} \tr BA^*\underline{\hat{S}}_N^{-1}A$ for each Hermitian $B$. In particular, taking $B_1\in\left\{ \left[\begin{smallmatrix} 0 & \frac12 \\ \frac12 & 0\end{smallmatrix}\right],\left[\begin{smallmatrix} 0 & \frac{\imath}2 \\ -\frac{\imath}2 & 0\end{smallmatrix}\right]\right\}$ to retrieve the asymptotic variances of $\sqrt{N}\Re[y^*\underline{\hat{S}}_N^{-1}(\underline{\rho})p]$ and $\sqrt{N}\Im[y^*\underline{\hat{S}}_N^{-1}(\underline{\rho})p]$, respectively, gives
\begin{align*}
\Delta_N^2(B_1;p) &=\frac1{2 \underline{\rho}^2} p^*C_NQ_N^2(\underline\rho)p \frac{c m(-\underline{\rho})^2(1-\underline{\rho})^2 \frac1N\tr C_N^2Q_N^2(\underline\rho) }{1-c m(-\underline{\rho})^2(1-\underline{\rho})^2 \frac1N\tr C_N^2Q_N^2(\underline\rho)} \\
\Delta_N^{\prime 2}(B_1;p) &=\frac1{2\underline{\rho}^2} p^*C_NQ_N^2(\underline\rho)p
\end{align*}
and thus $\sqrt{N}(\Re[y^*\underline{\hat{S}}_N^{-1}(\underline{\rho})p],\Im[y^*\underline{\hat{S}}_N^{-1}(\underline{\rho})p])$ is asymptotically equivalent to a Gaussian vector with zero mean and covariance matrix
\begin{align*}
(\Delta_N^2(B_1;p)+\Delta_N^{\prime 2}(B_1;p))I_2 &= \frac1{2\underline{\rho}^2} \frac{p^*C_NQ_N^2(\underline\rho)p}{1-c m(-\underline{\rho})^2(1-\underline{\rho})^2 \frac1N\tr C_N^2Q_N^2(\underline\rho)} I_2.
\end{align*}
We are now in position to apply Theorem~\ref{th:bilin}. Reminding that $\hat{S}_N^{-1}(\rho) (\rho + \frac1{\gamma_N(\rho)} \frac{1-\rho}{1-(1-\rho)c}) =\underline{\hat{S}}_N^{-1}(\underline\rho)$, we have by Theorem~\ref{th:bilin} for $k=-1$
\begin{align*}
\sqrt{N} A^*\left[\hat{C}_N^{-1}(\rho)-\frac{\underline{\hat{S}}_N(\underline\rho)^{-1}}{\rho + \frac1{\gamma_N(\rho)} \frac{1-\rho}{1-(1-\rho)c}}\right] A &\asto 0.
\end{align*}
Since almost sure convergence implies weak convergence, $\sqrt{N} A^*\hat{C}_N^{-1}(\rho)A$ has the same asymptotic fluctuations as $\sqrt{N} A^*\underline{\hat{S}}_N^{-1}(\underline\rho)A/(\frac1N\tr \hat{S}_N(\rho))$.
Also, as $T_N(\rho)$ remains identical when scaling $\hat{C}_N^{-1}(\rho)$ by $\frac1N\tr \hat{S}_N(\rho)$, only the fluctuations of $\sqrt{N} A^*\underline{\hat{S}}_N^{-1}(\underline\rho)A$ are of interest, which were previously derived. We then finally conclude by the delta method (or more directly by Slutsky's lemma) that
\begin{align*}
\sqrt{\frac{N}{y^*\hat{C}_N^{-1}(\rho)y p^*\hat{C}_N^{-1}(\rho)p}} \begin{bmatrix} \Re\left[ y^*\hat{C}_N^{-1}(\rho)p \right] \\ \Im\left[ y^*\hat{C}_N^{-1}(\rho)p \right] \end{bmatrix} - \sigma_N(\underline{\rho}) Z' = o_P(1)
\end{align*}
for some $Z'\sim \mathcal N(0,I_2)$ and
\begin{align*}
\sigma_N^2(\underline\rho) &\triangleq \frac12 \frac{ p^*C_NQ_N^2(\underline\rho)p}{ p^*Q_N(\underline\rho)p\cdot \frac1N\tr C_NQ_N(\underline\rho)\cdot \left(1-c m(-\underline{\rho})^2(1-\underline{\rho})^2 \frac1N\tr C_N^2Q_N^2(\underline\rho)\right) }.
\end{align*}
It unfolds that, for $\gamma>0$,
\begin{align}
\label{eq:conv_proba_rho}
P\left( T_N(\rho) > \frac{\gamma}{\sqrt{N}} \right) - \exp \left( - \frac{\gamma^2}{2\sigma_N^2(\underline\rho)} \right) \to 0
\end{align}
as desired.
\bigskip
The second step of the proof is to generalize \eqref{eq:conv_proba_rho} to uniform convergence across $\rho\in\mathcal R_\kappa$. To this end, somewhat similar to above, we shall transfer the distribution $P(\sqrt{N}T_N(\rho) > \gamma)$ to $P(\sqrt{N}\underline T_N(\rho) > \gamma)$ by exploiting the uniform convergence of Theorem~\ref{th:bilin}, where we defined\begin{align*}
\underline T_N(\rho) &\triangleq \frac{\left|y^*\underline{\hat S}_N(\underline\rho) p\right|}{\sqrt{y^*\underline{\hat S}_N(\underline\rho) y}\sqrt{p^*\underline{\hat S}_N(\underline\rho) p}}
\end{align*}
and exploit a $\rho$-Lipschitz property of $\sqrt{N}\underline T_N(\rho)$ to reduce the uniform convergence over $\mathcal R_\kappa$ to a uniform convergence over finitely many values of $\rho$.
The $\rho$-Lipschitz property we shall need is as follows: for each $\varepsilon>0$
\begin{align}
\label{eq:tightness_condition}
\lim_{\delta\to 0} \lim_{N\to\infty} P\left( \sup_{\substack{\rho,\rho'\in\mathcal R_\kappa \\ |\rho-\rho'|<\delta} } \sqrt{N}\left|T_N(\rho)-T_N(\rho')\right| > \varepsilon \right) &= 0.
\end{align}
Let us prove this result. By Theorem~\ref{th:bilin}, since almost sure convergence implies convergence in distribution, we have
\begin{align*}
P\left( \sup_{\rho\in\mathcal R_\kappa} \sqrt{N}\left|T_N(\rho)-\underline T_N(\rho)\right| > \varepsilon \right) &\to 0.
\end{align*}
Applying this result to \eqref{eq:tightness_condition} induces that it is sufficient to prove \eqref{eq:tightness_condition} for $\underline T_N(\rho)$ in place of $T_N(\rho)$.
Let $\eta>0$ small and $\mathcal A_N^\eta\triangleq \{\exists \underline\rho \in \mathcal R_\kappa,y^*\underline{\hat{S}}_N^{-1}(\underline\rho)yp^*\underline{\hat{S}}_N^{-1}(\underline\rho)p<\eta\}$. Developing the difference $\underline T_N(\rho)-\underline T_N(\rho')$ and isolating the denominator according to its belonging to $\mathcal A_N^\eta$ or not, we may write
\begin{align*}
& P\left( \sup_{\substack{\rho,\rho'\in\mathcal R_\kappa \\ |\rho-\rho'|<\delta} } \sqrt{N}\left|\underline T_N(\rho)-\underline T_N(\rho')\right| > \varepsilon \right) \nonumber \\
&\leq P\left( \mathcal A_N^\eta \right) + P\left( \sup_{\substack{\rho,\rho'\in\mathcal R_\kappa \\ |\rho-\rho'|<\delta} } \sqrt{N} V_N(\rho,\rho')>\varepsilon\eta \right)
\end{align*}
where
\begin{align*}
V_N(\rho,\rho') &\triangleq \left| y^*\underline{\hat{S}}_N^{-1}(\underline\rho)p \right| \sqrt{y^*\underline{\hat{S}}_N^{-1}(\underline\rho')y}\sqrt{p^*\underline{\hat{S}}_N^{-1}(\underline\rho')p} \nonumber \\
&- \left| y^*\underline{\hat{S}}_N^{-1}(\underline\rho')p \right| \sqrt{y^*\underline{\hat{S}}_N^{-1}(\underline\rho)y}\sqrt{p^*\underline{\hat{S}}_N^{-1}(\underline\rho)p}.
\end{align*}
From classical random matrix results, $P( \mathcal A_N^\eta )\to 0$ for a sufficiently small choice of $\eta$. To prove that $\lim_\delta \limsup_n P(\sup_{|\rho-\rho'|<\delta} \sqrt{N} V_N(\rho,\rho')>\varepsilon\eta)=0$, it is then sufficient to show that
\begin{align}
\label{eq:rho-rho'}
\lim_{\delta\to 0} \limsup_n P\left( \sup_{\substack{\rho,\rho'\in\mathcal R_\kappa \\ |\rho-\rho'|<\delta}} \sqrt{N}|y^*\underline{\hat S}_N(\underline\rho)^{-1}p-y^*\underline{\hat S}_N(\underline\rho')^{-1}p| > \varepsilon' \right) = 0
\end{align}
for any $\varepsilon'>0$ and similarly for $y^*\underline{\hat S}_N(\underline\rho)^{-1}y-y^*\underline{\hat S}_N(\underline\rho')^{-1}y$ and $p^*\underline{\hat S}_N(\underline\rho)^{-1}p-p^*\underline{\hat S}_N(\underline\rho')^{-1}p$.
Let us prove \eqref{eq:rho-rho'}, the other two results following essentially the same line of arguments.
For this, by \cite[Corollary~16.9]{KAL02} (see also \cite[Theorem~12.3]{BIL68}), it is sufficient to prove, say
\begin{align*}
\sup_{\substack{\rho,\rho'\in \mathcal R_\kappa\\ \rho\neq \rho'}} \sup_n \frac{\EE \left[\sqrt{N}\left|y^*\underline{\hat S}_N(\underline\rho)^{-1}p-y^*\underline{\hat S}_N(\underline\rho')^{-1}p \right|^2\right]}{|\rho-\rho'|^2} < \infty.
\end{align*}
But then, remarking that
\begin{align*}
&\sqrt{N} y^*\underline{\hat S}_N(\underline\rho)^{-1}p-y^*\underline{\hat S}_N(\underline\rho')^{-1}p \nonumber \\
&= (\underline\rho'-\underline\rho) \sqrt{N} y^*\underline{\hat S}_N(\underline\rho)^{-1}\left( I_N - \frac1n\sum_{i=1}^n z_iz_i^* \right)\underline{\hat S}_N(\underline\rho')^{-1}p
\end{align*}
this reduces to showing that
\begin{align*}
\sup_{\rho,\rho'\in \mathcal R_\kappa} \sup_n \EE \left[ N \left| y^*\underline{\hat S}_N(\underline\rho)^{-1}\left( I_N - \frac1n\sum_{i=1}^n z_iz_i^* \right)\underline{\hat S}_N(\underline\rho')^{-1}p \right|^2\right] <\infty.
\end{align*}
Conditioning first on $z_1,\ldots,z_n$, this further reduces to showing
\begin{align*}
\sup_{\rho,\rho'\in \mathcal R_\kappa} \sup_n \EE \left[ \left\Vert \underline{\hat S}_N(\underline\rho)^{-1}\left( I_N - \frac1n\sum_{i=1}^n z_iz_i^* \right)\underline{\hat S}_N(\underline\rho')^{-1}p \right\Vert^2\right] <\infty.
\end{align*}
But this is yet another standard random matrix result, obtained e.g., by noticing that
\begin{align*}
\left\Vert \underline{\hat S}_N(\underline\rho)^{-1}\left( I_N - \frac1n\sum_{i=1}^n z_iz_i^* \right)\underline{\hat S}_N(\underline\rho')^{-1}p \right\Vert^2 &\leq \frac1{\kappa^4}\left\Vert I_N - \frac1n\sum_{i=1}^n z_iz_i^* \right\Vert^2
\end{align*}
which remains of uniformly finite expectation (left norm is vector Euclidean norm, right norm is matrix spectral norm). This completes the proof of \eqref{eq:tightness_condition}.
\smallskip
Getting back to our original problem, let us now take $\varepsilon>0$ arbitrary, $\rho_1<\ldots<\rho_K$ be a regular sampling of $\mathcal R_\kappa$, and $\delta=1/K$. Then by \eqref{eq:conv_proba_rho}, $K$ being fixed, for all $n>n_0(\varepsilon)$,
\begin{align}
\label{eq:Kfoldconv}
\max_{1\leq k\leq K} \left| P\left( T_N(\rho_i) > \frac{\gamma}{\sqrt{N}} \right) - \exp\left( -\frac{\gamma^2}{2\sigma_N^2(\rho_i)} \right) \right| &< \varepsilon.
\end{align}
Also, from \eqref{eq:tightness_condition}, for small enough $\delta$,
\begin{align*}
&\max_{1\leq k\leq K} P\left( \sup_{\substack{\rho\in\mathcal R_\kappa \\ |\rho-\rho_k|<\delta}} \sqrt{N}|T_N(\rho)-T_N(\rho_k)| > \gamma \zeta \right) \nonumber \\
&\leq P\left( \sup_{\substack{\rho,\rho'\in\mathcal R_\kappa \\ |\rho-\rho'|<\delta}} \sqrt{N}|T_N(\rho)-T_N(\rho')| > \gamma \zeta \right) \\
&< \varepsilon
\end{align*}
for all large $n>n_0'(\varepsilon,\zeta)>n_0(\varepsilon)$ where $\zeta>0$ is also taken arbitrarily small. Thus we have, for each $\rho\in \mathcal R_\kappa$ and for $n>n_0'(\varepsilon,\zeta)$
\begin{align*}
P\left( T_N(\rho) > \frac{\gamma}{\sqrt{N}} \right) &\leq P\left( T_N(\rho_i) > \frac{\gamma(1-\zeta)}{\sqrt{N}} \right)+ P\left( \sqrt{N} |T_N(\rho)-T_N(\rho_i)|>\gamma\zeta\right) \\
&\leq P\left( T_N(\rho_i) > \frac{\gamma(1-\zeta)}{\sqrt{N}} \right) + \varepsilon
\end{align*}
for $i\leq K$ the unique index such that $|\rho-\rho_i|<\delta$ and where the inequality holds uniformly on $\rho\in \mathcal R_\kappa$. Similarly, reversing the roles of $\rho$ and $\rho_i$,
\begin{align*}
P\left( T_N(\rho) > \frac{\gamma}{\sqrt{N}} \right) &\geq P\left( T_N(\rho_i) > \frac{\gamma(1+\zeta)}{\sqrt{N}} \right) - \varepsilon.
\end{align*}
As a consequence, by \eqref{eq:Kfoldconv}, for $n>n_0'(\varepsilon,\zeta)$, uniformly on $\rho\in\mathcal R_\kappa$,
\begin{align*}
P\left( T_N(\rho) > \frac{\gamma}{\sqrt{N}} \right) &\leq \exp\left( -\frac{\gamma^2(1-\zeta)^2}{2\sigma_N^2(\rho_i)} \right) + 2\varepsilon \\
P\left( T_N(\rho) > \frac{\gamma}{\sqrt{N}} \right) &\geq \exp\left( -\frac{\gamma^2(1+\zeta)^2}{2\sigma_N^2(\rho_i)} \right) - 2\varepsilon
\end{align*}
which, by continuity of the exponential and of $\rho\mapsto \sigma_N(\rho)$,\footnote{Note that it is unnecessary to ensure $\liminf_N \sigma_N(\rho)>0$ as the exponential would tend to zero anyhow in this scenario.} letting $\zeta$ and $\delta$ small enough (up to growing $n_0'(\varepsilon,\zeta)$), leads to
\begin{align*}
\sup_{\rho\in\mathcal R_\kappa}\left| P\left( \sqrt{N} T_N(\rho) > \gamma \right) - \exp\left( -\frac{\gamma^2}{2\sigma_N^2(\rho)} \right)\right| &\leq 3\varepsilon
\end{align*}
for all $n>n_0'(\varepsilon,\zeta)$, which completes the proof.
\subsection{Around empirical estimates}
This section is dedicated to the proof of Proposition~\ref{prop:1} and Corollary~\ref{co:1}.
We start by showing that $\hat{\sigma}^2_N(1)$ is well defined. It is easy to observe that the ratio defining $\hat{\sigma}^2_N(\underline\rho)$ converges to an undetermined form (zero over zero) as $\underline\rho\uparrow 1$. Applying l'Hospital's rule to the ratio, using the differentiation $\frac{d}{d\underline\rho} \underline{\hat{S}}^{-1}_N(\underline\rho)=-\underline{\hat{S}}^{-2}_N(\underline\rho)(I_N-\frac1n\sum_i z_iz_i^*)$ and the limit $\underline{\hat{S}}^{-1}_N(\underline\rho)\to I_N$ as $\underline\rho\uparrow 1$, we end up with
\begin{align*}
\hat{\sigma}^2_N(\underline\rho) \to \frac12 \frac{p^*\left( \frac1n\sum_{i=1}^n z_iz_i^* \right)p}{\frac1N\tr \left( \frac1n\sum_{i=1}^n z_iz_i^* \right)}.
\end{align*}
Letting $\varepsilon>0$ arbitrary, since $p^*\frac1n\sum_i z_iz_i^*p - p^*C_Np\asto 0$, $\frac1N\tr \frac1n\sum_i z_iz_i^*\asto 1$ as $n\to\infty$, we immediately have, by continuity of both $\sigma^2_N(\underline\rho)$ and $\hat\sigma^2_N(\underline\rho)$,
\begin{align*}
\sup_{\rho \in (1-\kappa,1]}\left|\hat{\sigma}^2_N(\underline\rho) - \sigma^2_N(\underline\rho) \right| &\leq \varepsilon
\end{align*}
for all large $n$ almost surely. From now on, it then suffices to prove Proposition~\ref{prop:1} on the complementary set $\mathcal R_\kappa'\triangleq [\kappa+\min\{0,1-c^{-1}\},1-\kappa]$.
For this, we first recall the following results borrowed from \citep{COU14}:
\begin{align*}
\sup_{\rho\in\mathcal R_\kappa} \left\Vert \frac{\hat{C}_N(\rho)}{\frac1N\tr \hat{C}_N(\rho)} - \underline{\hat{S}}_N(\underline\rho) \right\Vert &\asto 0.
\end{align*}
Also, for $z\in\CC\setminus \RR^+$, defining
\begin{align*}
\underline{\underline{\hat S}}_N(z)&\triangleq (1-\underline\rho)\frac1n\sum_{i=1}^n z_iz_i^* -z I_N
\end{align*}
(so in particular $\underline{\underline{\hat S}}_N(-\underline\rho)=\underline{\hat S}_N(\underline\rho)$, for all $\underline\rho\in \mathcal R_\kappa$), we have, with $\mathcal C$ a compact set of $\CC\setminus\RR^+$ and any integer $k$,
\begin{align*}
\sup_{\bar z\in\mathcal C} \left| \frac{d^k}{dz^k} \left\{ \frac1N\tr \underline{\underline{\hat S}}_N^{-1}(z) - \frac1N\tr \left(-z \left[I_N + (1-\underline\rho)m_N(z)C_N\right]\right)^{-1} \right\}_{z=\bar z} \right| &\asto 0 \\
\sup_{\bar z\in\mathcal C} \left| \frac{d^k}{dz^k} \left\{ p^*\underline{\underline{\hat S}}_N^{-1}(z)p - p^*\left(-z \left[I_N + (1-\underline\rho)m_N(z)C_N\right]\right)^{-1}p \right\}_{z=\bar z} \right| &\asto 0
\end{align*}
where $m_N(z)$ is defined as the unique solution with positive (resp.\@ negative) imaginary part if $\Im[z]>0$ (resp.\@ $\Im[z]<0$) or unique positive solution if $z<0$ of
\begin{align*}
m_N(z) &= \left( -z + c \int \frac{(1-\underline\rho)t}{1+(1-\underline\rho)tm_N(z)}\nu_N(dt) \right)^{-1}
\end{align*}
(this follows directly from \citep{SIL95}).
This expression of $m_N(z)$ can be more rewritten under the more convenient form
\begin{align*}
m_N(z) &= -\frac{1-c}z + c \int \frac{\nu_N(dt)}{-z-z(1-\underline\rho)tm_N(z)} \\
&= -\frac{1-c}z + c \frac1N\tr \left(-z \left[I_N + (1-\underline\rho)m_N(z)C_N\right]\right)^{-1}
\end{align*}
so that, from the above relations
\begin{align*}
\sup_{\rho \in\mathcal R_\kappa'} \left| m_N(-\underline\rho) - \left( \frac{1-c_N}{\underline\rho} + c_N \frac1N\tr \hat{C}_N^{-1}(\rho)\cdot\frac1N\tr \hat{C}_N(\rho) \right) \right| &\asto 0 \\
\sup_{\rho \in\mathcal R_\kappa'} \left| \int \frac{t\nu_N(dt)}{1+(1-\underline\rho)m_N(-\underline\rho)t} - \frac{ 1 - \underline\rho \frac1N\tr \hat{C}_N^{-1}(\rho)\cdot\frac1N\tr \hat{C}_N(\rho)}{(1-\underline\rho)m_N(-\underline\rho)} \right| &\asto 0.
\end{align*}
Differentiating along $z$ the first defining identity of $m_N(z)$, we also recall that
\begin{align*}
m_N'(z) &= \frac{m_N^2(z)}{1-c \int \frac{m_N(z)^2(1-\underline\rho)^2t^2\nu_N(dt)}{(1-(1-\underline\rho)tm_N(-\underline\rho))^2}}.
\end{align*}
Now, remark that
\begin{align*}
p^*\underline{\underline{\hat S}}_N(\underline\rho)^{-2}p &= \frac{d}{dz} \left[ p^*\underline{\underline{\hat S}}_N(z)^{-1}p \right]_{z=-\underline\rho}
\end{align*}
which (by analyticity) is uniformly well approximated by
\begin{align*}
& \frac{d}{dz} \left[ p^*\left( -z \left[ I_N + (1-\underline\rho)m_N(z) C_N\right]\right)^{-1}p \right]_{z=-\underline\rho} \nonumber \\
&= \frac1{\underline\rho^2} p^*Q_N(\underline\rho)p - \frac1{\underline\rho} (1-\underline\rho) m_N'(-\underline\rho)p^*C_NQ^2_N(\underline\rho)p \\
&= \frac1{\underline\rho^2} p^*Q_N(\underline\rho)p - \frac1{\underline\rho} (1-\underline\rho) \frac{m_N^2(-\underline\rho)p^*C_NQ^2_N(\underline\rho)p}{1-c m_N(-\underline\rho)^2(1-\underline\rho)^2 \frac1N\tr Q^2_N(\underline\rho)}.
\end{align*}
(recall that $Q_N(\underline\rho)=\left(I_N+(1-\underline\rho)m_N(-\underline\rho)C_N \right)^{-1}$).
We then conclude
\begin{align*}
&\sup_{\rho\in\mathcal R_\kappa'} \left| \frac{p^*C_NQ^2_N(\underline\rho)p}{1-c m_N(-\underline\rho)^2(1-\underline\rho)^2 \frac1N\tr Q^2_N(\underline\rho)} \right. \nonumber \\
& \left. - \frac{p^*\hat{C}_N^{-1}(\rho)p\cdot \frac1N\tr \hat{C}_N(\rho)-\underline\rho p^*\hat{C}_N^{-2}(\rho)p\cdot \left( \frac1N\tr \hat{C}_N(\rho)\right)^2}{(1-\underline\rho)m_N(-\underline\rho)^2} \right| &\asto 0.
\end{align*}
Putting all results together, we obtain the expected result.
\bigskip
It now remains to prove Corollary~\ref{co:1}. This is easily performed thanks to Theorem~\ref{th:T} and Proposition~\ref{prop:1}.
From these, we indeed have the three relations
\begin{align*}
P\left( \sqrt{N} T_N(\hat\rho_N^*) > \gamma \right) - \exp\left(- \frac{\gamma^2}{2 \sigma_N^2(\underline{\hat\rho}_N^*)} \right) &\asto 0 \\
P\left( \sqrt{N} T_N(\rho_N^*) > \gamma \right) - \exp\left(- \frac{\gamma^2}{2 \sigma_N^2(\underline\rho_N^*)} \right) &\to 0 \\
\exp\left(- \frac{\gamma^2}{2 \sigma_N^2(\underline{\hat\rho}_N^*)} \right) - \exp\left(- \frac{\gamma^2}{2 \sigma_N^{*2}} \right) &\asto 0
\end{align*}
where we denoted $\rho_N^*$ any element in the argmin over $\rho$ of $P(\sqrt{N}T_N(\rho)>\gamma)$ (and $\underline\rho_N^*$ its associated value through the mapping $\rho\mapsto\underline\rho$) and $\sigma_N^{*2}$ the minimum of $\sigma_N(\underline\rho)$ (i.e.\@ the minimizer for $\exp(- \frac{\gamma^2}{2 \sigma_N^2(\underline\rho)})$). Note that the first two relations rely fundamentally on the uniform convergence $\sup_{\rho\in\mathcal R_\kappa}|P\left( \sqrt{N} T_N(\rho) > \gamma \right)-\exp\left(-\gamma^2/(2\sigma_N^2(\underline\rho)) \right)|\asto 0$. By definition of $\rho_N^*$ and $\sigma_N^{*2}$, we also have
\begin{align*}
\exp\left(- \frac{\gamma^2}{2 \sigma_N^{*2}} \right) &\leq \min\left\{ \exp\left(- \frac{\gamma^2}{2 \sigma_N^2(\underline{\hat\rho}_N^*)} \right) , \exp\left(- \frac{\gamma^2}{2 \sigma_N^2(\underline{\rho}_N^*)} \right) \right\} \\
P\left( \sqrt{N} T_N(\rho_N^*) > \gamma \right) &\leq P\left( \sqrt{N} T(\hat\rho_N^*) > \gamma \right).
\end{align*}
Putting things together then gives
\begin{align*}
P\left( \sqrt{N} T(\hat\rho_N^*) > \gamma \right) - P\left( \sqrt{N} T_N(\rho_N^*) > \gamma \right) \asto 0
\end{align*}
which is the expected result.
|
\section{Introduction}
General Relativity (GR) has been very successful at describing gravity
in a vast range of scales, from submillimeter ones~\cite{Adelberger:2002ic}, to those of the Earth, Solar system~\cite{Will:2005va}
and binary pulsars~\cite{1975ApJ...195L..51H,1975ApJ...199L..25B,Damour:1998jk,Antoniadis:2013pzd,Freire:2012mg}.
This beautiful theory, however, is known to be incomplete in the ultra-violet regime where
it must be replaced by a quantum theory of gravity. Furthermore, at
the infra-red cosmological scales, it needs to be supplemented with Dark Matter
and an unnaturally valued cosmological constant to explain observations, which can also be interpreted as a sign of failure.
These reasons
have spurred intense, and ongoing, efforts exploring how to describe gravity at
both classical and quantum levels, the latter being one of the most
challenging enterprises of modern physics.
Restricting to the classical regime --where both laboratory experiments and astrophysical observations can help constrain
possible alternatives-- a large number of putative theories have already been significantly constrained or ruled out altogether
(see e.g. Refs.~\cite{Will:2005yc,1995werp.book.....W}). In the particular case of astrophysical observations, binaries
involving pulsars have proved especially well suited for these studies~\cite{1975ApJ...195L..51H,1975ApJ...199L..25B,Damour:1998jk,Antoniadis:2013pzd,Freire:2012mg}.
Indeed, exquisite electromagnetic observations of
the pulsar signal allow following the binary's orbital dynamics and comparing it with predictions
from GR and other theories. To date this task has necessarily involved binaries
at relatively large separations and, correspondingly, low orbital velocities ($v/c \ll 1)$. Binaries in such configurations, as a consequence,
do not fully explore possible discrepancies that might arise at relativistic velocities $v/c \simeq 1$.
Such discrepancies include dipolar emission of non-tensorial gravitational waves~\cite{eardley1975,Will:1989sk,Damour:1992we,Damour:1993hw,Damour:1996ke},
as well as dynamical violations of the (strong) equivalence principle
and enhancement of the strength of the gravitational attraction in the last, highly relativistic stages of the binary inspiral and plunge~\cite{Barausse:2012da,Shibata:2013pra,Palenzuela:2013hsa,Taniguchi:2014fqa}.
Such status of affairs will soon be radically changed thanks to a
network of gravitational wave detectors that will allow analyzing compact binaries in highly relativistic velocity regimes. The detection
of gravitational waves from these systems will not only allow testing GR, but will additionally provide
important clues about the physical nature of these binaries, as well as identify the source's location. This knowledge may help -- both
directly and indirectly -- identify
electromagnetic counterparts (e.g. Refs.~\cite{2009arXiv0902.1527B,2013PhRvD..87l3004K,Andersson:2013mrx,Lehner:2014asa}) to these systems.
The synergy of gravitational and electromagnetic observations will permit in-depth ``multi-messenger'' investigation
of the behavior of gravity in such highly relativistic binaries.
For this analysis to be possible, on the gravitational-wave side it is
important that
possible deviations from the predictions of GR are either
understood, or suitably parametrized~\footnote{See discussion in e.g. Ref.~\cite{Sampson:2013jpa}.}, so as to guide
the detection and analysis of possible signals. Activities on both these fronts have recently been gaining significant momentum:
In the ``parametrized'' approach particular formalisms have been presented, motivated by specific theories
and phenomenological considerations. Such formalisms have been applied to derive bounds on the relevant
parameters describing deviations from GR~\cite{Yunes:2009ke,Li:2011cg,Loutrel:2014vja}.
In the ``direct'' approach, deviations from the gravitational waves predicted by GR are computed in specific bona fide gravity theories
(i.e. ones with a well-defined initial value problem)~\cite{Barausse:2012da,2012CQGra..29w2002H,Mirshekari:2013vb,Palenzuela:2013hsa,Shibata:2013pra,Berti:2013gfa}, and the prospects for
detecting these deviations are analyzed~\cite{Sampson:2013jpa,Sampson:2014qqa}. Among these theories, scalar-tensor
(ST) theories~\cite{Fierz:1956zz,jordan,Brans:1961sx,Bergmann:1968ve,Wagoner:1970vr,Nordtvedt:1970uv} -- where gravity is mediated not only by a metric tensor but also
by a scalar field -- have received the most attention, because the presence of a scalar field in nature is motivated
by e.g., the low-energy limit of string theories, the observation of the Higgs boson, and cosmological phenomenology (i.e. inflation and Dark Energy).
Among the most likely sources of gravitational waves for Earth-based detectors is the coalescence of binary
neutron stars. The study of such systems in ST theories has been traditionally undertaken via suitable perturbation
expansions~\cite{Will:1989sk,Damour:1992we,Damour:1996ke,Mirshekari:2013vb}, and more recently via numerical simulations~\cite{Barausse:2012da,Shibata:2013pra} or a hybrid
approach~\cite{Palenzuela:2013hsa}. These works have not only provided definitive predictions for the expected signals
but have also helped illustrate that strong deviations from GR could arise, in certain classes of ST theories, because of dynamically-induced
effects as the orbit tightens~\cite{Barausse:2012da}. A first analysis indicating how such differences could be
detected in the near future (by the upcoming second generation of gravitational wave interferometers) has been presented
in Ref.~\cite{Sampson:2014qqa} (see also~\cite{Taniguchi:2014fqa}).
As mentioned, among the possible physical parameters that can be obtained via gravitational wave observations
are the time (and frequency) of the merger as well as the sky location, both of which would aid follow-up efforts
to capture counterparts in a wide range of electromagnetic bands. Within GR, much effort has been going into identifying promising near-coalescence scenarios
able to yield detectable signals in the electromagnetic spectrum, and a number of mechanisms have been proposed and explored in recent years
(e.g. in Refs.~\cite{Palenzuela:2013kra,Palenzuela:2013hu,Kyutoku:2012fv,2012ApJ...746...48M,2012arXiv1204.6242P,Tsang:2011ad,2013PhRvD..87l3004K}).
It is therefore interesting to consider whether electromagnetic signals (either on their own, or in combination with gravitational wave observations)
can provide additional clues as to whether gravity behaves as predicted by GR. This is especially important
as facilities for gravitational wave observations will be for years much more restricted in the frequency window they can access
in comparison to the large spectra provided by diverse electromagnetic observatories.
To this goal, we consider here whether electromagnetic
counterparts triggered during the coalescence stage of neutron-star binaries can provide such testing opportunities. In particular
we here study the electromagnetic energy flux produced as a result of magnetosphere interactions in GR, as well as
in the ST theories that were shown in Refs.~\cite{Barausse:2012da,Palenzuela:2013hsa,Shibata:2013pra} to yield significant
deviations from the GR behavior in the late stages of the evolution of binary neutron star systems.
We take advantage of a unipolar induction model to account for these magnetospheric effects, enhanced to include
the stars' own magnetization in defining the radius at which the induction takes place~\cite{Palenzuela:2013kra}.
The emitted electromagnetic Poynting luminosity -- studied in Refs.~\cite{Ponce:2014sza,Palenzuela:2013kra,Palenzuela:2013hu}
and related previous works~\cite{1969ApJ...156...59G,Hansen:2000am,Lai:2012qe} -- might be converted into potentially
observable x-ray and radio signals by several processes~\cite{Hansen:2000am,Palenzuela:2013kra}.
Additionally, the binary's dynamics needed for this model is described through a post-Newtonian (PN) treatment of the equations of
motions within scalar tensor theories, augmented by a set of equations that provide a description of the scalar charge of each binary's
component along the evolution~\cite{Palenzuela:2013hsa}.
This paper is organized as follows:
in section \ref{sec:Methods} we review the ST theories considered, and
we describe the procedure that we use to obtain the dynamics and estimate the electromagnetic luminosities;
in section \ref{sec:Results} we present the cases considered and the results obtained;
and finally in section \ref{sec:Discussion} we discuss the implications of our work.
\section{Methods}
\label{sec:Methods}
In this work we are primarily concerned with estimating possible
electromagnetic signals from the coalescence of magnetized binary neutron stars
in GR and certain ST theories. These theories are described in section~\ref{subsec:STdynamics}.
We here note that such theories do suffer from the infrared problems mentioned in our introduction
but do provide an interesting and well motivated model to explore possible ``ultraviolet'' (strong-field) deviations.
Since they also have a well defined initial value problem and yield a well posed physical problem,
even in the non-linear regime, they constitute an excellent arena to study possible deviations
from general relativity.
We obtain the binary's dynamics
by solving the 2.5 PN equations of motion for ST theories as
described in Ref.~\cite{Mirshekari:2013vb} (enhanced with the formalism
proposed and validated in Ref.~\cite{Palenzuela:2013hsa} to account for the scalarization
effects allowed by the theories). The electromagnetic radiation
induced by magnetospheric interactions of the magnetized neutron
stars is estimated using a phenomenological model
based on an extension of the unipolar inductor (we will refer to it as the ``unipolar model'' throughout this work). Such a model
captures magnetospheric effects by considering the electromotive force (emf)
that is induced as an otherwise non-magnetized conductor moves through a magnetic field~\cite{Hansen:2000am,Lai:2012qe}.
We have augmented this model recently in Ref.~\cite{Palenzuela:2013kra,Ponce:2014sza} to account
for an additional ``shielding effect'' that arises when both stars are magnetized, and which modifies
the radius at which the emf induction takes place.
\subsection{Scalar-tensor theories}
\label{subsec:STdynamics}
\subsubsection{Dynamics in ST theories and equations of motion}
The action for a ST theory can be written in the Jordan frame as
\begin{equation}
\label{Jframe_action}
S=\!\!\int d^4 x\frac{\sqrt{-g}}{2\kappa}\left[\phi R-\frac{\omega(\phi)}{\phi} \partial_\mu \phi \partial^\mu \phi
\right]+ S_M[g_{\mu\nu},\psi]\,,
\end{equation}
where $\kappa=8 \pi G$, $R$, $g$ and $\phi$ are respectively the Ricci scalar, the metric determinant and the scalar field, and the
theory has no potential for the scalar field, but is characterized by an arbitrary function $\omega(\phi)$ (and by the boundary conditions for the scalar field).
Note also that
in this action we have assumed that the matter degrees of freedom $\psi$ couple minimally to the metric (and not to the
scalar field) so as to enforce the weak equivalence principle (i.e. the universality of free fall for weakly gravitating bodies).
The Jordan frame action can be recast in a more convenient form by a conformal transformation to the ``Einstein frame'', i.e. by defining a new metric $g_{\mu\nu}^E$ and a new scalar $\varphi$
such that $g^E_{\mu\nu}=\phi\, g_{\mu\nu}$ and $({{\rm d}\log \phi}/{{\rm d}\varphi})^2={2\kappa}/[{3+2 \omega(\phi)}]$. This transformation casts the action \eqref{Jframe_action}
in the form
\begin{equation}
\label{einframe}\!
\!S=\!\!\int\!d^4 x \sqrt{-g^E} \left( \frac{R^E}{2\kappa}\!-\!\frac{1}{2}g_E^{\mu\nu} \partial_\mu\varphi \partial_\nu\varphi
\right) \! +S_M\!\!\left[\frac{g^E_{\mu\nu}}{\phi(\varphi)},\psi\right]\!\!\!
\end{equation}
where note that the matter degrees of freedom still couple only to the Jordan frame metric $g_{\mu\nu}=g^E_{\mu\nu}/\phi$ (i.e. test particles
follow geodesics of $g_{\mu\nu}$ and \textit{not} ones of $g^E_{\mu\nu}$), and that $g^E_{\mu\nu}$ and $\varphi$ are coupled minimally in the absence of matter
(which explains why using Einstein frame variables is advantageous).
By varying the Einstein frame action, one obtains the field equations
\begin{gather}\label{einstein}
G_{\mu\nu}^{E}=\kappa \left(T^\varphi_{\mu\nu} + T^E_{\mu\nu} \right),\\
\Box^E \varphi = \frac12
\frac{{\rm d}\log \phi}{{\rm d}\varphi} T_E\label{KG} \,,\\
\nabla_\mu^E T_E^{\mu\nu}=-\frac{1}{2} T_E
\frac{{\rm d}\log \phi}{{\rm d}\varphi}
g_E^{\mu\nu} \partial_\mu \varphi\,,\label{scalarT}
\end{gather}
where we assume that indices are raised and lowered with $g^E_{\mu\nu}$ and define $T_E\equiv T_E^{\mu\nu}g^E_{\mu\nu}$.
Note that we also define the stress energy tensors appearing in the field equations as
\begin{eqnarray}
T_E^{\mu\nu}&=&\frac{2}{\sqrt{-g^E}}\frac{\delta S_M}{\delta g^E_{\mu\nu}}=T^{\mu\nu}{\phi^{-3}} \, \\
T^\varphi_{\mu\nu} &=&\partial_\mu \varphi \partial_\nu \varphi- \frac{g^E_{\mu\nu}}{2} g_E^{\alpha\beta} \partial_\alpha \varphi
\partial_\beta \varphi\, \label{Tphi}
\end{eqnarray}
where $T^{\mu\nu}$ is the Jordan frame stress-energy tensor of all the matter degrees of freedom.
Solutions to Eqs.~\eqref{einstein}--\eqref{scalarT} for binary systems of compact objects (e.g. neutron stars or black holes) can be obtained numerically in the late
stages of the inspiral and during the merger~\cite{Barausse:2012da,Shibata:2013pra,nico}, but in order to describe more widely separated systems such as observed binary pulsars, it is more convenient to expand the field equations in
PN orders. In such a scheme, one approximates the two objects as point particles with masses $m_i$ and sensitivity parameters $s_i$~\cite{eardley1975} (or equivalently
scalar charges~\cite{Damour:1992we} $\alpha_i= - (2 s_i-1)/(3+2\omega_0)^{1/2}$, with $\omega_0$ the value of the function $\omega(\phi)$ far from the binary system). The sensitivities can be calculated
from isolated solutions for the compact objects, and depend on the ST theory and on the object's compactness (e.g. $s_i\approx 0$ for white dwarfs and less compact stars, $s_i=1/2$ for black holes, while
for neutron stars the sensitivity depends critically on the star's compactness and the ST theory under consideration). The binary's
dynamics is then expanded in orders of $v/c$ ($v$ being
the binary's relative velocity), and to 2.5 PN order the resulting equations take the schematic form~\cite{Mirshekari:2013vb,Damour:1996ke,Damour:1992we,Will:1989sk}
\begin{eqnarray}
\label{eqn:ST-2.5pn}
&&\frac{d^2 \mathbf{x}}{dt^2} = -\frac{G_{\rm eff} M}{r^2} \mathbf{n} \nonumber\\
&&+ \frac{G_{\rm eff} M}{r^2} \left[\left(\frac{\mathcal{A}_{PN}}{c^2}+\frac{\mathcal{A}_{2PN}}{c^4}\right)\mathbf{n}\
+ \left( \frac{\mathcal{B}_{PN}}{c^2} + \frac{\mathcal{B}_{2PN}}{c^4} \right) \dot{r} \bf{v} \right] \nonumber\\
&& + \frac{8}{5} \eta \frac{\left(G_{\rm eff} M\right)^2}{r^3} \left[\left( \frac{\mathcal{A}_{1.5PN}}{c^3} +\frac{ \mathcal{A}_{2.5PN}}{c^5}\right) \dot{r} \mathbf{n} \right. \nonumber\\
&& \left.- \left( \frac{\mathcal{B}_{1.5PN}}{c^3} + \frac{\mathcal{B}_{2.5PN}}{c^5}\right) \mathbf{v}\right]
\end{eqnarray}
where $\mathbf{x} = \mathbf{x}_1 - \mathbf{x}_2$ is the binary separation,
$r = |\mathbf{x}|$, $\mathbf{n} = \mathbf{x}/r$,
$\mathbf{v}=\mathbf{v}_1 - \mathbf{v}_2$ is the relative velocity, $\dot{r} = dr/dt$, $M=m_1+m_2$ is the total mass of the system, and
$\eta = (m_1 m_2)/M^2$ is the symmetric mass ratio.
The ``effective'' gravitational constant $G_{\rm eff}$ is
related to the gravitational constant $G_N$ measured locally (e.g. by a Cavendish-type experiment) by
$G_{\rm eff} \approx G_N ( 1+\alpha_1\alpha_2 )$.
Explicit expressions for the functions $\{\mathcal{A_I},\mathcal{B_I}\}$ are given
in Ref.~\cite{Mirshekari:2013vb} and also depend on the
sensitivities/scalar charges of the binary components, e.g. the presence of dissipative 1.5PN terms (which are
absent in GR) in Eq.~\eqref{eqn:ST-2.5pn} is due to the scalar charges, which source the emission of
dipolar gravitational radiation with energy flux
\begin{equation}\label{dipole_flux}
\dot{E}_{\rm dipole} \approx \frac{G_N}{3 c^3} \left(\frac{G_{\rm eff} m_1 m_2}{r^2}\right)^2 (\alpha_1-\alpha_2)^2\,.
\end{equation}
\subsubsection{Spontaneous, induced and dynamical scalarization}
From the dependence of the PN equations on the sensitivities/scalar charges, which in turn depend on the nature and compactness of the binary's components, it is clear that
the PN evolution of a compact-object binary depends on the nature of its components. Therefore, the strong equivalence principle, defined as the universality of free fall
for strongly-gravitating objects, is violated in ST theories already in the PN inspiral. Recently, however, Ref.~\cite{Barausse:2012da} highlighted the existence of other violations of the strong
equivalence principle in the last stages of the inspiral of binary neutron stars and for a particular class of ST theories. More specifically, Ref.~\cite{Barausse:2012da} considered theories with
$\omega(\phi)=-3/2-\kappa/(4\beta \log\phi)$ (or equivalently $\phi = \exp(-\beta \varphi^2)$), which are known~\cite{Damour:1992we,Damour:1993hw} to give rise
to the ``spontaneous scalarization'' of isolated neutron stars, i.e. allow for scalar charges $\alpha\sim 0.1-1$ for sufficiently
compact neutron stars and for marginally viable values of the theory's parameters $\tilde{\beta}=\beta/(4\pi G)\gtrsim -4.5$ and $\varphi_0\lesssim 10^{-2}$ ($\varphi_0$ being the scalar's value far from the system).\footnote{We stress that the observational constraint $\tilde{\beta}\gtrsim -4.5$ depends
somewhat on the equation of state of neutron stars. Indeed, binary-pulsar observations essentially rule out spontaneous scalarization
for the neutron-star masses corresponding to the components of observed binaries, placing constraints on $\tilde{\beta}$ \textit{once} an equation of state is assumed.
For instance, Ref.~\cite{Shibata:2013pra} shows that values of $\tilde{\beta}$ as low as $-5$ may be allowed with certain equations of state.
{Here, we follow Refs.~\cite{Barausse:2012da} and~\cite{Palenzuela:2013hsa}, and adopt
a polytropic equation of state with $K=123 G^3 M_\odot^2/c^6$
and $\Gamma=2$. Since this equation of state is not realistic but just a simple toy model, our results should be
interpreted as qualitative.}}
Ref.~\cite{Barausse:2012da} showed that at sufficiently small binary separations, a spontaneously scalarized star (thus bearing a significant scalar charge)
can excite a scalar charge in the other star, even if that star had not spontaneously scalarized and thus did not have a scalar charge to start with. This ``induced scalarization''
was shown to be capable of triggering earlier binary plunges and mergers relative to GR, an effect potentially observable with advanced GW detectors~\cite{Sampson:2014qqa}. Even more strikingly, Ref.~\cite{Barausse:2012da} showed that significant
scalar charges can be produced in the last inspiral stages of binary systems of unscalarized neutron stars, i.e. in binaries whose components have little or no scalar charges at large separations.
This ``dynamical scalarization'' produces a sudden build-up of the scalar charges at small separations, quickly triggering a plunge/merger at frequencies within the reach of advanced GW detectors (c.f. Ref.~\cite{Sampson:2014qqa,Taniguchi:2014fqa}
for the detectability of this effect with GW detectors). It is also important to note that dynamical scalarization, being
an effect that turns on at small separations, can evade (to a certain extent) the constraints posed
by binary-pulsar observations, which only exclude the presence of \textit{spontaneously} scalarized stars in widely-separated binaries and for the specific observed values
of the neutron-star masses. In particular, dynamical scalarization may happen for values of $\tilde{\beta}$ between $-4.3$ and $-4.5$ (or lower), a window still allowed by binary-pulsar observations.
Remarkably, even though both induced and dynamical scalarization are strongly non-linear effects, they can still be understood and reproduced in their main qualitative features by a minimal modification of the PN expansion
scheme outline above. More precisely, Ref.~\cite{Palenzuela:2013hsa} describes the evolution of a neutron-star binary system by the PN equations of motion \eqref{eqn:ST-2.5pn}, but introduces a new way of computing
the sensitivities or scalar charges that appear in those equations. Instead of computing those parameters from isolated neutron-star solutions, as done in the ``classic'' PN scheme, Ref.~\cite{Palenzuela:2013hsa} introduced a
formalism that includes the interaction between the two stars in the calculation of the scalar charges, thus accounting for both induced and dynamical scalarization. In practice, this scheme starts from the
standard calculation of the scalar charges for the binary components
in isolation, and then uses that calculation to define a system of non-linear algebraic equations, which can be solved iteratively at each step of the PN orbital evolution to yield the charges of
both stars including non-linear effects. This procedure was validated by
comparing it with the fully non-linear simulations of Ref.~\cite{Barausse:2012da}.
\subsubsection{Electromagnetic coupling in ST theories}
One purpose of this paper is to extend the formalism of Ref.~\cite{Palenzuela:2013hsa} to include the effect of an electromagnetic field.
Let us consider a binary system of magnetized neutron stars surrounded by a plasma in ST theories. Solving the full non-linear problem in the Jordan frame would require solving the curved spacetime Maxwell
equations
\begin{gather}
\partial_{[\mu} F_{\nu\gamma]}=0\,,\label{max1}\\
\nabla_\nu F^{\nu \mu} = j^\mu\,,\label{max2}
\end{gather}
and including the stress-energy tensor of the electromagnetic field in the source of the Einstein equations. (Note that because of the weak equivalence principle, which is reflected
in the structure of the matter action written in Eq.~\eqref{Jframe_action}, the electromagnetic field only couples to the Jordan frame metric and to the plasma's electric charges, and not directly to the scalar field.)
Defining the Einstein-frame electromagnetic tensor $F^E_{\mu\nu}=F_{\mu\nu}$ and the plasma's current $j^\mu_E=j^\mu/\phi^2$, the Einstein-frame Maxwell equations take the same form as in the Jordan frame, i.e.
\begin{gather}
\partial_{[\mu} F^E_{\nu\gamma]}=0\label{m1}\,,\\
\nabla^E_\nu F_E^{\nu \mu} = j_E^\mu\,,\label{m2}
\end{gather}
and the Einstein-frame field equations \eqref{einstein}--\eqref{scalarT} become
\begin{gather}\label{einsteinBis}
G_{\mu\nu}^{E}=\kappa \left(T^\varphi_{\mu\nu}+ T^{{\rm em},E}_{\mu\nu}+ T^{{\rm pl},E}_{\mu\nu}+ T^E_{\mu\nu} \right),\\
\Box^E \varphi = \frac12
\frac{{\rm d}\log \phi}{{\rm d}\varphi} (T_E+T_E^{\rm pl})\label{KGbis} \,,\\
\nabla_\mu^E (T_E^{\mu\nu}+T_{{\rm pl},E}^{\mu\nu}) =F^\nu_{E\mu}j_E^\mu-\frac{1}{2} (T_E+T_E^{\rm pl})
\frac{{\rm d}\log \phi}{{\rm d}\varphi}
g_E^{\mu\nu} \partial_\mu \varphi\,,\label{scalarTbis}
\end{gather}
where $T_{{\rm em},E}^{\mu\nu}=T_{\rm em}^{\mu\nu}{\phi^{-3}}$ is the Einstein-frame stress-energy tensor of the electromagnetic field~\footnote{We recall that the explicit form of the stress-energy tensor of the electromagnetic
field in the Jordan frame is
\begin{equation*}
T_{\rm em}^{\mu\nu}=\frac{1}{4 \pi} \left(F^{\mu\alpha} F^{\nu}_{\phantom{a}\alpha}-\frac14 g^{\mu\nu} F_{\alpha\beta}F^{\alpha\beta}\right)\,.
\end{equation*}
Similarly, the Einstein-frame stress energy tensor is given by
\begin{equation*}
T_{\rm em,\,E}^{\mu\nu}=\frac{1}{4 \pi} \left(F_E^{\mu\alpha} F^{\nu}_{E\,\alpha}-\frac14 g_E^{\mu\nu} F^E_{\alpha\beta}F_E^{\alpha\beta}\right)\,,
\end{equation*}
which satisfies indeed $T_{{\rm em},E}^{\mu\nu}=T_{\rm em}^{\mu\nu}{\phi^{-3}}$.},
while $T_{E}^{\mu\nu}$, $T_{{\rm pl},E}^{\mu\nu}$
and $T_{{\rm em},E}^{\mu\nu}$ are those of
the neutron-star matter, of the matter of the plasma and of the electromagnetic field.
(Again, all indices are raised and lowered with the Einstein-frame metric.)
As discussed below, the stress energy of the plasma matter will be negligible
for our cases of interest, but for the moment
we keep it to make our treatment general and more clear.
Note that in deriving these equations we have used
\begin{equation}\label{emCons}
\nabla_\mu^E T_{{\rm em},E}^{\mu\nu}=-F^\nu_{E\mu}j_E^\mu
\end{equation}
(which follows from the Maxwell equations), and the fact that the trace of $T_{{\rm em},E}^{\mu\nu}$ is
zero. Note also that Eq.~\eqref{KGbis} can be recast in the form
\begin{equation}
\nabla^E_\mu T_\varphi^{\mu\nu}= \frac12
\frac{{\rm d}\log \phi}{{\rm d}\varphi} (T_E^{\rm pl}+T_E) g_E^{\mu\nu} \partial_\mu \varphi\,,
\end{equation}
which shows that in the Einstein frame (as in the Jordan frame) there is no direct energy or momentum transfer from the scalar field to the electromagnetic field.
Postponing the solution to the full non-linear problem to future work, let us note that for astrophysically realistic systems, the plasma and the electromagnetic field
in the magnetosphere are too small to significantly backreact on the metric and on the binary and scalar field evolution (``force-free approximation'', c.f. discussion in the next section),
i.e. to lowest order Eqs.~\eqref{einsteinBis}--\eqref{scalarTbis} reduce to Eqs.~\eqref{einstein}--\eqref{scalarT}. To next order, by combining Eqs.~\eqref{scalarTbis} and \eqref{scalarT} one then obtains
\begin{equation}\label{plasmaMotion}
\nabla_\mu^E T_{{\rm pl},E}^{\mu\nu} =F^\nu_{E\mu}j_E^\mu-\frac{1}{2} T_E^{\rm pl}
\frac{{\rm d}\log \phi}{{\rm d}\varphi}g_E^{\mu\nu} \partial_\mu \varphi \approx F^\nu_{E\mu}j_E^\mu
\end{equation}
where we have used the fact that $T_E^{\rm pl}\approx 0$ for the plasma (because the particles of which it is made typically move at speeds close to the speed of light)
and in any case ${{\rm d}\log \phi}/{{\rm d}\varphi}\approx0$ outside the neutron stars (where the plasma moves) because $\varphi$ is small there. Equation~\eqref{plasmaMotion} regulates the motion
of the plasma, while the electromagnetic field satisfies the Maxwell equations \eqref{m1}--\eqref{m2}. In both Eq.~\eqref{plasmaMotion} and Eqs.~\eqref{m1}--\eqref{m2},
the metric is determined by the evolution of the binary and scalar field alone [i.e.~by Eqs.~\eqref{einstein}--\eqref{scalarT}].
In practice, as can be seen from Eq.~\eqref{Tphi}, the scalar field stress-energy vanishes at linear order in the field's perturbation over a constant background, and is thus negligible
outside the neutron stars (although it is \text{not} always negligible inside the stars, where it can grow non-linear and give rise to scalar charges $\alpha\sim 0.1-1$ in scalarized systems).
Therefore, it is natural to approximate the metric outside the neutron stars with the \textit{general-relativistic} PN metric of two point particles (representing the neutron stars),
whose trajectories are calculated (\textit{including} the effect of the scalar charges) with the formalism of Ref.~\cite{Palenzuela:2013hsa}.
Therefore, because of Eqs.~\eqref{m1}, \eqref{m2} and \eqref{plasmaMotion}, the calculation of the electromagnetic fluxes can proceed as in GR, except for the modified binary trajectory. We will present a standard GR approximate strategy to
calculate such fluxes (the ``unipolar inductor'' model) in the next section.
\subsection{Magnetosphere and plasma treatment}
\label{sec:magnetospherePlasma}
As discussed in Ref.~\cite{1969ApJ...157..869G}, neutron stars are surrounded by a magnetosphere with
a plasma density $\rho^{pl} \simeq -\vec{\Omega}\cdot\vec{B}/(2\pi c)$, where $\vec{\Omega}$ represents the
rotational frequency of the plasma, $\vec{B}$ is the magnetic field present in the region, and $c$ is the speed of light.
The interaction of a rotating magnetized star with
its own magnetosphere is responsible for electromagnetic emissions in pulsars. The analysis of such an interaction
is a delicate subject, because the plasma dynamics may be intricate, and complex simulations are typically required to
fully capture its behavior. Fortunately, a useful approximation can be adopted that captures important
aspects of the system. This approximation relies on the observation that in the magnetosphere region
the inertia of the plasma is negligible with respect to the electromagnetic energy density, i.e. $T_{\mu\nu}^{pl} \ll T_{\mu\nu}^{em}$.
Through Eqs.~(\ref{emCons}) and~(\ref{plasmaMotion}), this in turn implies
$F^{\nu}_{E\mu} j^{\mu}_E\approx F^{\nu}_{\phantom{E}\mu} j^{\mu}\approx0\approx\nabla_{\mu}^{E} T_{em,E}^{\mu \nu}\approx\nabla_{\mu} T_{em}^{\mu \nu}$,
where in the last passage we have exploited the fact that Eq.~(\ref{emCons}) also holds in the Jordan frame (as it follows directly from the Maxwell equations~\eqref{max1} and \eqref{max2}).\footnote{One
can also show directly that $\nabla^E_\mu T^{\mu}_{em,E\nu}=\nabla_\mu T^{\mu}_{em\,\nu}/\phi^2$, using the conformal transformation between the Einstein and Jordan frames.}
These are known as the force-free conditions for the plasma~\cite{1969ApJ...156...59G,1977MNRAS.179..433B,2002MNRAS.336..759K}, and the resulting
electrodynamics equations,
while simpler to deal with as now one only needs to consider the behavior of electromagnetic
fields constrained by the force-free condition, still represent a non-linear coupled system of partial differential equations.
The electrodynamics equations are then augmented by those describing the dynamical behavior of the spacetime and the neutron-star matter, thus
complex and time-consuming simulations are typically needed to study the system's evolution. Nevertheless,
for the scenario of interest here -- i.e. the late stages of a magnetized binary merger -- and for the purpose of our work, we can make
use of a hybrid approach, combining the formalism of Ref.~\cite{Palenzuela:2013hsa} described above (whereby the neutron stars are treated as point-like objects
satisfying ordinary differential equations of motion that incorporate the relevant gravitational and scalarization effects) together
with a \textit{unipolar model} to account for magnetospheric effects. Both these approaches are supported by
simulations of the complete problem in the context of neutron star mergers (for ST theories and non-magnetized systems~\cite{Palenzuela:2013hsa}) and
binary neutron star and black hole-neutron star mergers in the context of magnetosphere interactions
in GR~\cite{Ponce:2014sza,Palenzuela:2013kra,Palenzuela:2013hu,Paschalidis:2013jsa}. As argued
in the previous section, a reliable model accounting for magnetosphere interactions within GR
should also be applicable to the case of ST theories of gravity. In what follows, we therefore describe the main aspects of
the unipolar model.
\subsubsection*{Magnetosphere interactions and luminosity}
\label{sec:magnetosphereInteractions}
A useful model to estimate the electromagnetic energy radiated by a
magnetized neutron star binary is based on the \textit{unipolar inductor}~\cite{1969ApJ...156...59G,Hansen:2000am}.
Such a model has recently been further analyzed for neutron star binaries in Ref.~\cite{Lai:2012qe} and confronted with
fully dynamical simulations including plasma effects, finding good agreement in the
obtained luminosity~\cite{Palenzuela:2013kra,Palenzuela:2013hu,Ponce:2014sza}. Similar conclusions
have been obtained in the model's application to black hole-neutron star binaries~\cite{2011ApJ...742...90M,Paschalidis:2013jsa}.
We next describe briefly the main ingredients required by this model for our purposes.
We assume that both stars are magnetized, their magnetic fields are
dipolar and are non-spinning\footnote{Spins introduce only minor modifications --as tidal locking can not occur--
and thus will not affect the conclusions of this work.}. Further, we assume that one star has a larger magnetization
than its companion, and study the electromagnetic radiation due to the interaction
of their magnetospheres. As the orbit tightens, as described
in section \ref{subsec:STdynamics}, ST effects will cause deviations from the GR orbital behavior, which induce a stronger Poynting flux.
It is important to stress here that for realistic magnetic field strengths,
electromagnetic effects do not backreact on the orbital evolution of the binary~\cite{Ioka:2000yb,Anderson:2008zp}.
Under such assumptions, one can estimate the luminosity of the binary~\cite{Lai:2012qe} as:
\begin{multline}
\label{eq:lums-vrel_sepn}
\mathcal{L} \approx 10^{38} \left(\frac{v_{rel}}{c}\right)^2
\left(\frac{B_*}{10^{11}{\rm G}}\right)^2
\left(\frac{R_*}{10{\rm km}}\right)^6\\\times \left(\frac{R_{\rm eff}}{10{\rm km}}\right)^2
\left(\frac{r}{100{\rm km}}\right)^{-6}
{\rm erg/s}
\end{multline}
where $v_{rel}$ is the relative velocity of the binary (obtained with the PN expansion
as discussed in~\cite{Palenzuela:2013hsa}),
$B_*$ and $R_*$ are the magnetic field and the radius of the primary star,
and $r$ is the separation between the stars.
Induction takes place on the secondary star at a radius $R_{\rm eff}$, which is equal
to the star's radius $R_c$ when the secondary is unmagnetized. Otherwise,
$R_{\rm eff}$ is larger, as the {secondary's field} shields a region around it. We account
for this effect by defining (see Refs.~\cite{2012PhRvD..86j4035L,Palenzuela:2013hu,Palenzuela:2013kra}),
\begin{equation}
\label{eq:effRadius}
R_{\rm eff} = \max \left( r\left(\frac{B_c}{B_*}\right)^{1/3} , R_c\right),
\end{equation}
{where $B_c$ is the magnetic field of the secondary}.
Naturally, this effect is relevant at large separations, while for separations
$r \leq R_c \left(\frac{B_c}{B_*}\right)^{-1/3}$ the effective radius reduces to the star's radius $R_c$.
Within GR calculations of the quasi-adiabatic regime of binary neutron star systems (i.e. at large separations),
estimates have been obtained from Eq.~(\ref{eq:lums-vrel_sepn}) by replacing $v_{rel}$ with its Keplerian
expression and $R_{\rm eff}=R_c$, giving rise to a dependence $\mathcal{L} \simeq r^{-7}$
(e.g. Refs.~\cite{Hansen:2000am,Lai:2012qe}). When the shielding effect given by Eq.~(\ref{eq:effRadius})
is considered, the luminosity follows a softer dependence $\mathcal{L} \simeq r^{-5}$
~\cite{Palenzuela:2013kra,Ponce:2014sza}. However, in ST theories, deviations from
Keplerian motion are possible, and we thus employ both Eqs.~(\ref{eq:lums-vrel_sepn}) and (\ref{eq:effRadius})
in our calculations.
It is important at this point to stress two limitations of our model.
First, some mechanism must act to convert the obtained Poynting flux to observable radiation, and this will
involve some conversion efficiency. This work does not address this issue, but rather focuses on
estimating the Poynting flux alone. The fact that the system that we study shares many common features with
pulsars, where observable radiation in multiple bands is observed, gives us some degree of confidence that
some energy conversion to observable radiation will take place~\cite{Palenzuela:2013kra}.
Second, this work concentrates on possible emissions prior to the merger. As the merger takes
place, the two stars will become tidally disrupted, merge into a hypermasive neutron star and
possibly form an accreting black hole. This rich dynamics will naturally have strong associated
luminosities, which are not accounted for here, as we concentrate on the pre-merger stage.
\section{Results}
\label{sec:Results}
\subsection{Quasi-circular case}
We first study binary systems in quasi-circular (i.e. zero eccentricity) orbits and
consider four different sets of masses. These configurations are chosen so that
they undergo at least one of the key scalarization processes
described in section~\ref{subsec:STdynamics}, {while
still being consistent with available solar-system and binary-pulsar data.}
More specifically, the configurations we consider are:
\begin{itemize}
\item case {\bf LE}, with low- and equal-mass stars ($M_1 = M_2 = 1.41M_\odot$),
which undergo dynamical scalarization but do not produce dipolar radiation.
\item case {\bf HE}, with high- and equal-masses stars ($M_1 = M_2 = 1.74M_\odot$),
which undergo spontaneous scalarization for $\tilde{\beta}=-4.5$ and dynamical scalarization for $\tilde{\beta}=-4.2$,
but do not produce dipolar radiation.
\item case {\bf LU}, with low- and unequal-mass stars ($M_1 = 1.41 M_\odot$, $M_2 = 1.64 M_\odot$), which
undergo dynamical as well as induced scalarization, and produce dipolar radiation.
\item case {\bf HU}, with high- and unequal-mass stars ($M_1 = 1.52 M_\odot$, $M_2 = 1.74 M_\odot$), which
undergo dynamical and induced scalarization (in the lower-mass star) and
spontaneous scalarization (in the higher-mass star), and produce dipolar radiation.
\end{itemize}
Additionally, we consider different possible magnetizations of
each star, and examine the characteristics of the resulting electromagnetic luminosity.
Henceforth, we will refer to the primary/companion star as the more/less massive one and, in the case of
equal mass configurations, as the more/less strongly magnetized one.
To adopt realistic configurations, we recall that
the standard formation channel of neutron-star binaries indicates that
the most likely configurations involve a magnetically
dominant (primary) star with a significantly less magnetized companion
(secondary)~\cite{2001ApJ...557..958C,lrr-2008-8,lrr-2006-6}.
To explore a range of possible options, we consider three ratios of the magnetizations between the stars,
namely $b \equiv B_c/B_* = (0.1, 0.01, 0.001)$.
Also, for simplicity we assume that the stars' magnetic dipoles are aligned with the orbital angular momentum.
Notice that
this is not a restrictive assumption,
as it yields reasonably good estimates for the expected power in more general configurations~\cite{Ponce:2014sza}.
Finally, we restrict our analysis to the ST theories
that yield the largest differences in the binary dynamics, while satisfying existing experimental constraints, i.e.
we take the coupling parameter of the ST theory to be $\tilde{\beta} = -4.5$ or $\tilde{\beta}=-4.2$ (so as to satisfy binary pulsar constraints
\footnote{{Note however, as already mentioned,
that the viable range for $\beta$ depends slightly on the equation of state adopted for the neutron stars.}}), and we adopt
a small value for the asymptotic value of the scalar field $\varphi_0=10^{-5}$ to pass solar system tests.
For comparison purposes, we also include the
corresponding GR results.
The list of the cases considered, as well as a summary of the main
results (e.g., the total radiated electromagnetic energy for each case),
is given in Table~\ref{table:cases_results}.
\begin{table*}[!]
\begin{tabular}{c||c|c|c|c |c}
\hline
& Masses & Magnetic field & & \multicolumn{2}{c}{Total $E_{rad}$ [erg]} \\
Case &(in $M_\odot$) & ratio, $b \equiv B_c/B_*$ & Theory & at 35 km & at 30 km \\
\hline\hline
low equal-mass
& $M_1 = 1.41 $ & $0.1$ &
& 1.74 $\times10^{41}$
& (1.90,1.93,1.93)$\times10^{41}$
\\
(\bf{LE}) & $M_2 = 1.41 $ & $0.01$ & $\tilde{\beta}=(-4.5, -4.2$); GR
& 4.02$\times10^{40}$
& (4.72,4.83,4.83)$\times10^{40}$
\\
& & $0.001$ &
& 2.31$\times10^{40}$
& (3.02,3.12,3.12)$\times10^{40}$
\\
\hline
high equal-mass
& $M_1 = 1.74 $ & $ 0.1$ &
& (1.27,1.49,1.55)$\times10^{41}$
& (1.44,1.66,1.78)$\times10^{41}$
\\
(\bf{HE}) & $M_2 = 1.74$ & $ 0.01$ & $\tilde{\beta}=(-4.5, -4.2)$; GR
& (3.06,3.61,3.87)$\times10^{40}$
& (4.24,4.78,5.42)$\times10^{40}$
\\
& & $ 0.001$ &
& (1.74,2.19,2.45)$\times10^{40}$
& (2.92,3.37,4.01)$\times10^{40}$
\\
\hline
low unequal-mass
& $M_1 = 1.41 $ & $ 0.1$ &
& $(1.18,1.29,1.41)\times10^{41}$
& $(1.27,1.49,1.55)\times10^{41}$
\\
(\bf{LU}) & $M_2 = 1.64 $ & $ 0.01$ & $\tilde{\beta}=(-4.5, -4.2)$; GR
& $(2.65,3.16,3.24)\times10^{40}$
& $(3.07,3.61,3.87)\times10^{40}$
\\
& & $0.001$ &
& $(1.32,1.74,1.83)\times10^{40}$
& $(1.73,2.19,2.45)\times10^{40}$
\\
\hline
high unequal-mass
& $M_1 = 1.52 $ & $ 0.1$ &
& $(0.913,1.05,1.18)\times10^{41}$
& $(0.996,1.13,1.3)\times10^{41}$
\\
(\bf{HU}) & $M_2 = 1.74 $ & $0.01$ & $\tilde{\beta}=(-4.5, -4.2)$; GR
& $(2.07,2.37,2.7)\times10^{40}$
& $(2.45,2.75,3.22)\times10^{40}$
\\
& & $ 0.001$ &
& $(1.06,1.2,1.5)\times10^{40}$
& $(1.43,1.58,2.02)\times10^{40}$
\\
\hline\hline
\end{tabular}
\caption{Quasi-circular cases studied in this work.
The last two columns show the total electromagnetic energy radiated ($E_{rad} = \int \mathcal{L} dt$)
for binaries starting at an initial separation of 180 km apart,
until a separation of 35 and 30 km respectively.
}
\label{table:cases_results}
\end{table*}
As mentioned above, the binary dynamics in the ST theories that we study can show clear departures from the GR behavior.
In particular, the binary's orbital frequency/separation can increase faster than in GR
for high neutron-star masses and low values of $\tilde{\beta}$, because of the enhanced gravitational attraction due to scalar effects
and the possible dipolar emission of scalar waves. { Such behavior is illustrated in
Fig.~\ref{fig:sepn-time}, which shows the separation for the four binaries considered, in
ST theories with $\tilde{\beta} =-4.5$ and $\tilde{\beta}=-4.2$, as well as in GR, as a function of
the time remaining until the merger. (The merger is defined
as the time at which the separation equals $R_*+R_c$). As illustrated in this figure,
in the ST theories that we consider the time to merger from a given separation is
always either shorter or equal than in GR. As we discuss next, in the former case differences in the electromagnetic flux of energy arise.}
The complementary Fig.~\ref{fig:thetadot-sepn} shows how all binary evolutions cover approximately the same frequency range,
although for higher masses and lower values of $\tilde{\beta}$, any given frequency is achieved at a larger separation.
As discussed, these effects are caused
by the scalar charges (or equivalently the sensitivities) that each star
acquires as the orbits tightens. The behavior of the scalar charge with respect to the orbital frequency for
each case considered is shown in Fig.~\ref{fig:ScChg-sepn_thetadot},
where the orbital frequency is determined instantaneously (i.e. as the derivative of the
azimuthal coordinate).
From this figure, it is clear that scalarization effects are quite weak in the
\textbf{LE} case, because the scalar charges remain negligible for most of the evolution and only rise
to $\alpha\sim {\cal O}(1)$ at relatively short separations (i.e. high frequencies).
In particular, the scalar charges remain $\lesssim 10^{-3}$ essentially until
the final plunge toward merger.
This behavior is in contrast with the remaining cases, where scalarization effects
are considerably stronger at the earlier stages, and therefore have a clear impact on
the dynamics, especially for $\tilde{\beta}=-4.5$.
Such behavior comes about because in the {\bf HE}, {\bf LU} and {\bf HU} cases the more massive star either spontaneously scalarizes
already in isolation,
which induces a time-dependent running of the scalar charge of the companion (induced scalarization), or both stars scalarize dynamically in the late inspiral
(after which the scalar charges further grow by mutual induced scalarization). As mentioned, as the scalar charges grow,
the gravitational attraction between the stars gets stronger than in GR, and for unequal-mass binaries the system
also emits dipolar radiation, clearly affecting the binary's dynamics.
\begin{figure*}[!]
\includegraphics[width=0.4\textwidth]{{{./sepn-time_14-14-loglog}}}
\includegraphics[width=0.4\textwidth]{{{./sepn-time_174-174-loglog}}}
\\
\vspace{2mm}
\includegraphics[width=0.4\textwidth]{{{./sepn-time_141-164-loglog}}}
\includegraphics[width=0.4\textwidth]{{{./sepn-time_152-174-loglog}}}
\caption{
Separation shown as a function of $t_{\rm{merger}} -t $ in log-scale,
being $t_{\rm{merger}}$ the time at which
the stars come into contact (i.e. the separation equals $R_*+R_c$).
The panels display: the \textbf{LE} (top-left panel), \textbf{HE} (top-right panel),
\textbf{LU} (bottom-left panel) and \textbf{HU} (bottom-right panel) binaries,
for $\tilde{\beta}=-4.5$ (dotted lines), $-4.2$ (dashed lines) and GR (solid lines).
Recall that magnetic field effects on the binary's motion are negligible,
thus the differences in the trajectories are solely due to the underlying gravity theory.
In the \textbf{LE} binary, the trajectories for the three cases are almost identical, as
there is almost no scalarization until very late in the inspiral.
{These examples show that at any given separation, the time to merger is shorter (or equal) in ST
theories than in GR.}
}
\label{fig:sepn-time}
\end{figure*}
\begin{figure*}
\includegraphics[width=0.4\textwidth]{{{./thetadot-sepn_14-14}}}
~
\includegraphics[width=0.4\textwidth]{{{./thetadot-sepn_174-174}}}
\\
\vspace{2mm}
\includegraphics[width=0.4\textwidth]{{{./thetadot-sepn_141-164}}}
~
\includegraphics[width=0.4\textwidth]{{{./thetadot-sepn_152-174}}}
\caption{Orbital frequency as a function of separation (in log-scale), for the same parameters shown in Fig.~\ref{fig:sepn-time}.
}
\label{fig:thetadot-sepn}
\end{figure*}
\begin{figure*}[!]
\includegraphics[width=0.4\textwidth]{{{./ScChg-thetadot_14-14}}}
~\includegraphics[width=0.4\textwidth]{{{./ScChg-thetadot_174-174}}}
\\
\vspace{2mm}
\includegraphics[width=0.4\textwidth]{{{./ScChg-thetadot_141-164}}}
~\includegraphics[width=0.4\textwidth]{{{./ScChg-thetadot_152-174}}}
\caption{Scalar charge for the same binaries as in Fig.~\ref{fig:sepn-time},
as a function of the orbital frequency,
for ${\tilde{\beta}}=-4.5$ (dotted lines) and $-4.2$ (dashed lines).
The top panels correspond to equal-mass binaries, and both stars have
the same scalar charge for a given value of ${\tilde{\beta}}$.
On the other hand, the bottom panels correspond to unequal-mass binaries, and
the scalar charges are different.
In these cases in particular, the upper line with the same style
(i.e. the same value of ${\tilde{\beta}}$) refers to the more massive star.
}
\label{fig:ScChg-sepn_thetadot}
\end{figure*}
The dynamical behavior that we have just described has direct consequences on the Poynting flux produced by
the system as the stars' magnetospheres interact. As described in Sec.~\ref{sec:magnetospherePlasma},
we estimate such flux via the enhanced unipolar inductor model, which depends on the orbital evolution as well
as on the magnetization of the binary components. The results for the luminosity
are displayed in Figs.~\ref{fig:lum-time} and~\ref{fig:lum-freq} as a function
of time and orbital frequency.
At a broad level, since all cases cover similar frequency ranges and separations -- from $\approx 180$ km to merger for the cases considered --
all the obtained luminosities are comparable and of the order of $10^{39}$--$10^{41} \mbox{erg/s}$
(the precise value depending on the binary's mass and the magnetic field ratio between the stars)
for a primary with magnetic field of $10^{11}\mbox{G}$.
A closer inspection, e.g.~of Fig.~\ref{fig:lum-time}, shows that ST effects are evident for the more massive
binaries and for lower values of $\tilde{\beta}$, especially near the merger.
This behavior is also visible
in the luminosity rate of change with time, as illustrated in Fig.~\ref{fig:dLdt-time}. As can be seen,
for higher masses and lower values of $\tilde{\beta}$, the dynamics proceeds at a faster pace in the ST theories under consideration than in GR,
thus inducing a less powerful electromagnetic flux {at a given time to merger (since for a given $t_{\rm merger}-t$, the binary's separation is larger
in the ST theories that we consider than in GR, c.f. Fig.~\ref{fig:sepn-time})}. This conclusion is further supported by
the {total radiated electromagnetic energy, as a function of the time elapsed from an initial separation of 180 km (Fig.~\ref{fig:Erad-time}). }
As time progresses, binaries governed by the ST theories show a clear departure from the general-relativistic behavior.
To further illustrate the differences in luminosities obtained, Fig.~\ref{fig:FracChg_lum-freq}
presents the \textit{fractional change in luminosity},
$FCL \equiv {|\mathcal{L}_{\rm ST}-\mathcal{L}_{\rm GR}|}/{\mathcal{L}_{\rm GR}}$. (For concreteness we adopt $b=0.1$, as
all other values of $b$ display a similar behavior.)
As can be seen, the \textbf{LE} case displays at most a $5\%$ difference, while the other ones
(\textbf{LU}, \textbf{HE} and \textbf{HU}) show departures of up to 40\% away from the GR prediction. Importantly, the relative differences display a frequency
dependent behavior, growing strongly toward higher frequencies.
\begin{figure*}[!]
\includegraphics[width=0.4\textwidth]{{{./beta_-4.5-4.2-GR--1.41-1.41--time}}}
\includegraphics[width=0.4\textwidth]{{{./beta_-4.5-4.2-GR--1.74-1.74--time}}}
\\
\vspace{2mm}
\includegraphics[width=0.4\textwidth]{{{./beta_-4.5-4.2-GR--1.41-1.64--time}}}
\includegraphics[width=0.4\textwidth]{{{./beta_-4.5-4.2-GR--1.52-1.74--time}}}
\caption{Luminosity vs. time to merger, for the same parameters shown in Fig.~\ref{fig:sepn-time}.
Different lines in each panel indicate the \textit{ratio} between
the secondary and primary magnetic fields (from top to bottom):
$b=0.1$ (in red), $b=0.01$ (in green), $b=0.001$ (in blue)
and a non-magnetized secondary (i.e. $b=0$, in black).
}
\label{fig:lum-time}
\end{figure*}
\begin{figure*}[!]
\includegraphics[width=0.4\textwidth]{{{./beta_-4.5-4.2-GR--1.41-1.41--freq_linInset}}}
\includegraphics[width=0.4\textwidth]{{{./beta_-4.5-4.2-GR--1.74-1.74--freq_linInset}}}
\\
\vspace{2mm}
\includegraphics[width=0.4\textwidth]{{{./beta_-4.5-4.2-GR--1.41-1.64--freq_linInset}}}
\includegraphics[width=0.4\textwidth]{{{./beta_-4.5-4.2-GR--1.52-1.74--freq_linInset}}}
\caption{Luminosity as a function of orbital frequency, for the same parameters shown in Fig.~\ref{fig:lum-time}.
}
\label{fig:lum-freq}
\end{figure*}
\begin{figure*}[!]
\includegraphics[width=0.4\textwidth]{{{./dLdt_avg--beta_-4.5-4.2-GR--1.41-1.41--time}}}
~\includegraphics[width=0.4\textwidth]{{{./dLdt_avg--beta_-4.5-4.2-GR--1.74-1.74--time}}}
\\
\vspace{1mm}
\includegraphics[width=0.4\textwidth]{{{./dLdt_avg--beta_-4.5-4.2-GR--1.41-1.64--time}}}
~\includegraphics[width=0.4\textwidth]{{{./dLdt_avg--beta_-4.5-4.2-GR--1.52-1.74--time}}}
\caption{Time derivative of the luminosity as a function of time to merger,
for the same parameters shown in Fig.~\ref{fig:lum-time}.
Note that the differences are less significative for the low-mass binaries.
}
\label{fig:dLdt-time}
\end{figure*}
\begin{figure*}
\includegraphics[width=0.4\textwidth]{{{./intLum--beta_-4.5-4.2-GR--1.41-1.41--time_wINset}}}
~\includegraphics[width=0.4\textwidth]{{{./intLum--beta_-4.5-4.2-GR--1.74-1.74--time_wINset}}}
\\
\vspace{2mm}
\includegraphics[width=0.4\textwidth]{{{./intLum--beta_-4.5-4.2-GR--1.41-1.64--time_wINset}}}
~\includegraphics[width=0.4\textwidth]{{{./intLum--beta_-4.5-4.2-GR--1.52-1.74--time_wINset}}}
\caption{
Total energy radiated in electromagnetic waves (i.e. the time integral of the luminosity)
as a function of time elapsed from an initial separation of approximately 180 km.
for the same parameters shown in Fig.~\ref{fig:lum-time}.
Insets show zoom-ins of the total energy radiated at late times, when the binary is close to the merger.
}
\label{fig:Erad-time}
\end{figure*}
\begin{figure*}[!]
\includegraphics[width=0.4\textwidth]{{{./FrcChg_LE}}}
~
\includegraphics[width=0.4\textwidth]{{{./FrcChg_HE}}}
\\
\vspace{2mm}
\includegraphics[width=0.4\textwidth]{{{./FrcChg_LU}}}
~
\includegraphics[width=0.4\textwidth]{{{./FrcChg_HU}}}
\caption{\textit{Fractional change} of the luminosity,
$FCL \equiv {|\mathcal{L}_{\rm ST}-\mathcal{L}_{\rm GR}|}/{\mathcal{L}_{\rm GR}}$,
for the \textbf{LE} (top-left panel), \textbf{HE} (top-right panel),
\textbf{LU} (bottom-left panel) and \textbf{HU} (bottom-right panel) binaries
as a function of orbital and GW frequencies (assuming $b = 0.1$ for concreteness and
clarity's sake).
The solid black line shows the $FCL$ of a ST with $\tilde{\beta}=-4.5$ with respect to GR,
while the dashed red line represents the $FCL$ of a ST with $\tilde{\beta}=-4.2$ (again with respect to GR).
}
\label{fig:FracChg_lum-freq}
\end{figure*}
\subsection{Eccentric binaries}
We now turn our attention to binaries with non-negligible eccentricity. Such configurations
are not expected to represent a large fraction of the sources detectable by advanced gravitational-wave detectors,
as gravitational emission tends to circularize binaries by the time they enter the detector's sensitive band.
Nevertheless, as already illustrated e.g.~in Ref.~\cite{Healy:2009zm,Gold:2011df,East:2012ww}, eccentric binaries provide
an excellent laboratory to test diverse and extreme physics. As we show here, this is also
the case for possible electromagnetic counterparts driven by magnetosphere interactions in non-GR gravity theories. To illustrate
this point, we here consider the following two cases:
an equal-mass binary with low masses $M_1 = M_2 = 1.52 M_\odot$,
and an unequal-mass binary with $M_1 =1.52 M_\odot$ and $M_2 = 1.74 M_\odot$.
Following Ref.~\cite{East:2012ww}, we study the dynamics and the emitted electromagnetic flux,
in two different configurations:
``mild-eccentricity'' orbits, where the initial apastron/periastron are at separations of
about 220 km/80 km respectively (i.e. $e \sim 0.47$);
and,
``high-eccentricity'' orbits, where the initial apastron/periastron are at separations of
about 220 km/50 km respectively (i.e. $e \sim 0.63$).
These eccentric binaries reveal another interesting phenomenon allowed in the ST theories under study, namely
a sequence of successive ``scalarization/descalarization'' cycles~\cite{Palenzuela:2013hsa}. These
produce a strong modulation of the scalar charges (see Fig.~\ref{fig:sepn_eccs}),
with a consequent impact on the dynamics, and possibly observable signatures both in the gravitational
and electromagnetic signals. For instance, the binary's orbit circularizes more rapidly for smaller values of
$\tilde{\beta}$ as illustrated in Fig.~\ref{fig:sepn_eccs}. As can also be seen in that figure, except for
the low-, equal-mass binary with mild eccentricity (which shows negligible differences between GR and ST theories), all the other cases display increasingly
marked deviations from GR as $\tilde{\beta}$ decreases and higher masses are considered. This behavior
has a direct impact on the electromagnetic flux, as shown in Fig.~\ref{fig:lums_eccs}.
Of particular interest is the fact that the orbital eccentricity induces, in all cases, an oscillatory
behavior in the flux intensity with a frequency and amplitude modulated by a growing trend as the orbit shrinks.
{Interestingly, the growth rate is different in the cases governed by ST theories for sufficiently low values of $\tilde{\beta}$.}
Note { however} that misalignment of the stars' dipole moments induces oscillations in the resulting luminosity
even in the quasi-circular case~\cite{Ponce:2014sza}.
{Finally, we note that the more rapid orbital evolution of strongly scalarized binaries
can cause the same total energy to be radiated within a significantly shorter time. This is shown
in Fig.~\ref{fig:intLum_eccs}, where we stress that we are showing the total radiated energy as a function
of time (and \textit{not} time to merger).
}
\begin{figure*}
\includegraphics[width=0.4\textwidth]{{{./sepnScChg-time_151-151_lowEcc}}}
\includegraphics[width=0.4\textwidth]{{{./sepnScChg-time_151-151_hghEcc}}}
\\
\vspace{1mm}
\includegraphics[width=0.4\textwidth]{{{./sepnScChg-time_152-174_lowEcc}}}
\includegraphics[width=0.4\textwidth]{{{./sepnScChg-time_152-174_hghEcc}}}
\caption{Eccentric binaries: evolution of the separation for a binary governed
by GR and ST with $\tilde \beta=-4.5$ and scalar charges of the binary's components (for binaries
governed by ST with $\tilde \beta=-4.5$ and $\tilde \beta=-4.2$ for comparison purposes)
as a function of time to merger, for an equal-mass binary (top panels) and
an unequal-mass binary (bottom panels), for low-eccentricity orbits (left column) and
high-eccentricity orbits (right column) [see text for the details on the exact masses and eccentricities].
As the orbit progresses, and scalar charges become significant, the eccentricity is reduced more rapidly in the ST cases
as illustrated by the increasingly reduced differences between local maxima and minima of the binary's separation.
}
\label{fig:sepn_eccs}
\end{figure*}
\begin{figure*}
\includegraphics[width=0.4\textwidth]{{{./beta_-4.5-4.2-GR--1.51-1.51_lowEcc-b_0-0.1}}}
\includegraphics[width=0.4\textwidth]{{{./beta_-4.5-4.2-GR--1.51-1.51_hghEcc-b_0-0.1}}}
\\
\vspace{2mm}
\includegraphics[width=0.4\textwidth]{{{./beta_-4.5-4.2-GR--1.52-1.74_lowEcc-b_0-0.1}}}
\includegraphics[width=0.4\textwidth]{{{./beta_-4.5-4.2-GR--1.52-1.74_hghEcc-b_0-0.1}}}
\caption{Luminosity vs time to merger, for the same parameters shown in Fig.~\ref{fig:sepn_eccs}.
For clarity's sake we show here the two extremal magnetic ratios considered,
$b = 0.1$ and $b = 0$ (i.e. a non-magnetized companion).
}
\label{fig:lums_eccs}
\end{figure*}
\begin{figure*}
\includegraphics[width=0.4\textwidth]{{{./intLum--beta_-4.5-4.2-GR--1.51-1.51_lowEcc_x3}}}
\includegraphics[width=0.4\textwidth]{{{./intLum--beta_-4.5-4.2-GR--1.51-1.51_hghEcc_x3}}}
\\
\vspace{2mm}
\includegraphics[width=0.4\textwidth]{{{./intLum--beta_-4.5-4.2-GR--1.52-1.74_lowEcc_x3}}}
\includegraphics[width=0.4\textwidth]{{{./intLum--beta_-4.5-4.2-GR--1.52-1.74_hghEcc_x3}}}
\caption{Energy radiated by eccentric binaries,
as a function of the time elapsed from an initial binary separation of approximately 220 km,
for the same parameters shown in Fig.~\ref{fig:sepn_eccs}.
We show, in contiguous panels, three different magnetic ratios (from left to right within each plot):
$b = 0.1$ (in red), $0.01$ (in green), and $0.001$ (in blue).
Note that in the case of low-eccentricity orbits (left column), the radiated energy has a smoother (continuous) trend,
while in the case of high-eccentricity orbits (right column) the radiated energy is more jagged;
this is a consequence of the oscillatory features in the luminosity, and ultimately the effect of the different dynamics of each binary.
}
\label{fig:intLum_eccs}
\end{figure*}
\section{Discussion}
\label{sec:Discussion}
The results shown in this work indicate that, within the context of electromagnetic emission
induced by magnetosphere interactions, binary neutron star systems could produce signals with
clear deviations from the GR expectation, depending on the gravity theory.
Our results open up the possibility of exploiting compact binary systems
to test gravitational theories by combining electromagnetic and gravitational signals. For instance, in the particular case of the ST theories of gravity considered here,
deviations in the gravitational-wave signals away from the GR prediction could be detected for binaries that undergo scalarization sufficiently
early in the advanced LIGO band~\cite{Sampson:2014qqa,Taniguchi:2014fqa} (naturally, how early in frequency
depends strongly on the signal-to-noise ratio of the signal). Such binaries, as illustrated here, also show distinctive departures
in the Poynting flux luminosity emitted from the system, when compared to the behavior within GR. The fact that deviations
could be observed with advanced LIGO if they take place at sufficiently low frequencies in the gravitational-wave signal
is a consequence of the detector's sensitivity curve (i.e. sufficiently early scalarization is needed to build up
enough signal-to-noise ratio in band). For this reason, scalarization effects
at high frequencies in the ST theories that we consider would be difficult to detect by gravitational-wave signals alone, unless the detector
is tuned for higher sensitivity in the higher frequency band. Without such tuning, however, complementary electromagnetic
signals could be exploited. As we illustrated here, such an opportunity does indeed appear to be possible.
In particular, our results indicate deviations in the expected electromagnetic luminosity in ST theories for {sufficiently low}
values of the coupling $\tilde{\beta}<0$. This observation
is not enough (in itself) to assess the nature of the underlying gravitational theory,
unless a good estimate of the stars' magnetization is available.
However, we find that the luminosity's strength and -- most importantly -- rate of change with time
exhibit differences {from the GR behavior} in the ST theories that we study, and this could provide precious information that
can be used to test the gravity theory. Indeed, in many respects, the rate of change of the luminosity would provide
analog information to the ``chirping'' gravitational signal, though in this case not limited by advanced LIGO's/VIRGO's noise
curve, which in its canonical configuration raises sharply before the kHz frequency at which the merger takes place.
As discussed in Ref.~\cite{Palenzuela:2013kra}, the Poynting flux from binary systems can indeed induce high-energy signals from the system, which can be exploited for this goal. Naturally, an important question to consider
is the distance at which these sources could be detected in the electromagnetic band. Such a distance
estimate must take into account both the type of expected radiation from the system, as well as the typical sensitivities of
the various observational facilities. A rough estimate can be computed by using the peak luminosities
$\simeq 10^{40} (B/10^{11}{\rm G})^2$ erg/s prior to the merger, and assuming that a relativistically expanding
electron-positron wind (sourced by energy dissipation and magnetohydrodynamical waves in between the stars)
produces an x-ray signal~\cite{Hansen:2000am}, preceding or coincident with the merger.
These signals may be detectable with ISS-Lobster, by virtue of its high sensitivity and wide field of view,
and may be seen up to distances of $\sim (B/10^{11}G)$~Mpc, if one assumes a fiducial 10\% efficient conversion of Poynting flux.
However, it is important to stress that the advent of gravitational-wave {detectors} would
potentially allow one to increase the accessible distances. Focused efforts by both
the gravitational-wave and electromagnetic-signal communities are ongoing to make the most of these ``multimessenger'' opportunities.
\vskip0.3cm
\textit{Acknowledgments:}
We thank N. Cornish, L. Sampson and N. Yunes for interesting discussions.
EB acknowledges support from
the European Union's Seventh Framework Programme (FP7/PEOPLE-2011-CIG)
through the Marie Curie Career Integration Grant GALFORMBHS PCIG11-GA-2012-321608.
LL acknowledges support by NSERC through a Discovery Grant and CIFAR. LL thanks
the Institut d'Astrophysique de Paris and the ILP LABEX (ANR-10-LABX-63),
for hospitality during a visit supported through the Investissements d'Avenir program under
reference ANR-11-IDEX-0004-02.
Research at Perimeter Institute is supported through Industry Canada
and by the Province of Ontario through the Ministry of Research and Innovation.
\normalem
|
\section{Introduction}
After their introduction over a quarter-century ago \cite{Aharonov1988}, quantum weak values \cite{Dressel2014,Kofman2012} have consistently found themselves at the center of controversy \cite{Aharonov2008,Aharonov2010}. Indeed, the original paper \cite{Aharonov1988} details how one can postselect a weak (i.e., noisy) measurement of a spin-$1/2$ operator for an electron (using a sequence of two Stern-Gerlach apparatuses) to obtain a \textit{conditioned expectation value} that approximates a weak value with an anomalously large value of $100$. The question whether this strange average value has any physical meaning pertaining to the spin has since plagued the concept of the weak value (e.g., \cite{Duck1989}).
The most recent addition to this controversy \cite{Ferrie2014b} considers a superficially similar example consisting of a classical coin that has its two faces noisily measured, then \textit{disturbed}, and finally conditioned to produce an anomalous average value of $100$ heads. The conclusion drawn from their study (which has been heavily criticized \cite{Ferrie2014bcom1,Ferrie2014bcom2,Ferrie2014bcom3,Ferrie2014bcom4,Ferrie2014bcom5}) is that strange weak values may be understood entirely as classical disturbance effects, making them not ``quantum.'' In fact, every element of this simple example of how intermediate disturbance can cause strange postselected averages of noisy signals has been previously demonstrated, and corroborates our published work: Not only did we emphasize a similar disturbance example using a colored marble in our systematic investigation of generalized observable measurements \cite{Dressel2010,Dressel2012b}, but we also carefully highlighted the potential role of invasive measurements in studies linking strange conditioned averages (including weak values) to violations of generalized Leggett-Garg inequalities \cite{Williams2008,Goggin2011,Dressel2011,Groen2013,Dressel2014c} (which were designed to test for ``quantum'' behavior in macroscropic systems \cite{Leggett1985,Leggett2002,Emary2014}). It is now well-established that any hidden-variable model that can produce strange conditioned averages like the weak value must include some form of intermediate disturbance (see also \cite{Tollaksen2007,Ipsen2014}). The more interesting question to raise is not whether a particular strange conditioned average may be explained as classical disturbance, but rather whether such models of disturbance can also reproduce the complete behavior of the weak value as its physical parameters are varied.
In this paper, we revisit this question in order to dispel the abundant confusion about weak values still present in the literature, and emphasize that a strange weak value is nonclassical in precisely the same manner that a single quantum particle can be considered to be nonclassical. Specifically, strange weak values fundamentally arise from \emph{interference} (i.e., superoscillations \cite{Aharonov2011,Berry2012}), and thus also appear in any wave-like field theory, such as classical optics \cite{Ritchie1991,Kocsis2011,Lundeen2011}, in a straightforward way. In such a classical field theory, anomalous weak values do faithfully indicate physical wave properties, despite how counter-intuitive their predictions may seem. For example, the orbital part of the Poynting vector field of optical vortex beams, or evanescent fields, can show anomalous local momentum distributions that are precisely equal to strange weak values \cite{Bliokh2013a,Bliokh2014,Dressel2014b}. Therefore, as with any quantum interference effect, only the fact that discrete and independent random events can be measured (as opposed to attenuated wave intensities) will distinguish whether the statistics producing a strange weak value are truly quantum mechanical in origin. (Alternatively, entangling the degrees of freedom of distinct particles will not have a simple classical field interpretation, e.g., \cite{Dressel2011}.) Nevertheless, even in the case of discrete measurement events the large number of measurements needed to statistically resolve such a weak value still imply that it is best considered as a dynamical physical variable for the effective (classical) mean field, and not necessarily to each individual quantum particle \emph{a priori} \cite{Dressel2014b}. To emphasize these subtle points, we carefully review several complementary approaches to deriving and understanding the weak value, paying special attention to its role as an ideal estimate, an experimentally measurable conditioned average, and as a classical dynamical variable for reduced quantum state evolution. This detailed treatment aims to supplement the simplified introduction to the experimental applications of weak values in \cite{Dressel2014} with an expanded theoretical discussion that highlights their pervasive and under-appreciated role throughout the quantum formalism.
In what follows we also emphasize the often overlooked connection between weak values and joint quasiprobability distributions (such as the Kirkwood-Dirac \cite{Kirkwood1933,Dirac1945,Chaturvedi2006,Lundeen2012,Hofmann2012b,Lundeen2014}, Terletsky-Margenau-Hill \cite{Terletsky1937,Margenau1961,Johansen2004,Johansen2004c}, and the various Moyal phase space distributions \cite{Wigner1932,Moyal1949,Glauber1963,Sudarshan1963}) that determine conditioned observable estimates. Notably, to obtain a strange weak value outside the usual eigenvalue bounds, these joint quasiprobability distributions must become negative as a consequence of the nonclassical quantum interference between probability amplitudes. The best known examples of this intrinsic negativity from quantum interference occur in the Moyal phase space distributions, such as the Wigner distribution for quadratures, or the Glauber-Sudarshan $P$ distribution for coherent state amplitudes, which show such negativity with nonclassical optical states \cite{Wigner1932,Glauber1963,Sudarshan1963,Moyal1949}. Indeed, for single particles such negativity in quasiprobability distributions has been proven to be an equivalent notion of ``nonclassicality'' as the need for \textit{contextual} hidden variables \cite{Spekkens2008,Ferrie2011} (in the sense of Bell-Kochen-Specker \cite{Kochen1967,Mermin1993}), and formally arises from the usual operator non-commutativity of quantum mechanics. This connection between contextuality and strange weak values was emphasized in a recent proof by Pusey \cite{Pusey2014}, as well as an earlier study by Tollaksen \cite{Tollaksen2007}, and is consistent with the established understanding that strange weak values arise fundamentally from (quantum) interference. It follows that if a classical model as in \cite{Ferrie2014b} could really mimic the detailed functional structure of the weak value, then it would also be able to simulate other features that are normally considered to be quantum mechanical. We further emphasize this latter point by reviewing the deep connections between weak values and the classical dynamical variables of the Hamilton-Jacobi formalism, both in its fully quantum generalization that is equivalent to the Schr\"odinger equation, and in the resulting classical limit.
This paper is organized as follows. In Section~\ref{sec:estimates} we derive and discuss how the weak value is the best statistical estimate for the average (but unmeasured) observable value for the times between two known measurement events. In Section~\ref{sec:measurement} we discuss three approaches for experimentally verifying the weak value as the appropriate such estimate: weak von Neumann coupling, weak generalized observable measurements, and as the physical dynamical variables for reduced state evolution. In Section~\ref{sec:quasiprob} we explicitly connect the weak value to quasiprobabilities, focusing on the Terletsky-Margenau-Hill, Kirkwood-Dirac, and Wigner distributions, and explicitly connect weak values to classical mean-field dynamical variables using the Hamilton-Jacobi quantum-classical correspondence. We conclude in Section~\ref{sec:conclusion}.
\begin{figure}[t]
\includegraphics[width=0.8\columnwidth]{disturbance.eps}
\caption{The weak value $Z_w(t) = \text{Re}\bra{f}\op{U}_{T-t}\op{Z}\op{U}_t\ket{i}/\bra{f}\op{U}_T\ket{i}$ estimates $\op{Z}$ conditioned on two boundaries that bracket the time interval $[0,T]$. For the Pauli $\op{Z} = \pr{1}-\pr{0}$ qubit operator prepared in $\ket{i} = \ket{1}$, postselected in $\bra{f} = \bra{0}$, and evolving with $\op{U}_t = \exp[i \omega t (\ket{1}\bra{0}+\ket{0}\bra{1})]$, the weak value $Z_w(t)$ (thin, green) coincides with the expectation value $Z(t) = \bra{i}\op{U}_t^\dagger\op{Z}\op{U}_t\ket{i}$ (dot-dashed, black) when the postselection is consistent with the natural oscillation. Otherwise, $Z(t)$ (dashed, black) displays a jump at time $T$, while $Z_w(t)$ (red) smoothly connects the boundaries, still passing through the same points of certain $Z$ (blue dots). The shaded regions exceed the eigenvalue bounds of $\pm 1$, indicating the inconsistency between the natural evolution and the observed boundaries.}
\label{fig:disturb}
\end{figure}
\section{Weak values as estimates}\label{sec:estimates}
Most of the controversy surrounding weak values rests upon their common (but unnecessary) association with an alternative time-symmetric approach to the quantum theory that involves two state vectors \cite{Watanabe1955,Aharonov1964,Aharonov1990,Aharonov1991,Aharonov2005,Aharonov2009}. In this time-symmetric approach, one forward-propagates a state-vector $\ket{i}$ from an initial time $0$ to $t$ in the usual way; however, one also \textit{back-propagates} a second state-vector $\bra{f}$ from a final time $T$ to $t$. While the initial state vector $\ket{i}$ corresponds to a preparation procedure, the final state vector $\bra{f}$ corresponds to a \textit{postselection} procedure.
Interestingly, the best estimate \cite{Hall2001,Johansen2004b,Hall2004} of the average (unmeasured) value for an observable $\op{A}$ at any time $t$ in the interval $[0,T]$ is then not the expectation value $A(t) = \bra{i}\op{U}_t^\dagger \op{A} \op{U}_t\ket{i}$ (which neglects the information about the postselection), but is rather the \textit{weak value}
\begin{align}\label{eq:wv}
A_w(t) &= \text{Re}\,\frac{\bra{f}\op{U}_{T-t}\op{A}\op{U}_t\ket{i}}{\bra{f}\op{U}_T\ket{i}},
\end{align}
as we will derive shortly. Here $\op{U}_t = \exp(-i\op{H} t/\hbar)$ is the unitary propagator for a time-interval $t$ that is generated by the Hamiltonian $\op{H}$. Note that the imaginary part of the weak value, while independently interesting as a measure of intrinsic measurement disturbance \cite{Johansen2004b,Dressel2012d,Hofmann2014}, is unrelated to the estimation of $\op{A}$ as an observable, so we will ignore it for now.
The problematic feature of Eq.~\eqref{eq:wv} as an estimate is that it may exceed the eigenvalue range of $\op{A}$; such strange behavior is illustrated as the shaded areas in Figure~\ref{fig:disturb}. As discussed in the introduction, a classical conditioned estimate may show such anomalous behavior only if the estimation procedure is \textit{noisy} and if what is being estimated is \textit{disturbed} in the interval $[0,T]$ \cite{Dressel2012b,Tollaksen2007,Williams2008,Dressel2011,Ipsen2014}. The question raised in Ref.~\cite{Ferrie2014b} is whether such a classical model with noisy estimation and disturbance is sufficient to explain Eq.~\eqref{eq:wv}. As we will show in what follows, such a classical disturbance explanation is difficult to defend when the many roles of this weak value expression are examined in detail.
\subsection{Derivation of the best estimate}
For completeness we now show how this weak value formula can be obtained as the optimal estimate that minimizes a statistical uncertainty metric, justifying its interpretation as a best estimate.
Suppose one wishes to estimate the best average values for an observable $\op{A}$ given a initial preparation $\ket{i}$, followed by a projective measurement in a particular basis $\ket{f}$ that does not commute with $\op{A}$. For each $f$, we can guess a value $\bar{a}_f$ that estimates the average of $\op{A}$ given that the specific result $f$ was observed. This procedure formally constructs a Hermitian observable $\op{A}_{\text{est}} = \sum_f \bar{a}_f\,\pr{f}$ that is measured by the procedure, which contains the estimates for every $f$. The goal is then to determine the optimal such estimates $\bar{a}^{\text{opt}}_f$ of $\op{A}$ conditioned on each measured result $f$.
To accomplish this goal, we must define a measure for how close the estimate $\op{A}_{\text{est}}$ is to the target observable $\op{A}$. A natural choice for such a measure is the weighted trace distance between two observables \cite{Hall2001,Johansen2004,Hall2004}
\begin{align}
\mathcal{D}_\rho(\op{A},\op{B}) &= \Tr{\op{\rho}\,(\op{A}-\op{B})^2},
\end{align}
which can generally depend upon any positive prior bias $\op{\rho}$. We can interpret this measure as specifying the shortest \emph{geometric distance} in operator space, weighted by the \emph{statistical prior} bias of $\op{\rho}$. Since $\op{A}$ and $\op{B}$ need not commute, there is generally no straightforward operational interpretation for this distance \cite{Busch2013,Dressel2014d,Busch2014}, but it does formally provide a reasonable definition of the geometric ``closeness'' for the two operators.
Now suppose we have a definite prior state $\op{\rho} = \pr{i}$ and choose $\op{B} = \op{A}_{\text{est}}$. Computing the weighted trace distance yields \cite{Hall2001,Johansen2004,Hall2004}
\begin{align}
\mathcal{D}_i(\op{A},\op{A}_{\text{est}}) &= \bra{i}\left[\op{A}^2 + \op{A}_{\text{est}}^2 - (\op{A}\op{A}_{\text{est}} + \op{A}_{\text{est}}\op{A})\right]\ket{i}, \nonumber \\
&= \bra{i}\op{A}^2\ket{i} - \textstyle{\sum_f} \abs{\braket{f}{i}}^2\,\left[\text{Re}\frac{\bra{f}\op{A}\ket{i}}{\braket{f}{i}}\right]^2 \nonumber \\
&\quad + \textstyle{\sum_f}\abs{\braket{f}{i}}^2\,\left[\bar{a}_f - \text{Re}\frac{\bra{f}\op{A}\ket{i}}{\braket{f}{i}}\right]^2.
\end{align}
Only the final term depends on the choice of estimates $\bar{a}_f$ for each $f$, and is positive definite. Therefore, the trace distance is minimized when this term vanishes, which in turn implies that the optimal estimate for each independent $f$ must be the weak value formula
\begin{align}
\bar{a}^{\text{opt}}_f &= \text{Re}\frac{\bra{f}\op{A}\ket{i}}{\braket{f}{i}}.
\end{align}
Note that if the basis $\ket{f}$ is chosen to be the eigenbasis of $\op{A}$, then these optimal estimates reduce identically to the eigenvalues of $\op{A}$ and the trace distance vanishes.
To obtain Eq.~\eqref{eq:wv} including intermediate time-evolution, we can choose a particular $\ket{f} \mapsto \op{U}_{T-t}\ket{f}$ and initial state $\ket{i} \mapsto \op{U}_t\ket{i}$ that take into account Hamiltonian propagation $\op{U}_t = \exp(t\op{H}/i\hbar)$ from the observed results for a specific preparation and postselection.
This pure-state derivation may also be generalized in a straightforward way \cite{Hall2004}, which produces a formulation of the weak value suitable for generalized measurements and mixed states
\begin{align}\label{eq:wvgen}
A_w(t) &= \text{Re}\,\frac{\Tr{\op{E}_{T-t}\,\op{A}\,\op{\rho}_t}}{\Tr{\op{E}_{T-t}\,\op{\rho}_t}}.
\end{align}
Here the back-propagating operator $\op{E}_{T-t}$ is often called a ``retrodictive state,'' or ``effect matrix,'' in contrast to the forward-propagating ``predictive state'' $\op{\rho}_t$, or ``density matrix.'' Recently, the estimate in Eq.~\eqref{eq:wvgen} has been used to great effect experimentally \cite{Campagne-Ibarcq2014,Weber2014,Tan2014,Rybarczyk2014} for ``quantum smoothing'' \cite{Tsang2009,Tsang2012} and ``past quantum state'' analyses \cite{Gammelmark2013,Gammelmark2014} of continuously measured signals (e.g., it was used to track individual photon emissions into a monitored cavity \cite{Rybarczyk2014}). Both $\op{E}_{T-t}$ and $\op{\rho}_t$ generally evolve according to open-system master equations \cite{Breuer2007,Wiseman2009} that can also include the effects from additional (discrete or continuous-in-time) stochastic measurement-results \cite{Dressel2013b,Gammelmark2013,Chantasri2013}, in contrast to the closed-system (unitary) Schr\"odinger-von Neumann dynamics usually assumed with Eq.~\eqref{eq:wv}. Note that if the effect matrix $\op{E}_{T-t}$ is the identity $\op{1}$, then no posterior conditioning has been performed, so the usual expectation value is also recovered as a special case.
\subsection{Interpreting and generalizing the estimate}
As a philosophical side note, for those who believe that the state-vector represents the complete physical (ontic) reality (e.g., adherents to the many-worlds interpretation \cite{Vaidman2014}), this time-symmetric estimate prompts several more radical speculations: The existence of the second state vector $\bra{f}\op{U}_{T-t}$ in Eq.~\eqref{eq:wv} seems to imply not only that the state $\op{U}_t\ket{i}$ is an incomplete description of reality at time $t$, but also that there seems to be a causal effect on the time $t$ from the future time $T$ \cite{Aharonov2012}. Such a \emph{retro-causal} interpretation is similar in spirit to the interpretations of anti-particles in quantum field theory as field-excitations that move backwards through time \cite{Feynman1949}. However, just as with anti-particles, one does not need to invoke such controversial philosophical concepts as physical state-vectors or retro-causation to meaningfully interpret the weak value in Eq.~\eqref{eq:wv} as the best available statistical estimate given only the information about the specified boundary conditions.
A more pragmatic attitude (which we shall adopt here) is to treat the estimate in Eq.~\eqref{eq:wv} as \textit{subjective} (epistemic), and pertaining to a time interval $[0,T]$ that has already occurred in the past. That is, one performs an experiment that prepares $\ket{i}$ at time $0$, waits a duration $T$, then makes a projective measurement that shows a result corresponding to the state $\bra{f}$. One then interprets Eq.~\eqref{eq:wv} as the best estimate of the (unmeasured) average value of $\op{A}$ within that time interval \cite{Hall2001,Johansen2004b,Hall2004}, given only the knowledge of $\ket{i}$, $\bra{f}$, and $\op{H}$. We emphasize that this approach is no different in character than stating that the expectation value $\bra{i}\op{U}^\dagger_t\op{A}\op{U}_t\ket{i}$ is the best estimate for the (unmeasured) average value of $\op{A}$, given only the knowledge of the preparation $\ket{i}$ and $\op{H}$. Indeed, such a counterfactual interpretation of the expectation value in the absence of measurement is at the core of the Ehrenfest theorem that equates quantum expectation values with mean-field classical dynamical variables \cite{Ehrenfest1927}. If we interpret weak values as similar (but additionally constrained) classical dynamical variables, then an anomalous weak value should indicate the presence of some interesting intermediate physical process that must have occurred in order to satisfy both boundary conditions that bracket the time interval $[0,T]$ (see Figure~\ref{fig:disturb}).
Supporting this point of view is the fact that similar bidirectional (in time) estimates about unknown properties of structured stochastic processes (e.g., hidden Markov models) during such an interval are now well-established in classical computational mechanics \cite{Shalizi2001,Crutchfield2009,Ellison2009,Mahoney2011}. There it is shown that one should use both forward and reverse ``causal states'' (i.e., probability distributions) that contain information gathered both before \textit{and after} each time $t$ to optimally estimate the properties of an evolving stochastic process. Similarly, classical statistics and filtering theory also use bidirectional states to provide the best estimate for information contained in noisy data confined to a time interval (called optimally ``smoothing'' the noise) \cite{Simonoff1998,Einicke2012}. Since quantum theory is closely related to probability theory \cite{Dressel2012b,Leifer2013}, it is logical that similar estimation methods can be applied. Indeed, upgrading these estimation schemes to the quantum realm \cite{Pegg2002a,Dressel2013b,Leifer2013} produces both states in Eq.~\eqref{eq:wv}, as well as the mixed-state generalization of Eq.~\eqref{eq:wvgen} \cite{Pegg2002b,Wiseman2002,Coecke2012}.
\section{Measuring weak values}\label{sec:measurement}
The confidence that estimations like the expectation value, or the weak value in Eqs.~\eqref{eq:wv} and \eqref{eq:wvgen}, reflect something meaningful about the physical world (and are not merely fevered hallucinations of the mind) follows from verification of their predictions by experimental measurements. In the case of the expectation value, any unbiased estimation of $\op{A}$ will suffice, corroborating the predicted result. In the case of the weak value, however, the presence of the posterior boundary condition additionally constrains the form of the possible measurements that can verify the estimate.
Specifically, those measurements must be ``weak,'' meaning that they should not appreciably perturb the evolution of the quantum system. Since information extraction necessarily disturbs the quantum state, only minimally informative (i.e., \emph{noisy}) measurements will leave the state mostly unperturbed \cite{Aharonov1988,Wiseman2009}, and thus faithfully reproduce the assumptions made about the evolution during the time interval $[0,T]$ by the formulas in Eqs.~\eqref{eq:wv} and \eqref{eq:wvgen}. The surprising fact is that averaging such weak observable measurements can indeed consistently verify the weak value as the correctly estimated average, even when it predicts anomalous averages.
In what follows we will detail the standard von Neumann approach for measuring the weak value, as well as a more general approach that solidifies its interpretation as a conditioned average in the limit of negligible disturbance to the quantum state. We will also detail how the weak value appears as a classical dynamical variable for reduced system evolution even outside the usual context of postselected weak measurements, further cementing its interpretation as the appropriate mean-field variable that physically estimates the observable when the natural state evolution is unchanged by external influences.
\subsection{von Neumann interaction}
The standard approach for performing a weak observable measurement \cite{Aharonov1988,Dressel2012e,Kofman2012,Dressel2014}, originally due to von Neumann \cite{vonNeumann1932}, is to couple the \emph{system} observable of interest $\op{A}$ (such as the spin of a particle) to a \emph{detector} observable $\op{F}$ (such as the transverse momentum of the same particle) for an independent degree of freedom, using a simple linear interaction Hamiltonian
\begin{align}\label{eq:vnham}
\op{H}_{DS}(t) &= \hbar\, g(t)\,\op{F}\otimes\op{A}.
\end{align}
The time-dependent coupling profile $g(t)$ is typically assumed to be zero outside a short interval of duration $\delta t$ and to be impulsive, i.e., short on the timescale of the natural dynamics of both the system and the detector. In the interaction picture for the system and detector, this coupling Hamiltonian produces a joint unitary rotation of the joint state that entangles the system with the detector
\begin{align}\label{eq:interaction}
\op{V}_{DS} &= \exp(-ig\,\op{F}\otimes\op{A}),
\end{align}
where $g = \int_0^{\delta t}\!\! g(t')\,dt'$ is the effective coupling strength for the impulsive interaction.
For a concrete example of such an interaction, in the optical experiments \cite{Ritchie1991,Kocsis2011,Lundeen2011} a wafer of birefringent crystal was used to couple the polarization (system) of a paraxial beam to its continuous transverse momentum (detector). In contrast, the optical experiments \cite{Pryde2005,Goggin2011,Dressel2011} used polarization-dependent reflection from partially transmitting optical elements to couple the polarization (system) of a paraxial beam to its binary orbital which-path degree of freedom (detector). In the more recent superconducting qubit experiment \cite{Groen2013} the energy basis of one transmon qubit (system) was coupled via an intermediate bus stripline resonator to the binary energy basis of a second and physically separated transmon qubit (detector) using microwave pulses.
\begin{figure}[t]
\includegraphics[width=\columnwidth]{vonneumann.eps}
\caption{Impulsive von Neumann interaction for measuring $\op{A} = \sum_a a\,\pr{a}$. A system degree of freedom $\ket{i}$ (blue, top) is prepared at time $0$, and a detector degree of freedom $\ket{d}$ (red, bottom) is prepared at a potentially different time $t_0$. After independent unitary propagation $\op{U}_t$ and $\op{U}_{t-t_0}^{(D)}$, respectively, to the intermediate time $t$ they interact impulsively (green, wavy) with a joint unitary interaction $\op{V}_{DS}$ for a short duration $\delta t$. After the interaction, both system and detector continue to propagate until the detector is measured at a time $T'$ to find the result $\bra{x}$, and the system is measured at a potentially different time $T$ to find the result $\bra{f}$. This entire coupling procedure may be represented as a single system operation (lightly shaded region) $\op{M}_x = \bra{x}\op{U}^{(D)}_{T'-t}\op{V}_{DS}\op{U}^{(D)}_{t-t_0}\ket{d}$ that includes the detector preparation, evolution, coupling, and measurement together, producing a net effect only on the evolving system state. If the identity $\op{A} = \sum_x \alpha_x\,\op{M}^\dagger_x\op{M}_x$ can be satisfied for some calibrated signal values $\alpha_x$ for the detector, then the system observable $\op{A}$ can be faithfully estimated by the detector results $x$ in an unbiased way. For strong measurements by the detector, the effective system operations become eigenstate projections $\op{M}_x \to \pr{a}$ with an eigenvalued signal $\alpha_x \to a$, while for weak measurements the operations approximate the identity $\op{M}_x \approx \op{1}$ with a noisy signal $|\alpha_x|\gg|a|$, leaving the state essentially unperturbed. Averaging this noisy signal $\alpha_x$ for such weak measurements conditioned on a particular $f$ approximates the weak value $\text{Re}\bra{f}\op{U}_{T-t}\op{A}\op{U}_t\ket{i}/\bra{f}\op{U}_T\ket{i}$.}
\label{fig:vonneumann}
\end{figure}
To derive the effect of such an interaction, consider the coupling procedure in Figure~\ref{fig:vonneumann}. Suppose that the impulsive coupling begins at a time $t$, and that the joint state of the detector-system degrees of freedom is initially a product state $\ket{d'}\ket{i'}$, where the initial detector and system states at time $t$
\begin{align}
\ket{d'} &= \op{U}_{t-t_0}^{(D)}\ket{d}, & \ket{i'} &= \op{U}_t\ket{i}
\end{align}
may have propagated from previous states $\ket{i}$ and $\ket{d}$ that were prepared at the possibly different earlier times $t=0$ and $t=t_0$, respectively, following the independent Hamiltonian evolution
\begin{align}
\op{U}^{(D)}_t &= \exp(-it\op{H}_D/\hbar), & \op{U}_t &= \exp(-it\op{H}_S/\hbar).
\end{align}
After the entangling interaction of Eq.~\eqref{eq:interaction}, the detector and system are allowed to again evolve freely for (potentially different) times $T-t$ and $T'-t$, respectively, after which they are independently measured projectively in the bases $\bra{x}$ and $\bra{f}$ to obtain a pair of detector-system results $(x,f)$. Back-propagating these measured states from $T'$ and $T$ to the time immediately following the impulsive coupling $t+\delta t\approx t$ produces the effective final states
\begin{align}
\bra{x'} &= \bra{x}\op{U}^{(D)}_{T'-t}, & \bra{f'} &= \bra{f}\op{U}_{T-t}.
\end{align}
The joint probability distribution for the results $(x,f)$ can then be written in the compact (scattering) form
\begin{align}\label{eq:jointvn}
p_{x,f} &= \abs{\bra{x',f'}\op{V}_{DS}\ket{d',i'}}^2.
\end{align}
Importantly, changing the durations of time $T$ and $T'$ before measuring the system and detector does not affect the general form of the joint distribution in Eq.~\eqref{eq:jointvn}; only the back-propagated states $\ket{x'}$ and $\ket{f'}$ will change from the free evolution when these durations are varied. For a concrete example of this effect, in the optical case of \cite{Ritchie1991,Kocsis2011,Lundeen2011} the free evolution of the transverse momentum (detector) produces diffraction effects after the birefringent crystal, which alters the effective back-propagated state implied by a later transverse position measurement.
For a sufficiently small coupling strength $g$, the interaction only \emph{weakly} perturbs the initial system states, and we can expand the joint interaction $\op{V}_{DS}$ perturbatively. To good approximation, we find that the relative change of the joint probability in Eq.~\eqref{eq:jointvn} due to the weak interaction has the form
\begin{align}\label{eq:jointwv}
\frac{p_{x,f}}{\abs{\braket{x',f'}{d',i'}}^2} &\approx 1 + 2g\,\text{Im}F_wA_w + g^2\,|F_w|^2|A_w|^2,
\end{align}
which involves only the \emph{first-order} (complex) detector and system weak values
\begin{align}
F_w &= \frac{\bra{x'}\op{F}\ket{d'}}{\braket{x'}{d'}}, & A_w &= \frac{\bra{f'}\op{A}\ket{i'}}{\braket{f'}{i'}}.
\end{align}
Note that continuing this expansion will produce an infinite series characterized entirely by higher-order weak values involving all powers of $\op{A}$ and $\op{F}$ \cite{Dressel2014}; however, this truncation that involves only the first-order weak values is remarkably accurate for sufficiently small coupling strengths that satisfy $g|F_w||A_w| \ll 1$ \cite{DiLorenzo2012}.
This joint distribution also determines the relative change of the marginalized distributions for the detector, $p_x = \sum_f\,p_{x,f}$, and system, $p_f = \sum_x\,p_{x,f}$, statistics alone
\begin{align}\label{eq:detectorwv}
\frac{p_{x}}{\abs{\braket{x'}{d'}}^2} &\approx 1 + 2g\,\mean{A}\text{Im}F_w + g^2\,\mean{A^2}|F_w|^2, \\
\frac{p_{f}}{\abs{\braket{f'}{i'}}^2} &\approx 1 + 2g\,\mean{F}\text{Im}A_w + g^2\,\mean{F^2}|A_w|^2.
\end{align}
Critically, note that the reduced \emph{detector} statistics involve the expectation value of the \emph{system} operator $\mean{A} = \bra{i'}\op{A}\ket{i'}$, as well as its second moment $\mean{A^2} = \bra{i'}\op{A}^2\ket{i'}$. This dependence means that by examining only the statistics of the detector, we can \emph{indirectly estimate} the system observable $\op{A}$. Similarly, the system statistics involve the expectation value of the detector operator $\mean{F} = \bra{d'}\op{F}\ket{d'}$, as well as its second moment $\mean{F^2} = \bra{d'}\op{F}^2\ket{d'}$, so can be used to indirectly estimate the detector observable $\op{F}$ in a symmetric way.
To perform such an estimation using only the detector statistics in the linear response regime (where we can neglect the terms that are second-order in $g$), an experimenter weights each of the outcomes $x$ of the detector by some scaling value $\alpha_x$, which constructs an effective detector readout observable \cite{Dressel2010,Kofman2012}
\begin{align}
\op{R} &= \textstyle{\sum_x} \alpha_x\,\pr{x},
\end{align}
and produces the detector average (keeping all system results)
\begin{align}\label{eq:estimate}
\textstyle{\sum_{x}}\,\alpha_x\,p_{x} &\approx \mean{R} + 2g\, \mean{A}\text{Im}\mean{RF},
\end{align}
in terms of the detector quantities $\mean{R} = \bra{d'}\op{R}\ket{d'}$ and $\mean{RF} = \bra{d'}\op{R}\op{F}\ket{d'}$. Hence, by calibrating a known initial detector state $\ket{d'}$ and choosing the observables $\op{F}$ and $\op{R}$ strategically, one can extract the expectation value $\mean{A}$ in an unbiased way using only the detector statistics. Note that typically the detector state $\ket{d'}$ is chosen such that the relevant observables have zero mean prior to the interaction, $\mean{F} = \mean{R} = 0$.
Now suppose our experimental setup additionally filters the system outcomes so that only $f$ may occur. (Alternatively, we can select only those particular events in the post-processing of data that includes more outcomes.) The statistics of the laboratory detector $x$ measurements that are properly conditioned on a particular system $f$ outcome will then have the usual form from Bayes' theorem
\begin{align}
p_{x|f} &= \frac{p_{x,f}}{\textstyle{\sum_x} p_{x,f}}.
\end{align}
The approximate relative change of this conditional distribution is thus
\begin{align}\label{eq:wvmeas}
\frac{p_{x|f}}{\abs{\braket{x'}{d'}}^2} &\approx \frac{1 + 2g\,\text{Im}F_wA_w + g^2\,|F_w|^2|A_w|^2}{1 + 2g\,\mean{F}\text{Im}A_w + g^2\,\mean{F^2}|A_w|^2}.
\end{align}
When the terms of order $g^2$ can be neglected, and when $\mean{F}=0$, we thus have the linear response result that can be compared to the unfiltered estimation in Eq.~\eqref{eq:estimate}
\begin{align}\label{eq:condestimate}
\textstyle{\sum_x}\,\alpha_x\,p_{x|f} &\approx \mean{R} + 2g\, \text{Im}A_w\mean{RF}, \\
&= \mean{R} + 2g\, [\text{Re}A_w\text{Im}\mean{RF} + \text{Im}A_w\text{Re}\mean{RF}]. \nonumber
\end{align}
Importantly, the weak value factor $\text{Re}A_w$ that scales $\text{Im}\mean{RF}$ corresponds directly to the expectation value $\mean{A}$ in Eq.~\eqref{eq:estimate}. That is, the filtering procedure partitions the total average $\mean{A}$ into subensembles that have conditioned averages of $\text{Re}A_w$, which precisely matches what we expect from the best estimate of $\op{A}$ in Eq.~\eqref{eq:wv}. The final term involving $\text{Im}A_w$ averages to zero in the total ensemble, and corresponds to the intrinsic symmetric backaction of the detector on the system due to the joint interaction of Eq.~\eqref{eq:interaction}; it does not correspond to the estimation of $\op{A}$ \cite{Steinberg1995,Steinberg1995b,Dressel2012d}, and can be removed from the detector signal in practice while preserving the estimation of $\mean{A}$ by choosing the detector observables such that $\text{Re}\mean{RF} = 0$.
\subsection{Observable estimation}
The preceding discussion of the von Neumann coupling is traditional in the weak value literature \cite{Aharonov1988,Dressel2014,Kofman2012} and is often sufficient for describing experimental implementations that measure weak values (e.g., \cite{Ritchie1991,Kocsis2011,Lundeen2011}). However, this perturbative derivation makes the association between the estimations of the total observable average $\mean{A}$ and the real part of the weak value $\text{Re}A_w$ somewhat inferential, prompting skepticism about the appropriateness of the connection. Since $\text{Re}A_w$ has such counterintuitive properties, having a more direct link between the estimation of $\op{A}$ and $\text{Re}A_w$ as a conditioned average value more generally is desirable. Thankfully, we can indeed demonstrate that this is the proper association by rephrasing the conditional estimation procedure using the formalism of generalized measurements \cite{Dressel2010,Dressel2012b,Dressel2013b}.
To do this, we rewrite the joint probability of Eq.~\eqref{eq:jointvn} entirely in the system space
\begin{align}\label{eq:joint}
p_{x,f} = \abs{\bra{f'}\op{M}_x\ket{i'}}^2,
\end{align}
by defining a \emph{measurement (Kraus) operator} that encodes the entire coupling and measurement procedure into a single system operator \cite{Nielsen2000,Wiseman2009}
\begin{align}
\op{M}_x &= \bra{x'}\op{V}_{DS}\ket{d'}.
\end{align}
Notice that this measurement operator has the form of a partial matrix element, but only contracts out the detector part of the joint unitary $\op{V}_{DS}$ to leave a purely system operator, which directly corresponds to how the complete measurement \emph{procedure} affects the system. Intuitively, for strong measurements of $\op{A}$, the measurement operators $\op{M}_x \to \pr{a}$ will become projectors onto the eigenstates of $\op{A}$, while for \emph{weak} measurements (in the sense of \cite{Aharonov1988}) the measurement operators $\op{M}_x \approx \op{1}$ will approximate the identity operator for all $x$, which leaves the initial state nearly unperturbed.
It follows that the marginalized distribution of only the detector results can be written
\begin{align}
p_x &= \textstyle{\sum_f}\, p_{x,f} = \bra{i'}\op{M}_x^\dagger\op{M}_x\ket{i'},
\end{align}
which has the form of the \emph{system} expectation of a probability operator
\begin{align}
\op{P}_x &= \op{M}^\dagger_x\op{M}_x
\end{align}
for the \emph{detector} result $x$. These probability operators are positive, and form a resolution of the identity in the system space
\begin{align}\label{eq:pom}
\textstyle{\sum_x} \op{P}_x = \bra{d'}\op{V}_{DS}^\dagger\left[\textstyle{\sum_x} \pr{x'}\right]\op{V}_{DS}\ket{d'} = \op{1},
\end{align}
making them a \emph{probability operator-valued measure} (POM, or POVM). Such a POM is the operator version of a properly normalized probability distribution. Indeed, when $\op{P}_x$ commutes with $\op{A}$, its diagonal elements will be precisely the classical conditional probabilities $p_{x|a}$ describing the likelihoods of each detector result $x$ given a definite preparation of $a$ \cite{Dressel2012b}. For each $a$, these probabilities then independently satisfy $\sum_x p_{x|a} = 1$ according to Eq.~\eqref{eq:pom}.
Generally speaking, any purity-preserving generalized measurement can be expressed as such a measurement operator $\op{M}_x$ and associated POM $\op{P}_x = \op{M}_x^\dagger\op{M}_x$ \cite{Nielsen2000,Wiseman2009}, with the von Neumann interaction being a special case for implementing such a measurement with a concrete Hamiltonian. For example, the polarization-dependent reflection measurement in \cite{Dressel2011} did not use a von Neumann Hamiltonian description of the interaction of the entangled photon pair with the glass microscope coverslip in the experimental analysis, but rather characterized the relevant POM operators $\op{P}_x$ directly with a series of separate calibration measurements by \emph{measuring} the conditional probabilities $p_{x|a}$ for known preparations of the eigenstates $\ket{a}$. Up to additional phases that were also calibrated in separate measurements, the diagonal elements of $\op{M}_x$ then had the form $\sqrt{p_{x|a}}$. The benefit of this direct approach is that the effect of the actual measurement procedure may be experimentally measured, without requiring a more detailed model of the extended detector space.
Now suppose that we can use the measured probabilities $p_x$ to estimate the expectation value of $\op{A}$ in an unbiased way. To do this, we must weight the outcomes $x$ of the detector with appropriately scaled values $\alpha_x$ (e.g., by rescaling a low-visibility signal in the usual way) \cite{Dressel2010}
\begin{align}\label{eq:expect}
\textstyle{\sum_x} \alpha_x\,p_x &= \bra{i'}[\textstyle{\sum_x} \alpha_x\, \op{P}_x]\ket{i'}.
\end{align}
Evidently, to produce the expectation value $\mean{A}=\bra{i'}\op{A}\ket{i'}$ for any initial state $\ket{i'}$, we must be able to choose appropriate values $\alpha_x$ that will calibrate the detector to probe the generalized spectral expansion \cite{Dressel2010,Dressel2012b}
\begin{align}\label{eq:identity}
\op{A} = \textstyle{\sum_x} \alpha_x \op{P}_x.
\end{align}
If it can be arranged, this operator identity will guarantee that the values $\alpha_x$ and measurement operators $\op{M}_x$ (and thus the probability operators $\op{P}_x$) will directly estimate $\op{A}$ in an unbiased way. As a concrete example, in \cite{Aharonov1988} the detector was chosen such that $\op{F} = \op{p}$ was the momentum operator, $\op{R} = \op{x}$ was the position operator, and $\braket{x}{d'} = (2\pi\sigma^2)^{-1/4}\exp(-x^2/4\sigma^2)$ was a zero-mean Gaussian distribution, such that $\mean{R}=\mean{F}=0$, $F_w = ix/2\sigma^2$, and thus $\mean{\op{R}\op{F}} = i/2$ in Eq.~\eqref{eq:estimate}; setting the simple scaled values of $\alpha_x = x/g$ then directly yields an unbiased estimation of $\mean{A}$ for all $g$. Note, however, that this common choice for Gaussian measurements is generally not a unique choice for producing an unbiased estimation, since the dimension of the detector often exceeds that of the system \cite{Dressel2010,Dressel2012b}.
Once we fix the weights $\alpha_x$ to achieve the estimation identity of Eq.~\eqref{eq:identity}, we also fix the partial average for each postselection $f$ according to Eq.~\eqref{eq:joint}
\begin{align}\label{eq:partial}
\textstyle{\sum_x} \alpha_x\,p_{x,f} &= \bra{i'}\op{O}_t\ket{i'},
\end{align}
which we write compactly as an expectation value for an effective system operator at time $t$
\begin{align}
\op{O}_t &\equiv \textstyle{\sum_x}\alpha_x \op{M}^\dagger_x\op{\Pi}_f\op{M}_x, \\
&= \frac{1}{2}(\op{A}\op{\Pi}_f + \op{\Pi}_f\op{A}) + \textstyle{\sum_x} \alpha_x\,\mathcal{L}[\op{M}^\dagger_x]\op{\Pi}_f, \nonumber
\end{align}
that we write in turn as a symmetric (Jordan) product \cite{Jordan1934} between the operator $\op{A}$ and the postselection projection operator $\op{\Pi}_f \equiv \pr{f'}$, modified by a sum of weighted Lindblad (dissipation) operations \cite{Dressel2013b}
\begin{align}\label{eq:lindblad}
\mathcal{L}[\op{M}^\dagger_x](\cdot) &\equiv \frac{1}{2}\left([\op{M}^\dagger_x,\cdot]\op{M}_x + \op{M}^\dagger_x[\cdot,\op{M}_x]\right),
\end{align}
familiar from open-system dynamics \cite{Wiseman2009,Breuer2007}. These Lindblad terms quantify the perturbation introduced by the measurement. Note that for \textit{weak} measurements (i.e., $\op{M}_x \approx \op{1}$) the commutators in the Lindblad terms approximately vanish to leave only the symmetric product with $\op{A}$, signifying that both the initial system state and postselection are essentially unaffected by the measurement results $x$ that are being used to estimate $\op{A}$ \cite{Dressel2012b,Dressel2013b}.
Expanding the partial average in Eq.~\eqref{eq:partial} produces
\begin{align}\label{eq:partial2}
\textstyle{\sum_x} \alpha_x\,p_{x,f} &= \text{Re}\,\bra{f'}\op{A}\ket{i'}\braket{i'}{f'} + \mathcal{E}[\alpha],
\end{align}
where we have introduced the Lindblad error terms
\begin{align}
\mathcal{E}[\alpha] = \textstyle{\sum_x} \alpha_x \bra{i'}\,(\mathcal{L}[\op{M}_x]\op{\Pi}_f)\,\ket{i'}
\end{align}
that are produced entirely by the perturbation from the measurement. Conditioning this partial average on obtaining a particular $f$ then yields
\begin{align}\label{eq:condav}
\frac{\sum_x \alpha_x\,p_{x,f}}{\sum_x\,p_{x,f}} &= \frac{\text{Re}\,\bra{f'}\op{A}\ket{i'}\braket{i'}{f'} + \mathcal{E}[\alpha]}{\braket{f'}{i'}\braket{i'}{f'} + \mathcal{E}[1]}.
\end{align}
When the error terms $\mathcal{E}$ are small enough to be neglected \cite{Dressel2012,Dressel2012b} (meaning that the initial system state is negligibly perturbed), the \emph{real part} of the weak value in Eq.~\eqref{eq:wv} is unambiguously recovered as the \textit{measured} conditioned estimate for $\op{A}$, verifying our derivation of this real part as a best estimate. As expected, the imaginary part is unrelated to the estimation of $\op{A}$, so does not contribute to Eq.~\eqref{eq:condav}, which justifies our interpretation of the terms in the von Neumann linear response of Eq.~\eqref{eq:condestimate}. Deriving Eq.~\eqref{eq:wvgen} as a measured estimation is a similar exercise \cite{Dressel2012e}.
Importantly, nothing about the derivation of Eq.~\eqref{eq:condav} changes when the time $t$, the initial system state $\ket{i}$, the system postselection $\bra{f}$, or even the system Hamiltonian $\op{H}$ are varied, as long one keeps fixed the measurement procedure set by the choice of calibration weights $\alpha_x$ and corresponding $\op{M}_x$ (e.g., the detector states $\ket{d'}$ and $\bra{x'}$, Hamiltonian $\op{H}^{(D)}$, and coupling interaction $\op{V}_{DS}$). This robustness of the derivation implies that the same weak measurement procedure can approximate the \emph{entire functional dependence} of the weak value in Eq.~\eqref{eq:wv}, in contrast to the single arbitrary value produced by the coin disturbance scheme in Ref.~\cite{Ferrie2014b}. Moreover, the weak value in Eq.~\eqref{eq:wv} no longer depends upon the specific measurement procedure, just like the expectation value in Eq.~\eqref{eq:expect}, so this approximation will work for \emph{any} unbiased weak measurement procedure. The only requirement for consistently recovering the weak value Eq.~\eqref{eq:wv} as the limiting value of the conditioned average in Eq.~\eqref{eq:condav} is for the Lindblad perturbation terms in Eq.~\eqref{eq:lindblad} to be small enough to neglect \cite{Dressel2012e,Dressel2013b}, meaning that the quantum state is approximately unperturbed. (For a physical example where the state disturbance from the coupling may not always be neglected, see \cite{Dressel2012c}.).
\subsection{Dynamical weak values}
\begin{figure}[t]
\includegraphics[width=0.9\columnwidth]{perturb.eps}
\caption{Energy level perturbations. The eigenvalues $E$ of a Hamiltonian $\op{H}$ are shifted to new eigenvalues $E'$ when a Hamiltonian perturbation $\op{\Delta}$ is added. These shifts in energy levels are exactly (real) weak values $\bra{E}\op{\Delta}\ket{E'}/\braket{E}{E'}$, and thus may lie outside the spectrum of the perturbation $\op{\Delta}$.}
\label{fig:perturb}
\end{figure}
Thus far we have carefully discussed the most well-known role of weak values in experiment, namely as complex parameters that characterize a von Neumann interaction, and as conditioned estimations of observable averages. However, these are not the only places that weak values naturally appear. Here we consider two more common cases that are usually overlooked: weak values as eigenvalue perturbations, and weak values as classical dynamical variables in reduced system dynamics.
The first case is mathematically trivial to show, but has nontrivial implications. Consider the case of a Hamiltonian $\op{H}$ with energy eigenstates $\ket{E}$ and eigenvalues $\op{H}\ket{E} = E\ket{E}$. Suppose this Hamiltonian becomes perturbed by a new contribution $\op{\Delta}$, producing new eigenstates $(\op{H}+\op{\Delta})\ket{E'} = E'\ket{E'}$. This latter eigenvalue equation can then be contracted with an unperturbed eigenstate $\bra{E}$ and rearranged to find the following relation between the eigenvalues:
\begin{align}
E' &= E + \frac{\bra{E}\op{\Delta}\ket{E'}}{\braket{E}{E'}},
\end{align}
which is illustrated in Figure~\ref{fig:perturb}. That is, the (purely real) weak value of the perturbation $\op{\Delta}$ determines the shift in energy for the eigenstates of the Hamiltonian, and thus may lie outside the spectrum of $\op{\Delta}$. One can understand this weak value as the best estimate of the average energy perturbation required to move from the old eigenstate $\ket{E}$ to the new eigenstate $\ket{E'}$. No measurement is being performed here, so this shift constitutes a dynamical effect where the weak value indeed represents the physical energy shift. This shift can be verified by measuring the eigenenergies before and after such a perturbation is added to the system.
\begin{figure}[t]
\includegraphics[width=\columnwidth]{cqed.eps}
\caption{Superconducting qubit capacitively coupled to a one-sided stripline resonator with energy-decay rate $\kappa$, which is pumped through a circulator with a coherent microwave source $\varepsilon(t)$ at a frequency detuned by $\Delta$ from the bare resonator frequency $\omega_r$. The resonator frequency is shifted by $\pm \chi$ depending on the qubit state due to the dispersive coupling. As such, each definite qubit state $\ket{0}$ or $\ket{1}$ approximately correlates to a distinct resonator state $\ket{\psi_0}$ or $\ket{\psi_1}$. The reduced qubit state $\op{\rho}_q(t)$ then displays coherence oscillations at a frequency $\omega_q + 2\chi\, \text{Re}\,n_w(t)$ due to the ac Stark shift, which depends on the real part of the weak value $n_w(t) = \bra{\psi_1}\op{a}^\dagger\op{a}\ket{\psi_0}/\braket{\psi_1}{\psi_0}$ of the resonator population. The qubit coherence similarly displays decay at an average rate $\Gamma = 2\chi\,\text{Im}\,n_w(t)$ that depends on the imaginary part of $n_w(t)$, indicating measurement-dephasing from ensemble-averaging the fluctuations of the resonator population. Importantly, the complex weak value $n_w(t)$ is \emph{physically} the relevant mean-field classical dynamical variable for the resonator population that affects the reduced qubit state at every point in time $t$, with the real part corresponding to the best classical estimation of the ensemble-averaged resonator population probed by coherent qubit superpositions of $\ket{0}$ and $\ket{1}$.}
\label{fig:cqed}
\end{figure}
The second case of a dynamical weak value is best illustrated by an explicit example, shown in Figure~\ref{fig:cqed}. Suppose we couple a qubit dispersively to a single-mode resonator (e.g., a superconducting qubit setup like the one used in \cite{Groen2013}). The simplest Hamiltonian for how such a joint system naturally evolves is
\begin{align}\label{eq:qubitham}
\op{H} &= \frac{\hbar\,\omega_q}{2}\,\op{\sigma}_z + \hbar\,\omega_r\, \op{a}^\dagger\op{a} + \hbar\,\chi\,\op{\sigma}_z\,\op{a}^\dagger\op{a},
\end{align}
where $\op{\sigma}_z = \pr{1} - \pr{0}$ is the Pauli Z-operator between the qubit energy levels, $\op{a}$ is the lowering operator of the resonator mode satisfying $[\op{a},\op{a}^\dagger]=1$, $\omega_q$ and $\omega_r$ are the oscillation frequencies of the qubit and resonator, and $\pm \chi$ is the dispersive frequency shift of the resonator that depends on the qubit state. Note that the interaction term between the qubit and the resonator has the general von Neumann form of Eq.~\eqref{eq:vnham}, but it is no longer impulsive.
Assuming a pure state for the joint qubit-resonator system, we can make the following ansatz for the form of the joint state
\begin{align}\label{eq:ansatz}
\ket{\Psi} &= c_0(t)\,\ket{0}\ket{\psi_0(t)} + c_1(t)\,\ket{1}\ket{\psi_1(t)},
\end{align}
where $c_{0,1}(t)$ are complex amplitudes, and $\ket{\psi_{0,1}(t)}$ are normalized resonator states that are correlated to each definite qubit state. It follows that the reduced qubit density matrix $\op{\rho}_q = \text{Tr}_r\pr{\Psi}$ after tracing out the resonator has the following diagonal populations
\begin{align}\label{eq:qubitpops}
p_0(t) &= |c_0(t)|^2, & p_1(t) &= |c_1(t)|^2,
\end{align}
as well as an off-diagonal coherence
\begin{align}\label{eq:qubitcoh}
\rho_{01}(t) &= c_1^*(t)c_0(t)\braket{\psi_1(t)}{\psi_0(t)},
\end{align}
that depends explicitly on the overlap between the two distinct and dynamically evolving resonator states that are correlated to definite qubit populations.
The simple Hamiltonian considered in Eq.~\eqref{eq:qubitham} is already diagonal in the energy basis of the qubit, so the populations in Eq.~\eqref{eq:qubitpops} do not change in time. However, the coherence in Eq.~\eqref{eq:qubitcoh} will display phase oscillations due to both the natural qubit energy splitting and the added influence of the dispersive resonator coupling. After extracting the components of the joint Schr\"odinger equation $i\hbar\partial_t \ket{\Psi} = \op{H}\ket{\Psi}$ by differentiating Eq.~\eqref{eq:ansatz}, and a bit of algebra, it is straightforward to derive the evolution equation for the coherence directly from differentiating Eq.~\eqref{eq:qubitcoh}
\begin{align}\label{eq:coherence}
\partial_t \rho_{01}(t) &= i[\omega_q + 2\chi\, n_w(t)]\,\rho_{01}(t).
\end{align}
Notably, the (complex) weak value of the resonator population naturally appears
\begin{align}\label{eq:resonatorwv}
n_w(t) &= \frac{\bra{\psi_1(t)}\op{a}^\dagger\op{a}\ket{\psi_0(t)}}{\braket{\psi_1(t)}{\psi_0(t)}} = \bar{n}(t) + i\bar{n}_\gamma(t).
\end{align}
This weak value can be understood as a \emph{classical dynamical variable} that completely determines the ensemble-averaged \emph{dynamical} influence of the resonator on the reduced qubit state at each point in time $t$, even in the absence of any explicit preselection or postselection measurements.
The real part $\bar{n}(t)$ of this weak value is the best estimate of the population of the resonator mode if the qubit transitions between its ground and excited states at time $t$. According to Eq.~\eqref{eq:coherence}, this weak value produces a shift $2\chi\bar{n}(t)$ of the natural qubit frequency, commonly known as the \emph{ac Stark shift} \cite{Bonch-Bruevich1969,Brune2004,Schuster2005}. Notably, this shift does not involve either of the average mode populations $n_0 = \bra{\psi_0}\op{a}^\dagger\op{a}\ket{\psi_0}$ or $n_1 = \bra{\psi_1}\op{a}^\dagger\op{a}\ket{\psi_1}$ that one might naively expect, since the qubit coherence does not pertain to a definite population. Any imaginary part $\bar{n}_\gamma(t)$ of the weak value does not contribute to the ac Stark shift in Eq.~\eqref{eq:coherence}, but instead produces a decay of the qubit coherence that indicates the resonator coupling is \emph{dephasing} the qubit at a rate $2\chi\bar{n}_\gamma(t)$.
For an explicit example of this reduced evolution, to a good approximation \cite{Gambetta2006,Gambetta2008,Boissonneault2009} a one-sided resonator with energy-decay rate $\kappa$ (as shown in Figure~\ref{fig:cqed}) that is pumped with a coherent state $\varepsilon$ detuned by $\Delta$ from the bare resonator frequency will reach a steady state that approximates the pure-state ansatz of Eq.~\eqref{eq:ansatz}, with $\ket{\psi_0}$ and $\ket{\psi_1}$ approximating coherent states with complex classical amplitudes
\begin{align}
\psi_0 &\equiv \bra{\psi_0}\op{a}\ket{\psi_0} = \frac{2\varepsilon}{\kappa}\frac{1}{1 + i2(\Delta - \chi)/\kappa}, \\
\psi_1 &\equiv \bra{\psi_1}\op{a}\ket{\psi_1} = \frac{2\varepsilon}{\kappa}\frac{1}{1 + i2(\Delta + \chi)/\kappa}.
\end{align}
The weak value of the resonator population in Eq.~\eqref{eq:resonatorwv} correspondingly has the simple steady-state form (assuming a wide-bandwidth resonator for brevity, with $\chi,\Delta \ll \kappa$)
\begin{align}
n_w = \psi_1^*\psi_0 \approx \frac{4\varepsilon^2}{\kappa^2}\left[1 + i\frac{4\chi}{\kappa}\right],
\end{align}
which produces the ac Stark shift $2\chi\bar{n}$ of the qubit frequency with $\bar{n} = 4\varepsilon^2/\kappa^2$, as well as the dephasing rate $\Gamma = 2\chi\bar{n}_\gamma = 8\chi^2\bar{n}/\kappa$. These expressions for the ac Stark shift and ensemble-average dephasing rate agree with the known results for dispersive qubit measurements in circuit quantum electrodynamics (cQED) \cite{Gambetta2006,Gambetta2008,Boissonneault2009,Korotkov2011}.
We emphasize that the complex weak value of the resonator population in Eq.~\eqref{eq:resonatorwv} is \emph{physically} the relevant classical dynamical variable that controls the behavior of the ensemble-averaged reduced qubit evolution in Eq.~\eqref{eq:coherence}. Indeed, the ac Stark shift of the qubit frequency is the primary method used in cQED for extracting the average population $\bar{n}$ in the resonator at steady state (e.g., \cite{Jeffrey2014}), but we see here that in actuality such a dynamical method probes the \emph{weak value} of that population, and not the population associated with any particular qubit state, or even the total average population $\bar{n}_{\text{tot}} = \bra{\Psi}\op{a}^\dagger\op{a}\ket{\Psi}$. The weak value arises from the interference between the two fields $\ket{\psi_0}$ and $\ket{\psi_1}$ in the resonator that are correlated with the two definite qubit states, on average.
At each time $t$ the real part of $n_w$ indicates the best estimate of the average resonator excitation number seen by a qubit that is not in a definite energy eigenstate, while its imaginary part indicates the best estimate of the backaction on the qubit dynamics caused by the fluctuations of the resonator population around that average value. These latter fluctuations result in ensemble-average dephasing of the reduced qubit state, reinforcing the observation made for weak measurements that a weak value can only describes the average (i.e., classical mean field) state of affairs for an ensemble of realizations \cite{Dressel2014b}. Indeed, when the time-dependent leakage from the resonator is accounted for, the qubit will not dephase in this manner, but will instead follow a pure \emph{quantum trajectory} \cite{Gambetta2008,Korotkov2011,Murch2013,Weber2014} that depends on the leakage record. As such, we must necessarily interpret the weak value $n_w$ here as implicitly averaging over many such realizations in practice to produce a classical dynamical variable associated with the classical mean field in the resonator.
\section{Disturbance, quasiprobabilities, and Hamilton-Jacobi}\label{sec:quasiprob}
Given the consistent role of the real part of a weak value as a best conditioned observable estimate, we can now observe an intriguing logical tension inherent to the weak value. On one hand, any classical conditioned average must include disturbance to obtain anomalous values \cite{Williams2008,Dressel2011,Dressel2012b,Tollaksen2007,Ipsen2014}: the larger the disturbance, the more strange the average can become. On the other hand, the strangeness of the conditioned average in Eq.~\eqref{eq:condav} is greatest when the quantum state is \emph{least} disturbed by an intermediate measurement \cite{Dressel2012e}, and even persists when there is no added disturbance to the natural dynamical evolution in the example of Eq.~\eqref{eq:coherence}.
These two statements imply that any classical (hidden-variable) explanation of a strange weak value as a disturbance effect must satisfy one of two properties: either (a) the quantum state must be a \textit{subjective} (epistemic) quantity that is completely insensitive to whatever physical (ontic) disturbance is occurring, or (b) the relevant disturbance occurs entirely during the \textit{postselection}, and not the intermediate measurement \cite{Dressel2012b}.
Classical fields that produce strange weak values in weak measurement experiments satisfy this second property, where the disturbance at the postselection filter causes interference between previously independent field components \cite{Ritchie1991,Bliokh2013a,Bliokh2014,Dressel2014b}. However, quantum systems can display similar interference without wavelike intensities \cite{Pryde2005,Goggin2011,Groen2013}, and permit additional entanglement effects \cite{Dressel2011} such as the dynamical evolution involving weak values emphasized in the previous section. Especially for such a dynamical role of the weak value, it does not seem possible to ascribe strange weak values to classical disturbance mechanisms without also demanding that the quantum state is itself a fundamentally subjective collection of some more physical microstates; such a demand, while not impossible, must contend with the Pusey-Barrett-Rudolph theorem \cite{Pusey2012} and the other no-go theorems (reviewed in \cite{Leifer2014}) for states of such epistemic character.
\subsection{Terletsky-Margenau-Hill and Kirkwood-Dirac}
To quantify this logical tension, we can express weak values in a more established and familiar way by rewriting Eq.~\eqref{eq:wv} using the spectral expansion $\op{A} = \sum_a \,a\,\pr{a}$ to find $A_w(t) = \sum_a\,a\,\tilde{p}_{a|i,f}$, where
\begin{align}\label{eq:quasicond}
\tilde{p}_{a|i,f} &= \frac{\text{Re}\,\braket{f'}{a}\braket{a}{i'}\braket{i'}{f'}}{|\braket{f'}{i'}|^2}.
\end{align}
This is a conditional \textit{quasiprobability} distribution that weights the eigenvalues of $\op{A}$ in $A_w$, and satisfies the normalization $\sum_a \tilde{p}_{a|i,f} = 1$. As a result, if a strange weak value $|A_w| > ||\op{A}||$ is estimated, then at least one quasiprobability must be negative: $\tilde{p}_{a|i,f} < 0$.
Since the conditioning denominator of Eq.~\eqref{eq:quasicond} is positive-definite, we infer that the joint quasiprobability
\begin{align}\label{eq:kirkwood}
\tilde{p}_{a,f|i} &= \text{Re}\,\braket{f'}{a}\braket{a}{i'}\braket{i'}{f'}
\end{align}
in the numerator of Eq.~\eqref{eq:quasicond} [also appearing directly in the experimental partial average of Eq.~\eqref{eq:partial2}] must be negative. This joint quasiprobability distribution is precisely the \emph{Terletsky-Margenau-Hill distribution} \cite{Terletsky1937,Margenau1961,Johansen2004,Johansen2004c} that has been used since the late 1930s.
Interestingly, the Terletsky-Margenau-Hill distribution is the real part of the complex quasiprobability distribution introduced even earlier by Kirkwood \cite{Kirkwood1933,Dirac1945,Chaturvedi2006,Lundeen2011,Lundeen2012,Hofmann2012b,Lundeen2014} as an alternative to the Wigner distribution \cite{Wigner1932}. This Kirkwood distribution is also known as the \emph{standard-ordering distribution} for quantum phase space \cite{Mehta1964}. In fact, Dirac later considered this distribution specifically to discuss the classical-to-quantum transition \cite{Dirac1945}, observing that the negativity arises directly from the usual operator noncommutativity of quantum mechanics. Notably, the fully complex weak value that appears in reduced state dynamics is nothing more than a conditioned version of this complex Kirkwood-Dirac quasiprobability distribution.
An important feature of the Kirkwood-Dirac distribution that has recently come to light \cite{Lundeen2011,Lundeen2012} is that any quantum state can be written in an operator basis such that this distribution forms its components. That is, if we define a suitable operator basis
\begin{align}
\Gamma_{a,f} = \frac{\ket{a}\bra{f}}{\braket{f}{a}}
\end{align}
then we can write any quantum state using the expansion
\begin{align}
\op{\rho} = \sum_{a,f} \,\braket{f}{a}\bra{a}\op{\rho}\ket{f}\,\Gamma_{a,f}.
\end{align}
As such, the quasiprobabilities $\rho(f,a) = \braket{f}{a}\bra{a}\op{\rho}\ket{f}$ of the complex Kirkwood-Dirac distribution are a complete quantum state representation for an arbitrary density matrix that is analogous to a complex wavefunction $\psi(x) = \braket{x}{\psi}$ for a pure state. Unlike the usual wavefunction, however, the Kirkwood-Dirac distribution is directly compatible with Bayes' theorem [as used in Eq.~\eqref{eq:quasicond}], and thus behaves like a true probability distribution. This notable feature has enabled alternative methods of quantum state tomography by directly measuring complex weak values using von Neumann interactions \cite{Lundeen2011,Lundeen2012,Salvail2013,Malik2014,Howland2014,Mirhosseini2014,Hofmann2014}, and also permits fully Bayesian quasiprobabilistic reformulations of coherent quantum dynamics \cite{Hofmann2012b,Lundeen2014}.
\subsection{Wigner distribution and negativity}
Importantly, the negativity in such a quasiprobability representation of a quantum state is closely associated to traditional measures of nonclassicality \cite{Spekkens2008,Ferrie2011}. The usual examples of this criterion for nonclassicality are the negativity of the Wigner distribution \cite{Wigner1932}, or the Glauber-Sudarshan $P$-distribution \cite{Glauber1963,Sudarshan1963}; however, the Terletsky-Margenau-Hill distribution in Eq.~\eqref{eq:kirkwood} has the same feature \cite{Johansen2004c}.
To emphasize this point, we can in fact relate weak values directly to the Wigner distribution when we are considering infinite-dimensional systems. To see this, consider the Wigner distribution for an initially pure state $\ket{i}$ \cite{Wigner1932}
\begin{align}
W_i(x,p) &= \int_{-\infty}^\infty\!\! \braket{x-\frac{y}{2}}{i}\braket{i}{x+\frac{y}{2}}e^{ipy/\hbar}\frac{dy}{2\pi\hbar},
\end{align}
where $x$ and $p$ represent the usual classical position and momentum variables. Now suppose we compute the partial average of the momentum $p$ for a fixed $x$ using this joint quasiprobability distribution
\begin{align}\label{eq:wignerpartial}
&\int_{-\infty}^\infty\!\! p\,W_i(x,p)dp = {} \\
&\quad = \iint_{-\infty}^\infty\!\! \braket{x-\frac{y}{2}}{i}\braket{i}{x+\frac{y}{2}}(-i\hbar\partial_y e^{ipy/\hbar})\frac{dydp}{2\pi\hbar}, \nonumber \\
&\quad = \int_{-\infty}^\infty\!\! i\hbar\partial_y \left[\braket{x-\frac{y}{2}}{i}\braket{i}{x+\frac{y}{2}}\right] \int_{-\infty}^\infty\!\! e^{ipy/\hbar}\frac{dp}{2\pi\hbar}dy, \nonumber \\
&\quad = \int_{-\infty}^\infty\!\! \frac{(-i\hbar\partial_x\braket{x-\frac{y}{2}}{i})\braket{i}{x+\frac{y}{2}} + \text{c.c.}}{2}\, \delta(y)\,dy, \nonumber \\
&\quad = \text{Re}(-i\hbar\partial_x\braket{x}{i})\braket{i}{x}, \nonumber \\
&\quad = \text{Re}\bra{x}\op{p}\ket{i}\braket{i}{x}, \nonumber \\
&\quad = \int_{-\infty}^\infty\!\! p\, \text{Re}\braket{x}{p}\braket{p}{i}\braket{i}{x}\,dp. \nonumber
\end{align}
In other words, after performing integration-by-parts the partial average of the Wigner function yields precisely the same result as averaging $p$ over the Terletsky-Margenau-Hill distribution of Eq.~\eqref{eq:kirkwood}. It follows that conditioning this partial average directly produces a momentum weak value as the proper local average of momentum at the position $x$
\begin{align}\label{eq:wignerwv}
\frac{\int_{-\infty}^\infty p\,W_i(x,p)dp}{\int_{-\infty}^\infty W_i(x,p)dp} = \text{Re}\frac{\bra{x}\op{p}\ket{i}}{\braket{x}{i}}.
\end{align}
As with Eq.~\eqref{eq:quasicond}, the denominator is a positive-definite probability, so only the partial average of Eq.~\eqref{eq:wignerpartial} in the numerator of Eq.~\eqref{eq:wignerwv} can produce nonclassical behavior.
The inescapable conclusion is that weak values are intimately related to quasiprobability distributions for the quantum formalism in a fundamental and unavoidable way \cite{Aharonov2005b}. Moreover, these conditioned averages consistently confirm the best estimate that we expected from the derivation of Eq.~\eqref{eq:wv}. Similar relationships can be established for the other quantum phase space Moyal distributions \cite{Moyal1949}, of which the Wigner function and the Glauber-Sudarshan $P$-distribution \cite{Glauber1963,Sudarshan1963} are special-cases. (See \cite{Bednorz2013} for an example that emphasizes this point using non-symmetrized correlation functions.)
We therefore have the following observation: if strange conditioned averages approximate the functional dependence of the weak value in Eq.~\eqref{eq:wv}, then classical hidden variable models will be unable to satisfactorily explain that dependence. If one could, then it would also be able to reproduce other nonclassical statistical features of the quantum theory that arise from the negativity of these quasiprobability distributions (see also \cite{Tollaksen2007,Pusey2014}). Such negativity, in turn, arises from intrinsic quantum interference that is not present in classical systems of particles.
\subsection{Hamilton-Jacobi Formalism}\label{sec:hamilton}
As a last, albeit poignant, illustration of the pervasive theoretical role of weak values in the quantum mechanical formalism, let us re-examine the Schr\"odinger equation for a nonrelativistic particle
\begin{align}\label{eq:schrodinger}
i\hbar \partial_t \ket{\psi} &= \left[\frac{\op{p}^2}{2m} + V(\op{x})\right]\ket{\psi}.
\end{align}
We can expand this equation into a wavefunction form that is local in the coordinate $x$ by contracting it with $\bra{x}$ and rearranging
\begin{align}
\partial_t (i\hbar \ln\braket{x}{\psi}) &= \frac{1}{2m}\frac{\bra{x}\op{p}^2\ket{\psi}}{\braket{x}{\psi}} + V(x).
\end{align}
This local form automatically involves the weak values of the kinetic and potential energies.
Noting the familiar structure of this equation, we can then define a complex version of \emph{Hamilton's principle function} (i.e., the classical action) as
\begin{align}
S(x,t) &\equiv -i\hbar \ln\braket{x}{\psi},
\end{align}
with real and imaginary parts
\begin{align}
\text{Re}S(x,t) &= \hbar \Phi(x,t), \\
\text{Im}S(x,t) &= -\frac{\hbar}{2}\ln \rho(x,t),
\end{align}
that naturally isolate the phase (i.e., \emph{eikonal}) $\Phi$ and probability density $\rho$ components of the wavefunction $\braket{x}{\psi} = \sqrt{\rho(x,t)}\,\exp[i\Phi(x,t)]$.
From this complex action, we can define the classically \emph{conjugate momentum field} $p(x,t)$ that corresponds to the local coordinate $x$ in the usual way
\begin{align}\label{eq:momentum}
p(x,t) &= \partial_x S(x,t) = -i\hbar\partial_x \ln\braket{x}{\psi} = \frac{\bra{x}\op{p}\ket{\psi}}{\braket{x}{\psi}},
\end{align}
which naturally produces a complex momentum weak value as the appropriate classical dynamical variable for the local momentum field. The real part of this weak value is the \emph{Bohmian momentum} \cite{Bohm1952a,Bohm1952b,Wiseman2007,Kocsis2011}
\begin{align}\label{eq:bohmmom}
\text{Re}\,p(x,t) &= \hbar \partial_x \Phi(x,t),
\end{align}
while the imaginary part is the \emph{osmotic momentum} \cite{Nelson1966,Bohm1989}
\begin{align}\label{eq:osmom}
\text{Im}\,p(x,t) &= -\frac{\hbar}{2}\partial_x \ln \rho(x,t),
\end{align}
that indicates the logarithmic change of the probability distribution $\rho(x,t) = |\braket{x}{\psi}|^2$ as $x$ is varied \cite{Dressel2012d}. These local momentum fields jointly act as the average fluid-like momentum that corresponds to a fluctuating classical mean field for the quantum particle \cite{Madelung1926,Madelung1927}.
Using this complex action, we can thus identically rewrite Schr\"odinger's Eq.~\eqref{eq:schrodinger} as a straightforward \emph{Hamilton-Jacobi} equation for the classical mean-field $x$ and $p(x,t)$ dynamical variables \cite{Hiley2012}
\begin{align}\label{eq:hamjac}
0 &= \partial_t S(x,t) + H[x,p(x,t),t],
\end{align}
where the classical Hamiltonian function has the form
\begin{align}
H[x,p(x,t),t] &= \frac{1}{2m}\frac{\bra{x}\op{p}^2\ket{\psi}}{\braket{x}{\psi}} + V(x),
\end{align}
involving local position weak values of the kinetic and potential energies.
Splitting the complex Hamilton-Jacobi equation of Eq.~\eqref{eq:hamjac} into its independent real and imaginary parts produces
\begin{align}
\label{eq:hamjacreal}
0 &= \hbar \partial_t \Phi(x,t) + \frac{[\text{Re}\,p(x,t)]^2}{2m} + V(x) + Q(x), \\
\label{eq:continuity}
0 &= \partial_t \rho(x,t) + \partial_x \left[\rho(x,t)\frac{p(x,t)}{m}\right].
\end{align}
Note that Eq.~\eqref{eq:continuity} is nothing more than the usual continuity equation for the probability density $\rho(x,t)$ in terms of the local Bohmian velocity field $v(x,t) = p(x,t)/m$, while Eq.~\eqref{eq:hamjacreal} is the Bohmian Hamilton-Jacobi equation that contains the local kinetic energy, potential energy, and what is usually known as the \emph{quantum potential energy}
\begin{align}\label{eq:qpotential}
Q(x,t) &\equiv \frac{\text{Var}[\text{Re}\,p(x,t)]}{2m} = -\frac{\hbar^2}{2m}\frac{\partial_x^2\sqrt{\rho(x,t)}}{\sqrt{\rho(x,t)}},
\end{align}
which is proportional to the \emph{weak momentum variance}
\begin{align}\label{eq:wvar}
\text{Var}[\text{Re}\,p(x,t)] &= \text{Re}\frac{\bra{x}[\op{p}-\text{Re}p(x,t)]^2\ket{\psi}}{\braket{x}{\psi}}, \\
&= \text{Re}\frac{\bra{x}\op{p}^2\ket{\psi}}{\braket{x}{\psi}} - \left[\text{Re}\frac{\bra{x}\op{p}\ket{\psi}}{\braket{x}{\psi}}\right]^2. \nonumber
\end{align}
The classical particle trajectory limit corresponds to the \emph{eikonal approximation}, or \emph{ray approximation}, (familiar from geometrical optics) where we can neglect rapid spatial changes in the probability density $\rho(x,t)$ that arise from the wave-like quantum interference. This approximation allows us to neglect the spatial derivatives of $\rho(x,t)$ that appear in the weak momentum variance of Eq.~\eqref{eq:wvar} (or quantum potential of Eq.~\eqref{eq:qpotential}), as well as in the osmotic momentum of Eq.~\eqref{eq:osmom}, implying that the local Bohmian momentum field in Eq.~\eqref{eq:bohmmom} has precise and non-fluctuating values at each definite position. In this limit, Eq.~\eqref{eq:hamjac} becomes purely real and identified with the usual classical Hamilton-Jacobi equation, while the momentum weak value of Eq.~\eqref{eq:momentum} reduces to the purely real Bohmian momentum and becomes identified with the usual classical momentum field, which describes families of parallel classical particle trajectories.
Evidently, at every stage of the usual quantum-classical correspondence through the Hamilton-Jacobi formalism, weak values directly produce the proper classical mean-field dynamical variables, so are the appropriate conditional estimates akin to the expectation values found in the Ehrenfest theorem. This fundamental role of the weak value corroborates the various interpretations we have encountered in this paper, and makes it clear that weak values are an inextricable feature of the quantum formalism that directly pertains to the classical mean-field limit \cite{Dressel2014b}.
\section{Conclusion}\label{sec:conclusion}
Quantum weak values have endured a controversial history, despite the fact that they are essential to the quantum formalism, with their seeming strangeness being a direct consequence of quantum interference. As estimates of observable averages within a bracketed time-window, or as classical dynamical variables for the mean field evolution, their potential for having anomalous values reflects the nonclassicality of the quantum probabilistic structure, and is equivalent to the need for negative quasiprobabilities. This negativity fundamentally arises as an interference effect, both for the classical mean fields and for manifestly quantum systems that permit discrete detection events and entanglement. No classical model can faithfully reproduce the functional dependence of weak value anomalies without also simulating the corresponding features of quantum mechanics.
\begin{acknowledgments}
The author is grateful for discussions with A. N. Jordan, P. B. Dixon, F. Nori, and E. A. Sete. The research was funded by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), through the Army Research Office (ARO) Grant No. W911NF-10-1-0334, and also supported from the ARO MURI Grant No. W911NF-11-1-0268.
\end{acknowledgments}
|
\section{Introduction}
The notion of the topological entropy was introduced by Adler, Konheim and McAndrew in 1965 in \cite{bib:AdlKonMcA65}. Another approach was presented by Bowen \cite{bib:Bow71} in early 70's. Ghys, Langevin and Walczak, in \cite{bib:GhyLanWal88}, extended this notion to the topological entropy for finitely generated groups and pseudogroups of continuous transformations, as well as the geometric entropy of foliation on a compact foliated Riemannian manifold. The entropy of foliation has more geometric nature, because it depends on a Riemannian metric chosen for foliated manifold. On the other hand, the dynamics of Finsler spaces have become a subject of a research of mathematicians in late 90's and recently. However, the research in this field is rather in the initial phase.
The aim of this paper is to extend the notion of the geometric entropy of foliations to the foliated manifolds equipped with leafwise Finsler structure. In paragraphs 2 and 3, one can find all necessary definitions and properties related to entropy and foliations with leafwise Finsler metric. Next paragraph describes relations between geometric and topological entropy. Fifth part of the paper refers to foliations with leafwise Randers norm. Last paragraph describes the entropy of one dimensional foliations defined by a unit vector field with leafwise Randers metric.
\section{Leafwise Finsler structures}
Let us recall that a {\it Minkowski norm} on a vector space $V$ is a non-negative function $F:V\to[0,\infty)$ such that
\begin{enumerate}
\item $F$ is $C^{\infty}$ on $V\setminus\{0\}$,
\item $F(\lambda v)=\lambda F(v)$ for any $\lambda>0$ and $v\in V$,
\item for every $y\in V\setminus \{0\}$, the symmetric bilinear form
\[
g_y (u,v):= \frac{1}{2} \frac{\partial ^2}{\partial t\partial s} F^2(y+su+tv)|_{t=s=0}
\]
is positively defined.
\end{enumerate}
Now, let $M$ be a smooth manifold. A function $F:TM\to [0,\infty)$ is called a {\it Finsler norm} if
\begin{enumerate}
\item $F$ is $C^{\infty}$ on the tangent bundle with removed the zero section $TM\setminus \{0\}$,
\item for any $x\in M$ the restricted norm $F_x=F|_{T_xM}$ is a Minkowski norm.
\end{enumerate}
The pair $(M,F)$ is called a Finsler space.
\begin{ex}\label{ex:Randers norm}
Let $(M,g)$ be a Riemannian manifold, and let $\beta:TM\to\mathbb{R}$ be a $1$-form. Let $\alpha:TM\to [0,\infty)$ be the norm defined by $g$, that is, $\alpha(v)=\sqrt{g_x(v,v)}$ for all $v\in T_x M$. Suppose that the $g$-norm of $\beta$ satisfies $||\beta||_g < 1$. We set $F(v)=\alpha(v)+\beta(v)$. $F$ is a Finsler norm and it is called a {\it Randers norm}.
\end{ex}
Note that the Finsler norm induces a function $d:M\times M\to [0,\infty)$ by the formula
\[
d(x,y)=\inf_{\gamma} \int_0^1 F(\dot{\gamma}(t)) dt
\]
where the infimum is taken over all curves $\gamma:[0,1]\to M$ linking $x$ and $y$. Function $d$ is a quasi-metric, that is,
\[
(d(x,y)=0 \textrm{ iff } x=y) \textrm{ and } d(x,y)+d(y,z)\geq d(x,z).
\]
Let $(M,\f,g)$ be a foliated Riemaniann manifold. Having $g$, we decompose the tangent bundle to the orthogonal sum of the bundle tangent to $\f$ and the orthogonal bundle, that is, $TM=T\f \oplus T\f^{\perp}$. We replace the norm induced in $T\f$ by the Riemannian structure $g|_{T\f}$ by a Finsler norm $F_{\f}$. Denote by $\pi_1:TM\to T\f$ and $\pi_2:TM\to T\f^{\perp}$ the natural projections. We set
\[
F(v)=\sqrt{F_{\f}^2(\pi_1(v))+g(\pi_2(v),\pi_2(v))}.
\]
$F$ is a Finsler norm on $TM$ and coincides with $\sqrt{g(v,v)}$ for $v\in T\f^{\perp}$ and with $F_{\f}$ for $v\in T\f$. We call $F$ a \emph{leafwise Finsler structure} on $(M,\f)$.
\section{Geometric entropy of foliations with leafwise Finsler metric}
Let $(M,\f,F)$ be a foliated manifold with leafwise Finsler structure. Let $\u$ be a \textit{nice covering}, i.e., a covering by the domains $D_{\varphi}$ of the charts of a nice foliated atlas $\a$, that is an atlas satisfying
\begin{enumerate}
\item the covering $\{D_{\varphi}: \varphi\in\a\}$ is locally finite,
\item for any $\varphi\in\a$, the set $R_{\varphi}=\varphi(D_{\varphi})\subset \rr^n$ is an open cube,
\item if $\varphi,\psi\in\a$, and $D_{\varphi}\cap D_{\psi}\neq\emptyset$, then there exists a chart $\chi=(\chi',\chi'')$ such that for any leaf $L$ of $\f$ the connected components of $L\cap D_{\chi}$ are given by the equation $\chi''={\rm const}$, and $R_{\chi}$ is an open cube, $D_{\chi}$ contains the closure of $D_{\varphi}\cup D_{\psi}$ and $\varphi=\chi|_{D_{\varphi}}$ and $\psi=\chi|_{D_{\psi}}$.
\end{enumerate}
Let $U\in\u$. Equip the space of plaques $T_U=U/_{\f|U}$ with the quotient topology. The disjoint union $T=\bigsqcup \{T_U; U\in\u\}$ is called a {\it complete transversal} for $\f$. Note that each of $T_U$ can be mapped homeomorphically onto a $C^r$-submanifold $T'_U\subset U$ transverse to $\f$.
Following \cite{bib:GhyLanWal88}, let us recall that for a given nice covering $\u$ of $(M,\f)$ there exists an $\varepsilon_0>0$ such that any point $x\in U$, $U\in \u$, can be projected orthogonally in the unique way to the plaque $P_y\subset U$ passing through a point $y\in U$ if only $\max\{ d(x,y),d(y,x)\}<\varepsilon_0$.
Let $\gamma:[0,1]\to L$ be a leafwise curve beginning at $x\in U$. For any $y\in U$ lying within the distance $\varepsilon<\varepsilon_0$, we can project orthogonally an initial part of the curve $\gamma$ to the plaque $P_y$ passing through $y$. Replacing $x$ and $y$ by the endpoints of the already projected piece and its image $\gamma_1$, we can continue this process as long as the distance between $\gamma$ and $\gamma'$ does not exceed $\varepsilon_0$. We will denote the projection of $\gamma$ by $p_y\gamma$.
Let $\u$ be a nice covering and let $T$ be the complete transversal for $\u$. Let $\varepsilon\in (0,\varepsilon_0)$, and let $d$ denotes the metric induced by the Finsler structure.
\begin{definition}
We say that $x,y\in T$ are $(R,\varepsilon)$-separated by $\f$ with respect to $F$ if either
\begin{itemize}
\item $\max\{d(x,y),d(y,x)\}\geq\varepsilon_0$
\end{itemize}
or
\begin{itemize}
\item there exists a leaf curve $\gamma:[0,1]\to L_x$ such that $\gamma(0)=x$, $l(\gamma)=\int_0^1 F(\dot{\gamma}(t))dt \leq R$ and
\[
\max\{d(\gamma(1),p_y\gamma(1)),d(p_y\gamma(1),\gamma(1))\}\geq \varepsilon.
\]
(or a leaf curve $\gamma:[0,1]\to L_y$ such that $\gamma(0)=y$, $l(\gamma)\leq R$, and $\max\{d(\gamma(1),p_x\gamma(1)),d(p_x\gamma(1),\gamma(1))\}\geq \varepsilon$).
\end{itemize}
A subset $A\subset T$ is called $(R,\varepsilon)$-separated if any two points $x,y\in A$, $x\neq y$, are $(R,\varepsilon)$-separated. Let $s(R,\varepsilon,\f)$ denote the maximum cardinality of an $(R,\varepsilon)$-separated subset of $T$. We set $s(\varepsilon,\f) = \limsup\limits_{R\to\infty} \frac{1}{R} \log s(R,\varepsilon,\f)$, and
\[
h(\f,F) = \lim_{\varepsilon\to 0^+} s(\varepsilon,\f).
\]
\end{definition}
\begin{rem}
The number $h(\f,F)$ does not depend on the choice of the nice covering $\u$. Let $\u'$ and $T'$ be another nice covering and complete transversal. Let $\varepsilon>0$ be small enough, and let us denote by $d_{\f}$ the leafwise metric induced by the Finsler structure $F_{\f}$. Since $M$ is compact, the geometry of $\f$ is bounded. Hence, one can project $T$ onto $T'$ in such a way that any $(R,\varepsilon)$-separated points $x,y\in T$ are projected to $x',y'\in T'$, respectively, which are $(R+R_0,\varepsilon)$-separated with $R_0$ being the maximum of the numbers $d_{\f}(x,x')$ and $d_{\f}(x',x)$, $x\in T\cap U$, $x'\in T'\cap U'$, $U\in\u$, $U'\in \u'$, and the plaques $P_x\subset U$ and $P_{x'}\subset U'$ intersects. Thus
\[
s'(R-R_0,\varepsilon,\f)\leq s(R,\varepsilon,\f)\leq s'(R+R_0,\varepsilon,\f),
\]
and both numbers $h(\f,F)$ and $h'(\f, F)$ are equal.
\end{rem}
\begin{rem}
Since any two Riemannian structures $g$ and $g'$ on a compact manifold satisfies
\[
c^{-2} g(v,w) \leq g'(v,w)\leq c^2 g(v,w)
\]
for some constant $c>1$, then the number $h(\f,F)$ does not depend on the choice of the Riemannian part of $F$. Indeed, there exists a constant $a>1$ such that for any leaf curve $\gamma$ and its orthogonal projections $p_y\gamma$ and $p_y'\gamma$, with respect to $g$ and $g'$ respectively, satisfies
\[
d(\gamma(t),p_y\gamma(t)) \leq a\cdot d'(\gamma(t),p_y'\gamma(t)),
\]
if $d(\gamma(t),p_y'\gamma(t))<\varepsilon$ for sufficiently small $\varepsilon>0$. Thus any two $(R,\varepsilon)$-separated points with respect to $F = \sqrt{F_{\f}^2+g}$ are $(R,\frac{\varepsilon}{a})$-separated with respect to $F' = \sqrt{F_{\f}^2+g'}$, and $ h(\f,F) \leq h(\f,F')$. Analogically we show that $h(\f,F') \leq h(\f,F)$.
\end{rem}
Since any two leafwise Finsler structures $F$ and $F'$ on a compact foliated manifold satisfies
\[
\frac{1}{c} F(v) \leq F'(v) \leq c\cdot F(v)
\]
for some constant $c\geq 1$, the geometric entropies $h(\f,F)$ and $h(\f,F')$ are both either equal to zero or not. The number $h(\f,F)$ is called the {\it geometric entropy of foliation with leafwise Finsler structure}. In further consideration we will denote by $F$ both, the structure $F_{\f}$ and the leafwise Finsler structure $F=\sqrt{F_{\f}^2+g}$.
\section{Relation between geometric entropy and topological entropy of holonomy pseudogroup}
Let $(M,\f)$ be a compact foliated manifold. Following \cite{bib:GhyLanWal88} or \cite{bib:Wal04}, one can define the topological entropy of the holonomy pseudogroup $\h_{\u}$ defined by the nice covering $\u$. The symbol $D_f$ denotes here the domain
|
\section{Introduction}
Several decay modes of $B$ decays into three pseudoscalar octet mesons PPP have been measured\cite{3p-data-babar,3p-data-lhcb}. $B\to PPP$ has been a subject of theoretical studies\cite{he3}. The new data have raised new interests in related theoretical studies\cite{yadong,gronau,xu-li-he1,he-li-xu2,hy-cheng,gronau-full}. With more data from LHCb, one can expect that the study of $B\to PPP$ will provide more important information for B decays in the standard model (SM).
A powerful tool to analyze B decays is flavor $SU(3)$ symmetry\cite{su31}.
Some of the interesting features of using flavor $SU(3)$ are the predictions of relations among different decay modes which can be experimentally tested. The flavor $SU(3)$ symmetry is, however, expected to be only an approximate symmetry because $u$, $d$ and $s$ quarks have different masses. Since the strange quark has a relative larger mass compared with those of up and down quarks, it is the larger source of symmetry breaking. If up and down quark masses are neglected, a non-zero strange quark mass breaks flavor $SU(3)$ symmetry down to the isospin symmetry. When up and down quark mass difference is kept, isospin symmetry is also broken. The $SU(3)$ breaking effect is at the level of 20 percent for the $\pi$ and $K$ decay constants $f_\pi$ and $f_K$. For 2-body pseudoscalar octet meson $B$ decays, although there are some $SU(3)$ breakings~\cite{su32}, it works reasonably well, such as rate differences between some of the $\Delta S=0$ and $\Delta S=1$ two-body pseudoscalar meson $B$ decays~\cite{he1,he2}. Analysis has also been carried out for $B\to PPP$ decays using flavor $SU(3)$ recently. It has been shown that the decay and CP asymmetry patterns for the charged $B^+$ decays into $K^+K^-K^+$, $K^+K^-\pi^+$, $K^+\pi^-\pi^+$ and $\pi^+\pi^-\pi^+$ do not follow $SU(3)$ predictions. To explain data, large $SU(3)$ breaking effects are needed\cite{xu-li-he1,he-li-xu2}. Usually isospin breaking effects are much smaller because up and down quark masses are much smaller than the strange quark mass and the QCD scale.
Because of possible large flavor $SU(3)$ breaking effects for $B \to PPP$, the predicted relations among different decay modes can only provide limited information. One wonders whether there exist relations which are immuned from $SU(3)$ or even isospin breaking effects due to $u$, $d$ and $s$ quark mass differences. To this end we carried out an analysis for $B \to PPP$ decays using flavor $SU(3)$ symmetry to identify possible relations, and then include $SU(3)$ breaking effects due to a strange quark mass, and also up and down quark masses to see whether some relations still remain to hold. We find that the relations between several fully-symmetric $B\to PPP$ decay amplitudes studied in Ref. \cite{gronau-full} are not affected by the flavor $SU(3)$ breaking effects due a non-zero strange quark mass, and some of them are not even affected by isospin breaking effects. These relations when measured experimentally can provide useful information about $B$ decays in the SM. In the following we provide some details.
\section{$SU(3)$ conserving amplitudes}
We start with the description of $B$ decays into three pseudoscalar octet mesons from flavor $SU(3)$ symmetry.
The leading quark level effective Hamiltonian up to one loop level in
electroweak interaction for hadronic charmless $B$ decays in the SM can be written as
\begin{eqnarray}
H_{eff}^q = {4 G_{F} \over \sqrt{2}} [V_{ub}V^{*}_{uq} (c_1 O_1 +c_2 O_2)
- \sum_{i=3}^{12}(V_{ub}V^{*}_{uq}c_{i}^{uc} +V_{tb}V_{tq}^*
c_i^{tc})O_{i}],
\end{eqnarray}
where $q$ can be $d$ or $s$the coefficients
$c_{1,2}$ and $c_i^{jk}=c_i^j-c_i^k$, with $j$ and $k$ indicate the internal quark,
are the Wilson Coefficients (WC). The tree WCs are of order one with, $c_1=-0.31$, and $c_2 = 1.15$. The penguin
WCs are much smaller with the largest one $c_6$ to be $-0.05$. These
WC's have been evaluated by several groups~\cite{heff}. $V_{ij}$ are the KM matrix elements.
In the above the factor $V_{cb}V_{cq}^*$ has
been eliminated using the unitarity property of the KM matrix.
The operators $O_i$ are given by
\begin{eqnarray}
\begin{array}{ll}
O_1=(\bar q_i u_j)_{V-A}(\bar u_i b_j)_{V-A}\;, &
O_2=(\bar q u)_{V-A}(\bar u b)_{V-A}\;,\\
O_{3,5}=(\bar q b)_{V-A} \sum _{q'} (\bar q' q')_{V \mp A}\;,&
O_{4,6}=(\bar q_i b_j)_{V-A} \sum _{q'} (\bar q'_j q'_i)_{V \mp A}\;,\\
O_{7,9}={ 3 \over 2} (\bar q b)_{V-A} \sum _{q'} e_{q'} (\bar q' q')_{V \pm A}\;,\hspace{0.3in} &
O_{8,10}={ 3 \over 2} (\bar q_i b_j)_{V-A} \sum _{q'} e_{q'} (\bar q'_j q'_i)_{V \pm A}\;,\\
O_{11}={g_s\over 16\pi^2}\bar q \sigma_{\mu\nu} G^{\mu\nu} (1+\gamma_5)b\;,&
O_{12}={Q_b e\over 16\pi^2}\bar q \sigma_{\mu\nu} F^{\mu\nu} (1+\gamma_5)b.
\end{array}
\end{eqnarray}
where $(\bar a b)_{V-A} = \bar a \gamma_\mu (1-\gamma_5) b$, $G^{\mu\nu}$ and
$F^{\mu\nu}$ are the field strengths of the gluon and photon, respectively.
At the hadron level, the decay amplitude can be generically written as
\begin{eqnarray}
A = \langle final\;state|H_{eff}^q|\bar {B}\rangle = V_{ub}V^*_{uq} T(q) + V_{tb}V^*_{tq}P(q)\;,
\end{eqnarray}
where $T(q)$ contains contributions from the $tree$ as well as $penguin$ due to charm and up
quark loop corrections to the matrix elements,
while $P(q)$ contains contributions purely from
one loop $penguin$ contributions. $B$ indicates one of the $B^+$, ${B}^0$ and ${B}^0_s$. $B_i = (B^+, B^0, B^0_s)$ forms an $SU(3)$ triplet.
The flavor $SU(3)$ symmetry transformation properties for operators $O_{1,2}$, $O_{3-6, 11,12}$, and $O_{7-10}$ are: $\bar 3_a + \bar 3_b +6 + \overline {15}$,
$\bar 3$, and $\bar 3_a + \bar 3_b +6 + \overline {15}$, respectively.
We indicate these representations by matrices in $SU(3)$ flavor space by $H(\bar 3)$, $H(6)$ and $H(\overline{15})$.
For $q=d$, the non-zero entries of the matrices $H(i)$ are given by~\cite{he1}
\begin{eqnarray}
H(\bar 3)^2 &=& 1\;,\;\;
H(6)^{12}_1 = H(6)^{23}_3 = 1\;,\;\;H(6)^{21}_1 = H(6)^{32}_3 =
-1\;,\nonumber\\
H(\overline {15} )^{12}_1 &=& H(\overline {15} )^{21}_1 = 3\;,\; H(\overline
{15} )^{22}_2 =
-2\;,\;
H(\overline {15} )^{32}_3 = H(\overline {15} )^{23}_3 = -1\;.
\end{eqnarray}
And for $q = s$, the non-zero entries are
\begin{eqnarray}
H(\bar 3)^3 &=& 1\;,\;\;
H(6)^{13}_1 = H(6)^{32}_2 = 1\;,\;\;H(6)^{31}_1 = H(6)^{23}_2 =
-1\;,\nonumber\\
H(\overline {15} )^{13}_1 &=& H(\overline {15} ) ^{31}_1 = 3\;,\; H(\overline
{15} )^{33}_3 =
-2\;,\;
H(\overline {15} )^{32}_2 = H(\overline {15} )^{23}_2 = -1\;.
\end{eqnarray}
These properties enable one to
write the decay amplitudes for $B \to PPP$ decays in only a few $SU(3)$ invariant amplitudes~\cite{su31}. Here $P$ is one of the mesons in the pseudoscalar octet meson $M=(M_{ij})$ which is given by,
\begin{eqnarray}
M= \left ( \begin{array}{ccc}
{\pi^0\over \sqrt{2}} + {\eta_8\over \sqrt{6}}&\pi^+&K^+\\
\pi^-&-{\pi^0\over \sqrt{2}} + {\eta_8\over \sqrt{6}}& K^0\\
K^-&\bar K^0&-{2\eta_8\over \sqrt{6}}
\end{array} \right ).
\end{eqnarray}
Construction of $B \to PPP$ decay amplitude can be done order by order by using three $M$'s, $B$, and the Hamiltonian $H$, and also derivatives on the mesons to form $SU(3)$. The $SU(3)$ conserving
momentum independent amplitudes can be constructed by the following.
For the $T(q)$ amplitude, we have\cite{xu-li-he1}
\begin{eqnarray}
T(q)&&=a^T(\overline{3})B_{i}H^{i}(\overline{3})M^{j}_{k}M^{k}_{l}M^{l}_{j}+b^T(\overline{3})H^{i}(\overline{3})M^{j}_{i}B_{j}M^{k}_{l}M^{l}_{k} +c^T(\overline{3})H^{i}(\overline{3})M^{l}_{i}M^{j}_{l}M^{k}_{j}B_{k}\nonumber\\
&&+a^T(6)B_{i}H^{ij}_{k}(6)M^{k}_{j}M^{l}_{n}M^{n}_{l}+b^T(6)B_{i}H^{ij}_{k}(6)M^{k}_{l}M^{l}_{n}M^{n}_{j}\nonumber\\
&&+c^T(6)B_{i}H^{jk}_{l}(6)M^{i}_{j}M^{n}_{k}M^{l}_{n}+d^T(6)B_{i}H^{jk}_{l}(6)M^{i}_{n}M^{l}_{j}M^{n}_{k}\nonumber\\
&&+a^T(\overline {15})B_{i}H^{ij}_{k}(\overline{15})M^{k}_{j}M^{l}_{n}M^{n}_{l}+b^T(\overline{15})B_{i}H^{ij}_{k}(\overline{15})M^{k}_{l}M^{l}_{n}M^{n}_{j}\nonumber\\
&&+c^T(\overline{15})B_{i}H^{jk}_{l}(\overline{15})M^{i}_{j}M^{n}_{k}M^{l}_{n}+d^T(\overline{15})B_{i}H^{jk}_{l}(\overline{15})M^{i}_{n}M^{l}_{j}M^{n}_{k}\;.\label{su3s}
\end{eqnarray}
One can write similar amplitude $P(q)$ for the penguin contributions.
The coefficients $a(i)$, $b(i)$, $c(i)$ and $d(i)$ are constants which contain the WCs and information about QCD dynamics. Expanding the above $T(q)$ amplitude, one can extract the decay amplitudes for specific decays in terms of these coefficients.
In the above we have described how to obtain flavor $SU(3)$ amplitudes which are momentum independent.
However, due to the three body decay nature, in general, there are momentum dependence in the decay amplitudes. The momentum dependence can in principle be determined by analysing Dalitz plots for the decays.
The lowest order terms with derivatives lead to two powers of momentum dependence.
One can obtain relevant terms by taking two times of derivatives on each of the terms in Eq.(\ref{su3s}) and then collecting them together. It has been shown\cite{xu-li-he1} that there are six independent ways of taking derivatives for each of the terms listed in eq. (\ref{su3s}). For example after taking derivatives for $B_i H^i(\overline 3) M^j_k M^k_l M^l_j$, we have the following independent terms
\begin{eqnarray}
&&(\partial_\mu B_i) H^i(\overline 3) (\partial^\mu M^j_k) M^k_l M^l_j,\;\;(\partial_\mu B_i) H^i(\overline 3) M^j_k (\partial^\mu M^k_l) M^l_j,\;\;(\partial_\mu B_i) H^i(\overline 3) M^j_k M^k_l (\partial^\mu M^l_j)\;,\nonumber\\
&&B_i H^i(\overline 3) (\partial_\mu M^j_k) (\partial^\mu M^k_l) M^l_j,\;\;B_i H^i(\overline 3) (\partial_\mu M^j_k ) M^k_l (\partial^\mu M^l_j),\;\;B_i H^i(\overline 3) M^j_k (\partial_\mu M^k_l) (\partial^\mu M^l_j)\;.\nonumber\\
\end{eqnarray}
The full list of the possible terms have been obtained in Ref.\cite{xu-li-he1} in the Appendix B. We will not repeat them here.
Using the above $SU(3)$ decay amplitudes, one can find some interesting relations among different decays\cite{xu-li-he1}.
It has been recently pointed out that there are additional relations among the fully-symmetric final states B decay amplitudes ${\cal A}_{FS}$\cite{gronau-full}. Study of these relations can provide further information about flavor $SU(3)$ symmetry in $B$ decays.
The fully-symmetric $B \to PPP$ amplitudes ${\cal A}_{FS}$ is related to the usual decay amplitudes $A(P_1(p_1) P_2(p_2) P_3(p_3))$ for the final mesons $P_{1,2,3}$ carrying momenta $p_{1,2,3}$, for all three final mesons are distinctive, by
\begin{eqnarray}
{\cal A}_{FS}(P_1 &P_2&P_3) = \\
&{1\over \sqrt{3}}& \left ( A(P_1(p_1) P_2(p_2) P_3(p_3)) + A(P_1(p_2) P_2(p_3) P_3(p_1))+A(P_1(p_3) P_2(p_1) P_3(p_2)) \right )\;,\nonumber
\end{eqnarray}
For the cases that two of them or all three of them are identical particles, the identical particle factorial factors should be taken cared. In Ref.\cite{gronau-full}, how the fully-symmetric amplitudes can be determined experimentally has been discussed in detail. We will not repeat the discussions here. We concentrate on how these amplitudes are derived in the framework of flavor SU(3) symmetry and how they are affected by SU(3) breaking effects due to finite quark masses for u, d and s quarks.
To understand that why there are new relations between the fully-symmetric amplitudes for different decay modes, let us consider $B^+\rightarrow K^0\pi^+\pi^0$ and $B^0_d\rightarrow K^+\pi^0\pi^-$ decays as examples.
Expanding eq. (\ref{su3s}), one obtains
\begin{eqnarray}
T(B^{+} \to K^{0}\pi^{+}\pi^{0})&&={\sqrt{2}}\left ( c(6)+d(6)+{2}c(\overline{15})+{2}d(\overline{15})\right )\;,
\end{eqnarray}
and
\begin{eqnarray}
T(B^{0} \to K^{+}\pi^{-}\pi^{0}) = T(B^{+} \to K^{0}\pi^{+}\pi^{0})\;.
\end{eqnarray}
From which we get $T_{FS}(B^{+} \to K^{0}\pi^{+}\pi^{0})=T_{FS}(B^{0} \to K^{+}\pi^{-}\pi^{0})$.
As the decay amplitudes may have momentum dependence, we should also check if the equality of the above two amplitudes are equal when taking into account of momentum dependence in the amplitudes. Expanding terms in Appendix B of Ref.\cite{xu-li-he1}, we find
\begin{eqnarray}
&&T^p(B^{+} \to K^{0}\pi^{+}\pi^{0})=\nonumber\\
&&\qquad\qquad\alpha_1p_{B}\cdot p_{1}+\alpha_2p_{B}\cdot p_{2}+\alpha_3p_{B}\cdot p_{3}+\alpha_4p_{1}\cdot p_{2}+\alpha_5p_{1}\cdot p_{3}+\alpha_6p_{2}\cdot p_{3}\;,\nonumber\\
&&T^p(B^{0}_d \to K^{+}\pi^{0}\pi^{-})=\nonumber\\
&&\qquad\qquad\beta_1p_{B}\cdot p_{1}+\beta_2p_{B}\cdot p_{2}+\beta_3p_{B}\cdot p_{3}+\beta_4p_{1}\cdot p_{2}+\beta_5p_{1}\cdot p_{3}+\beta_6p_{2}\cdot p_{3}\;.
\end{eqnarray}
The coefficients $\alpha_i$ and $\beta_i$ are given by,
\begin{eqnarray}
\alpha_1&=&\sqrt{2}\left ( c^{\prime}(6)_{2}+2c^{\prime}(\overline{15})_{2}+d^{\prime}(6)_{3}+2d^{\prime}(\overline{15})_{3}\right )\;,\nonumber\\
\alpha_2&=&\frac{1}{\sqrt{2}}\left (-b^{\prime}(6)_{1}+b^{\prime}(6)_{2}-3 b^{\prime}(\overline{15})_{1}+3b^{\prime}(\overline{15})_{2}+c^{\prime}(\overline{3})_{2}-c^{\prime}(\overline{3})_{3} +c^{\prime}(6)_{1}\right .\nonumber\\&+&\left . c^{\prime}(6)_{3}+c^{\prime}(\overline{15})_{1}+3c^{\prime}(\overline{15})_{3}+d^{\prime}(6)_{1}+d^{\prime}(6)_{3} +5d^{\prime}(\overline{15})_{1}-d^{\prime}(\overline{15})_{3} \right )\;,\nonumber\\
\alpha_3&=&\frac{1}{\sqrt{2}}\left (b^{\prime}(6)_{1}-b^{\prime}(6)_{2}+3b^{\prime}(\overline{15})_{1}-3b^{\prime}(\overline{15})_{2}-c^{\prime}(\overline{3})_{2}+c^{\prime}(\overline{3})_{3}+c^{\prime}(6)_{1}+c^{\prime}(6)_{3} \right .\\&+&\left . 3c^{\prime}(\overline{15})_{1}+c^{\prime}(\overline{15})_{3}+d^{\prime}(6)_{1}+2d^{\prime}(6)_{2}-d^{\prime}(6)_{3}-d^{\prime}(\overline{15})_{1}+4 d^{\prime}(\overline{15})_{2}+d^{\prime}(\overline{15})_{3}\right )\;,\nonumber\\
\alpha_4&=&\frac{1}{\sqrt{2}}\left (- b^{\prime\prime}(6)_{2}+b^{\prime\prime}(6)_{3}-3b^{\prime\prime}(\overline{15})_{2}+3b^{\prime\prime}(\overline{15})_{3}+c^{\prime\prime}(\overline{3})_{1}-c^{\prime\prime}(\overline{3})_{2}+c^{\prime\prime}(6)_{1}+c^{\prime\prime}(6)_{3}\right .\nonumber\\
&+&\left . c^{\prime\prime}(\overline{15})_{1}+3c^{\prime\prime}(\overline{15})_{2}-d^{\prime\prime}(6)_{1}+2d^{\prime\prime}(6)_{2}+d^{\prime\prime}(6)_{3}+d^{\prime\prime}(\overline{15})_{1}-d^{\prime\prime}(\overline{15})_{2}+4d^{\prime\prime}(\overline{15})_{3}\right )\;,\nonumber\\
\alpha_5&=&\frac{1}{\sqrt{2}}\left (b^{\prime\prime}(6)_{2}-b^{\prime\prime}(6)_{3}+3b^{\prime\prime}(\overline{15})_{2}-3b^{\prime\prime}(\overline{15})_{3}-c^{\prime\prime}(\overline{3})_{1}+c^{\prime\prime}(\overline{3})_{2}+c^{\prime\prime}(6)_{1}\right .\nonumber\\
&+&\left . c^{\prime\prime}(6)_{3}+3c^{\prime\prime}(\overline{15})_{1}+c^{\prime\prime}(\overline{15})_{2}+d^{\prime\prime}(6)_{1}+d^{\prime\prime}(6)_{3}-d^{\prime\prime}(\overline{15})_{1}+d^{\prime\prime}(\overline{15})_{2}\right )\;,\nonumber\\
\alpha_6&=&\sqrt{2}\left ( c^{\prime\prime}(6)_{2}+2 c^{\prime\prime}(\overline{15})_{3}+d^{\prime\prime}(6)_{1}+2d^{\prime\prime}(\overline{15})_{1}\right )\;. \nonumber
\end{eqnarray}
and
\begin{eqnarray}
\beta_1&=&\sqrt{2}\left (c^{\prime}(6)_{2}+2c^{\prime}(\overline{15})_{2}+d^{\prime}(6)_{3}+2d^{\prime}(\overline{15})_{3}\right )\;,\nonumber\\
\beta_2&=&\frac{1}{\sqrt{2}}\left (b^{\prime}(6)_{1}-b^{\prime}(6)_{2}+b^{\prime}(\overline{15})_{1}-b^{\prime}(\overline{15})_{2}+c^{\prime}(\overline{3})_{2}-c^{\prime}(\overline{3})_{3}+c^{\prime}(6)_{1}+c^{\prime}(6)_{3}\right .\nonumber\\
&+&\left .c^{\prime}(\overline{15})_{1}+ 3c^{\prime}(\overline{15})_{3}+d^{\prime}(6)_{1}+2d^{\prime}(6)_{2}-d^{\prime}(6)_{3}-3d^{\prime}(\overline{15})_{1}+4d^{\prime}(\overline{15})_{2}+3d^{\prime}(\overline{15})_{3}\right )\;,\nonumber\\
\beta_3&=&\frac{1}{\sqrt{2}}\left ( -b^{\prime}(6)_{1}+b^{\prime}(6)_{2}-b^{\prime}(\overline{15})_{1}+b^{\prime}(\overline{15})_{2}-c^{\prime}(\overline{3})_{2}+c^{\prime}(\overline{3})_{3}+c^{\prime}(6)_{1}\right .\\
&+&\left . c^{\prime}(6)_{3}+3c^{\prime}(\overline{15})_{1}+c^{\prime}(\overline{15})_{3}+d^{\prime}(6)_{1}+d^{\prime}(6)_{3}+7 d^{\prime}(\overline{15})_{1}-3d^{\prime}(\overline{15})_{3}\right )\;,\nonumber\\
\beta_4&=&\frac{1}{\sqrt{2}}\left (b^{\prime\prime}(6)_{2}-b^{\prime\prime}(6)_{3}+b^{\prime\prime}(\overline{15})_{2}-b^{\prime\prime}(\overline{15})_{3}+c^{\prime\prime}(\overline{3})_{1}-c^{\prime\prime}(\overline{3})_{2}+c^{\prime\prime}(6)_{1}\right .\nonumber\\
&+& \left . c^{\prime\prime}(6)_{3}+c^{\prime\prime}(\overline{15})_{1}+3c^{\prime\prime}(\overline{15})_{2}+d^{\prime\prime}(6)_{1}+d^{\prime\prime}(6)_{3}-3d^{\prime\prime}(\overline{15})_{1}+7d^{\prime\prime}(\overline{15})_{2}\right )\;,\nonumber\\
\beta_5&=&\frac{1}{\sqrt{2}}\left ( - b^{\prime\prime}(6)_{2}+b^{\prime\prime}(6)_{3}-b^{\prime\prime}(\overline{15})_{2}+b^{\prime\prime}(\overline{15})_{3}-c^{\prime\prime}(\overline{3})_{1}+c^{\prime\prime}(\overline{3})_{2}+c^{\prime\prime}(6)_{1}+ c^{\prime\prime}(6)_{3}\right .\nonumber\\
&+&\left . 3c^{\prime\prime}(\overline{15})_{1}+c^{\prime\prime}(\overline{15})_{2}-d^{\prime\prime}(6)_{1}+2 d^{\prime\prime}(6)_{2}+d^{\prime\prime}(6)_{3}+3d^{\prime\prime}(\overline{15})_{1}-3d^{\prime\prime}(\overline{15})_{2}+4d^{\prime\prime}(\overline{15})_{3}\right )\;,\nonumber\\
\beta_6&=&\sqrt{2}\left (c^{\prime\prime}(6)_{2}+2c^{\prime\prime}(\overline{15})_{3}+d^{\prime\prime}(6)_{1}+2d^{\prime\prime}(\overline{15})_{1}\right )\;.\nonumber
\end{eqnarray}
One can see from the above that $T^p(B^{+} \to K^{0}\pi^{+}\pi^{0})$ is no longer equal to $T^p(B^{0}_d \to K^{+}\pi^{0}\pi^{-})$. However, one can readily see from the above equations, that
\begin{eqnarray}
\alpha_1+\alpha_2+\alpha_3=\beta_1+\beta_2+\beta_3,~~\alpha_4+\alpha_5+\alpha_6=\beta_4+\beta_5+\beta_6\;.
\end{eqnarray}
This fact makes the fully-symmetric amplitudes to satisfy
\begin{eqnarray}
T^p(B^{+} \to K^{0}\pi^{+}\pi^{0})_{FS} = T^p(B^{0}_d \to K^{+}\pi^{0}\pi^{-})_{FS}\;.
\end{eqnarray}
Similarly, the penguin amplitudes $P$ and $P^p$ have the same properties discussed above for the tree amplitudes, $T$ and $T^p$.
The total fully-symmetric amplitudes ${\cal A}_{FS} = V_{ub} V^*_{uq} (T_{FS} + T^p_{FS}) + V_{tb}V_{tq}^* (P_{FS} + P^p_{FS})$ then have the relation
\begin{eqnarray}
{\cal A}(B^{+} \to K^{0}\pi^{+}\pi^{0})_{FS} = {\cal A}(B^{0}_d \to K^{+}\pi^{0}\pi^{-})_{FS}\;.
\end{eqnarray}
Enlarging the amplitudes to fully-symmetric ones, indeed produce more relations.
Expanding eq. (\ref{su3s}) and equations in Appendix B of Ref. \cite{xu-li-he1}, we obtain the following relations confirming those obtained in Ref.\cite{gronau-full}.
For $\bar b \to \bar s$ induced $B\to PPP$ amplitudes, we have
\\
$ 1.~B\to K\pi\pi$
\\
$S1.1 = {\cal A}(B^+\to K^0\pi^+\pi^0)_{\rm FS} - {\cal A}(B^0 \to K^+\pi^0\pi^-)_{\rm FS} =0$,
$S1.2 = \sqrt{2}{\cal A}(B^+\to K^0\pi^+\pi^0)_{\rm FS} - {\cal A}(B^0 \to K^0\pi^+\pi^-)_{\rm FS} + 2{\cal A}(B^0\to K^0\pi^0\pi^0)_{\rm FS} =0$,
$S1.3 = \sqrt{2}{\cal A}(B^0\to K^+\pi^0\pi^-)_{\rm FS} + {\cal A}(B^+\to K^+\pi^+\pi^-)_{\rm FS} - 2{\cal A}(B^+\to K^+\pi^0\pi^0)_{\rm FS} =0$.
\\
$2.~B\to KK{\bar K}$
\\
$S2.1 = - {\cal A}(B^+\to K^+K^+K^-)_{\rm FS} + {\cal A}(B^+\to K^+K^0{\bar K}^0)_{\rm FS} $
$\;\;\;\;\;\;\;\;+ {\cal A}(B^0 \to K^0K^+K^-)_{\rm FS} - {\cal A}(B^0 \to K^0 K^0{\bar K}^0)_{\rm FS} =0$.
\\
$3.~B^0_{s}\to \pi K{\bar K}$
\\
$S3.1 = \sqrt{2}{\cal A}(B^0_s\to \pi^0 K^+K^-)_{\rm FS} - \sqrt{2}{\cal A}(B^0_s\to \pi^0 K^0{\bar K}^0)_{\rm FS} $
$\;\;\;\;\;\;\;\;-~ {\cal A}(B^0_s\to \pi^- K^+ {\bar K}^0)_{\rm FS} - {\cal A}(B^0_s\to \pi^+ K^-K^0)_{\rm FS} =0 $.
\\
$4.~B^0_{s}\to \pi\pi\pi$
\\
$S4.1 = 2{\cal A}(B^0_s\to \pi^0\pi^0\pi^0)_{\rm FS} - {\cal A}(B^0_s \to \pi^0\pi^+\pi^-)_{\rm FS}=0 $.
\\
For $\bar b \to \bar d$ induced $B\to PPP$ amplitudes, we have
\\
$1.~B\to \pi K{\bar K}$
\\
$D1.1 = -\sqrt{2} {\cal A}(B^0 \to \pi^0 K^+K^-)_{\rm FS}
+ {\cal A}(B^0\to \pi^+ K^0K^-)_{\rm FS}
- {\cal A}(B^+ \to \pi^+ K^+K^-)_{\rm FS}$
$\;\;\;\;\;\;\;\;+~\sqrt{2} {\cal A}(B^0 \to \pi^0 K^0{\bar K}^0)_{\rm FS}
+ {\cal A}(B^0 \to \pi^- K^+{\bar K}^0)_{\rm FS} $
$\;\;\;\;\;\;\;\; +~{\cal A}(B^+ \to \pi^+ K^0{\bar K}^0)_{\rm FS}
- \sqrt{2} {\cal A}(B^+ \to \pi^0 K^+{\bar K}^0)_{\rm FS} = 0 $.
\\
$2.~B \to \pi\pi\pi$
\\
$D2.1 = 2 {\cal A}(B^0\to \pi^0\pi^0\pi^0)_{\rm FS} - {\cal A}(B^0\to \pi^+\pi^0\pi^-)_{\rm FS}=0 $,
$D2.2 = 2 {\cal A}(B^+\to \pi^+\pi^0\pi^0)_{\rm FS} - {\cal A}(B^+\to \pi^-\pi^+\pi^+)_{\rm FS}=0 $.
\\
$3.~B^0_{s}\to K\pi\pi$
\\
$D3.1 = -2{\cal A}(B^0_s\to {\bar K}^0\pi^0\pi^0)_{\rm FS} + {\cal A}(B^0_s\to{\bar K}^0\pi^+\pi^-)_{\rm FS} - \sqrt{2}{\cal A}(B^0_s\to K^-\pi^+\pi^0)_{\rm FS} = 0 $.
\\
In the above, we have considered some relations among decay processes with the same $\Delta S$. There are also some other relations among tree and penguin amplitudes but with different $\Delta S$. Some of them will be discussed later in the conclusions.
Note that we have different normalizations than those used in Ref.\cite{gronau-full} for some of the final meson states and also identical particle combinatorial factors. One can easily obtain relations in the form in Ref. \cite{gronau-full} by
multiplying a ``-1'' to the amplitudes when $\pi^{-},K^{-},\pi^{0}$ appear each time as one of the final states, and a factor $1/\sqrt{2}$ and $1/\sqrt{6}$ in our formulation for the corresponding amplitudes, respectively, when the decays involve two and three identical particles.
\section{$SU(3)$ and Isospin breaking due to quark mass differences}
The main source for flavor $SU(3)$ symmetry breaking effects comes from difference in
masses of $u$, $d$ and $s$ quarks. Under $SU(3)$, the mass matrix
can be viewed as combinations of representations from $3\times \bar
3$, to matching the ($u$, $d$, $s$) transformation property as a
fundamental representation, which contains an 1 and an 8 irreducible
representations. The diagonalized mass matrix can be expressed as a
linear combination of the identity matrix $I$, and the Gell-Mann
matrices $\lambda_3$ and $\lambda_8$. We have
\begin{eqnarray}
\left ( \begin{array}{ccc}
m_u & 0 & 0\\
0 & m_d& 0 \\
0 & 0 & m_s
\end{array}
\right ) = {1\over 3}(m_u+m_d+m_s) I + {1\over 2}(m_u - m_d) X + {1\over 6}(m_u+m_d - 2 m_s) W\;,
\end{eqnarray}
with $X$ and $W$ given by
\begin{eqnarray}
X=\left ( \begin{array}{ccc}
1 & 0 & 0\\
0 & -1& 0 \\
0 & 0 & 0
\end{array}
\right )\;,\;\;\;\;W=\left ( \begin{array}{ccc}
1 & 0 & 0\\
0 & 1& 0 \\
0 & 0 & -2
\end{array}
\right )\;.
\end{eqnarray}
Compared with $s$-quark mass
$m_s$, the $u$ and $d$ quark masses $m_{u,d}$ are much smaller, $SU(3)$ breaking effects due to a non-zero $m_s$
dominates the $SU(3)$ breaking effects. When up and down quark mass difference is neglected, the residual symmetry of $SU(3)$ becomes the isospin symmetry. In that case when studying $SU(3)$ breaking effects,
the term proportional to
$X$ can be dropped. The identity $I$ part contributes to the $B$ decay amplitudes in a similar way as that given in eq. (\ref{su3s}) which can be absorbed into the coefficients $a(i)$ to $d(i)$. Only $W$ piece will contribute to the $SU(3)$ breaking effects. We will first discuss this case to first order in $W$, and then also study the isospin breaking effects by including the first order term proportional to $X$.
\subsection{SU(3) breaking due a non-zero $m_s$}
To construct relevant decay amplitudes for $B\to PPP$ decays, one first breaks the contraction of indices at any joint in eq. (\ref{su3s}), and inserts a W in between, and then contracts all indices appropriately. For example corresponding to the first term in eq. (\ref{su3s}), there are two ways to insert $W$,
\begin{eqnarray}
B_{i}H^{a}(\overline{3})W^{i}_{a}M^{j}_{k}M^{k}_{l}M^{l}_{j}\;,\;\;\;\;B_{i}H^{i}(\overline{3})M^{j}_{k}M^{k}_{l}M^{a}_{j}W^{l}_{a}\;.
\end{eqnarray}
The full list of possible independent terms are given in Appendix A of Ref.\cite{xu-li-he1}.
Extracting the $SU(3)$ breaking terms for $B^{+} \to K^{0}\pi^{+}\pi^{0}$ and $B^{0} \to K^{+}\pi^{0}\pi^{-}$ decays, we have the corrections for the decay amplitudes, $\Delta T$, as
\begin{eqnarray}
\triangle T(B^{+} &\to& K^{0}\pi^{+}\pi^{0}
) =\sqrt{2}\left ( c^T_{1}(6)+\sqrt{2}c^T_{2}(6)-2c^T_{3}(6)+\sqrt{2}c^T_{4}(6)+c^T_{5}(6)+c^T_{1}(\overline{15})\right .\nonumber\\&+&\left . c^T_{2}(\overline{15})-2c^T_{3}(\overline{15})+ c^T_{4}(\overline{15})+c^T_{5}(\overline{15})+d^T_{1}(6)+d^T_{2}(6)-2d^T_{3}(6)+d^T_{4}(6)\right .\nonumber\\
&+&\left . d^T_{5}(6)+d^T_{1}(\overline{15})+d^T_{2}(\overline{15})-2d^T_{3}(\overline{15})+d^T_{4}(\overline{15})+d^T_{5}(\overline{15})\right )\;,
\end{eqnarray}
and
\begin{eqnarray}
\triangle T(B^{0} \to K^{+}\pi^{0}\pi^{-})=\triangle T(B^{+} \to K^{0}\pi^{+}\pi^{0})\;,
\end{eqnarray}
which leads to the equality of the fully-symmetric amplitudes for these two decays.
Therefore,
\begin{eqnarray}
S1.1 = {\cal A}(B^{0} \to K^{+}\pi^{0}\pi^{-})_{FS}-{\cal A}(B^{+} \to K^{0}\pi^{+}\pi^{0})_{FS}=0\;,
\end{eqnarray}
still holds.
Note that even $SU(3)$ breaking effects affect each of the decay amplitudes, the relation of the fully-symmetric amplitudes of these two decays are not affected {by the first order $SU(3)$ breaking effects.}
Expanding terms in Appendix A of Ref.\cite{xu-li-he1}, one can study relations discussed above. We find that all the relations among the fully-symmetric amplitudes still hold, that is
\begin{eqnarray}
&&S1.1 = 0\;,\;\;S2.1 = 0\;,\;\;S1.3=0\;\,\;\;S2.1=0\;,\;\;S3.1=0\;,\;\;S4.1=0\;,\nonumber\\
&&D1.1=0\;,\;\;D2.1=0\;,\;\;D2.2=0\;,\;\;D3.1=0\;,\label{resu3}
\end{eqnarray}
are still true even if one include $SU(3)$ breaking effects due a non-zero strange quark mass.
This actually is not a surprise because the relations discussed can be obtained by isospin symmetry considerations.
Experimental verification of these relations may provide important tests for the validity of flavor $SU(3)$ for $B$ decays.
\subsection{Isospin breaking due to up and down quark mass difference}
It would be interesting to investigate what happens when mass difference between up and down quark, which breaks isospin symmetry, is also included. We now discuss these isospin breaking effects for the relations discussed before.
One can obtain the corrections by replacing $W$ by $X$ in Appendix A of Ref.\cite{xu-li-he1}. We indicate the coefficients in a similar way as that $SU(3)$ breaking effects due to a non-zero $m_s$, but with a superscript $I$ to indicate the effects of isospin breaking, for example for tree operator corrections by $a^{T^I}_i$, $b^{T^I}_i$, $c^{T^I}_i$, and $d^{T^I}_i$. The correction to the decay amplitude will also be indicated by a superscript $I$, $\Delta T^I$.
Expanding all terms, we obtain the corrections due to isospin breaking effects for all the decay amplitudes discussed previously. We find that except that the relation $S4.1$ still holds, all other relations for the $B\to PPP$ decay amplitudes induced by $\bar b \to \bar s$ and $\bar b\to \bar d$ interactions discussed earlier are broken.
In fact each of the decay modes relevant in S4.1 is affected by isospin breaking effects,
\begin{eqnarray}
&&\Delta T^{I}(B^0_{s}\rightarrow \pi^0\pi^0\pi^0)=\nonumber\\
&&\qquad\quad\frac{\sqrt{2}}{2}(a^{T^{I}}_{2}(\overline{3})+2a^{T^{I}}_{2}(\overline{15})+2a^{T^{I}}_{3}(\overline{15})+b^{T^{I}}_{2}(\overline{15})+b^{T^{I}}_{3}(\overline{15})+b^{T^{I}}_{4}(\overline{15})+b^{T^{I}}_{5}(\overline{15}))\;,\nonumber\\
&&\Delta T^{I}(B^0_{s}\rightarrow \pi^+\pi^-\pi^0)=\\
&&\qquad\quad\sqrt{2}(a^{T^{I}}_{2}(\overline{3})+2a^{T^{I}}_{2}(\overline{15})+2a^{T^{I}}_{3}(\overline{15})+b^{T^{I}}_{2}(\overline{15})+b^{T^{I}}_{3}(\overline{15})+b^{T^{I}}_{4}(\overline{15})+b^{T^{I}}_{5}(\overline{15}))\;,\nonumber
\end{eqnarray}
but they are affected in such a way that the equality of the amplitudes is not affected. That is, we still have:
\begin{eqnarray}
S4.1=2 {\cal A}(B^0_s\to \pi^0\pi^0\pi^0)_{\rm FS} - {\cal A}(B^0_s \to \pi^0\pi^+\pi^-)_{\rm FS}=0.\label{reis1}
\end{eqnarray}
This makes this relation special because that this relation is
not affected by first order $SU(3)$ breaking effects due to a non-zero strange quark and isospin breaking due to up and down quark mass difference. It should hold to a high precision. Experimental test of this relation can provide important information about $B \to PPP$.
We also found some other interesting relations even isospin violating effects are included, namely the corrections for some of the relations discussed above are related to others.
For $b\to s$ interaction induced decay modes,
we have an additional relation which relate $S1.2$ and $S1.3$ because the isospin breaking effects satisfy
\begin{eqnarray}
&&[\sqrt{2}\Delta T^{I}(B^+\rightarrow K^0\pi^+\pi^0)-\Delta T^{I}(B^0\rightarrow K^0\pi^+\pi^-)+2\Delta T^{I}(B^0\rightarrow K^0\pi^0\pi^0)]\nonumber\\
&&=-[\sqrt{2}\Delta T^{I}(B^0\rightarrow K^+\pi^-\pi^0)+\Delta T^{I}(B^+\rightarrow K^+\pi^+\pi^-)-2\Delta T^{I}(B^+\rightarrow K^+\pi^0\pi^0)]\;.\nonumber\\
\end{eqnarray}
Although the right hand sides of $S1.2$ and $S1.3$ are not zero anymore, the above relation leads to,
\begin{eqnarray}
S1.2 =-S1.3\neq 0\;.\label{reis2}
\end{eqnarray}
For $b\to d$ interactions induced decay modes, we have
\begin{eqnarray}
&&[2\Delta T^{I}(B^+\rightarrow \pi^0\pi^0\pi^+)-\Delta T^{I}(B^+\rightarrow \pi^-\pi^+\pi^+)]\nonumber\\
&&=-[-2\Delta T^{I}(B^0_{s}\rightarrow \bar{K^0}\pi^0\pi^0)+\Delta T^{I}(B^0_{s}\rightarrow \bar{K^0}\pi^+\pi^-)-\sqrt{2}\Delta T^{I}(B^0_{s}\rightarrow K^-\pi^+\pi^0)]\;,\nonumber\\
&&\sqrt{2}[2\Delta T^{I}(B^0\rightarrow \pi^0\pi^0\pi^0)-\Delta T^{I}(B^0\rightarrow \pi^+\pi^-\pi^0)]\\
&&=-[2\Delta T^{I}(B^+\rightarrow \pi^0\pi^0\pi^+)-\Delta T^{I}(B^+\rightarrow \pi^-\pi^+\pi^+)]\;.\nonumber
\end{eqnarray}
Due to isospin breaking effects, the right hand sides of $D2.1$, $D2.2$ and $D3.1$ are non-zero. However, the above relations imply
\begin{eqnarray}
\sqrt{2} D2.1 = -D2.2\neq 0\;,\;\;\;\;D2.2= -D3.1 \neq 0\;.\label{reis3}
\end{eqnarray}
We would like to emphasize that since the above relations hold even when first order isospin effects have been taken into account, they can provide useful information about $B$ decays in the SM in a way independent of flavor $SU(3)$ and isospin breaking effects.
\subsection{Momentum dependent $SU(3)$ and isospin breaking amplitudes}
There are also momentum dependent amplitudes at the same order to the $SU(3)$ and isospin breaking effects discussed in the previous subsections. We find that all the relations eq. (\ref{resu3}), and eqs. (\ref{reis1}), (\ref{reis2}) and (\ref{reis3}) still hold when $SU(3)$ breaking effects due to a non-zero $m_s$, and isospin breaking effects due to $m_u$ and $m_d$ mass difference discussed earlier, respectively.
The analysis is similar to the case for $SU(3)$ conserving momentum amplitudes. We will not give details here, but just outline how the analysis can be carried out. The leading ones are constructed by taking two powers of derivatives on each term of the $SU(3)$ breaking amplitudes which have been shown in the Appendix A of Ref.\cite{xu-li-he1}. For example for the term, $a^T_{1}(\overline{3}) B_{i}H^{a}(\overline{3})W^{i}_{a}M^{j}_{k}M^{k}_{l}M^{l}_{j}$, we obtain the following six independent terms with two derivatives:
\begin{eqnarray}
&&a^{\prime}_{1}(\overline{3})_{1}(\partial_\mu B_{i}) H^{a}(\overline{3})W^{i}_{a} (\partial_\mu M^{j}_{k}) M^{k}_{l}M^{l}_{j},\;\;
a^{\prime}_{1}(\overline{3})_{2}(\partial_\mu B_{i}) H^{a}(\overline{3})W^{i}_{a}M^{j}_{k} (\partial_\mu M^{k}_{l}) M^{l}_{j},\;\nonumber\\
&&a^{\prime}_{1}(\overline{3})_{3}(\partial_\mu B_{i} ) H^{a}(\overline{3})W^{i}_{a}M^{j}_{k}M^{k}_{l} (\partial_\mu M^{l}_{j}), \;\;
a^{\prime\prime}_{1}(\overline{3})_{1} B_{i}H^{a}(\overline{3}) W^{i}_{a}(\partial_\mu M^{j}_{k} ) (\partial_\mu M^{k}_{l}) M^{l}_{j},\;\nonumber\\
&& a^{\prime\prime}_{1}(\overline{3})_{2}B_{i}H^{a}(\overline{3}) W^{i}_{a}(\partial_\mu M^{j}_{k} ) M^{k}_{l} (\partial_\mu M^{l}_{j}),\;\;
a^{\prime\prime}_{1}(\overline{3})_{3} B_{i}H^{a}(\overline{3}) W^{i}_{a} M^{j}_{k} (\partial_\mu M^{k}_{l})(\partial_\mu M^{l}_{j}).
\end{eqnarray}
Here, $a^{\prime}_{1}(\overline{3})_{1}, a^{\prime}_{1}(\overline{3})_{2}, a^{\prime}_{1}(\overline{3})_{3}, a^{\prime\prime}_{1}(\overline{3})_{1}, a^{\prime\prime}_{1}(\overline{3})_{2}, a^{\prime\prime}_{1}(\overline{3})_{3}$ are constants. We then extend similar definition of constants for other $SU(3)$ breaking terms in the Appendix A and B of Ref.\cite{xu-li-he1}. There are both tree and penguin amplitudes which can be further labeled by superscripts $T$ and $P$. We will omit writing them out with the understanding that what described below work for both tree and penguin amplitudes.
One obtains the relevant terms for isospin breaking effects by replacing $W$ by $X$.
Expanding all possible terms, one obtains the amplitudes. Taking the amplitudes in $S1.1$ for illustration, one finds that the momentum dependent amplitudes can be written as
\begin{eqnarray}
&&\Delta T^p(B^{+} \to K^{0}\pi^{+}\pi^{0})=\nonumber\\
&&\qquad\qquad \Delta\alpha_1p_{B}\cdot p_{1}+\Delta\alpha_2p_{B}\cdot p_{2}+\Delta\alpha_3p_{B}\cdot p_{3}+\Delta\alpha_4p_{1}\cdot p_{2}+\Delta\alpha_5p_{1}\cdot p_{3}+\Delta\alpha_6p_{2}\cdot p_{3}\;,\nonumber\\
&& \Delta T^p(B^{0}_d \to K^{+}\pi^{0}\pi^{-})=\\
&&\qquad\qquad\Delta\beta_1p_{B}\cdot p_{1}+\Delta\beta_2p_{B}\cdot p_{2}+\Delta\beta_3p_{B}\cdot p_{3}+\Delta\beta_4p_{1}\cdot p_{2}+\Delta\beta_5p_{1}\cdot p_{3}+\Delta\beta_6p_{2}\cdot p_{3}\;.\nonumber
\end{eqnarray}
The coefficients $\Delta\alpha_{i},\Delta\beta_{i}$ are collections of coefficients from all possible terms.
Our detailed calculations show that
\begin{eqnarray}
&&\Delta\alpha_1+\Delta\alpha_2+\Delta\alpha_3=\Delta\beta_1+\Delta\beta_2+\Delta\beta_3\nonumber\\
&&=\sqrt{2}[2c^{\prime}_{1}(\overline {15})_{1}+2c^{\prime}_{1}(\overline {15})_{2}+2c^{\prime}_{1}(\overline {15})_{3}+
2c^{\prime}_{2}(\overline {15})_{1}+2c^{\prime}_{2}(\overline {15})_{2}+2c^{\prime}_{2}(\overline {15})_{3}-
4c^{\prime}_{3}(\overline {15})_{1}\nonumber\\&&-4c^{\prime}_{3}(\overline {15})_{2}
-4c^{\prime}_{3}(\overline {15})_{3}+2c^{\prime}_{4}(\overline {15})_{1}+2c^{\prime}_{4}(\overline {15})_{2}+2c^{\prime}_{4}(\overline {15})_{3}+2c^{\prime}_{5}(\overline {15})_{1}+2c^{\prime}_{5}(\overline {15})_{2}+2c^{\prime}_{5}(\overline {15})_{3}\nonumber\\&&+c^{\prime}_{1}(6)_{1}+c^{\prime}_{1}(6)_{2}
+c^{\prime}_{1}(6)_{3}+c^{\prime}_{2}(6)_{1}+c^{\prime}_{2}(6)_{2}+c^{\prime}_{2}(6)_{3}-2c^{\prime}_{3}(6)_{1}-2c^{\prime}_{3}(6)_{2}
-2c^{\prime}_{3}(6)_{3}+c^{\prime}_{4}(6)_{1}\nonumber\\&&+c^{\prime}_{4}(6)_{2}+c^{\prime}_{4}(6)_{3}
+c^{\prime}_{5}(6)_{1}+c^{\prime}_{5}(6)_{2}+c^{\prime}_{5}(6)_{3}+2d^{\prime}_{1}(\overline {15})_{1}+2d^{\prime}_{1}(\overline {15})_{2}+2d^{\prime}_{1}(\overline {15})_{3}+2d^{\prime}_{2}(\overline {15})_{1}\nonumber\\&&+2d^{\prime}_{2}(\overline {15})_{2}+
2d^{\prime}_{2}(\overline {15})_{3}
-4d^{\prime}_{3}(\overline {15})_{1}-4d^{\prime}_{3}(\overline {15})_{2}-
4d^{\prime}_{3}(\overline {15})_{3}+2d^{\prime}_{4}(\overline {15})_{1}+2d^{\prime}_{4}(\overline {15})_{2}+
2d^{\prime}_{4}(\overline {15})_{3}\nonumber\\&&+2d^{\prime}_{5}(\overline {15})_{1}+2d^{\prime}_{5}(\overline {15})_{2}
+2d^{\prime}_{5}(\overline {15})_{3}
+d^{\prime}_{1}(6)_{1}+d^{\prime}_{1}(6)_{2}+d^{\prime}_{1}(6)_{3}+
+d^{\prime}_{2}(6)_{1}+d^{\prime}_{2}(6)_{2}+d^{\prime}_{2}(6)_{3}\nonumber\\&&-2d^{\prime}_{3}(6)_{1}-2d^{\prime}_{3}(6)_{2}-2d^{\prime}_{3}(6)_{3}
+d^{\prime}_{4}(6)_{1}+d^{\prime}_{4}(6)_{2}+d^{\prime}_{4}(6)_{3}+d^{\prime}_{5}(6)_{1}+d^{\prime}_{5}(6)_{2}+d^{\prime}_{5}(6)_{3}]\nonumber\\
&&\Delta\alpha_4+\Delta\alpha_5+\Delta\alpha_6=\Delta\beta_4+\Delta\beta_5+\Delta\beta_6\nonumber\\
&&=\sqrt{2}[2c^{\prime\prime}_{1}(\overline {15})_{1}+2c^{\prime\prime}_{1}(\overline {15})_{2}+2c^{\prime\prime}_{1}(\overline {15})_{3}+
2c^{\prime\prime}_{2}(\overline {15})_{1}+2c^{\prime\prime}_{2}(\overline {15})_{2}+2c^{\prime\prime}_{2}(\overline {15})_{3}-
4c^{\prime\prime}_{3}(\overline {15})_{1}\nonumber\\&&-4c^{\prime\prime}_{3}(\overline {15})_{2}
-4c^{\prime\prime}_{3}(\overline {15})_{3}+2c^{\prime\prime}_{4}(\overline {15})_{1}+2c^{\prime\prime}_{4}(\overline {15})_{2}+2c^{\prime\prime}_{4}(\overline {15})_{3}+2c^{\prime\prime}_{5}(\overline {15})_{1}+2c^{\prime\prime}_{5}(\overline {15})_{2}+2c^{\prime\prime}_{5}(\overline {15})_{3}\nonumber\\&&+c^{\prime\prime}_{1}(6)_{1}+c^{\prime\prime}_{1}(6)_{2}
+c^{\prime\prime}_{1}(6)_{3}+c^{\prime\prime}_{2}(6)_{1}+c^{\prime\prime}_{2}(6)_{2}+c^{\prime\prime}_{2}(6)_{3}-2c^{\prime\prime}_{3}(6)_{1}-2c^{\prime\prime}_{3}(6)_{2}
-2c^{\prime\prime}_{3}(6)_{3}+c^{\prime\prime}_{4}(6)_{1}\nonumber\\&&+c^{\prime\prime}_{4}(6)_{2}+c^{\prime\prime}_{4}(6)_{3}
+c^{\prime\prime}_{5}(6)_{1}+c^{\prime\prime}_{5}(6)_{2}+c^{\prime\prime}_{5}(6)_{3}+2d^{\prime\prime}_{1}(\overline {15})_{1}+2d^{\prime\prime}_{1}(\overline {15})_{2}+2d^{\prime\prime}_{1}(\overline {15})_{3}+2d^{\prime\prime}_{2}(\overline {15})_{1}\nonumber\\&&+2d^{\prime\prime}_{2}(\overline {15})_{2}+
2d^{\prime\prime}_{2}(\overline {15})_{3}
-4d^{\prime\prime}_{3}(\overline {15})_{1}-4d^{\prime\prime}_{3}(\overline {15})_{2}-
4d^{\prime\prime}_{3}(\overline {15})_{3}+2d^{\prime\prime}_{4}(\overline {15})_{1}+2d^{\prime\prime}_{4}(\overline {15})_{2}+
2d^{\prime\prime}_{4}(\overline {15})_{3}\nonumber\\&&+2d^{\prime\prime}_{5}(\overline {15})_{1}+2d^{\prime\prime}_{5}(\overline {15})_{2}
+2d^{\prime\prime}_{5}(\overline {15})_{3}
+d^{\prime\prime}_{1}(6)_{1}+d^{\prime\prime}_{1}(6)_{2}+d^{\prime\prime}_{1}(6)_{3}+
+d^{\prime\prime}_{2}(6)_{1}+d^{\prime\prime}_{2}(6)_{2}+d^{\prime\prime}_{2}(6)_{3}\nonumber\\&&-2d^{\prime\prime}_{3}(6)_{1}-2d^{\prime\prime}_{3}(6)_{2}-2d^{\prime\prime}_{3}(6)_{3}
+d^{\prime\prime}_{4}(6)_{1}+d^{\prime\prime}_{4}(6)_{2}+d^{\prime\prime}_{4}(6)_{3}+d^{\prime\prime}_{5}(6)_{1}+d^{\prime\prime}_{5}(6)_{2}+d^{\prime\prime}_{5}(6)_{3}]\nonumber
\end{eqnarray}
With these facts, after symmetrizing the amplitude to the fully-symmetric one, we find $S1.1 =0$ still holds.
We find that the other relations of eq.(\ref{resu3}) also hold.
In a very similar way one can obtain the momentum dependent corrections to the isospin breaking effects by replacing $W$ by $X$ as what have been done for $SU(3)$ case. We find that all the relations of eq.(\ref{reis1}), eq.(\ref{reis2}) and eq.(\ref{reis3}) still hold.
Before close this section, we would like to make a comment about finite mass effects of $m^2_\pi$ and $m^2_K$. In practical extraction of the amplitudes, one should also consider $SU(3)$ corrections in phase space due to final state meson mass differences which come in order $m^2_{\pi, K}/m^2_{B, B_s}$ since $m^2_\pi \sim m_u, m_d$ and $m^2_K \sim m_s$ which are the same order of $SU(3)$ and isospin breaking effects considered earlier. This can be done systematically when extracting the fully-symmetric amplitudes by Dalitz plot analysis. In the momentum dependent amplitudes discussed in section II, when express the amplitudes, for example those in Eq.(2.12), in terms of the $s$, $t$ and $u$ variables, terms proportional to $m^2_\pi$ and $m_K^2$ will be generated. However, these will not generate new terms compared with those already included in the $SU(3)$ and isospin breaking effects considered earlier in this section. The conclusions drawn above will not be changed.
\section{Conclusions and Discussions}
Flavor $SU(3)$ and isospin symmetries have been considered to be powerful tools in analyzing B decays. Such analyses are usually hampered by a relatively large strange quark mass which breaks $SU(3)$ symmetry down to isospin symmetry. The isospin symmetry also breaks down when up and down quark mass difference is kept. It is therefore interesting to find relations which are not sensitive to $SU(3)$ and isospin breaking effects. We have carried out detailed analyses including $SU(3)$ and isospin breaking effects due to u, d and s quark mass differences for $B \to PPP$ decays. We find that a class of relations in fully-symmetric amplitudes are not broken by $SU(3)$ breaking effects due to a non-zero strange quark mass, and
the relations
\begin{eqnarray}
S4.1=0\;,\;\;S1.2+S1.3 = 0\;,\;\;\sqrt{2}D2.1+D2.2=0\;,\;\;D2.2+D3.1=0\;,
\end{eqnarray}
hold even isospin breaking effects due to up and down quark mass difference is included.
Measurements for these relations will provide important information about B decays in the SM.
We would like to end the paper by commenting $SU(3)$ breaking effects on the $U$-spin symmetry relations in the following
\begin{eqnarray}
&&T_{\triangle s=-1}(B^+\rightarrow K^+K^+K^-)=T_{\triangle s=0}(B^+\rightarrow \pi^+\pi^+\pi^-)\;,\nonumber\\
&&T_{\triangle s=-1}(B^+\rightarrow K^+\pi^+\pi^-)=T_{\triangle s=0}(B^+\rightarrow \pi^+K^+K^-)\;.
\label{uspin}
\end{eqnarray}
The momentum dependent terms also respect the above relations. The above equalities also hold for the fully-symmetric amplitudes for corresponding pairs of decay modes in the $SU(3)$ limit.
These relations imply in the SM that the CP violating rate asymmetries defined by
$A_{asy} = \Gamma(B\to PPP) - \Gamma(\bar B \to \bar P\bar P\bar P)$ are equal but opposite in sign for each pair of decay modes above.
For the fully-symmetric amplitudes of these decays modes, we also have
\begin{eqnarray}
&&T_{\triangle s=-1}(B^+\rightarrow K^+K^+K^-)_{FS}=T_{\triangle s=-1}(B^+\rightarrow K^+\pi^+\pi^-)_{FS}\;,\nonumber\\
&&T_{\triangle s=0}(B^+\rightarrow \pi^+\pi^+\pi^-)_{FS}=T_{\triangle s=0}(B^+\rightarrow \pi^+K^+K^-)_{FS}\;.
\label{uspin-fs}
\end{eqnarray}
Unlike the other fully-symmetric amplitudes studied in previous sections, the relations in eq.(\ref{uspin}) and eq.(\ref{uspin-fs}) are broken when $SU(3)$ breaking effects due to a non-zero strange quark mass is included. Therefore there may be sizeable deviation for these relations.
Relations in eq.(\ref{uspin}) have been discussed recently. It was found that indeed there are large $SU(3)$ breaking effects\cite{xu-li-he1, he-li-xu2,gronau}. The relations in eq. (\ref{uspin}) and eq.(\ref{uspin-fs}) will not provide as much insight as those from the fully-symmetric amplitudes which still hold when isospin breaking effects are included discussed earlier.
However, we find that the $SU(3)$ breaking effects due to a non-zero strange quark mass and the isospin breaking effects due to the difference of up and down quark masses are equal for some of the above relations with
\begin{eqnarray}
&&\Delta T(B^{+}\rightarrow K^+ \pi^+ \pi^-)-\Delta T(B^{+}\rightarrow K^+ K^+ K^- )\nonumber\\
&& =- [\Delta T(B^{+}\rightarrow \pi^+ K^+ K^-)-\Delta T(B^{+}\rightarrow \pi^+ \pi^+ \pi^- )]\\
&&=3[a^T_{4}(6)+3a^T_{4}(\overline{15})+b^T_{3}(\overline{3})+b^T_{4}(6)+3b^T_{4}(\overline{15})+c^T_{2}(\overline{3})-c^T_{2}(6)+c^T_{3}(6)\nonumber\\
&&\quad-c^T_{5}(6)-c^T_{2}(\overline{15})-c^T_{3}(\overline{15})-2c^T_{4}(\overline{15})+3c^T_{5}(\overline{15})-d^T_{5}(6)+3d^T_{5}(\overline{15})]\;,\nonumber
\end{eqnarray}
and the isospin breaking effects satisfy
\begin{eqnarray}
&&\Delta T^{I}(B^{+}\rightarrow K^+ \pi^+ \pi^-)-\Delta T^{I}(B^{+}\rightarrow K^+ K^+ K^- )\nonumber\\
&&=- [\Delta T^{I}(B^{+}\rightarrow \pi^+ K^+ K^-)-\Delta T^{I}(B^{+}\rightarrow \pi^+ \pi^+ \pi^- )]\\
&&=-[a^{T^{I}}_{4}(6)+3a^{T^{I}}_{4}(\overline{15})+b^{T^{I}}_{3}(\overline{3})+b^{T^{I}}_{4}(6)+3b^{T^{I}}_{4}(\overline{15})+c^{T^{I}}_{2}(\overline{3})-c^{T^{I}}_{2}(6)+c^{T^{I}}_{3}(6)\nonumber\\
&&\quad-c^{T^{I}}_{5}(6)-c^{T^{I}}_{2}(\overline{15})-c^{T^{I}}_{3}(\overline{15})-2c^{T^{I}}_{4}(\overline{15})+3c^{T^{I}}_{5}(\overline{15})-d^{T^{I}}_{5}(6)+3d^{T^{I}}_{5}(\overline{15})].\nonumber
\end{eqnarray}
With the momentum dependent corrections to the $SU(3)$ and isospin breaking effects, detailed analyses similar to those carried out in Sectin III, we find that the following relation is still true to the order we can considering
\begin{eqnarray}
&&\Delta T^{p}(B^{+}\rightarrow K^+ \pi^+ \pi^-)-\Delta T^{p}(B^{+}\rightarrow K^+ K^+ K^- )\nonumber\\
&& =- [\Delta T^{p}(B^{+}\rightarrow \pi^+ K^+ K^-)-\Delta T^{p}(B^{+}\rightarrow \pi^+ \pi^+ \pi^- )],
\end{eqnarray}
Here $\Delta T^{p}$ including both the momentum dependent corrections to the $SU(3)$ and isospin breaking effects.
The above leads to the following relation which is not affected by first order $SU(3)$ breaking effects due to strange, up and down quark mass differences,
\begin{eqnarray}
&&T (B^{+}\rightarrow K^+ \pi^+ \pi^- )_{FS}-T(B^{+}\rightarrow K^+ K^+ K^- )_{FS} \nonumber\\
&&=T(B^{+}\rightarrow \pi^+ \pi^+ \pi^-)_{FS} - T(B^{+}\rightarrow \pi^+ K^+ K^- )_{FS}\neq 0\;,
\end{eqnarray}
and similarly for penguin amplitudes $P_{FS}$.
When the relevant decay amplitudes are measured precisely, one can also obtain useful information for $B$ decays in the SM.
\begin{acknowledgments}
The work was supported in part by MOE Academic Excellent Program (Grant No: 102R891505) and MOST of ROC, and in part by NNSF(Grant No:11175115) and Shanghai Science and Technology Commission (Grant No: 11DZ2260700) of PRC.
\end{acknowledgments}
|
\section{Introduction}
Modern power systems, such as distributed and smart grids, increasingly rely on network-wide information for which signal processing and communication technologies are needed to estimate the power quality parameters. The information of interest includes system states and operating conditions at different nodes in the system. Modern decentralised and semi-autonomous systems aim to make full use of such information to provide enhanced reliability, efficiency and power quality at both the transmission and distribution side. In addition, the emergence of smart and distributed grids (e.g. microgrids) has brought to light a number of power quality problems related to the unpredictability of power demand/supply and the accompanying system imbalance {\cite{Bollen_DSP_Power_IEEE_SPM_2009}}.
Power quality refers to the ``fitness'' of electrical power delivered to the consumer, and is characterised by continuity of service, harmonic behaviour, variations of amplitudes of line voltages, and fluctuations of system frequency around its nominal value {\cite{Bollen_Book}}. Maintaining the system frequency within its prescribed tolerance range is a prerequisite for the health of power systems, as frequency variations are indicative of unbalanced system conditions, such as generation-consumption imbalances or faults. Accurate and robust frequency estimation is therefore a key parameter for both the control and protection of power system and to maintain power quality {\cite{Power_Standard}}.
Currently used frequency estimation techniques include: i) Fourier transform approaches \cite{Wang_DFT_2004,Lobos_Freq_DFT_1997}, ii) gradient decent and least squares adaptive estimation \cite{Xiao_Freq_est_LMS_IEEE_Trans_2005}, and iii) state space methods and Kalman filters \cite{Huang_Freq_est_CEKF_IEEE_Trans_2008,Dash_freq_est_1999,Hu_Kuh_State_MicroGrids_IJCNN_2011,Mangesius_Obradovic_power_grid_2012}.
However, these are typically designed for single-phase systems operating under balanced conditions (equal voltage amplitudes) and are not adequate for the demands of modern three--phase and dynamically optimised power systems.
The modelling of three-phase systems requires simultaneous measurement of the three phase voltages and a mathematical framework to reduce redundancy through transformations of the variables {\cite{Clarke_Inst_Currents_Voltages_1951}}{\cite{Paap_Symmetrical_Components_2000}. This is typically achieved using Clarke's $\alpha\beta$ transform {\cite{Clarke_Book}}, the output of which is a complex variable $v=v_{\alpha} + \jmath \, v_{\beta}$, that represents balanced systems without loss of information. However, when operating in unbalanced system conditions, standard (strictly linear) complex-valued estimators inevitably introduce biased estimates together with spurious frequency estimates at twice the system frequency \cite{Beeman_1955}. This is attributed to noncircular phasor trajectories of unbalanced $\alpha\beta$ voltages which require widely linear estimators to model the system dynamics \cite{ Xia_Adaptive_Frequency_Freq_2012_IEEE_SP_Mag,mandic2014patent, Dini_Three_Phase_Freq_IEEE_IM_2013}.
The existing widely linear frequency estimators are theoretically rigorous but practically restrictive, as they only apply to a single node of a power system. Indeed, distributed and smart grids require cooperation of geographically distributed nodes equipped with local data acquisition and learning capabilities {\cite{Bollen_DSP_Power_IEEE_SPM_2009}}. Distributed estimation and fusion has already found application in both military and civilian scenarios \cite{Stadter_Dist_Spacecraft_IEEEMag_2002,Mandic_Fusion_Book_Short_2008, Cattivelli_Sayed_IEEETranAC_Dist_KF_2010, Olfati_Saber_Flocking_2006}, as cooperation between the nodes (sensors) provides more accurate and robust estimation over the independent nodes, while approaching the performance of centralised systems at much reduced communication overheads. {Recent approaches include distributed least-mean-square estimation \cite{Lopes_Sayed_TSP_Dist_LMS_2008, Yili_Dist_ACLMS_2011} and Kalman filtering \cite{Cattivelli_Sayed_IEEETranAC_Dist_KF_2010, Olfati_Saber_Dist_KF, Khan_Dist_KF_IEEETransSP_2008}, however, these references considered models with circular measurement noise without cross-nodal correlations, which is not typical in real world power systems. }
To this end, we introduce a class of distributed sequential state estimators for the generality of complex signals, both second order circular (proper) and second order noncircular (improper). The state space structure of the proposed distributed augmented (widely linear) complex Kalman filter (D-ACKF) and the distributed augmented extended Kalman filter (D-ACEKF) also accounts for the correlation between the observation noises at neighbouring nodes, encountered when node signals are exposed to common sources of interference (harmonics, fluctuations of reacive power). {The aim of this work is therefore two--fold: (i) to extend existing widely linear frequency estimators {\cite{Dini_Class_WLKF_IEEE_TNNLS_2012,Dini_Three_Phase_Freq_IEEE_IM_2013}} to the distributed scenario; (ii) to investigate the advantages of D-ACKF and D-AECKF over the existing D-CKFs \cite{Olfati_Saber_Dist_KF, Cattivelli_Sayed_IEEETranAC_Dist_KF_2010} through analysis and comprehensive case studies on distributed frequency estimation under balanced and unbalanced conditions, a key issue in the control and management of microgrids, electricity islands and smart grids.}
\section{Background on widely linear modelling}
Complex-valued signal processing underpins a number of real-world applications, including wireless communications \cite{Gao_Dist_Cooperative_IEEETranComm_2011}\cite{Mao_Wireless_Comm_IEEETranIFS_2007} and power systems \cite{Xia_WL_Freq_2011}. However, standard approaches are generally suboptimal and are adequate only for a restrictive class of proper (second order circular) complex processes, for which probability distributions are rotation invariant \cite{Picinbono97} \cite{Moreno_2008}, \cite{ErikssonAugmentID06}. A zero-mean proper complex signal, $\mathbf{x}$, has equal powers in the real and imaginary part so that the covariance matrix, $\mathbf{R}_{\mathbf{x}}=E\{\mathbf{x}\mathbf{x}^H\}$, is sufficient to represent its complete second-order statistics. However, full statistical description of improper (noncircular) complex signals requires {\em augmented complex statistics} which includes the second order moments called the pseudocovariance, $\mathbf{P}_{\mathbf{x}}=E\{\mathbf{x}\mathbf{x}^T\}$, which models both the power difference and cross-correlation between the real and imaginary parts of a complex signal \cite{Picinbono97}.
{\bf Widely linear model.} Consider the minimum mean square error (MSE) estimator of a zero-mean real valued random vector $\mathbf{y}$ in terms of an observed zero-mean real vector $\mathbf{x}$, that is, $\hat{\mathbf{y}} = E\{\mathbf{y}|\mathbf{x}\}$. For jointly normal $\mathbf{y}$ and $\mathbf{x}$, the optimal linear estimator is given by
\begin{equation}
\hat{\mathbf{y}} = \mathbf{A}\mathbf{x}
\label{Eq:Linear_Estimator}
\end{equation}
where $\mathbf{A} =\mathbf{R}_{\mathbf{yx}}\mathbf{R}_{\mathbf{x}}^{-1}$ is a coefficient matrix, and $\mathbf{R}_{\mathbf{yx}} = E\{\mathbf{y}\mathbf{x}^H\}$. Standard, `strictly linear' estimation in ${\mathbb C}$ assumes the same model but with complex valued $\mathbf{y}, \mathbf{x}$, and $\mathbf{A}$. However, when $\mathbf{y}$ and $\mathbf{x}$ are jointly improper $\mathbf{P}_{\mathbf{yx}} = E\{\mathbf{y}\mathbf{x}^T\} \neq \mathbf{0}$, and since $\mathbf{x}$ is also improper $\mathbf{P}_{\mathbf{x}} \neq \mathbf{0}$. The optimal estimator for both proper and improper data is then represented by the widely linear estimator\footnote{The `widely linear' model is associated with the signal generating system, whereas ``augmented statistics'' describe statistical properties of measured signals. Both the terms `widely linear' and `augmented' are used to name the resulting algorithms - in our work we mostly use the term `augmented'.}, given by \cite{Picinbono97}
\begin{equation}
\hat{\mathbf{y}} = \mathbf{B}\mathbf{x} + \mathbf{C}\mathbf{x}^*
= \mathbf{W} \mathbf{x}^{a}
\label{Eq:WL_model}
\end{equation}
where $\mathbf{B} = \mathbf{R}_{\mathbf{yx}}\mathbf{D} + \mathbf{P}_{\mathbf{yx}}\mathbf{E}^*$ and $\mathbf{C} = \mathbf{R}_{\mathbf{yx}}\mathbf{E} + \mathbf{P}_{\mathbf{yx}}\mathbf{D}^*$ are coefficient matrices, with $\mathbf{D} = (\mathbf{R}_{\mathbf{x}} - \mathbf{P}_{\mathbf{x}}\mathbf{R}_{\mathbf{x}}^{*-1}\mathbf{P}_{\mathbf{x}}^*)^{-1}$ and $\mathbf{E} = -(\mathbf{R}_{\mathbf{x}} - \mathbf{P}_{\mathbf{x}}\mathbf{R}_{\mathbf{x}}^{*-1}\mathbf{P}_{\mathbf{x}}^*)^{-1}\mathbf{P}_{\mathbf{x}}\mathbf{R}_{\mathbf{x}}^{*-1}$, while $\mathbf{x}^a = [\mathbf{x}^T, \mathbf{x}^H]^T$ is the augmented input vector, and $\mathbf{W} = [\mathbf{B}, \mathbf{C}]$ the optimal coefficient matrix. Based on ({\ref{Eq:WL_model}}) the full second order statistics for the generality of complex data (proper and improper) is therefore contained in the augmented covariance matrix
\begin{eqnarray}
\label{rza1}
\mathbf{R}^a_{\mathbf{x}} = E\{\mathbf{x}^a\mathbf{x}^{aH}\} = \begin{bmatrix} \mathbf{R}_{\mathbf{x}} & \mathbf{P}_{\mathbf{x}}\\\mathbf{P}^*_{\mathbf{x}} & \mathbf{R}^*_{\mathbf{\mathbf{x}}} \end{bmatrix}
\end{eqnarray}
that also includes the pseudocovariance \cite{Picinbono97}\cite{Dini_Class_WLKF_IEEE_TNNLS_2012}\cite{Moreno_2009}.
\section{Diffusion Kalman Filtering}
Every node $i$ in a distributed system (see Figure {\ref{fig:NetwotkTopologyGeneral}}) can be described by a standard linear state space model, given by {\cite{Hayes96}}
\begin{subequations}\label{linearstatespace}
\begin{eqnarray}
\mathbf{x}_{n} &=& \mathbf{F}_{n-1}\mathbf{x}_{n-1} + \mathbf{w}_{n} \label{s1}\\
\mathbf{y}_{i,n} &=& \mathbf{H}_{i,n}\mathbf{x}_{n} + \mathbf{v}_{i,n} \label{dm1}
\end{eqnarray}
\end{subequations}
where $\mathbf{x}_{n} \in \mathbb{C}^L$ and $\mathbf{y}_{i,n} \in \mathbb{C}^{K}$ are respectively the state vector at time instant $n$ and observation (measurement) vector at node $i$. The symbol $\mathbf{F}_n$ denotes the state transition matrix, $\mathbf{w}_{n} \in \mathbb{C}^L$ white state noise, while $\mathbf{H}_{i,n}$ is the observation matrix, and $\mathbf{v}_{i,n} \in \mathbb{C}^{K}$ is white measurement measurement noise (both at node $i$). Standard state space models assume the noises $\mathbf{w}_{n}$ and $\mathbf{v}_{i,n}$ to be uncorrelated and zero-mean, so that their covariances and pseudocovariance matrices are
\begin{eqnarray}
E \begin{bmatrix} \mathbf{w}_{n} \\ \mathbf{v}_{i,n} \end{bmatrix} \begin{bmatrix} \mathbf{w}_{k} \\ \mathbf{v}_{i,k} \end{bmatrix}^H
=
\begin{bmatrix} \mathbf{Q}_n & \mathbf{0} \\ \mathbf{0} & \mathbf{R}_{i,n} \end{bmatrix} \delta_{nk} \\
E \begin{bmatrix} \mathbf{w}_{n} \\ \mathbf{v}_{i,n} \end{bmatrix} \begin{bmatrix} \mathbf{w}_{k} \\ \mathbf{v}_{i,k} \end{bmatrix}^T
=
\begin{bmatrix} \mathbf{P}_n & \mathbf{0} \\ \mathbf{0} & \mathbf{U}_{i,n} \end{bmatrix} \delta_{nk}
\end{eqnarray}
where $\delta_{nk}$ is the standard Kronecker delta function.
\subsection{Distributed Complex Kalman Filter}\label{Sec:D-CKF}
{The distinguishing feature of the proposed class of distributed Kalman filters is that we have used the diffusion strategy in \cite{Cattivelli_Sayed_IEEETranAC_Dist_KF_2010} with more general system and noise models that do not restrict the correlation properties of the cross-nodal observation noises or the signal and noise circularity at different nodes; this generalises existing distributed Kalman filtering algorithms \cite{Cattivelli_Sayed_IEEETranAC_Dist_KF_2010, Olfati_Saber_Dist_KF, Kar_Moura_Dist_KF_IEEETransSP_2011} to wider application scenarios.} Figure \ref{fig:NetwotkTopologyGeneral} illustrates the distributed estimation scenario; the highlighted neighbourhood of node $i$ compromises the set of nodes, denoted by $\mathcal{N}_i$, that communicate with the node $i$ (including Node $i$ itself). The state estimate at node $i$ with a complex Kalman filter (CKF) is then based on all the data from the neighbourhood $\mathcal{N}_i$ consisting of $M = |\mathcal{N}_i|$ nodes, and is denoted by $\mathbf{\widehat{\underline{x}}}_{i,n|n}$, where the symbol $|\mathcal{N}_i|$ denotes the number of nodes in the neighbourhood $\mathcal{N}_i$. Finally, the collective neighbourhood observation equation at node $i$ is given by
\begin{eqnarray} \label{c_neighbourhood_obs_eqn}
\mathbf{\underline{y}}_{i,n} = \mathbf{\underline{H}}_{i,n}\mathbf{x}_{n} + \mathbf{\underline{v}}_{i,n}
\end{eqnarray}
while the collective (neighbourhood) variables are defined as
\begin{align*}
\mathbf{\underline{y}}_{i,n} &= \begin{bmatrix} \mathbf{y}_{i_1,n}^T, \mathbf{y}_{i_2,n}^T, \ldots , \mathbf{y}_{i_M,n}^T \end{bmatrix}^T\\
%
\mathbf{\underline{H}}_{i,n} &= \begin{bmatrix} \mathbf{H}_{i_1,n}^T, \mathbf{H}_{i_2,n}^T, \ldots , \mathbf{H}_{i_M,n}^T \end{bmatrix}^T\\
%
\mathbf{\underline{v}}_{i,n} &= \begin{bmatrix} \mathbf{v}_{i_1,n}^T, \mathbf{v}_{i_2,n}^T, \ldots , \mathbf{v}_{i_M,n}^T \end{bmatrix}^T
\end{align*}
where $\{i_1, i_2, \ldots, i_M\}$ are the nodes in the neighbourhood $\mathcal{N}_i$. The covariance and pseudocovariance matrices of the collective observation noise vector then become
\begin{align*}
&\mathbf{\underline{R}}_{i,n} = E\{\mathbf{\underline{v}}_{i,n}\mathbf{\underline{v}}_{i,n}^H \}
=
\begin{bmatrix}[l] \mathbf{R}_{i_1,n} & \mathbf{R}_{i_1i_2,n} & \cdots & \mathbf{R}_{i_1i_M,n} \\
\mathbf{R}_{i_2i_1,n} & \mathbf{R}_{i_2,n} & \cdots & \mathbf{R}_{i_2i_M,n} \\
\vdots & \vdots & \ddots & \vdots \\
\mathbf{R}_{i_Mi_1,n} & \mathbf{R}_{i_Mi_2,n} & \cdots & \mathbf{R}_{i_M,n}
\end{bmatrix} \\
&\mathbf{\underline{U}}_{i,n} = E\{\mathbf{\underline{v}}_{i,n}\mathbf{\underline{v}}_{i,n}^T \}
=
\begin{bmatrix}[l] \mathbf{U}_{i_1,n} & \mathbf{U}_{i_1i_2,n} & \cdots & \mathbf{U}_{i_1i_M,n} \\
\mathbf{U}_{i_2i_1,n} & \mathbf{U}_{i_2,n} & \cdots & \mathbf{U}_{i_2i_M,n} \\
\vdots & \vdots & \ddots & \vdots \\
\mathbf{U}_{i_Mi_1,n} & \mathbf{U}_{i_Mi_2,n} & \cdots & \mathbf{U}_{i_M,n}
\end{bmatrix}
\end{align*}
where $\mathbf{R}_{i_a,n} = E\{\mathbf{v}_{i_a,n}\mathbf{v}_{i_a,n}^H\}$, $\mathbf{R}_{i_ai_b,n} = E\{\mathbf{v}_{i_a,n}\mathbf{v}_{i_b,n}^H\}$, $\mathbf{U}_{i_a,n} = E\{\mathbf{v}_{i_a,n}\mathbf{v}_{i_a,n}^T\}$ and $\mathbf{U}_{i_ai_b,n} = E\{\mathbf{v}_{i_a,n}\mathbf{v}_{i_b,n}^T\}$, for $a,b \in \{1,2,\ldots , M\}$.
\vspace{3mm}\\
\noindent
{\bf Diffusion step.} The so found local neighbourhood state estimates are followed by the diffusion (combination) step,
\begin{figure}[t]
\centering
\includegraphics[trim = 4mm 30mm 4mm 9mm,scale=0.8]{{Network_topology_General}.eps}
\caption{An example of a distributed network topology.}
\label{fig:NetwotkTopologyGeneral}
\end{figure}
\begin{eqnarray}
\mathbf{\widehat{x}}_{i,n|n} = \sum_{k \in \mathcal{N}_i} c_{k,i}\mathbf{\widehat{\underline{x}}}_{k,n|n} \label{eq:diff_step}
\end{eqnarray}
which calculates the diffused state estimates $\mathbf{\widehat{x}}_{i,n|n}$ as a weighted sum of the estimates from the neighbourhood $\mathcal{N}_i$, where $c_{k,i} \geq 0$ are the weighting coefficients satisfying $\sum_{k \in \mathcal{N}_i} c_{k,i} = 1$.
{The combination weights $c_{k,i}$ used by the diffusion step in \eqref{eq:diff_step} can follow the Metropolis \cite{Lopes_Sayed_TSP_Dist_LMS_2008}, Laplacian \cite{Bru_Convergence_1994} or the nearest neighbour \cite{Cattivelli_Sayed_IEEETranAC_Dist_KF_2010} rules. However, finding the set of optimal weights is a challenging task, though progress has been made in important cases\cite{Li_Optimal_Dist_fusion_2003, Sayed_Proceedings_2014, Sayed_arXiv}.}
The distributed complex Kalman filter (D-CKF) aims to approach the performance of a centralised Kalman filter (access to data from all the nodes) via neighbourhood collaborations and diffusion, and is summarised in Algorithm \ref{alg:D-CKF}. Each node within D-CKF forms a collective observation as in \eqref{c_neighbourhood_obs_eqn}, using information from its neighbours; thereafter, each node computes a neighbourhood state estimate which is again shared with neighbours in order to be used for the diffusion step.
\begin{algorithm}[t]
Initialisation: For each node $i=1,2,\ldots,N$
\begin{eqnarray}
\mathbf{\widehat{x}}_{i,0|0} &=& E\{\mathbf{x}_{0}\} \nonumber\\
\mathbf{M}_{i,0|0} &=& E\{(\mathbf{x}_{0}-E\{\mathbf{x}_{0}\})(\mathbf{x}_{0}-E\{\mathbf{x}_{0}\})^H\} \nonumber
\end{eqnarray}
For every time instant $n=1,2,\ldots$
$-$ Evaluate at each node $i=1,2,\ldots,N$
\begin{align}
\mathbf{\widehat{x}}_{i,n|n-1} &= \mathbf{F}_{n-1}\mathbf{\widehat{x}}_{i,n-1|n-1} \label{px_est}\\
%
\mathbf{M}_{i,n|n-1} &= \mathbf{F}_{n-1}\mathbf{M}_{i,n-1|n-1}\mathbf{F}^{H}_{n-1} + \mathbf{Q}_n \label{pMSE}\\
%
\mathbf{G}_{i,n} &= \mathbf{M}_{i,n|n-1}\mathbf{\underline{H}}^{H}_{i,n}\big(\mathbf{\underline{H}}_{i,n}\mathbf{M}_{i,n|n-1}\mathbf{\underline{H}}^{H}_{i,n} + \mathbf{\underline{R}}_{i,n}\big)^{-1} \label{gain}\\
%
\mathbf{\widehat{\underline{x}}}_{i,n|n} &= \mathbf{\widehat{x}}_{i,n|n-1} + \mathbf{G}_{i,n}\big(\mathbf{\underline{y}}_{i,n} - \mathbf{\underline{H}}_{i,n}\mathbf{\widehat{x}}_{i,n|n-1}\big)\label{x_est}\\
%
\mathbf{M}_{i,n|n} &= (\mathbf{I} - \mathbf{G}_{i,n}\mathbf{\underline{H}}_{i,n})\mathbf{M}_{i,n|n-1}\label{MSE}
\end{align}
$-$ For every node $i$, compute the diffusion update as
\begin{align}\label{Diff}
\mathbf{\widehat{x}}_{i,n|n} = \sum_{k \in \mathcal{N}_i} c_{k,i}\mathbf{\widehat{\underline{x}}}_{k,n|n}
\end{align}
\caption{The D-CKF}
\label{alg:D-CKF}
\end{algorithm}
\\ \vspace{1mm}
\textbf{Remark \#1:} The D-CKF algorithm\footnote{The matrices $\mathbf{M}_{i,n|n}$ and $\mathbf{M}_{i,n|n-1}$ do not represent the covariances of $\mathbf{\widehat{x}}_{i,n|n}$ and $\mathbf{\widehat{x}}_{i,n|n-1}$, as is the case for the standard Kalman filter operating on linear Gaussian systems. This is due to the use of the suboptimal diffusion step, which updates the state estimate and not the covariance matrix $\mathbf{M}_{i,n|n}$.} given in Algorithm \ref{alg:D-CKF} is a variant of \cite{Cattivelli_Sayed_IEEETranAC_Dist_KF_2010}. It employs the standard (strictly linear) state space model \eqref{linearstatespace}, and therefore does not cater for widely linear complex state space models or noncircular state and observation noises ($\mathbf{P}_{n} \neq \mathbf{0}$ and $\mathbf{U}_{i,n} \neq \mathbf{0}$ for $i=1,\ldots,N$).
{Unlike existing distributed complex Kalman filters, the proposed D-CKF algorithm in \eqref{px_est} -- \eqref{Diff} caters for the correlations between the neighbourhood observation noises. When no such correlations exits, is identical to Algorithm \ref{alg:D-CKF} in \cite{Cattivelli_Sayed_IEEETranAC_Dist_KF_2010}.}
\subsection{Distributed Augmented Complex Kalman Filter} \label{Sec:D-ACKF}
As stated in Remark \#1, current state space algorithms do not cater for widely linear state and observation models or for improper measurements, states, and state and observation noises. To this end, we next employ the widely linear model in \eqref{Eq:WL_model} to introduce the widely linear version of the standard, strictly linear, distributed state space model\footnote{Observe that the noise models can also be widely linear, in which case:\\ $\mathbf{w}_{n} =\mathbf{D}_n\acute{\mathbf{w}}_{n} + \mathbf{E}_n\acute{\mathbf{w}}^*_{n}\quad$ and $\quad\mathbf{v}_{n} =\mathbf{F}_n\acute{\mathbf{v}}_{n} + \mathbf{H}_n\acute{\mathbf{v}}^*_{n}$, where $\mathbf{D, E, F, H}$ are coefficient matrices and $\acute{\mathbf{w}}_{n}$ and $\acute{\mathbf{v}}_{n}$ are proper or improper noise models.} in \eqref{linearstatespace} (see also \cite{Dini_Class_WLKF_IEEE_TNNLS_2012}, \cite{Dini_Dist_ACKF_ASILOMAR_2012})
\begin{subequations}\label{WLstatespace}
\begin{eqnarray}
\mathbf{x}_{n} &=& \mathbf{F}_{n-1}\mathbf{x}_{n-1} + \mathbf{A}_{n-1}\mathbf{x}^*_{n-1} + \mathbf{w}_{n} \\
\mathbf{y}_{i,n} &=& \mathbf{H}_{i,n}\mathbf{x}_{n} + \mathbf{B}_{i,n}\mathbf{x}_{n}^* + \mathbf{v}_{i,n} \label{WLdm1}
\end{eqnarray}
\end{subequations}
The compact, augmented representation, of this model is
\begin{subequations}\label{AUGstatespace}
\begin{eqnarray}
\mathbf{x}^a_{n} &=& \mathbf{F}^a_{n-1}\mathbf{x}^a_{n-1} + \mathbf{w}^a_{n} \label{as1}\\
\mathbf{y}_{i,n}^a &=& \mathbf{H}_{i,n}^a\mathbf{x}_{n}^a + \mathbf{v}_{i,n}^a
\end{eqnarray}
\end{subequations}
where $\mathbf{x}^a_n = [\mathbf{x}^T_n, \mathbf{x}^H_n]^T$ and $\mathbf{y}^a_n = [\mathbf{y}^T_n, \mathbf{y}^H_n]^T$, while
\begin{eqnarray*}
\mathbf{F}^a_{n} = \begin{bmatrix} \mathbf{F}_{n} & \mathbf{A}_{n}\\ \mathbf{A}^*_{n} & \mathbf{F}^*_{n} \end{bmatrix}
\text{ and }
\mathbf{H}^a_{i,n} = \begin{bmatrix} \mathbf{H}_{i,n} & \mathbf{B}_{i,n}\\ \mathbf{B}^*_{i,n} & \mathbf{H}^*_{i,n} \end{bmatrix}
\end{eqnarray*}
For \textit{strictly linear systems}, $\mathbf{A}_{n} = \mathbf{0}$ and $\mathbf{B}_{i,n} = \mathbf{0}$, so that the widely linear (augmented) state space model degenerates into a strictly linear one. However, the augmented state space representation above is still preferred in order to account for the pseudocovariances which reflect the \textit{impropriety of the signals} (\textit{cf.} widely linear nature of systems).
The augmented covariance matrices of $\mathbf{w}^a_n = [\mathbf{x}^T_n, \mathbf{w}^H_n]^T$ and $\mathbf{v}^a_{i,n} = [\mathbf{v}^T_{i,n}, \mathbf{v}^H_{i,n}]^T$ are then given by
\begin{eqnarray}
\mathbf{Q}^a_n &=& E\{\mathbf{w}^a_{n}\mathbf{w}^{aH}_{n}\} =\begin{bmatrix} \mathbf{Q}_n & \mathbf{P}_n\\ \mathbf{P}^*_n & \mathbf{Q}^*_n \end{bmatrix} \label{aQ}\\
\mathbf{R}^a_{i,n} &=& E\{\mathbf{v}^a_{i,n}\mathbf{v}^{aH}_{i,n}\} =\begin{bmatrix} \mathbf{R}_{i,n} & \mathbf{U}_{i,n}\\ \mathbf{U}^*_{i,n} & \mathbf{R}^*_{i,n} \end{bmatrix}\label{aR}
\end{eqnarray}
\noindent
{\bf Neighbourhood variables.} For collaborative estimation of the state within distributed networks, neighbourhood observation equations use all available neighbourhood observation data, to give
\begin{eqnarray} \label{ac_neighbourhood_obs_eqn}
\mathbf{\underline{y}}_{i,n} = \mathbf{\underline{H}}_{i,n}\mathbf{x}_{n} + \mathbf{\underline{B}}_{i,n}\mathbf{x}_{n}^* +\mathbf{\underline{v}}_{i,n}
\end{eqnarray}
where the symbol $\mathbf{\underline{B}}_{i,n} = \big[\mathbf{B}_{i_1,n}^T, \mathbf{B}_{i_2,n}^T, \ldots, \mathbf{B}_{i_M,n}^T \big]^T$ denotes the conjugate state matrix , and $\{i_1, i_2, \ldots, i_M\} \in \mathcal{N}_i$. The augmented neighbourhood observation equations now become
\begin{eqnarray}\label{aug_y_collective}
\mathbf{\underline{y}}_{i,n}^a = \mathbf{\underline{H}}_{i,n}^a\mathbf{x}_{n}^a + \mathbf{\underline{v}}_{i,n}^a
\end{eqnarray}
with the augmented neighbourhood variables defined as
\begin{eqnarray}
\mathbf{\underline{y}}_{i,n}^a = \begin{bmatrix} \mathbf{\underline{y}}_{i,n} \\ \mathbf{\underline{y}}_{i,n}^* \end{bmatrix},
\quad
\mathbf{\underline{H}}_{i,n}^a = \begin{bmatrix} \mathbf{\underline{H}}_{i,n} & \mathbf{\underline{B}}_{i,n}\\
\mathbf{\underline{B}}_{i,n}^* & \mathbf{\underline{H}}_{i,n}^*\end{bmatrix},
\quad
\mathbf{\underline{v}}_{i,n}^a = \begin{bmatrix} \mathbf{\underline{v}}_{i,n} \\ \mathbf{\underline{v}}_{i,n}^* \end{bmatrix}
\end{eqnarray}
Consequently, the covariance of the augmented neigbourhood observation noise $\mathbf{\underline{v}}_{i,n}^a$ takes the form
\begin{eqnarray}\label{aug_R_collective}
\mathbf{\underline{R}}_{i,n}^a
= E\{\mathbf{\underline{v}}^a_{i,n}\mathbf{\underline{v}}^{aH}_{i,n}\}
= \begin{bmatrix} \mathbf{\underline{R}}_{i,n} & \mathbf{\underline{U}}_{i,n}\\ \mathbf{\underline{U}}^*_{i,n} & \mathbf{\underline{R}}^*_{i,n} \end{bmatrix}
\end{eqnarray}
{\textbf{Remark \#2:} Observe that \eqref{aug_R_collective} caters for both the covariances $E\{\mathbf{v}_{i,n}\mathbf{v}_{i,n}^H\}$ and cross-correlations $E\{\mathbf{v}_{i,n}\mathbf{v}_{k,n}^H\}$, $i\neq k$ between the nodal observation noises. This is achieved through the covariance matrix $\mathbf{\underline{R}}_{i,n}$ and the pseudocovariances $E\{\mathbf{v}_{i,n}\mathbf{v}_{i,n}^T\}$, while the cross-pseudocorrelations $E\{\mathbf{v}_{i,n}\mathbf{v}_{k,n}^T\}$ are accounted for through the pseudocovariance matrix $\mathbf{\underline{U}}_{i,n}$.}
Finally, the augmented diffused state estimate becomes
\begin{eqnarray}
\mathbf{\widehat{x}}_{i,n|n}^a = \sum_{k \in \mathcal{N}_i} c_{k,i}\mathbf{\widehat{\underline{x}}}_{k,n|n}^a
\end{eqnarray}
and represents a weighted average of the augmented (neighbourhood) state estimates. The proposed distributed augmented complex Kalman filter (D-ACKF), based on the widely linear state space model, is summarised in Algorithm \ref{alg:D-ACKF}.
For strictly linear systems ($\mathbf{A}_{n} = \mathbf{0}$ and $\mathbf{B}_{i,n} = \mathbf{0}$ for all $n$ and $i$) and in the presence of circular state and observation noises ($\mathbf{P}_{n} = \mathbf{0}$ and $\mathbf{U}_{i,n} = \mathbf{0}$ for all $n$ and $i$), the D-ACKF and D-CKF algorithms yield identical state estimates for all time instants $n$. However, the D-ACKF is more general than the D-CKF, since it also caters for the noncircular data and noise natures together with correlated state and observation noises.
\begin{algorithm}[t]
Initialisation: For each node $i=1,2,\ldots,N$
\begin{eqnarray*}
\mathbf{\widehat{x}}_{i,0|0}^a &=& \big[E\{\mathbf{x}_{0}\}^T, E\{\mathbf{x}_{0}\}^H\big]^T \\
\mathbf{M}_{i,0|0}^a &=& E\big\{(\mathbf{x}_{0}^a-\mathbf{\widehat{x}}_{i,0|0}^a)(\mathbf{x}_{0}^a-\mathbf{\widehat{x}}_{i,0|0}^a)^{aH}\big\}
\end{eqnarray*}
For every time instant $n=1,2,\ldots$
\hspace{0.15cm}$-$ Evaluate at each node $i=1,2,\ldots,N$
\begin{align}
\mathbf{\widehat{x}}_{i,n|n-1}^a &= \mathbf{F}_{n-1}^a\mathbf{\widehat{x}}_{i,n-1|n-1}^a \label{augpx_est}\\
%
\mathbf{M}_{i,n|n-1}^a &= \mathbf{F}_{n-1}^a\mathbf{M}_{i,n-1|n-1}^a\mathbf{F}^{aH}_{n-1} + \mathbf{Q}_n^a \label{augpMSE}\\
%
\mathbf{G}_{i,n}^a &= \mathbf{M}_{i,n|n-1}^a\mathbf{\underline{H}}^{aH}_{i,n}\big(\mathbf{\underline{H}}_{i,n}^a\mathbf{M}_{i,n|n-1}^a\mathbf{\underline{H}}^{aH}_{i,n} + \mathbf{\underline{R}}_{i,n}^a\big)^{-1} \label{aug_gain}\\
%
\mathbf{\widehat{\underline{x}}}_{i,n|n}^a &= \mathbf{\widehat{x}}_{i,n|n-1}^a + \mathbf{G}_{i,n}^a\big(\mathbf{\underline{y}}_{i,n}^a - \mathbf{\underline{H}}_{i,n}^a\mathbf{\widehat{x}}_{i,n|n-1}^a\big)\label{augx_est}\\
%
\mathbf{M}_{i,n|n}^a &= (\mathbf{I} - \mathbf{G}_{i,n}^a\mathbf{\underline{H}}_{i,n}^a)\mathbf{M}_{i,n|n-1}^a\label{augMSE}
\end{align}
\hspace{0.15cm}$-$ For every node $i$, compute the diffusion update as
\begin{align}
\mathbf{\widehat{x}}_{i,n|n}^a = {\sum}_{k \in \mathcal{N}_i} c_{k,i}\mathbf{\widehat{\underline{x}}}_{k,n|n}^a \label{augdx}
\end{align}
\caption{The D-ACKF}
\label{alg:D-ACKF}
\end{algorithm}
\newline
\textbf{Remark \#3:} The information form of the D-ACKF, given in Algorithm \ref{alg:D-ACKF_info}, can be used to cater for the noncircularity of data when observation noises at different nodes are uncorrelated. Moreover, nodes in the distributed network can switch between the general D-ACKF in Algorithm \ref{alg:D-ACKF} and the information form of D-ACKF in Algorithm \ref{alg:D-ACKF_info}, depending on the correlation between the observation noises.
\begin{algorithm}[t]
Initialisation: For each node $i=1,2,\ldots,N$
\begin{eqnarray*}
\mathbf{\widehat{x}}_{i,0|0}^a &=& \big[E\{\mathbf{x}_{0}\}^T, E\{\mathbf{x}_{0}\}^H\big]^T \\
\mathbf{M}_{i,0|0}^a &=& E\big\{(\mathbf{x}_{0}^a-\mathbf{\widehat{x}}_{i,0|0}^a)(\mathbf{x}_{0}^a-\mathbf{\widehat{x}}_{i,0|0}^a)^{aH}\big\}
\end{eqnarray*}
For every time instant $n=1,2,\ldots$
\hspace{0.15cm}$-$ Evaluate at each node $i=1,2,\ldots,N$
\begin{align}
\mathbf{\widehat{x}}_{i,n|n-1}^a &= \mathbf{F}_{n-1}^a\mathbf{\widehat{x}}_{i,n-1|n-1}^a \label{augpx_est_info}\\
%
\mathbf{M}_{i,n|n-1}^a &= \mathbf{F}_{n-1}^a\mathbf{M}_{i,n-1|n-1}^a\mathbf{F}^{aH}_{n-1} + \mathbf{Q}_n^a \label{augpMSE_info}\\
%
\mathbf{S}_{i,n}^a &= {\sum}_{k \in \mathcal{N}_i} \mathbf{H}^{aH}_{k,n}(\mathbf{R}_{k,n}^a)^{-1}\mathbf{H}_{k,n}^a \\
%
\mathbf{r}_{i,n}^a &= {\sum}_{k \in \mathcal{N}_i} \mathbf{H}^{aH}_{k,n}(\mathbf{R}_{k,n}^a)^{-1}\mathbf{y}_{k,n}^a \\
%
(\mathbf{M}_{i,n|n}^a)^{-1} &= (\mathbf{M}_{i,n|n-1}^a)^{-1} + \mathbf{S}_{i,n}^a\\
\mathbf{\widehat{\chi}}_{i,n|n}^a &= \mathbf{\widehat{x}}_{i,n|n-1}^a + \mathbf{M}_{i,n|n}^a\big(\mathbf{r}_{i,n}^a - \mathbf{S}_{i,n}^a\mathbf{\widehat{x}}_{i,n|n-1}^a\big)
%
\end{align}
\hspace{0.15cm}$-$ For every node $i$, compute the diffusion update as
\begin{align}
\mathbf{\widehat{x}}_{i,n|n}^a = {\sum}_{k \in \mathcal{N}_i} c_{k,i}\mathbf{\widehat{\chi}}_{i,n|n}^a \label{augdx_info}
\end{align}
\caption{The D-ACKF Information Form}
\label{alg:D-ACKF_info}
\end{algorithm}
\subsection{Bias Analysis of the D-ACKF Estimates}
Consider the following augmented complex variables: the local (non-diffused) error at node $i \in [1,N]$ given by $\mathbf{\underline{e}}_{i,n|n}^a = \mathbf{x}_{n}^a - \mathbf{\widehat{\underline{x}}}_{i,n|n}^a$, the prediction error $\mathbf{e}_{i,n|n-1}^a = \mathbf{x}_{n}^a - \mathbf{\widehat{x}}_{i,n|n-1}^a$, and the diffused error $\mathbf{e}_{i,n|n}^a = \mathbf{x}_{n}^a - \mathbf{\widehat{x}}_{i,n|n}^a$.
Then%
\begin{align}\label{MeanaugDiffError2}
E\{\mathbf{e}_{i,n|n}^a\} &= \sum_{k \in \mathcal{N}_i} c_{k,i}
\mathbf{M}_{k,n|n}^a(\mathbf{M}_{k,n|n-1}^a)^{-1}\mathbf{F}_{n-1}^aE\{\mathbf{e}_{k,n-1|n-1}^a\} \nonumber\\
&= \mathbf{0}
\end{align}
\noindent
\textbf{Remark \#4:} The expression \eqref{MeanaugDiffError2} shows that the D-ACKF is an unbiased estimator of both proper and improper complex random signals.
\section{Distributed Augmented Complex Extended Kalman Filter} \label{Sec:D-ACEKF}
We next introduce the distributed augmented complex extended Kalman filter (D-ACEKF) for nonlinear state space models of the form
\begin{subequations}\label{nonlinearSS}
\begin{eqnarray}
\mathbf{x}_n &=& \mathbf{f}[\mathbf{x}_{n-1}] + \mathbf{w}_n \label{nsm1} \\
\mathbf{y}_{i,n} &=& \mathbf{h}_i[\mathbf{x}_n] + \mathbf{v}_{i,n} \label{nom1}
\end{eqnarray}
\end{subequations}
where the nonlinear functions $\mathbf{f}[\cdot]$ and $\mathbf{h}_i[\cdot]$ are respectively the (possibly time varying) process model and observation model at node $i$, the remaining variables are as defined above. Within the extended Kalman filter (EKF) framework, the nonlinear state and observation functions are approximated by their first order Taylor series expansions (TSE) about the state estimates $\mathbf{\widehat{x}}_{i,n-1|n-1}$ and $\mathbf{\widehat{x}}_{i,n|n-1}$ for each node $i$, so that {\cite{Dini_ACEKF_UDRC_2011}}
\begin{subequations}\label{linearisedSS}
\begin{align}
\mathbf{x}_n &\approx \mathbf{F}_{i,n-1}\mathbf{x}_{n-1} + \mathbf{A}_{i,n-1}\mathbf{x}^*_{n-1} + \mathbf{r}_{i,n-1} + \mathbf{w}_n \label{eqn-TSEstate}\\
\mathbf{y}_{i,n} &\approx \mathbf{H}_{i,n}\mathbf{x}_{n} + \mathbf{B}_{i,n}\mathbf{x}^*_{n} + \mathbf{z}_{i,n} + \mathbf{v}_{i,n} \label{eqn-TSEobs}
\end{align}
\end{subequations}
where the Jacobians of functions $\mathbf{f}[\cdot]$ and $\mathbf{h}_i[\cdot]$ are defined as
\begin{align}
\mathbf{F}_{i,n} &= \frac{\partial \mathbf{f}[\mathbf{x}]}{\partial\mathbf{x}}\Big|_{\mathbf{x} = \mathbf{\widehat{x}}_{i,n|n}}
\hspace{0.1cm}\text{,}\hspace{0.3cm}\qquad
\mathbf{A}_{i,n} = \frac{\partial \mathbf{f}[\mathbf{x}]}{\partial\mathbf{x}^*}\Big|_{\mathbf{x}^* = \mathbf{\widehat{x}}^*_{i,n|n}}
\text{,}\nonumber\\
\mathbf{H}_{i,n} &= \frac{\partial \mathbf{h}_i[\mathbf{x}]}{\partial\mathbf{x}}\Big|_{\mathbf{x} = \mathbf{\widehat{x}}_{i,n|n-1}}
\hspace{0.1cm}\text{and}\hspace{0.2cm}
\mathbf{B}_{i,n} = \frac{\partial \mathbf{h}_i[\mathbf{x}]}{\partial\mathbf{x}^*}\Big|_{\mathbf{x}^* = \mathbf{\widehat{x}}^*_{i,n|n-1}} \nonumber
\end{align}
and the vectors
\begin{align*}
\mathbf{r}_{i,n} &= \mathbf{f}[\mathbf{\widehat{x}}_{i,n-1|n-1}] - \mathbf{F}_{i,n-1}\mathbf{\widehat{x}}_{i,n-1|n-1} - \mathbf{A}_{i,n-1}\mathbf{\widehat{x}}^*_{i,n-1|n-1} \\
\mathbf{z}_{i,n} &= \mathbf{h}_i[\mathbf{\widehat{x}}_{i,n|n-1}] - \mathbf{H}_{i,n}\mathbf{\widehat{x}}_{i,n|n-1} - \mathbf{B}_{i,n}\mathbf{\widehat{x}}^*_{i,n|n-1}
\end{align*}
are deterministic inputs calculated from the state space model and state estimate. In order to cater for the full second order statistics of the variables in the linearised state space in \eqref{linearisedSS}, we shall consider its compact (augmented) version given by
\begin{subequations}\label{augLinearisedSS}
\begin{align}
\mathbf{x}^a_n &\approx \mathbf{F}^a_{i,n-1}\mathbf{x}^a_{n-1} + \mathbf{r}^a_{i,n-1} + \mathbf{w}^a_n \\
\mathbf{y}^a_{i,n} &\approx \mathbf{H}^a_{i,n}\mathbf{x}^a_{n} + \mathbf{z}^a_{i,n} + \mathbf{v}^a_{i,n}
\end{align}
\end{subequations}
where $\mathbf{r}^a_{i,n} = \big[ \mathbf{r}^T_{i,n} , \mathbf{r}^H_{i,n} \big]^T$, $\mathbf{z}^a_{i,n} = \big[ \mathbf{z}^T_{i,n} , \mathbf{z}^H_{i,n} \big]^T$, while\\\\
$\mathbf{F}^a_{i,n} = \begin{bmatrix} \mathbf{F}_{i,n} & \mathbf{A}_{i,n}\\ \mathbf{A}^*_{i,n} & \mathbf{F}^*_{i,n} \end{bmatrix}$ \quad and \quad
$\mathbf{H}^a_{i,n} = \begin{bmatrix} \mathbf{H}_{i,n} & \mathbf{B}_{i,n}\\ \mathbf{B}^*_{i,n} & \mathbf{H}^*_{i,n} \end{bmatrix}$.\\\\
Observe that the collective neighbourhood augmented observation equation for node $i$ takes the form
\begin{eqnarray} \label{neighbourhood_obs_eqn_NL}
\mathbf{\underline{y}}^a_{i,n} = \mathbf{\underline{h}}_i^a[\mathbf{x}_{n}] + \mathbf{\underline{v}}^a_{i,n}
\end{eqnarray}
while the collective observation function defined as
\begin{eqnarray*}
\mathbf{\underline{h}}_i^a[\mathbf{x}_{n}] &=& \Big[\mathbf{\underline{h}}^T_i[\mathbf{x}_{n}], \mathbf{\underline{h}}^H_i[\mathbf{x}_{n}]\Big]^T \\
\mathbf{\underline{h}}_i[\mathbf{x}_{n}] &=& \Big[\mathbf{h}^T_{i_1}[\mathbf{x}_{n}], \mathbf{h}^T_{i_2}[\mathbf{x}_{n}], \ldots,
\mathbf{h}^T_{i_M}[\mathbf{x}_{n}]\Big]^T
\end{eqnarray*}
where $i \in \{i_1, i_2, \ldots, i_M\}$ are all the nodes in the neighbourhood $\mathcal{N}_i$. The first order approximation of \eqref{neighbourhood_obs_eqn_NL} is then
\begin{align}
\mathbf{\underline{y}}^a_{i,n} &\approx \mathbf{\underline{H}}^a_{i,n}\mathbf{x}^a_{n} + \mathbf{\underline{z}}^a_{i,n} + \mathbf{\underline{v}}^a_{i,n}
\end{align}
with the Jacobian of the collective observation function
\begin{eqnarray*}
\mathbf{\underline{H}}^a_{i,n} = \begin{bmatrix} \mathbf{\underline{H}}_{i,n} & \mathbf{\underline{B}}_{i,n}\\ \mathbf{\underline{B}}^*_{i,n} & \mathbf{\underline{H}}^*_{i,n} \end{bmatrix}
\end{eqnarray*}
where $\mathbf{\underline{H}}_{i,n} = \big[\mathbf{H}_{i_1,n}^T, \mathbf{H}_{i_2,n}^T, \ldots, \mathbf{H}_{i_M,n}^T \big]^T$ and
$\mathbf{\underline{B}}_{i,n} = \big[\mathbf{B}_{i_1,n}^T, \mathbf{B}_{i_2,n}^T, \ldots, \mathbf{B}_{i_M,n}^T \big]^T$, wherein
\begin{align}
\mathbf{H}_{i_k,n} &= \frac{\partial \mathbf{h}_{i_k}[\mathbf{x}]}{\partial\mathbf{x}}\Big|_{\mathbf{x} = \mathbf{\widehat{x}}_{i,n|n-1}}
\hspace{0.1cm}\text{and}\hspace{0.2cm}
\mathbf{B}_{i_k,n} = \frac{\partial \mathbf{h}_{i_k}[\mathbf{x}]}{\partial\mathbf{x}^*}\Big|_{\mathbf{x}^* = \mathbf{\widehat{x}}^*_{i,n|n-1}} \nonumber
\end{align}
\begin{algorithm}[t]
Initialisation: For each node $i=1,2,\ldots,N$
\begin{eqnarray*}
\mathbf{\widehat{x}}_{i,0|0}^a &=& \big[E\{\mathbf{x}_{0}\}^T, E\{\mathbf{x}_{0}\}^H\big]^T \\
\mathbf{M}_{i,0|0}^a &=& E\big\{(\mathbf{x}_{0}^a-\mathbf{\widehat{x}}_{i,0|0}^a)(\mathbf{x}_{0}^a-\mathbf{\widehat{x}}_{i,0|0}^a)^{aH}\big\}
\end{eqnarray*}
For every time instant $n=1,2,\ldots$
\hspace{0.15cm}$-$ Evaluate at each node $i=1,2,\ldots,N$
\begin{align}
\mathbf{\widehat{x}}_{i,n|n-1}^a &= \Big[\mathbf{f}^T[\mathbf{\widehat{x}}_{i,n-1|n-1}], \mathbf{f}^H[\mathbf{\widehat{x}}_{i,n-1|n-1}]\Big]^T \label{augpex_est}\\
%
\mathbf{M}_{i,n|n-1}^a &= \mathbf{F}_{i,n-1}^a\mathbf{M}_{i,n-1|n-1}^a\mathbf{F}^{aH}_{i,n-1} + \mathbf{Q}_n^a \label{augpeMSE}\\
%
\mathbf{G}_{i,n}^a &= \mathbf{M}_{i,n|n-1}^a\mathbf{\underline{H}}^{aH}_{i,n}\big(\mathbf{\underline{H}}_{i,n}^a\mathbf{M}_{i,n|n-1}^a\mathbf{\underline{H}}^{aH}_{i,n} + \mathbf{\underline{R}}_{i,n}^a\big)^{-1} \label{aug_egain}\\
%
\mathbf{\widehat{\underline{x}}}_{i,n|n}^a &= \mathbf{\widehat{x}}_{i,n|n-1}^a + \mathbf{G}_{i,n}^a\big(\mathbf{\underline{y}}_{i,n}^a - \mathbf{\underline{h}}_{i}^a[\mathbf{\widehat{x}}_{i,n|n-1}]\big)\label{augex_est}\\
%
\mathbf{M}_{i,n|n}^a &= (\mathbf{I} - \mathbf{G}_{i,n}^a\mathbf{\underline{H}}_{i,n}^a)\mathbf{M}_{i,n|n-1}^a\label{augeMSE}
\end{align}
\hspace{0.15cm}$-$ For every node $i$, compute the diffusion update as
\begin{align}
\mathbf{\widehat{x}}_{i,n|n}^a = \sum_{k \in \mathcal{N}_i} c_{k,i}\mathbf{\widehat{\underline{x}}}_{k,n|n}^a \label{augdex}
\end{align}
\caption{Distributed Augmented Complex EKF}
\label{alg:D-ACEKF}
\end{algorithm}
Algorithm \ref{alg:D-ACEKF} summarises the proposed distributed augmented complex extended Kalman filter, where each node $i$ shares its (nonlinear) observation model $\mathbf{h}_i[\cdot]$ with its neighbours.
\newline
{
{\bf Remark \#5:} The D-ACEKF algorithm in Algorithm \ref{alg:D-ACEKF} extends the Distributed Extended Kalman filter in \cite{Sayed_EKF_2010} by using the widely linear model, and caters for the second order statistical moments of the state and noise models, together with the correlations present between the nodal observation noises.}
\section{Distributed Widely Linear Frequency Estimation}
The proposed augmented state space models are particularly suited for frequency estimation in power grid, as due to system inertia, the frequency can be assumed identical over the network of measurement nodes, while unbalanced systems generate noncircular measurements \cite{ mandic2014patent, Xia_Adaptive_Frequency_Freq_2012_IEEE_SP_Mag}. {For a three phase system, the instantaneous voltages at a node $i$ are given by}
\begin{eqnarray}
{ v_{a,i,n} } & {=}& { V_{a,n}\cos(\omega nT + \phi) + z_{a,i,n} }\nonumber\\
{v_{b,i,n}} & {=}& {V_{b,n}\cos(\omega nT + \phi - 2\pi/3 + \Delta_b) + z_{b,i,n}} \nonumber\\
{v_{c,i,n}} & {=}& {V_{c,n}\cos(\omega nT + \phi + 2\pi/3 + \Delta_c) + z_{c,i,n} }
\label{Eq:3_phase_voltages}
\end{eqnarray}
where $V_{a,n}$, $V_{b,n}$ and $V_{c,n}$ are the amplitudes of the three-phase voltages at time instant $n$, $\omega = 2\pi f$ the angular frequency, $f$ the system frequency, $T$ the sampling interval, and $\phi$ the phase of the fundamental component, while $ z_{a,i,n}$, $ z_{a,i,n}$ and $ z_{a,i,n}$ are zero-mean observation noise processes. {The terms $\Delta_b, \Delta_c$ are used to indicate the phase distortions relative to a balanced three-phase system.} The phase voltages in ({\ref{Eq:3_phase_voltages}}) are first mapped to the complex voltage of orthogonal $\alpha$ and $\beta$ components using Clarke's $\alpha \beta$ transformation, to give {
\begin{eqnarray}
\begin{bmatrix} v_{\alpha,i,n} \\ v_{\beta,i,n}\end{bmatrix} = \sqrt{\frac{2}{3}}\begin{bmatrix} 1 & -\frac{1}{2} & -\frac{1}{2} \\ 0 & \frac{\sqrt{3}}{2} & -\frac{\sqrt{3}}{2} \end{bmatrix} \begin{bmatrix} v_{a,i,n} \\ v_{b,i,n} \\ v_{c,i,n}\end{bmatrix}.
\label{Eq:Clarke_transform}
\end{eqnarray}
}
The $\alpha \beta$ voltage is then converted to a scalar complex signal $v_{i,n}=v_{\alpha,i,n} + jv_{\beta,i,n}$. For balanced systems, in which $V_{a,n}=V_{b,n}=V_{c,n}$, { and $\Delta_b = \Delta_c = 0$}, the variables
\begin{eqnarray*}
v_{\alpha,i,n}&=& A_n\cos(\omega nT + \phi) + z_{\alpha,i,n} \\
v_{\beta,i,n} &=& A_n\cos(\omega nT + \phi + \frac{\pi}{2}) + z_{\beta,i,n}
\end{eqnarray*}
where $A_n = \frac{\sqrt{6}}{2}V_{a,n}$. The noises mapped via the $\alpha \beta$ transform now become
\begin{eqnarray*}
z_{\alpha,i,n} &=& \sqrt{2/3}\Big(z_{a,i,n} - \frac{1}{2}z_{b,i,n} - \frac{1}{2}z_{c,i,n}\Big) \\
z_{\beta,i,n} &=& \sqrt{2/3}\Big( \frac{\sqrt{3}}{2}z_{b,i,n} - \frac{\sqrt{3}}{2}z_{c,i,n}\Big)
\end{eqnarray*}
For balanced systems, this scalar complex model takes a recursive form, so that for every node\footnote{The usual assumption in this type of estimation is that for a sampling frequency $>> 50$Hz, we have $A_n\approx A_{n-1}$.}
\begin{eqnarray}
v_{i,n} &=& v_{\alpha,i,n} + jv_{\beta,i,n} \nonumber\\
&=& A_ne^{j(\omega nT + \phi)} + z_{i,n} \nonumber\\
&=& v_{i,n-1}e^{j\omega T} + z_{i,n}
\label{Eq:output_voltage2}
\end{eqnarray}
where $z_{i,n} = z_{\alpha,i,n} + j \, z_{\beta,i,n}$.
The state space model for this system at a node $i$ is shown in \eqref{Eq:ss_state4} and \eqref{Eq:ss_obs4}, where the state variables $x_{k}$ and $u_n$ are used to estimate the exponential $e^{j\omega T}$ and the observation $v_{i,n}$ respectively, while $\mathbf{w}_n$ and $z_{i,n}$ are the state and observation noises respectively.
The system frequency is then derived from the state variable $x$ as follows:
\begin{eqnarray}
\hat{f}_n &=& \frac{1}{2\pi T}\arcsin\big(\Im(x_n)\big)
\label{Eq:freq_estimate1}
\end{eqnarray}
where $\Im(\cdot)$ is the imaginary part of a complex quantity.
\begin{algorithm}[t]
\caption{A Strictly Linear State Space (StSp-SL)}
State equation:
\begin{subequations}\label{alg:SS1_L}
\begin{align}
\begin{bmatrix}x_n \\ u_n\end{bmatrix}
= \begin{bmatrix}x_{n-1} \\ u_{n-1}x_n \end{bmatrix} + \mathbf{w}_{n-1}
\label{Eq:ss_state4}
\end{align}
Observation equation:
\begin{align}
v_{i,n} = \begin{bmatrix} 0 & 1 \end{bmatrix} \begin{bmatrix}x_n \\ u_n\end{bmatrix} + z_{i,n}
\label{Eq:ss_obs4}
\end{align}
\end{subequations}
\end{algorithm}
Figure \ref{fig:AB_trajectory} illustrates the trajectory of the transformed voltage (a rotating vector -- phasor), indicating that for a balanced system, Clarke's voltage $v_{i,n}$ in \eqref{Eq:output_voltage2} has a circular trajectory.
However, the model in \eqref{Eq:ss_state4} and \eqref{Eq:ss_obs4} becomes inaccurate when the system is operating under unbalanced conditions, in which case, the voltage amplitudes $V_{a,n}$, $V_{b,n}$ and $V_{c,n}$ are no longer equal or if the condition $\Delta_b = \Delta_c = 0$ is not satisfied, and the system trajectory becomes noncircular (ellipse in Figure \ref{fig:AB_trajectory}). For unbalanced systems, therefore, the correct system model is widely linear, and is given by
{
\begin{align} \nonumber
v_{i,n} = & {} \sqrt{\frac{2}{3}} \begin{bmatrix} 1 & j \end{bmatrix}
\begin{bmatrix} 1 & -\frac{1}{2} & -\frac{1}{2} \\ 0 & \frac{\sqrt{3}}{2} & -\frac{\sqrt{3}}{2} \end{bmatrix} \begin{bmatrix} v_{a,i,n} \\ v_{b,i,n} \\ v_{c,i,n}\end{bmatrix} \\ \label{eq:cmplx_voltage}
= & {} \sqrt{\frac{2}{3}} \begin{bmatrix} 1 & e^{j2\pi/3} & e^{-j2\pi/3} \end{bmatrix}
\begin{bmatrix} v_{a,i,n} \\ v_{b,i,n} \\ v_{c,i,n}\end{bmatrix}.
\end{align}
For the compactness of notation, we set $\theta_n = \omega n T + \phi$. Using the relationship $\cos(y) = \frac{e^{jy} + e^{-jy}}{2}$, the three phase voltages at each node $i$ can be rewritten as
\begin{align} \nonumber
\begin{bmatrix} v_{a,i,n} \\ v_{b,i,n} \\ v_{c,i,n} \end{bmatrix} = {} &
\frac{1}{2}
\begin{bmatrix}
V_{a,n} \left( e^{j\theta_n} +e^{-j\theta_n} \right)
\\
V_{b,n} \left(e^{j(\theta_n - 2\pi/3 + \Delta_b) } + e^{-j(\theta_n - 2\pi/3 + \Delta_b) } \right)
\\
V_{c, n} \left( e^{j(\theta_n + 2\pi/3 + \Delta_c) } + e^{-j(\theta_n + 2\pi/3 + \Delta_c) } \right )
\end{bmatrix}
\\
& {}
+ \begin{bmatrix} z_{a,i,n} \\ z_{b,i,n} \\ z_{c,i,n} \end{bmatrix} \label{eq:polarform}
\end{align}
Substituting \eqref{eq:polarform} into \eqref{eq:cmplx_voltage} gives
\begin{align}
v_{i,n}
= {} & A_{n}e^{j\theta_n} + B_{n}e^{-j\theta_n} + z_{i,n}
\label{Eq:output_voltageWL1}
\end{align}
where
\begin{align}
\label{Eq:A_B_coefficents}
A_n = {} & \frac{\sqrt{6}}{6}\left(V_{a,n} + V_{b,n}e^{j\Delta_b} + V_{c,n}e^{j \Delta_c} \right)\\ \nonumber
B_n = {} & \frac{\sqrt{6}}{6} \left(V_{a,n} + V_{b,n}e^{-j(\Delta_b + 2\pi/3)} + V_{c,n}e^{-j( \Delta_c - 2 \pi/3)} \right).
\end{align}
}
\begin{figure}[t]
\centering
\subfloat{\includegraphics[scale=0.7]{Voltage_Trajectory_SNR_20}} \vspace{-0.35cm}
\caption{Noncircularity in power systems. For a balanced system, characterised by $V_{a,n}=V_{b,n}=V_{c,n}$ and $\Delta_b = \Delta_c = 0$, the trajectory of Clarke's voltage $v_{i,n}$ is circular (blue line). For unbalanced systems, the voltage trajectory is noncircular (red line), such as in the case of a 100\% single-phase voltage sag illustrated by the ellipse. In both cases the signal to noise ratio (SNR) was $20$dB.}
\label{fig:AB_trajectory}
\end{figure}
For a balanced system under nominal conditions, we have $V_{a,n}=V_{b,n}=V_{c,n}$ and $\Delta_b = \Delta_c = 0$, and the coefficient $B_n$ in ({\ref{Eq:A_B_coefficents}}) vanishes, so that the system is adequately characterised by the strictly linear model in \eqref{Eq:output_voltage2}. For unbalanced systems, $B_n\neq0$ and the Clarke's voltage $v_{i,n}$ is noncircular, so that the expression in \eqref{Eq:output_voltageWL1} is second order optimal for both balanced and unbalanced conditions. This expression can be written in the form of a widely linear recursive mode
\begin{eqnarray}
v_{i,n} &=& v_{n-1}h_{n-1} + v^*_{n-1}g_{n-1} + z_{i,n}
\label{Eq:output_voltageWL2}
\end{eqnarray}
which is a first-order widely linear autoregressive model with coefficients $h_{k}$ and $g_{k}$. The widely linear (augmented) state space model corresponding to \eqref{Eq:output_voltageWL2} is defined in \eqref{Eq:ss_state3}, where the state vector consists of the strictly linear and conjugate weights $h_n$ and $g_n$, and the variable $u_n$ which corresponds to the noise-free widely linear observation $v_n$. The system frequency can be computed from the state variables $h_n$ and $g_n$ as follows:
\begin{eqnarray}
\hat{f}_n &=& \frac{1}{2\pi T}\arcsin\big(\Im(h_n + a_ng_n)\big)
\label{Eq:freq_estimate2}
\end{eqnarray}
\begin{eqnarray*}
a_n &=& \frac{-j\Im(h_n)+ j\sqrt{\Im^2(h_n) - |g_n|^2}}{g_n}
\label{Eq:a1}
\end{eqnarray*}
The state space model in \eqref{Eq:ss_state3} and \eqref{Eq:ss_obs3} provides a realistic and robust characterisation of real world power systems, as it represents both balanced and unbalanced systems, while its nonlinear state equation also models the coupling between state variables. The StSp-WL model in \eqref{alg:SS1_WL} can be implemented using the proposed distributed augmented complex extended Kalman filter in Section {\ref{Sec:D-ACEKF}}.
\begin{algorithm}[t]
\caption{A Widely Linear State Space (StSp-WL)}
State equation:\vspace{-0.1cm}
\begin{subequations}\label{alg:SS1_WL}
\begin{align}
\begin{bmatrix}h_n \\ g_n\\ u_n \\ h^*_n \\ g^*_n\\ u^*_n\end{bmatrix}
\!=\! \begin{bmatrix}h_{n-1} \\ g_{n-1}\\ u_{n-1}h_{n-1} + u^*_{n-1}g_{n-1}\\ h^*_{n-1} \\ g^*_{n-1}\\ \!u^*_{n-1}h^*_{n-1} \!+\! u_{n-1}g^*_{n-1}\!\end{bmatrix} \!+\! \mathbf{w}_{n-1}
\label{Eq:ss_state3}
\end{align}
Observation equation: \vspace{-0.1cm}
\begin{align}
\begin{bmatrix} v_{i,n}\\ v^*_{i,n} \end{bmatrix} = \begin{bmatrix} 0 & 0 & 1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 1\end{bmatrix} \begin{bmatrix}h_n \\ g_n\\ u_n \\ h^*_n \\ g^*_n\\ u^*_n\end{bmatrix} + \begin{bmatrix} z_{i,n}\\ z^*_{i,n} \end{bmatrix}
\label{Eq:ss_obs3}
\end{align}
\end{subequations}
\end{algorithm}
\section{Frequency Estimation Examples}
\begin{figure}[t]
\centering
\subfloat{ {\includegraphics[clip = true, trim = 25mm 40mm 270mm 15mm,scale=0.3]{Network_Power}}}
\caption{A distributed { power network} with $N=5$ nodes {(Sub-stations) }used in the simulations.}
\label{fig:NetwotkTopologyN5}
\end{figure}
\begin{figure}[t]
\centering
\subfloat{\includegraphics[scale=0.39]{Type_C_sag_Circularity}}
\subfloat{\includegraphics[scale=0.39]{Type_C_sag_Phasor}}\\ \vspace{-0.5cm}
\subfloat{\includegraphics[scale=0.39]{Type_D_sag_Circularity}}
\subfloat{\includegraphics[scale=0.39]{Type_D_sag_Phasor}}
\caption{Geometric (left) and phasor (right) views of Type C and D unbalanced voltage sags. The real-imaginary phasor plots illustrate the noncircularity of Clarke's voltage in unbalanced conditions, indicated by the elliptical shapes of circularity plots. The parameters of this ellipse (degree of noncircularity) serve to identify the type of fault (in this case a voltage sag).}
\label{fig:circularity_phasor_sags_C_D}
\end{figure}
\begin{figure*}[tbh]
\centering
\subfloat{\includegraphics[scale=0.7]{ACEKF_5nodes_SNR_40_typeC_then_typeD}}
\subfloat{\includegraphics[scale=0.7]{D-ACEKF_5nodes_SNR_40_typeC_then_typeD}}
\vspace{-0.35cm}
\caption{Frequency estimation performance of single node (CEKF and ACEKF) and distributed (D-CEKF and D-ACEKF) algorithms for a system at $40$dB SNR. The system is balanced up to $0.1$s, it then undergoes a Type C voltage sag followed by a Type D voltage sag at 0.3s.}
\label{fig:sag}\vspace{-0.5cm}
\end{figure*}
\begin{figure*}[tbh]
\centering
\subfloat{\includegraphics[scale=0.7]{ACEKF_5nodes_SNR_30}}
\subfloat{\includegraphics[scale=0.7]{D-ACEKF_5nodes_SNR_30}}
\vspace{-0.35cm}
\caption{Frequency estimation performance of single node (CEKF and ACEKF) and distributed (D-CEKF and D-ACEKF) algorithms for a balanced system in the presence of doubly white circular Gaussian noises at $30$dB SNR. As expected, the strictly and widely linear algorithms had similar performance.}
\label{fig:freq_SNR30}\vspace{-0.5cm}
\end{figure*}
\begin{figure*}[tbh]
\centering
\subfloat{\includegraphics[scale=0.7]{ACEKF_5nodes_Spike_noise}}
\subfloat{\includegraphics[scale=0.7]{D-ACEKF_5nodes_Spike_noise}}
\vspace{-0.35cm}
\caption{Frequency estimation performance of single node (CEKF and ACEKF) and distributed (D-CEKF and D-ACEKF) algorithms when the phase voltages at one of the nodes in the network are contaminated with random spike noise at $20$\% p.u.}
\label{fig:spikeNoise}\vspace{-0.5cm}
\end{figure*}
\begin{figure*}[tbh]
\centering
\subfloat{\includegraphics[scale=0.7]{ACEKF_5nodes_SNR_40_stepFreqChange}}
\subfloat{\includegraphics[scale=0.7]{D-ACEKF_5nodes_SNR_40_stepFreqChange}}
\vspace{-0.35cm}
\caption{Frequency estimation performance of single node (CEKF and ACEKF) and distributed (D-CEKF and D-ACEKF) algorithms for a power system at $40$dB SNR, which experiences a step change in system frequency to $51$Hz.}
\label{fig:stepChange}
\end{figure*}
\begin{figure*}[tbh]
\centering
\subfloat {\includegraphics[clip = true, trim =0mm 0mm 0mm 10mm, scale=0.2]{hilbert_typeD}}
\subfloat {\includegraphics[clip = true, trim =0mm 0mm 0mm 10mm, scale=0.2]{d_hilbert}}
\vspace{-0.35cm}
\caption{{Frequency estimation performance comparison between Hilbert Transform based instantaneous frequency estimate and ACEKF estimate. \textit{Left}: Single node based estimate (Hilbert and ACEKF). \textit{Right}: Distributed implementations of the algorithm (D-Hilbert and D-ACEKF) }}
\label{fig:hilbert}\vspace{-0.5cm}
\end{figure*}
The performance of the algorithms was evaluated over both comprehensive illustrative simulation studies and a real world example. Within the synthetic data, the power system under consideration had a nominal frequency of $50$Hz, and was sampled at a rate of $5$kHz while the state vectors of all the nodes in the network were initialised to $50.5$Hz. Without loss in generality, we used the distributed network topology shown in Figure \ref{fig:NetwotkTopologyN5}. The strictly linear state space model in \eqref{alg:SS1_L} and the widely linear state space model in \eqref{alg:SS1_WL} were implemented using the proposed widely linear D-ACEKF and its strictly linear version D-CEKF. For rigour, the uncooperative CEKF and ACEKF were also considered.
\noindent \textbf{Case Study \#1: Voltage sags.} In the first set of simulations, the performances of the algorithms were evaluated for an initially balanced system which became unbalanced after undergoing a Type C voltage sag starting at $0.1$s, characterised by a $20$\% voltage drop and $10^o$ phase offset on both the $v_b$ and $v_c$ channels, followed by a Type D sag starting at $0.3$s, characterised by a $20$\% voltage drop at line $v_a$ and a $10$\% voltage drop on both $v_b$ and $v_c$ with a $5^o$ phase angle offset. The degrees of noncircularity of these system imbalances are illustrated in Figure \ref{fig:circularity_phasor_sags_C_D}. Figure \ref{fig:sag} shows that, conforming with the analysis, the widely linear algorithms, ACEKF and D-ACEKF, were able to converge to the correct system frequency for both balanced and unbalanced operating conditions, while the strictly linear algorithms, CEKF and D-CEKF, were unable to accurately estimate the frequency during the voltage sag due to under-modeling of the system (not accounting for its widely linear nature) -- see \eqref{Eq:output_voltageWL2}. As expected, the widely linear and strictly linear algorithms had similar performances under balanced conditions, as illustrated in the time interval $0$-$0.1$s. The distributed algorithms, D-CEKF and D-ACEKF, outperformed their uncooperative counterparts, CEKF and ACEKF, owing to the sharing of information between neighbouring nodes.
\noindent \textbf{Case Study \#2: Presence of noise.} Figure \ref{fig:freq_SNR30} illustrates frequency estimation for a balanced system in the presence of white Gaussian noise at $30$dB SNR, while Figure \ref{fig:spikeNoise} illustrates frequency estimation in the presence of random spike noise, which typically occurs in the presence of switching devices or lightning. The distributed algorithms, D-CEKF and D-ACEKF, outperformed their uncooperative counterparts, CEKF and ACEKF, because neighbouring nodes were able to share information to facilitate better estimation performances. In both cases, the distributed algorithms exhibited lower variance in the frequency estimate.
\noindent \textbf{Case Study \#3: Frequency jumps.} Figure \ref{fig:stepChange} illustrates the performance of both single node and distributed algorithms (strictly and widely linear) when a power system is contaminated with white noise at $40$dB SNR and undergoes a step change in system frequency, a typical scenario when generation does not match the load (microgrids and islanding). Although all the algorithms had similar responses to the step change in frequency, the distributed algorithms exhibited enhanced frequency tracking.
\begin{figure*}[ht!]
\centering
\subfloat{\includegraphics[scale=0.7]{errorBias_typeDsag}}
\subfloat{\includegraphics[scale=0.7]{errorVariance_typeDsag}}
\vspace{-0.35cm}
\caption{Bias and variance analysis of the proposed distributed state space frequency estimators for an unbalanced system undergoing a type D voltage sag. {\it Left:} Estimation bias. {\it Right:} Estimation variance.}
\label{fig:BiasVar}
\end{figure*}
{
\noindent \textbf{Case Study \#4: Comparison with Hilbert method.} To further illustrate the advantages of widely linear modelling in a three-phase setting, the next case study compares the performance of the widely linear ACEKF to the Hilbert Transform based instantaneous frequency estimate from one of the three-phases. The three phase voltage signal at each node was simulated with circular white noise at $30$dB SNR and Type D imbalance after $0.25$s. Figure \ref{fig:hilbert} shows that the ``Hilbert" frequency estimate, found by differentiating the instantaneous phase angle of the single phase voltage that underwent the Hilbert Transform, was poor since the differentiation step (high pass filter) in the Hilbert method is not robust to noise. This was observed both in single-node and distributed settings.
}
\noindent \textbf{Case Study \#5: Bias and variance of the proposed estimators.} For rigour, Figure \ref{fig:BiasVar} provides the analysis of the bias and variance for the proposed distributed frequency estimators. The algorithms were evaluated at different SNR levels for an unbalanced system undergoing a Type D voltage sag (see also Figure {\ref{fig:circularity_phasor_sags_C_D}}). Observe that both the single- and multiple-node widely linear algorithms, ACEKF and D-ACEKF, were asymptotically unbiased (left panel, see Remark \#4) while both the single- and multiple-node strictly linear algorithms were biased. In terms of the variance of the estimators (right panel), both the distributed estimation algorithms outperformed their non-cooperative counterparts, while the only consistent estimator was the proposed distributed augmented complex extended Kalman filter.
{
\noindent \textbf{Real World Case Study.} We next assessed the performance of the proposed algorithms on a real world case study using three-phase voltage measurements from two adjacent sub-stations in Malaysia\footnote{ {Due to data provenance issues we are unable to reveal which particular sub-stations the measurements originate from.}} during a brief line-to-earth fault on the 29th June 2014. This caused voltage sags, similar to those in Case Study \#1. The three-phase measurements were sampled at 5kHz and the voltage values were normalized. The left panel in Figure \ref{fig:realWorld} shows the normalized $\alpha \beta$ voltages at one of the sub-stations. The fault that occurs in phase A around 0.1s is reflected in the voltage dip in $v_{\alpha}$. The right panel in Figure \ref{fig:realWorld} shows the frequency estimate from the D-ACEKF and D-CEKF. Conforming with the analysis and the single node scenario in Figure \ref{fig:sag}, the collaborative widely linear D-ACEKF was able to track the real world frequency of a power network under both balanced and unbalanced conditions, whereas the strictly linear D-CEKF was unable to track the frequency after 0.1s when the line-to-earth fault (non-circularity) occurred.
\begin{figure*}[tbh]
\centering
\subfloat {\includegraphics[clip = true, trim =0mm 0mm 0mm 10mm, scale=0.2]{VoltageSub}}
\subfloat {\includegraphics[clip = true, trim =0mm 0mm 0mm 10mm, scale=0.2]{ACLMS_RealWorld}}
\vspace{-0.35cm}
\caption{Real world case study. \textit{Left}: The $\alpha \beta$ voltages at Sub-station 1 before and during the fault event. \textit{Right}: Frequency estimation using the proposed algorithms before and during the fault event.}
\label{fig:realWorld}
\end{figure*}
}
\section{Conclusions}
We have proposed a novel class of diffusion based distributed complex valued Kalman filters for cooperative frequency estimation in power systems. To cater for the general case of improper states, observations, and state and observation noises, we have introduced the distributed (widely linear) augmented complex Kalman filter (D-ACKF) and its nonlinear version, the distributed augmented complex Kalman filter (D-ACEKF). These have been shown to provide sequential state estimation of the generality of complex signals, both circular and noncircular, within a general and unifying framework which also caters for correlated nodal observation noises.
This novel widely linear framework has been applied for distributed state space based frequency estimation in the context of three-phase power systems, and has been shown to be optimal for both balanced and unbalanced operating conditions. Simulations over a range of both balanced and unbalanced power system conditions {for both synthetic and real world measurements} have illustrated that the proposed distributed state space algorithms are consistent estimators, offering unbiased and minimum frequency estimation in both balanced an unbalanced system conditions, together with simultaneous frequency estimation and fault identification.
\bibliographystyle{ieeetr}
\balance
\section{Introduction}
Modern power systems, such as distributed and smart grids, increasingly rely on network-wide information for which signal processing and communication technologies are needed to estimate the power quality parameters. The information of interest includes system states and operating conditions at different nodes in the system. Modern decentralised and semi-autonomous systems aim to make full use of such information to provide enhanced reliability, efficiency and power quality at both the transmission and distribution side. In addition, the emergence of smart and distributed grids (e.g. microgrids) has brought to light a number of power quality problems related to the unpredictability of power demand/supply and the accompanying system imbalance {\cite{Bollen_DSP_Power_IEEE_SPM_2009}}.
Power quality refers to the ``fitness'' of electrical power delivered to the consumer, and is characterised by continuity of service, harmonic behaviour, variations of amplitudes of line voltages, and fluctuations of system frequency around its nominal value {\cite{Bollen_Book}}. Maintaining the system frequency within its prescribed tolerance range is a prerequisite for the health of power systems, as frequency variations are indicative of unbalanced system conditions, such as generation-consumption imbalances or faults. Accurate and robust frequency estimation is therefore a key parameter for both the control and protection of power system and to maintain power quality {\cite{Power_Standard}}.
Currently used frequency estimation techniques include: i) Fourier transform approaches \cite{Wang_DFT_2004,Lobos_Freq_DFT_1997}, ii) gradient decent and least squares adaptive estimation \cite{Xiao_Freq_est_LMS_IEEE_Trans_2005}, and iii) state space methods and Kalman filters \cite{Huang_Freq_est_CEKF_IEEE_Trans_2008,Dash_freq_est_1999,Hu_Kuh_State_MicroGrids_IJCNN_2011,Mangesius_Obradovic_power_grid_2012}.
However, these are typically designed for single-phase systems operating under balanced conditions (equal voltage amplitudes) and are not adequate for the demands of modern three--phase and dynamically optimised power systems.
The modelling of three-phase systems requires simultaneous measurement of the three phase voltages and a mathematical framework to reduce redundancy through transformations of the variables {\cite{Clarke_Inst_Currents_Voltages_1951}}{\cite{Paap_Symmetrical_Components_2000}. This is typically achieved using Clarke's $\alpha\beta$ transform {\cite{Clarke_Book}}, the output of which is a complex variable $v=v_{\alpha} + \jmath \, v_{\beta}$, that represents balanced systems without loss of information. However, when operating in unbalanced system conditions, standard (strictly linear) complex-valued estimators inevitably introduce biased estimates together with spurious frequency estimates at twice the system frequency \cite{Beeman_1955}. This is attributed to noncircular phasor trajectories of unbalanced $\alpha\beta$ voltages which require widely linear estimators to model the system dynamics \cite{ Xia_Adaptive_Frequency_Freq_2012_IEEE_SP_Mag,mandic2014patent, Dini_Three_Phase_Freq_IEEE_IM_2013}.
The existing widely linear frequency estimators are theoretically rigorous but practically restrictive, as they only apply to a single node of a power system. Indeed, distributed and smart grids require cooperation of geographically distributed nodes equipped with local data acquisition and learning capabilities {\cite{Bollen_DSP_Power_IEEE_SPM_2009}}. Distributed estimation and fusion has already found application in both military and civilian scenarios \cite{Stadter_Dist_Spacecraft_IEEEMag_2002,Mandic_Fusion_Book_Short_2008, Cattivelli_Sayed_IEEETranAC_Dist_KF_2010, Olfati_Saber_Flocking_2006}, as cooperation between the nodes (sensors) provides more accurate and robust estimation over the independent nodes, while approaching the performance of centralised systems at much reduced communication overheads. {Recent approaches include distributed least-mean-square estimation \cite{Lopes_Sayed_TSP_Dist_LMS_2008, Yili_Dist_ACLMS_2011} and Kalman filtering \cite{Cattivelli_Sayed_IEEETranAC_Dist_KF_2010, Olfati_Saber_Dist_KF, Khan_Dist_KF_IEEETransSP_2008}, however, these references considered models with circular measurement noise without cross-nodal correlations, which is not typical in real world power systems. }
To this end, we introduce a class of distributed sequential state estimators for the generality of complex signals, both second order circular (proper) and second order noncircular (improper). The state space structure of the proposed distributed augmented (widely linear) complex Kalman filter (D-ACKF) and the distributed augmented extended Kalman filter (D-ACEKF) also accounts for the correlation between the observation noises at neighbouring nodes, encountered when node signals are exposed to common sources of interference (harmonics, fluctuations of reacive power). {The aim of this work is therefore two--fold: (i) to extend existing widely linear frequency estimators {\cite{Dini_Class_WLKF_IEEE_TNNLS_2012,Dini_Three_Phase_Freq_IEEE_IM_2013}} to the distributed scenario; (ii) to investigate the advantages of D-ACKF and D-AECKF over the existing D-CKFs \cite{Olfati_Saber_Dist_KF, Cattivelli_Sayed_IEEETranAC_Dist_KF_2010} through analysis and comprehensive case studies on distributed frequency estimation under balanced and unbalanced conditions, a key issue in the control and management of microgrids, electricity islands and smart grids.}
\section{Background on widely linear modelling}
Complex-valued signal processing underpins a number of real-world applications, including wireless communications \cite{Gao_Dist_Cooperative_IEEETranComm_2011}\cite{Mao_Wireless_Comm_IEEETranIFS_2007} and power systems \cite{Xia_WL_Freq_2011}. However, standard approaches are generally suboptimal and are adequate only for a restrictive class of proper (second order circular) complex processes, for which probability distributions are rotation invariant \cite{Picinbono97} \cite{Moreno_2008}, \cite{ErikssonAugmentID06}. A zero-mean proper complex signal, $\mathbf{x}$, has equal powers in the real and imaginary part so that the covariance matrix, $\mathbf{R}_{\mathbf{x}}=E\{\mathbf{x}\mathbf{x}^H\}$, is sufficient to represent its complete second-order statistics. However, full statistical description of improper (noncircular) complex signals requires {\em augmented complex statistics} which includes the second order moments called the pseudocovariance, $\mathbf{P}_{\mathbf{x}}=E\{\mathbf{x}\mathbf{x}^T\}$, which models both the power difference and cross-correlation between the real and imaginary parts of a complex signal \cite{Picinbono97}.
{\bf Widely linear model.} Consider the minimum mean square error (MSE) estimator of a zero-mean real valued random vector $\mathbf{y}$ in terms of an observed zero-mean real vector $\mathbf{x}$, that is, $\hat{\mathbf{y}} = E\{\mathbf{y}|\mathbf{x}\}$. For jointly normal $\mathbf{y}$ and $\mathbf{x}$, the optimal linear estimator is given by
\begin{equation}
\hat{\mathbf{y}} = \mathbf{A}\mathbf{x}
\label{Eq:Linear_Estimator}
\end{equation}
where $\mathbf{A} =\mathbf{R}_{\mathbf{yx}}\mathbf{R}_{\mathbf{x}}^{-1}$ is a coefficient matrix, and $\mathbf{R}_{\mathbf{yx}} = E\{\mathbf{y}\mathbf{x}^H\}$. Standard, `strictly linear' estimation in ${\mathbb C}$ assumes the same model but with complex valued $\mathbf{y}, \mathbf{x}$, and $\mathbf{A}$. However, when $\mathbf{y}$ and $\mathbf{x}$ are jointly improper $\mathbf{P}_{\mathbf{yx}} = E\{\mathbf{y}\mathbf{x}^T\} \neq \mathbf{0}$, and since $\mathbf{x}$ is also improper $\mathbf{P}_{\mathbf{x}} \neq \mathbf{0}$. The optimal estimator for both proper and improper data is then represented by the widely linear estimator\footnote{The `widely linear' model is associated with the signal generating system, whereas ``augmented statistics'' describe statistical properties of measured signals. Both the terms `widely linear' and `augmented' are used to name the resulting algorithms - in our work we mostly use the term `augmented'.}, given by \cite{Picinbono97}
\begin{equation}
\hat{\mathbf{y}} = \mathbf{B}\mathbf{x} + \mathbf{C}\mathbf{x}^*
= \mathbf{W} \mathbf{x}^{a}
\label{Eq:WL_model}
\end{equation}
where $\mathbf{B} = \mathbf{R}_{\mathbf{yx}}\mathbf{D} + \mathbf{P}_{\mathbf{yx}}\mathbf{E}^*$ and $\mathbf{C} = \mathbf{R}_{\mathbf{yx}}\mathbf{E} + \mathbf{P}_{\mathbf{yx}}\mathbf{D}^*$ are coefficient matrices, with $\mathbf{D} = (\mathbf{R}_{\mathbf{x}} - \mathbf{P}_{\mathbf{x}}\mathbf{R}_{\mathbf{x}}^{*-1}\mathbf{P}_{\mathbf{x}}^*)^{-1}$ and $\mathbf{E} = -(\mathbf{R}_{\mathbf{x}} - \mathbf{P}_{\mathbf{x}}\mathbf{R}_{\mathbf{x}}^{*-1}\mathbf{P}_{\mathbf{x}}^*)^{-1}\mathbf{P}_{\mathbf{x}}\mathbf{R}_{\mathbf{x}}^{*-1}$, while $\mathbf{x}^a = [\mathbf{x}^T, \mathbf{x}^H]^T$ is the augmented input vector, and $\mathbf{W} = [\mathbf{B}, \mathbf{C}]$ the optimal coefficient matrix. Based on ({\ref{Eq:WL_model}}) the full second order statistics for the generality of complex data (proper and improper) is therefore contained in the augmented covariance matrix
\begin{eqnarray}
\label{rza1}
\mathbf{R}^a_{\mathbf{x}} = E\{\mathbf{x}^a\mathbf{x}^{aH}\} = \begin{bmatrix} \mathbf{R}_{\mathbf{x}} & \mathbf{P}_{\mathbf{x}}\\\mathbf{P}^*_{\mathbf{x}} & \mathbf{R}^*_{\mathbf{\mathbf{x}}} \end{bmatrix}
\end{eqnarray}
that also includes the pseudocovariance \cite{Picinbono97}\cite{Dini_Class_WLKF_IEEE_TNNLS_2012}\cite{Moreno_2009}.
\section{Diffusion Kalman Filtering}
Every node $i$ in a distributed system (see Figure {\ref{fig:NetwotkTopologyGeneral}}) can be described by a standard linear state space model, given by {\cite{Hayes96}}
\begin{subequations}\label{linearstatespace}
\begin{eqnarray}
\mathbf{x}_{n} &=& \mathbf{F}_{n-1}\mathbf{x}_{n-1} + \mathbf{w}_{n} \label{s1}\\
\mathbf{y}_{i,n} &=& \mathbf{H}_{i,n}\mathbf{x}_{n} + \mathbf{v}_{i,n} \label{dm1}
\end{eqnarray}
\end{subequations}
where $\mathbf{x}_{n} \in \mathbb{C}^L$ and $\mathbf{y}_{i,n} \in \mathbb{C}^{K}$ are respectively the state vector at time instant $n$ and observation (measurement) vector at node $i$. The symbol $\mathbf{F}_n$ denotes the state transition matrix, $\mathbf{w}_{n} \in \mathbb{C}^L$ white state noise, while $\mathbf{H}_{i,n}$ is the observation matrix, and $\mathbf{v}_{i,n} \in \mathbb{C}^{K}$ is white measurement measurement noise (both at node $i$). Standard state space models assume the noises $\mathbf{w}_{n}$ and $\mathbf{v}_{i,n}$ to be uncorrelated and zero-mean, so that their covariances and pseudocovariance matrices are
\begin{eqnarray}
E \begin{bmatrix} \mathbf{w}_{n} \\ \mathbf{v}_{i,n} \end{bmatrix} \begin{bmatrix} \mathbf{w}_{k} \\ \mathbf{v}_{i,k} \end{bmatrix}^H
=
\begin{bmatrix} \mathbf{Q}_n & \mathbf{0} \\ \mathbf{0} & \mathbf{R}_{i,n} \end{bmatrix} \delta_{nk} \\
E \begin{bmatrix} \mathbf{w}_{n} \\ \mathbf{v}_{i,n} \end{bmatrix} \begin{bmatrix} \mathbf{w}_{k} \\ \mathbf{v}_{i,k} \end{bmatrix}^T
=
\begin{bmatrix} \mathbf{P}_n & \mathbf{0} \\ \mathbf{0} & \mathbf{U}_{i,n} \end{bmatrix} \delta_{nk}
\end{eqnarray}
where $\delta_{nk}$ is the standard Kronecker delta function.
\subsection{Distributed Complex Kalman Filter}\label{Sec:D-CKF}
{The distinguishing feature of the proposed class of distributed Kalman filters is that we have used the diffusion strategy in \cite{Cattivelli_Sayed_IEEETranAC_Dist_KF_2010} with more general system and noise models that do not restrict the correlation properties of the cross-nodal observation noises or the signal and noise circularity at different nodes; this generalises existing distributed Kalman filtering algorithms \cite{Cattivelli_Sayed_IEEETranAC_Dist_KF_2010, Olfati_Saber_Dist_KF, Kar_Moura_Dist_KF_IEEETransSP_2011} to wider application scenarios.} Figure \ref{fig:NetwotkTopologyGeneral} illustrates the distributed estimation scenario; the highlighted neighbourhood of node $i$ compromises the set of nodes, denoted by $\mathcal{N}_i$, that communicate with the node $i$ (including Node $i$ itself). The state estimate at node $i$ with a complex Kalman filter (CKF) is then based on all the data from the neighbourhood $\mathcal{N}_i$ consisting of $M = |\mathcal{N}_i|$ nodes, and is denoted by $\mathbf{\widehat{\underline{x}}}_{i,n|n}$, where the symbol $|\mathcal{N}_i|$ denotes the number of nodes in the neighbourhood $\mathcal{N}_i$. Finally, the collective neighbourhood observation equation at node $i$ is given by
\begin{eqnarray} \label{c_neighbourhood_obs_eqn}
\mathbf{\underline{y}}_{i,n} = \mathbf{\underline{H}}_{i,n}\mathbf{x}_{n} + \mathbf{\underline{v}}_{i,n}
\end{eqnarray}
while the collective (neighbourhood) variables are defined as
\begin{align*}
\mathbf{\underline{y}}_{i,n} &= \begin{bmatrix} \mathbf{y}_{i_1,n}^T, \mathbf{y}_{i_2,n}^T, \ldots , \mathbf{y}_{i_M,n}^T \end{bmatrix}^T\\
%
\mathbf{\underline{H}}_{i,n} &= \begin{bmatrix} \mathbf{H}_{i_1,n}^T, \mathbf{H}_{i_2,n}^T, \ldots , \mathbf{H}_{i_M,n}^T \end{bmatrix}^T\\
%
\mathbf{\underline{v}}_{i,n} &= \begin{bmatrix} \mathbf{v}_{i_1,n}^T, \mathbf{v}_{i_2,n}^T, \ldots , \mathbf{v}_{i_M,n}^T \end{bmatrix}^T
\end{align*}
where $\{i_1, i_2, \ldots, i_M\}$ are the nodes in the neighbourhood $\mathcal{N}_i$. The covariance and pseudocovariance matrices of the collective observation noise vector then become
\begin{align*}
&\mathbf{\underline{R}}_{i,n} = E\{\mathbf{\underline{v}}_{i,n}\mathbf{\underline{v}}_{i,n}^H \}
=
\begin{bmatrix}[l] \mathbf{R}_{i_1,n} & \mathbf{R}_{i_1i_2,n} & \cdots & \mathbf{R}_{i_1i_M,n} \\
\mathbf{R}_{i_2i_1,n} & \mathbf{R}_{i_2,n} & \cdots & \mathbf{R}_{i_2i_M,n} \\
\vdots & \vdots & \ddots & \vdots \\
\mathbf{R}_{i_Mi_1,n} & \mathbf{R}_{i_Mi_2,n} & \cdots & \mathbf{R}_{i_M,n}
\end{bmatrix} \\
&\mathbf{\underline{U}}_{i,n} = E\{\mathbf{\underline{v}}_{i,n}\mathbf{\underline{v}}_{i,n}^T \}
=
\begin{bmatrix}[l] \mathbf{U}_{i_1,n} & \mathbf{U}_{i_1i_2,n} & \cdots & \mathbf{U}_{i_1i_M,n} \\
\mathbf{U}_{i_2i_1,n} & \mathbf{U}_{i_2,n} & \cdots & \mathbf{U}_{i_2i_M,n} \\
\vdots & \vdots & \ddots & \vdots \\
\mathbf{U}_{i_Mi_1,n} & \mathbf{U}_{i_Mi_2,n} & \cdots & \mathbf{U}_{i_M,n}
\end{bmatrix}
\end{align*}
where $\mathbf{R}_{i_a,n} = E\{\mathbf{v}_{i_a,n}\mathbf{v}_{i_a,n}^H\}$, $\mathbf{R}_{i_ai_b,n} = E\{\mathbf{v}_{i_a,n}\mathbf{v}_{i_b,n}^H\}$, $\mathbf{U}_{i_a,n} = E\{\mathbf{v}_{i_a,n}\mathbf{v}_{i_a,n}^T\}$ and $\mathbf{U}_{i_ai_b,n} = E\{\mathbf{v}_{i_a,n}\mathbf{v}_{i_b,n}^T\}$, for $a,b \in \{1,2,\ldots , M\}$.
\vspace{3mm}\\
\noindent
{\bf Diffusion step.} The so found local neighbourhood state estimates are followed by the diffusion (combination) step,
\begin{figure}[t]
\centering
\includegraphics[trim = 4mm 30mm 4mm 9mm,scale=0.8]{{Network_topology_General}.eps}
\caption{An example of a distributed network topology.}
\label{fig:NetwotkTopologyGeneral}
\end{figure}
\begin{eqnarray}
\mathbf{\widehat{x}}_{i,n|n} = \sum_{k \in \mathcal{N}_i} c_{k,i}\mathbf{\widehat{\underline{x}}}_{k,n|n} \label{eq:diff_step}
\end{eqnarray}
which calculates the diffused state estimates $\mathbf{\widehat{x}}_{i,n|n}$ as a weighted sum of the estimates from the neighbourhood $\mathcal{N}_i$, where $c_{k,i} \geq 0$ are the weighting coefficients satisfying $\sum_{k \in \mathcal{N}_i} c_{k,i} = 1$.
{The combination weights $c_{k,i}$ used by the diffusion step in \eqref{eq:diff_step} can follow the Metropolis \cite{Lopes_Sayed_TSP_Dist_LMS_2008}, Laplacian \cite{Bru_Convergence_1994} or the nearest neighbour \cite{Cattivelli_Sayed_IEEETranAC_Dist_KF_2010} rules. However, finding the set of optimal weights is a challenging task, though progress has been made in important cases\cite{Li_Optimal_Dist_fusion_2003, Sayed_Proceedings_2014, Sayed_arXiv}.}
The distributed complex Kalman filter (D-CKF) aims to approach the performance of a centralised Kalman filter (access to data from all the nodes) via neighbourhood collaborations and diffusion, and is summarised in Algorithm \ref{alg:D-CKF}. Each node within D-CKF forms a collective observation as in \eqref{c_neighbourhood_obs_eqn}, using information from its neighbours; thereafter, each node computes a neighbourhood state estimate which is again shared with neighbours in order to be used for the diffusion step.
\begin{algorithm}[t]
Initialisation: For each node $i=1,2,\ldots,N$
\begin{eqnarray}
\mathbf{\widehat{x}}_{i,0|0} &=& E\{\mathbf{x}_{0}\} \nonumber\\
\mathbf{M}_{i,0|0} &=& E\{(\mathbf{x}_{0}-E\{\mathbf{x}_{0}\})(\mathbf{x}_{0}-E\{\mathbf{x}_{0}\})^H\} \nonumber
\end{eqnarray}
For every time instant $n=1,2,\ldots$
$-$ Evaluate at each node $i=1,2,\ldots,N$
\begin{align}
\mathbf{\widehat{x}}_{i,n|n-1} &= \mathbf{F}_{n-1}\mathbf{\widehat{x}}_{i,n-1|n-1} \label{px_est}\\
%
\mathbf{M}_{i,n|n-1} &= \mathbf{F}_{n-1}\mathbf{M}_{i,n-1|n-1}\mathbf{F}^{H}_{n-1} + \mathbf{Q}_n \label{pMSE}\\
%
\mathbf{G}_{i,n} &= \mathbf{M}_{i,n|n-1}\mathbf{\underline{H}}^{H}_{i,n}\big(\mathbf{\underline{H}}_{i,n}\mathbf{M}_{i,n|n-1}\mathbf{\underline{H}}^{H}_{i,n} + \mathbf{\underline{R}}_{i,n}\big)^{-1} \label{gain}\\
%
\mathbf{\widehat{\underline{x}}}_{i,n|n} &= \mathbf{\widehat{x}}_{i,n|n-1} + \mathbf{G}_{i,n}\big(\mathbf{\underline{y}}_{i,n} - \mathbf{\underline{H}}_{i,n}\mathbf{\widehat{x}}_{i,n|n-1}\big)\label{x_est}\\
%
\mathbf{M}_{i,n|n} &= (\mathbf{I} - \mathbf{G}_{i,n}\mathbf{\underline{H}}_{i,n})\mathbf{M}_{i,n|n-1}\label{MSE}
\end{align}
$-$ For every node $i$, compute the diffusion update as
\begin{align}\label{Diff}
\mathbf{\widehat{x}}_{i,n|n} = \sum_{k \in \mathcal{N}_i} c_{k,i}\mathbf{\widehat{\underline{x}}}_{k,n|n}
\end{align}
\caption{The D-CKF}
\label{alg:D-CKF}
\end{algorithm}
\\ \vspace{1mm}
\textbf{Remark \#1:} The D-CKF algorithm\footnote{The matrices $\mathbf{M}_{i,n|n}$ and $\mathbf{M}_{i,n|n-1}$ do not represent the covariances of $\mathbf{\widehat{x}}_{i,n|n}$ and $\mathbf{\widehat{x}}_{i,n|n-1}$, as is the case for the standard Kalman filter operating on linear Gaussian systems. This is due to the use of the suboptimal diffusion step, which updates the state estimate and not the covariance matrix $\mathbf{M}_{i,n|n}$.} given in Algorithm \ref{alg:D-CKF} is a variant of \cite{Cattivelli_Sayed_IEEETranAC_Dist_KF_2010}. It employs the standard (strictly linear) state space model \eqref{linearstatespace}, and therefore does not cater for widely linear complex state space models or noncircular state and observation noises ($\mathbf{P}_{n} \neq \mathbf{0}$ and $\mathbf{U}_{i,n} \neq \mathbf{0}$ for $i=1,\ldots,N$).
{Unlike existing distributed complex Kalman filters, the proposed D-CKF algorithm in \eqref{px_est} -- \eqref{Diff} caters for the correlations between the neighbourhood observation noises. When no such correlations exits, is identical to Algorithm \ref{alg:D-CKF} in \cite{Cattivelli_Sayed_IEEETranAC_Dist_KF_2010}.}
\subsection{Distributed Augmented Complex Kalman Filter} \label{Sec:D-ACKF}
As stated in Remark \#1, current state space algorithms do not cater for widely linear state and observation models or for improper measurements, states, and state and observation noises. To this end, we next employ the widely linear model in \eqref{Eq:WL_model} to introduce the widely linear version of the standard, strictly linear, distributed state space model\footnote{Observe that the noise models can also be widely linear, in which case:\\ $\mathbf{w}_{n} =\mathbf{D}_n\acute{\mathbf{w}}_{n} + \mathbf{E}_n\acute{\mathbf{w}}^*_{n}\quad$ and $\quad\mathbf{v}_{n} =\mathbf{F}_n\acute{\mathbf{v}}_{n} + \mathbf{H}_n\acute{\mathbf{v}}^*_{n}$, where $\mathbf{D, E, F, H}$ are coefficient matrices and $\acute{\mathbf{w}}_{n}$ and $\acute{\mathbf{v}}_{n}$ are proper or improper noise models.} in \eqref{linearstatespace} (see also \cite{Dini_Class_WLKF_IEEE_TNNLS_2012}, \cite{Dini_Dist_ACKF_ASILOMAR_2012})
\begin{subequations}\label{WLstatespace}
\begin{eqnarray}
\mathbf{x}_{n} &=& \mathbf{F}_{n-1}\mathbf{x}_{n-1} + \mathbf{A}_{n-1}\mathbf{x}^*_{n-1} + \mathbf{w}_{n} \\
\mathbf{y}_{i,n} &=& \mathbf{H}_{i,n}\mathbf{x}_{n} + \mathbf{B}_{i,n}\mathbf{x}_{n}^* + \mathbf{v}_{i,n} \label{WLdm1}
\end{eqnarray}
\end{subequations}
The compact, augmented representation, of this model is
\begin{subequations}\label{AUGstatespace}
\begin{eqnarray}
\mathbf{x}^a_{n} &=& \mathbf{F}^a_{n-1}\mathbf{x}^a_{n-1} + \mathbf{w}^a_{n} \label{as1}\\
\mathbf{y}_{i,n}^a &=& \mathbf{H}_{i,n}^a\mathbf{x}_{n}^a + \mathbf{v}_{i,n}^a
\end{eqnarray}
\end{subequations}
where $\mathbf{x}^a_n = [\mathbf{x}^T_n, \mathbf{x}^H_n]^T$ and $\mathbf{y}^a_n = [\mathbf{y}^T_n, \mathbf{y}^H_n]^T$, while
\begin{eqnarray*}
\mathbf{F}^a_{n} = \begin{bmatrix} \mathbf{F}_{n} & \mathbf{A}_{n}\\ \mathbf{A}^*_{n} & \mathbf{F}^*_{n} \end{bmatrix}
\text{ and }
\mathbf{H}^a_{i,n} = \begin{bmatrix} \mathbf{H}_{i,n} & \mathbf{B}_{i,n}\\ \mathbf{B}^*_{i,n} & \mathbf{H}^*_{i,n} \end{bmatrix}
\end{eqnarray*}
For \textit{strictly linear systems}, $\mathbf{A}_{n} = \mathbf{0}$ and $\mathbf{B}_{i,n} = \mathbf{0}$, so that the widely linear (augmented) state space model degenerates into a strictly linear one. However, the augmented state space representation above is still preferred in order to account for the pseudocovariances which reflect the \textit{impropriety of the signals} (\textit{cf.} widely linear nature of systems).
The augmented covariance matrices of $\mathbf{w}^a_n = [\mathbf{x}^T_n, \mathbf{w}^H_n]^T$ and $\mathbf{v}^a_{i,n} = [\mathbf{v}^T_{i,n}, \mathbf{v}^H_{i,n}]^T$ are then given by
\begin{eqnarray}
\mathbf{Q}^a_n &=& E\{\mathbf{w}^a_{n}\mathbf{w}^{aH}_{n}\} =\begin{bmatrix} \mathbf{Q}_n & \mathbf{P}_n\\ \mathbf{P}^*_n & \mathbf{Q}^*_n \end{bmatrix} \label{aQ}\\
\mathbf{R}^a_{i,n} &=& E\{\mathbf{v}^a_{i,n}\mathbf{v}^{aH}_{i,n}\} =\begin{bmatrix} \mathbf{R}_{i,n} & \mathbf{U}_{i,n}\\ \mathbf{U}^*_{i,n} & \mathbf{R}^*_{i,n} \end{bmatrix}\label{aR}
\end{eqnarray}
\noindent
{\bf Neighbourhood variables.} For collaborative estimation of the state within distributed networks, neighbourhood observation equations use all available neighbourhood observation data, to give
\begin{eqnarray} \label{ac_neighbourhood_obs_eqn}
\mathbf{\underline{y}}_{i,n} = \mathbf{\underline{H}}_{i,n}\mathbf{x}_{n} + \mathbf{\underline{B}}_{i,n}\mathbf{x}_{n}^* +\mathbf{\underline{v}}_{i,n}
\end{eqnarray}
where the symbol $\mathbf{\underline{B}}_{i,n} = \big[\mathbf{B}_{i_1,n}^T, \mathbf{B}_{i_2,n}^T, \ldots, \mathbf{B}_{i_M,n}^T \big]^T$ denotes the conjugate state matrix , and $\{i_1, i_2, \ldots, i_M\} \in \mathcal{N}_i$. The augmented neighbourhood observation equations now become
\begin{eqnarray}\label{aug_y_collective}
\mathbf{\underline{y}}_{i,n}^a = \mathbf{\underline{H}}_{i,n}^a\mathbf{x}_{n}^a + \mathbf{\underline{v}}_{i,n}^a
\end{eqnarray}
with the augmented neighbourhood variables defined as
\begin{eqnarray}
\mathbf{\underline{y}}_{i,n}^a = \begin{bmatrix} \mathbf{\underline{y}}_{i,n} \\ \mathbf{\underline{y}}_{i,n}^* \end{bmatrix},
\quad
\mathbf{\underline{H}}_{i,n}^a = \begin{bmatrix} \mathbf{\underline{H}}_{i,n} & \mathbf{\underline{B}}_{i,n}\\
\mathbf{\underline{B}}_{i,n}^* & \mathbf{\underline{H}}_{i,n}^*\end{bmatrix},
\quad
\mathbf{\underline{v}}_{i,n}^a = \begin{bmatrix} \mathbf{\underline{v}}_{i,n} \\ \mathbf{\underline{v}}_{i,n}^* \end{bmatrix}
\end{eqnarray}
Consequently, the covariance of the augmented neigbourhood observation noise $\mathbf{\underline{v}}_{i,n}^a$ takes the form
\begin{eqnarray}\label{aug_R_collective}
\mathbf{\underline{R}}_{i,n}^a
= E\{\mathbf{\underline{v}}^a_{i,n}\mathbf{\underline{v}}^{aH}_{i,n}\}
= \begin{bmatrix} \mathbf{\underline{R}}_{i,n} & \mathbf{\underline{U}}_{i,n}\\ \mathbf{\underline{U}}^*_{i,n} & \mathbf{\underline{R}}^*_{i,n} \end{bmatrix}
\end{eqnarray}
{\textbf{Remark \#2:} Observe that \eqref{aug_R_collective} caters for both the covariances $E\{\mathbf{v}_{i,n}\mathbf{v}_{i,n}^H\}$ and cross-correlations $E\{\mathbf{v}_{i,n}\mathbf{v}_{k,n}^H\}$, $i\neq k$ between the nodal observation noises. This is achieved through the covariance matrix $\mathbf{\underline{R}}_{i,n}$ and the pseudocovariances $E\{\mathbf{v}_{i,n}\mathbf{v}_{i,n}^T\}$, while the cross-pseudocorrelations $E\{\mathbf{v}_{i,n}\mathbf{v}_{k,n}^T\}$ are accounted for through the pseudocovariance matrix $\mathbf{\underline{U}}_{i,n}$.}
Finally, the augmented diffused state estimate becomes
\begin{eqnarray}
\mathbf{\widehat{x}}_{i,n|n}^a = \sum_{k \in \mathcal{N}_i} c_{k,i}\mathbf{\widehat{\underline{x}}}_{k,n|n}^a
\end{eqnarray}
and represents a weighted average of the augmented (neighbourhood) state estimates. The proposed distributed augmented complex Kalman filter (D-ACKF), based on the widely linear state space model, is summarised in Algorithm \ref{alg:D-ACKF}.
For strictly linear systems ($\mathbf{A}_{n} = \mathbf{0}$ and $\mathbf{B}_{i,n} = \mathbf{0}$ for all $n$ and $i$) and in the presence of circular state and observation noises ($\mathbf{P}_{n} = \mathbf{0}$ and $\mathbf{U}_{i,n} = \mathbf{0}$ for all $n$ and $i$), the D-ACKF and D-CKF algorithms yield identical state estimates for all time instants $n$. However, the D-ACKF is more general than the D-CKF, since it also caters for the noncircular data and noise natures together with correlated state and observation noises.
\begin{algorithm}[t]
Initialisation: For each node $i=1,2,\ldots,N$
\begin{eqnarray*}
\mathbf{\widehat{x}}_{i,0|0}^a &=& \big[E\{\mathbf{x}_{0}\}^T, E\{\mathbf{x}_{0}\}^H\big]^T \\
\mathbf{M}_{i,0|0}^a &=& E\big\{(\mathbf{x}_{0}^a-\mathbf{\widehat{x}}_{i,0|0}^a)(\mathbf{x}_{0}^a-\mathbf{\widehat{x}}_{i,0|0}^a)^{aH}\big\}
\end{eqnarray*}
For every time instant $n=1,2,\ldots$
\hspace{0.15cm}$-$ Evaluate at each node $i=1,2,\ldots,N$
\begin{align}
\mathbf{\widehat{x}}_{i,n|n-1}^a &= \mathbf{F}_{n-1}^a\mathbf{\widehat{x}}_{i,n-1|n-1}^a \label{augpx_est}\\
%
\mathbf{M}_{i,n|n-1}^a &= \mathbf{F}_{n-1}^a\mathbf{M}_{i,n-1|n-1}^a\mathbf{F}^{aH}_{n-1} + \mathbf{Q}_n^a \label{augpMSE}\\
%
\mathbf{G}_{i,n}^a &= \mathbf{M}_{i,n|n-1}^a\mathbf{\underline{H}}^{aH}_{i,n}\big(\mathbf{\underline{H}}_{i,n}^a\mathbf{M}_{i,n|n-1}^a\mathbf{\underline{H}}^{aH}_{i,n} + \mathbf{\underline{R}}_{i,n}^a\big)^{-1} \label{aug_gain}\\
%
\mathbf{\widehat{\underline{x}}}_{i,n|n}^a &= \mathbf{\widehat{x}}_{i,n|n-1}^a + \mathbf{G}_{i,n}^a\big(\mathbf{\underline{y}}_{i,n}^a - \mathbf{\underline{H}}_{i,n}^a\mathbf{\widehat{x}}_{i,n|n-1}^a\big)\label{augx_est}\\
%
\mathbf{M}_{i,n|n}^a &= (\mathbf{I} - \mathbf{G}_{i,n}^a\mathbf{\underline{H}}_{i,n}^a)\mathbf{M}_{i,n|n-1}^a\label{augMSE}
\end{align}
\hspace{0.15cm}$-$ For every node $i$, compute the diffusion update as
\begin{align}
\mathbf{\widehat{x}}_{i,n|n}^a = {\sum}_{k \in \mathcal{N}_i} c_{k,i}\mathbf{\widehat{\underline{x}}}_{k,n|n}^a \label{augdx}
\end{align}
\caption{The D-ACKF}
\label{alg:D-ACKF}
\end{algorithm}
\newline
\textbf{Remark \#3:} The information form of the D-ACKF, given in Algorithm \ref{alg:D-ACKF_info}, can be used to cater for the noncircularity of data when observation noises at different nodes are uncorrelated. Moreover, nodes in the distributed network can switch between the general D-ACKF in Algorithm \ref{alg:D-ACKF} and the information form of D-ACKF in Algorithm \ref{alg:D-ACKF_info}, depending on the correlation between the observation noises.
\begin{algorithm}[t]
Initialisation: For each node $i=1,2,\ldots,N$
\begin{eqnarray*}
\mathbf{\widehat{x}}_{i,0|0}^a &=& \big[E\{\mathbf{x}_{0}\}^T, E\{\mathbf{x}_{0}\}^H\big]^T \\
\mathbf{M}_{i,0|0}^a &=& E\big\{(\mathbf{x}_{0}^a-\mathbf{\widehat{x}}_{i,0|0}^a)(\mathbf{x}_{0}^a-\mathbf{\widehat{x}}_{i,0|0}^a)^{aH}\big\}
\end{eqnarray*}
For every time instant $n=1,2,\ldots$
\hspace{0.15cm}$-$ Evaluate at each node $i=1,2,\ldots,N$
\begin{align}
\mathbf{\widehat{x}}_{i,n|n-1}^a &= \mathbf{F}_{n-1}^a\mathbf{\widehat{x}}_{i,n-1|n-1}^a \label{augpx_est_info}\\
%
\mathbf{M}_{i,n|n-1}^a &= \mathbf{F}_{n-1}^a\mathbf{M}_{i,n-1|n-1}^a\mathbf{F}^{aH}_{n-1} + \mathbf{Q}_n^a \label{augpMSE_info}\\
%
\mathbf{S}_{i,n}^a &= {\sum}_{k \in \mathcal{N}_i} \mathbf{H}^{aH}_{k,n}(\mathbf{R}_{k,n}^a)^{-1}\mathbf{H}_{k,n}^a \\
%
\mathbf{r}_{i,n}^a &= {\sum}_{k \in \mathcal{N}_i} \mathbf{H}^{aH}_{k,n}(\mathbf{R}_{k,n}^a)^{-1}\mathbf{y}_{k,n}^a \\
%
(\mathbf{M}_{i,n|n}^a)^{-1} &= (\mathbf{M}_{i,n|n-1}^a)^{-1} + \mathbf{S}_{i,n}^a\\
\mathbf{\widehat{\chi}}_{i,n|n}^a &= \mathbf{\widehat{x}}_{i,n|n-1}^a + \mathbf{M}_{i,n|n}^a\big(\mathbf{r}_{i,n}^a - \mathbf{S}_{i,n}^a\mathbf{\widehat{x}}_{i,n|n-1}^a\big)
%
\end{align}
\hspace{0.15cm}$-$ For every node $i$, compute the diffusion update as
\begin{align}
\mathbf{\widehat{x}}_{i,n|n}^a = {\sum}_{k \in \mathcal{N}_i} c_{k,i}\mathbf{\widehat{\chi}}_{i,n|n}^a \label{augdx_info}
\end{align}
\caption{The D-ACKF Information Form}
\label{alg:D-ACKF_info}
\end{algorithm}
\subsection{Bias Analysis of the D-ACKF Estimates}
Consider the following augmented complex variables: the local (non-diffused) error at node $i \in [1,N]$ given by $\mathbf{\underline{e}}_{i,n|n}^a = \mathbf{x}_{n}^a - \mathbf{\widehat{\underline{x}}}_{i,n|n}^a$, the prediction error $\mathbf{e}_{i,n|n-1}^a = \mathbf{x}_{n}^a - \mathbf{\widehat{x}}_{i,n|n-1}^a$, and the diffused error $\mathbf{e}_{i,n|n}^a = \mathbf{x}_{n}^a - \mathbf{\widehat{x}}_{i,n|n}^a$.
Then%
\begin{align}\label{MeanaugDiffError2}
E\{\mathbf{e}_{i,n|n}^a\} &= \sum_{k \in \mathcal{N}_i} c_{k,i}
\mathbf{M}_{k,n|n}^a(\mathbf{M}_{k,n|n-1}^a)^{-1}\mathbf{F}_{n-1}^aE\{\mathbf{e}_{k,n-1|n-1}^a\} \nonumber\\
&= \mathbf{0}
\end{align}
\noindent
\textbf{Remark \#4:} The expression \eqref{MeanaugDiffError2} shows that the D-ACKF is an unbiased estimator of both proper and improper complex random signals.
\section{Distributed Augmented Complex Extended Kalman Filter} \label{Sec:D-ACEKF}
We next introduce the distributed augmented complex extended Kalman filter (D-ACEKF) for nonlinear state space models of the form
\begin{subequations}\label{nonlinearSS}
\begin{eqnarray}
\mathbf{x}_n &=& \mathbf{f}[\mathbf{x}_{n-1}] + \mathbf{w}_n \label{nsm1} \\
\mathbf{y}_{i,n} &=& \mathbf{h}_i[\mathbf{x}_n] + \mathbf{v}_{i,n} \label{nom1}
\end{eqnarray}
\end{subequations}
where the nonlinear functions $\mathbf{f}[\cdot]$ and $\mathbf{h}_i[\cdot]$ are respectively the (possibly time varying) process model and observation model at node $i$, the remaining variables are as defined above. Within the extended Kalman filter (EKF) framework, the nonlinear state and observation functions are approximated by their first order Taylor series expansions (TSE) about the state estimates $\mathbf{\widehat{x}}_{i,n-1|n-1}$ and $\mathbf{\widehat{x}}_{i,n|n-1}$ for each node $i$, so that {\cite{Dini_ACEKF_UDRC_2011}}
\begin{subequations}\label{linearisedSS}
\begin{align}
\mathbf{x}_n &\approx \mathbf{F}_{i,n-1}\mathbf{x}_{n-1} + \mathbf{A}_{i,n-1}\mathbf{x}^*_{n-1} + \mathbf{r}_{i,n-1} + \mathbf{w}_n \label{eqn-TSEstate}\\
\mathbf{y}_{i,n} &\approx \mathbf{H}_{i,n}\mathbf{x}_{n} + \mathbf{B}_{i,n}\mathbf{x}^*_{n} + \mathbf{z}_{i,n} + \mathbf{v}_{i,n} \label{eqn-TSEobs}
\end{align}
\end{subequations}
where the Jacobians of functions $\mathbf{f}[\cdot]$ and $\mathbf{h}_i[\cdot]$ are defined as
\begin{align}
\mathbf{F}_{i,n} &= \frac{\partial \mathbf{f}[\mathbf{x}]}{\partial\mathbf{x}}\Big|_{\mathbf{x} = \mathbf{\widehat{x}}_{i,n|n}}
\hspace{0.1cm}\text{,}\hspace{0.3cm}\qquad
\mathbf{A}_{i,n} = \frac{\partial \mathbf{f}[\mathbf{x}]}{\partial\mathbf{x}^*}\Big|_{\mathbf{x}^* = \mathbf{\widehat{x}}^*_{i,n|n}}
\text{,}\nonumber\\
\mathbf{H}_{i,n} &= \frac{\partial \mathbf{h}_i[\mathbf{x}]}{\partial\mathbf{x}}\Big|_{\mathbf{x} = \mathbf{\widehat{x}}_{i,n|n-1}}
\hspace{0.1cm}\text{and}\hspace{0.2cm}
\mathbf{B}_{i,n} = \frac{\partial \mathbf{h}_i[\mathbf{x}]}{\partial\mathbf{x}^*}\Big|_{\mathbf{x}^* = \mathbf{\widehat{x}}^*_{i,n|n-1}} \nonumber
\end{align}
and the vectors
\begin{align*}
\mathbf{r}_{i,n} &= \mathbf{f}[\mathbf{\widehat{x}}_{i,n-1|n-1}] - \mathbf{F}_{i,n-1}\mathbf{\widehat{x}}_{i,n-1|n-1} - \mathbf{A}_{i,n-1}\mathbf{\widehat{x}}^*_{i,n-1|n-1} \\
\mathbf{z}_{i,n} &= \mathbf{h}_i[\mathbf{\widehat{x}}_{i,n|n-1}] - \mathbf{H}_{i,n}\mathbf{\widehat{x}}_{i,n|n-1} - \mathbf{B}_{i,n}\mathbf{\widehat{x}}^*_{i,n|n-1}
\end{align*}
are deterministic inputs calculated from the state space model and state estimate. In order to cater for the full second order statistics of the variables in the linearised state space in \eqref{linearisedSS}, we shall consider its compact (augmented) version given by
\begin{subequations}\label{augLinearisedSS}
\begin{align}
\mathbf{x}^a_n &\approx \mathbf{F}^a_{i,n-1}\mathbf{x}^a_{n-1} + \mathbf{r}^a_{i,n-1} + \mathbf{w}^a_n \\
\mathbf{y}^a_{i,n} &\approx \mathbf{H}^a_{i,n}\mathbf{x}^a_{n} + \mathbf{z}^a_{i,n} + \mathbf{v}^a_{i,n}
\end{align}
\end{subequations}
where $\mathbf{r}^a_{i,n} = \big[ \mathbf{r}^T_{i,n} , \mathbf{r}^H_{i,n} \big]^T$, $\mathbf{z}^a_{i,n} = \big[ \mathbf{z}^T_{i,n} , \mathbf{z}^H_{i,n} \big]^T$, while\\\\
$\mathbf{F}^a_{i,n} = \begin{bmatrix} \mathbf{F}_{i,n} & \mathbf{A}_{i,n}\\ \mathbf{A}^*_{i,n} & \mathbf{F}^*_{i,n} \end{bmatrix}$ \quad and \quad
$\mathbf{H}^a_{i,n} = \begin{bmatrix} \mathbf{H}_{i,n} & \mathbf{B}_{i,n}\\ \mathbf{B}^*_{i,n} & \mathbf{H}^*_{i,n} \end{bmatrix}$.\\\\
Observe that the collective neighbourhood augmented observation equation for node $i$ takes the form
\begin{eqnarray} \label{neighbourhood_obs_eqn_NL}
\mathbf{\underline{y}}^a_{i,n} = \mathbf{\underline{h}}_i^a[\mathbf{x}_{n}] + \mathbf{\underline{v}}^a_{i,n}
\end{eqnarray}
while the collective observation function defined as
\begin{eqnarray*}
\mathbf{\underline{h}}_i^a[\mathbf{x}_{n}] &=& \Big[\mathbf{\underline{h}}^T_i[\mathbf{x}_{n}], \mathbf{\underline{h}}^H_i[\mathbf{x}_{n}]\Big]^T \\
\mathbf{\underline{h}}_i[\mathbf{x}_{n}] &=& \Big[\mathbf{h}^T_{i_1}[\mathbf{x}_{n}], \mathbf{h}^T_{i_2}[\mathbf{x}_{n}], \ldots,
\mathbf{h}^T_{i_M}[\mathbf{x}_{n}]\Big]^T
\end{eqnarray*}
where $i \in \{i_1, i_2, \ldots, i_M\}$ are all the nodes in the neighbourhood $\mathcal{N}_i$. The first order approximation of \eqref{neighbourhood_obs_eqn_NL} is then
\begin{align}
\mathbf{\underline{y}}^a_{i,n} &\approx \mathbf{\underline{H}}^a_{i,n}\mathbf{x}^a_{n} + \mathbf{\underline{z}}^a_{i,n} + \mathbf{\underline{v}}^a_{i,n}
\end{align}
with the Jacobian of the collective observation function
\begin{eqnarray*}
\mathbf{\underline{H}}^a_{i,n} = \begin{bmatrix} \mathbf{\underline{H}}_{i,n} & \mathbf{\underline{B}}_{i,n}\\ \mathbf{\underline{B}}^*_{i,n} & \mathbf{\underline{H}}^*_{i,n} \end{bmatrix}
\end{eqnarray*}
where $\mathbf{\underline{H}}_{i,n} = \big[\mathbf{H}_{i_1,n}^T, \mathbf{H}_{i_2,n}^T, \ldots, \mathbf{H}_{i_M,n}^T \big]^T$ and
$\mathbf{\underline{B}}_{i,n} = \big[\mathbf{B}_{i_1,n}^T, \mathbf{B}_{i_2,n}^T, \ldots, \mathbf{B}_{i_M,n}^T \big]^T$, wherein
\begin{align}
\mathbf{H}_{i_k,n} &= \frac{\partial \mathbf{h}_{i_k}[\mathbf{x}]}{\partial\mathbf{x}}\Big|_{\mathbf{x} = \mathbf{\widehat{x}}_{i,n|n-1}}
\hspace{0.1cm}\text{and}\hspace{0.2cm}
\mathbf{B}_{i_k,n} = \frac{\partial \mathbf{h}_{i_k}[\mathbf{x}]}{\partial\mathbf{x}^*}\Big|_{\mathbf{x}^* = \mathbf{\widehat{x}}^*_{i,n|n-1}} \nonumber
\end{align}
\begin{algorithm}[t]
Initialisation: For each node $i=1,2,\ldots,N$
\begin{eqnarray*}
\mathbf{\widehat{x}}_{i,0|0}^a &=& \big[E\{\mathbf{x}_{0}\}^T, E\{\mathbf{x}_{0}\}^H\big]^T \\
\mathbf{M}_{i,0|0}^a &=& E\big\{(\mathbf{x}_{0}^a-\mathbf{\widehat{x}}_{i,0|0}^a)(\mathbf{x}_{0}^a-\mathbf{\widehat{x}}_{i,0|0}^a)^{aH}\big\}
\end{eqnarray*}
For every time instant $n=1,2,\ldots$
\hspace{0.15cm}$-$ Evaluate at each node $i=1,2,\ldots,N$
\begin{align}
\mathbf{\widehat{x}}_{i,n|n-1}^a &= \Big[\mathbf{f}^T[\mathbf{\widehat{x}}_{i,n-1|n-1}], \mathbf{f}^H[\mathbf{\widehat{x}}_{i,n-1|n-1}]\Big]^T \label{augpex_est}\\
%
\mathbf{M}_{i,n|n-1}^a &= \mathbf{F}_{i,n-1}^a\mathbf{M}_{i,n-1|n-1}^a\mathbf{F}^{aH}_{i,n-1} + \mathbf{Q}_n^a \label{augpeMSE}\\
%
\mathbf{G}_{i,n}^a &= \mathbf{M}_{i,n|n-1}^a\mathbf{\underline{H}}^{aH}_{i,n}\big(\mathbf{\underline{H}}_{i,n}^a\mathbf{M}_{i,n|n-1}^a\mathbf{\underline{H}}^{aH}_{i,n} + \mathbf{\underline{R}}_{i,n}^a\big)^{-1} \label{aug_egain}\\
%
\mathbf{\widehat{\underline{x}}}_{i,n|n}^a &= \mathbf{\widehat{x}}_{i,n|n-1}^a + \mathbf{G}_{i,n}^a\big(\mathbf{\underline{y}}_{i,n}^a - \mathbf{\underline{h}}_{i}^a[\mathbf{\widehat{x}}_{i,n|n-1}]\big)\label{augex_est}\\
%
\mathbf{M}_{i,n|n}^a &= (\mathbf{I} - \mathbf{G}_{i,n}^a\mathbf{\underline{H}}_{i,n}^a)\mathbf{M}_{i,n|n-1}^a\label{augeMSE}
\end{align}
\hspace{0.15cm}$-$ For every node $i$, compute the diffusion update as
\begin{align}
\mathbf{\widehat{x}}_{i,n|n}^a = \sum_{k \in \mathcal{N}_i} c_{k,i}\mathbf{\widehat{\underline{x}}}_{k,n|n}^a \label{augdex}
\end{align}
\caption{Distributed Augmented Complex EKF}
\label{alg:D-ACEKF}
\end{algorithm}
Algorithm \ref{alg:D-ACEKF} summarises the proposed distributed augmented complex extended Kalman filter, where each node $i$ shares its (nonlinear) observation model $\mathbf{h}_i[\cdot]$ with its neighbours.
\newline
{
{\bf Remark \#5:} The D-ACEKF algorithm in Algorithm \ref{alg:D-ACEKF} extends the Distributed Extended Kalman filter in \cite{Sayed_EKF_2010} by using the widely linear model, and caters for the second order statistical moments of the state and noise models, together with the correlations present between the nodal observation noises.}
\section{Distributed Widely Linear Frequency Estimation}
The proposed augmented state space models are particularly suited for frequency estimation in power grid, as due to system inertia, the frequency can be assumed identical over the network of measurement nodes, while unbalanced systems generate noncircular measurements \cite{ mandic2014patent, Xia_Adaptive_Frequency_Freq_2012_IEEE_SP_Mag}. {For a three phase system, the instantaneous voltages at a node $i$ are given by}
\begin{eqnarray}
{ v_{a,i,n} } & {=}& { V_{a,n}\cos(\omega nT + \phi) + z_{a,i,n} }\nonumber\\
{v_{b,i,n}} & {=}& {V_{b,n}\cos(\omega nT + \phi - 2\pi/3 + \Delta_b) + z_{b,i,n}} \nonumber\\
{v_{c,i,n}} & {=}& {V_{c,n}\cos(\omega nT + \phi + 2\pi/3 + \Delta_c) + z_{c,i,n} }
\label{Eq:3_phase_voltages}
\end{eqnarray}
where $V_{a,n}$, $V_{b,n}$ and $V_{c,n}$ are the amplitudes of the three-phase voltages at time instant $n$, $\omega = 2\pi f$ the angular frequency, $f$ the system frequency, $T$ the sampling interval, and $\phi$ the phase of the fundamental component, while $ z_{a,i,n}$, $ z_{a,i,n}$ and $ z_{a,i,n}$ are zero-mean observation noise processes. {The terms $\Delta_b, \Delta_c$ are used to indicate the phase distortions relative to a balanced three-phase system.} The phase voltages in ({\ref{Eq:3_phase_voltages}}) are first mapped to the complex voltage of orthogonal $\alpha$ and $\beta$ components using Clarke's $\alpha \beta$ transformation, to give {
\begin{eqnarray}
\begin{bmatrix} v_{\alpha,i,n} \\ v_{\beta,i,n}\end{bmatrix} = \sqrt{\frac{2}{3}}\begin{bmatrix} 1 & -\frac{1}{2} & -\frac{1}{2} \\ 0 & \frac{\sqrt{3}}{2} & -\frac{\sqrt{3}}{2} \end{bmatrix} \begin{bmatrix} v_{a,i,n} \\ v_{b,i,n} \\ v_{c,i,n}\end{bmatrix}.
\label{Eq:Clarke_transform}
\end{eqnarray}
}
The $\alpha \beta$ voltage is then converted to a scalar complex signal $v_{i,n}=v_{\alpha,i,n} + jv_{\beta,i,n}$. For balanced systems, in which $V_{a,n}=V_{b,n}=V_{c,n}$, { and $\Delta_b = \Delta_c = 0$}, the variables
\begin{eqnarray*}
v_{\alpha,i,n}&=& A_n\cos(\omega nT + \phi) + z_{\alpha,i,n} \\
v_{\beta,i,n} &=& A_n\cos(\omega nT + \phi + \frac{\pi}{2}) + z_{\beta,i,n}
\end{eqnarray*}
where $A_n = \frac{\sqrt{6}}{2}V_{a,n}$. The noises mapped via the $\alpha \beta$ transform now become
\begin{eqnarray*}
z_{\alpha,i,n} &=& \sqrt{2/3}\Big(z_{a,i,n} - \frac{1}{2}z_{b,i,n} - \frac{1}{2}z_{c,i,n}\Big) \\
z_{\beta,i,n} &=& \sqrt{2/3}\Big( \frac{\sqrt{3}}{2}z_{b,i,n} - \frac{\sqrt{3}}{2}z_{c,i,n}\Big)
\end{eqnarray*}
For balanced systems, this scalar complex model takes a recursive form, so that for every node\footnote{The usual assumption in this type of estimation is that for a sampling frequency $>> 50$Hz, we have $A_n\approx A_{n-1}$.}
\begin{eqnarray}
v_{i,n} &=& v_{\alpha,i,n} + jv_{\beta,i,n} \nonumber\\
&=& A_ne^{j(\omega nT + \phi)} + z_{i,n} \nonumber\\
&=& v_{i,n-1}e^{j\omega T} + z_{i,n}
\label{Eq:output_voltage2}
\end{eqnarray}
where $z_{i,n} = z_{\alpha,i,n} + j \, z_{\beta,i,n}$.
The state space model for this system at a node $i$ is shown in \eqref{Eq:ss_state4} and \eqref{Eq:ss_obs4}, where the state variables $x_{k}$ and $u_n$ are used to estimate the exponential $e^{j\omega T}$ and the observation $v_{i,n}$ respectively, while $\mathbf{w}_n$ and $z_{i,n}$ are the state and observation noises respectively.
The system frequency is then derived from the state variable $x$ as follows:
\begin{eqnarray}
\hat{f}_n &=& \frac{1}{2\pi T}\arcsin\big(\Im(x_n)\big)
\label{Eq:freq_estimate1}
\end{eqnarray}
where $\Im(\cdot)$ is the imaginary part of a complex quantity.
\begin{algorithm}[t]
\caption{A Strictly Linear State Space (StSp-SL)}
State equation:
\begin{subequations}\label{alg:SS1_L}
\begin{align}
\begin{bmatrix}x_n \\ u_n\end{bmatrix}
= \begin{bmatrix}x_{n-1} \\ u_{n-1}x_n \end{bmatrix} + \mathbf{w}_{n-1}
\label{Eq:ss_state4}
\end{align}
Observation equation:
\begin{align}
v_{i,n} = \begin{bmatrix} 0 & 1 \end{bmatrix} \begin{bmatrix}x_n \\ u_n\end{bmatrix} + z_{i,n}
\label{Eq:ss_obs4}
\end{align}
\end{subequations}
\end{algorithm}
Figure \ref{fig:AB_trajectory} illustrates the trajectory of the transformed voltage (a rotating vector -- phasor), indicating that for a balanced system, Clarke's voltage $v_{i,n}$ in \eqref{Eq:output_voltage2} has a circular trajectory.
However, the model in \eqref{Eq:ss_state4} and \eqref{Eq:ss_obs4} becomes inaccurate when the system is operating under unbalanced conditions, in which case, the voltage amplitudes $V_{a,n}$, $V_{b,n}$ and $V_{c,n}$ are no longer equal or if the condition $\Delta_b = \Delta_c = 0$ is not satisfied, and the system trajectory becomes noncircular (ellipse in Figure \ref{fig:AB_trajectory}). For unbalanced systems, therefore, the correct system model is widely linear, and is given by
{
\begin{align} \nonumber
v_{i,n} = & {} \sqrt{\frac{2}{3}} \begin{bmatrix} 1 & j \end{bmatrix}
\begin{bmatrix} 1 & -\frac{1}{2} & -\frac{1}{2} \\ 0 & \frac{\sqrt{3}}{2} & -\frac{\sqrt{3}}{2} \end{bmatrix} \begin{bmatrix} v_{a,i,n} \\ v_{b,i,n} \\ v_{c,i,n}\end{bmatrix} \\ \label{eq:cmplx_voltage}
= & {} \sqrt{\frac{2}{3}} \begin{bmatrix} 1 & e^{j2\pi/3} & e^{-j2\pi/3} \end{bmatrix}
\begin{bmatrix} v_{a,i,n} \\ v_{b,i,n} \\ v_{c,i,n}\end{bmatrix}.
\end{align}
For the compactness of notation, we set $\theta_n = \omega n T + \phi$. Using the relationship $\cos(y) = \frac{e^{jy} + e^{-jy}}{2}$, the three phase voltages at each node $i$ can be rewritten as
\begin{align} \nonumber
\begin{bmatrix} v_{a,i,n} \\ v_{b,i,n} \\ v_{c,i,n} \end{bmatrix} = {} &
\frac{1}{2}
\begin{bmatrix}
V_{a,n} \left( e^{j\theta_n} +e^{-j\theta_n} \right)
\\
V_{b,n} \left(e^{j(\theta_n - 2\pi/3 + \Delta_b) } + e^{-j(\theta_n - 2\pi/3 + \Delta_b) } \right)
\\
V_{c, n} \left( e^{j(\theta_n + 2\pi/3 + \Delta_c) } + e^{-j(\theta_n + 2\pi/3 + \Delta_c) } \right )
\end{bmatrix}
\\
& {}
+ \begin{bmatrix} z_{a,i,n} \\ z_{b,i,n} \\ z_{c,i,n} \end{bmatrix} \label{eq:polarform}
\end{align}
Substituting \eqref{eq:polarform} into \eqref{eq:cmplx_voltage} gives
\begin{align}
v_{i,n}
= {} & A_{n}e^{j\theta_n} + B_{n}e^{-j\theta_n} + z_{i,n}
\label{Eq:output_voltageWL1}
\end{align}
where
\begin{align}
\label{Eq:A_B_coefficents}
A_n = {} & \frac{\sqrt{6}}{6}\left(V_{a,n} + V_{b,n}e^{j\Delta_b} + V_{c,n}e^{j \Delta_c} \right)\\ \nonumber
B_n = {} & \frac{\sqrt{6}}{6} \left(V_{a,n} + V_{b,n}e^{-j(\Delta_b + 2\pi/3)} + V_{c,n}e^{-j( \Delta_c - 2 \pi/3)} \right).
\end{align}
}
\begin{figure}[t]
\centering
\subfloat{\includegraphics[scale=0.7]{Voltage_Trajectory_SNR_20}} \vspace{-0.35cm}
\caption{Noncircularity in power systems. For a balanced system, characterised by $V_{a,n}=V_{b,n}=V_{c,n}$ and $\Delta_b = \Delta_c = 0$, the trajectory of Clarke's voltage $v_{i,n}$ is circular (blue line). For unbalanced systems, the voltage trajectory is noncircular (red line), such as in the case of a 100\% single-phase voltage sag illustrated by the ellipse. In both cases the signal to noise ratio (SNR) was $20$dB.}
\label{fig:AB_trajectory}
\end{figure}
For a balanced system under nominal conditions, we have $V_{a,n}=V_{b,n}=V_{c,n}$ and $\Delta_b = \Delta_c = 0$, and the coefficient $B_n$ in ({\ref{Eq:A_B_coefficents}}) vanishes, so that the system is adequately characterised by the strictly linear model in \eqref{Eq:output_voltage2}. For unbalanced systems, $B_n\neq0$ and the Clarke's voltage $v_{i,n}$ is noncircular, so that the expression in \eqref{Eq:output_voltageWL1} is second order optimal for both balanced and unbalanced conditions. This expression can be written in the form of a widely linear recursive mode
\begin{eqnarray}
v_{i,n} &=& v_{n-1}h_{n-1} + v^*_{n-1}g_{n-1} + z_{i,n}
\label{Eq:output_voltageWL2}
\end{eqnarray}
which is a first-order widely linear autoregressive model with coefficients $h_{k}$ and $g_{k}$. The widely linear (augmented) state space model corresponding to \eqref{Eq:output_voltageWL2} is defined in \eqref{Eq:ss_state3}, where the state vector consists of the strictly linear and conjugate weights $h_n$ and $g_n$, and the variable $u_n$ which corresponds to the noise-free widely linear observation $v_n$. The system frequency can be computed from the state variables $h_n$ and $g_n$ as follows:
\begin{eqnarray}
\hat{f}_n &=& \frac{1}{2\pi T}\arcsin\big(\Im(h_n + a_ng_n)\big)
\label{Eq:freq_estimate2}
\end{eqnarray}
\begin{eqnarray*}
a_n &=& \frac{-j\Im(h_n)+ j\sqrt{\Im^2(h_n) - |g_n|^2}}{g_n}
\label{Eq:a1}
\end{eqnarray*}
The state space model in \eqref{Eq:ss_state3} and \eqref{Eq:ss_obs3} provides a realistic and robust characterisation of real world power systems, as it represents both balanced and unbalanced systems, while its nonlinear state equation also models the coupling between state variables. The StSp-WL model in \eqref{alg:SS1_WL} can be implemented using the proposed distributed augmented complex extended Kalman filter in Section {\ref{Sec:D-ACEKF}}.
\begin{algorithm}[t]
\caption{A Widely Linear State Space (StSp-WL)}
State equation:\vspace{-0.1cm}
\begin{subequations}\label{alg:SS1_WL}
\begin{align}
\begin{bmatrix}h_n \\ g_n\\ u_n \\ h^*_n \\ g^*_n\\ u^*_n\end{bmatrix}
\!=\! \begin{bmatrix}h_{n-1} \\ g_{n-1}\\ u_{n-1}h_{n-1} + u^*_{n-1}g_{n-1}\\ h^*_{n-1} \\ g^*_{n-1}\\ \!u^*_{n-1}h^*_{n-1} \!+\! u_{n-1}g^*_{n-1}\!\end{bmatrix} \!+\! \mathbf{w}_{n-1}
\label{Eq:ss_state3}
\end{align}
Observation equation: \vspace{-0.1cm}
\begin{align}
\begin{bmatrix} v_{i,n}\\ v^*_{i,n} \end{bmatrix} = \begin{bmatrix} 0 & 0 & 1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 1\end{bmatrix} \begin{bmatrix}h_n \\ g_n\\ u_n \\ h^*_n \\ g^*_n\\ u^*_n\end{bmatrix} + \begin{bmatrix} z_{i,n}\\ z^*_{i,n} \end{bmatrix}
\label{Eq:ss_obs3}
\end{align}
\end{subequations}
\end{algorithm}
\section{Frequency Estimation Examples}
\begin{figure}[t]
\centering
\subfloat{ {\includegraphics[clip = true, trim = 25mm 40mm 270mm 15mm,scale=0.3]{Network_Power}}}
\caption{A distributed { power network} with $N=5$ nodes {(Sub-stations) }used in the simulations.}
\label{fig:NetwotkTopologyN5}
\end{figure}
\begin{figure}[t]
\centering
\subfloat{\includegraphics[scale=0.39]{Type_C_sag_Circularity}}
\subfloat{\includegraphics[scale=0.39]{Type_C_sag_Phasor}}\\ \vspace{-0.5cm}
\subfloat{\includegraphics[scale=0.39]{Type_D_sag_Circularity}}
\subfloat{\includegraphics[scale=0.39]{Type_D_sag_Phasor}}
\caption{Geometric (left) and phasor (right) views of Type C and D unbalanced voltage sags. The real-imaginary phasor plots illustrate the noncircularity of Clarke's voltage in unbalanced conditions, indicated by the elliptical shapes of circularity plots. The parameters of this ellipse (degree of noncircularity) serve to identify the type of fault (in this case a voltage sag).}
\label{fig:circularity_phasor_sags_C_D}
\end{figure}
\begin{figure*}[tbh]
\centering
\subfloat{\includegraphics[scale=0.7]{ACEKF_5nodes_SNR_40_typeC_then_typeD}}
\subfloat{\includegraphics[scale=0.7]{D-ACEKF_5nodes_SNR_40_typeC_then_typeD}}
\vspace{-0.35cm}
\caption{Frequency estimation performance of single node (CEKF and ACEKF) and distributed (D-CEKF and D-ACEKF) algorithms for a system at $40$dB SNR. The system is balanced up to $0.1$s, it then undergoes a Type C voltage sag followed by a Type D voltage sag at 0.3s.}
\label{fig:sag}\vspace{-0.5cm}
\end{figure*}
\begin{figure*}[tbh]
\centering
\subfloat{\includegraphics[scale=0.7]{ACEKF_5nodes_SNR_30}}
\subfloat{\includegraphics[scale=0.7]{D-ACEKF_5nodes_SNR_30}}
\vspace{-0.35cm}
\caption{Frequency estimation performance of single node (CEKF and ACEKF) and distributed (D-CEKF and D-ACEKF) algorithms for a balanced system in the presence of doubly white circular Gaussian noises at $30$dB SNR. As expected, the strictly and widely linear algorithms had similar performance.}
\label{fig:freq_SNR30}\vspace{-0.5cm}
\end{figure*}
\begin{figure*}[tbh]
\centering
\subfloat{\includegraphics[scale=0.7]{ACEKF_5nodes_Spike_noise}}
\subfloat{\includegraphics[scale=0.7]{D-ACEKF_5nodes_Spike_noise}}
\vspace{-0.35cm}
\caption{Frequency estimation performance of single node (CEKF and ACEKF) and distributed (D-CEKF and D-ACEKF) algorithms when the phase voltages at one of the nodes in the network are contaminated with random spike noise at $20$\% p.u.}
\label{fig:spikeNoise}\vspace{-0.5cm}
\end{figure*}
\begin{figure*}[tbh]
\centering
\subfloat{\includegraphics[scale=0.7]{ACEKF_5nodes_SNR_40_stepFreqChange}}
\subfloat{\includegraphics[scale=0.7]{D-ACEKF_5nodes_SNR_40_stepFreqChange}}
\vspace{-0.35cm}
\caption{Frequency estimation performance of single node (CEKF and ACEKF) and distributed (D-CEKF and D-ACEKF) algorithms for a power system at $40$dB SNR, which experiences a step change in system frequency to $51$Hz.}
\label{fig:stepChange}
\end{figure*}
\begin{figure*}[tbh]
\centering
\subfloat {\includegraphics[clip = true, trim =0mm 0mm 0mm 10mm, scale=0.2]{hilbert_typeD}}
\subfloat {\includegraphics[clip = true, trim =0mm 0mm 0mm 10mm, scale=0.2]{d_hilbert}}
\vspace{-0.35cm}
\caption{{Frequency estimation performance comparison between Hilbert Transform based instantaneous frequency estimate and ACEKF estimate. \textit{Left}: Single node based estimate (Hilbert and ACEKF). \textit{Right}: Distributed implementations of the algorithm (D-Hilbert and D-ACEKF) }}
\label{fig:hilbert}\vspace{-0.5cm}
\end{figure*}
The performance of the algorithms was evaluated over both comprehensive illustrative simulation studies and a real world example. Within the synthetic data, the power system under consideration had a nominal frequency of $50$Hz, and was sampled at a rate of $5$kHz while the state vectors of all the nodes in the network were initialised to $50.5$Hz. Without loss in generality, we used the distributed network topology shown in Figure \ref{fig:NetwotkTopologyN5}. The strictly linear state space model in \eqref{alg:SS1_L} and the widely linear state space model in \eqref{alg:SS1_WL} were implemented using the proposed widely linear D-ACEKF and its strictly linear version D-CEKF. For rigour, the uncooperative CEKF and ACEKF were also considered.
\noindent \textbf{Case Study \#1: Voltage sags.} In the first set of simulations, the performances of the algorithms were evaluated for an initially balanced system which became unbalanced after undergoing a Type C voltage sag starting at $0.1$s, characterised by a $20$\% voltage drop and $10^o$ phase offset on both the $v_b$ and $v_c$ channels, followed by a Type D sag starting at $0.3$s, characterised by a $20$\% voltage drop at line $v_a$ and a $10$\% voltage drop on both $v_b$ and $v_c$ with a $5^o$ phase angle offset. The degrees of noncircularity of these system imbalances are illustrated in Figure \ref{fig:circularity_phasor_sags_C_D}. Figure \ref{fig:sag} shows that, conforming with the analysis, the widely linear algorithms, ACEKF and D-ACEKF, were able to converge to the correct system frequency for both balanced and unbalanced operating conditions, while the strictly linear algorithms, CEKF and D-CEKF, were unable to accurately estimate the frequency during the voltage sag due to under-modeling of the system (not accounting for its widely linear nature) -- see \eqref{Eq:output_voltageWL2}. As expected, the widely linear and strictly linear algorithms had similar performances under balanced conditions, as illustrated in the time interval $0$-$0.1$s. The distributed algorithms, D-CEKF and D-ACEKF, outperformed their uncooperative counterparts, CEKF and ACEKF, owing to the sharing of information between neighbouring nodes.
\noindent \textbf{Case Study \#2: Presence of noise.} Figure \ref{fig:freq_SNR30} illustrates frequency estimation for a balanced system in the presence of white Gaussian noise at $30$dB SNR, while Figure \ref{fig:spikeNoise} illustrates frequency estimation in the presence of random spike noise, which typically occurs in the presence of switching devices or lightning. The distributed algorithms, D-CEKF and D-ACEKF, outperformed their uncooperative counterparts, CEKF and ACEKF, because neighbouring nodes were able to share information to facilitate better estimation performances. In both cases, the distributed algorithms exhibited lower variance in the frequency estimate.
\noindent \textbf{Case Study \#3: Frequency jumps.} Figure \ref{fig:stepChange} illustrates the performance of both single node and distributed algorithms (strictly and widely linear) when a power system is contaminated with white noise at $40$dB SNR and undergoes a step change in system frequency, a typical scenario when generation does not match the load (microgrids and islanding). Although all the algorithms had similar responses to the step change in frequency, the distributed algorithms exhibited enhanced frequency tracking.
\begin{figure*}[ht!]
\centering
\subfloat{\includegraphics[scale=0.7]{errorBias_typeDsag}}
\subfloat{\includegraphics[scale=0.7]{errorVariance_typeDsag}}
\vspace{-0.35cm}
\caption{Bias and variance analysis of the proposed distributed state space frequency estimators for an unbalanced system undergoing a type D voltage sag. {\it Left:} Estimation bias. {\it Right:} Estimation variance.}
\label{fig:BiasVar}
\end{figure*}
{
\noindent \textbf{Case Study \#4: Comparison with Hilbert method.} To further illustrate the advantages of widely linear modelling in a three-phase setting, the next case study compares the performance of the widely linear ACEKF to the Hilbert Transform based instantaneous frequency estimate from one of the three-phases. The three phase voltage signal at each node was simulated with circular white noise at $30$dB SNR and Type D imbalance after $0.25$s. Figure \ref{fig:hilbert} shows that the ``Hilbert" frequency estimate, found by differentiating the instantaneous phase angle of the single phase voltage that underwent the Hilbert Transform, was poor since the differentiation step (high pass filter) in the Hilbert method is not robust to noise. This was observed both in single-node and distributed settings.
}
\noindent \textbf{Case Study \#5: Bias and variance of the proposed estimators.} For rigour, Figure \ref{fig:BiasVar} provides the analysis of the bias and variance for the proposed distributed frequency estimators. The algorithms were evaluated at different SNR levels for an unbalanced system undergoing a Type D voltage sag (see also Figure {\ref{fig:circularity_phasor_sags_C_D}}). Observe that both the single- and multiple-node widely linear algorithms, ACEKF and D-ACEKF, were asymptotically unbiased (left panel, see Remark \#4) while both the single- and multiple-node strictly linear algorithms were biased. In terms of the variance of the estimators (right panel), both the distributed estimation algorithms outperformed their non-cooperative counterparts, while the only consistent estimator was the proposed distributed augmented complex extended Kalman filter.
{
\noindent \textbf{Real World Case Study.} We next assessed the performance of the proposed algorithms on a real world case study using three-phase voltage measurements from two adjacent sub-stations in Malaysia\footnote{ {Due to data provenance issues we are unable to reveal which particular sub-stations the measurements originate from.}} during a brief line-to-earth fault on the 29th June 2014. This caused voltage sags, similar to those in Case Study \#1. The three-phase measurements were sampled at 5kHz and the voltage values were normalized. The left panel in Figure \ref{fig:realWorld} shows the normalized $\alpha \beta$ voltages at one of the sub-stations. The fault that occurs in phase A around 0.1s is reflected in the voltage dip in $v_{\alpha}$. The right panel in Figure \ref{fig:realWorld} shows the frequency estimate from the D-ACEKF and D-CEKF. Conforming with the analysis and the single node scenario in Figure \ref{fig:sag}, the collaborative widely linear D-ACEKF was able to track the real world frequency of a power network under both balanced and unbalanced conditions, whereas the strictly linear D-CEKF was unable to track the frequency after 0.1s when the line-to-earth fault (non-circularity) occurred.
\begin{figure*}[tbh]
\centering
\subfloat {\includegraphics[clip = true, trim =0mm 0mm 0mm 10mm, scale=0.2]{VoltageSub}}
\subfloat {\includegraphics[clip = true, trim =0mm 0mm 0mm 10mm, scale=0.2]{ACLMS_RealWorld}}
\vspace{-0.35cm}
\caption{Real world case study. \textit{Left}: The $\alpha \beta$ voltages at Sub-station 1 before and during the fault event. \textit{Right}: Frequency estimation using the proposed algorithms before and during the fault event.}
\label{fig:realWorld}
\end{figure*}
}
\section{Conclusions}
We have proposed a novel class of diffusion based distributed complex valued Kalman filters for cooperative frequency estimation in power systems. To cater for the general case of improper states, observations, and state and observation noises, we have introduced the distributed (widely linear) augmented complex Kalman filter (D-ACKF) and its nonlinear version, the distributed augmented complex Kalman filter (D-ACEKF). These have been shown to provide sequential state estimation of the generality of complex signals, both circular and noncircular, within a general and unifying framework which also caters for correlated nodal observation noises.
This novel widely linear framework has been applied for distributed state space based frequency estimation in the context of three-phase power systems, and has been shown to be optimal for both balanced and unbalanced operating conditions. Simulations over a range of both balanced and unbalanced power system conditions {for both synthetic and real world measurements} have illustrated that the proposed distributed state space algorithms are consistent estimators, offering unbiased and minimum frequency estimation in both balanced an unbalanced system conditions, together with simultaneous frequency estimation and fault identification.
\bibliographystyle{ieeetr}
\balance
|
\section{Introduction}
The discovery of the celebrated Penrose tilings \cite{P}, see also
\cite{GS, BG}, and of physical quasicrystals \cite{SBGC} gave rise to
the development of a mathematical theory of aperiodic order. Objects
of study are nonperiodic structures (i.e. not fixed by any
nontrivial translation) that nevertheless possess a high degree of
local and global order. In many cases the structures under
consideration are either discrete point sets (Delone sets,
see for instance \cite{DLS}) or tilings (tilings are also called
tesselations, see \cite{GS} for a wealth of results about tilings).
\footnote{{\em UDC Classification} 514}
Three frequently used construction methods for nonperiodic tilings are
local matching rules, cut-and-project schemes and tile substitutions,
see \cite{BG} and references therein for all three methods. In this
paper we describe an additional way to construct nonperiodic
tilings, the ``inductive rotation'', found in 2008 by the second author
who is not a scientist but an artist.
Variants of this construction have been considered
before, see for instance \cite{FH} $\to$ ``People'' $\to$ ``Petra Gummelt'',
but up to the knowledge of the authors this construction does not appear
in the existing literature.
After describing the construction we will prove that the resulting
tilings are nonperiodic, aperiodic and limitperiodic, that they
can be described as model sets,
hence are pure point diffractive, that they possess uniform patch
frequencies and that they are substitution tilings. (For an
explanation of these terms see below). The latter property will be
the key for most of the other results. We end with some remarks and
open questions. In order to keep the
paper as much self-contained as possible we try to explain all
terms here, but only as far as required to state the
results. Where needed we provide references for further information.
Let us fix some notation. A {\em tiling} of $\ensuremath{\mathbb R}^2$ is a collection of
tiles $\{ Q_i \, | \, i \in \ensuremath{\mathbb N} \}$ that is a covering (i.e.
$\bigcup_{i \in \ensuremath{\mathbb N}} Q_i = \ensuremath{\mathbb R}^2$) as well as a packing (i.e. the
intersection of the interiors of two distinct tiles $Q_i$ and $Q_j$
is empty). The $Q_i$ are called {\em tiles}
of the tiling. For our purposes it is fine to think of the tiles as
nice compact sets like squares or triangles. If all tiles in the tiling
belong to finitely many congruence classes $[T_1], \ldots, [T_m]$ then
we call the representatives $T_1, \ldots , T_m$ {\em prototiles} of the
tiling. Any finite subset of a tiling
is called a {\em patch}. Examples of patches of a tiling are obtained by
intersecting a tiling $\ensuremath{{\mathcal T}}$ with a ball $B_r(x)$, i.e. with an open ball of
radius $r$ with centre $x$. The intersection is defined as
\[ \ensuremath{{\mathcal T}} \cap B_r(x) := \{ T \in \ensuremath{{\mathcal T}} \, | \, T \subset B_r(x) \}. \]
Sometimes we want to
equip the tiles with an additional attribute like colour or decoration.
Then we consider the prototiles equipped with the different colours
or decorations, too. One can formalise this by considering pairs $(T,i)$,
where $T$ is some tile and $i$ encodes the additional attribute. In
order to keep notation simple we keep in mind to distinguish a black
unit square from a red unit square when necessary without writing
down the additional label $i$.
A vector $t \in \ensuremath{\mathbb R}^2$ such that $\ensuremath{{\mathcal T}} +t = \ensuremath{{\mathcal T}}$ is called {\em period} of $\ensuremath{{\mathcal T}}$.
(If $\ensuremath{{\mathcal T}}=\{ Q_i \, | \, i \in \ensuremath{\mathbb N} \}$ then $\ensuremath{{\mathcal T}}+t$ means $\{Q_i+t \, | \,
i \in \ensuremath{\mathbb N} \}$.) A tiling $\ensuremath{{\mathcal T}}$ is called {\em periodic} if it has a
nontrivial period, i.e. a period $t \ne 0$.
A tiling $\ensuremath{{\mathcal T}}$ is called {\em 2-periodic} if $\ensuremath{{\mathcal T}}$ possesses two linear
independent periods. A tiling $\ensuremath{{\mathcal T}}$ is called {\em nonperiodic} if its
only period is the trivial period $t=0$.
A {\em substitution rule} $\sigma$ is a simple method to generate
nonperiodic tilings. The basic idea is to substitute each prototile
$T_i$ with a patch $\sigma(T_i)$ consisting of congruent copies of
some of the prototiles $T_1, \ldots, T_m$.
The Penrose tilings can be generated by a substitution rule with
two prototiles, see \cite{BG}, \cite{FH} or \cite{WIK}. A simpler
example using just one prototiles is shown
in Figure \ref{fig:subst-bsp}: a substitution rule for the {\em chair
tiling}. The chair tiling is nonperiodic, even though it contains large
2-periodic subsets \cite{BG}.
\begin{figure}
\includegraphics[width=100mm]{chair-subst.pdf}
\caption{The substitution rule for the aperiodic chair tiling (left)
and the first two iterates of this rule applied to a single
prototile. \label{fig:subst-bsp}}
\end{figure}
Given a substitution $\sigma$ with prototiles $T_1, \ldots, T_m$
a patch of the form $\sigma(T_i)$ is called {\em supertile}.
More generally, a patch of the form $\sigma^k(T_i)$ is called
{\em level $k$ supertile}.
A substitution rule is called {\em primitive} if there is $k \in \ensuremath{\mathbb N}$
such that each level $k$ supertile contains congruent copies of all prototiles.
In the sequel we present a construction method for nonperiodic tilings
that is similar but not equal to a substitution rule.
Before we give a more precise description let us first illustrate the
idea of the construction. We start with a square $G_0$ of edge length two
centred in the origin, see Figure \ref{fig:gen0-3} left. (We may imagine
this and further squares cut out of cardboard or similar.) In the next
step we remove $G_0$ (but keep its position in mind); we take four
translates of $G_0$, and we place the first one one unit to the left
with respect to $G_0$. The next square we rotate by
$\frac{-\pi}{2}$ and place it one unit up with respect to $G_0$,
shuffling it partly {\em below} the first square. The third square we
rotate by ${-\pi}$, place it one unit to the right with respect to $G_0$
shuffling it partly below the second square. The last square is rotated by
$\frac{- 3 \pi}{2}$, placed one unit down with respect to $G_0$ and shuffled
below the third square. The result is shown in Figure \ref{fig:gen0-3}
second from the left. Let us call this constellation of four overlapping
squares $G_1$.
\begin{figure}
\includegraphics[width=110mm]{nackt-r0-r3.pdf}
\caption{The first three iterates of the construction for non-decorated
tiles. The black dot indicates the origin. \label{fig:gen0-3}}
\end{figure}
We proceed in similar way: We take four translates of $G_1$ and put
one translate two units to the left with respect to $G_1$, a second
translate (rotated by $\frac{-\pi}{2}$) two units up with respect to
$G_1$ and below the first translate, a third translate (rotated by
$-\pi$) two units to the right with respect to $G_1$ and
below the second translate, and a fourth translate (rotated by
$\frac{-3 \pi}{2}$) two units down with respect to $G_1$ and below
the third translate. The result $G_2$, as seen from above,
is shown in Figure \ref{fig:gen0-3} second from the right.
The next generations are constructed analogously,
just replace $G_1$ by $G_n$ and 2 by $2^n$ in the last paragraph.
In this way we can cover arbitrary large parts of the plane. In order
to translate this covering into a tiling we consider only the visible (parts
of) squares. Thus the tiling has four prototiles: a large square
with edge length two, a small square with edge length one, a $1 \times 2$
rectangle, and a ``chair'', i.e. a non-convex hexagon that is the
union of three small squares. The more subtle question how this
sequence of finite patterns yields an infinite
tiling of the plane is answered below. Briefly, we use the fact that the
central patches of the $G_i$ are fixed under the iteration. These patches
yield a nested sequence $S_2 \subset S_3 \subset S_4 \cdots$ of patches
with a well defined limit which is an infinite tiling.
\section{The Tilings}
Now we give a more precise description of the construction of the
sequence of patches.
Let $\phi$ denote a rotation through $-\pi/2$ about the origin.
Start with a square $Q:=[-1,1]^2$. In the following we need
to keep track about which squares lie ``above'' (parts of) other
squares. In order to achieve this we may equip each square with an
additional label. One may write $(Q,i)$ and consider the label $i$
as the ``height'' of $Q$ in some orthogonal direction. Anyway, this idea
leads to a tedious description of the construction that we
omit here. For our purposes it is sufficient just to
keep track which squares are higher respectively lower than other squares.
Let $P_0:=\{ Q \}$. Let
\[ P_1: = \{ Q+(-1,0), \phi Q + (0,1), \phi^2 Q+(1,0), \phi^3 Q+(0,-1) \}, \]
where the first square is on top of the other squares, the second square
is below the first one but above the third and the fourth square,
and the fourth square is on bottom.
In the $n$th step let $P_n$ consist of four congruent copies of $P_{n-1}$:
\[ P_n : = \{ P_{n-1}+(-2^{n-1},0), \phi P_{n-1}+(0,2^{n-1}),
\phi^2 P_{n-1}+(2^{n-1},0), \phi^3 P_{n-1}+(0,-2^{n-1}) \}. \]
All squares in the first set are on top of the squares in the other three sets,
the squares in the second set all are below the squares in the first set
but on top of all squares in the third set and the fourth set, and the
squares in the fourth set are all below the squares in the other three sets.
Two squares within one of the sets $P_{n-1}+t$ inherit their above-below
relation from the preceding steps in the iteration.
Note that the centres of the squares are contained in the lattice translate
\[ (1,0) + \Lambda := (1,0)+ \langle (1,1), (1,-1) \rangle_{\ensuremath{\mathbb Z}} =
\{ (x_1, x_2) \in \ensuremath{\mathbb R}^2 \, | \, x_1+x_2 \mbox{ odd } \} \]
Hence, by the construction, the centre of each of the congruent copies
of $P_1$ (i.e. the single point that is the intersection of
the four squares in the copy of $P_1$) is contained in $(2,0) + 2 \Lambda=
(2,0)+\langle (2,2), (2,-2) \rangle_{\ensuremath{\mathbb Z}}$.
Since each $x \in (1,1)+\Lambda$ has a unique presentation of the form
\[ x=y + \left\{ \begin{array}{l} (0, \pm 1) \\ (\pm1, 0 ) \end{array}
\right. , \; (y \in (2,0)+2\Lambda) \]
the following lemma is immediate.
\begin{lem} \label{lem:210}
The set $P_n$ consists of $4^n$ congruent copies of $Q$ with centres
\[ \{ (x_1,x_2) \in \ensuremath{\mathbb Z}^2 \, | \, x_1+x_2 \; \mbox{\rm odd }, \; |x_1|+|x_2|
\le 2^n-1 \} \]
\end{lem}
Because of Lemma \ref{lem:210} almost any $(x_1,x_2) \in \ensuremath{\mathbb R}^2$
(i.e.\ any with $x_1,x_2 \notin \ensuremath{\mathbb Z}$) is covered by exactly two
squares in $P_n$ (for $n$ large enough). Any $(x_1,x_2)$ with exactly one
integer coordinate is covered by three squares in $P_n$ (the interior of
one square and some common edge of two further squares; again if $n$ is
large enough), and any $(x_1,x_2)$ with
two integer coordinates is covered by four squares (midpoints of edges)
if $x_1+x_2$ is even, and by five squares (the centre of
one square and the common vertex of four further squares) if $x_1+x_2$
is odd. This yields the following result.
\begin{lem} \label{lem:covdeg2}
For all $n \in \ensuremath{\mathbb N}$ the covering degree of $P_n$ is two in
$\{ (x_1,x_2) \in \ensuremath{\mathbb R}^2 \; | \; |x_1+x_2| \le 2^n-1 \}$.
\end{lem}
Since we may decide to equip the squares used above with
some additional decoration
there are several ways in which this construction yields tilings of
$\ensuremath{\mathbb R}^2$. In the sequel we will mainly consider one particular tiling:
the {\em arrowed tiling} $\ensuremath{{\mathcal A}}$ that is obtained by decorating the
underlying square $R_0$ with arrows and colours as shown in Figure
\ref{fig:arrow} (left).
\begin{figure}
\includegraphics[width=110mm]{sw-pf-r0-r3.pdf}
\caption{The construction of Figure \ref{fig:gen0-3} applied to
squares with certain decorations. A single square $R_0$ is shown
on the left: the decoration divides $R_0$ into four unit squares of
distinct colours, where each unit square carries an arrow in addition.
These decorated unit squares yield the tiles of the tiling $\ensuremath{{\mathcal A}}$.
\label{fig:arrow}}
\end{figure}
The tiling $\ensuremath{{\mathcal A}}$ is obtained by considering what we see viewing $P_n$
from ``above'', where $P_n$ now consists of congruent copies of the
decorated square $R_0$. Formally,
each unit square $K=[k,k+1]\times [m,m+1]$ ($k,m \in \ensuremath{\mathbb Z}$) inherits
the decoration (colour and arrow) from the top square of the two
large squares in $P_n$ that contain $K$. We want to consider
$\ensuremath{{\mathcal A}}$ as a tiling by unit squares with decoration as prototiles
where the unit squares carry an
arrow as well as a colour (here: black, dark grey, light grey, white).
From now on the term ``tile'' always denotes such a unit square
with decoration, if not mentioned otherwise.
Regarding the point how to obtain a tiling of the entire plane from
the finite patches $P_n$, a very natural approach ---and a very useful
one---is considering the limit $\lim\limits_{n \to \infty} P_n$, where the
limit is taken with respect to the local topology \cite{BG}.
In particular, in this topology
two tilings are $\varepsilon$-close if they agree in a ball of radius
$1/\varepsilon$ centred in the origin. For our purposes we may define the
distance $d(\ensuremath{{\mathcal T}},\ensuremath{{\mathcal T}}')$ between two tilings $\ensuremath{{\mathcal T}}, \ensuremath{{\mathcal T}}'$ by
\[ \widetilde{d}(\ensuremath{{\mathcal T}}, \ensuremath{{\mathcal T}}') = \inf \{ \varepsilon > 0 \; | \; \exists x,y \in B_{\varepsilon}:
\, B_{1/\varepsilon} \cap (\ensuremath{{\mathcal T}}+x) = B_{1/\varepsilon} \cap (\ensuremath{{\mathcal T}}'+y) \} \]
where $B_r$ denotes the open ball of radius $r$ centred in 0.
In order to make $\widetilde{d}$ into a metric, i.e. in order to ensure
transitivity, we define
\[ d(\ensuremath{{\mathcal T}}, \ensuremath{{\mathcal T}}') = \min \{ \widetilde{d}(\ensuremath{{\mathcal T}}, \ensuremath{{\mathcal T}}') , \frac{1}{\sqrt{2}} \}, \]
compare \cite{LMS1}. We will see in the sequel that the central parts $S_n$
of $P_n$ for $n \ge 2$ form a nested sequence
\[ S_2 \subset S_3 \subset \cdots \subset S_n \subset \cdots, \]
where $S_n$ has support $\{ (x_1,x_2) \; | \; |x_1+x_2| < 2^{n-1}-1 \}$,
hence the sequence $P_n$---or any sequence of tilings $(\ensuremath{{\mathcal T}}_n)_n$ where
$\ensuremath{{\mathcal T}}_n$ contains $P_n$ as its central patch---converges with respect to
the metric $d$.
For the next steps it will be convenient to consider the arrowed tiling
$\ensuremath{{\mathcal A}}$. In order to avoid confusion let us call the corresponding
sequence $R_n$, i.e. the sequence $P_n$ where each tile is equipped
with the arrow decoration of Figure \ref{fig:arrow}.
The first observation is that the arrows on tiles at the boundary
of $R_n$ point outwards. In order to make the term ``boundary'' precise,
let $\bd(R_n)$ be the set of the tiles in $R_n$ that have edges
on the boundary of $\bigcup_{T \in R_n} T$. These are exactly the
tiles that have two edges that are not shared with other tiles
of $R_n$ ; or equivalently: these are exactly
the tiles with a vertex that is not vertex of another tile of $R_n$.
\begin{lem} \label{lem:pfeilaussen}
The arrows on all tiles at the boundary $\bd(R_n)$ of $R_n$
point outwards, i.e. in the direction of their vertex that is not a vertex
of any other tile of $R_n$.
\end{lem}
\begin{proof}
The claim is true for $R_0$ and $R_1$, hence it follows inductively
for all $R_n$ by considering the construction: The boundary of $R_n$
consists of the boundaries of $R_{n-1}$.
\end{proof}
\begin{lem} \label{lem:pfeildiag}
All arrows on the tiles in $R_n$ on the two diagonals $(x_1,x_2)$ with
$x_1=x_2$ or $x_1=-x_2$ respectively show in the following directions:
$(1,-1)$ for $x_1=x_2$, $(1,1)$ for $x_1=-x_2<0$, $(-1,-1)$ for $x_1=-x_2>0$.
\end{lem}
\begin{figure}
\includegraphics[width=100mm]{rn-rek.pdf}
\caption{The structure of the patches $R'_n$. Each $R'_n$
is build from four congruent copies of $R'_{n-1}$ and some further
tiles on its main diagonals. The arrows in the tiles on the diagonals
are pointing in the directions given in Lemma \ref{lem:pfeildiag}.
\label{fig:rn-rek}}
\end{figure}
This is an immediate consequence of Lemma \ref{lem:pfeilaussen} and
the construction. Figure \ref{fig:rn-rek} (right) illustrates the situation.
\begin{thm} \label{thm:lim-rn}
The sequence $R_n$ is convergent with respect to the metric $d$. Consequently,
any sequence of tilings $(\ensuremath{{\mathcal T}}_n)_n$ where $\ensuremath{{\mathcal T}}_n$ contains $R_n$ as its
central patch is convergent with respect to the metric $d$.
\end{thm}
\begin{proof}
The idea is to consider three steps of the iteration,
i.e. how $R_{n+2}$ is build up from congruent copies of $R_{n-1}$. This
shows that the central patch of $R_{n+1}$ reappears as
the central patch of $R_{n+2}$, compare Figure \ref{fig:rn-3rek}.
For $n \ge 1$ the set $R'_n:=R_n \setminus \bd (R_n)$ is nonempty
and reappears in $R_{n+1}$, see Figure \ref{fig:rn-rek}. Considering how
$R'_n$ is contained in $R_{n+1}$ we find that the central patch $S$ of
$R_{n+1}$ consists of four congruent copies of $R'_{n-1}$. (This is indicated in the
left part of Figure \ref{fig:rn-3rek}, the area shaded in darker grey.)
The central patch of $R_{n+2}$, indicated in Figure \ref{fig:rn-3rek} by
shading in lighter grey, is a translate of the central patch $S$ of
$R'_{n+1}$, since (a) it is build
up from congruent copies of $R'_{n-1}$ in the same manner (this yields
the bulky part), and (b) the arrows on the diagonal boundaries of the
$R'_{n-1}$ coincide by Lemma \ref{lem:pfeildiag} together with the
construction (this yields the
``skeleton'' part). The directions of the arrows on the tiles on
the boundaries of the $R'_{n-1}$ are indicated by small black squares
in the corresponding corners of the tiles in Figure \ref{fig:rn-3rek}.
Altogether this yields that the central patch of $R_{n+1}$
equals the central patch of $R_{n+2}$. Let us denote the central
patch of $R_{n+1}$ consisting of four congruent copies of $R'_{n-1}$
(plus the tiles on the two diagonals) by $S_{n+1}$. More precisely, let
\[ S_{n+1}:=R_{n+1} \cap \{ (x_1,x_2) \; | \; |x_1+x_2| \le 2^{n-1}-1 \}
=R_{n+2} \cap \{ (x_1,x_2) \; | \; |x_1+x_2| \le 2^{n-1}-1 \}. \]
The patches $S_{n+1}$ yield a nested sequence
\[ S_{2} \subset S_{3} \subset S_{3} \subset \cdots \subset S_{n}
\subset \cdots. \]
In particular we obtain that for any $n \in \ensuremath{\mathbb N}$ we have that
$R_{n+1}$ coincides with all $R_k$ ($k \ge n+1$) in a ball $B_{1/\varepsilon}$
for $\frac{1}{\varepsilon} \le \frac{1}{\sqrt{2}} (2^{n-1}-1)$, i.e. $\varepsilon \ge
\frac{\sqrt{2}}{2^{n-1}-1}$. Thus
\[ d(R_{n+1},R_k) \le \frac{\sqrt{2}}{2^{n-1}-1} \quad \mbox{for }
k \ge n+1. \]
Hence the sequence $R_n$, respectively any sequence
of tilings $(\ensuremath{{\mathcal T}}_n)_n$ where $\ensuremath{{\mathcal T}}_n$ contains $R_n$ as its central patch,
converges for $n \to \infty$ with respect to the metric $d$.
\end{proof}
\begin{figure}
\includegraphics[width=120mm]{rn-rek-3.pdf}
\caption{How congruent copies of $R'_{n-1}$ are located within $R_{n+1}$ (left
part) and $R_{n+2}$ (everything). \label{fig:rn-3rek}}
\end{figure}
Note that neither the sets $R_n$ nor the sets $R'_n$ form a nested
sequence, there are slight mismatches further away from the centre.
Since the sequence $R_n$, respectively any sequence of tilings $(\ensuremath{{\mathcal T}}_n)_n$
where each $\ensuremath{{\mathcal T}}_n$ contains $R_n$ as its
central patch, converges with respect to $d$, there is a unique limit
$\ensuremath{{\mathcal A}}= \lim_{n \to \infty} \ensuremath{{\mathcal T}}_n$. The tiling $\ensuremath{{\mathcal A}}$ is the desired infinite
tiling. Moreover we may now define the {\em hull} of $\ensuremath{{\mathcal A}}$. In the theory of
aperiodic tilings the hull turns out to be a central object of study.
The hull is the closure of the set of all translates of $\ensuremath{{\mathcal A}}$, i.e.
\[ \ensuremath{\mathbb X}(\ensuremath{{\mathcal A}}) = \overline{ \{ \ensuremath{{\mathcal A}}+t \; | \; t \in \ensuremath{\mathbb R}^2 \} }, \]
where $\overline{M}$ defines the closure of the set $M$ with respect to
$d$, see again \cite[Chapter 4]{BG} for details.
Since $\ensuremath{{\mathcal A}}$ has only finitely many prototiles, and all tiles are
vertex-to-vertex, the set of all finite patches in $\ensuremath{{\mathcal A}}$ up to some given
radius $r>0$ is finite, up to congurence (even up to translations). This
property is called {\em finite local complexity} \cite{BG}. It holds for
each tiling in $\ensuremath{\mathbb X}(\ensuremath{{\mathcal A}})$ and for the set $\ensuremath{\mathbb X}(\ensuremath{{\mathcal A}})$ as a whole. Hence by
standard reasoning (\cite{RW, SCH},
see also \cite[Chapter 4]{BG}), $\ensuremath{\mathbb X}(\ensuremath{{\mathcal A}})$ is compact with respect to $d$,
hence $(\ensuremath{\mathbb X}(\ensuremath{{\mathcal A}}),d)$ is a compact metric space.
\section{Properties of the tilings}
With the help of the metric $d$ above we are now able to give
a precise definition of aperiodic tilings. A tiling is
{\em aperiodic} if its hull does not contain any periodic tiling.
In particular, aperiodicity implies nonperiodicity. (The hull of
a periodic tiling $\ensuremath{{\mathcal T}}$ contains only translates of $\ensuremath{{\mathcal T}}$, so in
particular it contains {\em only} periodic tilings.)
Considering Figure \ref{fig:rn-3rek} one may get the impression that
$\ensuremath{{\mathcal A}}$ is periodic. At least the arrangement of the $R'_{n-1}$ in the image
seem to form a periodic pattern. This behaviour is well-known for
certain aperiodic structures like the chair tiling, the Robinson square
tiling or the one-dimensional period doubling sequence, see \cite{BG, FH}.
Loosely speaking, a limit-periodic tiling is one that is the union
of infinitely many periodic packings with larger and larger periods
where the lattices $\Lambda_i$ of periods are nested sequences
$\Lambda_1 \subset \Lambda_2 \subset \Lambda_3 \cdots$, possibly
up to a set of density zero. In fact, the exact definition of
limitperiodicity uses spectral properties of the hull. We will not explain
this in detail (compare~\cite{BG}) since there is a simple geometric
sufficient condition ensuring limitperiodicity that applies to our examples.
\begin{thm}
$\ensuremath{{\mathcal A}}$ is the union of 2-periodic packings $M_n$ and a set ${\mathcal Z}$
of density zero. More precisely, $\ensuremath{{\mathcal A}} = {\mathcal Z} \cup
\bigcup\limits_{n \in \ensuremath{\mathbb N}} M_n$, where
\begin{multline*} \label{eq:rn-per}
M_n:= \{ R'_n + \big((2^n,0)+2^{n+1} \Lambda), \phi R'_n + \big((0,-2^n)
+2^{n+1} \Lambda), \phi^2 R'_n + \big((-2^n,0)+2^{n+1} \Lambda),\\
\phi^3 R'_n + \big((0,2^n)+2^{n+1} \Lambda) \}
\end{multline*}
where $\Lambda = \langle (1,1), (1,-1) \rangle_{\ensuremath{\mathbb Z}}$ (i.e. the
integer span of the two vectors $(1,1)$ and $(1,-1)$) and
${\mathcal Z}$ is the set of all tiles in $\ensuremath{{\mathcal A}}$ whose diagonals
are contained in $\{ (x_1,x_2) \, | \, x_1= \pm x_2 \}$.
\end{thm}
\begin{proof}
By the proof of Theorem \ref{thm:lim-rn} we get that for any $n \in \ensuremath{\mathbb N}$
the central patch of $\ensuremath{{\mathcal A}}$ is $R_n$. Considering how patches of type
$R'_{n-3}$ are located in $R_n$ we obtain inductively arbitrary large
parts of 2-periodic packings consisting of congruent copies of
$R'_k$, $\phi R'_k$, $\phi^2 R'_k$ and
$\phi^3 R'_k$ for any $k \in \ensuremath{\mathbb N}$, compare Figure \ref{fig:rn-3rek}.
Figure \ref{fig:limitper} indicates the arrangement of congruent
copies of $R'_1$. Let $\Lambda = \langle (1,1), (1,-1) \rangle_{\ensuremath{\mathbb Z}}$.
Then the four 2-periodic sets
\begin{equation} \label{eq:r1-per}
R'_1 + \big((2,0)+4 \Lambda), \phi R'_1 + \big((0,-2)+4 \Lambda),
\phi^2 R'_1 + \big((-2,0)+4 \Lambda), \phi^3 R'_1 + \big((0,2)+4 \Lambda)
\end{equation}
yield already half of the tiles of $\ensuremath{{\mathcal A}}$. Each of the four sets has
period vectors $(4,4)$ and $(4,-4)$. This situation is
indicated in Figure \ref{fig:limitper}. The arrangement of $R_n$ is
similar on all levels $2^n$, hence further periodic structures
in $\ensuremath{{\mathcal A}}$ are
\begin{multline} \label{eq:rn-per}
R'_n + \big((2^n,0)+2^{n+1} \Lambda), \phi R'_n + \big((0,-2^n)
+2^{n+1} \Lambda), \phi^2 R'_n + \big((-2^n,0)+2^{n+1} \Lambda),\\
\phi^3 R'_n + \big((0,2^n)+2^{n+1} \Lambda)
\end{multline}
For each $n$ the union of the four sets has density $\frac{1}{2}$.
Each of the four sets has period vectors $(2^{n+1},2^{n+1})$ and
$(2^{n+1},-2^{n+1})$. It is easy to see that the union of these
sets for all $n \in \ensuremath{\mathbb N}$ contains all tiles of $\ensuremath{{\mathcal A}}$ except tiles
along the diagonals $x_1 = \pm x_2$, i.e. all tiles in ${\mathcal Z}$.
This set has density zero.
\begin{figure}
\includegraphics[width=130mm]{limitper.pdf}
\caption{The congruent copies of $R'_1$ in $\ensuremath{{\mathcal A}}$ form a periodic subset of
the tiling with density $\frac{1}{2}$. The centre of the shaded patch
is the origin. \label{fig:limitper}}
\end{figure}
\end{proof}
\begin{rem} \label{rem:limper+diag}
Note that Equation \eqref{eq:rn-per} yields all tiles in $\ensuremath{{\mathcal A}}$
except the tiles on the diagonals $x_1=\pm x_2$, hence
\eqref{eq:rn-per} together with Lemma \ref{lem:pfeildiag} yields
the entire tiling.
\end{rem}
The next theorem is the key to all further results. This holds
because the theory of substitution tilings is developed pretty
well, while there is no theory for the construction above yet.
\begin{thm} \label{thm:subst}
$\ensuremath{{\mathcal A}}$ can be generated by a primitive tile substitution rule.
\end{thm}
\begin{proof}
This result is achieved by giving an appropriate substitution
rule yielding the tiling $\ensuremath{{\mathcal A}}$. More precisely, we will show that this
rule yields a substitution tiling that possesses the particular
structure (union of periodic packings) described by
Remark \ref{rem:limper+diag}.
This substitution rule is shown in Figure \ref{fig:subst}.
Let us denote this substitution rule by $\sigma$. Its prototiles
are denoted by $T_1, T_2, T_3$ and $T_4$; they are the prototiles of $\ensuremath{{\mathcal A}}$.
Figure \ref{fig:pf-subst-2} shows the action of $\sigma,
\sigma^2, \sigma^3$ and $\sigma^4$ on $T_1$. Note that $\sigma(T_3)
= \sigma(T_4)$. Note also that there are no reflections involved:
all tiles in $\sigma(T_i)$ are direct congruent copies of the $T_i$, not
reflected congruent copies.
To obtain an infinite substitution tiling in $\ensuremath{\mathbb X}_{\sigma}$ it is useful
to consider a tiling $\ensuremath{{\mathcal S}}$ that is fixed under $\sigma$, i.e.
$\sigma(\ensuremath{{\mathcal S}})=\ensuremath{{\mathcal S}}$. The construction of such a tiling is standard: take
a legal ``seed'', i.e. a patch $P$ that is contained in some
$\sigma^n(T)$ for some prototile $T$, such that $\sigma(P)$ (or
$\sigma^2(P)$ or $\sigma^3(P)$ \ldots) contains a translate of
$P$ in its interior. Here we choose $P$ as the patch of four tiles
shown in Figure \ref{fig:pf-subst-3} left, consisting of three dark grey
tiles of type $T_2$ and one black tile of type $T_1$. $P$ occurs in
$\sigma^4(T_1)$, see Figure \ref{fig:pf-subst-2}.
Applying $\sigma$ to $P$ one gets that $P$ is exactly the central patch
of $\sigma(P)$, compare Figure \ref{fig:pf-subst-3}.
Hence $\sigma(P)$ is the central patch of $\sigma^2(P)$,
and inductively we get that $\sigma^n(P)$ is the central patch of
$\sigma^{n+1}(P)$. Hence the sequence $\sigma_n(P)$ (or any sequence of
tilings $\ensuremath{{\mathcal S}}_n$ having $\sigma^n(P)$ as their central patch respectively)
converge to some tiling in the local topology. Let us denote this tiling by $\ensuremath{{\mathcal S}}$.
\begin{figure}
\includegraphics[width=120mm]{pf-subst.pdf}
\caption{The substitution rule $\sigma$ yielding the tiling $\ensuremath{{\mathcal A}}$. \label{fig:subst}}
\end{figure}
In order to show that $\ensuremath{{\mathcal S}} = \ensuremath{{\mathcal A}}$ we use the particular structure
of $\ensuremath{{\mathcal A}}$ stated in Remark \ref{rem:limper+diag}.
Consider $\sigma^3(T_1), \sigma^3(T_2), \sigma^3(T_3)$ and $\sigma^3(T_4)$.
The interior of each of the four patches contains four congruent
copies of $R'_1$. On their boundaries these patches have halves of $R'_1$.
(See for instance Figure \ref{fig:pf-subst-2}, the patch on the right
is the union of one copy of $\sigma^3(T_1), \sigma^3(T_2), \sigma^3(T_3)$ and
$\sigma^3(T_4)$ each.) The entire constellation of congruent copies of
$R'_1$ and halves of $R'_1$ agrees on all four supertiles $\sigma^3(T_i)$,
and it is invariant under rotation by $\pi/2$ about the centre of each
$\sigma^3(T_i)$. All tiles in $\ensuremath{{\mathcal S}}$ lie edge to edge, hence all
``supertiles'' $\sigma^3(T_i)$ lie edge-to-edge. Hence $\ensuremath{{\mathcal S}}$ contains
the same periodic arrangement as $\ensuremath{{\mathcal A}}$, i.e.~the one in Equation
\eqref{eq:r1-per}.
\begin{figure}
\includegraphics[width=120mm]{pf-subst-2.pdf}
\caption{The image shows $T_1$, $\sigma(T_1)$, $\sigma^2(T_1)$,
$\sigma^3(T_1)$, and $\sigma^4(T_1)$ (from left to right).
On the right it is indicated how $\sigma^4(T_1)$ consists
of congruent copies of $\sigma^3(T_1)$, $\sigma^3(T_2)$,
$\sigma^3(T_3)$ and $\sigma^3(T_1)$.
A seed for $\ensuremath{{\mathcal A}}$ is found in $\sigma^4(T_1)$
(e.g. the middle of the fourth and the fifth row, compare
Figure \ref{fig:pf-subst-3}). \label{fig:pf-subst-2}}
\end{figure}
Observe that $\sigma(R_1)$ contains $R'_2$, and---more
generally---$\sigma(R_n)$ contains $R'_{n+1}$. (To see this one may
use Figure \ref{fig:arrow} together with Figure \ref{fig:subst}.)
Since $\sigma(\ensuremath{{\mathcal S}})=\ensuremath{{\mathcal S}}$ holds, the tiling $\ensuremath{{\mathcal S}}$
contains also the periodic arrangements from Equation
\eqref{eq:rn-per}. This shows
that $\ensuremath{{\mathcal S}}$ and $\ensuremath{{\mathcal A}}$ coincide everywhere except on the diagonals
$\{ (x,y) \, | \, x=\pm y \}$. Considering the action of $\sigma$
it is easy to see that $\ensuremath{{\mathcal S}}$ has
on each of the four branches of this set tiles with arrows that
show in the same directions, and that these directions coincide
with the ones in $\ensuremath{{\mathcal A}}$. Altogether we obtain $\ensuremath{{\mathcal S}}=\ensuremath{{\mathcal A}}$.
\end{proof}
\begin{figure}
\includegraphics[width=120mm]{pf-subst-3.pdf}
\caption{The iterates under $\sigma$ of the small patch on the left
converge to $\ensuremath{{\mathcal A}}$. \label{fig:pf-subst-3}}
\end{figure}
Note that there is also the concept of the hull $\ensuremath{\mathbb X}_{\sigma}$ of a
substitution. Since $\ensuremath{{\mathcal A}}$ can be generated by a primitive substitution
we have that $\ensuremath{\mathbb X}(\ensuremath{{\mathcal A}})=\ensuremath{\mathbb X}_{\sigma}$ (see \cite{BG} for details).
More important, we now obtain the following result easily.
\begin{thm}
All $\ensuremath{{\mathcal T}} \in \ensuremath{\mathbb X}(\ensuremath{{\mathcal A}})$ are aperiodic. In particular $\ensuremath{{\mathcal A}}$ is aperiodic
(hence nonperiodic).
\end{thm}
\begin{proof}
Since $\sigma$ is a primitive substitution and $\ensuremath{\mathbb X}(\ensuremath{{\mathcal A}})$ is of finite
local complexity, one can apply a classical result on substitution
tilings \cite[Thm 10.1.1]{GS}: a primitive substitution tiling is
aperiodic if in this tiling the level 1 supertiles can be identified
in a unique way. (See also Solomyak \cite[Thm 1.1]{Sol} for the proof
of the ``if and only if'' version of this result.)
For the tilings in $\ensuremath{\mathbb X}(\ensuremath{{\mathcal A}})$ this is particularly simple:
any level 1 supertile consists of four tiles sharing a common vertex.
These four tiles have at least three distinct colours, and the
arrows on the tiles do neither point to this common vertex nor
do they point away from it. This leaves only one possibility:
the vertices of the congruent copies of $R'_1$ are the centres of the supertiles.
\end{proof}
There are two geometric properties of a tiling $\ensuremath{{\mathcal T}}$ that have strong consequences
on the dynamical properties of the hull $\ensuremath{\mathbb X}(\ensuremath{{\mathcal T}})$ of $\ensuremath{{\mathcal T}}$. A tiling $\ensuremath{{\mathcal T}}$ is
{\em repetitive} if for each $r>0$ there is $R>0$ such that each patch
of radius less than $r$ is contained in each patch of radius $R$.
If $R \in O(r)$ then $\ensuremath{{\mathcal T}}$ is called {\em linearly repetitive}.
Moreover, a tiling $\ensuremath{{\mathcal T}}$ has {\em uniform patch frequencies} if the
frequencies of all patches are well-defined. More precisely: if $P$
is a patch in $\ensuremath{{\mathcal T}}$ then let $N_P(B_r(x))$ denote the number of congruent
copies of $P$ in $\ensuremath{{\mathcal T}} \cap B_r(x)$, where $B_r(x)$ denotes the open ball
of radius $r$ centred in $x$. If for all patches $P$ in $\ensuremath{{\mathcal T}}$ the limit
\[ \lim_{r \to \infty} \frac{1}{\pi r^2} N_P(B_r(x)) \]
exists uniformly in $x$ then $\ensuremath{{\mathcal T}}$ has {\em uniform patch frequencies}.
For a more thorough discussion of repetitivity or uniform patch frequencies
see \cite{BG} or \cite{FR}. We omit it here since we need the terms only
to state the following result.
\begin{cor}
All $\ensuremath{{\mathcal T}} \in \ensuremath{\mathbb X}(\ensuremath{{\mathcal A}})$ are linearly repetitive and have uniform patch frequencies.
\end{cor}
\begin{proof}
By a result of Solomyak \cite[Lemma 2.3]{Sol} each primitive substitution
tiling in $\ensuremath{\mathbb R}^2$ with finite local complexity is linearly repetitive.
By a result of Lagarias and Pleasants \cite[Theorem 6.1]{LP} linear
repetitivity implies uniform patch frequencies.
\end{proof}
The fact that our tilings possess the particular structure
(union of periodic packings) described by Remark \ref{rem:limper+diag}
is a hint that they are possibly {\em limitperiodic}. In this
particular case this means that
they can be generated by a certain cut-and-project method using a
lattice in $\ensuremath{\mathbb R}^2 \times (\ensuremath{\mathbb Q}_2)^2$, where $\ensuremath{\mathbb Q}_2$ denotes the field
of 2-adic numbers. Structures of this kind are called
{\em model sets (with 2-adic internal space)}. Model sets are
relevant since by a result of Hof \cite{hof, SCH, BG} all models
sets show pure point diffraction.
It is beyond the scope of this paper to give details on this, but
we may formulate the result and prove it using a simple to check
sufficient condition.
Properly speaking, model sets are discrete point sets, not tilings.
Hence our tilings are not model sets, but they are strongly related:
they are mutually locally derivable (mld) with model sets, meaning
that there is a local rule transforming one into the other
(see \cite{BG}).
\begin{thm}
All $\ensuremath{{\mathcal T}} \in \ensuremath{\mathbb X}(\ensuremath{{\mathcal A}})$ are limitperiodic. Consequently, all
$\ensuremath{{\mathcal T}} \in \ensuremath{\mathbb X}(\ensuremath{{\mathcal A}})$ are mutually locally derivable with model sets,
hence pure point diffractive.
\end{thm}
\begin{proof}
Since $\ensuremath{{\mathcal A}}$ is a primitive substitution tiling with integer scaling
factor (in this case 2) and $\sigma$ is a {\em block substitution} (since
each unit square is mapped to four unit squares in a $2 \times 2$
grid) we can apply a result in \cite{LMS1} and \cite{LMS2}:
we need to show that $\sigma$ has a modular coincidence.
By a result in \cite{FS}
this is the case if there is a {\em coincidence} in the supertiles.
For this it suffices to note that the upper right tile in $\sigma(T_1),
\sigma(T_2), \sigma(T_3)$ and $\sigma(T_4)$ is a black tile $T_1$
with its arrow pointing down right, compare Figure \ref{fig:subst}.
(The lower left tile in all four cases is $T_2$, yielding a second
coincidence.) For further details see \cite{LMS1, LMS2, FS, BG}.
\end{proof}
The fact that $\ensuremath{{\mathcal A}}$ is pure point diffractive has consequences if one
imagines $\ensuremath{{\mathcal A}}$ as a physical solid: assume that in such a solid the four
different types of atoms (or
molecules) are arranged in the same pattern as the four different
tiles in $\ensuremath{{\mathcal A}}$. A diffraction experiment would then show a diffraction
image with bright spots (``Bragg peaks'') and (ideally) no diffuse parts.
Now that we have obtained several results on the arrowed tiling $\ensuremath{{\mathcal A}}$
we turn our attention to the naked tiling using no decoration at all,
compare Figure \ref{fig:gen0-3}. Let us called the tiling obtained
by the same construction but with undecorated squares ${\mathcal N}$.
A tiling $\ensuremath{{\mathcal T}}_1$ is
called {\em locally derivable} from a tiling $\ensuremath{{\mathcal T}}_2$ if there is a local
rule transforming $\ensuremath{{\mathcal T}}_2$ into $\ensuremath{{\mathcal T}}_1$. Two tilings $\ensuremath{{\mathcal T}}_1, \ensuremath{{\mathcal T}}_2$ are
called {\em mutually locally derivable} if $\ensuremath{{\mathcal T}}_1$ is locally
derivable from $\ensuremath{{\mathcal T}}_2$ and vice versa \cite{BG}.
\begin{lem}
The naked tiling ${\mathcal N}$ is locally derivable from the
arrowed tiling $\ensuremath{{\mathcal A}}$.
\end{lem}
\begin{proof}
It suffices to give a local rule to transform $\ensuremath{{\mathcal A}}$ into ${\mathcal N}$.
Considering that in the arrow decoration of $\ensuremath{{\mathcal A}}$ the arrows point always
away from the centre of the large square, the arrows determine
locally the edges of the large (overlapping) squares in a unique
way. The edges of the overlapping squares determine ${\mathcal N}$.
\end{proof}
Two tilings that are mutually locally derivable share a lot of
properties (aperiodicity, repetitivity, pure point diffraction,
uniform patch frequency...) Thus it would suffice to give a local rule
that transforms ${\mathcal N}$ into $\ensuremath{{\mathcal A}}$. We probably found such a rule,
but unfortunately we were yet unable to prove that it really works.
{\bf Problem:} Are $\ensuremath{{\mathcal A}}$ and ${\mathcal N}$ mutually locally derivable?
As an intermediate step, one may consider whether $\ensuremath{{\mathcal A}}$ is mutually
locally derivable with the tiling obtained from $\ensuremath{{\mathcal A}}$ by deleting the
colours. It is a simple exercise to see that this is indeed true.
\section{Remarks and Outlook}
The original construction of inductive rotation tilings found by the
second author used to place the origin close to the leftmost part of the
patches $P_n$. Hence the ``limit'' of this sequence fills only a
quarter plane. Four copies of this limit could be used to fill the
entire plane. The resulting tilings are the same as the ones described
in this paper. The construction in this paper is more adapted to
the notion of convergence of sequences of tilings used here.
Here we showed that the tiling $\ensuremath{{\mathcal A}}$ is nonperiodic by showing that
$\ensuremath{{\mathcal A}}$ is a substitution tiling and then applying the result of Solomyak
that a substitution tiling is nonperiodic if $\sigma^{-1}$ is unique.
We found an alternative proof of nonperiodicity by direct means, using
only the limitperiodic structure of $\ensuremath{{\mathcal A}}$. For the sake of briefness we
omit the alternative proof here.
The construction presented here uses squares and rotations by
$\frac{\pi}{2}$. The second author found similar constructions for
triangles and rotations by $\frac{2\pi}{3}$ as well as for hexagons
and rotations by $\frac{\pi}{3}$. For an artistic application of these
constructions see {\tt http://hofstetterkurt.net/ip}, see also
\cite{par}.
The substitution rule $\sigma$ used to generate the tiling $\ensuremath{{\mathcal A}}$
uses four prototiles, but $\sigma(T_3)=\sigma(T_4)$. Probably
one of these tiles is redundant and everything works for a substitution
for three prototiles as well. We prefer to use the four-colour
version since it carries more information, so it might make some
arguments more clear.
We studied the tilings $\ensuremath{{\mathcal A}}$ and ${\mathcal N}$. We have good evidence (but no
proof so far) that $\ensuremath{{\mathcal A}}$ and ${\mathcal N}$ are mutually locally derivable.
What about other decorations? There are decorations of the large square
$Q$ leading to 2-periodic tilings. So the construction yields at least
two distinct mld classes (one aperiodic, one 2-periodic). Are there
more mld classes?
The tilings ${\mathcal N}$ and $\ensuremath{{\mathcal A}}$ are obtained by looking on $R_n$ from
``above''. Are the tilings obtained by looking from below congruent to
${\mathcal N}$ respectively $\ensuremath{{\mathcal A}}$? Are they at least mutually locally
derivable with ${\mathcal N}$ respectively $\ensuremath{{\mathcal A}}$?
We may realise the sets $R_n$ also with tiles with ``thickness'' in $\ensuremath{\mathbb R}^3$.
Now above and below have a precise meaning. The construction for
the sets $R_n$ now has to be described in three dimensions. Is there
such a construction such that the height of all $R_n$ is bounded by a
common constant? Are there arbitrary high stairs?
\section*{Acknowledgement}
The second author wants to express his gratitude to the many people
from Mathematics, Art, Computer Science and Textile Technology
who are supporting the work on Inductive Rotation Tilings, enabling new
applications of aperiodic tilings in Arts and Science.
|
\section{Introduction}
An interesting new idea started emerging a decade ago (see, e.g., Treu et al.
2006; Grillo et al. 2008; Biesiada et al. 2010; but also see Futamase \&
Yoshida 2001) to use individual lensing galaxies in order to measure
cosmological parameters. In principle, the deflection of quasar light by
the intervening galaxy is known precisely from general relativity as
long as one has a good model for the mass distribution within the lens
(Bartelmann \& Schneider 1999; Refregier 2003). The Einstein radius,
inferred from the deflection angle, then provides a measure of the
angular-size distance, which may be used to discriminate between
competing cosmological models.
The key, of course, is how well we understand the distribution of
matter within the lens, and this appears to be the principal source
of error in this type of measurement. As of today, the observation
of some 70 or so lensing galaxy systems has provided the data
that, in principle, can be used to carry out this kind of study.
The results thus far are consistent with the standard ($\Lambda$CDM)
model, though the precision with which model parameters may be
determined with this appoach is not yet as good as that available in
other studies, e.g., using Type Ia SNe as standard candles (see,
e.g., Riess et al. 1998; Perlmutter et al. 1999).
In recent years, the application of model selection tools in one-on-one
comparisons between $\Lambda$CDM and a cosmology we refer to as the
$R_{\rm h}=ct$ Universe (Melia 2007; Melia \& Shevchuk 2012) has shown
that the data actually tend to favor the latter over the former. These
include the use of cosmic chronometers (Melia \& Maier 2013), high-$z$
quasars (Melia 2013), gamma ray bursts (Wei et al. 2013) and, most recently,
the Type Ia SNe themselves (Wei et al. 2014b). The simplest way to
view this cosmology is to start with $\Lambda$CDM and then apply the
additional constraint $p=-\rho/3$ on its total equation of state,
where $p$ and $\rho$ are the total pressure and energy density,
respectively. With these other kinds of data, $R_{\rm h}=ct$ is
favored over $\Lambda$CDM with a likelihood of $\sim 90\%$ versus
only $\sim 10\%$.
The principal goal of this paper is to broaden the comparison between
$R_{\rm h}=ct$ and $\Lambda$CDM by now including strong gravitational
lenses in this study. In \S~2 of this paper, we describe the method,
and then assemble the catalog of suitable lensing systems in \S~3.
We discuss our results in \S~4. We will find that the current strong
lensing sample is not yet large enough to differentiate between these
two competing models, and we show in \S~5 how large the source
catalog needs to be in order to rule out one or the other expansion
scenarios at a 3-sigma confidence level. We present our conclusions
in \S~6.
\section{Strong Lensing}
The lens model often fitted to the observed images is based on a
singular isothermal ellipsoid (SIE; Ratnatunga et al. 1999), in which
the projected mass distribution (at redshift $z_l$) is elliptical,
with semi-minor axis $\theta_1$ and semi-major axis $\theta_2$.
In this paper, we will adopt the simpler version, using a singular
isothermal sphere instead. For generality, we will describe the
approach using semi-major and semi-minor axes, though we
will later set the two angles $\theta_1$ and $\theta_2$ equal
to each other. The source lensed by this system is a quasar
at redshift $z_s>z_l$. The key expression in strong gravitational
lensing theory is the lens equation (Schneider et al. 1992), which
gives the mapping between positions $\beta$ in the source plane
and $\theta$ in the image plane, according to
\begin{equation}
\beta = \theta -\nabla_\theta \Phi\;.
\end{equation}
The lensing potential of the singular isothermal ellipsoid may
be written (Kormann et al. 1994)
\begin{equation}
\Phi=\theta_{\rm E}\sqrt{(1-\epsilon)\theta_1^2+(1+\epsilon)\theta_2^2}\;,
\end{equation}
where the Einstein radius $\theta_{\rm E}$ is defined below in terms of
the (one-dimensional) velocity dispersion, $\sigma_v$, in the lensing galaxy,
and the angular diameter distances, $D_A(z_l,z_s)$ and $D_A(0,z_s)$,
between lens and source and between source and observer, respectively.
The `ellipticity' $\epsilon$ is related to the eccentricity $e$ of the
critical line by
\begin{equation}
e=\sqrt{(1-\epsilon)/(1+\epsilon)}\;.
\end{equation}
In addition, the semi-major and semi-minor axes are related to
the Einstein radius
\begin{equation}
\theta_{\rm E}\equiv 4\pi\left({\sigma_v\over c}\right)^2{\mathcal{D}}\;,
\end{equation}
where
\begin{equation}
{\mathcal{D}}\equiv {D_A(z_l,z_s)\over D_A(0,z_s)}\;,
\end{equation}
via the relations
\begin{eqnarray}
\theta_1&=&\theta_{\rm E}\sqrt{1-\epsilon}\nonumber \\
\theta_2&=&\theta_{\rm E}\sqrt{1+\epsilon}\;.
\end{eqnarray}
As noted, earlier, we will here consider the simpler case of a single
isothermal sphere (SIS), for which $\theta_1=\theta_2$.
In principle, equation~(4) can be used to test cosmological models in a
rather unique way because, unlike other kinds of comparisons that rely
on the optimization of the Hubble constant $H_0$, this particular analysis
is completely independent of $H_0$.\footnote{One can in fact
determine $H_0$ directly using strong gravitational lensing, but only
when time delays are measured between the various images of a given
source (see, e.g., Paraficz \& Hjorth 2009; Suyu et al. 2013; Wei
et al. 2014a).} Nonetheless, even though knowledge of $H_0$ is
not necessary for this type of test, fits to the data do depend on
the reliability of lens modelling (e.g., via the assumption of a singular
isothermal sphere, or a singular isothermal ellipsoid) and the
measurement of the velocity dispersion.
When using these expressions, $\sigma_v$ (the total velocity dispersion
of stellar plus dark matter) cannot be obtained directly from the
surface-brightness weighted average of the line-of-sight velocity dispersion
that is actually measured. In practice, the central velocity dispersion $\sigma_0$
is estimated from the {\it stellar} velocity dispersion within $R_e/8$, where
$R_e$ is the optical effective radius (see, e.g., Treu et al. 2006; Grillo
et al. 2008), and is then used to represent the velocity dispersion
$\sigma_{SIS}$ for the corresponding singular isothermal sphere or
ellpisoid (for the total mass present).
This works rather well because inside one effective radius, massive
elliptical galaxies are kinematically indistinguishable from an
isothermal ellipsoid (Koopmans et al. 2009), which is quite remarkable
considering the fact that $\sigma_{SIS}$ and $\sigma_0$ need not be
the same. One reason is that dark matter halos appear to be dynamically
hotter than the luminous stars (based on X-ray observations), so the
former must necessarily have a greater velocity dispersion than the
latter (White \& Davis 1996).
Still, when one introduces the SIS equivalent value $\sigma_{SIS}$ obtained
from modelling the lens as a singular isothermal sphere, these two
measures of velocity dispersion agree very closely. Treu et al. (2006) used
the large and homogeneously selected sample of lenses identified by the Sloan
Lenses ACS Survey (SLACS; Bolton et al. 2005, 2006) to study in detail the
degree of homogeneity of the early-type galaxies by measuring the ratio
between stellar velocity dispersion and $\sigma_{SIS}$ that best fits the
geometry of the corresponding multiple images. They found that the ratio
${\sigma_0/ \sigma_{SIS}}$ is very close to unity; specifically, they
inferred a sample average value $\langle {\sigma_0/ \sigma_{SIS}}\rangle=
1.010\pm 0.017$, with a relatively small scatter of $\sim 0.06$. Similarly,
van de Ven et al. (2003) examined this ratio for a range of anisotropy
parameters and found that $0.96<{\sigma_0/ \sigma_{SIS}}<1.08$. The
conclusion from such studies is that on average the approximation
$\sigma_v=\sigma_{SIS}\approx\sigma_0$ works surprisingly well, due perhaps
to some as yet unknown mechanism that couples stellar and dark mass,
sometimes referred to as a bulge-halo `conspiracy'. A possible resolution
of this parity may be that since the NFW (Navarro et al. 1997) and observed
stellar mass profiles are nearly isothermal, the more concentrated
mass profiles for the baryon component than for dark matter may
simply be a consequence of dissipative star formation. The observation
of $\sigma_0/\sigma_{SIS}\sim 1$ may therefore not be so
mysterious. Nonetheless, significant departures from this are seen
in individual cases, so one cannot ignore the scatter in any discussion
concerning the propagated measurement error for ${\mathcal{D}}_{obs}$.
In this paper, we will follow Cao et al. (2012), and put
\begin{equation}
\theta_{\rm E}\equiv 4\pi\left({\sigma_{SIS}\over c}\right)^2{\mathcal{D}}\;,
\end{equation}
with
\begin{equation}
\sigma_{SIS}\equiv f_{SIS}\,\sigma_0\;.
\end{equation}
We will keep $f_{SIS}$ as a free parameter to be optimized in the
fits, since it mimics at least several effects that apparently give
rise to the observed scatter in the individually measured ratio
$\sigma_0/\sigma_{SIS}$ for each system. These include: (1) systematic
errors in the rms deviation of $\sigma_{SIS}$ from $\sigma_0$; (2)
an rms error associated with the assumption that the SIS model
allows the translation from observed image separation to $\theta_{\rm E}$;
and (3) a softened isothermal sphere potential, which tends to decrease
the typical image separations (Narayan \& Bartelmann 1996). In the
analysis we describe below, we will adopt a dispersion $\sigma_f=0.06\,f_{SIS}$,
based on the rms scatter of $\sim 6\%$ found from
the work of Treu et al. (2006) and van de Ven et al. (2003). Note,
however, that $\sigma_f$ may be as big as $\sim 0.2$ according to
Cao et al. (2012), though one might have expected such a large value
to have emerged directly from the aforementioned survey by Treu et
al. (2006).
The overall uncertainty associated with ${\mathcal{D}_{obs}}$ (calculated
from equation~7) is estimated through the propagation equation involving
errors in $\theta_{\rm E}$, $\sigma_0$, and $f_{SIS}$. According to
Grillo et al. (2008), the error in measuring the Einstein radius
$\theta_{\rm E}$ is $\sim 5\%$, so we will assume a dispersion
$\sigma_{\theta_{\rm E}}=0.05\,\theta_{\rm E}$ for this quantity. In principle, any
optimization of the model parameters (and $f_{SIS}$) carried out while
fitting the data should also include the dispersion $\sigma_z$ in the
measured redshifts $z_l$ and $z_s$ (since the theoretical values
$D_A(z_l,z_s)$ and $D_A(0,z_s)$ directly depend on these).
However, a careful analysis
of SDSS quasar spectra shows that $\sigma_z/(1+z)\sim 10^{-4}$ over a
broad range of redshifts (Hewett \& Wild 2010). This error is so small
compared to the other three uncertainties that we will ignore it.
So in total, we will calculate the dispersion $\sigma_{\mathcal{D}}$
in ${\mathcal{D}_{obs}}$ using the expression
\begin{equation}
\sigma_{\mathcal{D}}={\mathcal{D}_{obs}}\left[\left({\sigma_{\theta_{\rm E}}\over
\theta_{\rm E}}\right)^2+4\left({\sigma_{\sigma_{0}}\over \sigma_0}\right)^2
+4\left({\sigma_f\over f_{SIS}}\right)^2\right]^{1/2}\;.
\end{equation}
Note that since the uncertainty in $\sigma_0$ also appears to be $\sim 5\%$
(Grillo et al. 2008), the average dispersion in the measured value of $\mathcal{D}$
is expected to be $\sigma_{\mathcal{D}}\sim 0.16\,{\mathcal{D}_{obs}}$.
Now, in principle, only the range $0\le {\mathcal{D}}\le 1$ is physically
meaningful, but such a value of $\sigma_{\mathcal{D}}$ can result in at least
some measurements ${\mathcal{D}_{obs}}>1.0$. A quick inspection of
equation~(5) shows that the measurements most at risk for this type of
outcome involve lenses much closer to the observer than the source. As
we shall see below, several of the sources in our complete sample have
${\mathcal{D}_{obs}}>1.0$. Though unrealistic, such values are
consistent with the quoted error, so we will include them in our analysis.
But to demonstrate their negative impact on the optimization of the model fits,
we will also carry out the analysis for a reduced sample omitting these
sources. As the measurements become more precise, and the sample
of strong lensing sources grows, it may be possible to avoid systems
with ${\mathcal{D}_{obs}}>1.0$ altogether.
From a theoretical standpoint, one must assume a cosmological
model in order to calculate the angular diameter distances $D_A(z_l,z_s)$
and $D_A(0,z_s)$, once the redshifts $z_l$ and $z_s$ for a particular
lensing system are known. In $\Lambda$CDM, this distance depends
on several parameters, including $H_0$ and the mass fractions
$\Omega_{\rm m} \equiv \rho_{\rm m}/\rho_{\rm c}$, $\Omega_{\rm r}\equiv
\rho_{\rm r}/\rho_{\rm c}$, and $\Omega_{\rm de}\equiv \rho_{\rm de}/
\rho_{\rm c}$, defined in terms of the current matter ($\rho_{\rm m}$),
radiation ($\rho_{\rm r}$), and dark energy ($\rho_{\rm de}$) densities,
and the critical density $\rho_{\rm c}\equiv 3c^2H_0^2/8\pi G$.
Assuming zero spatial curvature, so that $\Omega_{\rm m}+\Omega_{\rm r}
+\Omega_{\rm de}=1$, the angular diameter distance between redshifts
$z_1$ and $z_2$ ($>z_1$) is given by the expression
\begin{equation}
D_A^{\Lambda{\rm CDM}}(z_1,z_2)={c\over H_0}{1\over (1+z_2)}\int_{z_1}^{z_2}
\left[\Omega_{\rm m}(1+z)^3+\Omega_{\rm r}(1+z)^4+\Omega_{\rm de}
(1+z)^{3(1+w_{\rm de})}\right]^{-1/2}\;dz\;,
\end{equation}
where $p_{\rm de}=w_{\rm de}\rho_{\rm de}$ is the dark-energy equation
of state. As noted earlier, $H_0$ cancels out when we divide
$D_A^{\Lambda{\rm CDM}}(z_l,z_s)$ by $D_A^{\Lambda{\rm CDM}}(0,z_s)$
to form the ratio $\mathcal{D}_{\Lambda{\rm CDM}}$,
so the essential remaining parameters in flat $\Lambda$CDM are
$\Omega_{\rm m}$ and $w_{\rm de}$. If we further assume that
dark energy is a cosmological constant with $w_{\rm de}=-1$,
then only the parameter $\Omega_{\rm m}$ is available to fit
the data.
In the $R_{\rm h}=ct$ Universe (Melia 2007; Melia \& Shevchuk 2012),
the angular diameter distance depends only on $H_0$, but since
here too the Hubble constant cancels out in the ratio $\mathcal{D}_{R_{\rm h}=ct}$,
there are actually no free parameters left for fitting the gravitational
lensing data. In this cosmology,
\begin{equation}
D_A^{R_{\rm h}=ct}(z_1,z_2)={c\over H_0}{1\over (1+z_2)}
\ln\left({1+z_2\over 1+z_1}\right)\;.
\end{equation}
\section{Strong Gravitational Lensing Systems}
Our sample is drawn from a compilation of 69 strong lensing systems
(listed in Table I), with
good spectroscopic measurements of the central velocity dispersion, using
the SLACS ({\it Sloan Lens ACS}) Survey (first introduced by Bolton et al.
2006; Treu et al. 2006; and Koopmans et al. 2006), and the LSD ({\it Lenses
Structure and Dynamics}) Survey (see, e.g., Bolton et al. 2008; Newton
et al. 2011). Some original contributions to these data sets may be found
in Young et al. (1980), Huchra et al. (1985), Leh\'ar et al. (1993),
Fassnacht et al. (1996), Tonry et al. (1998), Koopmans \& Treu (2002, 2003),
and Treu and Koopmans (2004). The velocity dispersion $\sigma_0$ and its
uncertainty (with the aforementioned average value of $\sim 5\%$) were
obtained from the {\it Sloan Digital Sky Survey Database}. The SLACS and
LENS surveys complement each other rather well, with the former comprised
primarily of lens galaxies at redshift up to $\sim 0.3$, while the latter
includes systems beyond $z\sim 0.5$.
It has already been noted before (see, e.g., Biesiada et al.
2010, 2011; Cao et al. 2012) that some of these lenses produce 2 images,
while others produce 4. Both 2-image and 4-image lens systems are
usually affected by external shear. This effect degenerates with the
ellipticity of the SIE component, which introduces some uncertainty
in estimating the Einstein radius of a given lens. For 2-image
systems this can be more problematic due to a lack of observational
constraints. On the other hand, 4-image systems are better constrained
observationally, so both their ellipticity and external shear may be
determined for a more accurate measurement of the Einstein radius.
To gauge whether there are any systematic effects associated with
one category or the other, we will here track the results using both
sets of lens system.
There is an additional drawback to the measurement of $\mathcal{D}$
with strong gravitational lenses that we must carefully consider here. By
its very definition, $\mathcal{D}$ is confined
to a very compact range of values $(0,1)$, regardless of the lens ($z_l$)
and quasar ($z_s$) redshifts. In addition, there is no monotonic progression
from low to high values of $\mathcal{D}$ as the sequence of gravitational
lenses approaches or recedes from us, since the sources may lie anywhere
beyond them. Fortunately, $\mathcal{D}$ does not depend on $H_0$, so a
comparison between theoretical values of this ratio and $\mathcal{D}_{\rm obs}$
is not inhibited by any uncertainty in the expansion rate itself.
However, the tight range in $\mathcal{D}$ and its lack of correlation
with $z$ make it difficult to optimize parameters such as $\Omega_{\rm m}$,
which produce only slight changes in $\mathcal{D}_{\Lambda{\rm CDM}}$ even
when they increase by a factor of 2 or more. In this paper, we will
therefore compare how well $R_{\rm h}=ct$ fits the data with several
specific variations of $\Lambda$CDM, though always assuming a flat spatial
curvature and a cosmological constant with $w_{\rm de}=-1$. The most prominent
\begin{deluxetable}{lllccclccl}
\tablewidth{520pt}
\tabletypesize{\footnotesize}
\tablecaption{Strong Gravitational Lensing Systems}\tablenum{1}
\tablehead{\colhead{Galaxy}&\colhead{$z_l$}&\colhead{$z_s$}&\colhead{$\theta_{\rm E}$}&\colhead{$\sigma_0$}&
\colhead{$\mathcal{D}_{\rm obs}$}&\colhead{$\sigma_{\mathcal{D}}$}&
\colhead{${\mathcal{D}_{R_{\rm h}=ct}}$}&\colhead{${\mathcal{D}_{\Lambda{\rm CDM}}}$}& \colhead{Refs.} \\
&&&({\rm arcsec})&(${\rm km}\;{\rm s}^{-1})$&$f_{\rm SIS}=1.02$&&& $(\Omega_{\rm m}=0.27)$&
} \startdata
\sidehead{\centerline{\qquad\quad Systems with Two Images}}
{\rm SDSS J}0037-0942 & 0.1955 & 0.6322 & 1.47 & 282$\pm$11 & 0.617 & 0.094 & 0.636 & 0.656 & 1--9 \\
{\rm SDSS J}0216-0813 & 0.3317 & 0.5235 & 1.15 & 349$\pm$24 & 0.316 & 0.060 & 0.320 & 0.336 & 1--9 \\
{\rm SDSS J}0737+3216 & 0.3223 & 0.5812 & 1.03 & 326$\pm$16 & 0.323 & 0.053 & 0.390 & 0.409 & 1--9 \\
{\rm SDSS J}0912+0029 & 0.1642 & 0.3240 & 1.61 & 325$\pm$12 & 0.509 & 0.076 & 0.458 & 0.474 & 1--9 \\
{\rm SDSS J}1250+0523 & 0.2318 & 0.7950 & 1.15 & 274$\pm$15 & 0.511 & 0.087 & 0.644 & 0.665 & 1--9 \\
{\rm SDSS J}1630+4520 & 0.2479 & 0.7933 & 1.81 & 279$\pm$17 & 0.776 & 0.138 & 0.621 & 0.642 & 1--9 \\
{\rm SDSS J}2300+0022 & 0.2285 & 0.4635 & 1.25 & 305$\pm$19 & 0.448 & 0.081 & 0.460 & 0.479 & 1--9 \\
{\rm SDSS J}2303+1422 & 0.1553 & 0.5170 & 1.64 & 271$\pm$16 & 0.745 & 0.131 & 0.654 & 0.673 & 1--9 \\
{\rm CFRS}03.1077 & 0.9380 & 2.9410 & 1.24 & 251$\pm$19 & 0.657 & 0.131 & 0.518 & 0.506 & 1--9 \\
{\rm HST} 15433 & 0.4970 & 2.0920 & 0.36 & 116$\pm$10 & 0.893 & 0.193 & 0.643 & 0.652 & 1--9 \\
{\rm MG}2016 & 1.004 & 3.263 & 1.56 & 328$\pm$32 & 0.484 & 0.113 & 0.521 & 0.504 & 1--9 \\
{\rm SDSS J}0044+0113 & 0.1196 & 0.1965 & 0.79 & 266$\pm$13 & 0.373 & 0.061 & 0.370 & 0.381 & 10,11 \\
{\rm SDSS J}0330-0020 & 0.3507 & 1.0709 & 1.10 & 212$\pm$21 & 0.817 & 0.194 & 0.587 & 0.608 & 10,11 \\
{\rm SDSS J}0935-0003 & 0.3475 & 0.4670 & 0.87 & 396$\pm$35 & 0.185 & 0.041 & 0.222 & 0.234 & 10,11 \\
{\rm SDSS J}0955+0101 & 0.1109 & 0.3159 & 0.91 & 192$\pm$13 & 0.824 & 0.155 & 0.617 & 0.632 & 10,11 \\
{\rm SDSS J}0959+4416 & 0.2369 & 0.5315 & 0.96 & 244$\pm$19 & 0.538 & 0.109 & 0.501 & 0.521 & 10 \\
{\rm SDSS J}1112+0826 & 0.2730 & 0.6295 & 1.48 & 320$\pm$20 & 0.486 & 0.088 & 0.506 & 0.527 & 10,11 \\
{\rm SDSS J}1142+1001 & 0.2218 & 0.5039 & 0.98 & 221$\pm$22 & 0.670 & 0.159 & 0.509 & 0.529 & 10,11 \\
{\rm SDSS J}1143-0144 & 0.1060 & 0.4019 & 1.68 & 269$\pm$13 & 0.775 & 0.126 & 0.702 & 0.718 & 10 \\
{\rm SDSS J}1204+0358 & 0.1644 & 0.6307 & 1.31 & 267$\pm$17 & 0.613 & 0.112 & 0.689 & 0.708 & 10,11 \\
{\rm SDSS J}1205+4910 & 0.2150 & 0.4808 & 1.22 & 281$\pm$14 & 0.516 & 0.084 & 0.504 & 0.524 & 10 \\
{\rm SDSS J}1213+6708 & 0.1229 & 0.6402 & 1.42 & 292$\pm$15 & 0.556 & 0.092 & 0.766 & 0.783 & 10,11 \\
{\rm SDSS J}1403+0006 & 0.1888 & 0.4730 & 0.83 & 213$\pm$17 & 0.611 & 0.126 & 0.553 & 0.573 & 10 \\
{\rm SDSS J}1436-0000 & 0.2852 & 0.8049 & 1.12 & 224$\pm$17 & 0.745 & 0.149 & 0.575 & 0.597 & 10,11\\
{\rm SDSS J}1443-0304 & 0.1338 & 0.4187 & 0.81 & 209$\pm$11 & 0.619 & 0.104 & 0.641 & 0.658 & 10,11 \\
{\rm SDSS J}1451-0239 & 0.1254 & 0.5203 & 1.04 & 223$\pm$14 & 0.698 & 0.126 & 0.718 & 0.735 & 10,11 \\
{\rm SDSS J}1525+3327 & 0.3583 & 0.7173 & 1.31 & 264$\pm$26 & 0.627 & 0.148 & 0.434 & 0.454 & 10,11 \\
{\rm SDSS J}1531-0105 & 0.1596 & 0.7439 & 1.71 & 279$\pm$14 & 0.733 & 0.120 & 0.734 & 0.753 & 10,11 \\
{\rm SDSS J}1538+5817 & 0.1428 & 0.5312 & 1.00 & 189$\pm$12 & 0.934 & 0.170 & 0.687 & 0.705 & 10,11 \\
{\rm SDSS J}1621+3931 & 0.2449 & 0.6021 & 1.29 & 236$\pm$20 & 0.773 & 0.165 & 0.535 & 0.556 & 10,11 \\
{\rm MG}1549+3047 & 0.11 & 1.17 & 1.15 & 227$\pm$18 & 0.745 & 0.153 & 0.865 & 0.878 & 12 \\
{\rm CY}2201-3201 & 0.32 & 3.90 & 0.41 & 130$\pm$20 & 0.810 & 0.270 & 0.825 & 0.764 & 2,4,5 \\
{\rm SDSS J}1432+6317 & 0.1230 & 0.6643 & 1.26 & 199$\pm$10 & 1.062 & 0.174 & 0.772 & 0.749 & 10,11 \\
{\rm SDSS J}2238-0754 & 0.1371 & 0.7126 & 1.27 & 198$\pm$11 & 1.081 & 0.185 & 0.761 & 0.736 & 10,11 \\
{\rm Q}0957+561 & 0.36 & 3.90 & 1.41 & 167$\pm$10 & 1.077 & 0.190 & 0.650 & 0.599 & 13 \\
\cline{1-10}
\sidehead{\centerline{\qquad\quad Systems with More than Two Images}}
{\rm SDSS J}0956+5100 & 0.2405 & 0.4700 & 1.32 & 318$\pm$17 & 0.436 & 0.073 & 0.441 & 0.459 & 1--9 \\
{\rm SDSS J}0959+0410 & 0.1260 & 0.5349 & 1.00 & 229$\pm$13 & 0.636 & 0.110 & 0.723 & 0.740 & 1--9 \\
{\rm SDSS J}1330-0148 & 0.0808 & 0.7115 & 0.85 & 195$\pm$10 & 0.746 & 0.124 & 0.855 & 0.868 & 1--9 \\
{\rm SDSS J}1402+6321 & 0.2046 & 0.4814 & 1.39 & 290$\pm$16 & 0.552 & 0.094 & 0.526 & 0.546 & 1--9 \\
{\rm SDSS J}1420+6019 & 0.0629 & 0.5352 & 1.04 & 206$\pm$5 & 0.818 & 0.113 & 0.858 & 0.869 & 1--9 \\
{\rm SDSS J}1627-0053 & 0.2076 & 0.5241 & 1.21 & 295$\pm$13 & 0.464 & 0.073 & 0.552 & 0.573 & 1--9 \\
{\rm SDSS J}2321-0939 & 0.0819 & 0.5324 & 1.57 & 245$\pm$7 & 0.873 & 0.124 & 0.816 & 0.829 & 1--9 \\
{\rm Q}0047-2808 & 0.4850 & 3.5950 & 1.34 & 229$\pm$15 & 0.853 & 0.157 & 0.741 & 0.738 & 1--9 \\
{\rm HST} 14176 & 0.8100 & 3.3990 & 1.41 & 224$\pm$15 & 0.938 & 0.175 & 0.599 & 0.587 & 1--9 \\
{\rm SDSS J}0029-0055 & 0.2270 & 0.9313 & 0.96 & 229$\pm$18 & 0.611 & 0.125 & 0.689 & 0.710 & 10,11 \\
{\rm SDSS J}0109+1500 & 0.2939 & 0.5248 & 0.69 & 251$\pm$19 & 0.366 & 0.073 & 0.389 & 0.407 & 10 \\
{\rm SDSS J}0728+3835 & 0.2058 & 0.6877 & 1.25 & 214$\pm$11 & 0.911 & 0.151 & 0.642 & 0.663 & 10,11 \\
{\rm SDSS J}0822+2652 & 0.2414 & 0.5941 & 1.17 & 259$\pm$15 & 0.582 & 0.101 & 0.536 & 0.557 & 10,11 \\
{\rm SDSS J}0841-3824 & 0.1159 & 0.6567 & 1.41 & 225$\pm$11 & 0.930 & 0.151 & 0.783 & 0.799 & 10,11 \\
{\rm SDSS J}0936+0913 & 0.1897 & 0.5880 & 1.09 & 243$\pm$12 & 0.616 & 0.101 & 0.624 & 0.645 & 10 \\
{\rm SDSS J}0946+1006 & 0.2219 & 0.6085 & 1.38 & 263$\pm$21 & 0.666 & 0.137 & 0.609 & 0.599 & 10,11 \\
{\rm SDSS J}1016+3859 & 0.1679 & 0.4349 & 1.09 & 247$\pm$13 & 0.596 & 0.100 & 0.570 & 0.589 & 10 \\
{\rm SDSS J}1020+1122 & 0.2822 & 0.5530 & 1.20 & 282$\pm$18 & 0.504 & 0.092 & 0.435 & 0.455 & 10 \\
{\rm SDSS J}1023+4230 & 0.1912 & 0.6960 & 1.41 & 242$\pm$15 & 0.804 & 0.144 & 0.669 & 0.808 & 10,11 \\
{\rm SDSS J}1029+0420 & 0.1045 & 0.6154 & 1.01 & 210$\pm$11 & 0.764 & 0.128 & 0.793 & 0.808 & 10 \\
{\rm SDSS J}1032+5322 & 0.1334 & 0.3290 & 1.03 & 296$\pm$15 & 0.392 & 0.065 & 0.560 & 0.576 & 10 \\
{\rm SDSS J}1103+5322 & 0.1582 & 0.7353 & 1.02 & 196$\pm$12 & 0.886 & 0.158 & 0.734 & 0.752 & 10,11 \\
{\rm SDSS J}1106+5228 & 0.0955 & 0.4069 & 1.23 & 262$\pm$13 & 0.598 & 0.098 & 0.733 & 0.748 & 10,11 \\
{\rm SDSS J}1134+6027 & 0.1528 & 0.4742 & 1.10 & 239$\pm$12 & 0.643 & 0.106 & 0.634 & 0.652 & 10 \\
{\rm SDSS J}1153+4612 & 0.1797 & 0.8751 & 1.05 & 226$\pm$15 & 0.686 & 0.127 & 0.737 & 0.756 & 10 \\
{\rm SDSS J}1416+5136 & 0.2987 & 0.8111 & 1.37 & 240$\pm$25 & 0.794 & 0.195 & 0.560 & 0.582 & 10,11 \\
{\rm SDSS J}1430+4105 & 0.2850 & 0.5753 & 1.52 & 322$\pm$32 & 0.489 & 0.116 & 0.448 & 0.468 & 10 \\
{\rm SDSS J}1636+4707 & 0.2282 & 0.6745 & 1.09 & 231$\pm$15 & 0.682 & 0.125 & 0.601 & 0.623 & 10 \\
{\rm PG}1115+080 & 0.3100 & 1.7200 & 1.21 & 281$\pm$25 & 0.511 & 0.113 & 0.730 & 0.745 & 14 \\
{\rm Q}2237+030 & 0.04 & 1.169 & 0.91 & 215$\pm$30 & 0.657 & 0.202 & 0.949 & 0.940 & 15 \\
{\rm B}1608+656 & 0.63 & 1.39 & 1.13 & 247$\pm$35 & 0.618 & 0.193 & 0.439 & 0.386 & 16 \\
{\rm SDSS J}0252+0039 & 0.2803 & 0.9818 & 1.04 & 164$\pm$12 & 1.290 & 0.253 & 0.639 & 0.599 & 10 \\
{\rm SDSS J}0405-0455 & 0.0753 & 0.8098 & 0.80 & 160$\pm$8 & 1.043 & 0.171 & 0.878 & 0.861 & 10 \\
{\rm SDSS J}2341+0000 & 0.186 & 0.807 & 1.44 & 207$\pm$13 & 1.122 & 0.203 & 0.712 & 0.681 & 10,11 \\
\enddata
\tablenotetext{References:\hskip0.2in} {(1) Treu \& Kooopmans (2002); (2) Koopmans \& Treu (2002);
(3) Treu \& Koopmans (2003); (4) Koopmans \& Treu (2003); (5) Treu \& Koopmans (2004); (6) Treu et al.
(2006); (7) Kooopmans et al. (2006); (8) Grillo et al. (2008); (9) Biesiada, Pi\'orkowska
\& Malec (2010); (10) Bolton et al. (2008); (11) Newton et al. (2011); (12) Leh\'ar et al. (1993);
(13) Young et al. (1980); (14) Tonry (1998); (15) Huchra et al. (1985); (16) Fassnacht et al. (1986)
}
\end{deluxetable}
\noindent comparison will be between $R_{\rm h}=ct$ and the concordance
model (with $\Omega_{\rm m}=0.27$), though we will also consider other
values of $\Omega_{\rm m}$, including the Einstein-de Sitter (E-deS) model
with $\Omega_{\rm m}=1$.
Related to the possible difficulty in using sources with $\mathcal{D}_{\rm obs} >1$
is the fact that the uncertainty $\sigma_{\mathcal{D}}$ in $\mathcal{D}_{\rm obs}$
carries significantly more weight when $\mathcal{D}_{\rm obs}\gtrsim 0.6$ than
elsewhere in its permitted range, because here big changes in $z_s$ produce
only very slight modifications to $\mathcal{D}_{\Lambda{\rm CDM}}$ and
$\mathcal{D}_{R_{\rm h}=ct}$. One therefore sees an increasing scatter
among the observed values of $\mathcal{D}_{\rm obs}$, as shown in
figures~1 and 2. In order to fully understand the impact of all of these
issues, we will analyze the quality of the theoretical fit for both the full
sample and a reduced sample with $\mathcal{D}_{\rm obs}<1$, and
in each case also a sub-sample of 2-image systems only. Figures~1 and 2
show the results of analyzing all lensing systems with 2 images only. The
figures corresponding to other sample selection criteria are very similar.
\begin{figure}[hp]
\centerline{\includegraphics[angle=0,scale=0.7]{f1.eps}}
\vskip-0.2in
\caption{Observed value of $\mathcal{D}$ versus that predicted in the
$R_{\rm h}=ct$ Universe for all lensing systems with 2 images only.
A perfect fit would correspond to the dashed
diagonal line. The optimized value of $f_{\rm SIS}$ in this case
is $1.02$, and the reduced $\chi^2$ is $1.22$, with $34-1=33$ degrees
of freedom.}
\end{figure}
\begin{figure}[hp]
\centerline{\includegraphics[angle=0,scale=0.7]{f2.eps}}
\vskip-0.2in
\caption{Same as figure~1, except now for the concordance $\Lambda$CDM
model, with $\Omega_{\rm m}=0.27$. The optimized value of $f_{\rm SIS}$
is $1.004$, and the reduced $\chi^2$ for this fit is $1.24$, with
$34-1=33$ degrees of freedom.}
\end{figure}
As we have already noted, the power to optimize model parameters,
such as $\Omega_{\rm m}$ in a multi-parameter context, is very
limited (Biesiada et al. 2010; Cao et al. 2012). This will become
quite apparent in our discussion below, where we compare the
quality of the fit for several variations of $\Lambda$CDM. For each
model we consider here, we will therefore optimize the fit using only
$f_{\rm SIS}$ as a free parameter, which we do by minimizing the
$\chi^2$ function
\begin{equation}
\chi^2(f_{\rm SIS})=\sum_i {\left(\mathcal{D}_{{\rm obs},i}[f_{\rm SIS}]-
\mathcal{D}_{{\rm th},i} \right)^2\over \sigma_{\mathcal{D},i}^2}\;,
\end{equation}
where the index $i$ runs over all lens systems in the sample, `th' stands
for either $\Lambda$CDM or $R_{\rm h}=ct$, and $\sigma_{\mathcal{D},i}^2$
is the variance of $\mathcal{D}_{{\rm obs},i}$ calculated from equation~(9).
\section{Discussion}
We have used the data shown in Table 1 to compare 3 variations of $\Lambda$CDM and
the $R_{\rm h}=ct$ Universe, though always for a flat universe ($k=0$) and
$w_{\rm de}=-1$. A summary of the results is provided in Tables 2 and 3 for the whole
sample (of 69 sources), and in Tables 4 and 5 for a reduced sample with
$\mathcal{D}_{\rm obs} <1$ only (63 systems). We note,
first of all, that the optimized value of $f_{\rm SIS}$ is very close to 1 in
every case, in complete agreement with earlier findings, e.g., by Treu et al. (2006),
and van de Ven et al. (2003). As such, we do not find any possible dependence of
the so-called bulge-halo `conspiracy' on the assumed cosmological model.
\begin{deluxetable}{lccc}
\tablewidth{277pt}
\tabletypesize{\footnotesize}
\tablecaption{Model Comparison for the Whole Sample}\tablenum{2}
\tablehead{Model&\colhead{$\qquad\Omega_{\rm m}\qquad$}&\colhead{$\qquad f_{\rm SIS}\qquad$}&
\colhead{$\qquad\chi^2_{\rm dof}\qquad$}
} \startdata
$R_{\rm h}=ct$ & .... & 1.023 & 1.22 \\
$\Lambda$CDM$^a$ & 0.20& 1.00 & 1.22 \\
{\rm Concordance}$^a$ & 0.27 & 1.01 & 1.24 \\
Einstein-de Sitter\qquad & 1.00 & 1.046 & 1.33 \\
\enddata
\tablenotetext{\null\hbox{$^a$}} {Assumes a cosmological constant with $w_\Lambda=-1$.
The concordance model is $\Lambda$CDM with $\Omega_{\rm m}=0.27$.
}
\end{deluxetable}
\begin{deluxetable}{lccc}
\tablewidth{277pt}
\tabletypesize{\footnotesize}
\tablecaption{Model Comparison for Two-image Sources}\tablenum{3}
\tablehead{Model&\colhead{$\qquad\Omega_{\rm m}\qquad$}&\colhead{$\qquad f_{\rm SIS}\qquad$}&
\colhead{$\qquad\chi^2_{\rm dof}\qquad$}
} \startdata
$R_{\rm h}=ct$ & .... & 1.033 & 1.27 \\
$\Lambda$CDM$^a$ & 0.20& 1.01& 1.28 \\
{\rm Concordance}$^a$ & 0.27 & 1.02 & 1.29 \\
Einstein-de Sitter\qquad & 1.00 & 1.059 & 1.33 \\
\enddata
\tablenotetext{\null\hbox{$^a$}} {Assumes a cosmological constant with $w_\Lambda=-1$.
The concordance model is $\Lambda$CDM with $\Omega_{\rm m}=0.27$.
}
\end{deluxetable}
\begin{deluxetable}{lccc}
\tablewidth{277pt}
\tabletypesize{\footnotesize}
\tablecaption{Model Comparison for all $\mathcal{D}_{\rm obs}<1$}\tablenum{4}
\tablehead{Model&\colhead{$\qquad\Omega_{\rm m}\qquad$}&\colhead{$\qquad f_{\rm SIS}\qquad$}&
\colhead{$\qquad\chi^2_{\rm dof}\qquad$}
} \startdata
$R_{\rm h}=ct$ & .... & 1.01 & 0.99 \\
$\Lambda$CDM$^a$ & 0.20& 0.99& 0.99 \\
{\rm Concordance}$^a$ & 0.27 & 1.00 & 1.00 \\
Einstein-de Sitter\qquad & 1.00 & 1.03 & 1.09 \\
\enddata
\tablenotetext{\null\hbox{$^a$}} {Assumes a cosmological constant with $w_\Lambda=-1$.
The concordance model is $\Lambda$CDM with $\Omega_{\rm m}=0.27$.
}
\end{deluxetable}
\begin{deluxetable}{lccc}
\tablewidth{277pt}
\tabletypesize{\footnotesize}
\tablecaption{Model Comparison for 2-images and $\mathcal{D}_{\rm obs}<1$}\tablenum{5}
\tablehead{Model&\colhead{$\qquad\Omega_{\rm m}\qquad$}&\colhead{$\qquad f_{\rm SIS}\qquad$}&
\colhead{$\qquad\chi^2_{\rm dof}\qquad$}
} \startdata
$R_{\rm h}=ct$ & .... & 1.02 & 0.92 \\
$\Lambda$CDM$^a$ & 0.20& 1.00& 0.92 \\
{\rm Concordance}$^a$ & 0.27 & 1.00 & 0.93 \\
Einstein-de Sitter\qquad & 1.00 & 1.05 & 0.99 \\
\enddata
\tablenotetext{\null\hbox{$^a$}} {Assumes a cosmological constant with $w_\Lambda=-1$.
The concordance model is $\Lambda$CDM with $\Omega_{\rm m}=0.27$.
}
\end{deluxetable}
We have compared the model fits using both the full sample of 69 entries in
Table 1 and, separately, using only the sub-sample of 34 2-image systems.
The quality of the fit, for every model we considered, is actually somewhat better
for the former. This may simply be a reflection of the fact that, though
technically an isothermal sphere should produce only 2 images, the other
possible effects described in \S~2 could be significant enough to result in
a considerable scatter about the average value $f_{\rm SIS}\sim 1$
for individual systems, that dwarfs all the other possible
sources of error in calculating $\mathcal{D}_{\rm obs}$. Indeed, Cao et
al. (2012) have argued that $\sigma_f$ could be as large as $\sim 20\%$,
and even though such a large scatter was not seen by Treu et al. (2006)
and van de Ven et al. (2003), they nonetheless did report an rms deviation
of at least $6-7\%$.
Given the universal result $f_{\rm SIS}\sim 1$, the entries
in column 6 of Table 1 are shown for only one value ($f_{\rm SIS}
=1.02$) of this fraction, corresponding to the optimized fit for the
$R_{\rm h}=ct$ model using only the 2-image lens systems (with
$\mathcal{D}_{\rm obs} <1$). Also, column 9 shows the entries for
$\mathcal{D}_{\Lambda{\rm CDM}}$
only for the concordance model, i.e., for $\Omega_{\rm m}=0.27$. These
values change somewhat for other choices of $\Omega_{\rm m}$, but not enough
to warrant showing all of them here.
Figures~1 and 2 demonstrate graphically how the observed values of $\mathcal{D}$
compare with those predicted by $R_{\rm h}=ct$ and the concordance model using
only the sub-sample of 2-image lens systems, but for all values of
$\mathcal{D}_{\rm obs}$. The optimized value of
$f_{\rm SIS}$ is $1.02$ for the former, and $1.01$ for the latter,
and the reduced $\chi^2_{\rm dof}$ (with $34-1=33$ degrees of freedom) is quite
similar for these two cases, i.e., $1.22$ for the former versus $1.24$ for the
latter. It is quite evident from these figures that the scatter in
$\mathcal{D}_{\rm obs}$ about the theoretical curves (the straight
dashed lines in these plots) increases significantly as $D_A(z_l,z_s)
\rightarrow D_A(0,z_s)$. In other words, it appears that measuring
$\mathcal{D}$ becomes progressively less precise as the distance to the
gravitational lens becomes a smaller and smaller fraction of the distance
to the quasar source. This may simply have to do with the fact that $\theta_{\rm E}$
changes less and less for large values of $z_s/z_l$ so, for the same
error in the Einstein angle, one gets less precision in the
measurement of $\mathcal{D}_{\rm obs}/\mathcal{D}_{\rm th}$.
The principal results of this paper are summarized in Tables~2, 3, 4 and 5.
The first two show how well the 4 models considered here fit the complete
sample of 69 lens systems in Table 1 (for all values of $\mathcal{D}_{\rm obs}$),
whereas the latter two give the corresponding results for the reduced
sample with $\mathcal{D}_{\rm obs} <1$. Based
on the general trends emerging from these numbers, it is safe to draw
the following conclusions: (1) Even though the power of $\mathcal{D}$
to discriminate between different values of $\Omega_{\rm m}$ in $\Lambda$CDM
is quite limited, this analysis indicates that values larger than
$\Omega_{\rm m}=0.27$ probably don't work as well as those below it,
though the differences in $\chi^2_{\rm dof}$ are still too small
to draw any firm conclusions. And (2), the $R_{\rm h}=ct$ fits the
strong gravitational lens data at least as well as $\Lambda$CDM.
Still, the differences between $R_{\rm h}=ct$ and $\Lambda$CDM are
small enough that one cannot choose one model over the other based
solely on this analysis, using the current sample of strong
gravitational lens systems. The other tests we have completed thus far,
using, e.g., cosmic chronometers (Melia \& Maier 2013), high-$z$
quasars (Melia 2013), and gamma ray bursts (Wei et al. 2013), have
all resulted in a clear preference for $R_{\rm h}=ct$ over
$\Lambda$CDM using statistical model selection tools. The
analysis of the strong gravitational lensing data does not
result in a comparable outcome yet, though it too does not
provide any evidence that the standard model is a better
fit to these observations than $R_{\rm h}=ct$.
\section{Monte Carlo Simulations with a Mock Sample}
In order to provide a detailed quantitative assessment of what kind of strong
lensing data are necessary to really distinguish the $R_{\rm h}=ct$ Universe
from the standard $\Lambda$CDM model, we will here produce mock samples
of strong gravitational lenses based on the current measurement accuracy.
Several information criteria commonly used to differentiate between different
cosmological models (see, e.g., Melia \& Maier 2013, and references cited
therein) include the Akaike Information Criterion, ${\rm AIC}=\chi^{2}+2n$,
where $n$ is the number of free parameters (Liddle 2007),
the Kullback Information Criterion, ${\rm KIC}=\chi^{2}+3n$ (Cavanaugh
2004), and the Bayes Information Criterion,
${\rm BIC}=\chi^{2}+(\ln N)n$, where $N$ is the number of data points
(Schwarz 1978). In the case of AIC, with ${\rm AIC}_\alpha$
characterizing model $\mathcal{M}_\alpha$,
the unnormalized confidence that this model is true is the Akaike
weight $\exp(-{\rm AIC}_\alpha/2)$. Model $\mathcal{M}_\alpha$ has likelihood
\begin{equation}
P(\mathcal{M}_\alpha)= \frac{\exp(-{\rm AIC}_\alpha/2)}
{\exp(-{\rm AIC}_1/2)+\exp(-{\rm AIC}_2/2)}
\end{equation}
of being the correct choice in this one-on-one comparison. Thus, the difference
$\Delta \rm AIC \equiv {\rm AIC}_2\nobreak-{\rm AIC}_1$ determines the extent
to which $\mathcal{M}_1$ is favoured over~$\mathcal{M}_2$. For Kullback
and Bayes, the likelihoods are defined analogously. In using the model selection tools,
the outcome $\Delta\equiv$ AIC$_1-$ AIC$_2$ (and analogously for KIC and BIC)
is judged `positive' in the range $\Delta=2-6$, `strong' for $\Delta=6-10$,
and `very strong' for $\Delta>10$.
In this section, we will estimate the sample size required to significantly strengthen the
evidence in favour of $R_{\rm h}=ct$ or $\Lambda$CDM, by conservatively
seeking an outcome even beyond $\Delta\simeq11.62$, i.e., we will see what is
required to produce a likelihood $\sim 99.7\%$ versus $\sim 0.3\%$,
corresponding to a $3\sigma$ confidence level.
We will consider two cases: one in which the background cosmology is
assumed to be $\Lambda$CDM, and a second in which it is $R_{\rm h}=ct$,
and we will attempt to estimate the number of strong gravitational lenses
required in each case in order to rule out the alternative (incorrect)
model at a $\sim 99.7\%$ confidence level. The synthetic strong gravitational lenses
are each characterized by a set of parameters denoted as ($z_{l}$, $z_{s}$,
$\sigma_{SIS}$, $\theta_{\rm E}$). We generate the synthetic sample using the
following procedure:
1. The simulations are carried out based on the current lens measurements.
We assign the lens redshift $z_{l}$ uniformly between $0.1$
and $1.1$, the source redshift $z_{s}$ uniformly between $1.5$ and $3.5$, and
the velocity dispersion $\sigma_{SIS}$ uniformly
between $100$ and $300$ km $\rm s^{-1}$,
as Paraficz \& Hjorth (2009) did in their simulations.
2. With the mock $z_{l}$, $z_{s}$ and $\sigma_{SIS}$, we first infer $\theta_{\rm E}$
from Equation~(7) corresponding either to the $R_{\rm h}=ct$ Universe (\S~5.1) or
a flat $\Lambda$CDM cosmology with $\Omega_{\rm m}=0.27$ (\S~5.2). We then assign a
deviation ($\Delta \theta_{\rm E}$) to the $\theta_{\rm E}$ value, i.e.,
we infer $\theta'_{\rm E}$ from a normal distribution whose center value is $\theta_{\rm E}$,
with a dispersion $\sigma=0.12\;\theta_{\rm E}$. The typical value of $\sigma$ is taken from
the current (observed) sample, which yields a mean and median deviation of $\sigma=0.15\;\theta_{\rm E}$
and $\sigma=0.11\;\theta_{\rm E}$, respectively. We constrain the mock sample to easily
detectable systems, so we include in the simulations only lenses with $\theta'_{\rm E}$
larger than $0.1$ arcsec.
\begin{figure}[h]
\centerline{\includegraphics[angle=0,scale=1.0]{f3.eps}}
\vskip-0.2in
\caption{Left: ``Observed" values of $\mathcal{D}$ versus that predicted in the
best-fit $\Lambda$CDM model. A perfect fit would correspond to the dashed
diagonal line. The optimized values of $\Omega_{\rm m}$ and $\Omega_{\rm de}$
in this case are $0.45$ and $0.69$, respectively, and the reduced $\chi^2$ is $1.32$, with $300-2=298$ degrees
of freedom. Right: The $1\sigma-3\sigma$ contours corresponding to the parameters
$\Omega_{\rm m}$ and $\Omega_{\rm de}$ in the best-fit $\Lambda$CDM model, using the simulated sample with
300 lens systems, assuming $R_{\rm h}=ct$ as the background cosmology.}
\end{figure}
3. Assign observational errors to $\sigma_{SIS}$ and $\theta'_{\rm E}$.
Since both the observed errors $\sigma_{\sigma_{SIS}}$ and $\sigma_{\theta_{\rm E}}$
are about $5\%$ of $\sigma_{SIS}$ and $\theta'_{\rm E}$, we will assign the dispersions
$\sigma_{\sigma_{SIS}}=0.05\;\sigma_{SIS}$ and $\sigma_{\theta_{\rm E}}=0.05\;\theta'_{\rm E}$
to the synthetic sample.
This sequence of steps is repeated for each lens system in the sample, which
is enlarged until the likelihood criterion discussed above is reached. As with
the real 60-lens sample, we optimize the model fits by minimizing the $\chi^{2}$ function
in Equation~(12).
\subsection{Assuming $R_{\rm h}=ct$ as the Background Cosmology}
We have found that a sample of at least 300 strong gravitational lenses is required
in order to rule out $\Lambda$CDM at the $\sim 99.7 \%$ confidence level, if the
background cosmology is in fact $R_{\rm h}=ct$.
The left-hand panel of Figure~3 show how the ``observed" values of $\mathcal{D}$ compare
with those predicted by the best-fit $\Lambda$CDM model using the simulated sample with
300 lens systems, assuming $R_{\rm h}=ct$ as the background cosmology. The
optimized parameters corresponding to the best-fit $\Lambda$CDM model for
these simulated data are displayed in the right-hand panel of Figure~3. To allow for the greatest
flexibility in this fit, we relax the assumption of flatness and allow
$\Omega_{\rm de}$ to be a free parameter along with $\Omega_{\rm m}$. The right-hand panel of Figure~3
shows the 2-D plots for the $1\sigma-3\sigma$ confidence regions for $\Omega_{\rm m}$
and $\Omega_{\rm de}$. The best-fit values for $\Lambda$CDM using the simulated sample with 300 lens systems
in the $R_{\rm h}=ct$ Universe are $\Omega_{\rm m}=0.45_{-0.24}^{+0.37}$ $(1\sigma)$ and
$\Omega_{\rm de}=0.69_{-0.06}^{+0.07}$ $(1\sigma)$, with a $\chi^2$ per degree of freedom of $\chi^2_{\rm dof}=394.04/298=1.32$.
\begin{figure}[hp]
\centerline{\hskip 0.5in\includegraphics[angle=0,scale=0.7]{f4.eps}}
\vskip-0.2in
\caption{``Observed" values of $\mathcal{D}$ versus that predicted in the
$R_{\rm h}=ct$ universe, using a sample of 300 lens systems, simulated with $R_{\rm h}=ct$ as the
background cosmology. A perfect fit would correspond to the dashed diagonal line.}
\end{figure}
In Figure~4, we show how the ``observed" values of $\mathcal{D}$ compare
with those predicted in the $R_{\rm h}=ct$ universe. Note that there are no free parameters
in $R_{\rm h}=ct$. With 300 degrees of freedom, the reduced $\chi^2$ is $\chi^2_{\rm dof}=
394.01/300=1.31$.
Since the number $N$ of data points in the sample is now much greater than one, the
most appropriate information criterion to use is the BIC. The logarithmic penalty
in this model selection tool strongly suppresses overfitting if $N$ is large
(the situation we have here, which is deep in the asymptotic regime). With $N=300$,
our analysis of the simulated sample shows that the BIC would favour the $R_{\rm h}=ct$
Universe over $\Lambda$CDM by the aforementioned likelihood of $99.7\%$ versus only $0.3\%$
(i.e., the prescribed $3\sigma$ confidence limit).
\begin{figure}[hp]
\centerline{\includegraphics[angle=0,scale=1.0]{f5.eps}}
\vskip-0.2in
\caption{Same as Figure~3, except now with a flat $\Lambda$CDM as the (assumed)
background cosmology. The simulated model parameter was $\Omega_{\rm m}=0.27$.}
\end{figure}
\subsection{Assuming $\Lambda$CDM as the Background Cosmology}
In this case, we assume that the background cosmology is $\Lambda$CDM,
and seek the minimum sample size to rule out $R_{\rm h}=ct$ at the
$3\sigma$ confidence level. We have found that a minimum of 200 strong
gravitational lenses are required to achieve this goal. To allow for the greatest flexibility
in the $\Lambda$CDM fit, here too we relax the assumption of flatness and
allow $\Omega_{\rm de}$ to be a free parameter along with $\Omega_{\rm m}$.
The left-hand panel of Figure~5 demonstrates how the ``observed" values of $\mathcal{D}$ compare
with those predicted by the best-fit $\Lambda$CDM model using the simulated sample with
200 lens systems, assuming $\Lambda$CDM as the background cosmology.
In the right-hand panel of Figure~5, we show 2-D plots of the $1\sigma-3\sigma$ confidence regions
for $\Omega_{\rm m}$ and $\Omega_{\rm de}$. The best-fit values for $\Lambda$CDM using this simulated sample
with 200 lens systems are $\Omega_{\rm m}=0.47_{-0.28}^{+0.43}$ $(1\sigma)$ and
$\Omega_{\rm de}=0.67_{-0.07}^{+0.07}$ $(1\sigma)$, with a $\chi^2$ per degree of freedom of $\chi^2_{\rm dof}=259.90/198=1.31$.
The ``observed" values of $\mathcal{D}$ compared with those predicted in the $R_{\rm h}=ct$ universe
are shown in Figure~6. The dashed diagonal line denotes the perfect fit.
With 200 degrees of freedom, the reduced $\chi^2$ is $\chi^2_{\rm dof}=281.86/200=1.41$.
With $N=200$, our analysis of the simulated sample shows that in this case the BIC would favour $\Lambda$CDM
over $R_{\rm h}=ct$ by the aforementioned likelihood of $99.7\%$ versus only $0.3\%$
(i.e., the prescribed $3\sigma$ confidence limit).
\begin{figure}[hp]
\centerline{\hskip 0.5in\includegraphics[angle=0,scale=0.7]{f6.eps}}
\vskip-0.2in
\caption{Same as Figure~4, except now with $\Lambda$CDM as the (assumed)
background cosmology.}
\end{figure}
\section{Conclusions}
The use of individual gravitational lenses to measure cosmological parameters
has been with us for over a decade now (see, e.g., Treu et al. 2006; Grillo et al. 2008;
Biesiada et al. 2010; and Biesiada et al. 2011) and the results, though less
precise than those from other kinds of data, have nonetheless been consistent with
the basic $\Lambda$CDM cosmology. Our principal goal in this paper has been to
carry out a comparative analysis of the available galaxy lens data using both $\Lambda$CDM
and the $R_{\rm h}=ct$ Universe, which may be thought of as $\Lambda$CDM with the
additional constraint $p=-\rho/3$ on its total equation of state. This analysis
has been motivated by other kinds of study showing that model selection tools
tend to favor $R_{\rm h}=ct$ over the standard model.
Insofar as the strong gravitational lenses are concerned, both $R_{\rm h}=ct$
and $\Lambda$CDM fit the data quite well. We have not found an inconsistency
between these results and those of previous studies using a variety of observations
at low and high redshifts. We may already be able to rule out values of $\Omega_{\rm m}$
much greater than the concordance value of $0.27$, but apparently not smaller than
this. Where things stand now is that gravitational lens data do not provide
conclusive evidence in favor of either model.
As much as we have learned about these lens systems, several sources of
uncertainty remain, including the need to properly model the mass distribution within
the lens and to better understand the source of the so-called bulge-halo conspiracy.
These errors appear to be more debilitating for lens systems with large values
of $z_s/z_l$, so a priority for future work ought to be the search for lens
systems with small distances between the lens and the source compared with
distances between the lens and observer. We have found that systems with
correspondingly small values of $\mathcal{D}_{\rm obs}$ provide significantly
greater precision in the measurement of cosmological parameters than those
with values approaching $1$ (see Figures~1 and 2).
Given the limitations of the current sample, we have also investigated how
big the catalog of measured lensing galaxies has to be in order for us to rule
out one (or more) of these models. We have considered
two synthetic samples with characteristics similar to those of the current observed
lens systems, one based on a $\Lambda$CDM background cosmology, the other on
$R_{\rm h}=ct$. From the analysis of these simulated lenses, we have
estimated that a sample of about 200 systems would be needed to rule out
$R_{\rm h}=ct$ at a $\sim 99.7\%$ confidence level if the real cosmology
were in fact $\Lambda$CDM, while a sample of at least 300 systems would
be needed to similarly rule out $\Lambda$CDM if the background cosmology
were instead $R_{\rm h}=ct$. The difference in required sample size
results from $\Lambda$CDM's greater flexibility in fitting the data, since
it has a larger number of free parameters.
Looking to the future, a convincing demonstration that $R_{\rm h}=ct$ is
the correct cosmology would provide sweeping new capabilities for
carrying out structural and evolutionary studies of lensing galaxies,
for the very simple reason that $\mathcal{D}$ in this cosmology is completely
independent of any model parameters, such as $H_0$ and $\Omega_{\rm m}$.
The quantity $\mathcal{D}_{\rm th}$ in this spacetime depends solely on
the observed values of $z_l$ and $z_s$ which, as we have noted in this paper,
are measured with much higher precision than any of the other lens-dependent
parameters. Imagine, therefore, the probative power of such measurements
on a determination of individual $f_{\rm SIS}$'s or, even better, on
providing the capability to probe the mass structure within these galaxies.
As of now, early-type galaxies appear to be well approximated by singular
isothermal ellipsoids. But this mass-density profile differs significantly
from cosmologically motivated ones (see, e.g., Navarro et al. 1997;
Moore et al. 1998), and also appears to require fine-tuning between the
distributions of baryonic and dark matter. This awkward situation begs
the question of how these structures formed in the first place. The
use of gravitational lensing within the $R_{\rm h}=ct$ framework may
finally break this deadlock and explain the origin of the bulge-halo
conspiracy.
\vskip-0.2in
\acknowledgments
We are grateful to the anonymous referee for providing a thoughtful review
and for suggesting several improvements to the manuscript.
This work is partially supported by the National Basic Research Program (``973" Program) of China
(Grants 2014CB845800 and 2013CB834900), the National Natural Science Foundation of China
(grants Nos. 11322328,11373068, 11173064 and 11233008), the One-Hundred-Talents Program
and the Youth Innovation Promotion Association, and the Strategic Priority Research Program
``The Emergence of Cosmological Structures'' (Grant No. XDB09000000) of the Chinese Academy
of Sciences, and the Natural Science Foundation of Jiangsu Province. F.M. is also grateful to
Amherst College for its support through a John Woodruff Simpson Lectureship, and to Purple
Mountain Observatory in Nanjing, China, for its hospitality while this work was being carried out.
This work was partially supported by grant 2012T1J0011 from The Chinese Academy of
Sciences Visiting Professorships for Senior International Scientists, and grant
GDJ20120491013 from the Chinese State Administration of Foreign Experts Affairs.
|
\chapter*{Introduction}
In this book, we study Gromov's metric geometric theory
\cite{Gromov}*{\S 3$\frac{1}{2}$}
on the space of metric measure spaces,
based on the idea of concentration of measure phenomenon
due to L\'evy and Milman.
Although most of the details are omitted in the original
article \cite{Gromov}*{\S 3$\frac{1}{2}$},
we present complete and detailed proofs for some main parts,
in which we prove several claims
that are not mentioned in any literature.
We also discuss concentration with a lower bound of curvature,
which is originally studied in \cite{FS}.
The concentration of measure phenomenon was first discovered by P.~L\'evy
\cite{Levy}
and further put forward by V.~Milman \cite{Mil:Dvoretzky,Mil:heritage}.
It has many applications in various areas, such as, geometry, analysis,
probability theory, and discrete mathematics
(see \cite{Ldx:book} and the references therein).
The phenomenon is stated as that
any $1$-Lipschitz continuous function is close to a constant
on a domain with almost full measure,
which is often observed for high-dimensional spaces.
As a most fundamental example, we observe it
in the high-dimensional unit spheres $S^n(1) \subset \field{R}^{n+1}$,
i.e., any $1$-Lipschitz continuous function on $S^n(1)$ is close to a constant
on a domain with almost full measure if $n$ is large enough.
In general, it is described for a sequence of
metric measure spaces.
In this book, we assume a metric measure space, an \emph{mm-space} for short,
to be a triple $(X,d_X,\mu_X)$, where
$(X,d_X)$ is a complete separable metric space
and $\mu_X$ a Borel
probability\footnote{In \cite{Gromov}*{\S 3$\frac{1}{2}$},
the measures of mm-spaces are not necessarily probability.
However, all our proofs easily extend to the case of non-probability
mm-spaces.} measure on $X$.
A sequence of mm-spaces $X_n$, $n=1,2,\dots$, is called a \emph{L\'evy family}
if
\[
\lim_{n\to\infty} \inf_{c\in\field{R}} \mu_{X_n}(|f_n-c| > \varepsilon) = 0
\]
for any sequence of $1$-Lipschitz continuous functions $f_n : X_n \to \field{R}$,
$n=1,2,\dots$, and for any $\varepsilon > 0$.
The sequence of the unit spheres $S^n(1)$, $n=1,2,\dots$, is a L\'evy family,
where the measure on $S^n(1)$ is taken to be the Riemannian volume measure
normalized as the total measure to be one.
One of central themes in this book is the study of the observable distance.
The \emph{observable distance $\dconc(X,Y)$ between
two mm-spaces $X$ and $Y$}
is, roughly speaking, the difference between $1$-Lipschitz functions on $X$
and those on $Y$ (see Definition \ref{defn:obs-dist} for the precise
definition).
A sequence of mm-spaces is a L\'evy family if and only if
it $\dconc$-converges to a one-point mm-space, where we note that
any $1$-Lipschitz function on a one-point mm-space is constant.
Thus, $\dconc$-convergence of mm-spaces
can be considered as a generalization of the L\'evy property.
We call $\dconc$-convergence of mm-spaces \emph{concentration of mm-spaces}.
A typical example of a concentration $X_n\to Y$ is obtained
by a fibration
\[
F_n \to X_n \to Y
\]
such that $\{F_n\}_{n=1}^\infty$ is a L\'evy family,
which example makes us to notice that
concentration of mm-spaces is an analogue of collapsing
of Riemannian manifolds. Concentration is strictly weaker than
measured Gromov-Hausdorff convergence
and is more suitable for the study of a sequence of manifolds whose dimensions
are unbounded.
Although $\dconc$ is not easy to investigate,
we have a more elementary distance, called the \emph{box distance},
between mm-spaces.
The box distance function is fit for well-known measured Gromov-Hausdorff
convergence of mm-spaces (see Remark \ref{rem:box-mGH}).
Concentration of mm-spaces is rephrased as convergence of associated pyramids
using the box distance function,
where a \emph{pyramid} is a family of mm-spaces
that forms a directed set with respect to some natural order relation
between mm-spaces, called the \emph{Lipschitz order}
(see Definitions \ref{defn:dom} and \ref{defn:pyramid}).
We have a metric $\rho$ on the set of pyramids, say $\Pi$,
induced from the box distance function (see Definition \ref{defn:metric-Pi}
and \cite{Shioya:mmlim}).
Each mm-space $X$ is associated with the pyramid, say $\mathcal{P}_X$,
consisting of all descendants of
the mm-space (i.e., smaller mm-spaces with respect to the Lipschitz order).
Denote the set of mm-spaces by $\mathcal{X}$.
We prove that the map
\[
\iota : \mathcal{X} \ni X \longmapsto \mathcal{P}_X \in \Pi
\]
is a $1$-Lipschitz continuous topological embedding map with respect to
$\dconc$ and $\rho$.
This means that concentration of mm-spaces is expressed
only by the box distance function,
since $\rho$ is induced from the box distance function.
We also prove that $\Pi$ is a compactification of $\mathcal{X}$ with $\dconc$.
Such a concrete compactification is far more valuable
than just an abstract one.
It is also interesting to study a sequence of mm-spaces
that $\dconc$-diverges but have proper asymptotic behavior.
A sequence of mm-spaces $X_n$, $n=1,2,\dots$, is said to be \emph{asymptotic}
if the associated pyramid $\mathcal{P}_{X_n}$ converges in $\Pi$.
We say that a sequence of mm-spaces \emph{asymptotically concentrates}
if it is a $\dconc$-Cauchy sequence.
Any asymptotically concentrating sequence of mm-spaces
is asymptotic.
For example, the sequence of the Riemannian product spaces
\[
S^1(1) \times S^2(1) \times \dots \times S^n(1),
\quad n=1,2,\dots,
\]
$\dconc$-diverges and asymptotically concentrates
(see Example \ref{ex:prod-sph}).
The sequence of the spheres $S^n(\sqrt{n})$ of radius $\sqrt{n}$,
$n=1,2,\dots$, does not even asymptotically concentrate
but is asymptotic (see Theorem \ref{thm:sphere-Gaussian},
Corollary \ref{cor:sphere-Gaussian}, and \cite{Shioya:mmlim}).
One of main theorems in this book states that the map $\iota : \mathcal{X} \to \Pi$
extends to the $\dconc$-completion of $\mathcal{X}$, so that
the space $\Pi$ of pyramids is also a compactification of
the $\dconc$-completion of the space $\mathcal{X}$ of mm-spaces
(see Theorem \ref{thm:emb-pyramid}).
Let $\gamma^n$ denote the standard Gaussian measure on $\field{R}^n$.
Then, the associated pyramids $\mathcal{P}_{S^n(\sqrt{n})}$ and $\mathcal{P}_{(\field{R}^n,\gamma^n)}$
both converge to a common pyramid as $n\to\infty$
(see Theorem \ref{thm:sphere-Gaussian} and \cite{Shioya:mmlim}),
which can be thought as a generalization of
the Maxwell-Boltzmann distribution law
(or the Poincar\'e limit theorem).
The spectral property is deeply related with
the asymptotic behavior of a sequence of mm-spaces.
The \emph{spectral compactness} of a family of mm-spaces
is defined by the Gromov-Hausdorff compactness
of the energy sublevel sets of $L_2$ functions
(see Definition \ref{defn:spec-cpt})
and is closely related with the notion of
asymptotic compactness of Dirichlet energy forms
(see \cite{KS}).
For a family of compact Riemannian manifolds,
it is equivalent to the discreteness of the limit set of the spectrums
of the Laplacians of the manifolds (see Proposition \ref{prop:spec-cpt}).
(In this book, manifolds may have nonempty boundary.)
We prove that any spectrally compact and asymptotic sequence of mm-spaces
is asymptotically concentrates if the observable diameter
is bounded from above (see Theorem \ref{thm:spec-conc}).
We say that a sequence of mm-spaces \emph{spectrally concentrates}
if it spectrally compact and asymptotically concentrates.
For example, let
\[
X_n := F_1 \times F_2 \times \dots \times F_n
\]
be the Riemannian product of compact Riemannian manifolds $F_n$, $n=1,2,\dots$.
If $\lambda_1(F_n)$ diverges to infinity as $n\to\infty$,
then $\{X_n\}$ spectrally concentrates (see Corollary \ref{cor:prod-spec-conc}).
There is a notion of dissipation for a sequence of mm-spaces,
which is opposite to concentration and means that
the mm-spaces disperse into many small pieces far apart each other.
A sequence of mm-spaces \emph{$\delta$-dissipates}, $\delta > 0$,
if and only if
any limit of the associated pyramids contains
all mm-spaces with diameter $\le \delta$.
The sequence \emph{infinitely dissipates} if and only if
the associated pyramid converges to the space of mm-spaces
(see Proposition \ref{prop:dissipate}).
On one hand, for a disconnected mm-space $F$,
the sequence of the $n^{th}$ power product spaces $F^n$, $n=1,2,\dots$,
with $l_\infty$ metric $\delta$-dissipates for some $\delta > 0$
(see Proposition \ref{prop:disconn-dissipation}).
On the other hand, the non-dissipation theorem
(Theorem \ref{thm:non-dissipation}) states that
the sequence $\{F^n\}$ does not $\delta$-dissipate for any $\delta > 0$
if $F$ is connected and locally connected.
The proof of the non-dissipation theorem relies on the study
of the obstruction condition for dissipation.
For example, a sequence of compact Riemannian manifolds $X_n$
does not dissipate if $\lambda_1(X_n)$ is bounded away from zero
(see Corollary \ref{cor:prod-diss}),
which is one of essential statements in the proof of
the non-dissipation theorem.
It is interesting to study the relation between curvature
and concentration.
The concept of Ricci curvature bounded below is generalized
to the \emph{curvature-dimension condition} for an mm-space
by Lott-Villani-Sturm \cites{LV,Sturm:geoI,Sturm:geoII}
via the optimal mass-transport theory.
We prove that if a sequence of mm-spaces satisfying
the curvature-dimension condition concentrates to an mm-space,
then the limit also satisfies the curvature-dimension condition
(see \cite{FS}).
This stability result of the curvature-dimension condition
has an important application to the eigenvalues of Laplacian
on Riemannian manifolds.
In fact, under the nonnegativity of Ricci curvature,
the $k^{th}$ eigenvalue of the Laplacian
of a closed Riemannian manifold
is dominated by a constant multiple of
the first eigenvalue, where the constant depends only on
$k$ and is independent of the dimension of the manifold.
This dimension-free estimate cannot be obtained
by the ordinary technique.
Combining this estimate with Gromov-V.~Milman's and
E.~Milman's results \cites{GroMil,Emil:isop,Emil:role},
we have the equivalence:
\begin{align*}
&\text{$\{X_n\}$ is a L\'evy family}\\
\Longleftrightarrow\ &\lambda_1(X_n) \to +\infty\\
\Longleftrightarrow\ &\lambda_k(X_n) \to +\infty
\ \text{for some $k$}
\end{align*}
for a sequence of closed Riemannian manifolds $X_n$, $n=1,2,\dots$,
with nonnegative Ricci curvature.
The organization of this book is as follows.
In Chapter \ref{chap:conv-meas}, we define
weak and vague convergence of measures,
the Prohorov distance, transportation, the Ky Fan metric,
convergence in measure of maps,
and present those basic facts.
Chapter \ref{chap:LM-conc} is devoted to
a minimal introduction to L\'evy-Milman concentration phenomenon.
We define the observable diameter, the separation distance,
and the Lipschitz order.
We prove the normal law \'a la L\'evy for $S^n(\sqrt{n})$ stating that
any limit of the push-forward of the normalized volume measure
on $S^n(\sqrt{n})$ by a $1$-Lipschitz continuous function on $S^n(\sqrt{n})$
as $n\to\infty$
is the push-forward of the $1$-dimensional standard Gaussian measure
by some $1$-Lipschitz continuous function on $\field{R}$.
From this we derive the asymptotic estimate of the observable diameter
of $S^n(1)$ and $\field{C} P^n$. We also prove the relation between
the $k^{th}$ eigenvalue of the Laplacian and the separation distance
for a compact Riemannian manifold,
which yields some examples of L\'evy families.
Most of the contents in this chapter are already known for specialists.
Chapter \ref{chap:GH-dist-matrix} presents some basic facts
on metric geometry, such as,
the Hausdorff distance and the Gromov-Hausdorff distance.
We also prove the equivalence between the Gromov-Hausdorff convergence
and the convergence of the distance matrices of compact metric spaces.
Chapter \ref{chap:box-dist} deals with the box distance between
mm-spaces, which is one of fundamental tools in this book.
We prove that the Lipschitz order is stable under box convergence,
and that any mm-space can be approximated by a monotone nondecreasing
sequence of finite-dimensional mm-spaces.
We investigate the convergence of finite product spaces to the infinite product.
Chapter \ref{chap:obs-dist-measurement}
discusses the observable distance and the measurements,
where the \emph{$N$-measurement of an mm-space} is defined to be
the set of push-forwards of the measure of the mm-space
by $1$-Lipschitz maps to $\field{R}^N$ with $l_\infty$ norm.
The measurements have the complete information of the mm-space
and can be treated easier than the mm-space itself.
We prove that the concentration of mm-spaces
is equivalent to
the convergence of the corresponding measurements,
which is one of the essential points for
the investigation of convergence of pyramids.
Chapter \ref{chap:pyramid} is devoted to the space of pyramids.
We define a metric on the space of pyramids
and prove its compactness.
The metric is originally due to \cite{Shioya:mmlim}.
In Chapter \ref{chap:asymp-conc},
we finally complete the proof of the theorem that the $\dconc$-completion of
the space of mm-spaces is embedded into the space of pyramids,
which is one of main theorems in this book.
We study the asymptotic concentration of finite product spaces
and the asymptotic property of
the pyramids $\mathcal{P}_{S^n(\sqrt{n})}$ and $\mathcal{P}_{(\field{R}^n,\gamma^n)}$ (see \cite{Shioya:mmlim}).
We also study spectral compactness
and prove that any spectrally compact and asymptotic sequence of mm-spaces
asymptotically concentrates if the observable diameter is bounded from above.
Chapter \ref{chap:dissipation} discusses dissipation.
After the basics of dissipation, we present some examples of dissipation.
One of the interesting examples is the sequence of the spheres $S^n(r_n)$
of radius $r_n$.
It infinitely dissipates if only if
$r_n/\sqrt{n} \to +\infty$ as $n\to\infty$ (see \cite{Shioya:mmlim}).
We also study some obstruction for dissipation
and prove the non-dissipation theorem.
The final Chapter \ref{chap:curv-conc} is an exposition of \cite{FS}.
We prove the stability theorem of the curvature-dimension condition
for concentration, and apply it to the study of the eigenvalues of
Laplacian on closed Riemannian manifolds.
We also prove the stability of a lower bound of Alexandrov curvature.
\begin{ack}
The author would like to thank Prof.~Mikhail Gromov and Prof.~Vitali Milman
for their comments and encouragement.
He also thanks to Prof.~Asuka Takatsu,
Mr.~Takuya Higashi, Mr.~Daisuke Kazukawa, Ms.~Yumi Kume,
and Mr.~Hirotaka Nakajima
for checking a draft version of the manuscript.
Thanks to Prof.~Atsushi Katsuda, Prof.~Takefumi Kondo, and
Mr.~Ryunosuke Ozawa for
valuable discussions and comments.
\end{ack}
\chapter{Preliminaries from measure theory}
\label{chap:conv-meas}
\section{Some basics}
In this section, we enumerate some basic definitions and
facts on measure theory.
We refer to \cites{Bog,Bil,Kechris} for more details.
\index{measure}
A \emph{measure} $\mu$ on a set $X$ is a nonnegative countably additive
set function on a $\sigma$-algebra over $X$.
\index{measure space}
We call a pair $(X,\mu)$ of a set $X$ and a measure $\mu$ on $X$
a \emph{measure space}.
\index{Borel measure}
A measure on a topological space $X$ is called
a \emph{Borel measure} if it is defined on the Borel $\sigma$-algebra
over $X$.
A measure $\mu$ on a set $X$ is said to be \emph{finite} if $\mu(X) < +\infty$.
A \emph{probability measure $\mu$ on $X$} is defined to be
a measure on $X$ with $\mu(X) = 1$.
\index{finite measure} \index{probability measure}
A map $f : X \to Y$ from a measure space $(X,\mu)$ to a topological space $Y$
is said to be ($\mu$-)\emph{measurable}
if for any Borel subset $A \subset Y$ the preimage $f^{-1}(A)$ belongs to
the associated $\sigma$-algebra of $(X,\mu)$.
A map $f : X \to Y$ between two topological spaces $X$ and $Y$
is \emph{Borel measurable}
if for any Borel subset $A \subset Y$ the preimage $f^{-1}(A)$
is a Borel subset of $X$.
\begin{defn}[Inner and outer regular measure]
\index{inner regular} \index{tight}
Let $\mu$ be a Borel measure on a topological space $X$.
$\mu$ is said to be \emph{inner regular} (or \emph{tight})
if for any Borel subset $A \subset X$ and for any real number
$\varepsilon > 0$,
there exists a compact set $K$ contained in $A$ such that
$\mu(A) \le \mu(K)+\varepsilon$.
\index{outer regular}
$\mu$ is said to be \emph{outer regular} if for any Borel subset
$A \subset X$ and for any $\varepsilon > 0$, there exists
an open set $U$ containing $A$
such that $\mu(A) \ge \mu(U) - \varepsilon$.
\end{defn}
\begin{thm}
Any finite Borel measure on a complete separable metric space
is inner and outer regular.
\end{thm}
\begin{defn}[Absolute continuity]
\index{absolutely continuous}
Let $\mu$ and $\nu$ be two Borel measures on a topological space $X$.
$\mu$ is said to be \emph{absolutely continuous with respect to $\nu$}
if $\mu(A) = 0$ for any Borel subset $A \subset X$ with $\nu(A) = 0$.
\end{defn}
\begin{thm}[Radon-Nikodym theorem] \label{thm:RN}
\index{Radon-Nikodym theorem}
Let $X$ be a topological space.
If a Borel measure $\mu$ on $X$ is absolutely continuous with respect to
a Borel measure $\nu$ on $X$, then there exists a Borel measurable function
$f : X \to [\,0,+\infty\,)$ such that
\[
\mu(A) = \int_A f \,d\nu
\]
for any Borel subset $A \subset X$.
Moreover, $f$ is unique $\nu$-a.e.
\end{thm}
\begin{defn}[Radon-Nikodym derivative]
\index{Radon-Nikodym derivative} \index{density}
The function $f$ in Theorem \ref{thm:RN} is called
the \emph{Radon-Nikodym derivative}
(or \emph{density}) \emph{of $\mu$ with respect to $\nu$}
and is denoted by
\[
\frac{d\mu}{d\nu}.
\]
\index{dmudnu@$\frac{d\mu}{d\nu}$}
\end{defn}
\begin{defn}[Push-forward]
Let $p : X \to Y$ be a measurable map from
a measure space $(X,\mu)$ to a topological space $Y$.
We define a Borel measure $p_*\mu$ on $Y$ by
\[
p_*\mu(A) := \mu(p^{-1}(A))
\]
for any Borel subset $A \subset Y$.
We call $p_*\mu$ the \emph{push-forward of $\mu$ by the map $p$}.
\end{defn}
\begin{thm}[Disintegration theorem] \label{thm:disintegration}
\index{disintegration theorem}
Let $p : X \to Y$ be a Borel measurable map between
two complete separable metric spaces $X$ and $Y$.
Then, for any finite Borel measure $\mu$ on $X$,
there exists a
family of probability measures $\mu_y$, $y\in Y$, on $X$ such that
\begin{enumerate}
\item the map $Y \ni y \mapsto \mu_y$ is Borel measurable, i.e.,
$Y \ni y \mapsto \mu_y(A)$ is a Borel measurable function for
any Borel subset $A \subset X$,
\item $\mu_y(X \setminus p^{-1}(y)) = 0$ for $p_*\mu$-a.e.~$y \in Y$,
\item for any Borel measurable function $f : X \to [\,0,+\infty\,)$,
\[
\int_X f(x) \, d\mu(x)
= \int_Y \int_{p^{-1}(y)} f(x) \, d\mu_y(x) d(p_*\mu)(y).
\]
\end{enumerate}
Moreover, $\{\mu_y\}_{y\in Y}$ is unique $p_*\mu$-a.e.
\end{thm}
\begin{defn}[Disintegration]
\index{disintegration}
The family $\{\mu_y\}_{y\in Y}$ as in Theorem \ref{thm:disintegration}
is called the \emph{disintegration of $\mu$ for $p : X \to Y$}.
\end{defn}
\begin{defn}[Median and L\'evy mean]
\index{median} \index{Levy mean@L\'evy mean}
Let $X$ be a measure space with probability measure $\mu$
and $f : X \to \field{R}$ a measurable function.
A real number $m$ is called a \emph{median of $f$}
if it satisfies
\begin{align*}
\mu(f\geq m) \geq \frac{1}{2} \quad \text{and}\quad
\mu(f\leq m) \geq \frac{1}{2},
\end{align*}
where $\mu(P)$ for a conditional formula $P$ denotes the $\mu$-measure
of the set of points where $P$ holds.
It is easy to see that the set of medians of $f$
is a closed and bounded interval.
The \emph{L\'evy mean of $f$ with respect to the measure $\mu$}
is defined to be \index{mf@$m_f$} \index{lm(f;muX)@$\lm(f;\mu)$}
\[
m_f := \lm(f;\mu) := \frac{a_f + b_f}{2},
\]
where $a_f$ is the minimum of medians of $f$
and $b_f$ the maximum of medians of $f$.
\end{defn}
\section{Convergence of measures}
\label{sec:obs-sphere}
For a while, let $X$ be a metric space with metric $d_X$.
\begin{defn}[Weak and vague convergence of measures]
\index{weak convergence} \index{converge weakly}
Let $\mu$ and $\mu_n$, $n=1,2,\dots$, be finite Borel measures on $X$.
We say that $\mu_n$ \emph{converges weakly} to $\mu$
and write \emph{$\mu_n \to \mu$ weakly} as $n\to\infty$ if
\begin{equation}
\label{eq:weak}
\lim_{n\to\infty} \int_X f \; d\mu_n = \int_X f \; d\mu
\end{equation}
for any bounded and continuous function $f : X \to \field{R}$.
\index{vague convergence} \index{converge vaguely}
We say that $\mu_n$ \emph{converges vaguely} to $\mu$
and write \emph{$\mu_n \to \mu$ vaguely} $n\to\infty$ if
\eqref{eq:weak} holds for any continuous function $f : X \to \field{R}$
with compact support.
\end{defn}
Any weakly convergent sequence is vaguely convergent,
but the converse is not necessarily true.
For example, the sequence of Dirac's delta measures $\delta_n$,
$n=1,2,\dots$, converges vaguely to the zero measure on $\field{R}$,
but it does not converge weakly, where \emph{Dirac's delta measure $\delta_x$
at a point $x$ in a space $X$} is defined by
\index{Dirac's delta measure} \index{deltax@$\delta_x$}
\[
\delta_x(A) :=
\begin{cases}
1 & \text{if $x \in A$},\\
0 & \text{if $x \notin A$}
\end{cases}
\]
for $A \subset X$.
For any weakly convergent sequence $\mu_n \to \mu$,
the total measure $\mu_n(X)$ converges to $\mu(X)$.
\begin{lem}
Assume that a sequence of finite Borel measures $\mu_n$, $n=1,2,\dots$,
converges weakly {\rm(}resp.~vaguely{\rm)}
to a finite Borel measure $\mu$ on a metric space
{\rm(}resp.~locally compact metric space{\rm)} $X$.
Then, for any Borel {\rm(}resp.~relatively compact Borel{\rm)}
subset $A \subset X$, we have
\[
\mu(A^\circ) \le
\liminf_{n\to\infty} \mu_n(A) \le \limsup_{n\to\infty} \mu_n(A) \le \mu(\bar{A}),
\]
where $A^\circ$ and $\bar{A}$ denote the interior and the closure
of $A$, respectively.
\end{lem}
\begin{defn}[Prohorov distance]
\index{Prohorov distance}
The \emph{Prohorov distance $d_P(\mu,\nu)$
between two Borel probability measures $\mu$ and $\nu$ on $X$}
is defined to be the infimum of $\varepsilon > 0$ satisfying
\begin{equation} \label{eq:Proh}
\mu(U_\varepsilon(A)) \ge \nu(A) - \varepsilon
\end{equation}
for any Borel subset $A \subset X$, where
\[
U_\varepsilon(A) := \{\; x \in X \mid d_X(x,A) < \varepsilon\;\}.
\]
\index{UepsilonA@$U_\varepsilon(A)$}
The distance function $d_P$ is called the \emph{Prohorov metric}.
\end{defn}
We have $d_P(\mu,\nu) \le 1$ for any two Borel probability measures
$\mu$ and $\nu$ on $X$.
Note that \eqref{eq:Proh} holds
for any Borel subset $A \subset X$ if and only if
\begin{equation}
\label{eq:Proh2}
\nu(U_\varepsilon(A)) \ge \mu(A) - \varepsilon
\end{equation}
for any Borel subset $A \subset X$.
In fact, since $U_\varepsilon(X \setminus U_\varepsilon(A)) \subset X \setminus A$,
\eqref{eq:Proh} for $X \setminus U_\varepsilon(A)$ yields
\[
\mu(X \setminus A) \ge \mu(U_\varepsilon(X \setminus U_\varepsilon(A)))
\ge \nu(X \setminus U_\varepsilon(A)) - \varepsilon,
\]
which implies \eqref{eq:Proh2}.
In particular, we have $d_P(\nu,\mu) = d_P(\mu,\nu)$.
\begin{prop}[cf.~\cite{Bil}*{\S 6}]
Let $X$ be a metric space.
The Prohorov metric $d_P$ is a metric on the set of Borel probability
measures on $X$.
\end{prop}
The following is sometimes useful.
\begin{lem}
For any two Borel probability measures $\mu$ and $\nu$ on $X$,
we have
\begin{align*}
d_P(\mu,\nu) = \inf\{\;\varepsilon \ge 0 &\mid
\mu(B_\varepsilon(A)) \ge \nu(A) - \varepsilon\\
&\quad\text{for any Borel subset $A \subset X$}\;\},
\end{align*}
where
\[
B_\varepsilon(A) := \{\;x \in X \mid d_X(x,A) \le \varepsilon\;\}.
\]
\index{BepsilonA@$B_\varepsilon(A)$}
\end{lem}
\begin{proof}
Since $U_\varepsilon(A) \subset B_\varepsilon(A)$,
the right-hand side is not greater than $d_P(\mu,\nu)$.
If $\mu(B_\varepsilon(A)) \ge \nu(A) - \varepsilon$
for a real number $\varepsilon > 0$, then
$\mu(U_{\varepsilon'}(A)) \ge \nu(A) - \varepsilon'$
for any $\varepsilon'$ with $\varepsilon' > \varepsilon$.
This proves that the right-hand side is not less than $d_P(\mu,\nu)$.
\end{proof}
\begin{lem}[cf.~\cite{Bil}*{\S 5--6}] \label{lem:conv-meas}
\begin{enumerate}
\item If $X$ is separable, then we have
\[
\mu_n \to \mu \ \text{weakly}\ \Longleftrightarrow\ d_P(\mu_n,\mu) \to 0
\]
for any Borel probability measures $\mu$ and $\mu_n$, $n=1,2,\dots$,
on $X$.
\item
If $X$ is separable {\rm(}resp.~separable and complete{\rm)},
then the set of Borel probability measures on $X$
is separable {\rm(}resp.~separable and complete{\rm)} with respect to $d_P$.
\item\label{it:conv-meas-w}
If $X$ is compact, then any sequence of Borel measures $\mu_n$,
$n=1,2,\dots$, on $X$ with $\sup_n\mu_n(X) < +\infty$
has a weakly convergent subsequence, and
in particular, the set of Borel probability measures
on $X$ is $d_P$-compact.
\item\label{it:conv-meas-v}
If $X$ is proper, then any sequence of Borel measures $\mu_n$,
$n=1,2,\dots$, on $X$ with $\sup_n\mu_n(X) < +\infty$
has a vaguely convergent subsequence,
where $X$ is said to be \emph{proper} if any bounded subset of $X$
is relatively compact. \index{proper metric space}
\end{enumerate}
\end{lem}
\begin{proof}
We refer to \cite{Bil}*{\S 5--6} for the proof of (1)--\eqref{it:conv-meas-w}.
We give the proof of \eqref{it:conv-meas-v}.
Let $\mu_n$, $n=1,2,\dots$, be Borel measures on a proper metric space $X$
with $\sup_n\mu_n(X) < +\infty$.
By \cite{Kechris}*{(5.3)}, the one-point compactification of $X$,
say $\hat{X}$, is metrizable.
Applying \eqref{it:conv-meas-w} yields that there exists
a weakly convergent subsequence of $\{\mu_n\}$ on $\hat{X}$,
which is a vaguely convergent sequence on $X$ if each $\mu_n$
is restricted on $X$.
This completes the proof.
\end{proof}
\begin{defn}[Tight]
\index{tight}
Let $\mathcal{M}$ be a family of Borel measures on a topological space $X$.
We say that $\mathcal{M}$ is \emph{tight} if
for any real number $\varepsilon > 0$ there exists a compact subset
$K_\varepsilon \subset X$ such that $\mu(X \setminus K_\varepsilon) < \varepsilon$
for every $\mu \in \mathcal{M}$.
\end{defn}
\begin{thm}[Prohorov's theorem]
\index{Prohorov's theorem} \label{thm:Proh}
Let $\mathcal{M}$ be a family of Borel probability measures
on a complete separable metric space.
Then the following {\rm(1)} and {\rm(2)} are equivalent to each other.
\begin{enumerate}
\item $\mathcal{M}$ is tight.
\item $\mathcal{M}$ is relatively compact with respect to $d_P$.
\end{enumerate}
\end{thm}
\begin{defn}[Transport plan]
\index{transport plan} \index{coupling}
Let $\mu$ and $\nu$ be two finite Borel measures on $X$.
A Borel measure $m$ on $X \times X$ is called
a \emph{transport plan} (or \emph{coupling}) \emph{between $\mu$ and $\nu$}
if
\[
m(A \times X) = \mu(A) \quad\text{and}\quad m(X \times A) = \nu(A)
\]
for any Borel subset $A \subset X$.
\end{defn}
\begin{defn}[$\varepsilon$-Transportation]
\index{epsilon transportation@$\varepsilon$-transportation}
\index{transportation}
Let $\mu$ and $\nu$ be two Borel probability measures on $X$.
A Borel measure $m$ on $X \times X$ is called
an \emph{$\varepsilon$-transportation between $\mu$ and $\nu$}
if the following (1) and (2) are satisfied.
\begin{enumerate}
\item There exist two Borel measures $\mu'$ and $\nu'$ on $X$
with $\mu' \le \mu$ and $\nu' \le \nu$
such that $m$ is a transport plan between $\mu'$ and $\nu'$.
\item We have
\[
\supp m \subset
\Delta_\varepsilon
:= \{\;(x,y) \in X \times X \mid d_X(x,y) \le \varepsilon\;\},
\]
where $\supp m$ is the \emph{support of $m$}, i.e.,
the set of points $x$ such that any open neighborhood of $x$
has positive $m$-measure.
\index{supp@$\supp$} \index{support of a measure}
\end{enumerate}
For an $\varepsilon$-transportation $m$ between $\mu$ and $\nu$,
the \emph{deficiency of $m$} is defined to be
\index{deficiency}
\[
\defi m := 1-m(X\times X).
\]
\end{defn}
\begin{thm}[Strassen's theorem; cf.~\cite{Villani:topics}*{Corollary 1.28}]
\label{thm:di-tra} \index{Strassen's theorem}
For any two Borel probability measures $\mu$ and $\nu$ on $X$, we have
\begin{align*}
d_P(\mu,\nu) = \inf\{\;\varepsilon > 0 &\mid
\text{There exists an $\varepsilon$-transportation $m$}\\
&\quad\text{between $\mu$ and $\nu$ with $\defi m \le \varepsilon$}\;\}.
\end{align*}
\end{thm}
\section{Convergence in measure of maps}
\begin{defn}[Convergence in measure, Ky Fan metric]
\index{convergence in measure} \index{Ky Fan metric}
\index{dKF mu@$d_{KF}^\mu$} \index{dKF@$\dKF$}
Let $(X,\mu)$ be a measure space and $Y$ a metric space.
For two $\mu$-measurable maps $f,g : X \to Y$, we define
$\dKF(f,g) = d_{KF}^\mu(f,g)$
to be the infimum of $\varepsilon \ge 0$ satisfying
\begin{align} \label{eq:me}
\mu(\{\;x \in X \mid d_Y(f(x),g(x)) > \varepsilon\;\}) \le \varepsilon.
\end{align}
We call $d_{KF}^\mu$ the \emph{Ky Fan metric} on the set of
$\mu$-measurable maps from $X$ to $Y$.
We say that a sequence of $\mu$-measurable maps $f_n : X \to Y$,
$n=1,2,\dots$, \emph{converges in measure} to a $\mu$-measurable map
$f : X \to Y$ if
\[
\lim_{n\to\infty} d_{KF}^\mu(f_n,f) = 0.
\]
\end{defn}
Note that $\dKF(f,g) \le 1$ for any $\mu$-measurable maps $f,g : X \to Y$
provided that $\mu$ is a probability measure.
\begin{rem} \label{rem:me}
Since $\varepsilon \mapsto
\mu(\{\;x \in X \mid d_Y(f(x),g(x)) > \varepsilon\;\})$
is right-continuous, there is the minimum of $\varepsilon \ge 0$
satisfying \eqref{eq:me}.
For a real number $\varepsilon$,
$\dKF(f,g) \le \varepsilon$ holds if and only if \eqref{eq:me} holds.
\end{rem}
\begin{lem}
Let $(X,\mu)$ be a measure space and $Y$ a metric space.
Then, $d_{KF}^\mu$ is a metric on the set of $\mu$-measurable maps from $X$ to $Y$
by identifying two measurable maps from $X$ to $Y$ if
they are equal to each other $\mu$-almost everywhere.
\end{lem}
\begin{proof}
Let $f, g, h : X \to Y$ be three $\mu$-measurable maps.
It is obvious that $f = g$ $\mu$-a.e. if and only if $\dKF(f,g) = 0$.
It is also clear that $\dKF(g,f) = \dKF(f,g)$.
We prove the triangle inequality $\dKF(f,h) \le \dKF(f,g) + \dKF(g,h)$.
Setting $\varepsilon := \dKF(f,g)$ and $\delta := \dKF(g,h)$, we have
(see Remark \ref{rem:me})
\begin{align*}
\mu(\{\;x \in X \mid d_Y(f(x),g(x)) > \varepsilon\;\}) &\le \varepsilon,\\
\mu(\{\;x \in X \mid d_Y(g(x),h(x)) > \delta\;\}) &\le \delta.
\end{align*}
If $d_Y(f(x),g(x))+d_Y(g(x),h(x)) > \varepsilon+\delta$ for a point $x \in X$,
then we have at least one of $d_Y(f(x),g(x)) > \varepsilon$ and
$d_Y(g(x),h(x)) > \delta$. Therefore,
\begin{align*}
&\mu(\{\;x \in X \mid d_Y(f(x),h(x)) > \varepsilon+\delta\;\})\\
&\le
\mu(\{\;x \in X \mid d_Y(f(x),g(x))+d_Y(g(x),h(x)) >
\varepsilon+\delta\;\})\\
&\le
\mu(\{\;x \in X \mid d_Y(f(x),g(x)) > \varepsilon\;\})\\
&\ +\mu(\{\;x \in X \mid d_Y(g(x),h(x)) > \delta\;\})\\
&\le \varepsilon+\delta,
\end{align*}
which implies $\dKF(f,h) \le \varepsilon+\delta$.
This completes the proof.
\end{proof}
\begin{lem} \label{lem:di-me}
Let $X$ be a topological space with a Borel probability measure $\mu$
and $Y$ a metric space.
For any two $\mu$-measurable maps $f,g : X \to Y$, we have
\[
d_P(f_*\mu,g_*\mu) \le d_{KF}^\mu(f,g).
\]
\end{lem}
\begin{proof}
Let $\varepsilon := d_{KF}^\mu(f,g)$.
We take any Borel subset $A \subset Y$.
It suffices to prove that $f_*\mu(B_\varepsilon(A)) \ge g_*\mu(A)-\varepsilon$.
Setting $X_0 := \{\;x \in X \mid d_Y(f(x),g(x)) \le \varepsilon\;\}$,
we have $\mu(X \setminus X_0) \le \varepsilon$.
We prove that $g^{-1}(A) \cap X_0 \subset f^{-1}(B_\varepsilon(A))$.
In fact, if we take any point $x \in g^{-1}(A) \cap X_0$, then
$g(x) \in A$ and $x \in X_0$, which imply
$f(x) \in B_\varepsilon(A)$ and so $x \in f^{-1}(B_\varepsilon(A))$.
Thus, $g^{-1}(A) \cap X_0 \subset f^{-1}(B_\varepsilon(A))$.
Since $\mu(g^{-1}(A) \setminus X_0) \le \mu(X \setminus X_0) \le \varepsilon$,
\begin{align*}
g_*\mu(A) &= \mu(g^{-1}(A))
= \mu(g^{-1}(A) \cap X_0) + \mu(g^{-1}(A) \setminus X_0)\\
&\le \mu(f^{-1}(B_\varepsilon(A))) + \varepsilon
= f_*\mu(B_\varepsilon(A)) + \varepsilon
\end{align*}
This completes the proof.
\end{proof}
The proof of the following lemma is left to the reader.
\begin{lem}
Let $X$ be a topological space with a Borel probability measure $\mu$,
and $Y$ a metric space.
For any Borel measurable map $f : X \to Y$ and for any point $c \in Y$,
we have
\[
d_P(f_*\mu,\delta_c) = d_{KF}^\mu(f,c).
\]
\end{lem}
\chapter{L\'evy-Milman concentration phenomenon}
\label{chap:LM-conc}
\section{Observation of spheres}
Let $S^n(r)$ \index{Snr@$S^n(r)$} be the sphere of radius $r > 0$
centered at the origin in the $(n+1)$-dimensional Euclidean space $\field{R}^{n+1}$
and $\sigma^n$ \index{sigman@$\sigma^n$} the Riemannian volume measure
on $S^n(r)$ normalized as $\sigma^n(S^n(r)) = 1$.
We assume the distance between points in $S^n(r)$
to be the geodesic distance.
Let $k \le n$.
Identifying $\field{R}^k$ with the subspace
$\field{R}^k \times \{(0,0,\dots,0)\} \subset \field{R}^{n+1}$,
we consider the orthogonal projection from $\field{R}^{n+1}$ to $\field{R}^k$,
and denote the restriction of it on $S^n(\sqrt{n})$ by
$\pi_{n,k} : S^n(\sqrt{n}) \to \field{R}^k$.
Note that $\pi_{n,k} : S^n(\sqrt{n}) \to \field{R}^k$ is $1$-Lipschitz
continuous, i.e., Lipschitz continuous functions with
Lipschitz constant $1$.
\index{standard Gaussian measure} \index{gammak@$\gamma^k$}
\index{Gaussian measure}
Denote by $\gamma^k$
the \emph{$k$-dimensional standard Gaussian measure on $\field{R}^k$},
i.e.,
\[
d\gamma^k(x) := \frac{1}{(2\pi)^{k/2}} e^{-\frac{1}{2}\|x\|_2^2} \; dx,
\qquad x \in \field{R}^k,
\]
where $\|x\|_2$ is the Euclidean norm of $x$
and $dx$ the $k$-dimensional Lebesgue measure on $\field{R}^k$.
\index{bar two@$\Vert\cdot\Vert_2$}
\begin{prop}[Maxwell-Boltzmann distribution law\footnotemark]
\label{prop:MB-law}
\index{Maxwell-Boltzmann distribution law} \index{Poincar\'e's limit}
\index{the Poincar\'e limit theorem}
\footnotetext{This is also called the Poincar\'e limit theorem
in many literature.
However, there is no evidence that Poincar\'e proved this
(see \cite{DF}).}
For any natural number $k$ we have
\[
\frac{d(\pi_{n,k})_*\sigma^n}{dx} \to \frac{d\gamma^k}{dx}
\qquad \text{as $n \to \infty$},
\]
where $(\pi_{n,k})_*\sigma^n$ is the push-forward of $\sigma^n$
by $\pi_{n,k}$. In particular,
\[
(\pi_{n,k})_*\sigma^n \to \gamma^k \ \text{weakly}
\qquad \text{as $n \to \infty$}.
\]
\end{prop}
\begin{proof}
Denote by $\vol_l$ the $l$-dimensional volume measure.
Since $\pi_{n,k}^{-1}(x)$, $x \in \field{R}^k$, is isometric to
$S^{n-k}((n-\|x\|_2^2)^{1/2})$, we have
\begin{align*}
\frac{d(\pi_{n,k})_*\sigma^n}{dx}
&= \frac{\vol_{n-k}\pi_{n,k}^{-1}(x)}{\vol_n S^n(\sqrt{n})}
= \frac{(n-\|x\|_2^2)^{(n-k)/2}}
{\int_{\|x\|_2 \le \sqrt{n}} (n-\|x\|_2^2)^{(n-k)/2} \; dx}\\
&\overset{n\to\infty}\longrightarrow
\frac{e^{-\frac{1}{2}\|x\|_2^2}}{\int_{\field{R}^k} e^{-\frac{1}{2}\|x\|_2^2}\,dx}
= \frac{1}{(2\pi)^{k/2}} e^{-\frac{1}{2}\|x\|_2^2}
= \frac{d\gamma^k}{dx}.
\end{align*}
\end{proof}
The purpose of this section is to prove the following
\begin{thm}[Normal law \`a la L\'evy] \label{thm:normal}
\index{normal law a la Levy@normal law \`a la L\'evy}
Let $f_n : S^n(\sqrt{n}) \to \field{R}$, $n=1,2,\dots$, be
$1$-Lipschitz functions.
Assume that, for a subsequence $\{f_{n_i}\}$ of $\{f_n\}$,
the push-forward $(f_{n_i})_*\sigma^{n_i}$ converges vaguely to
a Borel measure $\sigma_\infty$ on $\field{R}$.
Then, there exists a $1$-Lipschitz function $\alpha : \field{R} \to \field{R}$
such that
\[
\alpha_*\gamma^1 = \sigma_\infty
\]
unless $\sigma_\infty$ is identically equal to zero.
\end{thm}
Lemma \ref{lem:conv-meas}\eqref{it:conv-meas-v} implies
the existence of a subsequence $\{f_{n_i}\}$
such that $(f_{n_i})_*\sigma^{n_i}$ converges vaguely to
some finite Borel measure on $\field{R}$.
With the notation of Definition \ref{defn:dom}, we have
\[
(\field{R},|\cdot|,\sigma_\infty) \prec (\field{R},|\cdot|,\gamma^1).
\]
We need some claims for the proof of Theorem \ref{thm:normal}.
The following theorem is well-known.
\begin{thm}[L\'evy's isoperimetric inequality \cites{Levy,FLM}]
\label{thm:Levy-isop}
\index{Levys isoperimetric inequality@L\'evy's isoperimetric inequality}
For any closed subset $\Omega \subset S^n(1)$,
we take a metric ball $B_\Omega$ of $S^n(1)$ with
$\sigma^n(B_\Omega) = \sigma^n(\Omega)$.
Then we have
\[
\sigma^n(U_r(\Omega)) \ge \sigma^n(U_r(B_\Omega))
\]
for any $r > 0$.
\end{thm}
Assume the condition of Theorem \ref{thm:normal}.
We consider a natural compactification $\bar\field{R} := \field{R} \cup \{-\infty,+\infty\}$
of $\field{R}$. Then, by replacing $\{f_{n_i}\}$ with a subsequence,
$\{(f_{n_i})_*\sigma^{n_i}\}$ converges weakly to a Borel probability measure
$\bar{\sigma}_\infty$ on $\bar{\field{R}}$
(see Lemma \ref{lem:conv-meas}\eqref{it:conv-meas-w})
such that $\bar{\sigma}_\infty|_{\field{R}} = \sigma_\infty$.
We prove the following
\begin{lem}\label{lem:normal1}
Let $x$ and $x'$ be two given real numbers.
If $\gamma^1(\,-\infty,x\,] = \bar\sigma_\infty[\,-\infty,x'\,]$
and if $\sigma_\infty\{x'\} = 0$, then
\[
\sigma_\infty[\,x'-\varepsilon_1,x'+\varepsilon_2\,]
\ge \gamma^1[\,x-\varepsilon_1,x+\varepsilon_2\,]
\]
for all real numbers $\varepsilon_1,\varepsilon_2 \ge 0$.
In particular, if $\bar\sigma_\infty \neq \delta_{\pm\infty}$,
then $\bar\sigma_\infty\{-\infty,+\infty\} = 0$
and $\sigma_\infty$ is a probability measure on $\field{R}$.
\end{lem}
\begin{proof}
We set $\Omega_+ := \{\,f_{n_i} \ge x'\,\}$ and
$\Omega_- := \{\,f_{n_i} \le x'\,\}$.
We have $\Omega_+ \cup \Omega_- = S^{n_i}(\sqrt{n_i})$.
Let us prove
\begin{align}\label{eq:normal-omega1}
U_{\varepsilon_1}(\Omega_+) \cap U_{\varepsilon_2}(\Omega_-)
\subset \{\,x'-\varepsilon_1 \le f_{n_i} \le x'+\varepsilon_2\,\}.
\end{align}
In fact, for any point $\xi \in U_{\varepsilon_1}(\Omega_+)$,
there is a point $\xi' \in \Omega_+$ such that the geodesic distance
between $\xi$ and $\xi'$ is not greater than $\varepsilon_1$.
The $1$-Lipschitz continuity of $f_{n_i}$ proves that
$f_{n_i}(\xi) \ge f_{n_i}(\xi') - \varepsilon_1 \ge x'-\varepsilon_1$.
Thus we have
$U_{\varepsilon_1}(\Omega_+) \subset \{\,x'-\varepsilon_1 \le f_{n_i}\;\}$
and, in the same way,
$U_{\varepsilon_2}(\Omega_-) \subset \{\,f_{n_i} \le x'+\varepsilon_2\,\}$.
Combining these two inclusions implies \eqref{eq:normal-omega1}.
It follows from \eqref{eq:normal-omega1} and
$U_{\varepsilon_1}(\Omega_+) \cup U_{\varepsilon_2}(\Omega_-)= S^{n_i}(\sqrt{n_i})$
that
\begin{align*}
&(f_{n_i})_*\sigma^{n_i}[\,x'-\varepsilon_1,x'+\varepsilon_2\,]\\
&= \sigma^{n_i}(x'-\varepsilon_1 \le f_{n_i} \le x'+\varepsilon_2)
\ge \sigma^{n_i}(U_{\varepsilon_1}(\Omega_+) \cap U_{\varepsilon_2}(\Omega_-))\\
&= \sigma^{n_i}(U_{\varepsilon_1}(\Omega_+)) + \sigma^{n_i}(U_{\varepsilon_2}(\Omega_-))
-1.
\end{align*}
The L\'evy's isoperimetric inequality (Theorem \ref{thm:Levy-isop})
implies
$\sigma^{n_i}(U_{\varepsilon_1}(\Omega_+)) \ge \sigma^{n_i}(U_{\varepsilon_1}(B_{\Omega_+}))$
and
$\sigma^{n_i}(U_{\varepsilon_2}(\Omega_-)) \ge \sigma^{n_i}(U_{\varepsilon_2}(B_{\Omega_-}))$,
so that
\[
(f_{n_i})_*\sigma^{n_i}[\,x'-\varepsilon_1,x'+\varepsilon_2\,]
\ge \sigma^{n_i}(U_{\varepsilon_1}(B_{\Omega_+})) + \sigma^{n_i}(U_{\varepsilon_2}(B_{\Omega_-}))
-1.
\]
It follows from $\sigma_\infty\{x'\} = 0$ that $\sigma^{n_i}(\Omega_+)$
converges to $\bar\sigma_\infty(\,x',+\infty\,]$ as $n\to\infty$.
We besides have
$\bar\sigma_\infty(\,x',+\infty\,] = \gamma^1[\,x,+\infty\,) \neq 0,1$,
so that $\sigma^{n_i}(\Omega_+) \neq 0,1$ for all sufficiently large $i$.
Let $a_i$ and $b_i$ be two real numbers such that
$\sigma^{n_i}(\Omega_+) = (\pi_{n,1})_*\sigma^{n_i}[\,a_i,+\infty\,)$
and $\sigma^{n_i}(U_{\epsilon_1}(B_{\Omega_+}))
= (\pi_{n_i,1})_*\sigma^{n_i}[\,b_i,+\infty\,)$.
By the Maxwell-Boltzmann distribution law (Proposition \ref{prop:MB-law})
and by remarking that the radius of the sphere is divergent to infinity,
we see that $a_i$ and $b_i$ converges to $x$ and $x-\varepsilon_1$,
respectively, as $i\to\infty$.
In particular we obtain
\[
\lim_{i\to\infty} \sigma^{n_i}(U_{\varepsilon_1}(B_{\Omega_+}))
= \gamma^1[\,x-\varepsilon_1,+\infty\,)
\]
as well as
\[
\lim_{i\to\infty} \sigma^{n_i}(U_{\varepsilon_2}(B_{\Omega_-}))
= \gamma^1(\,-\infty,x+\varepsilon_2\,].
\]
Therefore,
\begin{align*}
\sigma_\infty[\,x'-\varepsilon_1,x'+\varepsilon_2\,]
&\ge \liminf_{i\to\infty} (f_{n_i})_*\sigma^{n_i}[\,x'-\varepsilon_1,x'+\varepsilon_2\,]\\
&\ge \gamma^1[\,x-\varepsilon_1,+\infty\,) + \gamma^1(\,-\infty,x+\varepsilon_2\,]
-1\\
&= \gamma^1[\,x-\varepsilon_1,x+\varepsilon_2\,].
\end{align*}
The first part of the lemma is obtained.
Assume that $\bar\sigma_\infty \neq \delta_{\pm\infty}$.
The rest of the proof is to show that $\sigma_\infty(\field{R}) = 1$.
Suppose $\sigma_\infty(\field{R}) < 1$.
Then, there is a non-atomic point $x' \in \field{R}$ of $\sigma_\infty$
(i.e., a point $x'$ with $\sigma_\infty\{x'\} = 0$)
such that
$0 < \bar\sigma_\infty[\,-\infty,x'\,) < 1$.
We find a real number $x$ in such a way that
$\gamma^1(\,-\infty,x\,] = \bar\sigma_\infty[\,-\infty,x'\,]$.
The first part of the lemma implies
$\sigma_\infty[\,x'-\varepsilon_1,x'+\varepsilon_2\,]
\ge \gamma^1[\,x-\varepsilon_1,x+\varepsilon_2\,]$
for all $\varepsilon_1,\varepsilon_2 \ge 0$.
Taking the limit as $\varepsilon_1,\varepsilon_2 \to +\infty$,
we obtain $\sigma_\infty(\field{R}) = 1$.
This completes the proof.
\end{proof}
\begin{lem} \label{lem:normal2}
$\supp\sigma_\infty$ is a closed interval.
\end{lem}
\begin{proof}
$\supp\sigma_\infty$ is a closed set by the definition of the support
of a measure.
It then suffices to prove the connectivity of $\supp\sigma_\infty$.
Suppose not. Then, there are numbers $x'$ and $\varepsilon > 0$
such that
$\sigma_\infty(\,-\infty,x'-\varepsilon\,) > 0$,
$\sigma_\infty[\,x'-\varepsilon,x'+\varepsilon\,] = 0$,
and $\sigma_\infty(\,x'+\varepsilon,+\infty\,) > 0$.
There is a number $x$ such that
$\gamma^1(\,-\infty,x\,] = \sigma_\infty(\,-\infty,x'\,]$.
Lemma \ref{lem:normal1} shows that
$\sigma_\infty[\,x'-\varepsilon,x'+\varepsilon\,]
\ge \gamma^1[\,x-\varepsilon,x+\varepsilon\,] > 0$,
which is a contradiction.
The lemma has been proved.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:normal}]
For any given real number $x$, there exists a smallest number $x'$
satisfying $\gamma^1(\,-\infty,x\,] \le \sigma_\infty(\,-\infty,x'\,]$.
The existence of $x'$ follows from the right-continuity and the monotonicity
of $y \mapsto \sigma_\infty(\,-\infty,y\,]$.
Setting
$\alpha(x) := x'$ we have a function
$\alpha : \field{R} \to \field{R}$, which is monotone nondecreasing.
It is easy to see that
$(\supp\sigma_\infty)^\circ \subset \alpha(\field{R}) \subset \supp\sigma_\infty$.
We first prove the continuity of $\alpha$ in the following.
Take any two numbers $x_1$ and $x_2$ with $x_1 < x_2$.
We have
$\gamma^1(\,-\infty,x_1\,] \le \sigma_\infty(\,-\infty,\alpha(x_1)\,]$
and $\gamma^1(\,-\infty,x_2\,] \ge \sigma_\infty(\,-\infty,\alpha(x_2)\,)$,
which imply
\begin{align} \label{eq:normal2}
\gamma^1[\,x_1,x_2\,] \ge \sigma_\infty(\,\alpha(x_1),\alpha(x_2)\,).
\end{align}
This shows that, as $x_1 \to a-0$ and $x_2 \to a+0$ for a number $a$,
we have
$\sigma_\infty(\,\alpha(x_1),\alpha(x_2)\,) \to 0$, which together with
Lemma \ref{lem:normal2} implies $\alpha(x_2) - \alpha(x_1) \to 0$.
Thus, $\alpha$ is continuous on $\field{R}$.
Let us next prove the $1$-Lipschitz continuity of $\alpha$.
We take two numbers $x$ and $\varepsilon > 0$ and fix them.
It suffices to prove that
\[
\Delta\alpha := \alpha(x+\varepsilon)-\alpha(x) \le \varepsilon.
\]
\begin{clm}
If $\sigma_\infty\{\alpha(x)\} = 0$, then $\Delta\alpha \le \varepsilon$.
\end{clm}
\begin{proof}
The claim is trivial if $\Delta\alpha = 0$.
We thus assume $\Delta\alpha > 0$.
Since $\sigma_\infty\{\alpha(x)\} = 0$, we have
$\gamma^1(\,-\infty,x\,] = \sigma_\infty(\,-\infty,\alpha(x)\,]$,
so that Lemma \ref{lem:normal1} implies that
\begin{align} \label{eq:normal1}
\sigma_\infty[\,\alpha(x),\alpha(x)+\delta\,]
\ge \gamma^1[\,x,x+\delta\,]
\end{align}
for all $\delta \ge 0$.
By \eqref{eq:normal2} and \eqref{eq:normal1},
\begin{align*}
\gamma^1[\,x,x+\varepsilon\,]
&\ge \sigma_\infty(\,\alpha(x),\alpha(x+\varepsilon)\,)\\
&= \sigma_\infty[\,\alpha(x),\alpha(x)+\Delta\alpha\,)\\
&= \lim_{\delta \to \Delta\alpha-0} \sigma_\infty[\,\alpha(x),\alpha(x)+\delta\,]\\
&\ge \lim_{\delta \to \Delta\alpha-0} \gamma^1[\,x,x+\delta\,]\\
&= \gamma^1[\,x,x+\Delta\alpha\,],
\end{align*}
which implies $\Delta\alpha \le \varepsilon$.
\end{proof}
We next prove that $\Delta\alpha \le \varepsilon$ in the case where
$\sigma_\infty\{\alpha(x)\} > 0$.
We may assume that $\Delta\alpha > 0$.
Let $x_+ := \sup \alpha^{-1}(\alpha(x))$.
It follows from $\alpha(x) < \alpha(x+\varepsilon)$
that $x_+ < x+\varepsilon$.
The continuity of $\alpha$ implies that $\alpha(x_+) = \alpha(x)$.
There is a sequence of positive numbers
$\varepsilon_i \to 0$ such that
$\sigma_\infty\{\alpha(x_++\varepsilon_i)\} = 0$.
By applying the claim above,
\[
\alpha(x_++\varepsilon_i+\varepsilon)-\alpha(x_++\varepsilon_i) \le \varepsilon.
\]
Moreover we have
$\alpha(x+\varepsilon) \le \alpha(x_++\varepsilon_i+\varepsilon)$
and $\alpha(x_++\varepsilon_i) \to \alpha(x_+) = \alpha(x)$ as $i\to\infty$.
Thus,
\[
\alpha(x+\varepsilon) - \alpha(x) \le \varepsilon
\]
and so $\alpha$ is $1$-Lipschitz continuous.
The rest is to prove that $\alpha_*\gamma^1 = \sigma_\infty$.
Take any number $x' \in \alpha(\field{R})$ and fix it.
Set $x := \sup \alpha^{-1}(x') \;(\le +\infty)$.
We then have $\alpha(x) = x'$ provided $x < +\infty$.
Since $x$ is the largest number to satisfy
$\gamma^1(\,-\infty,x\,] \le \sigma_\infty(\,-\infty,x'\,]$,
we have $\gamma^1(\,-\infty,x\,] = \sigma_\infty(\,-\infty,x'\,]$,
where we agree $\gamma^1(\,-\infty,+\infty\,] = 1$.
By the monotonicity of $\alpha$, we obtain
\[
\alpha_*\gamma^1(\,-\infty,x'\,] = \gamma^1(\alpha^{-1}(\,-\infty,x'\,])
= \gamma^1(\,-\infty,x\,] = \sigma_\infty(\,-\infty,x'\,],
\]
which implies that $\alpha_*\gamma^1 = \sigma_\infty$
because $\sigma_\infty$ is a Borel probability measure.
This completes the proof.
\end{proof}
\begin{cor}[L\'evy's lemma] \label{cor:Levy}
\index{Levy's lemma@L\'evy's lemma}
Let $f_n : S^n(1) \to \field{R}$, $n=1,2,\dots$, be $1$-Lipschitz functions
such that $\int_{S^n(1)} f_n \; d\sigma^n = 0$.
Then we have
\[
(f_n)_*\sigma^n \to \delta_0 \ \text{weakly as $n\to\infty$},
\]
or equivalently,
$f_n$ converges in measure to zero as $n\to\infty$.
\end{cor}
\begin{proof}
Suppose that the lemma is false, so that
we find $1$-Lipschitz functions $f_n : S^n(1) \to \field{R}$, $n=1,2,\dots$,
with $\int_{S^n(1)} f_n \; d\sigma^n = 0$,
and a subsequence $\{f_{n_i}\}$ of $\{f_n\}$ such that
\begin{equation}
\label{eq:Levy}
\liminf_{i\to\infty} d_P((f_{n_i})_*\sigma^{n_i},\delta_0) > 0.
\end{equation}
We denote by $\iota_n : S^n(\sqrt{n}) \to S^n(1)$ a natural map.
$\tilde{f}_n := \sqrt{n}\, f_n \circ \iota_n : S^n(\sqrt{n}) \to \field{R}$
is $1$-Lipschitz continuous.
Let $m_n$ be a median of $f_n$.
Then, $\sqrt{n}\,m_n$ is a median of $\tilde{f}_n$.
We consider the measure
$\tilde\sigma^{n_i} := (\tilde{f}_{n_i}-\sqrt{n_i}\,m_{n_i})_*\sigma^{n_i}$.
Lemma \ref{lem:conv-meas}\eqref{it:conv-meas-w} implies that
there is a subsequence of $\{\tilde\sigma^{n_i}\}$ that is weakly convergent on $\bar\field{R}$.
Replace $\{\tilde\sigma^{n_i}\}$ with such a subsequence.
Since
$\tilde\sigma^{n_i}(\,-\infty,0\,], \tilde\sigma^{n_i}[\,0,+\infty\,) \ge 1/2$
and by Lemma \ref{lem:normal1},
the limit of $\tilde\sigma^{n_i}$
is a Borel probability measure on $\field{R}$.
Therefore, $(f_{n_i} - m_{n_i})_*\sigma^{n_i}
= ((1/\sqrt{n_i})\tilde{f}_{n_i}-m_{n_i})_*\sigma^{n_i}$
converges weakly to $\delta_0$ as $i\to\infty$.
Since the geodesic distance between any two points in $S^n(1)$
is at most $\pi$
and $\int_{S^n(1)} f_n d\sigma^n = 0$,
we have $-\pi \le f_n \le \pi$, so that
$m_{n_i}$ converges to zero as $i\to\infty$.
We then obtain that $(f_{n_i})_*\sigma^{n_i}$ converges weakly to $\delta_0$
as $i\to\infty$, which contradicts \eqref{eq:Levy}.
This completes the proof.
\end{proof}
\section{mm-Isomorphism and Lipschitz order}
\begin{defn}[mm-Space]
\index{mm-space}
Let $(X,d_X)$ be a complete separable metric space
and $\mu_X$ a Borel probability measure on $X$.
We call the triple $(X,d_X,\mu_X)$ an \emph{mm-space}.
We sometimes say that $X$ is an mm-space, in which case
the metric and measure of $X$ are respectively indicated by $d_X$ and $\mu_X$.
\end{defn}
In this book, manifolds may have nonempty boundary
unless otherwise stated.
For a complete Riemannian manifold $X$ with finite volume,
we always equip $X$ with the Riemannian distance function $d_X$
and with the volume measure $\mu_X$ normalized as $\mu_X(X) = 1$,
i.e., $\mu_X := \vol_X/\vol_X(X)$,
where $\vol_X$ is the Riemannian volume measure on $X$.
Then, $(X,d_X,\mu_X)$ is an mm-space.
A complete Riemannian manifold with finite diameter is always compact.
However, an mm-space with finite diameter is not necessarily compact.
Such an example is obtained as the discrete countable space
$X = \{x_i\}_{i=1}^\infty$ with $d_X(x_i,x_j) = 1-\delta_{ij}$
and $\mu_X = \sum_{i=1}^\infty 2^{-i}\delta_{x_i}$,
where $\delta_{ii} = 1$ and $\delta_{ij} = 0$ if $i \neq j$.
\begin{defn}[mm-Isomorphism]
\index{mm-isomorphism} \index{mm-isomorphic}
Two mm-spaces $X$ and $Y$ are said to be \emph{mm-isomorphic}
to each other if there exists an isometry $f : \supp\mu_X \to \supp\mu_Y$
such that $f_*\mu_X = \mu_Y$.
Such an isometry $f$ is called an \emph{mm-isomorphism}.
The mm-isomorphism relation is an equivalence relation
on the set of mm-spaces.
Denote by $\mathcal{X}$ \index{X@$\mathcal{X}$}
the set of mm-isomorphism classes of mm-spaces.
\end{defn}
Any mm-isomorphism between mm-spaces is automatically surjective,
even if we do not assume it.
Note that $X$ is mm-isomorphic to $(\supp\mu_X,d_X,\mu_X)$.
\emph{We assume that any mm-space $X$ satisfies
\[
X = \supp\mu_X
\]
unless otherwise stated.}
\begin{defn}[Lipschitz order] \label{defn:dom}
\index{Lipschitz order}
Let $X$ and $Y$ be two mm-spaces.
We say that $X$ (\emph{Lipschitz}) \emph{dominates} $Y$
\index{dominate} \index{Lipschitz dominate} \index{less than@$\prec$}
and write $Y \prec X$ if
there exists a $1$-Lipschitz map $f : X \to Y$ satisfying
\[
f_*\mu_X = \mu_Y.
\]
We call the relation $\prec$ on $\mathcal{X}$ the \emph{Lipschitz order}.
\end{defn}
\begin{prop} \label{prop:Liporder}
The Lipschitz order $\prec$ is a partial order relation on $\mathcal{X}$, i.e.,
we have the following {\rm(1)}, {\rm(2)}, and {\rm(3)}
for any mm-spaces $X$, $Y$, and $Z$.
\begin{enumerate}
\item $X \prec X$.
\item If $X \prec Y$ and $Y \prec X$, then $X$ and $Y$ are
mm-isomorphic to each other.
\item If $X \prec Y$ and $Y \prec Z$, then $X \prec Z$.
\end{enumerate}
\end{prop}
(1) and (3) are obvious.
For the proof of (2), we need a lemma.
Let $\varphi : [\,0,+\infty\,) \to [\,0,+\infty\,)$
be a bounded, continuous, and strictly monotone increasing function.
For an mm-space $X$ we define
\[
\Avr_\varphi(X) := \int_{X \times X} \varphi(d_X(x,x'))\;d(\mu_X\otimes\mu_X)(x,x').
\]
\begin{lem} \label{lem:avr}
Let $X$ and $Y$ be two mm-spaces.
\begin{enumerate}
\item If $X \prec Y$, then $\Avr_\varphi(X) \le \Avr_\varphi(Y)$.
\item If $X \prec Y$ and if $\Avr_\varphi(X) = \Avr_\varphi(Y)$, then
$X$ and $Y$ are mm-isomorphic to each other.
\end{enumerate}
\end{lem}
\begin{proof}
We prove (1). By $X \prec Y$, we have a $1$-Lipschitz map $f : Y \to X$
with $f_*\mu_Y = \mu_X$.
We have
\begin{align*}
\Avr_\varphi(X) &= \int_{X \times X} \varphi(d_X(x,x'))
\;d(f_*\mu_Y\otimes f_*\mu_Y)(x,x')\\
&= \int_{Y \times Y} \varphi(d_X(f(y),f(y')))\;d(\mu_Y\otimes \mu_Y)(y,y')\\
&\le \int_{Y \times Y} \varphi(d_Y(y,y'))\;d(\mu_Y\otimes \mu_Y)(y,y')\\
&= \Avr_\varphi(Y).
\end{align*}
We prove (2). Since the equality holds in the above,
we have
\[
\varphi(d_X(f(y),f(y'))) = \varphi(d_Y(y,y'))
\]
$\mu_Y\otimes \mu_Y$-a.e.~$(y,y') \in Y \times Y$,
which implies that $f : Y \to X$ is isometric.
It follows from $f_*\mu_Y = \mu_X$ that
the image $f(Y)$ is dense in $X$,
which together with the completeness of $Y$
proves $f(Y) = X$.
Thus, $f$ is an mm-isomorphism between $X$ and $Y$.
This completes the proof.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop:Liporder}]
It suffices to prove (2).
Assume that $X \prec Y$ and $Y \prec X$.
By Lemma \ref{lem:avr}(1) we have $\Avr_\varphi(X) = \Avr_\varphi(Y)$,
so that Lemma \ref{lem:avr}(2) implies that
$X$ and $Y$ are mm-isomorphic to each other.
This completes the proof.
\end{proof}
\section{Observable diameter}
The observable diameter is one of the most fundamental invariants
of an mm-space.
\begin{defn}[Partial and observable diameter]
\index{observable diameter}
Let $X$ be an mm-space and $Y$ a metric space.
For a real number $\alpha \le 1$, we define
the \emph{partial diameter
$\diam(X;\alpha) = \diam(\mu_X;\alpha)$ of $X$}
\index{partial diameter}
\index{diam@$\diam(X;\alpha)$} \index{diam@$\diam(\mu_X;\alpha)$}
to be the infimum of $\diam A$,
where $A \subset X$ runs over all Borel subsets
with $\mu_X(A) \ge \alpha$, and the \emph{deameter $\diam A$ of $A$}
is defined by $\diam A := \sup_{x,y\in A} d_X(x,y)$ for $A \neq \emptyset$
and $\diam\emptyset := 0$.
For a real number $\kappa > 0$ we define
\begin{align*}
\ObsDiam_Y(X;-\kappa) &:= \sup\{\;\diam(f_*\mu_X;1-\kappa) \mid\\
&\qquad\qquad\text{$f : X \to Y$ is $1$-Lipschitz}\;\},\\
\ObsDiam_Y(X) &:= \inf_{\kappa > 0} \max\{\ObsDiam_Y(X;-\kappa),\kappa\}.
\end{align*}
\index{obsdiam@$\ObsDiam_Y(\cdots)$, $\ObsDiam(\cdots)$}
We call $\ObsDiam_Y(X)$ (resp.~$\ObsDiam_Y(X;-\kappa)$)
the \emph{observable diameter of $X$ with screen $Y$}
(resp.~\emph{$\kappa$-observable diameter of $X$ with screen $Y$}).
\index{screen}
The case $Y = \field{R}$ is most important and we set
\begin{align*}
\ObsDiam(X;-\kappa) &:= \ObsDiam_\field{R}(X;-\kappa),\\
\ObsDiam(X) &:= \ObsDiam_\field{R}(X).
\end{align*}
\end{defn}
The observable diameter is invariant under mm-isomorphism.
Note that $\ObsDiam_Y(X;-\kappa) = \diam(X;1-\kappa) = 0$ for $\kappa \ge 1$
and we always have $\ObsDiam_Y(X) \le 1$.
We see that $\diam(\mu_X;1-\kappa)$ and $\ObsDiam_Y(X;-\kappa)$ are
both monotone nonincreasing in $\kappa$.
\begin{defn}[L\'evy family]
\index{Levy family@L\'evy family}
A sequence of mm-spaces $X_n$, $n=1,2,\dots$,
is called a \emph{L\'evy family} if
\[
\lim_{n\to\infty} \ObsDiam(X_n) = 0,
\]
or equivalently
\[
\lim_{n\to\infty} \ObsDiam(X_n;-\kappa) = 0
\]
for any $\kappa > 0$.
\end{defn}
It follows from the definition that
$\{X_n\}_{n=1}^\infty$ is a L\'evy family if and only if
\begin{itemize}
\item for any $1$-Lipschitz functions $f_n : X_n \to \field{R}$, $n=1,2,\dots$,
there exist real numbers $c_n$ such that
\[
\lim_{n\to\infty} \dKF(f_n,c_n) = 0.
\]
\end{itemize}
In particular, L\'evy's lemma (Corollary \ref{cor:Levy}) implies
\begin{thm}
$\{S^n(1)\}_{n=1}^\infty$ is a L\'evy family.
\end{thm}
\begin{rem}
$\diam(S^n(1);1-\kappa)$ does not converge to zero
as $n\to\infty$ for $0 < \kappa < 1$.
\end{rem}
\begin{rem}
For a L\'evy family $\{X_n\}$,
the above constant $c_n$ can always be taken to be a median of $f_n$,
where the width of the interval of medians of $f_n$ shrinks to zero
as $n\to\infty$.
In the case where $\diam X_n$ is bounded from above, such as $S^n(1)$,
the difference between the average and a median of $f_n$ tends to
zero as $n\to\infty$ for the L\'evy family $\{X_n\}$.
However, this is not true in general.
For instance, considering the sequence of the measures
\[
\mu_n := (1-1/n)\delta_0+(1/n)\delta_n, \quad n=1,2,\dots,
\]
we see that $\{(\field{R},\mu_n)\}$ is a L\'evy family.
The map $f_n(x) = x$, $x \in \field{R}$, has $\mu_n$-average $1$ for any $n$,
but $0$ is the unique median of $f_n$ for $n \ge 3$.
\end{rem}
\begin{prop} \label{prop:diam-ObsDiam-dom}
Let $X$ and $Y$ be two mm-spaces and $\kappa > 0$ a real number.
\begin{enumerate}
\item If $X$ is dominated by $Y$, then
\[
\diam(X;1-\kappa) \le \diam(Y;1-\kappa).
\]
\item We have
\[
\ObsDiam(X;-\kappa) \le \diam(X;1-\kappa).
\]
\item If $X$ is dominated by $Y$, then
\[
\ObsDiam(X;-\kappa) \le \ObsDiam(Y;-\kappa).
\]
\end{enumerate}
\end{prop}
\begin{proof}
We prove (1).
Since $X \prec Y$, there is a $1$-Lipschitz map $F : Y \to X$
such that $F_*\mu_Y = \mu_X$.
Let $A$ be any Borel subset of $Y$ with $\mu_Y(A) \ge 1-\kappa$
and $\overline{F(A)}$ the closure of $F(A)$.
\index{overline@$\overline{\ \cdot\ }$}
We have $\mu_X(\overline{F(A)}) = \mu_Y(F^{-1}(\overline{F(A)}))
\ge \mu_Y(A) \ge 1-\kappa$ and,
by the $1$-Lipschitz continuity of $F$,
$\diam(\overline{F(A)}) \le \diam A$.
Therefore, $\diam(X;1-\kappa) \le \diam A$.
Taking the infimum of $\diam A$ over all $A$'s
yields (1).
We prove (2). Let $f : X \to \field{R}$ be any $1$-Lipschitz
function. Since $(\field{R},|\cdot|,f_*\mu_X)$ is dominated by $X$,
(1) implies that $\diam(f_*\mu_X;1-\kappa) \le \diam(X;1-\kappa)$.
This proves (2).
We prove (3).
By $X \prec Y$, there is a $1$-Lipschitz map $F : Y \to X$ with
$F_*\mu_Y = \mu_X$.
For any $1$-Lipschitz function $f : X \to \field{R}$,
we have $f_*\mu_X = f_*F_*\mu_Y = (f\circ F)_*\mu_Y$.
Since $f\circ F : Y \to \field{R}$ is also a $1$-Lipschitz function,
\[
\diam(f_*\mu_X;1-\kappa) = \diam((f\circ F)_*\mu_Y;1-\kappa)
\le \ObsDiam(Y;-\kappa).
\]
This completes the proof.
\end{proof}
\begin{prop} \label{prop:scale-ObsDiam}
Let $X$ be an mm-space.
Then, for any real number $t > 0$ we have
\[
\ObsDiam(tX;-\kappa) = t\ObsDiam(X;-\kappa),
\]
where $tX := (X,td_X,\mu_X)$. \index{tX@$tX$}
\end{prop}
\begin{proof}
We have
\begin{align*}
&\ObsDiam(tX;-\kappa)\\
&= \sup\{\;\diam(f_*\mu_X;1-\kappa) \mid
\text{$f : tX \to \field{R}$ $1$-Lipschitz}\;\}\\
&= \sup\{\;\diam(f_*\mu_X;1-\kappa) \mid
\text{$t^{-1}f : X \to \field{R}$ $1$-Lipschitz}\;\}\\
&= \sup\{\;\diam((tg)_*\mu_X;1-\kappa) \mid
\text{$g : X \to \field{R}$ $1$-Lipschitz}\;\}\\
&= t\ObsDiam(X;-\kappa).
\end{align*}
This completes the proof.
\end{proof}
Denote the \emph{$l_\infty$ norm} on $\field{R}^N$ by $\|\cdot\|_\infty$, i.e.,
\[
\|x\|_\infty := \max_{i=1}^N |x_i|
\]
for $x = (x_1,x_2,\dots,x_N) \in \field{R}^N$.
\index{bar infinity@$\Vert\cdot\Vert_\infty$}
\index{l infinity norm@$l_\infty$ norm}
\begin{lem} \label{lem:ObsDiamRN-ObsDiam}
Let $X$ be an mm-space.
For any real number $\kappa > 0$ and any natural number $N$,
we have
\[
\ObsDiam_{(\field{R}^N,\|\cdot\|_\infty)}(X;-N\kappa)
\le \ObsDiam(X;-\kappa).
\]
\end{lem}
\begin{proof}
Assume $\ObsDiam(X;-\kappa) < \varepsilon$ for a number
$\varepsilon$.
Let $F : X \to (\field{R}^N,\|\cdot\|_\infty)$ be any $1$-Lipschitz map.
By setting $(f_1,f_2,\dots,f_N) := F$,
each $f_i$ is $1$-Lipschitz continuous and so
$\diam((f_i)_*\mu_X;1-\kappa) < \varepsilon$.
There is a Borel subset $A_i \subset \field{R}$ for each $i$ such that
$(f_i)_*\mu_X(A_i) \ge 1-\kappa$ and $\diam A_i < \varepsilon$.
Letting $A := A_1 \times A_2 \times \dots \times A_N$,
we have
\[
F_*\mu_X(A) = \mu_X(F^{-1}(A))
= \mu_X(f_1^{-1}(A_1) \cap \dots \cap f_N^{-1}(A_N))
\ge 1-N\kappa.
\]
Since $\diam A < \varepsilon$,
we have $\diam(F_*\mu_X;1-N\kappa) < \varepsilon$.
This completes the proof.
\end{proof}
\begin{thm} \label{thm:ObsDiam-Sn}
For any real number $\kappa$ with $0 < \kappa < 1$,
we have
\begin{align}
\tag{1}
\lim_{n\to\infty} \ObsDiam(S^n(\sqrt{n});-\kappa)
&= \diam(\gamma^1;1-\kappa) = 2I^{-1}((1-\kappa)/2),\\
\tag{2}
\ObsDiam(S^n(1);-\kappa) &= O(n^{-1/2}),
\end{align}
where $I(r) := \gamma^1[\,0,r\,]$.
\end{thm}
The theorem implies the L\'evy property of $\{S^n(1)\}_{n=1}^\infty$.
\begin{proof}
(1) follows from the normal law \`a la L\'evy (Theorem \ref{thm:normal})
and the Maxwell-Boltzmann distribution law (Proposition \ref{prop:MB-law}).
We prove (2).
By Proposition \ref{prop:scale-ObsDiam} we have
\[
\ObsDiam(S^n(1);-\kappa) = \frac{1}{\sqrt{n}}\ObsDiam(S^n(\sqrt{n});\kappa),
\]
which together with (1) implies (2).
This completes the proof.
\end{proof}
Combining Theorem \ref{thm:ObsDiam-Sn} and
Proposition \ref{prop:scale-ObsDiam} proves the following
\begin{cor}[\cite{GroMil}*{\S 1.1}] \label{cor:ObsDiam-Sn}
Let $r_n$, $n=1,2,\dots$, be positive real numbers.
Then, $\{S^n(r_n)\}$ is a L\'evy family if and only if
$r_n/\sqrt{n} \to 0$ as $n\to\infty$.
\end{cor}
\begin{ex} \label{ex:CPn}
We see that the Hopf fibration $f_n : S^{2n+1}(1) \to \field{C} P^n$ is $1$-Lipschitz
continuous with respect to the Fubini-Study metric on $\field{C} P^n$, and
that the push-forward $(f_n)_*\sigma^{2n+1}$ coincides with
the normalized volume measure on $\field{C} P^n$ induced from the Fubini-Study
metric.
This together with Proposition \ref{prop:diam-ObsDiam-dom}(3) implies
\[
\ObsDiam(\field{C} P^n;-\kappa) \le \ObsDiam(S^{2n+1};-\kappa) = O(n^{-1/2})
\]
for any $\kappa$ with $0 < \kappa < 1$.
In particular, $\{\field{C} P^n\}_{n=1}^\infty$ is a L\'evy family.
\end{ex}
\section{Separation distance}
\begin{defn}[Separation distance]
\index{separation distance} \index{sep@$\Sep(\cdots)$}
Let $X$ be an mm-space.
For any real numbers $\kappa_0,\kappa_1,\cdots,\kappa_N > 0$
with $N\geq 1$,
we define the \emph{separation distance}
\[
\Sep(X;\kappa_0,\kappa_1, \cdots, \kappa_N)
\]
of $X$ as the supremum of $\min_{i\neq j} d_X(A_i,A_j)$
over all sequences of $N+1$ Borel subsets $A_0,A_2, \cdots, A_N \subset X$
satisfying that $\mu_X(A_i) \geq \kappa_i$ for all $i=0,1,\cdots,N$,
where $d_X(A_i,A_j) := \inf_{x\in A_i,y\in A_j} d_X(x,y)$.
If there exists no sequence $A_0,\dots,A_N \subset X$
with $\mu_X(A_i) \ge \kappa_i$, $i=0,1,\cdots,N$, then
we define
\[
\Sep(X;\kappa_0,\kappa_1, \cdots, \kappa_N) := 0.
\]
\end{defn}
We see that $\Sep(X;\kappa_0,\kappa_1, \cdots, \kappa_N)$ is
monotone nonincreasing in $\kappa_i$ for each $i=0,1,\dots,N$.
The separation distance is an invariant under mm-isomorphism.
\begin{lem} \label{lem:Sep-prec}
Let $X$ and $Y$ be two mm-spaces.
If $X$ is dominated by $Y$, then we have,
for any real numbers $\kappa_0,\dots,\kappa_N > 0$,
\[
\Sep(X;\kappa_0,\dots,\kappa_N) \le \Sep(Y;\kappa_0,\dots,\kappa_N).
\]
\end{lem}
\begin{proof}
If we assume $X \prec Y$, then there is a $1$-Lipschitz map
$f : Y \to X$ such that $f_*\mu_Y = \mu_X$.
We take any Borel subsets $A_0,A_1,\dots,A_N \subset X$
such that $\mu_X(A_i) \ge \kappa_i$ for any $i$.
If we have no such sequence of Borel subsets, the the lemma is trivial.
Since $f$ is $1$-Lipschitz continuous and since
$\mu_Y(f^{-1}(A_i)) = \mu_X(A_i) \ge \kappa_i$, we have
\[
\min_{i\neq j} d_X(A_i,A_j)
\le \min_{i\neq j} d_Y(f^{-1}(A_i),f^{-1}(A_j))
\le \Sep(Y;\kappa_0,\dots,\kappa_N).
\]
This completes the proof.
\end{proof}
\begin{prop} \label{prop:ObsDiam-Sep}
For any mm-space $X$ and any real numbers $\kappa$ and $\kappa'$
with $\kappa > \kappa' > 0$, we have
\begin{align}
\tag{1} &\ObsDiam(X;-2\kappa) \le \Sep(X;\kappa,\kappa),\\
\tag{2} &\Sep(X;\kappa,\kappa) \le \ObsDiam(X;-\kappa').
\end{align}
\end{prop}
\begin{proof}
We prove (1).
If $\kappa \ge 1/2$, then the left-hand side of (1)
becomes zero and (1) is trivial.
We assume $\kappa < 1/2$.
Let $f : X \to \field{R}$ be any $1$-Lipschitz function and set
\begin{align*}
\rho_- &:= \sup\{\;t \in \field{R} \mid f_*\mu_X(\,-\infty,t\,) \le \kappa\;\},\\
\rho_+ &:= \inf\{\;t \in \field{R} \mid f_*\mu_X(\,t,+\infty\,) \le \kappa\;\}.
\end{align*}
We then see that
\begin{align*}
& f_*\mu_X(\,-\infty,\rho_-\,) \le \kappa, \qquad
f_*\mu_X(\,-\infty,\rho_-\,] \ge \kappa,\\
& f_*\mu_X(\,\rho_+,+\infty\,) \le \kappa, \qquad
f_*\mu_X[\,\rho_+,+\infty\,) \ge \kappa,\\
& \qquad\qquad\qquad\text{and}\quad \rho_- \le \rho_+.
\end{align*}
Since $f_*\mu_X[\,\rho_-,\rho_+\,] \ge 1-2\kappa$, we have
\begin{align*}
\diam(f_*\mu_X;1-2\kappa) &\le \rho_+ - \rho_-
= d_{\field{R}}((\,-\infty,\rho_-\,],[\,\rho_+,+\infty\,))\\
&\le \Sep((\field{R},f_*\mu_X);\kappa,\kappa)
\le \Sep(X;\kappa,\kappa),
\end{align*}
where the last inequality follows from Lemma \ref{lem:Sep-prec}.
This proves (1).
We prove (2).
Let $A_i \subset X$, $i=1,2$, be any Borel subsets
with $\mu_X(A_i) \ge \kappa$.
We define $f(x) := d_X(x,A_1)$, $x \in X$.
Then, $f : X \to \field{R}$ is a $1$-Lipschitz function.
Let us estimate $\diam(f_*\mu_X;1-\kappa')$.
Any interval $I \subset \field{R}$ with $f_*\mu_X(I) \ge 1-\kappa'$
intersects both $f(A_1)$ and $f(A_2)$,
because
\[
f_*\mu_X(f(A_i)) + f_*\mu_X(I) \ge \mu_X(A_i) + f_*\mu_X(I)
\ge \kappa + 1-\kappa' > 1
\]
for $i=1,2$.
Therefore,
\[
\diam I \ge d_{\field{R}}(f(A_1),f(A_2)) = \inf_{x\in A_2} f(x) = d_X(A_1,A_2)
\]
and so $\ObsDiam(X;-\kappa') \ge \diam(f_*\mu_X;1-\kappa') \ge d_X(A_1,A_2)$.
This completes the proof of the proposition.
\end{proof}
\begin{rem}
Proposition \ref{prop:ObsDiam-Sep}(2) does not hold
for $\kappa = \kappa'$.
In fact, let us consider an mm-space $X := \{x_1,x_2\}$ with
$\mu_X := (1/2)\delta_{x_1} + (1/2)\delta_{x_2}$.
Then we have
\begin{align*}
\Sep(X;1/2,1/2) &= d_X(x_1,x_2),\\
\ObsDiam(X;-1/2) &= \diam(X;1/2) = 0.
\end{align*}
\end{rem}
It follows from Proposition \ref{prop:ObsDiam-Sep} that
$\{X_n\}$ is a L\'evy family if and only if
\begin{itemize}
\item for any Borel subsets $A_{in} \subset X_n$, $i=1,2$, $n=1,2,\dots$,
such that $\liminf_{n\to\infty} \mu_{X_n}(A_{in}) > 0$, we have
\[
\lim_{n\to\infty} d_{X_n}(A_{1n},A_{2n}) = 0.
\]
\end{itemize}
It is easy to see that this is also equivalent to
\begin{itemize}
\item for any Borel subsets $A_n \subset X_n$, $n=1,2,\dots$, such that\\
$\liminf_{n\to\infty} \mu_{X_n}(A_n) > 0$ and
$\limsup_{n\to\infty} \mu_{X_n}(A_n) < 1$, we have
\[
\lim_{n\to\infty} \mu_{X_n}(B_\varepsilon(A_n) \cap B_\varepsilon(X_n \setminus A_n)) = 1
\]
for any $\varepsilon > 0$.
\end{itemize}
In the case where $X_n$ are Riemannian manifolds, the L\'evy property
of $\{X_n\}$ means that
the measure $\mu_{X_n}$ on $X_n$ concentrates around the boundary of any such
$A_n$ for large $n$.
\begin{rem}
Let $X$ be an mm-space.
One more well-known invariant is
the \emph{concentration function} $\alpha_X(r)$, $r > 0$,
\index{concentration function} \index{alphaX@$\alpha_X$}
defined by
\[
\alpha_X(r) := \sup\{\;1-\mu_X(U_r(A)) \mid A \subset X : \text{Borel},
\ \mu_X(A) \ge 1/2\;\}.
\]
We have
\begin{align}
\tag{1} &\ObsDiam(X;-\kappa)
\le 2\,\inf\{\; r > 0 \mid \alpha_X(r) \le \kappa/2\;\},\\
\tag{2} &\alpha_X(r)
\le \sup\{\;\kappa > 0 \mid \ObsDiam(X;-\kappa) \ge r\;\}.
\end{align}
We refer to \cite{Ldx:book}*{Proposition 1.12} for the proof of (1).
We now prove (2).
Let $\varepsilon > 0$ be any small real number.
There is a Borel subset $A \subset X$ such that
$\mu_X(A) \ge 1/2$ and $\mu_X(X \setminus U_r(A)) > \alpha_X(r)-\varepsilon$.
Since $d_X(A,X \setminus U_r(A)) \ge r$,
we have $\Sep(X;1/2,\alpha_X(r)-\varepsilon) \ge r$.
By Proposition \ref{prop:ObsDiam-Sep},
$\ObsDiam(X;-(\alpha_X(r)-2\varepsilon)) \ge r$, so that
$\alpha_X(r)-2\varepsilon$ is not greater than the right-hand side
of (2) for any small $\varepsilon > 0$. This proves (2).
\end{rem}
\section{Comparison theorem for observable diameter}
In this section, we prove the following theorem by using
the L\'evy-Gromov isoperimetric inequality.
A manifold is said to be \emph{closed} if it is compact and has no boundary.
\index{closed manifold}
$\Ric_X$ denotes the Ricci curvature of a Riemannian manifold $X$.
\index{Ric@$\Ric_X$}
\begin{thm} \label{thm:comp}
Let $X$ be a closed and connected $n$-dimensional Riemannian manifold
of $\Ric_X \ge n-1$, $n \ge 2$.
Then, for any $\kappa$ with $0 < \kappa \le 1$, we have
\begin{align*}
\ObsDiam(X;-\kappa) &\le \ObsDiam(S^n(1);-\kappa) = \pi-2v^{-1}(\kappa/2)\\
&\le \frac{2\sqrt{2}}{\sqrt{n-1}} \sqrt{-\log\left(\sqrt{\frac{2}{\pi}}\kappa\right)},
\end{align*}
where
\[
v(r) := \frac{\int_0^r \sin^{n-1} t \; dt}{\int_0^\pi \sin^{n-1} t \; dt}
\]
is the $\sigma^n$-measure of a metric ball of radius $r$ in $S^n(1)$.
\end{thm}
\begin{rem}
Theorem \ref{thm:comp} also leads to Theorem \ref{thm:ObsDiam-Sn}.
\end{rem}
Throughout this section, let $X$ be a closed and connected $n$-dimensional
Riemannian manifold
of $\Ric_X \ge n-1$, $n \ge 2$, with the normalized volume measure $\mu_X$,
and let $0 < \kappa \le 1$.
The following is a generalization of L\'evy's isoperimetric inequality
(Theorem \ref{thm:Levy-isop}).
\begin{thm}[L\'evy-Gromov isoperimetric inequality
\cite{Gromov}*{Appendix $C_+$}]
For any closed subset $\Omega \subset X$, we take a metric ball $B_\Omega$
of $S^n(1)$ with $\sigma^n(B_\Omega) = \mu_X(\Omega)$. \label{thm:LGisop}
Then we have
\[
\mu_X(U_r(\Omega)) \ge \sigma^n(U_r(B_\Omega))
\]
for any $r > 0$.
\end{thm}
Using this theorem we prove
\begin{lem} \label{lem:comp1}
We have
\[
\Sep(X;\kappa/2,\kappa/2) \le \Sep(S^n(1);\kappa/2,\kappa/2)
= \pi - 2v^{-1}(\kappa/2).
\]
\end{lem}
\begin{proof}
Since the distance between the metric ball centered at the north pole of $S^n(1)$
with $\sigma^n$-measure $\kappa/2$ and the metric ball centered at the
south pole of $S^n(1)$ with $\sigma^n$-measure $\kappa/2$
is equal to $\pi - 2v^{-1}(\kappa/2)$,
we see that
\[
\Sep(S^n(1);\kappa/2,\kappa/2) \ge \pi - 2v^{-1}(\kappa/2).
\]
The rest of the proof is to show
$\Sep(X;\kappa/2,\kappa/2) \le \pi - 2v^{-1}(\kappa/2)$.
Let $\Omega, \Omega' \subset X$ be two mutually disjoint closed subsets
such that $\mu_X(\Omega) = \mu_X(\Omega') = \kappa/2$,
and let $r := d_X(\Omega,\Omega')$.
It suffices to prove that
$r \le \pi - 2v^{-1}(\kappa/2)$.
Since $\Omega'$ and $U_r(\Omega)$ are disjoint to each other,
the L\'evy-Gromov isoperimetric inequality (Theorem \ref{thm:LGisop}) proves
\begin{align*}
&\frac{\kappa}{2} = \mu_X(\Omega') \le 1 - \mu_X(U_r(\Omega))\\
&\le 1 - \sigma^n(U_r(B_\Omega))
= 1 - v(r+v^{-1}(\kappa/2))
\end{align*}
and hence
\[
r + v^{-1}(\kappa/2) \le v^{-1}(1-\kappa/2) = \pi - v^{-1}(\kappa/2).
\]
This completes the proof.
\end{proof}
\begin{lem} \label{lem:comp2}
We have
\[
\ObsDiam(S^n(1);-\kappa) = \Sep(S^n(1);\kappa/2,\kappa/2).
\]
\end{lem}
\begin{proof}
Proposition \ref{prop:ObsDiam-Sep}(1) implies
\[
\ObsDiam(S^n(1);-\kappa) \le \Sep(S^n(1);\kappa/2,\kappa/2).
\]
We prove the reverse inequality.
Let $f(x) := d_{S^n(1)}(x_0,x)$, $x \in S^n(1)$, be the distance function
from a fixed point $x_0 \in S^n(1)$, where $d_{S^n(1)}$
is the geodesic distance function on $S^n(1)$.
We see
\[
\frac{df_*\sigma^n}{dx}(r)
= \frac{\sin^{n-1}r}{\int_0^\pi \sin^{n-1}t \; dt},
\]
which implies $\diam(f_*\sigma^n;1-\kappa) = \pi - 2v^{-1}(\kappa/2)$.
Therefore,
\[
\ObsDiam(S^n(1);-\kappa) \ge \pi - 2v^{-1}(\kappa/2)
= \Sep(S^n(1);\kappa/2,\kappa/2)
\]
by Lemma \ref{lem:comp1}.
\end{proof}
Combining Proposition \ref{prop:ObsDiam-Sep},
Lemmas \ref{lem:comp1} and \ref{lem:comp2} yields
\[
\ObsDiam(X;-\kappa) \le \ObsDiam(S^n(1);-\kappa) = \pi-2v^{-1}(\kappa/2).
\]
The rest is to estimate $\pi-2v^{-1}(\kappa/2)$ from above.
For that, we need the following
\begin{lem} \label{lem:comp-tail}
We have
\[
\int_r^\infty \frac{1}{\sqrt{2\pi}} e^{-\frac{t^2}{2}} \; dt \le \frac{1}{2} e^{-\frac{r^2}{2}}.
\]
for any $r \ge 0$.
\end{lem}
\begin{proof}
Let
\[
f(r) := \frac{1}{2} e^{-\frac{r^2}{2}}
- \int_r^\infty \frac{1}{\sqrt{2\pi}} e^{-\frac{t^2}{2}}\;dt.
\]
Then, $f(0) = f(\infty) = 0$, $f'(r) \ge 0$ for $0 \le r \le 2/\sqrt{2\pi}$,
and $f'(r) \le 0$ for $r \ge 2/\sqrt{2\pi}$.
This completes the proof.
\end{proof}
The following lemma completes the proof of Theorem \ref{thm:comp}.
\begin{lem}
\[
\pi-2v^{-1}(\kappa/2)
\le \frac{2\sqrt{2}}{\sqrt{n-1}} \sqrt{-\log\left(\sqrt{\frac{2}{\pi}}\kappa\right)}.
\]
\end{lem}
\begin{proof}
Setting $r := \pi/2 - v^{-1}(\kappa/2)$ and
$w_n := \int_0^\pi \sin^{n-1} t \; dt$, we see
\begin{align*}
\frac{\kappa}{2} &= \frac{1}{w_n} \int_{\pi/2+r}^\pi \sin^{n-1} t\;dt
= \frac{1}{w_n} \int_r^{\pi/2} \cos^{n-1} t\;dt\\
&= \frac{1}{w_n\sqrt{n-1}} \int_{r\sqrt{n-1}}^{(\pi/2)\sqrt{n-1}}
\cos^{n-1}\frac{s}{\sqrt{n-1}}\;ds
\intertext{and, by $\cos x \le e^{-\frac{x^2}{2}}$ for $0 \le x \le \pi/2$
and by Lemma \ref{lem:comp-tail},}
&\le \frac{1}{w_n\sqrt{n-1}} \int_{r\sqrt{n-1}}^\infty e^{-\frac{s^2}{2}}\;ds
\le \frac{\sqrt{\pi}}{w_n\sqrt{2(n-1)}} e^{-\frac{n-1}{2}r^2}.
\end{align*}
Let $n \ge 2$.
By the integration by parts, we see
$w_{n+2} = n(n+1)^{-1} w_n \ge \sqrt{(n-1)(n+1)^{-1}}\, w_n$,
so that $\sqrt{n+1}\, w_{n+2} \ge \sqrt{n-1}\, w_n$.
This together with
$w_2 = 2$ and $\sqrt{2} \, w_3 = \pi/\sqrt{2} \ge 2$
implies $\sqrt{n-1} \, w_n \ge 2$.
Therefore,
\[
\kappa \le \frac{\sqrt{\pi}}{\sqrt{2}} e^{-\frac{n-1}{2}r^2},
\]
which proves the lemma.
\end{proof}
We have the following corollary to Theorem \ref{thm:comp}.
\begin{cor}
Let $n \ge 2$.
If a closed and connected $n$-dimensional Riemannian manifold $X$
satisfies $\Ric_X \ge K$ for a constant $K > 0$, then
\[
\ObsDiam(X;-\kappa) \le \sqrt{\frac{n-1}{K}} \ObsDiam(S^n(1);-\kappa)
= O(K^{-1/2})
\]
for any $\kappa > 0$.
\end{cor}
\begin{proof}
Assume that $\Ric_X \ge K > 0$.
Setting $t := \sqrt{K/(n-1)}$, we have $\Ric_{tX} \ge K/t^2 = n-1$.
Apply Theorem \ref{thm:comp} for $tX$
and use Proposition \ref{prop:scale-ObsDiam}.
\end{proof}
\begin{ex} \label{ex:SOSUSp}
Let $SO(n)$, $SU(n)$, and $Sp(n)$ be the special orthogonal group,
the special unitary group, and the compact symplectic group, respectively.
For $X_n = SO(n),SU(n),Sp(n)$ we have
\[
\Ric_{X_n} = \left(\frac{\beta(n+2)}{4}-1\right)g_{X_n},
\]
where $\beta = 1$ for $X_n = SO(n)$,
$\beta = 2$ for $X_n = SU(n)$, and $\beta = 4$ for $X_n = Sp(n)$
(see \cite{AGZ}*{(F.6)}).
Therefore,
$\{SO(n)\}_{n=1}^\infty$, $\{SU(n)\}_{n=1}^\infty$, and $\{Sp(n)\}_{n=1}^\infty$
are all L\'evy families with
\[
\ObsDiam(X_n;-\kappa) \le O(n^{-1/2}).
\]
\end{ex}
\section{Spectrum of Laplacian and separation distance}
\label{sec:spec-sep}
The first nonzero eigenvalue of the Laplacian is useful to
detect L\'evy families.
Let $X$ be a compact Riemannian manifold and
$\Delta$ the (nonnegative) Laplacian on $X$.
We equip $X$ with the volume measure $\mu_X$ normalized as $\mu_X(X) =1$
as always.
It is known that the spectrum of $\Delta$ consists of
eigenvalues \index{lambdakX@$\lambda_k(X)$} \index{lambda1X@$\lambda_1(X)$}
\[
0 = \lambda_0(X) \le \lambda_1(X) \le \lambda_2(X) \le \dots
\le \lambda_k(X) \le \dots,
\]
(with multiplicity),
and $\lambda_k(X)$ is divergent to infinity as $k\to\infty$.
Denote by $L_2(X)$ \index{L2X@$L_2(X)$}
the space of square integrable functions on $X$ with respect to $\mu_X$,
and by $\mathcal{L}{\it ip}(X)$ the set of Lipschitz functions on $X$.
A min-max principle \index{min-max principle} says that
\begin{align*}
\lambda_k(X) = \inf_L \sup_{u \in L \setminus \{0\}} R(u),
\end{align*}
where $L$ runs over all $(k+1)$-dimensional linear subspaces of
$L_2(X) \cap \mathcal{L}{\it ip}(X)$ and
$R(u) := \|\grad u\|_{L_2}^2/\|u\|_{L_2}^2$ is the \emph{Rayleigh quotient}.
\index{Rayleigh quotient}
Note that any Lipschitz function on $X$ is differentiable a.e.~by
Rademacher's theorem, so that its gradient vector filed is defined a.e.~on $X$.
The compactness of $X$ implies that $X$ has at most finitely many
connected components.
The number of connected components of $X$ coincides with
the smallest number $k$ with $\lambda_k(X) > 0$.
In particular, $X$ is connected if and only if $\lambda_1(X) > 0$.
Note that the Riemannian distance between two
different connected components is defined to be infinity and that
the separation distance takes values in $[\,0,+\infty\,]$.
\begin{prop} \label{prop:lamk-Sep}
Let $X$ be a compact Riemannian manifold.
Then we have
\[
\Sep(X;\kappa_0,\kappa_1,\dots,\kappa_k)
\le \frac{2}{\sqrt{\lambda_k(X)\min_{i=0,1,\dots,k}\kappa_i}}
\]
for any $\kappa_0,\kappa_1,\dots,\kappa_k > 0$.
\end{prop}
We see a more refined estimate in \cite{CGY}.
\begin{proof
We set $S := \Sep(X;\kappa_0,\dots,\kappa_k)$ for simplicity.
If $S = 0$, then the proposition is trivial.
Assume $S > 0$.
Let $r$ be any number with $0 < r < S$.
There are Borel subsets $A_0,\dots,A_k \subset X$
such that $\mu_X(A_i) \ge \kappa_i$ and $d_X(A_i,A_j) > r$
for any different $i$ and $j$.
Let
\[
f_i(x) := \max\{1-\frac{2}{r}d_X(x,A_i),0\}, \quad x \in X.
\]
$f_i$ are Lipschitz functions on $X$ and are differentiable
a.e.~on $X$.
We see that $f_0,f_1,\dots,f_k$ are $L_2$ orthogonal to each other,
$|\grad f_i| \le 2/r$ a.e., and $\|f_i\|_{L_2}^2 \ge \kappa_i$.
Denote by $L_0$ the linear subspace of $L_2(X)$ generated by
$f_0,\dots,f_k$. We obtain
\begin{align*}
\lambda_k(X) = \inf_{L\; :\;\dim L = k+1} \sup_{u \in L \setminus\{0\}} R(u)
\le \sup_{u \in L_0\setminus\{0\}} R(u),
\end{align*}
where
$R(u) := \|\grad u\|_{L_2}^2/\|u\|_{L_2}^2$.
It is easy to prove that, for any function
$u = \sum_{i=0}^k a_i f_i \in L_0 \setminus \{0\}$,
\[
\|u\|_{L_2}^2 \ge \min_{j=0,\dots,k} \kappa_j \sum_{i=0}^k a_i^2
\quad\text{and}\quad
\|\grad u\|_{L_2}^2 \le \frac{4}{r^2} \sum_{i=0}^k a_i^2,
\]
so that $R(u) \le 4/(r^2\min_i\kappa_i)$.
Therefore, $\lambda_k(X) \le 4/(r^2\min_i\kappa_i)$.
By the arbitrariness of $r$, we obtain the proposition.
\end{proof}
Propositions \ref{prop:ObsDiam-Sep} and \ref{prop:lamk-Sep}
together imply
\begin{cor} \label{cor:ObsDiam-spec}
Let $X$ be a compact Riemannian manifold.
For any $\kappa > 0$ we have
\[
\ObsDiam(X;-2\kappa) \le \Sep(X;\kappa,\kappa)
\le \frac{2}{\sqrt{\lambda_1(X)\,\kappa}}.
\]
In particular, if $\{X_n\}$ is a sequence of compact Riemannian manifolds
such that $\lambda_1(X_n) \to +\infty$ as $n\to\infty$,
then $\{X_n\}$ is a L\'evy family.
\end{cor}
\begin{rem}
It is a famous theorem of Lichnerowicz that
if the Ricci curvature of a closed and connected $n$-dimensional
Riemannian manifold $X$
satisfies $\Ric_X \ge n-1$, $n \ge 2$, then
\[
\lambda_1(X) \ge n.
\]
This together with Corollary \ref{cor:ObsDiam-spec} implies
\[
\ObsDiam(X;-2\kappa) \le \frac{2}{\sqrt{n\kappa}},
\]
which is weaker than Theorem \ref{thm:comp} for small $\kappa > 0$.
\end{rem}
\chapter{Gromov-Hausdorff distance
and distance matrix}
\label{chap:GH-dist-matrix}
\section{Net, covering number, and capacity}
In this and next sections, we define some of the basic terminologies
from metric geometry.
Assume that $X$ is a metric space.
\begin{defn}[Net, $\varepsilon$-covering number, and $\varepsilon$-capacity]
\index{net} \index{epsilon-covering number@$\varepsilon$-covering number}
\index{epsilon-capacity@$\varepsilon$-capacity}
We call a discrete subset of $X$ a \emph{net of $X$}.
Let $\varepsilon > 0$.
A net $\mathcal{N}$ of $X$ is called an \emph{$\varepsilon$-net}
\index{epsilon-net@$\varepsilon$-net}
if $B_\varepsilon(\mathcal{N}) = X$.
A net $\mathcal{N}$ of $X$ is said to be \emph{$\varepsilon$-discrete}
\index{epsilon-discrete net@$\varepsilon$-discrete net}
if $d_X(x,y) > \varepsilon$ for any different
$x,y \in \mathcal{N}$.
Define
\begin{align*}
\Cov_\varepsilon(X) &:= \inf\{\; \#\mathcal{N} \mid
\text{$\mathcal{N}$ is an $\varepsilon$-net of $X$}\;\},\\
\Cap_\varepsilon(X) &:= \sup\{\; \#\mathcal{N} \mid
\text{$\mathcal{N}$ is an $\varepsilon$-discrete net of $X$}\;\},
\end{align*}
where $\#\mathcal{N}$ is the number of points in $\mathcal{N}$.
$\Cov_\varepsilon(X)$ and $\Cap_\varepsilon(X)$ are respectively
called the \emph{$\varepsilon$-covering number}
and the \emph{$\varepsilon$-capacity} of $X$.
\end{defn}
\begin{lem}
For any $\varepsilon > 0$ we have
\[
\Cap_{2\varepsilon}(X) \le \Cov_\varepsilon(X) \le \Cap_\varepsilon(X).
\]
\end{lem}
\begin{proof}
Since a maximal $\varepsilon$-discrete net of $X$ is an $\varepsilon$-net,
we have $\Cov_\varepsilon(X) \le \Cap_\varepsilon(X)$.
To prove the other inequality, we take any $2\varepsilon$-discrete net
$\mathcal{N}$ of $X$ and any $\varepsilon$-net $\mathcal{N}'$ of $X$.
Since any ball of radius $\varepsilon$ covers at most one point in
$\mathcal{N}$, we have $\#\mathcal{N} \le \#\mathcal{N}'$,
which implies $\Cap_{2\varepsilon}(X) \le \Cov_\varepsilon(X)$.
\end{proof}
\begin{defn}[$\varepsilon$-Projection and nearest point projection]
\index{epsilon-projection@$\varepsilon$-projection}
\index{nearest point projection}
Let $A \subset X$ be a subset and let $\varepsilon \ge 0$.
A map $\pi : X \to A$ is called an \emph{$\varepsilon$-projection to $A$}
if
\[
d_X(x,\pi(x)) \le d_X(x,A) + \varepsilon
\]
for any $x \in X$.
A $0$-projection to $A$ is called a \emph{nearest point projection to $A$}.
\end{defn}
\begin{lem} \label{lem:Borel-proj}
For any finite net $\mathcal{N} \subset X$,
there exists a Borel measurable nearest point projection to $\mathcal{N}$.
\end{lem}
\begin{proof}
We set $\{a_i\}_{i=1}^N := \mathcal{N}$.
For a given $x \in X$, let $\{a_{i_j}\}_{j=1}^k$, $k \le N$, be
the set of nearest points in $\mathcal{N}$ to $x$, and let
\[
i(x) := \min_{j=1}^k i_j,\qquad \pi(x) := a_{i(x)}.
\]
This defines a nearest point projection $\pi : X \to \mathcal{N}$.
We prove that $\pi$ is Borel measurable.
In fact, for any $a_i \in \mathcal{N}$,
\begin{align*}
\pi^{-1}(a_i) = \{\;x \in X \mid
\ &d_X(x,a_i) < d_X(x,a_j)\ \text{for any $j < i$},\\
&d_X(x,a_i) \le d_X(x,a_j)\ \text{for any $j > i$}\;\}
\end{align*}
is a Borel subset of $X$, which implies the Borel measurability of
$\pi : X \to \mathcal{N}$.
\end{proof}
\begin{lem} \label{lem:eps-proj}
Let $X$ be a separable metric space.
For any subset $A \subset X$ and any $\varepsilon > 0$,
there exists a Borel measurable $\varepsilon$-projection $\pi$ to $A$
such that, if $A$ is a Borel subset of $X$, then $\pi|A = \id_A$.
\end{lem}
\begin{proof}
Since $X$ is separable, there is a
a dense countable subset $\{x_i\}_{i=1}^\infty \subset X$.
For each $x_i$ there is a point $a_i \in A$ such that
$d_X(x_i,a_i) \le d_X(x_i,A) + \varepsilon/3$.
By setting
\[
B_1 := B_{\varepsilon/3}(x_1) \quad\text{and}\quad
B_{i+1} := B_{\varepsilon/3}(x_{i+1}) \setminus \bigcup_{j=1}^i B_{\varepsilon/3}(x_j)
\]
for $i \ge 1$, the sequence $\{B_i\}_{i=1}^\infty$ is a disjoint covering of $X$.
For $i$ with $B_i \neq \emptyset$ and for $x \in B_i$,
we define $\pi(x) := a_i$ to obtain
a Borel measurable $\varepsilon$-projection $\pi : X \to A$.
In fact, the preimage of any set by $\pi$ is a Borel subset of $X$.
If $A$ is a Borel subset of $X$, then we obtain a desired
$\varepsilon$-projection $\pi'$ by
\[
\pi'(x) :=
\begin{cases}
\pi(x) &\text{if $x \in X \setminus A$},\\
x &\text{if $x \in A$,}
\end{cases}
\]
which is Borel measurable.
This completes the proof.
\end{proof}
\section{Hausdorff and Gromov-Hausdorff distance}
\begin{defn}[Hausdorff distance]
\index{Hausdorff distance} \index{dH@$d_H$}
Let $Z$ be a metric space and $X, Y \subset Z$ two subsets.
The \emph{Hausdorff distance $d_H(X,Y)$ between $X$ and $Y$}
is defined to be the infimum of $\varepsilon \ge 0$ such that
\[
X \subset B_\varepsilon(Y) \quad\text{and}\quad Y \subset B_\varepsilon(X).
\]
\end{defn}
\begin{lem}[cf. \cite{BBI}*{\S 7.3}] \label{lem:dH}
For a metric space $Z$ we denote by $\mathcal{F}(Z)$ the set of closed
subsets of $Z$.
\begin{enumerate}
\item If $Z$ is complete, then so is $(\mathcal{F}(Z),d_H)$.
\item If $Z$ is compact, then so is $(\mathcal{F}(Z),d_H)$.
\end{enumerate}
\end{lem}
\begin{defn}[Gromov-Hausdorff distance]
\index{Gromov-Hausdorff distance}
\index{dGH@$d_{GH}$}
Let $X$ and $Y$ be two compact metric spaces.
We embed $X$ and $Y$ into some metric space $Z$ isometrically
and define the \emph{Gromov-Hausdorff distance $d_{GH}(X,Y)$
between $X$ and $Y$}
to be the infimum of $d_H(X,Y)$ over all such $Z$ and
all such isometric embeddings $X, Y \hookrightarrow Z$.
\end{defn}
\begin{lem}[cf. \cite{BBI}*{\S 7.3}] \label{lem:dGH-complete}
The Gromov-Hausdorff metric $d_{GH}$ is a complete metric on
the set of isometry classes of compact metric spaces.
\end{lem}
\begin{defn}[$\varepsilon$-Isometric map and $\varepsilon$-isometry]
\index{epsilon-isometric map@$\varepsilon$-isometric map}
\index{epsilon-isometry@$\varepsilon$-isometry}
Let $X$ and $Y$ be two metric spaces and let $\varepsilon \ge 0$.
A map $f : X \to Y$ is said to be \emph{$\varepsilon$-isometric}
if
\[
|\;d_Y(f(x),f(y)) - d_X(x,y)\;| \le \varepsilon
\]
for any $x,y \in X$.
A map $f : X \to Y$ is called an \emph{$\varepsilon$-isometry}
if $f$ is $\varepsilon$-isometric and if
\[
B_\varepsilon(f(X)) = Y.
\]
\end{defn}
\begin{lem}[cf. \cite{BBI}*{\S 7.3}] \label{lem:dGH-isom}
For a real number $\varepsilon > 0$ we have the following
{\rm(1)} and {\rm(2)}.
\begin{enumerate}
\item If $d_{GH}(X,Y) < \varepsilon$, then
there exists a $2\varepsilon$-isometry $f : X \to Y$.
\item If there exists an $\varepsilon$-isometry $f : X \to Y$,
then $d_{GH}(X,Y) < 2\varepsilon$.
\end{enumerate}
\end{lem}
\begin{lem}[cf. \cite{BBI}*{\S 7.4}] \label{lem:dGH-precpt}
Let $\mathcal{C}$ be a set of isometry classes of compact metric spaces
such that $\sup_{X \in \mathcal{C}} \diam X < +\infty$,
where $\diam X := \sup_{x,y\in X} d_X(x,y)$ is the diameter of $X$.
Then, the following {\rm(1)}, {\rm(2)}, and {\rm(3)}
are all equivalent to each other.
\begin{enumerate}
\item $\mathcal{C}$ is $d_{GH}$-precompact.
\item For any $\varepsilon > 0$ we have
\[
\sup_{X \in \mathcal{C}} \Cap_\varepsilon(X) < +\infty.
\]
\item For any $\varepsilon > 0$ we have
\[
\sup_{X \in \mathcal{C}} \Cov_\varepsilon(X) < +\infty.
\]
\end{enumerate}
\end{lem}
\section{Distance matrix}
\begin{defn}[Distance matrix]
\index{distance matrix} \index{KNX@$K_N(X)$}
For a metric space $X$ and a natural number $N$,
the \emph{distance matrix $K_N(X)$ of $X$ of order $N$}
is defined to be the set of symmetric matrices $(d_X(x_i,x_j))_{ij}$
of order $N$, where $x_i$, $i=1,2,\dots,N$, run over all
points in $X$.
\end{defn}
If $X$ is compact, then $K_N(X)$ is compact for any $N$.
\begin{lem} \label{lem:KN-isom}
Let $X$ and $Y$ be two compact metric spaces.
If $K_N(X) = K_N(Y)$ for every natural number $N$,
then $X$ and $Y$ are isometric to each other.
\end{lem}
\begin{proof}
Assume that $K_N(X) = K_N(Y)$ for all natural number $N$.
We take any $\varepsilon > 0$ and fix it.
For any net $\{x_i\}_{i=1}^N \subset X$,
since $(d_X(x_i,x_j))_{ij} \in K_N(X) = K_N(Y)$,
there is a net $\{y_i\}_{i=1}^N \subset Y$ such that
$d_X(x_i,x_j) = d_Y(y_i,y_j)$ for all $i,j=1,2,\dots,N$.
It holds that $\{x_i\}_{i=1}^N$ is $\varepsilon$-discrete if and only if
so is $\{y_i\}_{i=1}^N$.
Therefore we have $\Cap_\varepsilon(X) = \Cap_\varepsilon(Y)$.
By setting $N := \Cap_\varepsilon(X) = \Cap_\varepsilon(Y)$,
there are two $\varepsilon$-discrete nets $\{x_i\}_{i=1}^N \subset X$ and
$\{y_i\}_{i=1}^N \subset Y$ such that
$d_X(x_i,x_j) = d_Y(y_i,y_j)$ for all $i,j=1,2,\dots,N$.
Note that $\{x_i\}_{i=1}^N$ and $\{y_i\}_{i=1}^N$ are $\varepsilon$-nets
of $X$ and $Y$, respectively.
For any given point $x \in X$ we find an $i$ in such a way that
$d_X(x_i,x) \le \varepsilon$ and then set $f(x) := y_i$.
This defines a map $f : X \to Y$.
We see that $f$ is a $2\varepsilon$-isometry.
Lemma \ref{lem:dGH-isom} implies that $d_{GH}(X,Y) < 4\varepsilon$.
By the arbitrariness of $\varepsilon > 0$, we have $d_{GH}(X,Y) = 0$,
so that $X$ and $Y$ are isometric to each other.
\end{proof}
The \emph{$l_\infty$ norm} \index{l infinity norm@$l_\infty$ norm}
of a square matrix $A = (a_{ij})$ of order $N$
is defined to be
\[
\|A\|_\infty := \max_{i,j=1}^N |a_{ij}|.
\]
Of course, the $l_\infty$ norm induces a metric,
called the \emph{$l_\infty$ metric}, \index{l infinity metric@$l_\infty$ metric}
on the set of square matrices of order $N$.
\begin{lem} \label{lem:KN-GH}
For any two compact metric spaces $X$ and $Y$
and for any natural number $N$, we have
\[
d_H(K_N(X),K_N(Y)) \le 2\,d_{GH}(X,Y),
\]
where $d_H$ in the left-hand side is the Hausdorff distance
defined by the $l_\infty$ metric on the set of square matrices of order $N$.
\end{lem}
\begin{proof}
Assume that $d_{GH}(X,Y) < \varepsilon$ for a number $\varepsilon$.
$X$ and $Y$ are embedded into some compact metric space $Z$
such that $d_H(X,Y) < \varepsilon$.
We take any matrix $A \in K_N(X)$ and set
$A = (d_X(x_i,x_j))_{ij}$, $\{x_i\}_{i=1}^N \subset X$.
By $d_H(X,Y) < \varepsilon$, there is a point $y_i \in Y$ for each $x_i$
such that $d_Z(x_i,y_i) < \varepsilon$.
Since
\[
|\;d_X(x_i,x_j)-d_Y(y_i,y_j)\;|
\le d_Z(x_i,y_i)+d_Z(x_j,y_j) < 2\varepsilon
\]
for any $i,j = 1,2,\dots,N$,
the matrix $B := (d_Y(y_i,y_j))_{ij}$ satisfies $B \in K_N(Y)$ and
$\|A-B\|_\infty < 2\varepsilon$.
This proves that $K_N(X) \subset B_{2\varepsilon}(K_N(Y))$.
Since this also holds by exchanging $X$ and $Y$, we obtain
\[
d_H(K_N(X),K_N(Y)) \le 2\varepsilon.
\]
This completes the proof.
\end{proof}
\begin{lem} \label{lem:KN-GH-2}
Let $X$ and $X_n$, $n=1,2,\dots$, be compact metric spaces.
If $K_N(X_n)$ Hausdorff converges to $K_N(X)$ as $n\to\infty$
with respect to the $l_\infty$ norm for any natural number $N$,
then $X_n$ Gromov-Hausdorff converges to $X$ as $n\to\infty$.
\end{lem}
\begin{proof}
We first prove that $\{X_n\}$ is $d_{GH}$-precompact.
Setting $\delta_n := d_H(K_N(X_n),K_N(X))$
we have
$K_N(X_n) \subset B_{\delta_n}(K_N(X))$.
This proves that, for any net $\{x_i\}_{i=1}^N \subset X_n$,
there is a net $\{y_i\}_{i=1}^N \subset X$ such that
\[
|\;d_{X_n}(x_i,x_j) - d_X(y_i,y_j)\;| \le \delta_n
\]
for all $i,j=1,2,\dots,N$.
We take any $\varepsilon > 0$ and any $n$ with $\delta_n < \varepsilon/2$.
If $\{x_i\}_{i=1}^N$ is $\varepsilon$-discrete, then
$\{y_i\}_{i=1}^N$ is $\varepsilon/2$-discrete.
Thus we have $\Cap_\varepsilon(X_n) \le \Cap_{\varepsilon/2}(X)$,
which holds for all sufficiently large $n$, so that
$\sup_{n=1,2,\dots} \Cap_\varepsilon(X_n) < +\infty$.
It is also easy to see that $\sup_{n=1,2,\dots} \diam X_n < +\infty$.
By Lemma \ref{lem:dGH-precpt}, $\{X_n\}$ is $d_{GH}$-precompact.
For any $d_{GH}$-convergent subsequence $\{X_{n_i}\}$ of $\{X_n\}$,
Lemma \ref{lem:KN-GH} implies that $d_H(K_N(X_{n_i}),K_N(X')) \to 0$
as $i\to\infty$ for any $N$,
where $X'$ is the limit of $\{X_{n_i}\}$.
Therefore we have $K_N(X) = K_N(X')$ for any $N$.
By Lemma \ref{lem:KN-isom},
$X$ and $X'$ are isometric to each other.
This proves that $X_n$ $d_{GH}$-converges to $X$ as $n\to\infty$.
\end{proof}
\chapter{Box distance}
\label{chap:box-dist}
\section{Basics for the box distance}
In this section, we define the box distance between mm-spaces
and prove its basic properties.
The proofs in this section are mostly taken from \cite{Funano:thesis}.
\begin{defn}[Parameter]
\index{parameter} \index{I@$I$}
Let $I := [\,0,1\,)$ and let $X$ be a topological space
with a Borel probability measure $\mu_X$.
A map $\varphi : I \to X$ is called a \emph{parameter of $X$}
if $\varphi$ is a Borel measurable map such that
\[
\varphi_*\mathcal{L}^1 = \mu_X,
\]
where $\mathcal{L}^1$ denotes the one-dimensional Lebesgue measure on $I$.
\index{L1@$\mathcal{L}^1$}
\end{defn}
Note that $\varphi(I)$ has full measure in $X$.
\begin{lem}
Any mm-space has a parameter.
\end{lem}
\begin{proof}
Let $X$ be an mm-space and let
$A = \{x_i\}_{i=1}^N$, $N \le \infty$, be the set of atoms of $\mu_X$,
where a point $x \in X$ is called an \emph{atom of $\mu_X$} \index{atom}
if $\mu_X\{x\} > 0$.
We assume that $x_i \neq x_j$ for $i\neq j$.
Putting $a_i := \mu_X\{x_i\}$, $b_0 := 0$, $b_i := \sum_{j=1}^i a_j$
for $1 \le i \le N$,
we have $b_N = \mu_X(A)$.
By \cite{Kechris}*{(17.41)},
there is a Borel isomorphism $\varphi : [\,b_N,1\,) \to X \setminus A$
such that $\varphi_*(\mathcal{L}^1|_{[\,b_N,1\,)}) = \mu_X - \sum_i a_i\delta_{x_i}$.
Setting $\varphi|_{[\,b_{i-1},b_i\,)} := x_i$ for $i \ge 1$ defines
a Borel measurable map $\varphi : I \to X$ such that
$\varphi_*\mathcal{L}^1 = \mu_X$.
\end{proof}
\begin{defn}[Pseudo-metric]
\index{pseudo-metric}
A \emph{pseudo-metric} $\rho$ on a set $S$ is defined to be
a function $\rho : S \times S \to [\,0,+\infty\,)$ satisfying that,
for any $x,y,z \in S$,
\begin{enumerate}
\item $\rho(x,x) = 0$,
\item $\rho(y,x) = \rho(x,y)$,
\item $\rho(x,z) \le \rho(x,y) + \rho(y,z)$.
\end{enumerate}
\end{defn}
Note that $\rho(x,y) = 0$ does not necessarily imply $x = y$.
A Lipschitz map between two spaces with pseudo-metrics
is defined in the same way as usual.
\begin{defn}[Box distance]
\index{box distance} \index{boxrho1rho2@$\square(\rho_1,\rho_2)$}
For two pseudo-metrics $\rho_1$ and $\rho_2$ on $I$, we define
$\square(\rho_1,\rho_2)$ to be the infimum of $\varepsilon \ge 0$
satisfying that there exists a Borel subset $I_0 \subset I$
such that
\begin{align}
& |\,\rho_1(s,t)-\rho_2(s,t)\,| \le \varepsilon
\quad\text{for any $s,t \in I_0$},\tag{1}\\
& \mathcal{L}^1(I_0) \ge 1-\varepsilon.\tag{2}
\end{align}
We define the \emph{box distance $\square(X,Y)$ between
two mm-spaces $X$ and $Y$} \index{boxXY@$\square(X,Y)$} to be
the infimum of $\square(\varphi^*d_X,\psi^*d_Y)$, where
$\varphi : I \to X$ and $\psi : I \to Y$ run over all parameters
of $X$ and $Y$, respectively, and where
$\varphi^*d_X(s,t) := d_X(\varphi(s),\varphi(t))$ for $s,t \in I$.
\index{phistardX@$\varphi^*d_X$}
\end{defn}
Note that $\square(\rho_1,\rho_2) \le 1$ for any two pseudo-metrics
$\rho_1$ and $\rho_2$,
so that $\square(X,Y) \le 1$ for any two mm-spaces $X$ and $Y$.
We are going to prove that $\square$ is a metric on the set of
mm-isomorphism classes of mm-spaces.
\begin{defn}[Distance matrix distribution]
\index{distance matrix distribution}
Denote by $M_N$ the set of real symmetric matrices of order $N$.
For an mm-space $X$, we define a map $\kappa_N : X^N \to M_N$ by
\[
\kappa_N(x_1,x_2,\dots,x_N) := (d_X(x_i,x_j))_{ij}
\]
for $(x_1,x_2,\dots,x_N) \in X^N$.
The \emph{distance matrix distribution of $X$ of order $N$}
is defined to be the push-forward
\[
\underline{\mu}_N^X := (\kappa_N)_*\mu_X^{\otimes N}
\]
\index{muNX@$\underline{\mu}_N^X$}
of the $N$-times product of $\mu_X$ by the map $\kappa_N$.
\end{defn}
\begin{lem} \label{lem:box}
If $\square(X,Y) = 0$ for two mm-spaces, then
\[
\underline{\mu}_N^X = \underline{\mu}_N^Y
\]
for any natural number $N$.
\end{lem}
\begin{proof}
Assume that $\square(X,Y) = 0$ for two mm-spaces $X$ and $Y$.
Then, there are parameters $\varphi_n$ and $\psi_n$ of $X$ and $Y$
respectively, $n=1,2,\dots$, such that
$\square(\varphi_n^*d_X,\psi_n^*d_Y) \to 0$ as $n\to\infty$.
We have $\varepsilon_n \to 0+$ and $I_n \subset I$ in such a way that
$\mathcal{L}^1(I_n) \ge 1-\varepsilon_n$ and
\[
|\;\varphi_n^*d_X(s,t) - \psi_n^*d_Y(s,t)\;| \le \varepsilon_n
\]
for any $s,t \in I_n$.
Let $f : M_N \to \field{R}$ be a uniformly continuous and bounded function
and set
\begin{align*}
f_{\varphi_n}(s_1,s_2,\dots,s_N) &:= f((\varphi_n^*d_X(s_i,s_j))_{ij}),\\
f_{\psi_n}(s_1,s_2,\dots,s_N) &:= f((\psi_n^*d_X(s_i,s_j))_{ij})
\end{align*}
for $(s_1,s_2,\dots,s_N) \in I^N$.
Note that $f_{\varphi_n}$ and $f_{\psi_n}$ are both Borel measurable functions.
The uniform continuity of $f$ shows that
\begin{equation}
\label{eq:dist-distr1}
\lim_{n\to\infty} \sup_{I_n^N} |\;f_{\varphi_n}-f_{\psi_n}\;| = 0.
\end{equation}
We have
\begin{equation}
\label{eq:dist-distr2}
\int_{M_N} f \; d\underline{\mu}_N^X
= \int_{X^N} f\circ\kappa_N \; d\mu_X^{\otimes N}
= \int_{I^N} f_{\varphi_n} \; d\mathcal{L}^N,
\end{equation}
where $\mathcal{L}^N$ denotes the $N$-dimensional Lebesgue measure.
\index{LN@$\mathcal{L}^N$}
\eqref{eq:dist-distr2} also holds
if we replace $X$ and $\varphi_n$ with $Y$ and $\psi_n$, respectively.
Since
\[
\limsup_{n\to\infty} \left| \int_{I^N \setminus I_n^N} f_{\varphi_n}
\; d\mathcal{L}^N \right|
\le \sup|f| \lim_{n\to\infty} \mathcal{L}^N(I^N \setminus I_n^N) = 0,
\]
\eqref{eq:dist-distr1} and \eqref{eq:dist-distr2} together imply
\begin{align*}
\int_{M_N} f \; d\underline{\mu}_N^X
&= \lim_{n\to\infty} \int_{I_n^N} f_{\varphi_n} \; d\mathcal{L}^N
= \lim_{n\to\infty} \int_{I_n^N} f_{\psi_n} \; d\mathcal{L}^N
= \int_{M_N} f \; d\underline{\mu}_N^Y.
\end{align*}
This completes the proof.
\end{proof}
\begin{thm}[mm-Reconstruction theorem] \label{thm:mm-reconst}
\index{mm-reconstruction theorem}
Let $X$ and $Y$ be two mm-spaces.
If $\underline{\mu}_N^X = \underline{\mu}_N^Y$ for any natural number $N$, then
$X$ and $Y$ are mm-isomorphic to each other.
\end{thm}
For the complete proof of this theorem, we refer to \cite{Gromov}*{3$\frac12$.5},
\cite{Vershik}*{\S 2}, and \cite{Kondo}.
The following is taken from \cite{Kondo}.
\begin{proof}[Sketch of Proof]
We define the distance matrix distribution
\[
\underline{\mu}_\infty^X := (\kappa_\infty)_*\mu_X^{\otimes\infty}
\]
of $X$ of infinite order
as an Borel probability measure
on the set, say $M_\infty$, of real symmetric matrices of infinite order
in the same manner as before,
where $\kappa_\infty : X^\infty \to M_\infty$ is defined by
\[
\kappa_\infty((x_i)_{i=1}^\infty) := (d_X(x_i,x_j))_{i,j=1}^\infty
\]
and $M_\infty$ is equipped with the product topology.
Taking the completion of $\underline{\mu}_\infty^X$,
we assume that $\underline{\mu}_\infty^X$ is a complete measure.
Assume that $\underline\mu_N^X = \underline\mu_N^Y$ for any natural number $N$.
Let us first prove $\underline{\mu}_\infty^X = \underline{\mu}_\infty^Y$.
In fact, it is easy to see that, for $1 \le N < N' \le \infty$, the measure
$\underline\mu_N^X$ coincides with the push-forward of
$\underline{\mu}_{N'}^X$ by the natural projection from $M_{N'}$ to $M_N$.
The Kolmogorov extension theorem tells us that
$\underline{\mu}_\infty^X$ is determined only by $\underline\mu_N^X$, $N = 1,2,\dots$.
Therefore, by the assumption, we have $\underline{\mu}_\infty^X = \underline{\mu}_\infty^Y$.
Let $E_X \subset X^\infty$ be the set of uniformly distributed sequences on $X$, i.e.,
$(x_i)_{i=1}^\infty \in E_X$ if and only if
\[
\lim_{N\to\infty} \frac{1}{N} \sum_{i=1}^N f(x_i)
= \int_X f \; d\mu_X
\]
for any bounded continuous function $f : X \to \field{R}$.
Note that each $(x_i)_{i=1}^\infty \in E_X$ is dense in $X$.
It is well-known that $\mu_X^{\otimes\infty}(E_X) = 1$.
Since $\kappa_\infty^{-1}(\kappa_\infty(E_X)) \supset E_X$,
the set $\kappa_\infty^{-1}(\kappa_\infty(E_X))$ is $\mu_X^{\otimes\infty}$-measurable,
which implies the $\underline{\mu}_\infty^X$-measurablity of $\kappa_\infty(E_X)$.
We also obtain the $\underline{\mu}_\infty^Y$-measurablity of $\kappa_\infty(E_Y)$
by the same reason.
We have $\underline{\mu}_\infty^X(\kappa_\infty(E_X))
= \underline{\mu}_\infty^Y(\kappa_\infty(E_Y)) = 1$, which together with
$\underline{\mu}_\infty^X = \underline{\mu}_\infty^Y$ imlies
$\kappa_\infty(E_X) \cap \kappa_\infty(E_Y) \neq \emptyset$.
Therefore, there are two sequences $(x_i)_{i=1}^\infty \in E_X$ and $(y_i)_{i=1}^\infty \in E_Y$
such that $d_X(x_i,x_j) = d_Y(y_i,y_j)$ for all $i,j = 1,2,\dots$.
The map $x_i \mapsto y_i$ extends to an isometry $F : X \to Y$.
For any bounded continuous function $f : Y \to \field{R}$, we have
\[
\int_Y f \; d\mu_Y = \lim_{N\to\infty} \frac{1}{N} \sum_{i=1}^N f(y_i)
= \lim_{N\to\infty} \frac{1}{N} \sum_{i=1}^N f(F(x_i))
= \int_X f\circ F \; d\mu_X,
\]
which implies $F_*\mu_X = \mu_Y$.
Thus, $X$ and $Y$ are mm-isomorphic to each other.
\end{proof}
\begin{lem} \label{lem:box-pseudo-tri}
$\square$ satisfies a triangle inequality between
pseudo-metrics.
\end{lem}
\begin{proof}
Assume that $\square(\rho_1,\rho_2) < \varepsilon$
and $\square(\rho_2,\rho_3) < \delta$ for three pseudo-metrics $\rho_1$,
$\rho_2$, and $\rho_3$ and for two numbers $\varepsilon, \delta > 0$.
It suffices to prove that $\square(\rho_1,\rho_3) \le \varepsilon+\delta$.
There are two Borel subsets $I_0,I_0' \subset I$ such that
\begin{align*}
&\mathcal{L}^1(I_0) \ge 1-\varepsilon,\qquad
\mathcal{L}^1(I_0') \ge 1-\delta,\\
&|\;\rho_1(s,t)-\rho_2(s,t)\;| \le \varepsilon
\quad\text{for any $s,t \in I_0$},\\
&|\;\rho_2(s,t)-\rho_3(s,t)\;| \le \delta
\quad\text{for any $s,t \in I_0'$}.
\end{align*}
Setting $I_0'' := I_0 \cap I_0'$, we have
\begin{align*}
\mathcal{L}^1(I_0'') &= \mathcal{L}^1(I_0) + \mathcal{L}^1(I_0')
- \mathcal{L}^1(I_0 \cup I_0')\\
&\ge 2-\varepsilon-\delta- \mathcal{L}^1(I_0 \cup I_0')
\ge 1-(\varepsilon+\delta)
\end{align*}
and, for any $s,t \in I_0''$,
\begin{align*}
|\;\rho_1(s,t)-\rho_3(s,t)\;| &\le
|\;\rho_1(s,t)-\rho_2(s,t)\;| + |\;\rho_2(s,t)-\rho_3(s,t)\;|\\
&\le \varepsilon + \delta.
\end{align*}
Therefore we obtain $\square(\rho_1,\rho_3) \le \varepsilon+\delta$.
This completes the proof.
\end{proof}
The following lemma is needed to prove a triangle inequality
for the box distance between mm-spaces.
\begin{lem} \label{lem:box-tri}
Let $\varphi : I \to X$ and $\psi : I \to X$ be two parameters
of an mm-space $X$. Then, for any $\varepsilon > 0$
there exist two Borel isomorphisms $f : I \to I$ and $g : I \to I$
such that
\begin{enumerate}
\item $f_*\mathcal{L}^1 = g_*\mathcal{L}^1 = \mathcal{L}^1$,
\item $\square((\varphi\circ f)^*d_X,(\psi\circ g)^*d_X) \le \varepsilon$.
\end{enumerate}
\end{lem}
\begin{proof}
Let $\varepsilon > 0$ be a given number.
There is a sequence $\{B_i\}_{i=1}^N$ of disjoint Borel subsets of $X$,
$N \le \infty$,
such that $\bigcup_{i=1}^N B_i = X$ and $\diam B_i \le \varepsilon/2$
for any $i$.
We set $a_i := \mu_X(B_i) = \mathcal{L}^1(\varphi^{-1}(B_i))$, $b_0 := 0$,
and $b_i := \sum_{j \le i} a_j$.
Since each $\varphi^{-1}(B_i)$ is a Borel subset of $I$,
it is a standard Borel space.
Thus, there is a Borel isomorphism
$f_i : [\,b_{i-1},b_i\,) \to \varphi^{-1}(B_i)$ for each $i \ge 1$
such that
$(f_i)_*(\mathcal{L}^1|_{[\,b_{i-1},b_i\,)}) = \mathcal{L}^1|_{\varphi^{-1}(B_i)}$.
Combining all $f_i$'s defines a Borel isomorphism
$f : I \to I$ with the property that $f([\,b_{i-1},b_i\,)) = \varphi^{-1}(B_i)$
for any $i \ge 1$ and $f_*\mathcal{L}^1 = \mathcal{L}^1$.
In the same way, we have a Borel isomorphism $g : I \to I$ such that
$g([\,b_{i-1},b_i\,)) = \psi^{-1}(B_i)$
for any $i \ge 1$ and $g_*\mathcal{L}^1 = \mathcal{L}^1$.
For any $s \in I$, both
$(\varphi\circ f)(s)$ and $(\psi\circ g)(s)$ belong to
a common $B_i$ and hence
$d_X((\varphi\circ f)(s),(\psi\circ g)(s)) \le \varepsilon/2$.
Therefore, for any $s,t \in I$,
\[
|\;d_X((\varphi\circ f)(s),(\varphi\circ f)(t)) -
d_X((\psi\circ g)(s),(\psi\circ g)(t))\;| \le \varepsilon,
\]
which implies that
$\square((\varphi\circ f)^*d_X,(\psi\circ g)^*d_X) \le \varepsilon$.
This completes the proof.
\end{proof}
We are now ready to prove the following
\begin{thm} \label{thm:box}
The box metric $\square$ is a metric on the set $\mathcal{X}$ of
mm-isomorphism classes of mm-spaces.
\end{thm}
\begin{proof}
Combining Lemma \ref{lem:box} and Theorem \ref{thm:mm-reconst} yields
that $\square(X,Y) = 0$ if and only if $X$ and $Y$ are mm-isomorphic to
each other.
The symmetricity for $\square$ is clear.
Let us prove the triangle inequality
$\square(X,Z) \le \square(X,Y) + \square(Y,Z)$
for three mm-spaces $X$, $Y$, and $Z$.
We take any parameters $\varphi : I \to X$, $\psi,\psi' : I \to Y$,
and $\xi : I \to Z$, and fix them.
Lemma \ref{lem:box-tri} implies that, for any $\varepsilon > 0$,
there are two Borel isomorphisms
$f : I \to I$ and $g : I \to I$ such that
$\square((\psi\circ f)^*d_Y,(\psi'\circ g)^*d_Y) \le \varepsilon$.
It follows from Lemma \ref{lem:box-pseudo-tri} that
\begin{align*}
&\square(\varphi^*d_X,\psi^*d_Y) + \square({\psi'}^*d_Y,\xi^*d_Z)\\
&= \square((\varphi\circ f)^*d_X,(\psi\circ f)^*d_Y)
+ \square((\psi'\circ g)^*d_Y,(\xi\circ g)^*d_Z)\\
&\ge \square((\varphi\circ f)^*d_X,(\xi\circ g)^*d_Z)
- \square((\psi\circ f)^*d_Y,(\psi'\circ g)^*d_Y)\\
&\ge \square(X,Z) - \varepsilon.
\end{align*}
Taking the infimum of the left-hand side for any parameters
$\varphi,\psi,\psi',\xi$ yields
$\square(X,Y) + \square(Y,Z) \ge \square(X,Z) - \varepsilon$.
Since $\varepsilon > 0$ is arbitrary, this completes the proof.
\end{proof}
Recall that $tX := (X,td_X,\mu_X)$
for an mm-space $X$ and for $t > 0$.
\begin{lem} \label{lem:box-scale}
For any two mm-spaces $X$ and $Y$, we have the following {\rm(1)}
and {\rm(2)}.
\begin{enumerate}
\item $\square(tX,tY)$ is monotone nondecreasing in $t > 0$.
\item $t^{-1}\square(tX,tY)$ is monotone nonincreasing in $t > 0$.
\end{enumerate}
\end{lem}
The proof of the lemma is left for the reader.
\begin{prop} \label{prop:box-di}
Let $X$ be a complete separable metric space.
For any two Borel probability measures $\mu$ and $\nu$ on $X$,
we have
\[
\frac{1}{2}\,\square((X,\mu),(X,\nu)) \le
\square((2^{-1}X,\mu),(2^{-1}X,\nu)) \le d_P(\mu,\nu).
\]
\end{prop}
\begin{proof}
The first inequality follows from Lemma \ref{lem:box-scale}(2).
We prove the second.
Assume that $d_P(\mu,\nu) < \varepsilon$ for a number $\varepsilon$.
It suffices to prove that
$\square((2^{-1}X,\mu),(2^{-1}X,\nu)) \le \varepsilon$.
By Strassen's theorem (Theorem \ref{thm:di-tra}),
there is an $\varepsilon$-transportation $m$ between $\mu$ and $\nu$
with $\defi m \le \varepsilon$.
Namely, $m$ is a transport plan between two measures $\mu'$ and $\nu'$
such that
$\mu' \le \mu$, $\nu' \le \nu$, $1-m(X\times X) \le \varepsilon$,
and $\supp m \subset \Delta_\varepsilon$.
If $m(X \times X) = 1$, then we set $m_+ := m$.
If otherwise, we set
\[
m_+ := m + \frac{(\mu-\mu') \times (\nu-\nu')}{1-m(X\times X)}.
\]
Note that $m(X\times X) = \mu'(X) = \nu'(X)$.
It is easy to see that $m_+$ is a transport plan between $\mu$ and $\nu$.
Let $\Phi : I \to X \times X$ be a parameter of $(X \times X,m_+)$,
and let $\varphi := \pr_1\circ\Phi : I \to X$ and
$\psi := \pr_2\circ\Phi : I \to X$, where
$\pr_i : X \times X \to X$, $i=1,2$, are the projections.
Let us prove that $\varphi$ is a parameter of $(X,\mu)$.
In fact, for any Borel subset $A \subset X$,
\begin{align*}
\varphi_*\mathcal{L}^1(A) &= \mathcal{L}^1(\varphi^{-1}(A))
= \mathcal{L}^1(\Phi^{-1}(A \times X))\\
&= \Phi_*\mathcal{L}^1(A \times X) = m_+(A \times X) = \mu(A).
\end{align*}
In the same way, we see that $\psi$ is a parameter of $(X,\nu)$.
Set $I_0 := \Phi^{-1}(\Delta_\varepsilon)$.
To obtain $\square((2^{-1}X,\mu),(2^{-1}X,\nu)) \le \varepsilon$,
it suffices to prove that
\begin{enumerate}
\item $\mathcal{L}^1(I_0) \ge 1-\varepsilon$,
\item $|d_X(\varphi(s),\varphi(t)) - d_X(\psi(s),\psi(t))| \le 2\varepsilon$
for any $s,t \in I_0$.
\end{enumerate}
For (1), we have
\[
\mathcal{L}^1(I_0) = \Phi_*\mathcal{L}^1(\Delta_\varepsilon)
= m_+(\Delta_\varepsilon) \ge m(\Delta_\varepsilon)
= m(X \times X) \ge 1-\varepsilon.
\]
We prove (2).
Take any $s,t \in I_0$.
Since $\Phi(s),\Phi(t) \in \Delta_\varepsilon$,
we have $d_X(\varphi(s),\psi(s)) \le \varepsilon$ and
$d_X(\varphi(t),\psi(t)) \le \varepsilon$, so that,
by a triangle inequality,
the left-hand side of (2) is
\[
\le d_X(\varphi(s),\psi(s)) + d_X(\varphi(t),\psi(t))
\le 2\varepsilon.
\]
This completes the proof.
\end{proof}
To obtain the completeness of $\square$ on $\mathcal{X}$,
we need the following
\begin{lem}[Union lemma] \label{lem:union}
\index{union lemma}
Let $X_n$, $n=1,2,\dots$, be mm-spaces such that
\[
\square(2^{-1}X_n,2^{-1}X_{n+1}) < \varepsilon_n
\]
for any $n$ and
for a sequence of real numbers $\varepsilon_n$, $n=1,2,\dots$.
Then, there exists a metric on the disjoint union of all $X_n$'s
that is an extension of each $X_n$ such that
\[
d_P(\mu_{X_n},\mu_{X_{n+1}}) \le \varepsilon_n
\]
for any $n=1,2,\dots$.
\end{lem}
\begin{proof}
Assume that $\square(2^{-1}X_n,2^{-1}X_{n+1}) < \varepsilon_n$
for any $n=1,2,\dots$.
Then, for each $n$ there are two parameters $\varphi_n : I \to X_n$,
$\psi_n : I \to X_{n+1}$, and a Borel subset $I_n \subset I$ such that
$\mathcal{L}^1(I_n) \ge 1-\varepsilon_n$ and
\[
|\;d_{X_n}(\varphi_n(s),\varphi_n(t))-d_{X_{n+1}}(\psi_n(s),\psi_n(t))\;|
\le 2\varepsilon_n
\]
for any $s,t \in I_n$.
Let $Z$ be the disjoint union of all $X_n$'s.
We define a function $d_Z : Z \times Z \to [\,0,+\infty\,)$ as follows:
For $x,y \in X_n$, we set
\[
d_Z(x,y) := d_{X_n}(x,y).
\]
For $x \in X_n$ and $y \in X_{n+1}$,
\[
d_Z(x,y) := \inf_{s \in I_n}(d_{X_n}(x,\varphi_n(s)) + d_{X_{n+1}}(\psi_n(s),y))
+ \varepsilon_n
\]
For $x \in X_n$ and $y \in X_m$ with $n + 2 \le m$,
\[
d_Z(x,y) := \inf \sum_{i=n}^{m-1} d_Z(x_i,x_{i+1}),
\]
where $x_n := x$, $x_m := y$, and where $x_i$ with $n < i < m$ runs over all
points in $X_i$.
For $x \in X_n$ and $y \in X_m$ with $n > m$,
\[
d_Z(x,y) := d_Z(y,x).
\]
It is easy to check that $d_Z$ is a metric on $Z$.
Let $f_n : I \to Z \times Z$ be a map defined by
$f_n(s) := (\varphi_n(s),\psi_n(s))$, and let $m := (f_n)_*\mathcal{L}^1|_{I_n}$.
Setting $\Delta_{\varepsilon_n}
:= \{\;(x,y) \in Z \times Z \mid d_Z(x,y) \le \varepsilon_n\;\}$,
we have
\begin{align*}
&m(Z \times Z \setminus \Delta_{\varepsilon_n})
= \mathcal{L}^1(I_n \cap f_n^{-1}(Z \times Z \setminus \Delta_{\varepsilon_n}))\\
&= \mathcal{L}^1(\{\;s \in I_n
\mid d_Z(\varphi_n(s),\psi_n(s)) > \varepsilon_n\;\}) = 0
\end{align*}
and also, for any Borel subset $A \subset Z$,
\[
m(A \times Z) = \mathcal{L}^1(\varphi_n^{-1}(A) \cap I_n)
\le \mathcal{L}^1(\varphi_n^{-1}(A)) = \mu_{X_n}(A)
\]
as well as $m(Z \times A) \le \mu_{X_{n+1}}(A)$.
Therefore $m$ is an $\varepsilon_n$-transportation between $\mu_{X_n}$
and $\mu_{X_{n+1}}$.
Note that $X_n \cup X_{n+1}$ is a complete separable subspace of $Z$.
Since $\defi m = 1-m(Z\times Z) = 1-\mathcal{L}^1(I_n) \le \varepsilon_n$,
Strassen's theorem (Theorem \ref{thm:di-tra}) tells us that
$d_P(\mu_{X_n},\mu_{X_{n+1}}) \le \varepsilon_n$.
This completes the proof.
\end{proof}
\begin{thm} \label{thm:box-complete}
The box metric $\square$ is complete on $\mathcal{X}$.
\end{thm}
\begin{proof}
Let $\{X_n\}_{n=1}^\infty$ be a $\square$-Cauchy sequence of mm-spaces.
It suffices to prove that $\{X_n\}_{n=1}^\infty$ has a $\square$-convergent
subsequence.
Replacing $\{X_n\}_{n=1}^\infty$ with a subsequence, we assume that
$\square(X_n,X_{n+1}) < 2^{-n}$ for any $n$.
By the union lemma (Lemma \ref{lem:union}), we have a metric $d_Z$
on the disjoint union
$Z$ of all $X_n$'s such that $d_P(\mu_{X_n},\mu_{X_{n+1}}) \le 2^{-n}$ for any $n$.
We see that $\{\mu_{X_n}\}_{n=1}^\infty$ is a $d_P$-Cauchy sequence.
Let $(\bar Z,d_{\bar Z})$ be the completion of $(Z,d_Z)$.
Note that $\bar Z$ is a complete separable metric space.
Since the set of Borel probability measures on $\bar Z$ is $d_P$-complete
(see Lemma \ref{lem:conv-meas}),
the sequence $\{\mu_{X_n}\}_{n=1}^\infty$ $d_P$-converges
to some Borel probability measure, say $\mu_\infty$, on $\bar Z$.
Proposition \ref{prop:box-di} proves that
$X_n$ $\square$-converges to $(\bar Z,\mu_\infty)$.
This completes the proof.
\end{proof}
\begin{ex}
We have $\lim_{n\to\infty} \square(S^n(1),S^{n+1}(1)) = 0$.
This is because, embedding $S^n(1)$ into $S^{n+1}(1)$ naturally, we have
\[
\lim_{n\to\infty} \sigma^{n+1}(S^{n+1}(1) \setminus U_\varepsilon(S^n(1))) = 0
\]
for any $\varepsilon > 0$, which implies
\[
\square(S^n(1),S^{n+1}(1)) \le 2 d_P(\sigma^n,\sigma^{n+1})
\to 0 \quad\text{as $n\to\infty$.}
\]
However, $\{S^n(1)\}$ is not a $\square$-convergent sequence
as is seen in Corollary \ref{cor:homog}.
\end{ex}
\begin{rem}
\index{Gromov-Prohorov distance} \index{dGP@$d_{GP}$}
We define the \emph{Gromov-Prohorov distance $d_{GP}(X,Y)$
between two mm-spaces $X$ and $Y$}
to be the infimum of $d_P(\mu_X,\mu_Y)$ for all metrics on
the disjoint union of $X$ and $Y$ that are extensions
of $d_X$ and $d_Y$.
As a direct consequence of
Proposition \ref{prop:box-di} and the union lemma,
we obtain
\[
\square(2^{-1}X,2^{-1}Y) = d_{GP}(X,Y)
\]
and, in particular,
\[
\frac{1}{2}\,\square(X,Y) \le d_{GP}(X,Y) \le \square(X,Y)
\]
for any two mm-spaces $X$ and $Y$.
This is proved by L\"ohr \cite{Lohr}.
\end{rem}
\section{Finite approximation}
In this section, we prove that
any mm-space can be approximated by a finite mm-space,
i.e., an mm-space consisting of at most finitely many points.
By using this, we study an $\varepsilon$-mm-isomorphism
and the $\square$-compactness of a family of mm-spaces.
\begin{defn}[$\varepsilon$-Supporting net]
\index{epsilon-supporting net@$\varepsilon$-supporting net}
Let $X$ be an mm-space and $\mathcal{N}$ a net of $X$.
For a real number $\varepsilon > 0$, we say that
$\mathcal{N}$ \emph{$\varepsilon$-supports $X$}
if
\[
\mu_X(B_\varepsilon(\mathcal{N})) \ge 1-\varepsilon.
\]
\end{defn}
\begin{lem} \label{lem:eps-supp}
Let $X$ be an mm-space.
For any $\varepsilon > 0$ there exists a finite net of $X$
that $\varepsilon$-supports $X$.
\end{lem}
\begin{proof}
Since $X$ is separable,
we have a dense countable subset $\{a_i\}_{i=1}^\infty \subset X$.
Since $\bigcup_{i=1}^\infty B_\varepsilon(a_i) = X$, we have
$\lim_{n\to\infty} \mu_X(\bigcup_{i=1}^n B_\varepsilon(a_i)) = \mu(X) = 1$.
Therefore, there is a number $n$ such that
$\mu_X(\bigcup_{i=1}^n B_\varepsilon(a_i)) \ge 1-\varepsilon$.
$\mathcal{N} := \{a_i\}_{i=1}^n$ is a desired net.
\end{proof}
Let $X$ be an mm-space
and $\mathcal{N}$ a finite net $\varepsilon$-supporting $X$
for a number $\varepsilon > 0$.
We have a Borel measurable nearest point projection
$\pi_{\mathcal{N}} : X \to \mathcal{N}$
(see Lemma \ref{lem:Borel-proj}).
\begin{lem} \label{lem:pi-di}
We have
\[
d_P((\pi_{\mathcal{N}})_*\mu_X,\mu_X) \le \varepsilon.
\]
\end{lem}
\begin{proof}
Since
\[
\mu_X(\{\;x \in X \mid d_X(\pi_{\mathcal{N}}(x),x) > \varepsilon\;\})
= \mu_X(X \setminus B_\varepsilon(\mathcal{N})) \le \varepsilon,
\]
we have $\dKF(\pi_{\mathcal{N}},\id_X) \le \varepsilon$,
where $\id_X : X \to X$ is the identity map.
By Lemma \ref{lem:di-me},
$d_P((\pi_{\mathcal{N}})_*\mu_X,\mu_X) \le \dKF(\pi_{\mathcal{N}},\id_X)
\le \varepsilon$.
\end{proof}
\begin{prop} \label{prop:fin-approx}
Let $X$ be an mm-space.
For any $\varepsilon > 0$ there exists a finite mm-space $\dot X$
such that
\[
\square(X,\dot X) \le \varepsilon.
\]
\end{prop}
\begin{proof}
By Lemma \ref{lem:eps-supp}, there is a finite net
$\mathcal{N} \subset X$ $(\varepsilon/2)$-supporting $X$.
By letting $\dot X := (\mathcal{N},d_X,(\pi_{\mathcal{N}})_*\mu_X)$,
Proposition \ref{prop:box-di} and Lemma \ref{lem:pi-di} together imply
that $\square(X,\dot X) \le 2 d_P(\mu_X,(\pi_{\mathcal{N}})_*\mu_X) \le \varepsilon$.
\end{proof}
Let $X$ and $Y$ be two mm-spaces and $f : X \to Y$ a Borel measurable map.
\begin{defn}[$\varepsilon$-mm-Isomorphism] \label{defn:mm-iso}
\index{epsilon-mm-isomorphism@$\varepsilon$-mm-isomorphism}
Let $\varepsilon \ge 0$ be a real number.
We say that $f$ is \emph{$\varepsilon$-mm-isomorphism} if
there exists a Borel subset $X_0 \subset X$ such that
\begin{align*}
& \mu_X(X_0) \ge 1-\varepsilon,\tag{1}\\
& |\;d_X(x,y) - d_Y(f(x),f(y))\;| \le \varepsilon
\qquad\text{for any $x,y \in X_0$,}\tag{2}\\
& d_P(f_*\mu_X,\mu_Y) \le \varepsilon.\tag{3}
\end{align*}
We call $X_0$ a \emph{non-exceptional domain of $f$}.
\index{non-exceptional domain}
\end{defn}
Clearly, a $0$-mm-isomorphism means an mm-isomorphism.
\begin{lem} \label{lem:box-eps-mm-iso}
\begin{enumerate}
\item If there exists an $\varepsilon$-mm-isomorphism $: X \to Y$,
then $\square(X,Y) \le 3\varepsilon$.
\item If $\square(X,Y) < \varepsilon$, then
there exists a $3\varepsilon$-mm-isomorphism $: X \to Y$.
\end{enumerate}
\end{lem}
\begin{proof}
We prove (1).
Assume that there is an $\varepsilon$-mm-isomorphism $f : X \to Y$.
Let $X_0 \subset X$ be a non-exceptional domain of $f$.
We take a parameter $\varphi : I \to X$.
We see that $\psi := f \circ \varphi : I \to Y$ is a parameter of
$(Y,f_*\mu_X)$.
Setting $I_0 := \varphi^{-1}(X_0)$, we have,
by Definition \ref{defn:mm-iso}(2),
\[
|d_X(\varphi(s),\varphi(t))-d_Y(\psi(s),\psi(t))| \le \varepsilon
\]
for any $s,t \in I_0$.
Since $\mathcal{L}^1(I_0) = \varphi_*\mathcal{L}^1(X_0) = \mu_X(X_0)
\ge 1-\varepsilon$, we have
$\square(X,(Y,f_*\mu_X)) \le \varepsilon$.
By Proposition \ref{prop:box-di} and Definition \ref{defn:mm-iso}(3),
$\square((Y,f_*\mu_X),Y) \le 2 d_P(f_*\mu_X,\mu_Y) \le 2\varepsilon$.
A triangle inequality proves that
$\square(X,Y) \le 3\varepsilon$.
We prove (2). Assume that $\square(X,Y) < \varepsilon$.
Let $\varepsilon' := (\varepsilon-\square(X,Y))/4$.
There is a finite net $\mathcal{N} \subset X$ $\varepsilon'$-supporting
$X$ such that, by setting
$\dot X := (\mathcal{N},d_X,(\pi_{\mathcal{N}})_*\mu_X)$,
we have $\square(X,\dot X) \le 2\varepsilon'$ and
$\mu_X(B_\varepsilon(\mathcal{N})) \ge 1-\varepsilon$.
A triangle inequality proves that
\[
\square(\dot X,Y) \le 2\varepsilon' + \square(X,Y) < \varepsilon.
\]
Therefore, there are two parameters $\varphi$ and $\psi$ of $\dot{X}$ and $Y$,
respectively, and a Borel subset $I_0 \subset I$ such that
\begin{align}
&\mathcal{L}^1(I_0) \ge 1-\varepsilon, \label{eq:box-eps-mm-iso-1}\\
&| d_{\dot{X}}(\varphi(s),\varphi(t)) - d_Y(\psi(s),\psi(t))| \le \varepsilon
\quad\text{for any $s,t \in I_0$} \label{eq:box-eps-mm-iso-2}.
\end{align}
For any point $x \in \dot X$ we take a point $s \in \varphi^{-1}(x)$.
We assume that $s \in \varphi^{-1}(x) \cap I_0$ if $x \in \varphi(I_0)$.
Letting $\varphi'(x) := s$ defines a map $\varphi' : \dot X \to I$.
It is obvious that $\varphi'$ is Borel measurable,
$\varphi'(\varphi(I_0)) \subset I_0$, and $\varphi\circ\varphi' = \id_{\dot X}$.
Set $\dot f := \psi\circ\varphi' : \dot{X} \to Y$
and $\dot X_0 := \varphi(I_0)$.
We have $\mu_{\dot X}(\dot X_0) = \varphi_*\mathcal{L}^1(\varphi(I_0))
= \mathcal{L}^1(\varphi^{-1}(\varphi(I_0))) \ge \mathcal{L}^1(I_0)
\ge 1-\varepsilon$.
Moreover, by \eqref{eq:box-eps-mm-iso-2},
\begin{align} \label{eq:box-eps-mm-iso-3}
|d_{\dot X}(x,y) - d_Y(\dot f(x),\dot f(y))| \le \varepsilon
\quad\text{for any $x,y \in \dot X_0$}.
\end{align}
Let us prove that $d_P(\dot f_*\mu_{\dot X},\mu_Y) \le \varepsilon$.
For that, it suffices to prove that
$\dot f_*\mu_{\dot X}(B_\varepsilon(A)) \ge \mu_Y(A) -\varepsilon$
for any Borel subset $A \subset Y$.
\begin{clm}
For any Borel subset $A \subset Y$ we have
\[
\psi^{-1}(A) \cap I_0 \subset
\varphi^{-1}({\varphi'}^{-1}(\psi^{-1}(B_\varepsilon(A)))).
\]
\end{clm}
\begin{proof}
Take any $s \in \psi^{-1}(A) \cap I_0$ and set
$t := \varphi'(\varphi(s))$.
Since $s \in I_0$, we have $t \in \varphi'(\varphi(I_0)) \subset I_0$
and so
\[
|d_{\dot X}(\varphi(s),\varphi(t))-d_Y(\psi(s),\psi(t))| \le \varepsilon.
\]
It follows from $\varphi(t) = \varphi(\varphi'(\varphi(s))) = \varphi(s)$
that $d_Y(\psi(s),\psi(t)) \le \varepsilon$.
By $\psi(s) \in A$ we see that
$\psi(\varphi'(\varphi(s))) = \psi(t) \in B_\varepsilon(A)$,
so that $s \in \varphi^{-1}({\varphi'}^{-1}(\psi^{-1}(B_\varepsilon(A))))$.
This proves the claim.
\end{proof}
By the claim and \eqref{eq:box-eps-mm-iso-1},
\begin{align*}
\dot f_*\mu_{\dot X}(B_\varepsilon(A))
&= \mu_{\dot X}({\varphi'}^{-1}(\psi^{-1}(B_\varepsilon(A))))
= \mathcal{L}^1(\varphi^{-1}({\varphi'}^{-1}(\psi^{-1}(B_\varepsilon(A)))))\\
&\ge \mathcal{L}^1(\psi^{-1}(A) \cap I_0)
= \mathcal{L}^1(\psi^{-1}(A)) - \mathcal{L}^1(\psi^{-1}(A) \setminus I_0)\\
&\ge \mathcal{L}^1(\psi^{-1}(A)) - \varepsilon
= \mu_Y(A) - \varepsilon
\end{align*}
and therefore $d_P(\dot f_*\mu_{\dot X},\mu_Y) \le \varepsilon$.
By setting $f := \dot f \circ \pi_{\mathcal{N}} : X \to Y$,
it is a Borel measurable map.
It follows from $\mu_{\dot X} = (\pi_{\mathcal{N}})_*\mu_X$ that
$f_*\mu_X = \dot f_* (\pi_{\mathcal{N}})_* \mu_X = \dot f_*\mu_{\dot X}$
and hence $d_P(f_*\mu_X,\mu_Y) \le \varepsilon$.
Let $X_0 := B_\varepsilon(\mathcal{N}) \cap \pi_{\mathcal{N}}^{-1}(\dot{X}_0)$.
Then, any point $x \in X_0$ satisfies that
$d_X(\pi_{\mathcal{N}}(x),x) \le \varepsilon$ and $\pi_{\mathcal{N}}(x) \in \dot{X}_0$,
which together with \eqref{eq:box-eps-mm-iso-3}
proves that $f$ is $3\varepsilon$-isometric on $X_0$.
We also have
\begin{align*}
\mu_X(X \setminus X_0)
&= \mu_X((X \setminus B_\varepsilon(\mathcal{N}))
\cup (X \setminus \pi_{\mathcal{N}}^{-1}(\dot{X}_0)))\\
&\le \mu_X(X \setminus B_\varepsilon(\mathcal{N}))
+ \mu_X(X \setminus \pi_{\mathcal{N}}^{-1}(\dot{X}_0))
\le 2\varepsilon.
\end{align*}
Thus, $f : X \to Y$ is a $3\varepsilon$-mm-isomorphism.
This completes the proof.
\end{proof}
\begin{prop} \label{prop:separable}
The set $\mathcal{X}$ of mm-isomorphism classes of mm-spaces
is separable with respect to the box metric $\square$.
\end{prop}
\begin{proof}
Let $n$ be a natural number.
We set
\begin{align*}
R_n &:= \{\;(r_{ij})_{i,j=1,\dots,n;\ i<j} \in \field{R}^{n(n-1)/2} \mid\\
&\qquad r_{ij} > 0,\ r_{ik} \le r_{ij} + r_{jk}
\ \text{for any $i < j < k$}\;\},\\
W_n &:= \{\;(w_k)_{k=1,\dots,n-1} \in \field{R}^{n-1} \mid
w_k > 0,\ \sum_{k=1}^{n-1} w_k < 1\;\}.
\end{align*}
For $r \in R_n$ and $w \in W_n$, let $X(r,w) = \{x_1,x_2,\dots,x_n\}$
be an $n$-point space equipped with the following mm-structure:
\[
d_{X(r,w)}(x_i,x_j) := r_{ij}
\quad\text{and}\quad
\mu_{X(r,w)} := \sum_{k=1}^n w_k \delta_{x_k},
\]
where $r_{ii} := 0$, $r_{ij} := r_{ji}$ for $i > j$,
and $w_n := 1- \sum_{k=1}^{n-1} w_k$.
Let $\varepsilon \ge 0$.
For any $r,r' \in R_n$ and $w,w' \in W_n$
such that
\[
| r_{ij} - r_{ij}' | \le \varepsilon
\quad\text{and}\quad
| w_k - w_k' | \le \varepsilon
\quad\text{for any $i < j$ and $k$},
\]
we see $d_P(\mu_{X(r,w)},\mu_{X(r',w')}) \le n\varepsilon$
and therefore, the identity map from $X(r,w)$ to $X(r',w')$
is an $n\varepsilon$-mm-isomorphism.
By Lemma \ref{lem:box-eps-mm-iso}, we have
$\square(X(r,w),X(r',w')) \le 3n\varepsilon$.
Denote by $\mathcal{X}_n$ the mm-isomorphism classes of $n$-point mm-spaces.
The map $R_n \times W_n \ni (r,w) \mapsto X(r,w) \in \mathcal{X}_n$
is Lipschitz continuous with respect to the $l_\infty$-norm on $R_n \times W_n$
and $\square$. Since $R_n \times W_n$ is separable,
so is $(\mathcal{X}_n,\square)$ for any natural number $n$.
Therefore, the set of mm-isomorphism classes of finite mm-spaces,
say $\mathcal{X}_{<\infty}$, is separable with respect to $\square$.
Since $\mathcal{X}_{<\infty}$ is dense in $\mathcal{X}$,
this completes the proof.
\end{proof}
\begin{lem} \label{lem:precpt}
For a family $\mathcal{Y} \subset \mathcal{X}$, the following are equivalent
to each other.
\begin{enumerate}
\item $\mathcal{Y}$ is $\square$-precompact.
\item For any $\varepsilon > 0$ there exists a positive number
$\Delta(\varepsilon)$ such that
for any $X \in \mathcal{Y}$ we have a finite mm-space $X' \in \mathcal{X}$
such that $\square(X,X') \le \varepsilon$, $\# X' \le \Delta(\varepsilon)$,
and $\diam X' \le \Delta(\varepsilon)$.
\item For any $\varepsilon > 0$ there exists a positive number
$\Delta(\varepsilon)$ such that
for any $X \in \mathcal{Y}$ we have a finite net $\mathcal{N} \subset X$
$\varepsilon$-supporting $X$
such that
$\#\mathcal{N} \le \Delta(\varepsilon)$ and
$\diam\mathcal{N} \le \Delta(\varepsilon)$.
\item For any $\varepsilon > 0$ there exists a positive number
$\Delta(\varepsilon)$ such that
for any $X \in \mathcal{Y}$ we have Borel subsets
$K_1,K_2,\dots,K_N \subset X$, $N \le \Delta(\varepsilon)$,
such that $\diam K_i \le \varepsilon$ for any $i$,
$\diam\bigcup_{i=1}^N K_i \le \Delta(\varepsilon)$, and
$\mu_X(X \setminus \bigcup_{i=1}^N K_i) \le \varepsilon$.
\end{enumerate}
\end{lem}
\begin{proof}
We prove `(1) $\implies$ (2)'.
By (1), for any $\varepsilon > 0$
there are finitely many mm-spaces $X_1,\dots,X_{N(\varepsilon)} \in \mathcal{Y}$
such that
$\mathcal{Y} \subset \bigcup_{i=1}^{N(\varepsilon)} B_{\varepsilon/2}(X_i)$.
By Proposition \ref{prop:fin-approx}, we have a finite mm-space
$X_i'$ for each $i$ such that
$\square(X_i,X_i') \le \varepsilon/2$.
We define $\Delta(\varepsilon)$ to be the maximum of
$\# X_i'$ and $\diam X_i'$ for all $i=1,2,\dots,N(\varepsilon)$.
We then obtain (2).
We prove `(2) $\implies$ (1)'.
For a number $D > 0$ we set
\[
\mathcal{X}_D := \{\;X \in \mathcal{X} \mid \# X \le D,\ \diam X \le D\;\}.
\]
It is obvious that $\mathcal{X}_D$ is $\square$-compact.
Take any number $\varepsilon > 0$.
By (2), there is a number $\Delta(\varepsilon/2)$ such that
$\mathcal{Y} \subset B_{\varepsilon/2}(\mathcal{X}_{\Delta(\varepsilon/2)})$.
The $\square$-compactness of $\mathcal{X}_{\Delta(\varepsilon/2)}$ implies that
there is an $(\varepsilon/2)$-net
$\{X_i\}_{i=1}^{N(\varepsilon)} \subset \mathcal{X}_{\Delta(\varepsilon/2)}$, i.e.,
$\mathcal{X}_{\Delta(\varepsilon/2)} \subset B_{\varepsilon/2}(\{X_i\}_{i=1}^{N(\varepsilon)})$.
We see that
\[
\mathcal{Y} \subset B_{\varepsilon/2}(\mathcal{X}_{\Delta(\varepsilon/2)})
\subset \bigcup_{i=1}^{N(\varepsilon)} B_\varepsilon(X_i),
\]
which implies (1).
We prove `(2) $\implies$ (4)'.
For any $\varepsilon > 0$ we have $\Delta(\varepsilon/4)$ as in (2).
Namely, for any $X \in \mathcal{Y}$ there is a finite mm-space
$X' \in \mathcal{X}$ such that $\square(X,X') \le \varepsilon/4$,
$\# X' \le \Delta(\varepsilon/4)$, and $\diam X' \le \Delta(\varepsilon/4)$.
We find an $\varepsilon$-mm-isomorphism $f : X \to X'$.
Let $\{x_1,\dots,x_N\} := X'$ and $K_i := f^{-1}(x_i) \cap X_0$,
where $X_0$ is a non-exceptional domain of $f$.
We have $\diam K_i \le \varepsilon$.
Since $\diam X' \le \Delta(\varepsilon/4)$,
it holds that
$\diam \bigcup_{i=1}^N K_i \le \Delta(\varepsilon/4) + \varepsilon$.
It follows from $\bigcup_{i=1}^N K_i = X_0$ that
$\mu_X(X \setminus \bigcup_{i=1}^N K_i) \le \varepsilon$.
(4) is obtained.
We prove `(4) $\implies$ (3)'.
For any $\varepsilon > 0$ there is a number $\Delta(\varepsilon)$
such that for any $X \in \mathcal{X}$ we have
subsets $K_1,\dots,K_N \subset X$ as in (4).
We take a point $x_i$ in each $K_i$ and set $\mathcal{N} := \{x_i\}_{i=1}^N$.
This satisfies (3).
We prove `(3) $\implies$ (2)'.
For any $\varepsilon > 0$ there is $\Delta(\varepsilon/2)$
such that for any $X \in \mathcal{X}$ we have a net $\mathcal{N} \subset X$ as in (3).
Let $\pi_{\mathcal{N}} : X \to \mathcal{N}$ be a Borel measurable
nearest point projection and let
$\dot X := (\mathcal{N},d_X,(\pi_{\mathcal{N}})_*\mu_X)$.
Proposition \ref{prop:box-di} and Lemma \ref{lem:pi-di} imply
$\square(X,\dot X) \le \varepsilon$.
Since $\#\dot X \le \Delta(\varepsilon/2)$
and $\diam\dot X \le \Delta(\varepsilon/2)$,
we obtain (2).
This completes the proof of the lemma.
\end{proof}
\begin{defn}[Uniform family of mm-spaces]
\index{uniform family}
A family $\mathcal{Y} \subset \mathcal{X}$ of mm-isomorphism classes of mm-spaces
is said to be \emph{uniform} if
\[
\sup_{X\in\mathcal{Y}} \diam X < + \infty \quad\text{and}\quad
\inf_{X\in\mathcal{Y}}\inf_{x \in X} \mu_X(B_\varepsilon(x)) > 0
\]
for any real number $\varepsilon > 0$.
\end{defn}
\begin{cor} \label{cor:precpt}
Any uniform family in $\mathcal{X}$ is $\square$-precompact.
\end{cor}
\begin{proof}
Let $\mathcal{Y} \subset \mathcal{X}$ be a uniform family.
We will check (3) of Lemma \ref{lem:precpt} for $\mathcal{Y}$.
Set
\[
D := \sup_{X\in\mathcal{Y}} \diam X < + \infty \quad\text{and}\quad
a_\varepsilon := \inf_{X\in\mathcal{Y}}\inf_{x \in X} \mu_X(B_\varepsilon(x)) > 0.
\]
For any $\varepsilon > 0$ and $X \in \mathcal{Y}$,
we find an $\varepsilon$-maximal net $\mathcal{N}$ of $X$.
The $\varepsilon$-maximality of $\mathcal{N}$ implies that
$B_\varepsilon(\mathcal{N}) = \mathcal{N}$ and in particular $\mathcal{N}$ $\varepsilon$-supports
$X$.
It is clear that $\diam\mathcal{N} \le D$.
Since $\mu_X(B_{\varepsilon/2}(x)) \ge a_{\varepsilon/2} > 0$ for any $x \in X$
and since $B_{\varepsilon/2}(x)$, $x \in \mathcal{N}$, are disjoint to each other,
the net $\mathcal{N}$ has at most $[1/a_{\varepsilon/2}]$ points,
where $[a]$ for a real number $a$ indicates
the largest integer not greater than $a$.
Lemma \ref{lem:precpt} proves the corollary.
\end{proof}
\begin{defn}[Doubling condition]
\index{doubling condition}
We say that an mm-space $X$ satisfies the \emph{doubling condition}
if there exists a constant $C > 0$ such that
\[
\mu_X(B_{2r}(x)) \le C \mu_X(B_r(x))
\]
for any $x \in X$ and $r > 0$.
The constant $C$ is called a \emph{doubling constant}.
\index{doubling constant}
\end{defn}
\begin{rem} \label{rem:precpt}
If an mm-space $X$ has diameter $\le D$ and doubling constant $C$,
then, for any $x \in X$ and $\varepsilon > 0$,
\[
\mu_X(B_\varepsilon(x)) \ge \frac{\mu_X(B_{2\varepsilon}(x))}{C}
\ge \dots \ge \frac{\mu_X(B_{2^n\varepsilon}(x))}{C^n} = C^{-n}
\]
for a natural number $n$ with $2^n\varepsilon \ge D$ and therefore
\[
\mu_X(B_\varepsilon(x)) \ge C^{-\log_2([D/\varepsilon]+1)}
= ([D/\varepsilon]+1)^{-\log_2C}.
\]
Thus, if a family of mm-isomorphism classes of mm-spaces has
an upper diameter bound and an upper bound of doubling constant,
then the family is uniform and is $\square$-precompact.
It follows from the Bishop-Gromov volume comparison theorem that
the set of isometry classes of closed Riemannian manifold
with an upper dimension bound, an upper diameter bound, and
a lower bound of Ricci curvature has an upper bound of doubling constant
and then it is uniform.
\end{rem}
\begin{defn}[Measured Gromov-Hausdorff convergence] \label{defn:mGH}
\index{measured Gromov-Hausdorff convergence}
Let $X$ and $X_n$, $n=1,2,\dots$, be compact mm-spaces.
We say that \emph{$X_n$ measured Gromov-Hausdorff converges to $X$}
as $n\to\infty$ if there exist Borel measurable
$\varepsilon_n$-isometries $f_n : X_n \to X$, $n=1,2,\dots$,
with $\varepsilon_n \to 0$ such that
$(f_n)_*\mu_{X_n}$ converges weakly to $\mu_X$ as $n\to\infty$.
\end{defn}
If $X_n$ measured Gromov-Hausdorff converges to $X$, then
$X_n$ Gromov-Hausdorff converges to $X$.
\begin{rem} \label{rem:box-mGH}
Let $X$ and $X_n$, $n=1,2,\dots$, be compact mm-spaces.
If $X_n$ measured Gromov-Hausdorff converges to $X$ as $n\to\infty$,
then the map $f_n$ as above is an $\varepsilon_n'$-mm-isomorphism
with $\varepsilon_n' \to 0$, and, by Lemma \ref{lem:box-eps-mm-iso},
$X_n$ $\square$-converges to $X$ as $n\to\infty$.
If $X_n$ $\square$-converges to $X$ as $n\to\infty$,
then $X_n$ does not necessarily converge to $X$
in the sense of measured Gromov-Hausdorff convergence.
In fact, we consider the sequence of mm-spaces $X_n$, $n=1,2,\dots$,
defined by
\begin{align*}
& X_n := \{x_i\}_{i=0}^n, \qquad d_{X_n}(x_i,x_j) := 1-\delta_{ij},\\
& \mu_{X_n} := \left(1-\frac{1}{n}\right)\delta_{x_0}
+ \sum_{i=1}^n \frac{1}{n}\delta_{x_i}.
\end{align*}
As $n\to\infty$, $X_n$ $\square$-converges to the one-point mm-space
$(\{x_0\},\delta_{x_0})$.
However, $\{X_n\}$ is not $d_{GH}$-precompact
and has no $d_{GH}$-convergent subsequence.
It is not difficult to prove that,
if $X_n$ $\square$-converges to $X$ as $n\to\infty$ and if
\[
\inf_n \inf_{x \in X_n} \mu_{X_n}(B_\varepsilon(x)) > 0
\]
for any $\varepsilon > 0$,
then $X_n$ measured Gromov-Hausdorff converges to $X$.
The proof is left to the reader.
\end{rem}
\section{Lipschitz order and box convergence}
The purpose of this section is to prove the following theorem.
\begin{thm} \label{thm:dom}
Let $X$, $Y$, $X_n$, and $Y_n$ be mm-spaces, $n=1,2,\dots$.
If $X_n$ and $Y_n$ $\square$-converge
to $X$ and $Y$ respectively as $n\to\infty$
and if $X_n \prec Y_n$ for any $n$, then $X \prec Y$.
\end{thm}
For the proof, we need some lemmas.
\begin{lem} \label{lem:push-di-Lip}
Let $X$ and $Y$ be two metric spaces, $f : X \to Y$ a Borel measurable map,
$\mu$ and $\nu$ two Borel probability measures on $X$,
and $\varepsilon > 0$ a real number.
If there exists a Borel subset $X_0 \subset X$ such that
$\mu(X_0) \ge 1-\varepsilon$, $\nu(X_0) \ge 1-\varepsilon$,
and
\[
d_Y(f(x),f(y)) \le d_X(x,y) + \varepsilon
\]
for any $x,y \in X_0$, then we have
\[
d_P(f_*\mu,f_*\nu) \le d_P(\mu,\nu) + 2\varepsilon.
\]
\end{lem}
\begin{proof}
Take any Borel subset $A \subset Y$ and a real number $\delta > 0$.
Let us first prove
\begin{align} \label{eq:push-di-Lip}
f^{-1}(U_{\varepsilon+\delta}(A)) \supset
U_\delta(f^{-1}(A) \cap X_0) \cap X_0.
\end{align}
In fact, taking any point
$x \in U_\delta(f^{-1}(A) \cap X_0) \cap X_0$
we find a point $x' \in f^{-1}(A) \cap X_0$ such that
$d_X(x,x') < \delta$.
By the assumption,
\[
d_Y(f(x),f(x')) \le d_X(x,x') + \varepsilon < \varepsilon + \delta.
\]
Since $f(x') \in A$, we have $f(x) \in U_{\varepsilon+\delta}(A)$
and so $x \in f^{-1}(U_{\varepsilon+\delta}(A))$.
\eqref{eq:push-di-Lip} has been proved.
We assume that $d_P(\mu,\nu) < \delta$ for a number $\delta$.
It follows from \eqref{eq:push-di-Lip} and
$\mu(X \setminus X_0),\nu(X \setminus X_0) \le \varepsilon$ that
\begin{align*}
&f_*\mu(U_{\varepsilon+\delta}(A)) = \mu(f^{-1}(U_{\varepsilon+\delta}(A)))
\ge \mu(U_\delta(f^{-1}(A) \cap X_0) \cap X_0)\\
&\ge \mu(U_\delta(f^{-1}(A) \cap X_0)) - \varepsilon
\ge \nu(f^{-1}(A) \cap X_0)) - \varepsilon - \delta\\
&\ge \nu(f^{-1}(A)) - 2\varepsilon - \delta
= f_*\nu(A) - 2\varepsilon - \delta,
\end{align*}
which proves that $d_P(f_*\mu,f_*\nu) \le 2\varepsilon + \delta$.
By the arbitrariness of $\delta$ we have the lemma.
\end{proof}
The following is a direct consequence of Lemma \ref{lem:push-di-Lip}.
\begin{cor} \label{cor:push-di-Lip}
Let $X$ and $Y$ be a metric space and
$f : X \to Y$ a $1$-Lipschitz map.
For any Borel probability measures $\mu$ and $\nu$ on $X$,
we have
\[
d_P(f_*\mu,f_*\nu) \le d_P(\mu,\nu).
\]
\end{cor}
\begin{defn}[$1$-Lipschitz up to an additive error] \label{defn:1-Lip-err}
\index{1-Lipschitz up to an additive error@$1$-Lipschitz up to an additive error}
\index{1-Lipschitz up to epsilon@$1$-Lipschitz up to $\varepsilon$}
\index{additive error}
Let $X$ and $Y$ be two mm-spaces.
A map $f : X \to Y$ is said to be \emph{$1$-Lipschitz up to}
(\emph{an additive error}) $\varepsilon$ if there exists a Borel subset
$X_0 \subset X$ such that
\begin{align*}
&\mu_X(X_0) \ge 1-\varepsilon,\tag{1}\\
&d_Y(f(x),f(y)) \le d_X(x,y) + \varepsilon
\qquad\text{for any $x,y \in X_0$.}\tag{2}
\end{align*}
We call such a set $X_0$ a \emph{non-exceptional domain of $f$} and
the complement $X \setminus X_0$ an \emph{exceptional domain of $f$}.
\index{non-exceptional domain} \index{exceptional domain}
Even in the case where the metrics of $X$ and $Y$ are pseudo-metrics,
a map that is $1$-Lipschitz up to an additive error
is defined in the same way as above.
\end{defn}
\begin{lem} \label{lem:dom1}
Let $X$ and $Y$ be two mm-spaces.
Assume that there exist Borel measurable maps $p_n : X \to Y$,
$n=1,2,\dots$, that are $1$-Lipschitz up to additive errors
$\varepsilon_n$ with $\varepsilon_n \to 0$ such that
\[
d_P((p_n)_*\mu_X,\mu_Y) < \varepsilon_n
\]
for any $n$.
Then, $X$ dominates $Y$.
\end{lem}
\begin{proof}
We find an increasing sequence of compact subsets $C_N \subset Y$,
$N=1,2,\dots$,
with the property that $\mu_Y(C_N) \ge 1-1/N$ for any $N$.
We have
\[
\mu_X(p_n^{-1}(B_{\varepsilon_n}(C_N))) = (p_n)_*\mu_X(B_{\varepsilon_n}(C_N))
\ge \mu_Y(C_N) - \varepsilon_n \ge 1-\frac{1}{N}-\varepsilon_n.
\]
For any natural number $n$, there is a Borel subset
$\tilde{X}_n \subset X$ such that
$\mu_X(\tilde{X}_n) \ge 1-\varepsilon_n$ and
\[
d_Y(p_n(x),p_n(y)) \le d_X(x,y) + \varepsilon_n
\]
for any points $x,y \in \tilde{X}_n$.
We see that
$\mu_X(p_n^{-1}(B_{\varepsilon_n}(C_N)) \cap \tilde{X}_n) \ge 1-1/N-2\varepsilon_n$.
Setting
\[
E_N := \bigcap_{m=1}^\infty \bigcup_{n=m}^\infty
p_n^{-1}(B_{\varepsilon_n}(C_N)) \cap \tilde{X}_n,
\]
we have
\[
\mu_X(E_N)
\ge \liminf_{n\to\infty} \mu_X(p_n^{-1}(B_{\varepsilon_n}(C_N)) \cap \tilde{X}_n)
\ge 1-\frac{1}{N}.
\]
Since $\bigcup_{N=1}^\infty E_N$ is fully measured, it is dense in $X$.
We take a dense countable subset
$\{x_k\}_{k=1}^\infty \subset \bigcup_{N=1}^\infty E_N$.
For any natural number $k$, we find a number $N(k)$ in such a way that
$x_1,x_2,\dots,x_k \in E_{N(k)}$.
Then, for any natural number $m$ there is a number $n \ge m$
such that $p_n(x_k) \in B_{\varepsilon_n}(C_{N(k)})$ and $x_k \in \tilde{X}_n$.
Therefore, there is a subsequence $\{p_{n^1_i}(x_1)\}_{i=1}^\infty$
of $\{p_n(x_1)\}$
that converges to a point in $C_{N(1)}$ and
$x_1 \in \tilde{X}_{n^1_i}$ for any $i$.
In the same way, there is a subsequence $\{p_{n^2_i}(x_2)\}_{i=1}^\infty$ of
$\{p_{n^1_i}(x_2)\}_{i=1}^\infty$ that converges to a point in $C_{N(2)}$
and $x_2 \in \tilde{X}_{n^2_i}$ for any $i$.
Repeating this procedure we have an infinite sequence of subsequences
$\{p_{n^1_i}\}, \{p_{n^2_i}\}, \{p_{n^3_i}\}, \dots$.
By a diagonal argument, we choose a common subsequence
$\{p_{n_i}\}$ of $\{p_n\}$ such that
$\{p_{n_i}(x_k)\}_{i=1}^\infty$ converges for each $k$ and
$x_k \in \tilde{X}_{n_i}$ for any $k$ and $i$.
Letting $f(x_k) := \lim_{i\to\infty} p_{n_i}(x_k)$ defines
a $1$-Lipschitz map $f : \{x_k\}_{k=1}^\infty \to Y$.
Since $\{x_k\}_{k=1}^\infty$ is dense in $X$, this map extends to
a $1$-Lipschitz map $f : X \to Y$.
For any $\varepsilon > 0$,
there is a number $k_0$ such that
$\mu_X( \bigcup_{k=1}^{k_0} B_\varepsilon(x_k)) \ge 1-\varepsilon$
and $1/N(k_0) \le \varepsilon$.
We see that
$\mu_X( \bigcup_{k=1}^{k_0} B_\varepsilon(x_k) \cap E_{N(k_0)})
\ge 1-1/N(k_0)-\varepsilon \ge 1-2\varepsilon$.
For the number $\varepsilon$, there is a natural number $i_0$
such that
for any number $i \ge i_0$ we have
$\varepsilon_{n_i} \le \varepsilon$ and
$d_Y(p_{n_i}(x_k),f(x_k)) \le \varepsilon$ for all $k=1,2,\dots,k_0$.
We set
\[
D_{\varepsilon,i} := \bigcup_{k=1}^{k_0}
B_\varepsilon(x_k) \cap E_{N(k_0)} \cap \tilde{X}_{n_i}.
\]
Let $i$ be any number with $i \ge i_0$.
It holds that $\mu_X(D_{\varepsilon,i}) \ge 1-3\varepsilon$.
For any point $x \in D_{\varepsilon,i}$ there is a number
$k(x) \in \{1,2,\dots,k_0\}$ such that $d_X(x,x_{k(x)}) \le \varepsilon$.
Since $x,x_{k(x)} \in \tilde{X}_{n_i}$,
we have $d_Y(p_{n_i}(x),p_{n_i}(x_{k(x)})) \le d_X(x,x_{k(x)})+\varepsilon_{n_i}
\le 2\varepsilon$.
We also have $d_Y(p_{n_i}(x_{k(x)}),f(x_{k(x)})) \le \varepsilon$.
It follows from the $1$-Lipschitz continuity of $f$ that
$d_Y(f(x_{k(x)}),f(x)) \le d_X(x,x_{k(x)}) \le \varepsilon$.
Combining these inequalities yields $d_Y(p_{n_i}(x),f(x)) \le 4\varepsilon$
for any $x \in D_{\varepsilon,i}$, so that, by Lemma \ref{lem:di-me},
$d_P((p_{n_i})_*\mu_X,f_*\mu_X) \le \dKF(p_{n_i},f) \le 4\varepsilon$.
As $i\to\infty$, $(p_{n_i})_*\mu_X$ converges weakly to $f_*\mu_X$.
Since the assumption implies the weak convergence
$(p_{n_i})_*\mu_X \to \mu_Y$, we obtain $f_*\mu_X = \mu_Y$.
This completes the proof.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:dom}]
By the assumption, there are a sequence $\varepsilon_n \to 0$,
$\varepsilon_n$-mm-isomorphisms $\varphi_n : X_n \to X$ and
$\psi_n : Y \to Y_n$.
In addition, there is a $1$-Lipschitz map $f_n : Y_n \to X_n$ with
$(f_n)_*\mu_{Y_n} = \mu_{X_n}$.
Since $d_P((\psi_n)_*\mu_Y,\mu_{Y_n}) \le \varepsilon_n$,
Corollary \ref{cor:push-di-Lip} implies
\begin{align}
\label{eq:dom-conv}
d_P((f_n\circ\psi_n)_*\mu_Y,\mu_{X_n})
&= d_P((f_n)_*(\psi_n)_*\mu_Y,(f_n)_*\mu_{Y_n})\\
\notag &\le d_P((\psi_n)_*\mu_Y,\mu_{Y_n}) \le \varepsilon_n.
\end{align}
For the $\varepsilon_n$-mm-isomorphism $\varphi_n : X_n \to X$,
there is a Borel subset $\tilde{X}_n \subset X_n$ such that
$\mu_{X_n}(\tilde{X}_n) \ge 1-\varepsilon_n$ and
$|d_X(\varphi_n(x),\varphi_n(x'))-d_{X_n}(x,x')| \le \varepsilon_n$
for any $x,x' \in \tilde{X}_n$.
By Lemma \ref{lem:eps-proj}, we have a Borel measurable
$\varepsilon_n$-projection $\pi_n : X_n \to \tilde{X}_n$
with $\pi_n|_{\tilde{X}_n} = \id_{\tilde{X}_n}$.
Let $\varphi_n' := \varphi_n\circ\pi_n : X_n \to X$.
It follows from $\mu_{X_n}(\tilde{X}_n) \ge 1-\varepsilon_n$ that
$d_{KF}^{\mu_{X_n}}(\varphi_n,\varphi_n') \le \varepsilon_n$, which together
with Lemma \ref{lem:di-me} implies
$d_P((\varphi_n)_*\mu_{X_n},(\varphi_n')_*\mu_{X_n}) \le \varepsilon_n$.
For any $x,x' \in B_{\varepsilon_n}(\tilde{X}_n)$ we have
\begin{align} \label{eq:varphi-prime}
|d_X(\varphi_n'(x),\varphi_n'(x'))-d_{X_n}(x,x')| \le 3\varepsilon_n.
\end{align}
Moreover, \eqref{eq:dom-conv} implies
\[
(f_n\circ\psi_n)_*\mu_Y(B_{\varepsilon_n}(\tilde{X}_n))
\ge \mu_{X_n}(\tilde{X}_n)-\varepsilon_n \ge 1-2\varepsilon_n.
\]
Let
\[
F_n := \varphi_n'\circ f_n\circ\psi_n : Y \to X.
\]
This is a Borel measurable map.
It follows from Lemma \ref{lem:push-di-Lip} that
\begin{align*}
d_P((F_n)_*\mu_Y,(\varphi_n')_*\mu_{X_n})
&= d_P((\varphi_n')_*(f_n\circ\psi_n)_*\mu_Y,(\varphi_n')_*\mu_{X_n})\\
&\le d_P((f_n\circ\psi_n)_*\mu_Y,\mu_{X_n}) + 6\varepsilon_n
\le 7\varepsilon_n.
\end{align*}
Since $d_P((\varphi_n')_*\mu_{X_n},\mu_X)
\le d_P((\varphi_n')_*\mu_{X_n},(\varphi_n)_*\mu_{X_n})
+ d_P((\varphi_n)_*\mu_{X_n},\mu_X) \le 2\varepsilon_n$,
we have
\begin{align} \label{eq:Fn-push}
d_P((F_n)_*\mu_Y,\mu_X) \le 9\varepsilon_n.
\end{align}
For the $\varepsilon_n$-mm-isomorphism $\psi_n : Y \to Y_n$,
there is a Borel subset $\tilde{Y}_n \subset Y$ such that
$\mu_Y(\tilde{Y}_n) \ge 1-\varepsilon_n$ and
$|d_{Y_n}(\psi_n(y),\psi_n(y'))-d_Y(y,y')| \le \varepsilon_n$
for any $y,y' \in \tilde{Y}_n$.
We set
\[
\tilde{Y}'_n := \tilde{Y}_n \cap
(f_n\circ\psi_n)^{-1}(B_{\varepsilon_n}(\tilde{X}_n)).
\]
Since $\mu_Y((f_n\circ\psi_n)^{-1}(B_{\varepsilon_n}(\tilde{X}_n))
= (f_n\circ\psi_n)_*\mu_Y(B_{\varepsilon_n}(\tilde{X}_n)) \ge 1-2\varepsilon_n$,
we have $\mu_Y(\tilde{Y}'_n) \ge 1-3\varepsilon_n$.
For any $y,y' \in \tilde{Y}'_n$, we have
\[
d_{X_n}(f_n\circ\psi_n(y),f_n\circ\psi_n(y'))
\le d_{Y_n}(\psi_n(y),\psi_n(y'))
\le d_Y(y,y')+\varepsilon_n,
\]
so that,
by \eqref{eq:varphi-prime},
\[
d_X(F_n(y),F_n(y')) \le d_Y(y,y') + 4\varepsilon_n.
\]
This proves that $F_n$ is $1$-Lipschitz up to $4\varepsilon_n$.
Applying Lemma \ref{lem:dom1} completes the proof.
\end{proof}
\section{Finite-dimensional approximation}
The purpose in this section is to prove that
any mm-space can be approximated by $\field{R}^N$ with a Borel probability measure.
\begin{defn}[$\mathcal{L}{\it ip}_1(X)$, $\underline\mu_N$, and $\underline{X}_N$]
\index{Lip1X@$\mathcal{L}{\it ip}_1(X)$}
\index{muN@$\underline\mu_N$}
\index{XN@$\underline{X}_N$}
Let $X$ be an mm-space.
Denote by $\mathcal{L}{\it ip}_1(X)$ the set of $1$-Lipschitz functions on $X$.
For $\varphi_i \in \mathcal{L}{\it ip}_1(X)$, $i=1,2,\dots,N$, we define
\begin{align*}
\Phi_N &:= (\varphi_1,\dots,\varphi_N) : X \to \field{R}^N,\\
\underline{\mu}_N &:= (\Phi_N)_*\mu_X,\\
\underline{X}_N &:= (\field{R}^N,\|\cdot\|_\infty,\underline{\mu}_N).
\end{align*}
\end{defn}
\begin{prop}
We have
\[
\underline{X}_1 \prec \underline{X}_2 \prec \dots \prec \underline{X}_N
\prec X.
\]
\end{prop}
\begin{proof}
We prove $\underline{X}_N \prec X$.
In fact, for any $x,y \in X$, we have
\[
\|\Phi_N(x)-\Phi_N(y)\|_\infty
= \max_{i=1,2,\dots,N} |\varphi_i(x)-\varphi_i(y)|
\le d_X(x,y),
\]
i.e., $\Phi_N : X \to \field{R}^N$ is $1$-Lipschitz continuous.
We therefore obtain $\underline{X}_N \prec X$.
We prove $\underline{X}_n \prec \underline{X}_{n+1}$.
Since the projection $\pr : \underline{X}_{n+1} \to \underline{X}_n$
is $1$-Lipschitz continuous and since $\pr \circ \Phi_{n+1} = \Phi_n$,
we have
\[
\pr_*\underline{\mu}_{n+1} = \pr_*(\Phi_{n+1})_*\mu_X
= (\pr\circ\Phi_{n+1})_*\mu_X = (\Phi_n)_*\mu_X = \underline{\mu}_n.
\]
Therefore, $\underline{X}_n \prec \underline{X}_{n+1}$.
This completes the proof.
\end{proof}
\begin{defn}[$\mathcal{L}_1(X)$]
\index{l1@$\mathcal{L}_1(X)$}
We define an action of $\field{R}$ on $\mathcal{L}{\it ip}_1(X)$ by
\[
\field{R} \times \mathcal{L}{\it ip}_1(X) \ni (c,f) \mapsto f+c \in \mathcal{L}{\it ip}_1(X).
\]
Let
\[
\mathcal{L}_1(X) := \mathcal{L}{\it ip}_1(X)/\field{R}
\]
be the quotient space of $\mathcal{L}{\it ip}_1(X)$ by the $\field{R}$-action.
We indicate by $[f]$ the $\field{R}$-orbit of $f \in \mathcal{L}{\it ip}_1(X)$.
For two orbits $[f], [g] \in \mathcal{L}_1(X)$,
we define
\[
\dKF([f],[g]) := \inf_{f'\in [f],\; g'\in [g]}
\dKF(f',g').
\]
\index{dKF@$\dKF$}
\end{defn}
Since any $\field{R}$-orbit is closed in $\mathcal{L}{\it ip}_1(X)$,
we see that $\dKF$ is a metric on $\mathcal{L}_1(X)$.
It is easy to prove the continuity of the $\field{R}$-action on $\mathcal{L}{\it ip}_1(X)$
with respect to $\dKF$, which implies the following lemma.
\begin{lem} \label{lem:me-const}
For any two functions $f,g \in \mathcal{L}{\it ip}_1(X)$,
there exists a real number $c$ such that
\[
\dKF([f],[g]) = \dKF(f,g+c).
\]
\end{lem}
To prove the $\dKF$-compactness of $\mathcal{L}_1(X)$ we need the following
\begin{lem} \label{lem:cpt-1Lip}
Let $Y$ be a proper metric space
and $F_i : X \to Y$, $i = 1,2,\dots$, $1$-Lipschitz maps.
If $\{F_i(x_0)\}_{i=1}^\infty$ has a convergent subsequence
for a point $x_0 \in X$, then
some subsequence of $\{F_i\}$ converges in measure
to a $1$-Lipschitz map $F : X \to Y$.
\end{lem}
\begin{proof}
We take a monotone decreasing sequence $\varepsilon_j \to 0$
as $j \to \infty$.
There is an increasing sequence of compact subsets
$K_1 \subset K_2 \subset \dots$ of $X$ such that
$x_0 \in K_1$, $\mu_X(X \setminus K_j) < \varepsilon_j$ for any $j$,
and $\bigcup_{j=1}^\infty K_j = X$.
By the Arzel\`a-Ascoli theorem,
$\{F_i\}$ has a subsequence $\{F_{i_1(k)}\}$ that
converges to a $1$-Lipschitz map $G_1$ uniformly on $K_1$.
$\{F_{i_1(k)}\}$ has a subsequence $\{F_{i_2(k)}\}$ that
converges to a $1$-Lipschitz map $G_2$ uniformly on $K_2$.
Repeat this procedure to get subsequences
$\{F_{i_j(k)}\}$, $j=1,2,\dots$, such that, for each $j$,
$\{F_{i_j(k)}\}$ converges to a $1$-Lipschitz map $G_j$ uniformly on $K_j$.
By the diagonal argument, there is a common subsequence $\{F_{i(k)}\}$
that converges to $G_j$ uniformly on each $K_j$.
We see that $G_{j+1}|_{K_j} = G_j$ for any $j$.
We define a map $F : X \to Y$ by $F|_{K_j} := G_j$ for any $j$.
Then, $F : X \to Y$ is $1$-Lipschitz continuous.
Since $F_{i(k)}$ converges to $F$ uniformly on each $K_j$,
it converges in measure to $F$ on $X$.
This completes the proof.
\end{proof}
\begin{prop}
$\mathcal{L}_1(X)$ is $\dKF$-compact.
\end{prop}
\begin{proof}
Take a sequence $[\varphi_i] \in \mathcal{L}_1(X)$, $i = 1,2,\dots$.
We may assume that $\varphi_i(x_0) = 0$ for each $i$.
We apply Lemma \ref{lem:cpt-1Lip} to obtain
a subsequence $\{\varphi_{i(j)}\}$ of $\{\varphi_i\}$
that converges in measure to a function $\varphi \in \mathcal{L}{\it ip}_1(X)$.
$[\varphi_{i(j)}]$ $\dKF$-converges to $[\varphi]$ as $j\to\infty$.
This completes the proof.
\end{proof}
\begin{thm} \label{thm:XN}
Let $X$ be an mm-space.
For any $\varepsilon > 0$, there exists a real number
$\delta = \delta(X,\varepsilon) > 0$ such that
if $\{[\varphi_i]\}_{i=1}^N \subset \mathcal{L}_1(X)$ is a $\delta$-net
with respect to $\dKF$, then
\[
\square(\underline{X}_N,X) \le \varepsilon.
\]
\end{thm}
\begin{proof}
Take any $\varepsilon > 0$ and fix it.
For a number $\rho > 0$ we set
\[
Z_\rho := \{\; x \in X \mid \mu_X(B_\varepsilon(x)) \le \rho\;\}.
\]
$Z_\rho$ is monotone nondecreasing in $\rho$.
Since $\mu_X(B_\varepsilon(x)) > 0$ for any $x \in X$, we have
$\bigcap_{\rho > 0} Z_\rho = \emptyset$ and so
$\lim_{\rho\to 0+} \mu_X(Z_\rho) = 0$.
Therefore, there is a number
$\delta = \delta(X,\varepsilon) > 0$ such that
$\delta \le \varepsilon$ and $\mu_X(Z_\delta) < \varepsilon$.
We take any two points $x,y \in X \setminus Z_\delta$ and fix them.
The definition of $Z_\delta$ implies that
\begin{align}
\mu_X(B_\varepsilon(x)) > \delta \quad\text{and}\quad
\mu_X(B_\varepsilon(y)) > \delta. \label{eq:XN}
\end{align}
Let $\{[\varphi_i]\}_{i=1}^N \subset \mathcal{L}_1(X)$ be a $\delta$-net.
There is a number $i_0 \in \{1,2,\dots,N\}$ such that
$\dKF([d_x],[\varphi_{i_0}]) \le \delta$,
where $d_x(z) := d_X(x,z)$ for $x,z \in X$.
We find a real number $c$ with $\dKF(d_x,\varphi_{i_0}^{(c)}) \le \delta$
(see Lemma \ref{lem:me-const}),
where $\varphi_{i_0}^{(c)} := \varphi_{i_0} + c$.
Since $\mu_X(|d_x-\varphi_{i_0}^{(c)}| > \delta) \le \delta$
(see Remark \ref{rem:me}) and by \eqref{eq:XN},
the two balls $B_\varepsilon(x)$ and $B_\varepsilon(y)$ both intersect
the set $\{\,|d_x-\varphi_{i_0}^{(c)}| \le \delta\,\}$,
so that there are two points $x' \in B_\varepsilon(x)$ and
$y' \in B_\varepsilon(y)$ such that
\begin{align}
|d_X(x,x')-\varphi_{i_0}^{(c)}(x')| \le \delta
\quad\text{and}\quad
|d_X(x,y')-\varphi_{i_0}^{(c)}(y')| \le \delta. \label{eq:XN2}
\end{align}
Therefore,
\begin{align}
&|d_X(x,y)-\varphi_{i_0}^{(c)}(y)| \label{eq:XN3}\\
&\le |d_X(x,y)-d_X(x,y')| + |d_X(x,y')-\varphi_{i_0}^{(c)}(y')|
+|\varphi_{i_0}^{(c)}(y')-\varphi_{i_0}^{(c)}(y)| \notag\\
&\le d_X(y,y') + \delta + d_X(y,y')
\le 2\varepsilon + \delta \le 3\varepsilon. \notag
\end{align}
The $1$-Lipschitz continuity of $|\varphi_{i_0}^{(c)}|$ and \eqref{eq:XN2}
show
\begin{align}
|\varphi_{i_0}^{(c)}(x)| \le |\varphi_{i_0}^{(c)}(x')| + d_X(x,x')
\le 2\, d_X(x,x')+\delta \le 3\varepsilon. \label{eq:XN4}
\end{align}
Combining \eqref{eq:XN3} and \eqref{eq:XN4} yields
\begin{align*}
d_X(x,y) &\le \varphi_{i_0}^{(c)}(y) + 3\varepsilon
\le \varphi_{i_0}^{(c)}(y) - \varphi_{i_0}^{(c)}(x) + 6\varepsilon\\
&= \varphi_{i_0}(y) - \varphi_{i_0}(x) + 6\varepsilon
\le \|\Phi_N(x) - \Phi_N(y)\|_\infty + 6\varepsilon,
\end{align*}
which together with the $1$-Lipschitz continuity of $\Phi_N$
implies that
\[
|\; \|\Phi_N(x)-\Phi_N(y)\|_\infty - d_X(x,y) \;| \le 6\varepsilon.
\]
for any $x,y \in X \setminus Z_\delta$.
By recalling $\mu_X(Z_\delta) < \varepsilon$,
the map $\Phi_N : X \to \underline{X}_N$ turns out to be
a $6\varepsilon$-mm-isomorphism.
By Lemma \ref{lem:box-eps-mm-iso}(1) we obtain
$\square(\underline{X}_N,X) \le 18\varepsilon$.
This completes the proof.
\end{proof}
\begin{cor} \label{cor:XN}
If a sequence $\{[\varphi_i]\}_{i=1}^\infty \subset \mathcal{L}_1(X)$ is $\dKF$-dense,
then $\underline{X}_N$ $\square$-converges to $X$ as $N \to \infty$.
\end{cor}
Note that the compactness of $\mathcal{L}_1(X)$ implies
the existence of a dense countable subset of $\mathcal{L}_1(X)$,
so that any mm-space $X$ can be approximated by
some finite-dimensional space $\underline{X}_N$.
\begin{rem}
The infinite-product $[\,0,1\,]^\infty$ of the interval $[\,0,1\,]$
is not separable with respect to $\|\cdot\|_\infty$,
the proof of which is seen in Remark \ref{rem:nonsep}.
Therefore, $([\,0,1\,]^\infty,\|\cdot\|_\infty,\mathcal{L}^\infty)$
is not an mm-space, where $\mathcal{L}^\infty$ is the infinite-product
of the one-dimensional Lebesgue measure on $[\,0,1\,]$.
Besides, $\mathcal{L}_1([\,0,1\,]^\infty,\|\cdot\|_\infty,\mathcal{L}^\infty)$
is not $\dKF$-compact.
In fact, letting $\varphi_i : [\,0,1\,]^\infty \to \field{R}$ be
the $i^{th}$ projection, we have
$\dKF([\varphi_i],[\varphi_j]) = c\,(1-\delta_{ij})$ for some constant $c > 0$,
where $\delta_{ii} = 1$ and $\delta_{ij} = 0$ if $i \neq j$.
\end{rem}
\begin{prop}
Let $X$ be an mm-space
and $\{[\varphi_i]\}_{i=1}^\infty \subset \mathcal{L}_1(X)$ an $\dKF$-dense countable subset.
We define
\begin{align*}
\Phi_\infty &:= (\varphi_1,\varphi_2,\dots) : X \to \field{R}^\infty,\\
\underline{\mu}_\infty &:= (\Phi_\infty)_*\mu_X,\\
\underline{X}_\infty &:= (\Phi_\infty(X),\|\cdot\|_\infty,\underline{\mu}_\infty).
\end{align*}
Then, $\Phi_\infty : X \to \underline{X}_\infty$ is an isometry.
In particular, the image $\Phi_\infty(X)$ is separable
with respect to $\|\cdot\|_\infty$,
the space $\underline{X}_\infty$ is an mm-space,
and $\Phi_\infty : X \to \underline{X}_\infty$
is an mm-isomorphism.
\end{prop}
\begin{proof}
It suffices to prove that $\Phi_\infty : X \to \underline{X}_\infty$
is an isometry.
The proof is similar to that of Theorem \ref{thm:XN}.
Since each $\varphi_i$ is $1$-Lipschitz,
$\Phi_\infty : X \to \underline{X}_\infty$ is $1$-Lipschitz continuous.
Take any two points $x,y \in X$.
Since $\{[\varphi_i]\}_{i=1}^\infty \subset \mathcal{L}_1(X)$ is $\dKF$-dense,
there is a subsequence $\{[\varphi_{i_j}]\}_{j=1}^\infty$ of
$\{[\varphi_i]\}_{i=1}^\infty$ such that
$\lim_{j\to\infty} \dKF([\varphi_{i_j}],[d_x]) = 0$
and hence $\varphi_{i_j}^{(c_j)}$ converges in measure to $d_x$ as $j\to\infty$
for some $c_j$, where $\varphi_{i_j}^{(c_j)} := \varphi_{i_j} + c_j$.
This proves that, for any $\varepsilon > 0$,
the two balls $B_\varepsilon(x)$ and $B_\varepsilon(y)$ both intersect
$\{\,|\varphi_{i_j}^{(c_j)}-d_x| \le \varepsilon\,\}$ for all $j$ large enough.
Therefore,
we find two sequences $\{x_j\}_{j=1}^\infty$ and $\{y_j\}_{j=1}^\infty$
of points in $X$ respectively converging to $x$ and $y$ such that
\[
\lim_{j\to\infty} \varphi_{i_j}^{(c_j)}(x_j) = d_x(x) = 0 \quad\text{and}\quad
\lim_{j\to\infty} \varphi_{i_j}^{(c_j)}(y_j) = d_x(y) = d_X(x,y).
\]
Thus,
\begin{align*}
&\|\Phi_\infty(x) - \Phi_\infty(y)\|_\infty
\ge |\varphi_{i_j}(x)-\varphi_{i_j}(y)|
= |\varphi_{i_j}^{(c_j)}(x)-\varphi_{i_j}^{(c_j)}(y)|\\
&\ge |\varphi_{i_j}^{(c_j)}(x_j)-\varphi_{i_j}^{(c_j)}(y_j)|
- d_X(x,x_j) - d_X(y,y_j)
\overset{j\to\infty}{\longrightarrow} d_X(x,y),
\end{align*}
which together with the $1$-Lipschitz continuity of $\Phi_\infty$
proves that $\Phi_\infty$ is an isometry.
This completes the proof.
\end{proof}
\section{Infinite product, I}
In this section, we prove that a finite product space
$\square$-converges to the infinite product.
Let $p$ be an extend real number with $1 \le p \le +\infty$,
and $F_n$, $n=1,2,\dots$, be mm-spaces.
We set
\begin{align*}
X_n := F_1 \times F_2 \times \dots \times F_n
\quad\text{and}\quad
X_\infty := \prod_{n=1}^\infty F_n.
\end{align*}
Define, for two points $x = (x_i)_{i=1}^n, y = (y_i)_{i=1}^n \in X_n$,
$1 \le n \le \infty$,
\[
d_{l_p}(x,y) :=
\begin{cases}
\left( \sum\limits_{i=1}^n d_{F_i}(x_i,y_i)^p \right)^{\frac{1}{p}}
&\text{if $p < +\infty$,}\\
\sup\limits_{i=1}^n d_{F_i}(x_i,y_i)
&\text{if $p = +\infty$,}
\end{cases}
\]
\index{dlp@$d_{l_p}$}
Note that $d_{l_p}(x,y) \le +\infty$.
\begin{asmp} \label{asmp:finite}
If $p < +\infty$, then
\[
\sum_{n=1}^\infty (\diam F_n)^p < +\infty.
\]
If $p = +\infty$, then
\[
\lim_{n\to\infty} \diam F_n = 0.
\]
\end{asmp}
\begin{lem} \label{lem:prod-top}
\begin{enumerate}
\item The topology on $X_\infty$ induced from $d_{l_p}$
is not weaker than the product topology.
\item Under Assumption \ref{asmp:finite},
the topology on $X_\infty$ induced from $d_{l_p}$
coincides with the product topology.
\end{enumerate}
\end{lem}
\begin{proof}
We prove (1).
Recall that the product topology is generated by
\[
\mathcal{B} := \left\{\prod_{n=1}^\infty O_n \mid
\text{$O_n \subset F_n$ is open,
$\exists k$ : $O_n = F_n$ for any $n \ge k$}\;\right\}
\]
as an open basis.
It suffices to prove that any set $\prod_{n=1}^\infty O_n \in \mathcal{B}$ is an open set
with respect to $d_{l_p}$.
We have $k$ such that $O_n = F_n$ for any $n \ge k$.
Take any point $x = (x_n)_{n=1}^\infty \in \prod_{n=1}^\infty O_n$.
Since each $O_n$ is open, for any number $n < k$ there is a number
$\delta_n > 0$ such that $U_{\delta_n}(x_n) \subset O_n$.
Let $\delta := \min\{\delta_1,\dots,\delta_k\}$.
For any point $y = (y_n)_{n=1}^\infty \in U_\delta(x)$,
since $d_{F_n}(x_n,y_n) \le d_{l_p}(x,y) < \delta$, we have $y_n \in O_n$
for any $n$.
$U_\delta(x)$ is contained in $\prod_{n=1}^\infty O_n$.
By the arbitrariness of $x \in \prod_{n=1}^\infty O_n$,
the set $\prod_{n=1}^\infty O_n$ is open with respect to $d_{l_p}$.
(1) is obtained.
We prove (2) in the case of $p < +\infty$.
((2) in the case of $p = +\infty$ is proved
in the same way.)
Recall that the family of open $d_{l_p}$-metric balls is an open basis
for the topology induced from $d_{l_p}$.
It suffices to prove that
any open metric ball $U_\varepsilon(x)$, $\varepsilon > 0$, $x \in X_\infty$,
is open with respect to the product topology.
Take any point $y \in U_\varepsilon(x)$.
Since
\[
\left( \sum_{n=1}^\infty d_{F_n}(x_n,y_n)^p \right)^{\frac{1}{p}} = d_{l_p}(x,y)
< \varepsilon,
\]
there is a natural number $k$ such that
\[
\left( \sum_{n=k+1}^\infty (\diam F_n)^p \right)^{\frac{1}{p}}
< \varepsilon - d_{l_p}(x,y).
\]
We find $k$ real numbers $\delta_1,\dots,\delta_k > 0$
in such a way that
\[
\left( \sum_{n=1}^k \delta_n^p \right)^{\frac{1}{p}}
< \varepsilon - d_{l_p}(x,y)
- \left( \sum_{n=k+1}^\infty (\diam F_n)^p \right)^{\frac{1}{p}}.
\]
The set
\[
O := U_{\delta_1}(y_1) \times \dots \times U_{\delta_k}(y_k) \times
F_{k+1} \times F_{k+2} \times \dots
\]
is open with respect to the product topology
and contains $y$.
It suffices to prove that $O \subset U_\varepsilon(x)$.
If $z \in O$, then
\begin{align*}
d_{l_p}(y,z)
&\le \left( \sum_{n=1}^k \delta_n^p
+ \sum_{n=k+1}^\infty (\diam F_n)^p \right)^{\frac{1}{p}}\\
&\le \left( \sum_{n=1}^k \delta_n^p \right)^{\frac{1}{p}}
+ \left( \sum_{n=k+1}^\infty (\diam F_n)^p \right)^{\frac{1}{p}}\\
&< \varepsilon - d_{l_p}(x,y)
\end{align*}
and so $d_{l_p}(x,z) \le d_{l_p}(x,y) + d_{l_p}(y,z) < \varepsilon$.
We obtain $O \subset U_\varepsilon(x)$.
This completes the proof of the lemma.
\end{proof}
\begin{rem}
Supposing that Assumption \ref{asmp:finite} does not hold,
the topology on $X_\infty$ induced from $d_{l_p}$ is strictly stronger than
the product topology.
\end{rem}
\begin{proof}
If $p < +\infty$ and if $\sum_{n=1}^\infty (\diam F_n)^p = +\infty$,
then any set in the open basis $\mathcal{B}$ as in the proof of
Lemma \ref{lem:prod-top}
has infinite $d_{l_p}$-diameter and is not contained in
any $d_{l_p}$-metric ball,
so that any $d_{l_p}$-metric open ball is not open with respect to
the product topology.
If $p = +\infty$ and if $\delta := \liminf_{n\to\infty} \diam F_n > 0$,
then any set in $\mathcal{B}$ has $d_{l_\infty}$-diameter $\ge \delta$
and is not contained in any $d_{l_\infty}$-metric ball of radius $< \delta/2$.
Any $d_{l_\infty}$-metric open ball of radius $< \delta/2$
is not open with respect to the product topology in this case.
This completes the proof.
\end{proof}
\begin{lem} \label{lem:comp-sep}
Under Assumption \ref{asmp:finite} we have the following.
\begin{enumerate}
\item If each $F_n$ is complete, then so is $(X_\infty,d_{l_p})$.
\item If each $F_n$ is separable, then so is $(X_\infty,d_{l_p})$.
\end{enumerate}
\end{lem}
\begin{proof}
We prove (1).
Let $\{x_i\}_{i=1}^\infty$ be a Cauchy sequence in $(X_\infty,d_{l_p})$.
We set $(x_{in})_{n=1}^\infty := x_i$.
Since $d_{F_n}(x_{in},x_{jn}) \le d_{l_p}(x_i,x_j)$,
the sequence $\{x_{in}\}_i$ is a Cauchy sequence in $F_n$ for each $n$,
so that it is a convergent sequence.
Let
\[
x_{\infty,n} := \lim_{i\to\infty} x_{in}
\quad\text{and}\quad
x_\infty := (x_{\infty,n})_{n=1}^\infty.
\]
If $p < +\infty$, then
\begin{align*}
d_{l_p}(x_i,x_\infty)^p
&= \sum_{n=1}^\infty \lim_{j\to\infty} d_{F_i}(x_{in},x_{jn})^p\\
&\le \liminf_{j\to\infty} \sum_{n=1}^\infty d_{F_n}(x_{in},x_{jn})^p\\
&= \liminf_{j\to\infty} d_{l_p}(x_i,x_j)^p\\
&\to 0 \ \text{as}\ i\to\infty,
\end{align*}
and therefore $x_i$ converges to $x_\infty$ as $i\to\infty$.
In the case of $p = +\infty$, we see that $x_i$ converges to $x_\infty$
in the same way.
(1) has been proved.
We prove (2).
By the second countability of each $F_n$,
there is a countable open basis $\{O_{ni}\}_{i=1}^\infty$ of $F_n$,
where we assume that $F_n$ belongs to $\{O_{ni}\}_{i=1}^\infty$.
By Lemma \ref{lem:prod-top}, the topology of $(X_\infty,d_{l_p})$
coincides with the product topology, so that
\begin{align*}
\mathcal{B} := \{\;& O_{1i_1} \times O_{2i_2} \times \dots \times O_{ki_k}
\times F_{k+1} \times F_{k+2} \times \dots\\
&|\ k=1,2,\dots, \ i_1,i_2,\dots,i_k=1,2,\dots\}
\end{align*}
is an open basis over $X_\infty$.
For each $k$, the family
\begin{align*}
\mathcal{B}_k := \{\;& O_{1i_1} \times O_{2i_2} \times \dots \times O_{ki_k}
\times F_{k+1} \times F_{k+2} \times \dots\\
&| \ i_1,i_2,\dots,i_k=1,2,\dots,k\;\}
\end{align*}
is finite and satisfies $\bigcup_{k=1}^\infty\mathcal{B}_k = \mathcal{B}$.
Therefore, $\mathcal{B}$ is a countable basis.
This completes the proof.
\end{proof}
\begin{rem} \label{rem:nonsep}
Let $F_n$, $n=1,2,\dots$, be metric spaces.
If $\inf_n \diam F_n > 0$, then
$(X_\infty,d_{l_\infty})$ is not separable.
In particular, the infinite product space $F^\infty$ of an mm-space $F$
with the $d_{l_p}$ metric and the infinite product measure
is not an mm-space whenever $F$ has
two different points.
\end{rem}
\begin{proof}
Let $\delta := \inf_n \diam F_n > 0$.
We take any countable set $\{x_k\}_{k=1}^\infty \subset X_\infty$
and put
\[
x_k = (x_{k1},x_{k2},\dots),\quad x_{kn} \in F_n.
\]
For any two natural numbers $k$ and $n$,
there is a point $y_{kn} \in F_n$ such that
$d_{F_n}(x_{kn},y_{kn}) \ge \delta/2$.
Setting $y := (y_{11},y_{22},\dots) \in X_\infty$ we have
\[
d_{l_\infty}(x_k,y) = \sup_n d_{F_n}(x_{kn},y_{nn})
\ge d_{F_k}(x_{kk},y_{kk}) \ge \delta/2.
\]
Therefore, $\{x_k\}_{k=1}^\infty$ is not dense in $X_\infty$.
This completes the proof.
\end{proof}
\begin{prop}
Let $F_n$, $n=1,2,\dots$, be mm-spaces and let $1 \le p \le +\infty$.
We equip $X_n$ and $X_\infty$ with the product measures
$\bigotimes_{k=1}^n \mu_{F_k}$ and $\bigotimes_{k=1}^\infty \mu_{F_k}$
and with the $d_{l_p}$ metrics.
Under Assumption \ref{asmp:finite}, we have the following
{\rm(1)}, {\rm(2)}, and {\rm(3)}.
\begin{enumerate}
\item $X_\infty$ is an mm-space.
\item We have
\[
X_1 \prec X_2 \prec \dots \prec X_n \prec X_\infty,
\quad n=1,2,\dots.
\]
\item $X_n$ $\square$-converges to $X_\infty$ as $n\to\infty$.
\end{enumerate}
\end{prop}
\begin{proof}
We prove (1).
Lemma \ref{lem:comp-sep} says that $X_\infty$ is complete separable
and hence an mm-space.
(2) is obvious.
We prove (3).
Fixing a point $x_0 = (x_{0n})_{n=1}^\infty \in X_\infty$,
we define an isometric embedding map
\[
\iota_n : X_n \ni (x_1,\dots,x_n) \mapsto
(x_1,\dots,x_n,x_{0,n+1},x_{0,n+2},\dots) \in X_\infty.
\]
According to \cite{Bog}*{8.2.16},
we obtain that $(\iota_n)_*\mu_{X_n}$ converges weakly to $\mu_{X_\infty}$
as $n\to\infty$.
Moreover, $X_n$ is mm-isomorphic to $(X_\infty,(\iota_n)_*\mu_{X_n})$.
Proposition \ref{prop:box-di} proves (3).
This completes the proof.
\end{proof}
\chapter{Observable distance and measurement}
\label{chap:obs-dist-measurement}
\section{Basics for the observable distance}
We define the observable distance $\dconc(X,Y)$
between two mm-spaces $X$ and $Y$,
and study its basic properties.
\begin{defn}[$\dconc$ between pseudo-metrics]
\index{dconcrho1rho2@$\dconc(\rho_1,\rho_2)$}
For a pseudo-metric $\rho$ on $I$, let $\mathcal{L}{\it ip}_1(\rho)$ denotes
the set of $1$-Lipschitz functions on $I$ with respect to $\rho$.
Denote by $\mathbf{D}$ the set of pseudo-metrics $\rho$ on $I$
such that every element of $\mathcal{L}{\it ip}_1(\rho)$ is a Lebesgue measurable function.
For two pseudo-metrics $\rho_1,\rho_2 \in \mathbf{D}$, we set
\[
\dconc(\rho_1,\rho_2) := d_H(\mathcal{L}{\it ip}_1(\rho_1),\mathcal{L}{\it ip}_1(\rho_2)),
\]
where the Hausdorff distance $d_H$
is defined with respect to the Ky Fan metric $\dKF$.
\end{defn}
Since $\dKF \le 1$, we have $\dconc(\rho_1,\rho_2) \le 1$
for any $\rho_1,\rho_2 \in \mathbf{D}$.
\begin{lem}
Let $X$ be an mm-space.
For any parameter $\varphi$ of $X$ we have
\[
\mathcal{L}{\it ip}_1(\varphi^*d_X) = \varphi^*\mathcal{L}{\it ip}_1(X)
:= \{\;f\circ\varphi \mid f \in \mathcal{L}{\it ip}_1(X)\;\}
\]
and, in particular, $\varphi^*d_X$ belongs to $\mathbf{D}$.
\end{lem}
\begin{proof}
It is easy to see that $\mathcal{L}{\it ip}_1(\varphi^*d_X) \supset \varphi^*\mathcal{L}{\it ip}_1(X)$.
We prove $\mathcal{L}{\it ip}_1(\varphi^*d_X) \subset \varphi^*\mathcal{L}{\it ip}_1(X)$.
For any function $f \in \mathcal{L}{\it ip}_1(\varphi^*d_X)$,
we have
\begin{align} \label{eq:Lip1}
|f(s)-f(t)| \le d_X(\varphi(s),\varphi(t))
\end{align}
for any $s,t \in I$.
In particular, if $\varphi(s) = \varphi(t)$, then $f(s) = f(t)$.
For a given point $x \in \varphi(I)$, we take a point $s \in I$
with $\varphi(s) = x$.
Then, $f(s)$ depends only on $x$ and independent of $s$.
We define $\tilde{f}(x) := f(s)$.
It follows from \eqref{eq:Lip1} that $\tilde{f} : \varphi(I) \to \field{R}$
is $1$-Lipschitz continuous and extends to a $1$-Lipschitz function
$\tilde{f} : X \to \field{R}$.
The definition of $\tilde{f}$ implies that $f = \tilde{f}\circ\varphi$.
This completes the proof.
\end{proof}
\begin{defn}[Observable distance $\dconc(X,Y)$] \label{defn:obs-dist}
\index{observable distance} \index{dconcXY@$\dconc(X,Y)$}
We define the \emph{observable distance $\dconc(X,Y)$ between
two mm-spaces $X$ and $Y$} by
\[
\dconc(X,Y) := \inf_{\varphi,\psi} \dconc(\varphi^*d_X,\psi^*d_Y),
\]
where $\varphi : I \to X$ and $\psi : I \to Y$ run over all parameters
of $X$ and $Y$, respectively.
We say that a sequence of mm-spaces $X_n$, $n=1,2,\dots$,
\emph{concentrates to} \index{concentrate} an mm-space $X$
if
\[
\lim_{n\to\infty} \dconc(X_n,X) = 0.
\]
\end{defn}
Note that $\dconc(X,Y) \le 1$ for any two mm-spaces $X$ and $Y$.
We prove that $\dconc$ is a metric on $\mathcal{X}$ later in Theorem \ref{thm:dconc}.
\begin{lem} \label{lem:Lip-approx}
Let $S$ be a topological space with a Borel probability measure $\mu_S$,
and $\rho$ a pseudo-metric on $S$ such that any $1$-Lipschitz
function on $S$ with respect to $\rho$ is Borel measurable.
For any Borel measurable map $f : S \to (\field{R}^N,\|\cdot\|_\infty)$
that is $1$-Lipschitz up to an additive error $\varepsilon \ge 0$
with respect to $\rho$,
there exists a $1$-Lipschitz map $\tilde f : S \to (\field{R}^N,\|\cdot\|_\infty)$
such that
\begin{align*}
\dKF(\tilde{f},f) \le \varepsilon.
\end{align*}
\end{lem}
\begin{proof}
By the assumption, there is a (non-exceptional) Borel subset $S_0 \subset S$
such that $\mu_S(S_0) \ge 1-\varepsilon$ and
\[
\|f(x)-f(y)\|_\infty \le \rho(x,y) + \varepsilon
\]
for any $x,y \in S_0$.
We set $(f_1,f_2,\dots,f_N) := f$.
For any $i$ and $x,y \in S_0$,
\[
|f_i(x)-f_i(y)| \le \|f(x)-f(y)\|_\infty \le \rho(x,y) + \varepsilon.
\]
We define, for $x \in S$,
\[
\tilde{f}_i(x) := \inf_{y \in S_0} (f_i(y)+\rho(x,y)).
\]
Then, for any $x,y \in S$,
\begin{align*}
\tilde{f}_i(x)-\tilde{f}_i(y)
&= \inf_{x' \in S_0} (f_i(x')+\rho(x,x'))
- \inf_{y' \in S_0} (f_i(y')+\rho(y,y'))\\
&\le \sup_{x' \in S_0} (f_i(x')+\rho(x,x')-f_i(x')-\rho(y,x'))\\
&\le \rho(x,y).
\end{align*}
Since this also holds if we exchange $x$ and $y$,
the function $\tilde{f}_i$ is $1$-Lipschitz continuous and so is
$\tilde{f} := (\tilde{f}_1,\tilde{f}_2,\dots,\tilde{f}_N)
: S \to (\field{R}^N,\|\cdot\|_\infty)$.
For any $x \in S_0$, we have $\tilde{f}_i(x) \le f_i(x)$
and
\[
f_i(x) - \tilde{f}_i(x) = \sup_{y \in S_0} (f_i(x)-f_i(y)-\rho(x,y))
\le \varepsilon.
\]
Since this holds for any $i$, we have
\[
\|\tilde{f}(x)-f(x)\|_\infty \le \varepsilon
\]
for any $x \in S_0$.
We obtain $\dKF(\tilde{f},f) \le \varepsilon$.
This completes the proof.
\end{proof}
\begin{prop} \label{prop:dconc-box}
\begin{enumerate}
\item For any two pseudo-metrics $\rho_1, \rho_2 \in \mathbf{D}$,
we have
\[
\dconc(\rho_1,\rho_2) \le \square(\rho_1,\rho_2).
\]
\item For any two mm-spaces $X$ and $Y$ we have
\[
\dconc(X,Y) \le \square(X,Y).
\]
\end{enumerate}
\end{prop}
\begin{proof}
(2) follows from (1).
We prove (1).
Assume that $\square(\rho_1,\rho_2) < \varepsilon$ for two pseudo-metrics
$\rho_1,\rho_2 \in \mathbf{D}$ and for a number $\varepsilon$.
There is a Borel subset $I_0 \subset I$ such that
$\mathcal{L}^1(I_0) > 1-\varepsilon$ and
\[
|\;\rho_1(s,t) - \rho_2(s,t)\;| < \varepsilon
\]
for any $s,t \in I_0$.
Take any function $f \in \mathcal{L}{\it ip}_1(\rho_1)$.
We then see that $f$ is $1$-Lipschitz up to the additive error $\varepsilon$
with respect to $\rho_2$.
Apply Lemma \ref{lem:Lip-approx} to obtain a function
$\tilde{f} \in \mathcal{L}{\it ip}_1(\rho_2)$ with $\dKF(\tilde{f},f) \le \varepsilon$.
Therefore, $\mathcal{L}{\it ip}_1(\rho_1) \subset B_\varepsilon(\mathcal{L}{\it ip}_1(\rho_2))$.
Since this also holds by exchanging $\rho_1$ and $\rho_2$, we have
$d_H(\mathcal{L}{\it ip}_1(\rho_1),\mathcal{L}{\it ip}_1(\rho_2)) \le \varepsilon$.
This completes the proof.
\end{proof}
We denote by $*$ a one-point mm-space, i.e.,
$*$ consists of a single point with trivial metric and Dirac's delta measure.
\begin{lem} \label{lem:dconc-pt}
Let $X$ be an mm-space. Then we have
\[
\dconc(X,*) = \sup_{f \in \mathcal{L}{\it ip}_1(X)}\inf_{c \in \field{R}} \dKF(f,c).
\]
\end{lem}
\begin{proof}
Let $\varphi : I \to X$ and $\psi : I \to *$ be two parameters.
Since $\psi^*\mathcal{L}{\it ip}_1(*)$ is the set of constant functions on $I$,
we have
\begin{align*}
d_H(\varphi^*\mathcal{L}{\it ip}_1(X),\psi^*\mathcal{L}{\it ip}_1(*))
&= \sup_{f \in \mathcal{L}{\it ip}_1(X)}\inf_{c\in\field{R}} \dKF(f\circ\varphi,c)\\
&= \sup_{f \in \mathcal{L}{\it ip}_1(X)}\inf_{c\in\field{R}} \dKF(f,c).
\end{align*}
\end{proof}
\begin{prop} \label{prop:ObsDiam-dconc}
For any mm-space $X$ we have
\[
\dconc(X,*) \le \ObsDiam(X) \le 2\dconc(X,*).
\]
\end{prop}
\begin{proof}
We prove the first inequality.
Assume that $\ObsDiam(X) < \varepsilon$ for a number $\varepsilon$.
Then we have $\ObsDiam(X;-\varepsilon) < \varepsilon$
and so $\diam(f_*\mu_X;1-\varepsilon) < \varepsilon$
for any function $f \in \mathcal{L}{\it ip}_1(X)$.
There are two numbers $a < b$ such that
$f_*\mu_X([\,a,b\,]) \ge 1-\varepsilon$ and $b-a < \varepsilon$.
We set $c := (a+b)/2$.
Since $[\,a,b\,] \subset (\,c-\varepsilon/2,c+\varepsilon/2\,)$,
we have
$\mu_X(|f-c| < \varepsilon/2) \ge 1-\varepsilon$,
which implies $\dKF(f,c) \le \varepsilon$.
By Lemma \ref{lem:dconc-pt}, $\dconc(X,*) \le \varepsilon$.
The first inequality has been proved.
We prove the second inequality.
Let $\varepsilon := \dconc(X,*)$.
By Lemma \ref{lem:dconc-pt},
for any function $f \in \mathcal{L}{\it ip}_1(X)$ there is a real number $c$
such that $\dKF(f,c) \le \varepsilon$.
Since $f_*\mu_X([\,c-\varepsilon,c+\varepsilon\,])
= \mu_X(|f-c| \le \varepsilon) \ge 1-\varepsilon$,
we have
$\diam(f_*\mu_X;1-\varepsilon) \le 2\varepsilon$
and so $\ObsDiam(X;-\varepsilon) \le 2\varepsilon$.
This completes the proof.
\end{proof}
Proposition \ref{prop:ObsDiam-dconc} implies the following
\begin{cor} \label{cor:ObsDiam-dconc}
Let $X_n$, $n=1,2,\dots$, be mm-spaces.
The sequence $\{X_n\}_{n=1}^\infty$ is a L\'evy family if and only if
it concentrates to a one-point mm-space.
\end{cor}
\begin{defn}[$h$-Homogeneous measure]
\index{h-homogeneous measure@$h$-homogeneous measure}
A Borel measure $\mu$ on a metric space $X$ is said to be
$h$-\emph{homogeneous}, $h \ge 1$,
if
\[
\mu(B_r(x)) \le h\,\mu(B_r(y))
\]
for any two points $x,y \in X$ and for any $r > 0$.
\end{defn}
Note that any $h$-homogeneous Borel measure on a metric space
has full support.
The following is a slight extension of a result in \cite{Funano:est-box}.
\begin{prop} \label{prop:homog}
Let $X_n$, $n=1,2,\dots$, be mm-spaces with $h$-homogeneous measure
for a real number $h \ge 1$.
If $\{X_n\}$ is a L\'evy family and if
the diameter of $X_n$ is bounded away from zero,
then $\{X_n\}$ has no $\square$-convergent subsequence.
\end{prop}
\begin{proof}
Let $X_n$, $n=1,2,\dots$, be mm-spaces with $h$-homogeneous measure,
and assume that $\{X_n\}$ is a L\'evy family and
the diameter of $X_n$ is bounded away from zero.
Suppose that $\{X_n\}$ has a $\square$-convergent subsequence.
Replacing $\{X_n\}$ with the subsequence,
we assume that $X_n$ is $\square$-convergent as $n\to\infty$.
It follows from Proposition \ref{prop:dconc-box} and Corollary \ref{cor:ObsDiam-dconc}
that $X_n$ $\square$-converges to the one-point mm-space $*$.
We set
\[
r := \min\left\{\frac{1}{2(1+h)},\,\frac{1}{5}\inf_n \diam X_n\right\}.
\]
There is a number $n_0$ such that $\square(X_n,*) < r$ for all $n \ge n_0$.
For each $n \ge n_0$,
there are a parameter $\varphi_n : I \to X_n$ and a measurable set
$I_n \subset I$
such that $\mathcal{L}^1(I_n) \ge 1-r$ and
$d_{X_n}(\varphi_n(s),\varphi_n(t)) \le r$
for any $s,t \in I_n$.
We take a point $s_n \in I_n$ for each $n$ and fix it.
Since $\diam X_n \ge 5r$,
there is a point $x_n \in X_n$ for each $n \ge n_0$ such that
$d_{X_n}(\varphi_n(s_n),x_n) > 2r$.
Since $B_r(\varphi_n(s_n))$ and $B_r(x_n)$ are disjoint to each other,
we have
\begin{align*}
1 &\ge \mu_{X_n}(B_r(\varphi_n(s_n))) + \mu_{X_n}(B_r(x_n))\\
&\ge (1+h^{-1})\mu_{X_n}(B_r(\varphi_n(s_n)))
= (1+h^{-1})\mathcal{L}^1(\varphi_n^{-1}(B_r(\varphi_n(s_n))))
\intertext{and, since $\varphi_n^{-1}(B_r(\varphi_n(s_n))) \supset I_n$,
the last term is}
&\ge (1+h^{-1})\mathcal{L}^1(I_n) \ge (1+h^{-1})(1-r)\\
&\ge \left(1+\frac{1}{h}\right)\left(1-\frac{1}{2(1+h)}\right)
= 1+ \frac{1}{2h} > 1,
\end{align*}
which is a contradiction.
This completes the proof.
\end{proof}
Since $\{S^n(1)\}$, $\{\field{C} P^n\}$, $\{SO(n)\}$, $\{SU(n)\}$,
and $\{Sp(n)\}$ are all L\'evy families of mm-spaces with $1$-homogeneous measure
(see Examples \ref{ex:CPn} and \ref{ex:SOSUSp}), we have the following.
\begin{cor} \label{cor:homog}
Any subsequence of $\{S^n(1)\}$, $\{\field{C} P^n\}$, $\{SO(n)\}$, $\{SU(n)\}$,
and $\{Sp(n)\}$ is $\square$-divergent.
\end{cor}
\section{$N$-Measurement
and nondegeneracy of the observable distance}
In this section, we prove that the observable distance function
is a metric on $\mathcal{X}$ by using measurements.
\begin{defn}[$N$-Measurement]
\index{measurement} \index{N-measurement@$N$-measurement}
Let $X$ be an mm-space and $N$ a natural number.
Denote by $\mathcal{M}(N)$ the set of Borel probability measures on $\field{R}^N$
equipped with the Prohorov metric $d_P$.
We call the subset of $\mathcal{M}(N)$
\[
\mathcal{M}(X;N) := \{\; F_*\mu_X \mid F : X \to (\field{R}^N,\|\cdot\|_\infty)
\ \text{is $1$-Lipschitz}\;\}
\]
\index{MXN@$\mathcal{M}(X;N)$}
the \emph{$N$-measurement of $X$}.
\end{defn}
\begin{lem}
The $N$-measurement $\mathcal{M}(X;N)$ of $X$ is a closed subset of $\mathcal{M}(N)$.
\end{lem}
\begin{proof}
Assume that a sequence $(F_i)_*\mu_X \in \mathcal{M}(X;N)$ $d_P$-converges
to a measure $\mu \in \mathcal{M}(N)$, where $F_i : X \to (\field{R}^N,\|\cdot\|_\infty)$, $i=1,2,\dots$,
are $1$-Lipschitz maps.
It suffices to prove that $\mu$ belongs to $\mathcal{M}(X;N)$.
Let us first prove that $\{F_i(x_0)\}_{i=1}^\infty$ is a bounded sequence
in $\field{R}^N$, where $x_0$ is a point in $X$.
In fact, the $1$-Lipschitz continuity of $F_i$
implies $F_i(B_1(x_0)) \subset B_1(F_i(x_0))$ and so
\[
(F_i)_*\mu_X(B_1(F_i(x_0))) = \mu_X(F_i^{-1}(B_1(F_i(x_0))))
\ge \mu_X(B_1(x_0)) > 0.
\]
If $\{F_i\}$ has a subsequence $\{F_{i_j}\}$ for which
$\|F_{i_j}(x_0)\|_\infty$ diverges to infinity as $j \to \infty$,
then, since $B_1(F_{i_j}(x_0))$ does not intersect
$U^N_R(o) := \{\,x \in \field{R}^N \mid \|x\|_\infty < R\,\}$ for any fixed $R > 0$
and for every sufficiently large $j$,
we have
\begin{align*}
\mu(\field{R}^N \setminus U^N_R(o))
&\ge \liminf_{j\to\infty} (F_{i_j})_*\mu_X(\field{R}^N \setminus U^N_R(o))\\
&\ge \liminf_{j\to\infty} (F_{i_j})_*\mu_X(B_1(F_{i_j}(x_0)))
\ge \mu_X(B_1(x_0)) > 0,
\end{align*}
which is a contradiction to
$\lim_{R\to +\infty} \mu(\field{R}^N \setminus U^N_R(o)) = 0$.
Therefore, $\{F_i(x_0)\}$ is bounded.
Applying Lemma \ref{lem:cpt-1Lip} yields that $\{F_i\}$ has
a subsequence $\{F_{i_j}\}$ that converges in measure to
a $1$-Lipschitz map $F : X \to (\field{R}^N,\|\cdot\|_\infty)$.
By Lemma \ref{lem:di-me},
$(F_{i_j})_*\mu_X$ $d_P$-converges to $F_*\mu_X$.
We thus obtain $\mu = F_*\mu_X \in \mathcal{M}(X;N)$.
\end{proof}
\begin{lem} \label{lem:dom-M}
For two mm-spaces $X$ and $Y$, the following {\rm(1)} and {\rm(2)}
are equivalent to each other.
\begin{enumerate}
\item $X$ is dominated by $Y$.
\item $\mathcal{M}(X;N) \subset \mathcal{M}(Y;N)$ for any natural number $N$.
\end{enumerate}
\end{lem}
\begin{proof}
We prove `(1) $\implies$ (2)'.
$X \prec Y$ implies that
there is a $1$-Lipschitz map $f : Y \to X$ such that $f_*\mu_Y = \mu_X$.
Take any $F_*\mu_X \in \mathcal{M}(X;N)$, where
$F : X \to (\field{R}^N,\|\cdot\|_\infty)$ is any $1$-Lipschitz map.
The composition $F \circ f : Y \to (\field{R}^N,\|\cdot\|_\infty)$ is $1$-Lipschitz
and
$F_*\mu_X = F_*f_*\mu_Y = (F \circ f)_*\mu_Y \in \mathcal{M}(Y;N)$,
so that we have $\mathcal{M}(X;N) \subset \mathcal{M}(Y;N)$.
We prove `(2) $\implies$ (1)'.
Assume that $\mathcal{M}(X;N) \subset \mathcal{M}(Y;N)$ for any natural number $N$.
According to Corollary \ref{cor:XN},
there is a sequence of measures $\underline{\mu}_N \in \mathcal{M}(X;N)$,
$N=1,2,\dots$, such that
$\underline{X}_N = (\field{R}^N,\|\cdot\|_\infty,\underline{\mu}_N)$
$\square$-converges to $X$ as $N\to\infty$.
The assumption implies $\underline{\mu}_N \in \mathcal{M}(Y;N)$,
so that we have a $1$-Lipschitz map $F : Y \to \field{R}^N$
with $\underline{\mu}_N = F_*\mu_Y$,
which means that $\underline{X}_N$ is dominated by $Y$.
By Theorem \ref{thm:dom}, $X$ is dominated by $Y$.
This completes the proof.
\end{proof}
\begin{lem} \label{lem:M-dconc}
Let $X$ and $Y$ be two mm-spaces.
For any natural number $N$ we have
\[
d_H(\mathcal{M}(X;N),\mathcal{M}(Y;N)) \le N\cdot\dconc(X,Y),
\]
where the Hausdorff distance $d_H$ is defined with respect to
the Prohorov metric $d_P$.
\end{lem}
\begin{proof}
Assume that $\dconc(X,Y) < \varepsilon$ for a number $\varepsilon$.
There are two parameters $\varphi : I \to X$ and $\psi : I \to Y$
such that
\begin{equation}
\label{eq:M-dconc}
d_H(\varphi^*\mathcal{L}{\it ip}_1(X),\psi^*\mathcal{L}{\it ip}_1(Y)) < \varepsilon.
\end{equation}
Let us prove that $\mathcal{M}(X;N) \subset B_{N\varepsilon}(\mathcal{M}(Y;N))$.
Take any $F_*\mu_X \in \mathcal{M}(X;N)$, where $F : X \to (\field{R}^N,\|\cdot\|_\infty)$
is a $1$-Lipschitz map.
Setting $(f_1,\dots,f_N) := F$ we have $f_i \in \mathcal{L}{\it ip}_1(X)$ and so
$f_i \circ\varphi \in \varphi^*\mathcal{L}{\it ip}_1(X)$.
By \eqref{eq:M-dconc},
there is a function $g_i \in \mathcal{L}{\it ip}_1(Y)$ such that
$\dKF(f_i\circ\varphi,g_i\circ\psi) < \varepsilon$.
Since $G := (g_1,\dots,g_N) : Y \to (\field{R}^N,\|\cdot\|_\infty)$ is $1$-Lipschitz,
we have $G_*\mu_Y \in \mathcal{M}(Y;N)$.
We prove $d_P(F_*\mu_X,G_*\mu_Y) \le N\varepsilon$ in the following.
For this, it suffices to prove
$F_*\mu_X(B_\varepsilon(A)) \ge G_*\mu_Y(A) - N\varepsilon$
for any Borel subset $A \subset \field{R}^N$.
Since $F_*\mu_X = (F\circ\varphi)_*\mathcal{L}^1$
and $G_*\mu_Y = (G\circ\psi)_*\mathcal{L}^1$, we have
\[
F_*\mu_X(B_\varepsilon(A)) = \mathcal{L}^1((F\circ\varphi)^{-1}(B_\varepsilon(A))),
\quad
G_*\mu_Y(A) = \mathcal{L}^1((G\circ\psi)^{-1}(A)).
\]
It is sufficient to prove
\[
\mathcal{L}^1((G\circ\psi)^{-1}(A) \setminus (F\circ\varphi)^{-1}(B_\varepsilon(A))) \le N\varepsilon.
\]
If we take
$s \in (G\circ\psi)^{-1}(A) \setminus (F\circ\varphi)^{-1}(B_\varepsilon(A))$,
then
$G\circ\psi(s) \in A$ and $F \circ\varphi(s) \notin B_\varepsilon(A)$
together imply
\[
\|F\circ\varphi(s) - G\circ\psi(s)\|_\infty > \varepsilon
\]
and therefore
\begin{align*}
&\mathcal{L}^1((G\circ\psi)^{-1}(A) \setminus
(F\circ\varphi)^{-1}(B_\varepsilon(A)))\\
&\le \mathcal{L}^1(\{\;s \in I \mid
\|F\circ\varphi(s) - G\circ\psi(s)\|_\infty > \varepsilon\;\})\\
&= \mathcal{L}^1\left( \bigcup_{i=1}^N \{\;s\in I \mid
|f_i\circ\varphi(s)-g_i\circ\psi(s)| > \varepsilon\;\}\right)\\
&\le \sum_{i=1}^N \mathcal{L}^1(\{\;s\in I \mid
|f_i\circ\varphi(s)-g_i\circ\psi(s)| > \varepsilon\;\})\\
&\le N\varepsilon,
\end{align*}
where the last inequality follows from
$\dKF(f_i\circ\varphi,g_i\circ\psi) < \varepsilon$.
We thus obtain $d_P(F_*\mu_X,G_*\mu_Y) \le N\varepsilon$,
so that $\mathcal{M}(X;N) \subset B_{N\varepsilon}(\mathcal{M}(Y;N))$.
Since this also holds if we exchange $X$ and $Y$,
we have
\[
d_H(\mathcal{M}(X;N),\mathcal{M}(Y;N)) \le N\varepsilon.
\]
This completes the proof.
\end{proof}
\begin{thm} \label{thm:dconc}
The function $\dconc$ is a metric on $\mathcal{X}$.
\end{thm}
\begin{proof}
The symmetricity is clear.
A triangle inequality is obtained in the same way as in the proof of
Theorem \ref{thm:box}, by using Lemma \ref{lem:box-tri}
and Proposition \ref{prop:dconc-box}.
We prove the nondegeneracy.
Assume that $\dconc(X,Y) = 0$ for two mm-spaces $X$ and $Y$.
Then, Lemma \ref{lem:M-dconc} implies that $\mathcal{M}(X;N) = \mathcal{M}(Y;N)$
for any $N$, which together with Lemma \ref{lem:dom-M} yields
that $X \prec Y$ and $Y \prec X$. By Proposition \ref{prop:Liporder},
$X$ and $Y$ are mm-isomorphic to each other.
This completes the proof.
\end{proof}
\begin{rem}
Without the mm-reconstruction theorem, we obtain the
nondegeneracy of the box metric $\square$ in the same way as in
the proof of Theorem \ref{thm:dconc}.
\end{rem}
\begin{defn}[Concentration topology]
\index{concentration topology}
We call the topology on $\mathcal{X}$ induced from $\dconc$
the \emph{concentration topology}.
\end{defn}
\begin{prop} \label{prop:box-dconc-precpt}
Let $\mathcal{Y} \subset \mathcal{X}$ be a $\square$-precompact family of
mm-isomorphism classes of mm-spaces.
Then, the concentration topology coincides with
the topology induced from the box metric
on the $\square$-closure of $\mathcal{Y}$.
\end{prop}
\begin{proof}
Let $\overline{\mathcal{Y}}^\square$ be the $\square$-closure of $\mathcal{Y}$.
It follows from the completeness of $\mathcal{X}$ (see Theorem \ref{thm:box-complete})
that $\overline{\mathcal{Y}}^\square$ is $\square$-compact.
Applying the homeomorphism theorem for the identity map
$\id_\mathcal{X} : (\overline{\mathcal{Y}}^\square,\square) \to (\overline{\mathcal{Y}}^\square,\dconc)$
yields that it is a homeomorphism.
This completes the proof.
\end{proof}
Combining Proposition \ref{prop:box-dconc-precpt}
and Corollary \ref{cor:precpt} implies
\begin{cor} \label{cor:box-dconc-precpt}
Let $\mathcal{Y} \subset \mathcal{X}$ be a uniform family of mm-isomorphism classes
of mm-spaces.
Then, the concentration topology coincides with
the topology induced from the box metric on the $\square$-closure of $\mathcal{Y}$.
In particular, if $\{X_n\}_{n=1}^\infty$ is a uniform sequence of mm-spaces
that concentrates to an mm-space $X$,
then $X_n$ $\square$-converges to $X$.
\end{cor}
\begin{rem}
Recall that $\{S^n(1)\}$ concentrates to a one-point space,
but has no $\square$-convergent subsequence (see Corollary \ref{cor:homog}).
It is a non-uniform sequence.
A non-uniform sequence of mm-spaces is more interesting
than uniform one for the study of concentration.
\end{rem}
\section{Convergence of $N$-measurements}
In this section we prove the following.
\begin{thm}[Observable criterion for concentration] \label{thm:A}
\index{observable criterion for concentration}
Let $X$ and $X_n$, $n=1,2,\dots$, be mm-spaces.
Then, the following {\rm(1)} and {\rm(2)} are equivalent to each other.
\begin{enumerate}
\item The sequence $\{X_n\}$ concentrates to $X$.
\item For any natural number $N$, the $N$-measurement $\mathcal{M}(X_n;N)$ of $X_n$
converges to $\mathcal{M}(X;N)$ with respect to
the Hausdorff distance defined from the Prohorov metric.
\end{enumerate}
\end{thm}
`(1) $\implies$ (2)' follows from Lemma \ref{lem:M-dconc}.
To prove `(2) $\implies$ (1)' we need several lemmas.
From now on let $X$ and $Y$ be two mm-spaces.
\begin{lem} \label{lem:pullback-me}
Let $p : X \to Y$ be a Borel measurable map such that $p_*\mu_X = \mu_Y$.
For any two Borel measurable functions $f,g : Y \to \field{R}$, we have
\[
\dKF(p^*f,p^*g) = \dKF(f,g),
\]
where $p^*f := f\circ p$.
\end{lem}
\begin{proof}
The lemma follows from
\begin{align*}
\mu_Y(|f-g| > \varepsilon) &= p_*\mu_X(|f-g| > \varepsilon)
= \mu_X(p^{-1}(\{|f-g| > \varepsilon\}))\\
&= \mu_X(|p^*f-p^*g| > \varepsilon).
\end{align*}
\end{proof}
\begin{defn}[Enforce $\varepsilon$-concentration]
\index{enforce epsilon-concentration@enforce $\varepsilon$-concentration}
A Borel measurable map $p : X \to Y$ is said to
\emph{enforce $\varepsilon$-concentration of $X$ to $Y$}
if
\[
d_H(\mathcal{L}{\it ip}_1(X),p^*\mathcal{L}{\it ip}_1(Y)) \le \varepsilon.
\]
\end{defn}
\begin{lem} \label{lem:enforce-conc}
If a Borel measurable map $p : X \to Y$ enforces $\varepsilon$-concentration
of $X$ to $Y$, then
\[
\dconc(X,Y) \le 2\,d_P(p_*\mu_X,\mu_Y) + \varepsilon.
\]
\end{lem}
\begin{proof}
We take a parameter $\varphi : I \to X$.
The map $\psi := p\circ\varphi : I \to Y$ is a parameter
of $(Y,p_*\mu_X)$.
By Lemma \ref{lem:pullback-me},
\begin{align*}
d_H(\varphi^*\mathcal{L}{\it ip}_1(X),\psi^*\mathcal{L}{\it ip}_1(Y))
&= d_H(\varphi^*\mathcal{L}{\it ip}_1(X),\varphi^*p^*\mathcal{L}{\it ip}_1(Y))\\
&= d_H(\mathcal{L}{\it ip}_1(X),p^*\mathcal{L}{\it ip}_1(Y)) \le \varepsilon,
\end{align*}
which implies $\dconc(X,(Y,p_*\mu_X)) \le \varepsilon$.
Since
\[
\dconc(Y,(Y,p_*\mu_X)) \le \square(Y,(Y,p_*\mu_X))
\le 2\,d_P(\mu_Y,p_*\mu_X),
\]
the lemma follows from a triangle inequality.
\end{proof}
\begin{lem}
For a Borel measurable map $p : X \to Y$,
the following {\rm(1)} and {\rm(2)} are equivalent to each other.
\begin{enumerate}
\item $p^*\mathcal{L}{\it ip}_1(Y) \subset \mathcal{L}{\it ip}_1(X)$.
\item $p : X \to Y$ is $1$-Lipschitz continuous.
\end{enumerate}
\end{lem}
\begin{proof}
`(2) $\implies$ (1)' is obvious.
We prove `(1) $\implies$ (2)'.
Take any two points $x,y \in X$ and fix them.
The function $f := d_Y(p(x),\cdot)$ belongs to $\mathcal{L}{\it ip}_1(Y)$,
which together with (1) implies
$p^*f \in p^*\mathcal{L}{\it ip}_1(Y) \subset \mathcal{L}{\it ip}_1(X)$.
Therefore,
\[
d_Y(p(x),p(y)) = |p^*f(x)-p^*f(y)| \le d_X(x,y).
\]
This completes the proof.
\end{proof}
\begin{lem} \label{lem:1-Lip-up-to-Lip1}
For a Borel measurable map $p : X \to Y$
and for two real numbers $\varepsilon,\delta > 0$,
we consider the two following conditions.
\begin{enumerate}
\item[(A${}_\varepsilon$)] $p^*\mathcal{L}{\it ip}_1(Y) \subset B_\varepsilon(\mathcal{L}{\it ip}_1(X))$.
\item[(B${}_\delta$)] $p$ is $1$-Lipschitz up to $\delta$.
\end{enumerate}
Then we have the following {\rm(1)} and {\rm(2)}.
\begin{enumerate}
\item There exists a real number $\delta = \delta(Y,\varepsilon) > 0$
for any $\varepsilon > 0$
such that $\lim_{\varepsilon\to 0} \delta(Y,\varepsilon) = 0$
and if {\rm(A${}_\varepsilon$)} holds and
if $d_P(p_*\mu_X,\mu_Y) < \varepsilon$,
then we have {\rm(B${}_\delta$)}.
\item
If {\rm(B${}_\delta$)} holds, then we have {\rm(A${}_\delta$)}.
\end{enumerate}
\end{lem}
\begin{proof}
We prove (1).
For a number $\varepsilon' > 0$, let
$N(\varepsilon')$ be the infimum of $\#\mathcal{N}$,
where $\mathcal{N}$ runs over all nets in $Y$ such that
there is a Borel subset $Y_0 \subset Y$
with the property that $\mu_Y(Y_0) \ge 1-\varepsilon'$
and $\mathcal{N} \subset Y_0$ is an $\varepsilon'$-net of $Y_0$.
Since we have a compact subset of $Y$ whose $\mu_Y$-measure
is arbitrarily close to $1$, the number $N(\varepsilon')$ is finite.
For any $\varepsilon > 0$,
there is a number $\varepsilon' = \varepsilon'(\varepsilon) > 0$
such that $\lim_{\varepsilon\to 0} \varepsilon' = 0$ and
$N(\varepsilon') \le 1/\sqrt{\varepsilon}$.
We find a Borel subset $Y_0 \subset Y$ and an $\varepsilon'$-net
$\mathcal{N} \subset Y_0$ such that
\[
\mu_Y(Y_0) \ge 1-\varepsilon' \quad\text{and}\quad
\#\mathcal{N} = N(\varepsilon') \le \frac{1}{\sqrt{\varepsilon}}.
\]
It follows from $d_P(p_*\mu_X,\mu_Y) < \varepsilon$
that
\[
\mu_X(p^{-1}(B_\varepsilon(Y_0))) = p_*\mu_X(B_\varepsilon(Y_0))
\ge \mu_Y(Y_0)-\varepsilon \ge 1-\varepsilon-\varepsilon'.
\]
Let $y \in \mathcal{N}$ be any point and set
$f_y := d_Y(y,\cdot)$.
(A${}_\varepsilon$) implies that
$p^*f_y \in p^*\mathcal{L}{\it ip}_1(Y) \subset B_\varepsilon(\mathcal{L}{\it ip}_1(X))$,
so that we have a function $g_y \in \mathcal{L}{\it ip}_1(X)$ with
$\dKF(p^*f_y,g_y) \le \varepsilon$, namely
$\mu_X(|p^*f_y - g_y| > \varepsilon) \le \varepsilon$.
Setting
\[
X_0 := p^{-1}(B_\varepsilon(Y_0)) \setminus
\bigcup_{y\in\mathcal{N}} \{\;|p^*f_y - g_y| > \varepsilon\;\},
\]
we have
\[
\mu_X(X \setminus X_0) \le \#\mathcal{N}\cdot\varepsilon + \varepsilon + \varepsilon'
\le \sqrt{\varepsilon} + \varepsilon + \varepsilon'.
\]
Let $\pi : Y \to \mathcal{N}$ be a Borel measurable
nearest point projection.
We take any two points $x,x' \in X_0$.
Since $p(x) \in B_\varepsilon(Y_0)$ and $\pi(p(x)) \in \mathcal{N}$,
we have
\begin{align*}
d_Y(p(x),p(x')) &\le d_Y(\pi(p(x)),p(x')) + \varepsilon + \varepsilon'
= p^*f_{\pi(p(x))}(x') + \varepsilon + \varepsilon'\\
&\le g_{\pi(p(x))}(x') + 2\varepsilon + \varepsilon',\\
g_{\pi(p(x))}(x) &\le p^*f_{\pi(p(x))}(x) + \varepsilon
= d_Y(\pi(p(x)),p(x)) + \varepsilon\\
&\le 2\varepsilon + \varepsilon'
\end{align*}
and hence
\begin{align*}
d_Y(p(x),p(x')) &\le g_{\pi(p(x))}(x') - g_{\pi(p(x))}(x)
+ 4\varepsilon+2\varepsilon'\\
&\le d_X(x,x') + 4\varepsilon+2\varepsilon'.
\end{align*}
Setting $\delta := \max\{4\varepsilon+2\varepsilon',
\sqrt{\varepsilon} + \varepsilon + \varepsilon'\}$,
the map $p$ is $1$-Lipschitz up to $\delta$.
We prove (2).
Take any $f \in \mathcal{L}{\it ip}_1(Y)$.
(B${}_\delta$) implies that $p^*f$ is $1$-Lipschitz up to $\delta$.
By Lemma \ref{lem:Lip-approx},
there is a function $\tilde{f} \in \mathcal{L}{\it ip}_1(X)$ such that
$\dKF(\tilde{f},p^*f) \le \delta$.
Therefore we have
\[
p^*\mathcal{L}{\it ip}_1(Y) \subset B_\delta(\mathcal{L}{\it ip}_1(X)).
\]
This completes the proof.
\end{proof}
\begin{lem} \label{lem:M-p}
Let $Y$ be an mm-space.
For any $\varepsilon > 0$,
there exists a natural number $N = N(Y,\varepsilon)$
depending only on $Y$ and $\varepsilon$ such that,
if $\mathcal{M}(Y;N) \subset B_\varepsilon(\mathcal{M}(X;N))$ for an mm-space $X$,
then there exists a Borel measurable map $p : X \to Y$
that is $1$-Lipschitz up to $5\varepsilon$ and satisfies
\[
d_P(p_*\mu_X,\mu_Y) \le 15\varepsilon.
\]
\end{lem}
\begin{proof}
Corollary \ref{cor:XN} implies that
there is a natural number $N(Y,\varepsilon)$ and
a measure $\underline{\mu}_N \in \mathcal{M}(Y;N)$
such that $\square(\underline{Y}_N,Y) < \varepsilon/3$,
where $\underline{Y}_N := (\field{R}^N,\|\cdot\|_\infty,\underline{\mu}_N)$.
By Lemma \ref{lem:box-eps-mm-iso},
there is an $\varepsilon$-mm-isomorphism $\Psi : \underline{Y}_N \to Y$.
We find a Borel subset $\underline{Y}_{N,0} \subset \underline{Y}_N$
such that $\underline{\mu}_N(\underline{Y}_{N,0}) \ge 1-\varepsilon$
and
\[
|\;d_Y(\Psi(u),\Psi(v)) - \|u-v\|_\infty\;| \le \varepsilon
\]
for any $u,v \in \underline{Y}_{N,0}$.
Since $\underline{\mu}_N \in \mathcal{M}(Y;N) \subset B_\varepsilon(\mathcal{M}(X;N))$,
there is a $1$-Lipschitz map $\Phi' : X \to (\field{R}^N,\|\cdot\|_\infty)$
such that
\[
d_P(\underline{\mu}_N,\Phi'_*\mu_X) \le \varepsilon.
\]
We see that
\[
\Phi'_*\mu_X(B_\varepsilon(\underline{Y}_{N,0}))
\ge \underline{\mu}_N(\underline{Y}_{N,0}) - \varepsilon
\ge 1-2\varepsilon.
\]
By Lemma \ref{lem:eps-proj}, there is a Borel measurable
$\varepsilon$-projection
$\pi : \field{R}^N \to \underline{Y}_{N,0}$ with
$\pi|_{\underline{Y}_{N,0}} = \id_{\underline{Y}_{N,0}}$.
Let $\Psi' := \Psi\circ\pi : \field{R}^N \to Y$.
For any $u,v \in B_\varepsilon(\underline{Y}_{N,0})$,
\begin{align*}
&|\;d_Y(\Psi'(u),\Psi'(v)) - \|u-v\|_\infty\;|\\
&\le |\;d_Y(\Psi(\pi(u)),\Psi(\pi(v))) - \|\pi(u)-\pi(v)\|_\infty\;|
+ 4\varepsilon\\
&\le 5\varepsilon.
\end{align*}
By Lemma \ref{lem:push-di-Lip},
\[
d_P(\Psi'_*\underline{\mu}_N,\Psi'_*\Phi'_*\mu_X) \le
d_P(\underline{\mu}_N,\Phi'_*\mu_X) + 10\varepsilon \le 11\varepsilon.
\]
It follows from $\underline{\mu}_N(\underline{Y}_{N,0}) \ge 1-\varepsilon$
that $d_P(\pi_*\underline{\mu}_N,\underline{\mu}_N)
\le d_{KF}^{\underline{\mu}_N}(\pi,\id_{\field{R}^N}) \le \varepsilon$.
Besides we have $\pi_*\underline{\mu}_N(\underline{Y}_{N,0}) = 1
\ge \underline{\mu}_N(\underline{Y}_{N,0}) \ge 1-\varepsilon$.
Applying Lemma \ref{lem:push-di-Lip} yields
\[
d_P(\Psi'_*\underline{\mu}_N,\Psi_*\underline{\mu}_N)
= d_P(\Psi_*\pi_*\underline{\mu}_N,\Psi_*\underline{\mu}_N)
\le d_P(\pi_*\underline{\mu}_N,\underline{\mu}_N) + 2\varepsilon
\le 3\varepsilon
\]
and hence, by a triangle inequality,
\[
d_P((\Psi'\circ\Phi')_*\mu_X,\Psi_*\underline{\mu}_N)
\le 14\varepsilon.
\]
Since $\Psi : \underline{Y}_N \to Y$ is an $\varepsilon$-mm-isomorphism,
we have $d_P(\Psi_*\underline{\mu}_N,\mu_Y) \le \varepsilon$,
which together with the above inequality implies
\begin{align*}
d_P((\Psi'\circ\Phi')_*\mu_X,\mu_Y)
\le 15\varepsilon.
\end{align*}
Setting $X_0 := {\Phi'}^{-1}(B_\varepsilon(\underline{Y}_{N,0})) \subset X$,
we have, for any $x,y \in X_0$,
\begin{align*}
d_Y(\Psi'\circ\Phi'(x),\Psi'\circ\Phi'(y))
\le \|\Phi'(x) - \Phi'(y)\|_\infty + 5\varepsilon
\le d_X(x,y) + 5\varepsilon.
\end{align*}
Moreover we have
\[
\mu_X(X_0) = \Phi'_*\mu_X(B_\varepsilon(\underline{Y}_{N,0}))
\ge 1-2\varepsilon.
\]
The desired map is $p := \Psi'\circ\Phi' : X \to Y$.
This completes the proof.
\end{proof}
\begin{lem} \label{lem:me-diff-di}
For any measurable maps $F = (f_1,\dots,f_N) : X \to \field{R}^N$,
$G = (g_1,\dots,g_N) : Y \to \field{R}^N$, and for any $i,j = 1,2,\dots,N$,
we have
\begin{align}
|\;\dKF(f_i,f_j) - \dKF(g_i,g_j)\;| &\le 2\,d_P(F_*\mu_X,G_*\mu_Y), \tag{1}\\
|\;\dKF([f_i],[f_j]) - \dKF([g_i],[g_j])\;| &\le 2\,d_P(F_*\mu_X,G_*\mu_Y).
\tag{2}
\end{align}
\end{lem}
\begin{proof}
We prove (1).
We take any $i$ and $j$ in $\{1,2,\dots,N\}$ and fix them.
(1) is trivial if $i = j$. We then assume $i \neq j$.
Suppose that $d_P(F_*\mu_X,G_*\mu_Y) < \varepsilon$ and
$\dKF(f_i,f_j) < \rho$ for two numbers $\varepsilon$ and $\rho$.
We set
\[
\rho' := \rho + 2\varepsilon, \quad
X_0 := \{\;|x_i-x_j| \ge \rho'\;\}, \quad
d_{X_0}(x) := \inf_{x' \in X_0} \|x-x'\|_\infty,
\]
where $\{\;|x_i-x_j| \ge \rho'\;\} :=
\{\;(x_1,x_2,\dots,x_N) \in \field{R}^N \mid |x_i-x_j| \ge \rho'\;\}$.
Let us now prove
\begin{align}
d_{X_0}(x) = \max\left\{\;\frac{1}{2}(\rho'-|x_i-x_j|),
0\;\right\}.
\label{eq:me-diff-di-1}
\end{align}
Let $r$ be the right-hand side of \eqref{eq:me-diff-di-1}.
Taking any point $x' = (x_1',\dots,x_N') \in X_0$, we have
$|x_i'-x_j'| \ge \rho'$.
If $r > 0$, then $\rho' = |x_i-x_j|+2r$ and so
$|x_i'-x_j'| \ge |x_i-x_j|+2r$. This implies
\begin{align*}
2\|x-x'\|_\infty \ge |x_i-x_i'| + |x_j-x_j'|
\ge |x_i'-x_j'| - |x_i-x_j| \ge 2r
\end{align*}
and thus $d_{X_0}(x) \ge r$.
We next prove $d_{X_0}(x) \le r$. This is trivial if $x \in X_0$.
We suppose $x \notin X_0$. Then,
\[
r = \frac{1}{2} (\rho'-|x_i-x_j|) > 0.
\]
Put
\begin{align*}
x_i' :=
\begin{cases}
x_i + r & \text{if $x_i \ge x_j$},\\
x_i - r & \text{if $x_i < x_j$},
\end{cases}
\quad\text{and}\quad
x_j' :=
\begin{cases}
x_j - r & \text{if $x_i \ge x_j$},\\
x_j + r & \text{if $x_i < x_j$}.
\end{cases}
\end{align*}
For any $k \in \{1,2,\dots,N\}$ with $k \neq i,j$, we set
$x_k' := x_k$.
Since $|x_i'-x_j'| = |x_i-x_j| + 2r = \rho'$,
the point $x' = (x_1',\dots,x_N')$ belongs to $X_0$,
which together with $\|x-x'\|_\infty = r$ implies $d_{X_0}(x) \le r$.
We thus obtain \eqref{eq:me-diff-di-1}.
It follows from \eqref{eq:me-diff-di-1} that
\begin{align*}
B_\varepsilon(X_0) = \{\;|x_i-x_j| \ge \rho\;\},
\end{align*}
which together with $\dKF(f_i,f_j) < \rho$ and
$d_P(F_*\mu_X,G_*\mu_Y) < \varepsilon$ leads to
\begin{align*}
\rho &\ge \mu_X(|f_i-f_j| \ge \rho) = F_*\mu_X(|x_i-x_j| \ge \rho)
= F_*\mu_X(B_\varepsilon(X_0))\\
&\ge G_*\mu_Y(X_0) - \varepsilon
= \mu_Y(|g_i-g_j| \ge \rho') - \varepsilon
\end{align*}
and so $\dKF(g_i,g_j) \le \rho + 2\varepsilon$.
Letting $\varepsilon \to d_P(F_*\mu_X,G_*\mu_Y)$ and
$\rho \to \dKF(f_i,f_j)$ yields
\[
\dKF(g_i,g_j) \le \dKF(f_i,f_j) + 2\,d_P(F_*\mu_X,G_*\mu_Y).
\]
Since this also hold if we exchange $i$ and $j$,
we obtain (1).
We prove (2).
It holds that for any $\mathbf{c} \in \field{R}^N$,
\[
d_P((F+\mathbf{c})_*\mu_X,(G+\mathbf{c})_*\mu_Y) = d_P(F_*\mu_X,G_*\mu_Y).
\]
This together with (1) implies that
\begin{align*}
\dKF([f_i],[f_j]) &\le \dKF(f_i+c,f_j+c')\\
&\le \dKF(g_i+c,g_j+c') + 2\,d_P(F_*\mu_X,G_*\mu_Y)
\end{align*}
for any real numbers $c$ and $c'$.
Taking the infimum of the right-hand side over all
$c$ and $c'$ yields
\begin{align*}
\dKF([f_i],[f_j]) \le \dKF([g_i],[g_j]) + 2\,d_P(F_*\mu_X,G_*\mu_Y).
\end{align*}
This completes the proof.
\end{proof}
\begin{lem} \label{lem:KN-Lip}
For any natural number $N$ we have
\[
d_H(K_N(\mathcal{L}_1(X)),K_N(\mathcal{L}_1(Y)))
\le 2 \, d_H(\mathcal{M}(X;N),\mathcal{M}(Y;N)).
\]
\end{lem}
\begin{proof}
Assume that $d_H(\mathcal{M}(X;N),\mathcal{M}(Y;N)) < \varepsilon$ for a number
$\varepsilon$.
It suffices to prove that $d_H(K_N(\mathcal{L}_1(X)),K_N(\mathcal{L}_1(Y))) \le 2\varepsilon$.
For this, we are going to prove
$K_N(\mathcal{L}_1(X)) \subset B_{2\varepsilon}(K_N(\mathcal{L}_1(Y)))$.
Take any matrix $A \in K_N(\mathcal{L}_1(X))$.
We find $N$ functions $f_i \in \mathcal{L}{\it ip}_1(X)$, $i=1,2,\dots,N$,
such that $A = (\dKF([f_i],[f_j]))_{ij}$.
Set $F := (f_1,\dots,f_N) : X \to \field{R}^N$.
By the assumption, there is a
$1$-Lipschitz map $G = (g_1,\dots,g_N) : X \to (\field{R}^N,\|\cdot\|_\infty)$
such that $d_P(F_*\mu_X,G_*\mu_Y) < \varepsilon$.
Lemma \ref{lem:me-diff-di}(2) implies that
\[
|\;\dKF([f_i],[f_j]) - \dKF([g_i],[g_j])\;| < 2\varepsilon
\]
for any $i,j = 1,\dots,N$.
Letting $B := (\dKF([g_i],[g_j]))_{ij}$ we have
$\|A - B\|_\infty < 2\varepsilon$ and
$B \in K_N(\mathcal{L}_1(Y))$.
We therefore obtain $K_N(\mathcal{L}_1(X)) \subset B_{2\varepsilon}(K_N(\mathcal{L}_1(Y)))$.
This also holds if we exchange $X$ and $Y$.
The proof of the lemma is completed.
\end{proof}
\begin{lem} \label{lem:pb-me-di}
Let $p : X \to Y$ be a Borel measurable map.
For any two functions $f,g \in \mathcal{L}{\it ip}_1(Y)$ we have
\begin{align}
|\;\dKF(p^*f,p^*g) - \dKF(f,g)\;| &\le 2\,d_P(p_*\mu_X,\mu_Y), \tag{1}\\
|\;\dKF([p^*f],[p^*g]) - \dKF([f],[g])\;| &\le 2\,d_P(p_*\mu_X,\mu_Y).\tag{2}
\end{align}
In particular, if $p_*\mu_X = \mu_Y$, then the map
\[
p^* : \mathcal{L}_1(Y) \ni [f] \mapsto p^*[f] := [p^*f]
\]
is isometric with respect to $\dKF$.
\end{lem}
\begin{proof}
We prove (1).
Let us first prove that
\begin{align} \label{eq:pb-me-di1}
B_\varepsilon(\{|f-g| \ge \rho+2\varepsilon\})
\subset \{\;|f-g| \ge \rho\;\}
\end{align}
for any $\rho,\varepsilon > 0$.
In fact, if we take a point
$y \in B_\varepsilon(\{|f-g| \ge \rho+2\varepsilon\})$,
then there is a point $y' \in Y$ such that $d_Y(y,y') \le \varepsilon$
and $|f(y')-g(y')| \ge \rho+2\varepsilon$,
which together with the $1$-Lipschitz continuity of $f$ and $g$
imply $|f(y)-g(y)| \ge \rho$.
This proves \eqref{eq:pb-me-di1}.
Assume that $\dKF(p^*f,p^*g) < \rho$ and $d_P(p_*\mu_X,\mu_Y) < \varepsilon$
for two numbers $\rho$ and $\varepsilon$.
It follows from \eqref{eq:pb-me-di1} that
\begin{align*}
\mu_Y(|f-g| \ge \rho+2\varepsilon)
&\le p_*\mu_X(B_\varepsilon(\{|f-g|\ge \rho+2\varepsilon\})) + \varepsilon\\
&\le p_*\mu_X(|f-g| \ge \rho) + \varepsilon\\
&= \mu_X(|p^*f-p^*g| \ge \rho) + \varepsilon
\le \rho + \varepsilon,
\end{align*}
which implies $\dKF(f,g) \le \rho + 2\varepsilon$.
Therefore,
\begin{equation}
\label{eq:pb-me-di2}
\dKF(f,g) \le \dKF(p^*f,p^*g) + 2\,d_P(p_*\mu_X,\mu_Y).
\end{equation}
Using \eqref{eq:pb-me-di1} we also have
\begin{equation}
\label{eq:pb-me-di3}
\dKF(p^*f,p^*g) \le \dKF(f,g) + 2\,d_P(p_*\mu_X,\mu_Y)
\end{equation}
in the same way as above.
Combining \eqref{eq:pb-me-di2} and \eqref{eq:pb-me-di3}
implies (1).
We prove (2). By (1), we have, for any two real numbers $c$ and $c'$,
\begin{align*}
\dKF([p^*f],[p^*g]) &\le \dKF(p^*f+c,p^*g+c')\\
&\le \dKF(f+c,g+c') + 2\,d_P(p_*\mu_X,\mu_Y).
\end{align*}
Taking the infimum of the right-hand side over all $c$ and $c'$ yields
\[
\dKF([p^*f],[p^*g]) \le \dKF([f],[g]) + 2\,d_P(p_*\mu_X,\mu_Y).
\]
In the same way we have
\[
\dKF([f],[g]) \le \dKF([p^*f],[p^*g]) + 2\,d_P(p_*\mu_X,\mu_Y).
\]
Combining these two inequalities implies (2).
This completes the proof.
\end{proof}
Let $\mathcal{F}$ be a metric space with metric $d_{\mathcal{F}}$.
We give an isometric action of a group $G$
on $\mathcal{F}$,
\[
G \times \mathcal{F} \ni (g,x) \mapsto g\cdot x \in \mathcal{F},
\]
with the property that any $G$-orbit is closed in $\mathcal{F}$.
Then, the quotient space $\mathcal{F}/G$ is a metric space,
where the metric $d_{\mathcal{F}/G}$ on $\mathcal{F}/G$ is defined by
\[
d_{\mathcal{F}/G}([x],[y]) := \inf_{x' \in [x],\ y' \in [y]} d_{\mathcal{F}}(x',y')
\]
for any $[x],[y] \in \mathcal{F}/G$.
We later apply the following lemma for $\mathcal{L}{\it ip}_1(X)$ and $\mathcal{L}_1(X)$.
\begin{lem} \label{lem:action-dH}
For any $G$-invariant subset
$\mathcal{L},\mathcal{L'} \subset \mathcal{F}$, we have
\begin{enumerate}
\item for any real number $\varepsilon \ge 0$,
we have the equivalence
\[
\mathcal{L}' \subset B_\varepsilon(\mathcal{L}) \Longleftrightarrow
\mathcal{L}'/G \subset B_\varepsilon(\mathcal{L}/G),
\]
\item $d_H(\mathcal{L},\mathcal{L}') = d_H(\mathcal{L}/G,\mathcal{L}'/G)$.
\end{enumerate}
\end{lem}
\begin{proof}
(2) follows from (1).
We prove (1).
Assume that $\mathcal{L}' \subset B_\varepsilon(\mathcal{L})$.
We take any point $x \in \mathcal{L}'$.
Since $x \in B_\varepsilon(\mathcal{L})$,
there is a sequence of points $x_n \in \mathcal{L}$, $n=1,2,\dots$,
such that
$\limsup_{n\to\infty} d_{\mathcal{F}}(x_n,x) \le \varepsilon$.
This implies
\[
\limsup_{n\to\infty} d_{\mathcal{F}}([x_n],[x])
\le \limsup_{n\to\infty} d_{\mathcal{F}}(x_n,x) \le \varepsilon
\]
and so $[x] \in B_\varepsilon(\mathcal{L}/G)$.
Therefore we have $\mathcal{L}'/G \subset B_\varepsilon(\mathcal{L}/G)$.
We conversely assume that
$\mathcal{L}'/G \subset B_\varepsilon(\mathcal{L}/G)$.
Take any point $x \in \mathcal{L}'$.
Since $[x] \in \mathcal{L}'/G \subset B_\varepsilon(\mathcal{L}/G)$,
there is a sequence of points $x_n \in \mathcal{L}$, $n=1,2,\dots$,
such that $\limsup_{n\to\infty} d_{\mathcal{F}/G}([x_n],[x]) \le \varepsilon$.
Hence, there are elements $g_n,h_n \in G$, $n=1,2,\dots$, such that
\[
\limsup_{n\to\infty} d_{\mathcal{F}}(g_n\cdot x,h_n\cdot x_n)
\le \varepsilon.
\]
Since $\mathcal{L}$ is $G$-invariant,
the point $x_n' := g_n^{-1}h_n\cdot x_n$ belongs to $\mathcal{L}$
and satisfies $d_{\mathcal{F}}(g_n\cdot x,h_n\cdot x_n) =
d_{\mathcal{F}}(x,x_n')$, so that
\[
\limsup_{n\to\infty} d_{\mathcal{F}}(x_n',x) \le \varepsilon.
\]
This implies that $x \in B_\varepsilon(\mathcal{L})$.
We therefore have $\mathcal{L}' \subset B_\varepsilon(\mathcal{L})$.
This completes the proof.
\end{proof}
\begin{lem} \label{lem:L-Ln}
Let $\mathcal{L}$ and $\mathcal{L}_n$, $n=1,2,\dots$, be
compact metric spaces such that
$\mathcal{L}_n$ Gromov-Hausdorff converges to $\mathcal{L}$ as $n\to\infty$.
Let $q_n : \mathcal{L} \to \mathcal{L}_n$, $n=1,2,\dots$, be
$\varepsilon_n$-isometric maps with $\varepsilon_n \to 0$.
Then, there exists a sequence of real numbers $\varepsilon_n' \to 0+$
such that $q_n(\mathcal{L})$ is $\varepsilon_n'$-dense in $\mathcal{L}_n$,
i.e., $B_{\varepsilon_n'}(q_n(\mathcal{L})) = \mathcal{L}_n$.
\end{lem}
\begin{proof}
By $\lim_{n\to\infty} d_{GH}(\mathcal{L}_n,\mathcal{L}) = 0$,
there are numbers $\varepsilon_n' \to 0+$ and
$\varepsilon_n'$-isometries $q_n' : \mathcal{L}_n \to \mathcal{L}$.
Suppose that the lemma does not hold.
Then, there is a number $\delta > 0$ and
a sequence of points $x_n \in \mathcal{L}_n$, $n=1,2,\dots$,
such that $d_{\mathcal{L}_n}(x_n,q_n(\mathcal{L})) \ge \delta$
for every sufficiently large $n$.
This implies that
$d_{\mathcal{L}}(q_n'(x_n),q_n'\circ q_n(\mathcal{L})) \ge \delta/2$
for every sufficiently large $n$.
There is a subsequence $\{n_i\}$ of $\{n\}$ such that
$q_{n_i}'\circ q_{n_i}$ and $q_{n_i}'(x_{n_i})$ both converges as $i\to\infty$.
The limit, say $f : \mathcal{L} \to \mathcal{L}$, of $q_{n_i}'\circ q_{n_i}$
is isometric. The image $f(\mathcal{L})$ does not contain
the limit of $q_{n_i}'(x_{n_i})$ and, in particular, $f$ is not surjective.
This is a contradiction (see \cite{BBI}*{Thm 1.6.14}).
\end{proof}
\begin{lem} \label{lem:A}
Assume {\rm(2)} of Theorem \ref{thm:A}.
Then, there exist Borel measurable maps $p_n : X_n \to Y$, $n=1,2,\dots$,
that enforce $\varepsilon_n$-concentration of $X_n$ to $Y$
and $d_P((p_n)_*\mu_{X_n},\mu_Y) \le \varepsilon_n$ for all $n$
and for some sequence $\varepsilon_n \to 0$.
\end{lem}
\begin{proof}
By the assumption and Lemma \ref{lem:M-p},
there are Borel measurable maps $p_n : X_n \to X$ that are
$1$-Lipschitz up to some $\alpha_n \to 0$,
$n=1,2,\dots$, such that $d_P((p_n)_*\mu_{X_n},\mu_X) \le \alpha_n$.
By Lemma \ref{lem:1-Lip-up-to-Lip1}(2),
we have
$p_n^*\mathcal{L}{\it ip}_1(X) \subset B_{\alpha_n}(\mathcal{L}{\it ip}_1(X_n))$.
By setting $\mathcal{L} := \mathcal{L}_1(X)$ and $\mathcal{L}_n := \mathcal{L}_1(X_n)$,
Lemma \ref{lem:action-dH}(1) yields
$p_n^*\mathcal{L}\;\subset B_{\alpha_n}(\mathcal{L}_n)$.
By Lemma \ref{lem:pb-me-di}(2),
the map $p_n^* : \mathcal{L}\;\to B_{\alpha_n}(\mathcal{L}_n)$
is $2\alpha_n$-isometric with respect to $\dKF$.
Let $\pi_n : B_{\alpha_n}(\mathcal{L}_n) \to \mathcal{L}_n$
be a nearest point projection.
Then it is a $2\alpha_n$-isometry.
Therefore, $\pi_n \circ p_n^* : \mathcal{L}\;\to \mathcal{L}_n$
is $4\alpha_n$-isometric for any $n$.
The assumption together with Lemma \ref{lem:KN-Lip} implies that
$\lim_{n\to\infty} d_H(K_N(\mathcal{L}_n),K_N(\mathcal{L})) = 0$
for any $N$.
By Lemma \ref{lem:KN-GH-2},
$\mathcal{L}_n$ Gromov-Hausdorff converges to $\mathcal{L}$ as $n\to\infty$.
From Lemma \ref{lem:L-Ln}, we find a sequence of numbers
$\alpha_n' \to 0+$ such that
$\pi_n \circ p_n^*(\mathcal{L})$ is $\alpha_n'$-dense in $\mathcal{L}_n$.
Thereby, $p_n^*\mathcal{L}$ is $\varepsilon_n$-dense in
$B_{\alpha_n'}(\mathcal{L}_n)$,
where $\varepsilon_n := \alpha_n + \alpha_n'$.
This implies $d_H(p_n^*\mathcal{L},\mathcal{L}_n) \le \varepsilon_n$.
By Lemma \ref{lem:action-dH}, we have
\[
d_H(p_n^*\mathcal{L}{\it ip}_1(X),\mathcal{L}{\it ip}_1(X_n)) \le \varepsilon_n,
\]
i.e., $p_n : X_n \to X$ enforces $\varepsilon_n$-concentration of $X_n$ to $X$.
This completes the proof.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:A}]
Recall that `(1) $\implies$ (2)' follows from Lemma \ref{lem:M-dconc}.
`(2) $\implies$ (1)' follows from
Lemma \ref{lem:A} and \ref{lem:enforce-conc}.
This completes the proof.
\end{proof}
\begin{cor} \label{cor:enforce}
Let $X_n$ and $Y$ be mm-spaces, where $n = 1,2,\dots$.
Then the following {\rm(1)} and {\rm(2)} are equivalent to each other.
\begin{enumerate}
\item $X_n$ concentrates to $Y$ as $n\to\infty$.
\item There exists a sequence of Borel measurable maps
$p_n : X_n \to Y$, $n=1,2,\dots$, that enforce $\varepsilon_n$-concentration
of $X_n$ to $Y$ and $d_P((p_n)_*\mu_{X_n},\mu_Y) \le \varepsilon_n$
for all $n$ and for some sequence $\varepsilon_n \to 0$.
\end{enumerate}
\end{cor}
\begin{proof}
`(1) $\implies$ (2)' follows from Theorem \ref{thm:A} and Lemma \ref{lem:A}.
`(2) $\implies$ (1)' follows from Lemma \ref{lem:enforce-conc}.
\end{proof}
\section{$(N,R)$-Measurement}
A main purpose of this section is to prove
that the convergence of $(N,R)$-measurement for any $R > 0$ is equivalent to
that of $N$-measurement, which is necessary in the later sections.
\begin{defn}[$(N,R)$-Measurement]
\index{measurement} \index{NR-measurement@$(N,R)$-measurement}
For an mm-space $X$, a natural number $N$, and a real number $R > 0$,
we define
\[
\mathcal{M}(X;N,R) := \{\;\mu \in \mathcal{M}(X;N) \mid
\supp\mu \subset B^N_R\;\},
\]
\index{MXNR@$\mathcal{M}(X;N,R)$}
where $B^N_R := \{\,x \in \field{R}^N \mid \|x\|_\infty \le R\,\}$.
\index{BNR@$B^N_R$}
We call $\mathcal{M}(X;N,R)$ the \emph{$(N,R)$-measurement of $X$}.
\end{defn}
$\mathcal{M}(X;N,R)$ is a compact subset of $\mathcal{M}(N)$ (see Lemma \ref{lem:conv-meas}).
\begin{defn}[$\pi_{\xi,R}$]
\index{pi@$\pi_{\xi,R}$}
For a point $\xi \in \field{R}^N$ and a real number $R \ge 0$,
we define a map $\pi_{\xi,R} : \field{R}^N \to \field{R}^N$ in the following.
For a given point $x = (x_1,\dots,x_N) \in \field{R}^N$
we determine a point $y = (y_1,\dots,y_N) \in \field{R}^N$ as,
for $i=1,\dots,N$,
\[
y_i :=
\begin{cases}
\xi_i+R &\text{if $x_i > \xi_i+R$},\\
\xi_i-R &\text{if $x_i < \xi_i-R$},\\
x_i &\text{if $\xi_i-R \le x_i \le \xi_i+R$}.
\end{cases}
\]
We then define $\pi_{\xi,R}(x) := y$.
\end{defn}
We see that $\pi_{\xi,R}(\field{R}^N) =
B^N_R(\xi) := \{\,x \in \field{R}^N \mid \|x-\xi\|_\infty \le R\,\}$
\index{BNRxi@$B^N_R(\xi)$}
and $\pi_{\xi,R}$ is a unique nearest point projection to $B^N_R(\xi)$
with respect to $\|\cdot\|_\infty$.
Letting $\pi_R := \pi_{o,R}$ \index{piR@$\pi_R$} we have
$\pi_{\xi,R}(x) = \pi_R(x-\xi)+\xi$ for any $x \in \field{R}^N$.
\begin{lem}
The map $\pi_{\xi,R} : \field{R}^N \to \field{R}^N$ is $1$-Lipschitz
with respect to $\|\cdot\|_\infty$.
\end{lem}
\begin{proof}
Take any two points $x,x' \in \field{R}^N$ and set
$y := \pi_{\xi,R}(x)$, $y' := \pi_{\xi,R}(x')$.
It follows from the definition of $\pi_{\xi,R}$ that
$|y_i-y_i'| \le |x_i-x_i'|$ for $i = 1,\dots,N$,
which implies $\|y-y'\|_\infty \le \|x-x'\|_\infty$.
This completes the proof.
\end{proof}
\begin{lem} \label{lem:npp}
Let $N$ be a natural number, $R > 0$ a real number,
and $\mu$ a Borel probability measure on $\field{R}^N$.
Then, for any two points $\xi,\eta \in \field{R}^N$ we have
\begin{align}
\sup_{x \in \field{R}^N} \|\pi_{\xi,R}(x) - \pi_{\eta,R}(x)\|_\infty
&\le \|\xi-\eta\|_\infty,\tag{1}\\
d_P((\pi_{\xi,R})_*\mu,(\pi_{\eta,R})_*\mu) &\le \|\xi-\eta\|_\infty.\tag{2}
\end{align}
\end{lem}
\begin{proof}
We prove (1).
Let $x \in \field{R}^N$ be any point and set
$y := \pi_{\xi,R}(x)$, $y' := \pi_{\eta,R}(x)$.
By the definition of $\pi_{\xi,R}$,
we have $|y_i-y_i'| \le |\xi_i-\eta_i|$ for $i=1,\dots,N$,
which implies $\|y-y'\|_\infty \le \|\xi-\eta\|_\infty$.
(1) has been obtained.
We prove (2). It follows from (1) that
\[
d_P((\pi_{\xi,R})_*\mu,(\pi_{\eta,R})_*\mu)
\le d_{KF}^\mu(\pi_{\xi,R},\pi_{\eta,R}) \le \|\xi-\eta\|_\infty.
\]
This completes the proof.
\end{proof}
\begin{defn}[Perfect set of measures]
\index{perfect set}
Let $N$ be a natural number.
A subset $\mathcal{A} \subset \mathcal{M}(N)$ is said to be \emph{perfect}
if for any $\mu \in \mathcal{A}$ and $\nu \in \mathcal{M}(N)$ with
$(\field{R}^N,\|\cdot\|_\infty,\nu) \prec (\field{R}^N,\|\cdot\|_\infty,\mu)$,
we have $\nu \in \mathcal{A}$.
A subset $\mathcal{A} \subset \mathcal{M}(N,R)$ is said to be \emph{perfect on $B^N_R$}
\index{perfect on BNR@perfect on $B^N_R$}
if for any $\mu \in \mathcal{A}$ and $\nu \in \mathcal{M}(N,R)$ with
$(B^N_R,\|\cdot\|_\infty,\nu) \prec (B^N_R,\|\cdot\|_\infty,\mu)$,
we have $\nu \in \mathcal{A}$.
\end{defn}
Note that $\mathcal{M}(X;N)$ is perfect and $\mathcal{M}(X;N,R)$
is perfect on $B^N_R$.
We see that $(\pi_R)_*\mathcal{M}(X;N) = \mathcal{M}(X;N,R)$.
\begin{lem} \label{lem:MR-half}
For any two perfect subsets $\mathcal{A},\mathcal{B} \subset \mathcal{M}(N)$
and for any real number $R > 0$, we have
\[
d_H((\pi_R)_*\mathcal{A},(\pi_R)_*\mathcal{B})
\le 2\,d_H(\mathcal{A},\mathcal{B}),
\]
where the Hausdorff distance $d_H$ is defined with respect to
the Prohorov metric $d_P$ on $\mathcal{M}(N)$
\end{lem}
\begin{proof}
Let $\varepsilon := d_H(\mathcal{A},\mathcal{B})$.
Note that the perfectness of $\mathcal{A}$ implies $(\pi_R)_*\mathcal{A} \subset \mathcal{A}$.
For any measure $\mu \in (\pi_R)_*\mathcal{A}$ there is a measure
$\nu \in \mathcal{B}$ such that $d_P(\mu,\nu) \le \varepsilon$.
We have
\begin{equation}
\label{eq:MR-half}
\nu(B_\varepsilon(B^N_R)) \ge \mu(B^N_R) - \varepsilon = 1-\varepsilon.
\end{equation}
Since $\pi_R|_{B^N_R} = \id_{B^N_R}$ and by \eqref{eq:MR-half}, we have
$d_P((\pi_R)_*\nu,\nu) \le d_{KF}^\nu(\pi_R,\id_{\field{R}^N}) \le \varepsilon$
and hence
\[
d_P(\mu,(\pi_R)_*\nu) \le d_P(\mu,\nu) + d_P(\nu,(\pi_R)_*\nu)
\le 2\varepsilon,
\]
so that $(\pi_R)_*\mathcal{A} \subset B_{2\varepsilon}((\pi_R)_*\mathcal{B})$.
Exchanging $\mathcal{A}$ and $\mathcal{B}$ yields
$(\pi_R)_*\mathcal{B} \subset B_{2\varepsilon}((\pi_R)_*\mathcal{A})$.
We thus obtain
\[
d_H((\pi_R)_*\mathcal{A},(\pi_R)_*\mathcal{B}) \le 2\varepsilon.
\]
This completes the proof.
\end{proof}
\begin{lem} \label{lem:diam-box}
Let $X$ and $Y$ be two mm-spaces and $\varepsilon > 0$ a real number.
If a real number $\delta$ satisfies $\square(X,Y) < \delta$, then
\[
\diam(Y;1-\varepsilon-\delta)
\le \diam(X;1-\varepsilon) + \delta.
\]
\end{lem}
\begin{proof}
By $\square(X,Y) < \delta$, there are two parameters
$\varphi : I \to X$, $\psi : I \to Y$, and a Borel subset
$I_0 \subset I$ such that
$\mathcal{L}^1(I_0) \ge 1-\delta$ and
\[
|\;\varphi^*d_X(s,t)-\psi^*d_Y(s,t)\;| \le \delta
\]
for any $s,t \in I_0$.
Therefore we have
\begin{align*}
&\diam(Y;1-\varepsilon-\delta)\\
&= \inf\{\;\diam A \mid A \subset Y,
\ \mu_Y(A) \ge 1-\varepsilon-\delta\;\}\\
&= \inf\{\;\diam(J,\psi^*d_Y) \mid J \subset I,
\ \mathcal{L}^1(J) \ge 1-\varepsilon-\delta\;\}\\
&\le \inf\{\;\diam(J,\psi^*d_Y) \mid J \subset I_0,
\ \mathcal{L}^1(J) \ge 1-\varepsilon-\delta\;\}\\
&\le \inf\{\;\diam(J,\varphi^*d_X) \mid J \subset I_0,
\ \mathcal{L}^1(J) \ge 1-\varepsilon-\delta\;\} + \delta\\
&\le \inf\{\;\diam(J_1\cup J_2,\varphi^*d_X) \mid J_1 \subset I_0,
\ J_2 \subset I \setminus I_0,\ \mathcal{L}^1(J_1) \ge 1-\varepsilon-\delta\;\}
+ \delta\\
&\le \inf\{\;\diam(J_1\cup J_2,\varphi^*d_X) \mid J_1 \subset I_0,
\ J_2 \subset I \setminus I_0,\ \mathcal{L}^1(J_1\cup J_2) \ge 1-\varepsilon\;\}
+ \delta\\
&= \diam(X;1-\varepsilon) + \delta.
\end{align*}
This completes the proof.
\end{proof}
\begin{lem} \label{lem:MR}
Let $\mathcal{A}$ and $\mathcal{A}_n$, $n=1,2,\dots$, be
perfect subsets of $\mathcal{M}(N)$ such that
\[
\sup_{\mu \in \mathcal{A}} \diam(\mu;1-\kappa) < +\infty
\]
for any real number $\kappa$ with $0 < \kappa < 1$,
where $\diam$ is defined for the $l_\infty$ norm on $\field{R}^N$.
Then, the following {\rm(1)} and {\rm(2)} are equivalent
to each other.
\begin{enumerate}
\item $\mathcal{A}_n$ Hausdorff converges to $\mathcal{A}$
as $n \to \infty$.
\item $(\pi_R)_*\mathcal{A}_n$ Hausdorff converges to $(\pi_R)_*\mathcal{A}$
as $n \to \infty$ for any real number $R > 0$.
\end{enumerate}
\end{lem}
\begin{proof}
`(1) $\implies$ (2)' follows from Lemma \ref{lem:MR-half}.
We prove `(2) $\implies$ (1)'.
Take any $\varepsilon > 0$ and fix it.
Let us first prove that
\begin{equation}
\label{eq:MR1}
\mathcal{A} \subset B_\varepsilon(\mathcal{A}_n)
\end{equation}
for every sufficiently large $n$.
For any $\mu \in \mathcal{A}$ there is a number $R > 0$
such that $d_P((\pi_R)_*\mu,\mu) < \varepsilon/2$.
(2) proves that $(\pi_R)_*\mu \in B_{\varepsilon/2}((\pi_R)_*\mathcal{A}_n)
\subset B_{\varepsilon/2}(\mathcal{A}_n)$
for every sufficiently large $n$.
Therefore, $\mu$ belongs to $B_\varepsilon(\mathcal{A}_n)$
for every sufficiently large $n$, which implies \eqref{eq:MR1}.
Let $0 < \varepsilon < 1$.
It suffices to prove that
$\mathcal{A}_n \subset B_{3\varepsilon}(\mathcal{A})$
if $n$ is large enough.
We take any sequence $\mu_n \in \mathcal{A}_n$, $n=1,2,\dots$.
Set
\[
R := \max\{\sup_{\mu \in \mathcal{A}} \diam(\mu;1-\varepsilon),1\}
\quad\text{and}\quad
R' := 100R.
\]
It follows from (2) that
\begin{align} \label{eq:MR2}
d_H((\pi_{R'})_*\mathcal{A}_n,(\pi_{R'})_*\mathcal{A}) < \varepsilon/2
\end{align}
for every $n$ large enough.
From now on we assume $n$ to be sufficiently large.
Let us prove the following claim.
\begin{clm}
We have
\begin{align} \label{eq:MR3}
\diam(\mu_n;1-2\varepsilon) \le R + \varepsilon.
\end{align}
\end{clm}
\begin{proof}
Let $x \in \field{R}^N$ be any point and
$\iota_x : \field{R}^N \to \field{R}^N$
the translation defined by $\iota_x(y) = y-x$, $y \in \field{R}^N$.
The perfectness of $\mathcal{A}_n$ proves
$(\iota_x)_*(\pi_{x,R'})_*\mu_n \in (\pi_{R'})_*\mathcal{A}_n$.
By \eqref{eq:MR2},
there is a measure $\nu_n \in (\pi_{R'})_*\mathcal{A}$ such that
$d_P((\iota_x)_*(\pi_{x,R'})_*\mu_n,\nu_n) < \varepsilon/2$.
Since $\iota_x$ is an isometry,
\begin{align*}
&\square((\field{R}^N,\|\cdot\|_\infty,(\pi_{x,R'})_*\mu_n),
(\field{R}^N,\|\cdot\|_\infty,\nu_n))\\
&\le 2\,d_P((\iota_x)_*(\pi_{x,R'})_*\mu_n,\nu_n) < \varepsilon.
\end{align*}
Applying Lemma \ref{lem:diam-box} yields
\[
\diam((\pi_{x,R'})_*\mu_n;1-2\varepsilon)
< \diam(\nu_n;1-\varepsilon) + \varepsilon \le R + \varepsilon.
\]
Therefore, there is a Borel subset $A_x \subset B^N_{R'}(x)$
such that
$(\pi_{x,R'})_*\mu_n(A_x) \ge 1-2\varepsilon$ and
$\diam(A_x,\|\cdot\|_\infty) \le R+\varepsilon$.
If there is a point $x_0 \in \field{R}^N$ such that $A_{x_0}$ belongs to
the interior $U^N_{R'}(x_0)$ of $B^N_{R'}(x_0)$, then we have \eqref{eq:MR3}.
We are going to prove the existence of such a point $x_0$.
Suppose that we have no such point $x_0$.
Then, since $\diam(A_x,\|\cdot\|_\infty) \le R+\varepsilon < 2R$,
the set $A_x$ does not intersect $U^N_{98R}(x)$ and
the Euclidean distance between $x$ and $A_x$ is not less than $98R$.
Let $a_x \in \field{R}^N$ be the center of mass of $A_x$
with respect to the $N$-dimensional Lebesgue measure.
We see that $\|a_x-x\|_2 \ge 96R$, where $\|\cdot\|_2$ denotes
the Euclidean or $l_2$ norm on $\field{R}^N$.
Let
\[
V_x := \frac{1}{\|a_x-x\|_2}(a_x-x).
\]
Then, $V$ is a (not necessarily continuous) unit vector field on $\field{R}^N$.
The continuity of the map $\field{R}^N \ni x \mapsto (\pi_{x,R'})_*\mu_n$
(see Lemma \ref{lem:npp}) proves that
if two points $x, y \in \field{R}^N$ are close enough to each other,
then $d_{\field{R}^N}(A_x,A_y)$ is small enough and so
the angle between $V_x$ and $V_y$ is less than $\pi/4$.
There is a compact subset $K \subset \field{R}^N$ such that
$\mu_n(K) > 2\varepsilon$.
The set $A_x$ intersects $\pi_{x,R'}(K)$
and, if $\|x\|_2$ is sufficiently large, then
$\pi_{x,R'}(K)$ is contained in
an $(N-1)$-dimensional face of $\partial B^N_{x,R'}$
containing $\pi_{x,R'}(o)$.
Therefore, $a_x$ belongs to the $4R$-neighborhood of the face
if $\|x\|_2$ is large enough.
This proves that
\[
\lim_{\|x\|_2 \to +\infty} \angle\left(V_x,-\frac{x}{\|x\|_2}\right)
\le \frac{\pi}{3},
\]
where $\angle(\cdot,\cdot)$ denotes the angle.
From a standard mollifier argument,
we find a $C^\infty$ unit vector filed $\tilde V$ on $\field{R}^N$
and a large number $C > 0$
in such a way that,
if $\|x\|_2 \ge C$, then the angle between $\tilde{V}_x$ and
$-\frac{x}{\|x\|_2}$ is less than $\pi/2$.
Applying the Poincar\'e-Hopf theorem to the vector field $\tilde{V}$
on the Euclidean ball $\{\,x \in \field{R}^n \mid \|x\|_2 \le C\,\}$,
we have a contradiction.
The claim follows.
\end{proof}
By \eqref{eq:MR3}, there is a point $x_n \in \field{R}^N$ such that
$\mu_n(B^N_{2R}(x_n)) \ge 1-2\varepsilon$.
For the translation $\iota_{x_n} : \field{R}^N \to \field{R}^N$, we have
\[
d_P((f_n)_*\mu_n,(\pi_{R'})_*(f_n)_*\mu_n) \le d_{KF}^{(f_n)_*\mu_n}(\pi_{R'},\id_{\field{R}^N})
\le 2\varepsilon.
\]
By \eqref{eq:MR2}, there is a measure $\nu_n \in (\pi_{R'})_*\mathcal{A}$ such that
\[
d_P((\pi_{R'})_*(f_n)_*\mu_n,\nu_n) < \varepsilon.
\]
A triangle inequality proves
\[
d_P(\mu_n,(f_n^{-1})_*\nu_n) = d_P((f_n)_*\mu_n,\nu_n) < 3\varepsilon.
\]
Since $(f_n^{-1})_*\nu_n \in \mathcal{A}$, we have
$\mu_n \in B_{3\varepsilon}(\mathcal{A})$.
Therefore we obtain $\mathcal{A}_n \subset B_{3\varepsilon}(\mathcal{A})$.
This completes the proof of the lemma.
\end{proof}
\chapter{The space of pyramids}
\label{chap:pyramid}
\section{Tail and pyramid}
\begin{defn}[Tail]
\index{tail} \index{TXn@$\mathcal{T}\{X_n\}$}
Let $\{X_n\}_{n=1}^\infty$ be a sequence of mm-spaces.
The \emph{tail $\mathcal{T}\{X_n\}$ of $\{X_n\}$} is defined to be
the set of mm-spaces that are the $\square$-limits of
mm-spaces $Y_n$, $n=1,2,\dots$, such that
each $Y_n$ is dominated by $X_n$.
\end{defn}
If $\{X_{n_i}\}_{i=1}^\infty$ is a subsequence of $\{X_n\}_{n=1}^\infty$,
then $\mathcal{T}\{X_{n_i}\} \supset \mathcal{T}\{X_n\}$.
\begin{prop} \label{prop:conc-tail}
Let $\{X_n\}_{n=1}^\infty$ be a sequence of mm-spaces.
\begin{enumerate}
\item If $X_n$ concentrates to $X$ as $n\to\infty$,
then $X$ is a maximal element of $\mathcal{T}\{X_n\}$,
i.e., $X \in \mathcal{T}\{X_n\}$ and
$Y \prec X$ for any $Y \in \mathcal{T}\{X_n\}$.
\item If $X$ is a maximal element of $\mathcal{T}\{X_n\}$
and if $\mathcal{T}\{X_{n_i}\} = \mathcal{T}\{X_n\}$
for any subsequence $\{X_{n_i}\}$ of $\{X_n\}$,
then $X_n$ concentrates to $X$.
\end{enumerate}
\end{prop}
\begin{proof}
We prove (1). Assume that $X_n$ concentrates to $X$ as $n\to\infty$.
By Corollary \ref{cor:XN},
there are measures $\underline{\mu}_N \in \mathcal{M}(X;N)$, $N=1,2,\dots$,
such that
$\underline{X}_N = (\field{R}^N,\|\cdot\|_\infty,\underline{\mu}_N)$
$\square$-converges to $X$ as $N\to\infty$.
Since Theorem \ref{thm:A} implies that $\mathcal{M}(X_n;N)$ Hausdorff
converges to $\mathcal{M}(X;N)$, we find $1$-Lipschitz maps
$F_{N,n} : X_n \to (\field{R}^N,\|\cdot\|_\infty)$, $n=1,2,\dots$, for each $N$
such that $\lim_{n\to\infty} d_P((F_{N,n})_*\mu_{X_n},\underline{\mu}_N) = 0$.
The mm-space $X_{N,n} := (\field{R}^N,\|\cdot\|_\infty,(F_{N,n})_*\mu_{X_n})$
is dominated by $X_n$ and $\square$-converges to
$\underline{X}_N = (\field{R}^N,\|\cdot\|_\infty,\underline{\mu}_N)$ as $n\to\infty$.
Therefore, there is a monotone nondecreasing function $m : \field{N} \to \field{N}$
such that $m(N) \to \infty$ as $N \to \infty$ and
$\square(X_{N,n},\underline{X}_N) < 1/N$ for any
$n,N \in \field{N}$ with $n \ge m(N)$.
Setting $M(n) := \max\{\;j \in \field{N} \mid m(j) \le n\;\}$ for $n \in \field{N}$,
we see that $M(n)$ is monotone nondecreasing in $n$ and
$M(n) \to \infty$ as $n \to \infty$.
By $m(M(n)) \le n$,
\[
\square(X_{M(n),n},\underline{X}_{M(n)}) < 1/M(n) \to 0 \ \text{as}\ n\to\infty
\]
and hence
\[
\square(X_{M(n),n},X) \le \square(X_{M(n),n},\underline{X}_{M(n)})
+ \square(\underline{X}_{M(n)},X) \to 0 \ \text{as}\ n \to \infty,
\]
which proves $X \in \mathcal{T}\{X_n\}$.
Let us prove the maximality of $X$ in $\mathcal{T}\{X_n\}$.
Take any mm-space $Y \in \mathcal{T}\{X_n\}$.
There is a sequence of mm-spaces $Y_n$, $n=1,2,\dots$,
with $Y_n \prec X_n$ that $\square$-converges to $Y$ as $n\to\infty$.
By Lemma \ref{lem:dom-M}, $\mathcal{M}(Y_n;N) \subset \mathcal{M}(X_n;N)$.
Theorem \ref{thm:A} implies that
$\mathcal{M}(X_n;N)$ and $\mathcal{M}(Y_n;N)$ Hausdorff converges
to $\mathcal{M}(X;N)$ and $\mathcal{M}(Y;N)$, respectively.
Therefore we have $\mathcal{M}(Y;N) \subset \mathcal{M}(X;N)$ for any $N$,
which together with Lemma \ref{lem:dom-M} implies that
$Y$ is dominated by $X$.
We thus obtain the maximality of $X$ in $\mathcal{T}\{X_n\}$.
(1) has been proved.
We prove (2).
Since $X \in \mathcal{T}\{X_n\}$, we find a sequence of mm-spaces $Y_n$,
$n=1,2,\dots$, with $Y_n \prec X_n$ that $\square$-converges to $X$.
By Proposition \ref{prop:dconc-box},
$Y_n$ concentrates to $X$ as $n\to\infty$,
so that, by Theorem \ref{thm:A} and Lemma \ref{lem:MR},
$\mathcal{M}(Y_n;N,R)$ Hausdorff converges to $\mathcal{M}(X;N,R)$ as $n\to\infty$
for any $N$ and $R > 0$.
We see $\mathcal{M}(Y_n;N,R) \subset \mathcal{M}(X_n;N,R)$ by Lemma \ref{lem:dom-M}.
Theorem \ref{thm:A} and Lemma \ref{lem:MR} together tell us that,
to prove the concentration of $X_n$ to $X$,
it suffices to show that $\mathcal{M}(X_n;N,R)$ Hausdorff converges to $\mathcal{M}(X;N,R)$.
Suppose that $\mathcal{M}(X_n;N,R)$ does not Hausdorff converge to $\mathcal{M}(X;N,R)$
for some $N$ and $R > 0$.
Then, there are a number $\delta > 0$,
a sequence of natural numbers $n_i \to \infty$,
and measures $\mu_{n_i} \in \mathcal{M}(X_{n_i};N,R)$, $i=1,2,\dots$,
such that $d_P(\mu_{n_i},\mathcal{M}(X;N,R)) \ge \delta$ for any $i$.
Replacing $\{\mu_{n_i}\}$ with a subsequence
we assume that $\mu_{n_i}$ converges weakly to some measure $\mu_\infty$
on $B^N_R$. The mm-space $(\field{R}^N,\|\cdot\|_\infty,\mu_{n_i})$
is dominated by $X_{n_i}$ and $\square$-converges to
$(\field{R}^N,\|\cdot\|_\infty,\mu_\infty)$, so that
$(\field{R}^N,\|\cdot\|_\infty,\mu_\infty)$ belongs to the tail
$\mathcal{T}\{X_{n_i}\} = \mathcal{T}\{X_n\}$.
The maximality of $X$ in $\mathcal{T}\{X_n\}$ yields
that $X$ dominates $(\field{R}^N,\|\cdot\|_\infty,\mu_\infty)$
and so $\mu_\infty$ is an element of $\mathcal{M}(X;N,R)$,
which is a contradiction.
This completes the proof.
\end{proof}
\begin{defn}[Pyramid] \label{defn:pyramid}
\index{pyramid}
Recall that $\mathcal{X}$ is the mm-isomorphism classes of mm-spaces.
A subset $\mathcal{P} \subset \mathcal{X}$ is called a \emph{pyramid}
if it satisfies the following (1), (2), and (3).
\begin{enumerate}
\item If $X \in \mathcal{P}$ and if $Y \prec X$, then $Y \in \mathcal{P}$.
\item For any two mm-spaces $X, X' \in \mathcal{P}$,
there exists an mm-space $Y \in \mathcal{P}$ such that
$X \prec Y$ and $X' \prec Y$.
\item $\mathcal{P}$ is nonempty and $\square$-closed.
\end{enumerate}
(2) is called the Moore-Smith property and a pyramid is a directed subfamily of $\mathcal{X}$.
We denote the set of pyramids by $\Pi$.
\index{Pi@$\Pi$}
For an mm-space $X$ we define
\[
\mathcal{P}_X := \{\;X' \in \mathcal{X} \mid X' \prec X\;\}.
\]
\index{PX@$\mathcal{P}_X$}
It follows from Theorem \ref{thm:dom} that $\mathcal{P}_X$ is a pyramid.
We call $\mathcal{P}_X$ the \emph{pyramid associated with $X$}.
\index{pyramid associated with X@pyramid associated with $X$}
\end{defn}
We see that $X \prec Y$ if and only if $\mathcal{P}_X \subset \mathcal{P}_Y$.
Note that for any two mm-spaces $X$ and $X'$ we always have
an mm-space dominating both $X$ and $X'$, in fact
$X \times X'$ with product measure and $l_p$ product metric,
$1 \le p \le +\infty$,
is an example of such an mm-space.
It is trivial that $\mathcal{X}$ itself is a pyramid.
In Gromov's book \cite{Gromov}, the definition of a pyramid
is only by (1) and (2) of Definition \ref{defn:pyramid}.
We here put (3) as an additional condition for later convenience.
\section{Weak Hausdorff convergence}
All the discussions in this section also work in the case where
$\mathcal{X}$ is a general separable metric space.
\begin{defn}[Weak (Hausdorff) convergence] \label{defn:w-conv}
\index{weak convergence} \index{weak Hausdorff convergence}
\index{converges weakly}
Denote by $\mathcal{F}(\mathcal{X})$ the set of $\square$-closed subsets of $\mathcal{X}$.
Let $\mathcal{Y}_n, \mathcal{Y} \in \mathcal{F}(\mathcal{X})$
(where $\mathcal{Y}_n$ are $\mathcal{Y}$ may be empty).
We say that \emph{$\mathcal{Y}_n$ converges weakly to $\mathcal{Y}$}
as $n\to\infty$
if the following (1) and (2) are both satisfied.
\begin{enumerate}
\item For any mm-space $X \in \mathcal{Y}$, we have
\[
\lim_{n\to\infty} \square(X,\mathcal{Y}_n) = 0.
\]
\item For any mm-space $X \in \mathcal{X} \setminus \mathcal{Y}$, we have
\[
\liminf_{n\to\infty} \square(X,\mathcal{Y}_n) > 0.
\]
\end{enumerate}
We here agree that $\square(X,\emptyset) = +\infty$.
\end{defn}
\begin{lem} \label{lem:w-subconv}
Any sequence $\{\mathcal{Y}_n\}_{n=1}^\infty \subset \mathcal{F}(\mathcal{X})$
has a weakly convergent subsequence.
\end{lem}
\begin{proof}
Let $\{\mathcal{Y}_n\}_{n=1}^\infty \subset \mathcal{F}(\mathcal{X})$ be a given sequence.
We first prove the following
\begin{clm}
There exists a subsequence $\{\mathcal{Y}_{n_i}\}$ of $\{\mathcal{Y}_n\}$
such that, for any mm-space $X \in \mathcal{X}$, the limit
\[
\lim_{i\to\infty} \square(X,\mathcal{Y}_{n_i}) \in [\,0,+\infty\,]
\]
exists.
\end{clm}
\begin{proof}
According to Proposition \ref{prop:separable},
we find a dense countable subset $\{X_k\}_{k=1}^\infty \subset \mathcal{X}$.
There is a subsequence $\{\mathcal{Y}_{n^1_i}\}$ of
$\{\mathcal{Y}_n\}$ such that
the limit $\lim_{i\to\infty} \square(X_1,\mathcal{Y}_{n^1_i}) \in [\,0,+\infty\,]$
exists.
There is also a subsequence $\{\mathcal{Y}_{n^2_i}\}$ of
$\{\mathcal{Y}_{n^1_i}\}$ such that
the limit $\lim_{i\to\infty} \square(X_2,\mathcal{Y}_{n^2_i}) \in [\,0,+\infty\,]$
exists.
Repeating this procedure we have the limits
$\lim_{i\to\infty} \square(X_j,\mathcal{Y}_{n^j_i}) \in [\,0,+\infty\,]$
for all $j=1,2,3,\dots$.
By a diagonal argument, we are able to choose a subsequence
$\{\mathcal{Y}_{n_i}\}$
of $\{\mathcal{Y}_n\}$ in such a way that the limit
$\lim_{i\to\infty} \square(X_j,\mathcal{Y}_{n_i}) \in [\,0,+\infty\,]$
exists for all $j=1,2,\dots$.
If
$\lim_{i\to\infty} \square(X_{j_0},\mathcal{Y}_{n_i}) = +\infty$ for some $j_0$,
then a triangle inequality proves that
$\lim_{i\to\infty} \square(X,\mathcal{Y}_{n_i}) = +\infty$
for any mm-space $X \in \mathcal{X}$.
Assume that, for any $j$, the limit
\[
r_j := \lim_{i\to\infty} \square(X_j,\mathcal{Y}_{n_i})
\]
is finite.
Let $X$ be any mm-space.
Since $\{X_k\}_{k=1}^\infty$ is $\square$-dense in $\mathcal{X}$,
there is a natural number $j_k$ for any $k$
such that $\square(X,X_{j_k}) < 1/k$.
By a triangle inequality,
\[
|\;\square(X,\mathcal{Y}_{n_i}) - \square(X_{j_k},\mathcal{Y}_{n_i})\;|
\le \square(X,X_{j_k}) < 1/k
\]
and therefore
\[
r_{j_k}-1/k \le \liminf_{i\to\infty} \square(X,\mathcal{Y}_{n_i})
\le \limsup_{i\to\infty} \square(X,\mathcal{Y}_{n_i}) \le r_{j_k}+1/k,
\]
which proves the existence of the limit
$\lim_{i\to\infty} \square(X,\mathcal{Y}_{n_i})$.
The claim has been proved.
\end{proof}
Let $\mathcal{Y}$ be the set of mm-spaces $X$
satisfying $\lim_{i\to\infty}\square(X,\mathcal{Y}_{n_i}) = 0$.
\begin{clm}
$\mathcal{Y}$ is $\square$-closed.
\end{clm}
\begin{proof}
Assume that a sequence of mm-spaces $X_j \in \mathcal{Y}$
$\square$-converges to an mm-space $X$.
A triangle inequality implies
\[
\square(X,\mathcal{Y}_{n_i})
\le \square(X,X_j) + \square(X_j,\mathcal{Y}_{n_i}),
\]
where we have $\lim_{j\to\infty} \square(X,X_j) = 0$ and
$\lim_{i\to\infty} \square(X_j,\mathcal{Y}_{n_i})
= 0$ because $X_j \in \mathcal{Y}$.
Therefore, $\lim_{i\to\infty} \square(X,\mathcal{Y}_{n_i}) = 0$
and $X \in \mathcal{Y}$.
We have the claim.
\end{proof}
From the definition of $\mathcal{Y}$,
it is easy to prove
the weak convergence of $\mathcal{Y}_{n_i}$ to $\mathcal{Y}$.
This completes the proof of Lemma \ref{lem:w-subconv}.
\end{proof}
\begin{prop}[Down-to-earth criterion for weak convergence] \label{prop:w-conv}
\index{down-to-earth criterion for weak convergence}
\ \\
For given $\mathcal{Y}_n,\mathcal{Y} \in \mathcal{F}(\mathcal{X})$, $n=1,2,\dots$,
the following {\rm(1)} and {\rm(2)} are equivalent
to each other.
\begin{enumerate}
\item $\mathcal{Y}_n$ converges weakly to $\mathcal{Y}$.
\item Let $\underline{\mathcal{Y}}_\infty$ be the set of the limits
of convergent sequences $Y_n \in \mathcal{Y}_n$,
and $\overline{\mathcal{Y}}_\infty$ the set of the limits of
convergent subsequences of $Y_n \in \mathcal{Y}_n$.
Then we have
\[
\mathcal{Y} = \underline{\mathcal{Y}}_\infty = \overline{\mathcal{Y}}_\infty.
\]
\end{enumerate}
\end{prop}
Note that we have
$\underline{\mathcal{Y}}_\infty \subset \overline{\mathcal{Y}}_\infty$ in general.
\begin{proof}
We prove `(1) $\implies$ (2)'.
Assume that $\mathcal{Y}_n$ converges weakly to $\mathcal{Y}$.
Let us first prove $\mathcal{Y} \subset \underline{\mathcal{Y}}_\infty$.
We take any mm-space $X \in \mathcal{Y}$.
From Definition \ref{defn:w-conv}(1),
we have $\lim_{n\to\infty} \square(X,\mathcal{Y}_n) = 0$.
There is a sequence of mm-spaces
$X_n \in \mathcal{Y}_n$ that $\square$-converges to $X$,
i.e., $X \in \underline{\mathcal{Y}}_\infty$.
Thus we obtain $\mathcal{Y} \subset \underline{\mathcal{Y}}_\infty$.
Let us next prove $\overline{\mathcal{Y}}_\infty \subset \mathcal{Y}$.
Take any mm-space $X \in \overline{\mathcal{Y}}_\infty$.
There is a sequence of mm-spaces $X_i \in \mathcal{Y}_{n_i}$
with $n_i \to \infty$ that $\square$-converges to $X$.
If $X$ does not belong to $\mathcal{Y}$, then,
by Definition \ref{defn:w-conv}(2), we have
$\liminf_{n\to\infty} \square(X,\mathcal{Y}_n) > 0$,
which contradicts that $\mathcal{Y}_{n_i} \ni X_i \overset{\square}{\to} X$
as $i\to\infty$.
Thus, $X$ belongs to $\mathcal{Y}$ and
we have $\overline{\mathcal{Y}}_\infty \subset \mathcal{Y}$.
We obtain (2).
We prove `(2) $\implies$ (1)'.
Assume that
$\mathcal{Y} = \underline{\mathcal{Y}}_\infty = \overline{\mathcal{Y}}_\infty$.
Let us verify Definition \ref{defn:w-conv}(1).
Take any mm-space $X \in \mathcal{Y}$.
Since $X \in \underline{\mathcal{Y}}_\infty$,
there is a sequence of mm-spaces $X_n \in \mathcal{Y}_n$, $n=1,2,\dots$,
that $\square$-converges to $X$.
Therefore,
\[
\limsup_{n\to\infty} \square(X,\mathcal{Y}_n)
\le \lim_{n\to\infty} \square(X,X_n) = 0.
\]
Let us verify Definition \ref{defn:w-conv}(2).
Suppose that $\liminf_{n\to\infty} \square(X,\mathcal{Y}_n) = 0$
for an mm-space $X$.
It suffices to prove that $X$ belongs to $\mathcal{Y}$.
We find a subsequence $\{\mathcal{Y}_{n_i}\}$ of $\{\mathcal{Y}_n\}$
in such a way that $\lim_{i\to\infty} \square(X,\mathcal{Y}_{n_i}) = 0$.
There is an mm-space $X_i \in \mathcal{Y}_{n_i}$ for each $i$
such that $X_i$ $\square$-converges to $X$ as $i\to\infty$.
Therefore, $X$ belongs to $\overline{\mathcal{Y}}_\infty = \mathcal{Y}$.
This completes the proof.
\end{proof}
\section{Weak convergence of pyramids}
To show that the weak limit of pyramids is also a pyramid,
we prove the following lemma.
\begin{lem} \label{lem:lim-pyramid}
\begin{enumerate}
\item If a sequence of mm-spaces $X_n$, $n=1,2,\dots$, $\square$-converges
to an mm-space $X$ and if $X$ dominates an mm-space $Y$,
then there exists a sequence of mm-spaces $Y_n$ $\square$-converging
to $Y$ such that $X_n$ dominates $Y_n$ for each $n$.
\item If two sequences of mm-spaces $X_n$ and $Y_n$, $n=1,2,\dots$,
$\square$-converge to $X$ and $Y$, respectively, and if $X_n$ and $Y_n$
are both dominated by an mm-space $\tilde{Z_n}$ for each $n$,
then there exists a sequence of mm-spaces $Z_n$
such that $X_n,Y_n \prec Z_n \prec \tilde{Z_n}$
and $\{Z_n\}$ has a $\square$-convergent subsequence.
\end{enumerate}
\end{lem}
Note that the limit of the $\square$-convergent subsequence of $\{Z_n\}$
dominates $X$ and $Y$ by Theorem \ref{thm:dom}.
\begin{proof}
We prove (1).
By Corollary \ref{cor:XN}, $Y$ is approximated by
$\underline{Y}_N = (\field{R}^N,\|\cdot\|_\infty,\underline{\mu}_N)$,
where $\Phi_N : Y \to (\field{R}^N,\|\cdot\|_\infty)$ is a $1$-Lipschitz map
and $\underline{\mu}_N := (\Phi_N)_*\mu_Y$.
Since $X \succ Y$, we find a $1$-Lipschitz map $F : X \to Y$
with $F_*\mu_X = \mu_Y$.
There is an $\varepsilon_n$-mm-isomorphism $p_n : X_n \to X$
with $\varepsilon_n \to 0$.
Then the composition $f_n := \Phi_N \circ F \circ p_n
: X_n \to \field{R}^N$
is $1$-Lipschitz up to $\varepsilon_n$.
Applying Lemma \ref{lem:Lip-approx} we find
a $1$-Lipschitz map $\tilde{f}_n : X_n \to (\field{R}^N,\|\cdot\|_\infty)$
in such a way that
\[
\dKF(\tilde{f}_n,f_n) \le \varepsilon_n.
\]
By Corollary \ref{cor:push-di-Lip},
\begin{align*}
d_P((f_n)_*\mu_{X_n},\underline{\mu}_N)
&= d_P((\Phi_N\circ F)_*(p_n)_*\mu_{X_n},(\Phi_N\circ F)_*\mu_X)\\
&\le d_P((p_n)_*\mu_{X_n},\mu_X) \le \varepsilon_n.
\end{align*}
Note that this holds for all $N$ and $n$.
Take a sequence of natural numbers $N_n$, $n=1,2,\dots$,
divergent to infinity. We see
\[
\lim_{n\to\infty} d_P((f_n)_*\mu_{X_n},\underline{\mu}_{N_n}) = 0.
\]
Letting $Y_n' := (\field{R}^{N_n},\|\cdot\|_\infty,(f_n)_*\mu_{X_n})$,
we have, by Proposition \ref{prop:box-di},
\[
\lim_{n\to\infty} \square(Y_n',\underline{Y}_{N_n}) = 0,
\]
which together with the convergence
$\underline{Y}_{N_n} \overset{\square}{\to} Y$ implies that
$Y_n'$ $\square$-converges to $Y$ as $n\to\infty$.
By setting $Y_n := (\field{R}^N,\|\cdot\|_\infty,(\tilde{f}_n)_*\mu_{X_n})$,
it follows from Proposition \ref{prop:box-di} and Lemma \ref{lem:di-me}
that
\[
\square(Y_n,Y_n') \le 2\,d_P((\tilde{f}_n)_*\mu_{X_n},(f_n)_*\mu_{X_n}) \le
2\dKF(\tilde{f}_n,f_n) \le 2\varepsilon_n,
\]
so that $Y_n$ $\square$-converges to $Y$.
Since $\tilde{f}_n : X_n \to Y_n$ is $1$-Lipschitz,
we have $Y_n \prec X_n$.
(1) has been proved.
We prove (2).
Since $X_n,Y_n \prec \tilde{Z}_n$,
we find $1$-Lipschitz maps $\tilde{f}_n : \tilde{Z}_n \to X_n$,
$\tilde{g}_n : \tilde{Z}_n \to Y_n$ such that
$(\tilde{f}_n)_*\mu_{\tilde{Z}_n} = \mu_{X_n}$ and
$(\tilde{g}_n)_*\mu_{\tilde{Z}_n} = \mu_{Y_n}$.
We set, for $x,y \in \tilde{Z}_n$,
\[
d_n(x,y) := \max\{\;d_{X_n}(\tilde{f}_n(x),\tilde{f}_n(y)),
d_{Y_n}(\tilde{g}_n(x),\tilde{g}_n(y))\;\}.
\]
It is easy to see that $d_n$ is a pseudo-metric on $\tilde{Z}_n$.
Let $\hat{Z}_n$ be the quotient space of $\tilde{Z}_n$
modulo $d_n=0$.
Then, $d_n$ induces a metric on $\hat{Z}_n$.
Let $(Z_n,d_{Z_n})$ be the completion of $\hat{Z}_n$.
Note that $\hat{Z}_n$ is naturally embedded in $Z_n$.
If $d_n(x,y) = 0$ for two points $x,y \in \tilde{Z}_n$,
then $\tilde{f}_n(x) = \tilde{f}_n(y)$ and $\tilde{g}_n(x) = \tilde{g}_n(y)$.
Therefore, setting $f_n([x]) := \tilde{f}_n(x)$
and $g_n([x]) := \tilde{g}_n(x)$ for an equivalence class $[x] \in \hat{Z}_n$,
we obtain two maps $f_n : \hat{Z}_n \to X_n$ and $g_n : \hat{Z}_n \to Y_n$
for each $n$.
It follows from the $1$-Lipschitz continuity of
$\tilde{f}_n$ and $\tilde{g}_n$ that $f_n$ and $g_n$ are both $1$-Lipschitz
continuous, so that both of them extend to $1$-Lipschitz
maps $f_n : Z_n \to X_n$ and $g_n : Z_n \to Y_n$.
Let $\pi_n : \tilde{Z}_n \to Z_n$ be the natural projection
and let $\mu_{Z_n} := (\pi_n)_*\mu_{\tilde{Z}_n}$.
Then, $Z_n = (Z_n,d_{Z_n},\mu_{Z_n})$ is an mm-space dominated by $\tilde{Z}_n$.
Since
\[
(f_n)_*\mu_{Z_n} = (\tilde{f}_n)_*\mu_{\tilde{Z}_n} = \mu_{X_n}
\quad\text{and}\quad
(g_n)_*\mu_{Z_n} = (\tilde{g}_n)_*\mu_{\tilde{Z}_n} = \mu_{Y_n},
\]
we have $X_n, Y_n \prec Z_n$.
The $\square$-precompactness of $\{X_n\}$ and $\{Y_n\}$ together with
Lemma \ref{lem:precpt} implies that for any $\varepsilon > 0$
there is a number $\Delta(\varepsilon)$ such that
for each $n$ we find Borel subsets $K_{nj} \subset X_n$,
$j=1,\dots,N$, and $K_{nj}' \subset Y_n$, $j=1,\dots,N'$,
with $N,N' \le \Delta(\varepsilon)$ in such a way that
\begin{align*}
\diam K_{nj}, \; &\diam K_{nj}' \le \varepsilon,\\
\diam \bigcup_{j=1}^N K_{nj},
\; &\diam \bigcup_{j=1}^{N'} K_{nj}' \le \Delta(\varepsilon),\\
\mu_{X_n}\left(X_n \setminus \bigcup_{j=1}^N K_{nj}\right),
\; &\mu_{Y_n}\left(Y_n \setminus \bigcup_{j=1}^{N'} K_{nj}\right) \le \varepsilon.
\end{align*}
Set $K_{njk} := f_n^{-1}(K_{nj}) \cap g_n^{-1}(K_{nk}')$
for $j=1,\dots,N$ and $k=1,\dots,N'$.
To prove the $\square$-precompactness of $\{Z_n\}$,
we are going to verify (4) of Lemma \ref{lem:precpt}.
Let us prove that $\diam K_{njk} \le \varepsilon$.
In fact, if we take two points $x,y \in K_{njk}$, then
since $f_n(x),f_n(y) \in K_{nj}$, $g_n(x),g_n(y) \in K_{nk}'$,
we have $d_{X_n}(f_n(x),f_n(y)), d_{Y_n}(g_n(x),g_n(y)) \le \varepsilon$
and so $d_{Z_n}(x,y) \le \varepsilon$.
Let us prove that
$\diam \bigcup_{1\le j\le N, 1\le k\le N'}K_{njk} \le \Delta(\varepsilon)$.
For any two points $x,y \in \bigcup_{1\le j\le N, 1\le k\le N'}K_{njk}$,
there are four numbers $j(x)$, $k(x)$, $j(y)$, and $k(y)$ such that
$x \in K_{nj(x)k(x)}$ and $y \in K_{nj(y)k(y)}$.
Since $f_n(x) \in K_{nj(x)}$, $g_n(x) \in K_{nk(x)}'$,
$f_n(y) \in K_{nj(y)}$, and $g_n(y) \in K_{nk(y)}'$, we have
\begin{align*}
d_{X_n}(f_n(x),f_n(y)) &\le \diam \bigcup_{j=1}^N K_{nj}
\le \Delta(\varepsilon),\\
d_{Y_n}(g_n(x),g_n(y)) &\le \diam \bigcup_{k=1}^N K_{nk}'
\le \Delta(\varepsilon),
\end{align*}
which implies $d_{Z_n}(x,y) \le \Delta(\varepsilon)$.
Let us prove that $\mu_{Z_n}(Z_n \setminus \bigcup_{j,k}K_{njk}) \le 2\varepsilon$.
In fact, since $Z_n \setminus \bigcup_{j,k} K_{njk}
= f_n^{-1}(X_n \setminus \bigcup_j K_{nj})
\cup g_n^{-1}(Y_n \setminus \bigcup_k K_{jk}')$,
we have
\begin{align*}
\mu_{Z_n}\left(Z_n \setminus \bigcup_{j,k}K_{njk}\right)
&\le \mu_{X_n}\left(X_n \setminus \bigcup_j K_{nj}\right)
+ \mu_{Y_n}\left(Y_n \setminus \bigcup_k K_{jk}'\right)\\
&\le 2\varepsilon.
\end{align*}
Applying Lemma \ref{lem:precpt} yields
the $\square$-precompactness of $\{Z_n\}$.
In particular, it has a $\square$-convergent subsequence.
This completes the proof.
\end{proof}
Combining Lemma \ref{lem:lim-pyramid} and
Theorem \ref{thm:dom} implies the following.
\begin{prop} \label{prop:lim-pyramid}
If a sequence of pyramids converges weakly,
then the weak limit is also a pyramid.
\end{prop}
Lemma \ref{lem:w-subconv} together with
this proposition implies the following theorem.
\begin{thm} \label{thm:cpt-pyramid}
The set $\Pi$ of pyramids is sequentially compact,
i.e., any sequence of pyramids has a subsequence that converges weakly to
a pyramid.
\end{thm}
\begin{prop} \label{prop:conc-pyramid}
For given mm-spaces $X$ and $X_n$, $n=1,2,\dots$,
the following {\rm(1)} and {\rm(2)} are equivalent to each other.
\begin{enumerate}
\item $X_n$ concentrates to $X$ as $n\to\infty$.
\item $\mathcal{P}_{X_n}$ converges weakly to $\mathcal{P}_X$ as $n\to\infty$.
\end{enumerate}
\end{prop}
\begin{proof}
We prove `(1) $\implies$ (2)'.
Assume that $X_n$ concentrates $X$ as $n\to\infty$.
Let us first prove that $\mathcal{P}_X = \mathcal{T}\{X_n\}$.
Proposition \ref{prop:conc-tail}(1) says that
$X$ is a maximal element of $\mathcal{T}\{X_n\}$, which implies
$\mathcal{T}\{X_n\} \subset \mathcal{P}_X$.
It follows from $X \in \mathcal{T}\{X_n\}$ that there is a sequence of mm-spaces
$X_n'$ with $X_n' \prec X_n$, $n=1,2,\dots$,
such that $X_n'$ $\square$-converges to $X$ as $n\to\infty$.
Let $Y \in \mathcal{P}_X$ be any mm-space.
Lemma \ref{lem:lim-pyramid}(1) tells us that
there is a sequence $Y_n$
$\square$-converging to $Y$ such that $Y_n \prec X_n' \prec X_n$.
This proves that $Y$ belongs to $\mathcal{T}\{X_n\}$.
We thus have $\mathcal{P}_X = \mathcal{T}\{X_n\}$.
We apply the above discussion to any subsequence $\{X_{n_i}\}$ of $\{X_n\}$
to obtain $\mathcal{P}_X = \mathcal{T}\{X_{n_i}\}$.
Therefore, $\mathcal{Y}_n := \mathcal{P}_{X_n}$ satisfies
$\underline{\mathcal{Y}}_\infty = \overline{\mathcal{Y}}_\infty = \mathcal{P}_X$,
so that Proposition \ref{prop:w-conv} proves
the weak convergence of $\mathcal{P}_{X_n}$ to $\mathcal{P}_X$.
We prove `(2) $\implies$ (1)'.
Assume that $\mathcal{P}_{X_n}$ converges weakly to $\mathcal{P}_X$.
Since $\mathcal{T}\{X_n\}$ is the set of the limits of mm-spaces in $\mathcal{P}_{X_n}$,
Proposition \ref{prop:w-conv} shows that,
for any subsequence $\{X_{n_i}\}$ of $\{X_n\}$,
we have $\mathcal{P}_X = \mathcal{T}\{X_n\} = \mathcal{T}\{X_{n_i}\}$.
In particular, $X$ is a maximal element of $\mathcal{T}\{X_n\}$.
By Proposition \ref{prop:conc-tail}(2),
$X_n$ concentrates to $X$.
\end{proof}
Proposition \ref{prop:conc-pyramid} means that
the map
\[
\iota : \mathcal{X} \ni X \longmapsto \mathcal{P}_X \in \Pi
\]
is a topological embedding map with respect to
the concentration topology on $\mathcal{X}$,
where the topology (metric) on the set $\Pi$ of pyramids
is introduced in the next section.
In Section \ref{sec:comp}, we prove that
the $\dconc$-completion of $\mathcal{X}$ is also embedded in $\Pi$.
\begin{cor}
Let $\{X_n\}_{n=1}^\infty$ be a L\'evy family.
Then, any mm-space $Y_n$ with $Y_n \prec X_n$
either $\square$-converges to a one-point space $*$,
or $\square$-diverges.
\end{cor}
\begin{proof}
Since $X_n$ concentrates to the one-point space $*$ as $n\to\infty$,
the associated pyramid $\mathcal{P}_{X_n}$ converges weakly to the pyramid $\{*\}$.
This proves the corollary.
\end{proof}
\begin{cor}
Any pyramid is $\dconc$-closed.
\end{cor}
\begin{proof}
Let $\mathcal{P}$ be a pyramid and
assume that a sequence of mm-spaces $X_n \in \mathcal{P}$, $n=1,2,\dots$,
concentrates to an mm-space $X$.
By Proposition \ref{prop:conc-pyramid},
$\mathcal{P}_{X_n}$ converges weakly to $\mathcal{P}_X$ as $n\to\infty$.
It follows from $\mathcal{P}_{X_n} \subset \mathcal{P}$ and the $\square$-closedness of $\mathcal{P}$
that $\mathcal{P}_X$ is contained in $\mathcal{P}$ and in particular
$X$ belongs to $\mathcal{P}$.
This completes the proof.
\end{proof}
\section{Metric on the space of pyramids}
A main purpose in this section is to introduce
a compact metric on the set of pyramids compatible with
weak convergence.
\begin{defn}[$\mathcal{X}(N,R)$]
\index{XNR@$\mathcal{X}(N,R)$}
Let $N$ be a natural number and $R$ a nonnegative extended real number,
i.e., $0 \le R \le +\infty$.
Denote by $\mathcal{M}(N,R)$ the set of $\mu \in \mathcal{M}(N)$ with
$\supp\mu \subset B^N_R$, where $B^N_\infty := \field{R}^N$.
\index{BNinfinity@$B^N_\infty$}
Note that $\mathcal{M}(N,\infty) = \mathcal{M}(N)$.
We define
\[
\mathcal{X}(N,R) := \{\;(B^N_R,\|\cdot\|_\infty,\mu) \mid \mu \in \mathcal{M}(N,R)\;\}.
\]
\end{defn}
If $R < +\infty$, then $\mathcal{X}(N,R)$ is $\square$-compact.
\begin{lem} \label{lem:X-dense}
$\bigcup_{N=1}^\infty \mathcal{X}(N,N)$ is $\square$-dense in $\mathcal{X}$.
\end{lem}
\begin{proof}
We take any mm-space $X$ and fix it.
By Corollary \ref{cor:XN}, there is a sequence of mm-spaces
$\underline{X}_N = (\field{R}^N,\|\cdot\|_\infty,\underline{\mu}_N)$, $N=1,2,\dots$,
with $\underline{\mu}_N \in \mathcal{M}(X;N)$ that $\square$-converges to $X$.
Let $\pi_R : \field{R}^N \to B^N_R$, $R > 0$, be the nearest point projection.
For $R \in (\,0,+\infty\,)$,
the push-forward measure $\underline{\mu}_{N,R} := (\pi_R)_*\underline{\mu}_N$
belongs to $\mathcal{M}(X;N,R)$ and converges weakly to $\underline{\mu}_N$
as $R \to +\infty$.
The mm-space
$\underline{X}_{N,R} := (B^N_R,\|\cdot\|_\infty,\underline{\mu}_{N,R})$
belongs to $\bigcup_{N=1}^\infty \mathcal{X}(N,N)$
and $\square$-converges to $\underline{X}_N$ as $R\to +\infty$
(see Proposition \ref{prop:box-di}).
By using a triangle inequality,
there is a sequence $R_N \to +\infty$ such that
$\underline{X}_{N,R_N}$ $\square$-converges to $X$ as $N\to\infty$.
This completes the poof.
\end{proof}
\begin{lem} \label{lem:down-XNR}
Let $N$ be a natural number and let $0 \le R \le +\infty$.
If a sequence of mm-spaces $X_n$, $n=1,2,\dots$, $\square$-converges
to $(B^N_R,\|\cdot\|_\infty,\mu)$ for a measure $\mu \in \mathcal{M}(N,R)$,
then there exists a sequence of measures $\mu_n \in \mathcal{M}(X_n;N,R)$,
$n=1,2,\dots$, converging weakly to $\mu$.
In particular, setting
$X_n' := (B^N_R,\|\cdot\|_\infty,\mu_n)$, we have
$X_n' \in \mathcal{X}(N,R)$, $X_n' \prec X_n$, and
$\lim_{n\to\infty} \square(X_n',X) = 0$.
\end{lem}
\begin{proof}
Assume that a sequence of mm-spaces $X_n$, $n=1,2,\dots$,
$\square$-converges to $(B^N_R,\|\cdot\|_\infty,\mu)$
for a measure $\mu \in \mathcal{M}(N,R)$.
Then, there is an
$\varepsilon_n$-mm-isomorphism $f_n : X_n \to X = (\field{R}^N,\|\cdot\|_\infty,\mu)$
with $\varepsilon_n \to 0$.
This satisfies $d_P((f_n)_*\mu_{X_n},\mu) \le \varepsilon_n$.
Apply Lemma \ref{lem:Lip-approx}
to obtain a $1$-Lipschitz map $\tilde{f}_n : X_n \to (\field{R}^N,\|\cdot\|_\infty)$
such that $\dKF(\tilde{f}_n,f_n) \le \varepsilon_n$.
We therefore have
\begin{align*}
d_P((\tilde{f}_n)_*\mu_{X_n},\mu)
&\le d_P((\tilde{f}_n)_*\mu_{X_n},(f_n)_*\mu_{X_n})
+ d_P((f_n)_*\mu_{X_n},\mu)\\
&\le \dKF(\tilde{f}_n,f_n) + \varepsilon_n
\le 2\varepsilon_n.
\end{align*}
In the case where $R = \infty$, the measure
$\mu_n := (\tilde{f}_n)_*\mu_{X_n}$ is a desired one.
In the case where $R < \infty$,
since $\supp\mu \subset B^N_R$,
the measure $\mu_n := (\pi_R)_*(\tilde{f}_n)_*\mu_{X_n}$
satisfies
$\mu_n \in \mathcal{M}(X_n;N,R)$ and $\mu_n \to \mu$ weakly,
where $\pi_R : \field{R}^N \to B^N_R$ is the nearest point projection.
We thus obtain the first part of the lemma.
The rest is clear.
This completes the proof.
\end{proof}
\begin{lem} \label{lem:pyramid-conv-dH}
For given pyramids $\mathcal{P}$ and $\mathcal{P}_n$, $n=1,2,\dots$,
the following {\rm(1)} and {\rm(2)} are equivalent to each other.
\begin{enumerate}
\item $\mathcal{P}_n$ converges weakly to $\mathcal{P}$ as $n\to\infty$.
\item For any natural number $N$ and for any real number $R \ge 0$,
the set $\mathcal{P}_n\cap\mathcal{X}(N,R)$ Hausdorff
converges to $\mathcal{P}\cap\mathcal{X}(N,R)$ as $n\to\infty$, where the Hausdorff distance
is induced from the box metric.
\end{enumerate}
\end{lem}
\begin{proof}
We prove `(1) $\implies$ (2)'.
Suppose that $\mathcal{P}_n$ converges weakly to $\mathcal{P}$, but
$\mathcal{P}_n\cap\mathcal{X}(N,R)$ does not Hausdorff converge to $\mathcal{P}\cap\mathcal{X}(N,R)$
for some $N$ and $R$.
We find a subsequence $\{\mathcal{P}_{n_i}\}$ of $\{\mathcal{P}_n\}$ in such a way that
$\liminf_{n\to\infty} d_H(\mathcal{P}_n\cap\mathcal{X}(N,R),\mathcal{P}\cap\mathcal{X}(N,R)) > 0$.
Since $\mathcal{X}(N,R)$ is $\square$-compact,
Lemma \ref{lem:dH}(2) tells us the $d_H$-compactness of
$\mathcal{F}(\mathcal{X}(N,R))$.
By replacing $\{\mathcal{P}_{n_i}\}$ with a subsequence,
$\mathcal{P}_{n_i} \cap \mathcal{X}(N,R)$ Hausdorff converges to
some compact subset $\mathcal{P}_\infty \subset \mathcal{X}(N,R)$,
that is different from $\mathcal{P}\cap\mathcal{X}(N,R)$.
Since any mm-space $X \in \mathcal{P}_\infty$ is the limit of
some $X_i \in \mathcal{P}_{n_i} \cap \mathcal{X}(N,R)$, $i=1,2,\dots$,
the set $\mathcal{P}_\infty$ is contained in $\mathcal{P}$, so that
$\mathcal{P}_\infty \subset \mathcal{P}\cap\mathcal{X}(N,R)$.
For any mm-space $X \in \mathcal{P} \cap \mathcal{X}(N,R)$,
there is a sequence of mm-spaces $X_i \in \mathcal{P}_{n_i}$ $\square$-converging
to $X$ as $i\to\infty$.
By Lemma \ref{lem:down-XNR},
we find a sequence of mm-spaces $X_i' \in \mathcal{X}(N,R)$ with $X_i' \prec X_i$
that $\square$-converges to $X$.
Since $X_i' \in \mathcal{P}_{n_i} \cap \mathcal{X}(N,R)$,
the space $X$ belongs to $\mathcal{P}_\infty$.
Thus we have $\mathcal{P}_\infty = \mathcal{P} \cap \mathcal{X}(N,R)$.
This is a contradiction.
We prove `(2) $\implies$ (1)'.
We assume (2).
Let $\underline{\mathcal{P}}_\infty$ be the set of the limits of
convergent sequences of mm-spaces $X_n \in \mathcal{P}_n$, and
$\overline{\mathcal{P}}_\infty$ the set of the limits of
convergent subsequences of mm-spaces $X_n \in \mathcal{P}_n$.
We have $\underline{\mathcal{P}}_\infty \subset \overline{\mathcal{P}}_\infty$ in general.
By Proposition \ref{prop:w-conv},
it suffices to prove that $\underline{\mathcal{P}}_\infty = \overline{\mathcal{P}}_\infty = \mathcal{P}$.
Let us first prove $\mathcal{P} \subset \underline{\mathcal{P}}_\infty$.
Take any mm-space $X \in \mathcal{P}$.
By Lemma \ref{lem:X-dense},
there is a sequence of mm-spaces $X_i \in \bigcup_{N=1}^\infty \mathcal{X}(N,N)$
that $\square$-converges to $X$.
For each $i$ we find a natural number $N_i$ with $X_i \in \mathcal{X}(N_i,N_i)$.
By (2), there is a sequence of mm-spaces
$X_{in} \in \mathcal{P}_n \cap \mathcal{X}(N_i,N_i)$, $n=1,2,\dots$,
that $\square$-converges to $X_i$ for each $i$.
There is a sequence $i_n \to \infty$ such that
$X_{i_nn}$ $\square$-converges to $X$, so that
$X$ belongs to $\underline{\mathcal{P}}_\infty$.
We obtain $\mathcal{P} \subset \underline{\mathcal{P}}_\infty$.
The rest of the proof is to show that $\overline{\mathcal{P}}_\infty \subset \mathcal{P}$.
Take any mm-space $X \in \overline{\mathcal{P}}_\infty$.
By Corollary \ref{cor:XN},
$X$ is approximated by
some $\underline{X}_N = (\field{R}^N,\|\cdot\|_\infty,\underline{\mu}_N)$,
$\underline{\mu}_N \in \mathcal{M}(X;N)$.
Since $(\pi_R)_*\underline{\mu}_N \to \underline{\mu}_N$
as $R \to +\infty$, where
$\pi_R : \field{R}^N \to B^N_R$ is the nearest point projection,
$X$ is approximated by some $X' \in \mathcal{X}(N,R)$ with $X' \prec X$.
By the $\square$-closedness of $\mathcal{P}$,
it suffices to prove that $X'$ belongs to $\mathcal{P}$.
It follows from $X \in \overline{\mathcal{P}}_\infty$ that
there are sequences $n_i \to \infty$ and $X_i \in \mathcal{P}_{n_i}$
such that $X_i$ $\square$-converges to $X$.
By Lemma \ref{lem:lim-pyramid}(1),
we find a sequence of mm-spaces
$X_i'$ with $X_i' \prec X_i$ that $\square$-converges to $X'$.
Lemma \ref{lem:down-XNR} implies
the existence of a sequence $X_i'' \in \mathcal{X}(N,R)$ such that
$X_i'' \prec X_i'$ for any $i$ and $X_i''$ converges to $X'$ as $i\to\infty$.
Since $\mathcal{P}_{n_i}$ is a pyramid, $X_i''$ belongs to $\mathcal{P}_{n_i}$.
By (2), $X'$ is an element of $\mathcal{P}$.
This completes the proof.
\end{proof}
\begin{lem} \label{lem:pyramid-conv-mono}
Let $\mathcal{P}$ and $\mathcal{P}_n$, $n=1,2,\dots$, be pyramids.
Let $N$ be a natural number and let $R' \ge R \ge 0$.
If $\mathcal{P}_n\cap\mathcal{X}(N,R')$ Hausdorff converges to $\mathcal{P}\cap\mathcal{X}(N,R')$
as $n\to\infty$, then
$\mathcal{P}_n\cap\mathcal{X}(N,R)$ Hausdorff converges to $\mathcal{P}\cap\mathcal{X}(N,R)$
as $n\to\infty$.
\end{lem}
\begin{proof}
Assume that $\mathcal{P}_n\cap\mathcal{X}(N,R')$ Hausdorff converges to $\mathcal{P}\cap\mathcal{X}(N,R')$
as $n\to\infty$.
By the compactness of $\mathcal{X}(N,R)$,
it suffices to prove that the limit set of any subsequence
of $\{\mathcal{P}_n\cap\mathcal{X}(N,R)\}$
coincides with $\mathcal{P}\cap\mathcal{X}(N,R)$.
Take any mm-space $X \in \mathcal{P}\cap\mathcal{X}(N,R)$.
By the assumption, there are mm-spaces $X_n \in \mathcal{P}_n\cap\mathcal{X}(N,R')$
such that $X_n$ $\square$-converges to $X$.
Lemma \ref{lem:down-XNR} implies that
there are mm-spaces $X_n' \in \mathcal{P}_n\cap\mathcal{X}(N,R)$
such that $X_n' \prec X_n$ for each $n$ and $X_n'$ $\square$-converges
to $X$. Therefore, the limit set of any subsequence of
$\{\mathcal{P}_n\cap\mathcal{X}(N,R)\}$ contains $\mathcal{P}\cap\mathcal{X}(N,R)$.
Let $\{\mathcal{P}_{n_i}\cap\mathcal{X}(N,R)\}$ be a subsequence of
$\{\mathcal{P}_n\cap\mathcal{X}(N,R)\}$,
and let $X_i \in \mathcal{P}_{n_i}\cap\mathcal{X}(N,R)$ $\square$-converge to
an mm-space $X$ as $i\to\infty$.
Since $\mathcal{P}_n\cap\mathcal{X}(N,R')$ Hausdorff converges to $\mathcal{P}\cap\mathcal{X}(N,R')$,
the limit mm-space $X$ belongs to $\mathcal{P}$, so that $X \in \mathcal{P}\cap\mathcal{X}(N,R)$.
This completes the proof.
\end{proof}
\begin{defn}[Metric on the space of pyramids]
\label{defn:metric-Pi}
\index{metric on the space of pyramids}
\index{rho, rhok@$\rho$, $\rho_k$}
Define
for a natural number $k$ and for two pyramids $\mathcal{P}$ and $\mathcal{P}'$,
\begin{align*}
\rho_k(\mathcal{P},\mathcal{P}') &:= \frac{1}{4k} d_H(\mathcal{P}\cap\mathcal{X}(k,k),\mathcal{P}'\cap\mathcal{X}(k,k)),\\
\rho(\mathcal{P},\mathcal{P}') &:= \sum_{k=1}^\infty 2^{-k} \rho_k(\mathcal{P},\mathcal{P}').
\end{align*}
\end{defn}
\begin{thm} \label{thm:metric-Pi}
$\rho$ is a metric on the space $\Pi$ of pyramids
that is compatible with weak convergence.
$\Pi$ is compact with respect to $\rho$.
\end{thm}
\begin{proof}
We first prove that $\rho$ is a metric.
Since $\square \le 1$, we have $\rho_k \le 1/(4k)$ for each $k$
and then $\rho \le 1/4$.
Each $\rho_k$ is a pseudo-metric on $\Pi$ and so is $\rho$.
If $\rho(\mathcal{P},\mathcal{P}') = 0$ for two pyramids $\mathcal{P}$ and $\mathcal{P}'$,
then $\rho_k(\mathcal{P},\mathcal{P}') = 0$ for any $k$,
which implies $\mathcal{P} = \mathcal{P}'$.
Thus, $\rho$ is a metric on $\Pi$.
We next prove the compatibility of the metric $\rho$
with weak convergence in $\Pi$.
It follows from Lemmas \ref{lem:pyramid-conv-dH} and
\ref{lem:pyramid-conv-mono} that
a sequence of pyramids $\mathcal{P}_n$, $n=1,2,\dots$,
converges weakly to a pyramid $\mathcal{P}$
if and only if $\lim_{n\to\infty} \rho_k(\mathcal{P}_n,\mathcal{P}) = 0$
for any $k$, which is also equivalent to
$\lim_{n\to\infty} \rho(\mathcal{P}_n,\mathcal{P}) = 0$.
Since $\Pi$ is sequentially compact
(see Theorem \ref{thm:cpt-pyramid}),
it is compact with respect to $\rho$.
This completes the proof.
\end{proof}
\begin{lem} \label{lem:PX-MNR-dH}
Let $X$ and $Y$ be two mm-spaces, $N$ a natural number,
and $R$ a nonnegative extended real number.
Then we have
\[
d_H(\mathcal{P}_X \cap \mathcal{X}(N,R),\mathcal{P}_Y \cap \mathcal{X}(N,R))
\le 2\,d_H(\mathcal{M}(X;N,R),\mathcal{M}(Y;N,R)).
\]
\end{lem}
\begin{proof}
The lemma follows from Proposition \ref{prop:box-di}.
\end{proof}
\begin{thm} \label{thm:rho-dconc}
For any two mm-spaces $X$ and $Y$, we have
\[
\rho(\mathcal{P}_X,\mathcal{P}_Y) \le \dconc(X,Y),
\]
i.e., the embedding map $\iota : \mathcal{X} \ni X \mapsto \mathcal{P}_X \in \Pi$
is $1$-Lipschitz continuous.
\end{thm}
\begin{proof}
By Lemmas \ref{lem:PX-MNR-dH}, \ref{lem:MR-half}, and \ref{lem:M-dconc},
\begin{align*}
& d_H(\mathcal{P}_X \cap \mathcal{X}(N,R),\mathcal{P}_Y \cap \mathcal{X}(N,R))\\
&\le 2 d_H(\mathcal{M}(X;N,R),\mathcal{M}(Y;N,R))\\
&\le 4 d_H(\mathcal{M}(X;N),\mathcal{M}(Y;N))
\le 4N \dconc(X,Y),
\end{align*}
so that $\rho_k(\mathcal{P}_X,\mathcal{P}_Y) \le \dconc(X,Y)$ for any $k$.
This proves the theorem.
\end{proof}
\begin{rem}
As is shown in Section \ref{sec:Spheres-Gaussians},
the sequence of spheres $X_n := S^n(\sqrt{n})$, $n=1,2,\dots$,
satisfies that $\mathcal{P}_{X_n}$ converges weakly (Theorem \ref{thm:sphere-Gaussian})
and that $\{X_n\}$ has no $\dconc$-Cauchy subsequence
(Corollary \ref{cor:sphere-Gaussian}).
This implies that it is impossible to estimate $\rho(\mathcal{P}_X,\mathcal{P}_Y)$
below by $\dconc(X,Y)$.
However, by Proposition \ref{prop:conc-pyramid} and
Theorem \ref{thm:metric-Pi},
the map $\iota : \mathcal{X} \ni X \mapsto \mathcal{P}_X \in \Pi$
is an embedding map with respect to $\dconc$ and $\rho$.
\end{rem}
\chapter{Asymptotic concentration}
\label{chap:asymp-conc}
\section{Compactification
of the space of ideal mm-spaces}
\label{sec:comp}
In this section, we prove the main theorem in this book
that the $\dconc$-completion of the space of mm-spaces
is embedded in the set of pyramids.
\begin{defn}[Asymptotic sequence of mm-spaces and asymptotic concentration]
\index{asymptotic} \index{asymptotic concentration}
A sequence of mm-spaces $X_n$, $n=1,2,\dots$, is said to be
\emph{asymptotic} if $\mathcal{P}_{X_n}$ converges weakly as $n\to\infty$.
We say that a sequence of mm-spaces \emph{asymptotically concentrates}
if it is a $\dconc$-Cauchy sequence.
\index{asymptotically concentrate}
\end{defn}
\begin{prop} \label{prop:asymp}
If a sequence of mm-spaces asymptotically concentrates,
then it is asymptotic.
\end{prop}
\begin{proof}
Let $\{X_n\}$ be a sequence of mm-spaces that asymptotically concentrates.
Then, Theorem \ref{thm:rho-dconc} proves that
$\{\mathcal{P}_{X_n}\}$ is a $\rho$-Cauchy sequence
and converges in $\Pi$ by the compactness of $(\Pi,\rho)$.
This completes the proof.
\end{proof}
It is clear that any monotone
nondecreasing (with respect to the Lipschitz order) sequence of mm-spaces,
$X_n$, $n=1,2,\dots$, is asymptotic,
where the limit pyramid is the $\square$-closure of
$\bigcup_{n=1}^\infty \mathcal{P}_{X_n}$.
In particular, for any given mm-spaces $F_n$, $n=1,2,\dots$,
the product space
\[
F_1 \times F_2 \times \dots \times F_n
\]
with $d_{l_p}$, $1 \le p \le +\infty$, and product measure
$\bigotimes_{i=1}^n \mu_{F_i}$ is asymptotic.
\begin{defn}[Completion of $\mathcal{X}$ and ideal mm-space]
\index{Completion of X@completion of $\mathcal{X}$}
\index{ideal mm-space}
Denote by $\bar{\mathcal{X}}$ the $\dconc$-completion of the set $\mathcal{X}$
\index{Xbar@$\bar{\mathcal{X}}$}
of mm-isomorphism classes of mm-spaces, and let
$\partial\mathcal{X} := \bar{\mathcal{X}} \setminus \mathcal{X}$.
\index{boundaryX@$\partial\mathcal{X}$}
We call each element of $\partial\mathcal{X}$ an \emph{ideal mm-space}.
\end{defn}
Since the map $\iota : \mathcal{X} \ni X \mapsto \mathcal{P}_X \in \Pi$ is $1$-Lipschitz
continuous
with respect to $\dconc$ and $\rho$ (Theorem \ref{thm:rho-dconc}),
this uniquely extends to a $1$-Lipschitz continuous
map
\[
\iota : \bar\mathcal{X} \ni \bar{X} \longmapsto \mathcal{P}_{\bar{X}} \in \Pi.
\]
A main purpose of this section is to prove that the map $\iota$
is a topological embedding map (see Theorem \ref{thm:emb-pyramid})
by using many statements proved in the previous sections.
Since $\Pi$ is compact and $\iota(\mathcal{X})$ is dense in $\Pi$,
this turns out to be a compactification of $\bar\mathcal{X}$ (and of $\mathcal{X}$)
with respect to the concentration topology.
\begin{defn}[Measurement of a pyramid]
\label{defn:m-pyramid} \index{measurement of a pyramid} \index{measurement}
Let $\mathcal{P}$ be a pyramid.
For a natural number $N$ and a nonnegative extended real number $R$,
we define
\begin{align*}
\mathcal{M}(\mathcal{P};N,R) &:= \{\;\mu \in \mathcal{M}(N,R) \mid
(B^N_R,\|\cdot\|_\infty,\mu) \in \mathcal{P}\;\},\\
\mathcal{M}(\mathcal{P};N) &:= \mathcal{M}(\mathcal{P};N,+\infty).
\end{align*}
$\mathcal{M}(\mathcal{P};N,R)$ and $\mathcal{M}(\mathcal{P};N)$ are respectively called
the \emph{$(N,R)$-measurement} and the \emph{$N$-measurement} of $\mathcal{P}$.
\index{NR-measurement@$(N,R)$-measurement}
\index{N-measurement@$N$-measurement}
\end{defn}
We see that $\mathcal{M}(\mathcal{P}_X;N,R) = \mathcal{M}(X;N,R)$ for any mm-space $X$.
It is obvious that $\mathcal{M}(\mathcal{P};N,R)$ is perfect on $B^N_R$ for any pyramid
$\mathcal{P}$.
\begin{lem} \label{lem:perfect-lim}
Let $\mathcal{A}_n,\mathcal{A} \subset \mathcal{M}(N,R)$, $n=1,2,\dots$,
be closed subsets, where
$N$ is a natural number and $R$ a nonnegative extended real number.
If $\mathcal{A}_n$ Hausdorff converges to $\mathcal{A}$ as $n\to\infty$
and if each $\mathcal{A}_n$ is perfect on $B^N_R$, then
$\mathcal{A}$ is perfect on $B^N_R$.
\end{lem}
\begin{proof}
Assume that $(B^N_R,\|\cdot\|_\infty,\nu) \prec (B^N_R,\|\cdot\|_\infty,\mu)$
for two measures $\mu \in \mathcal{A}$ and $\nu \in \mathcal{M}(N,R)$.
Since $d_H(\mathcal{A}_n,\mathcal{A}) \to 0$,
there is a sequence of measures $\mu_n \in \mathcal{A}_n$
that converges weakly to $\mu$.
Since $\square((B^N_R,\|\cdot\|_\infty,\mu_n),(B^N_R,\|\cdot\|_\infty,\mu))
\le d_P(\mu_n,\mu) \to 0$ as $n\to\infty$ and
by Lemma \ref{lem:lim-pyramid}(1),
there is a sequence of mm-spaces $X_n$ such that
$X_n \prec (B^N_R,\|\cdot\|_\infty,\mu_n)$
and $X_n$ $\square$-converges to $(B^N_R,\|\cdot\|_\infty,\nu)$.
Applying Lemma \ref{lem:down-XNR},
we find a sequence of measures $\nu_n \in \mathcal{M}(X_n;N,R)$
converging weakly to $\nu$.
Since $(B^N_R,\|\cdot\|_\infty,\nu_n) \prec X_n \prec
(B^N_R,\|\cdot\|_\infty,\mu_n)$, the perfectness of $\mathcal{A}_n$
implies that $\nu_n$ belongs to $\mathcal{A}_n$,
so that $\nu$ belongs to $\mathcal{A}$.
This completes the proof.
\end{proof}
\begin{thm}[Observable criterion for asymptotic concentration]
\label{thm:cri-asymp-conc}
\index{observable criterion for asymptotic concentration}
\ \\
Let $\{X_n\}_{n=1}^\infty$ be a sequence of mm-spaces
and $\bar{X} \in \partial\mathcal{X}$ an ideal mm-space.
Then, the following {\rm(1)}, {\rm(2)}, and {\rm(3)}
are equivalent to each other.
\begin{enumerate}
\item $X_n$ asymptotically concentrates to $\bar{X}$ as $n\to\infty$.
\item For any natural number $N$,
the $N$-measurement $\mathcal{M}(X_n;N)$ Hausdorff converges to $\mathcal{M}(\mathcal{P}_{\bar{X}};N)$
as $n\to\infty$.
\item For any natural number $N$ and any nonnegative real number $R$,
the $(N,R)$-measurement $\mathcal{M}(X_n;N,R)$ Hausdorff converges to
$\mathcal{M}(\mathcal{P}_{\bar{X}};N,R)$ as $n\to\infty$.
\end{enumerate}
\end{thm}
\begin{proof}[Proof of `{\rm(1)} $\implies$ {\rm(2)}']
Assume that a sequence of mm-spaces $X_n$, $n=1,2,\dots$,
asymptotically concentrates to $\bar{X}$.
We take any natural number $N$ and fix it.
Lemma \ref{lem:M-dconc} implies that
$\{\mathcal{M}(X_n;N)\}_n$ is $d_H$-Cauchy.
By Lemma \ref{lem:dH}(1),
$\{\mathcal{M}(X_n;N)\}_n$ Hausdorff converges to a closed subset
$\mathcal{A} \subset \mathcal{M}(N)$.
Let us prove
\begin{align} \label{eq:cL-barX}
\mathcal{P}_{\bar{X}} \cap \mathcal{X}(N) = \{\;(\field{R}^N,\|\cdot\|_\infty,\mu) \mid
\mu \in \mathcal{A}\;\}.
\end{align}
To prove `$\supset$', we take any measure $\mu \in \mathcal{A}$.
There is a sequence of measures $\mu_n \in \mathcal{M}(X_n;N)$, $n=1,2,\dots$,
converging weakly to $\mu$.
$(\field{R}^N,\|\cdot\|_\infty,\mu_n)$ $\square$-converges to
$(\field{R}^N,\|\cdot\|_\infty,\mu)$ as $n\to\infty$.
Since $(\field{R}^N,\|\cdot\|_\infty,\mu_n) \in \mathcal{P}_{X_n}$ and
since $\mathcal{P}_{X_n}$ converges weakly to $\mathcal{P}_{\bar{X}}$,
the space $(\field{R}^N,\|\cdot\|_\infty,\mu)$ belongs to $\mathcal{P}_{\bar{X}}$.
To prove `$\subset$', we take any mm-space
$(\field{R}^N,\|\cdot\|_\infty,\mu) \in \mathcal{P}_{\bar{X}} \cap \mathcal{X}(N)$.
Since $\mathcal{P}_{X_n}$ converges weakly to $\mathcal{P}_{\bar{X}}$ as $n\to\infty$,
there is a sequence of mm-spaces $X_n' \in \mathcal{P}_{X_n}$
that $\square$-converges to $(\field{R}^N,\|\cdot\|_\infty,\mu)$.
Applying Lemma \ref{lem:down-XNR},
we find a sequence of measures $\mu_n \in \mathcal{M}(X_n;N)$, $n=1,2,\dots$,
converging weakly to $\mu$.
We therefore have $\mu \in \mathcal{A}$.
\eqref{eq:cL-barX} has been proved.
By Lemma \ref{lem:perfect-lim}, $\mathcal{A}$ is perfect,
which together with \eqref{eq:cL-barX} proves that
$\mathcal{A} = \mathcal{M}(\mathcal{P}_{\bar{X}};N)$.
We obtain (2).
\end{proof}
For the rest of the proof of Theorem \ref{thm:cri-asymp-conc},
we need several lemmas.
\begin{lem} \label{lem:dGH-dconc}
For any two mm-spaces $X$ and $Y$ we have
\[
d_{GH}(\mathcal{L}_1(X),\mathcal{L}_1(Y)) \le \dconc(X,Y).
\]
\end{lem}
\begin{proof}
Let $\varphi : I \to X$ and $\psi : I \to Y$ be any parameters.
By Lemmas \ref{lem:pb-me-di} and \ref{lem:action-dH}(2),
\begin{align*}
d_{GH}(\mathcal{L}_1(X),\mathcal{L}_1(Y))
&\le d_H(\varphi^*\mathcal{L}_1(X),\psi^*\mathcal{L}_1(Y))\\
&= d_H(\varphi^*\mathcal{L}{\it ip}_1(X),\psi^*\mathcal{L}{\it ip}_1(Y)).
\end{align*}
Taking the infimum of the right-hand side over all $\varphi$ and $\psi$,
we have the lemma.
\end{proof}
Denote by $\mathcal{H}$ the set of isometry classes of compact metric spaces.
Lemma \ref{lem:dGH-dconc} tells us that the map
\[
\mathcal{L}_1 : \mathcal{X} \ni X \longmapsto \mathcal{L}_1(X) \in \mathcal{H}
\]
is $1$-Lipschitz continuous with respect to $\dconc$ and $d_{GH}$.
Since $(\mathcal{H},d_{GH})$ is a complete metric space
(see Lemma \ref{lem:dGH-complete}),
the map $\mathcal{L}_1$ extends to
\[
\mathcal{L}_1 : \bar{\mathcal{X}} \to \mathcal{H}
\]
as an $1$-Lipschitz map.
If a sequence of mm-spaces $X_n$, $n=1,2,\dots$,
asymptotically concentrates to an ideal mm-space $\bar{X} \in \partial\mathcal{X}$,
then $\mathcal{L}_1(X_n)$ Gromov-Hausdorff converges to $\mathcal{L}_1(\bar{X})$ as $n\to\infty$.
\index{L1(barX)@$\mathcal{L}_1(\bar{X})$}
\begin{lem} \label{lem:isom-L}
\begin{enumerate}
\item For any two mm-spaces $X$ and $Y$ with $Y \prec X$,
there exists an isometric embedding
$\iota_Y : \mathcal{L}_1(Y) \hookrightarrow \mathcal{L}_1(X)$.
\item For any ideal mm-space $\bar{X} \in \partial\mathcal{X}$
and for any mm-space $Y \in \mathcal{P}_{\bar{X}}$,
there exists an isometric embedding
$\iota_Y : \mathcal{L}_1(Y) \hookrightarrow \mathcal{L}_1(\bar{X})$.
\end{enumerate}
\end{lem}
\begin{proof}
We prove (1).
By $Y \prec X$, we find a $1$-Lipschitz map $p : X \to Y$
with $p_*\mu_X = \mu_Y$.
By Lemma \ref{lem:pb-me-di},
$\iota_Y := p^* : \mathcal{L}_1(Y) \to \mathcal{L}_1(X)$ is an isometric embedding.
We prove (2).
Let $\bar{X} \in \bar\mathcal{X}$ and $Y \in \mathcal{P}_{\bar{X}}$.
There is a sequence of mm-spaces $X_n$, $n=1,2,\dots$,
that asymptotically concentrates to $\bar{X}$.
Since $\mathcal{P}_{X_n}$ converges weakly to $\mathcal{P}_{\bar{X}}$,
there is a sequence of mm-spaces $Y_n \in \mathcal{P}_{X_n}$
that $\square$-converges to $Y$.
By $Y_n \prec X_n$ and (1), we find an isometric embedding
$\iota_{Y_n} : \mathcal{L}_1(Y_n) \hookrightarrow \mathcal{L}_1(X_n)$.
Since $\mathcal{L}_1(X_n)$ and $\mathcal{L}_1(Y_n)$ Gromov-Hausdorff converge to
$\mathcal{L}_1(\bar{X})$ and $\mathcal{L}_1(Y)$ respectively as $n\to\infty$,
some subsequence of $\{p_n^*\}$ converges to
an isometric embedding $\iota_Y : \mathcal{L}_1(Y) \hookrightarrow \mathcal{L}_1(\bar{X})$
(see \cite{Petersen}*{\S 10.1.3}).
\end{proof}
\begin{defn}[Concentrated pyramid]
\index{concentrated pyramid}
A pyramid $\mathcal{P}$ is said to be \emph{concentrated}
if $\{\mathcal{L}_1(X)\}_{X \in \mathcal{P}}$ is $d_{GH}$-precompact.
\end{defn}
Lemma \ref{lem:isom-L} implies
\begin{cor} \label{cor:concentrated}
For any {\rm(}ideal{\rm)} mm-space $\bar{X} \in \bar{\mathcal{X}}$,
the associated pyramid $\mathcal{P}_{\bar{X}}$ is concentrated.
\end{cor}
\begin{lem} \label{lem:Sep-CapLo}
For any mm-space $X$ and any real number $\kappa > 0$,
we have
\[
\Sep(X;\kappa,\kappa) \le \kappa(\Cap_\kappa(\mathcal{L}_1(X))+1).
\]
\end{lem}
\begin{proof}
If $\Sep(X;\kappa,\kappa) = 0$, then the lemma is trivial.
Assume $\Sep(X;\kappa,\kappa) > 0$, and
let $r$ be any real number with $0 < r < \Sep(X;\kappa,\kappa)$.
We find two Borel subsets $A, B \subset X$ in such a way that
$\mu_X(A),\mu_X(B) \ge \kappa$ and $d_X(A,B) > r$.
Let $f(x) := d_X(x,A)$, $x \in X$.
It is clear that
$\lambda f$ belongs to $\mathcal{L}{\it ip}_1(X)$ for any $\lambda \in [\,-1,1\,]$.
For two real numbers $\lambda$ and $\lambda'$
we estimate $\dKF([\lambda f],[\lambda'f])$.
Let $c$ be any real number.
If $x \in A$, then $|\lambda f(x)-\lambda'f(x)-c| = |c|$.
If $x \in B$, then $|\lambda f(x)-\lambda'f(x)-c|
\ge |\lambda-\lambda'|f(x)-|c| \ge |\lambda-\lambda'|r - |c|$.
Therefore, if $|c| \ge |\lambda-\lambda'| r/2$, then
\begin{equation}
\label{eq:sup-ObsDiam}
|\lambda f(x) - \lambda'f(x) - c| \ge \frac{|\lambda-\lambda'| r}{2}
\end{equation}
for all $x \in A$.
If $|c| < |\lambda-\lambda'| r/2$, then
\eqref{eq:sup-ObsDiam} holds for all $x \in B$.
We thus have
\[
\mu_X\left(|\lambda f(x) - \lambda'f(x) - c|
\ge \frac{|\lambda-\lambda'| r}{2}\right) \ge \kappa,
\]
so that, if $|\lambda-\lambda'| \ge 2\kappa/r$, then
$\dKF([\lambda f],[\lambda'f]) \ge \kappa$.
Setting $N := [r/(2\kappa)]$ and $\lambda_k := 2k\kappa/r$
for $k = 0,\pm 1,\pm 2,\dots,\pm N$,
we see that
$-1 \le \lambda_{-N} < \dots < \lambda_{N} \le 1$.
For two different integers $k$ and $l$ in $\{0,\pm 1,\pm 2,\dots,\pm N\}$,
we have $|\lambda_k-\lambda_l| \ge 2\kappa/r$
and so $\dKF([\lambda_k f],[\lambda_l f]) \ge \kappa$.
Namely, $\{[\lambda_k f]\}_k$ is a $\kappa$-discrete net
of $\mathcal{L}_1(X)$ and therefore $2N+1 \le \Cap_\kappa(X)$.
Since $N > r/(2\kappa) -1$, we have
\[
\frac{r}{\kappa}-1 < \Cap_\kappa(X).
\]
This completes the proof.
\end{proof}
\begin{cor} \label{cor:sup-diam-cP}
Let $\mathcal{P}$ be a concentrated pyramid.
Then we have
\[
\sup_{\mu\in\mathcal{M}(\mathcal{P};N)} \diam(\mu;1-\kappa) < +\infty.
\]
for any natural number $N$ and any real number $\kappa$ with $0 < \kappa < 1$.
\end{cor}
\begin{proof}
It follows from Lemmas \ref{lem:dGH-precpt} and \ref{lem:Sep-CapLo} that
\[
\sup_{X\in\mathcal{P}} \Sep(X;\kappa,\kappa) < +\infty
\]
for any $\kappa > 0$,
which together with Proposition \ref{prop:ObsDiam-Sep} and
Lemma \ref{lem:ObsDiamRN-ObsDiam} implies
\[
\sup_{X\in\mathcal{P}} \ObsDiam_{(\field{R}^N,\|\cdot\|_\infty)}(X;-\kappa) < +\infty
\]
for any $N$ and $\kappa$ with $0 < \kappa < 1$.
This proves the corollary.
\end{proof}
\begin{proof}[Proof of `{\rm(2)} $\Longleftrightarrow$ {\rm(3)}'
of Theorem \ref{thm:cri-asymp-conc}]
Let $\mathcal{A} := \mathcal{M}(X;N)$ and $\mathcal{A}_n := \mathcal{M}(X_n;N)$.
Since $\mathcal{P}_{\bar{X}}$ is concentrated (see Corollary \ref{cor:concentrated}),
Corollary \ref{cor:sup-diam-cP} implies
the assumption of Lemma \ref{lem:MR}.
Applying Lemma \ref{lem:MR} completes the proof.
\end{proof}
\begin{defn}[Approximate a pyramid]
\index{approximate a pyramid}
Let $\mathcal{P}$ be a pyramid and $\{Y_m\}_{m=1}^\infty$ a sequence of mm-spaces.
We say that $\{Y_m\}_{m=1}^\infty$ \emph{approximates $\mathcal{P}$} if
\[
Y_1 \prec Y_2 \prec \dots \prec Y_m \prec \dots\quad\text{and}\quad
\overline{\bigcup_{m=1}^\infty \mathcal{P}_{Y_m}}^\square = \mathcal{P},
\]
where the upper bar with $\square$ means the $\square$-closure.
\end{defn}
We see that, if $\{Y_m\}_{m=1}^\infty$ approximates a pyramid $\mathcal{P}$,
then $\mathcal{P}_{Y_m}$ converges weakly to $\mathcal{P}$ as $m\to\infty$.
\begin{lem} \label{lem:approx-pyramid}
For any pyramid $\mathcal{P}$,
there exists a sequence of mm-spaces that approximates $\mathcal{P}$.
\end{lem}
\begin{proof}
The $\square$-separability of $\mathcal{X}$ (see
Proposition \ref{prop:separable}) implies that
there is a dense countable set $\{Y_m'\}_{m=1}^\infty \subset \mathcal{P}$.
Let $Y_1 := Y_1'$.
There is an mm-space $Y_2 \in \mathcal{P}$ dominating $Y_1$ and $Y_2'$.
There is also an mm-space $Y_3 \in \mathcal{P}$
dominating $Y_2$ and $Y_3'$.
Repeating this procedure, we have a sequence of mm-spaces
$Y_m \in \mathcal{P}$, $m=1,2,\dots$, with the property that
$Y_1 \prec Y_2 \prec \dots \prec Y_m \prec \cdots$
and $Y_m' \prec Y_m \in \mathcal{P}$ for any $m$.
We see that
\[
\{Y_m'\}_{m=1}^\infty \subset \bigcup_{m=1}^\infty \mathcal{P}_{Y_m}
\subset \mathcal{P}.
\]
This completes the poof.
\end{proof}
Let $\mathcal{P}$ be a concentrated pyramid and
$\{Y_m\}_{m=1}^\infty$ a sequence of mm-spaces
that approximates $\mathcal{P}$.
By the monotonicity of $\{Y_m\}$ and by Lemma \ref{lem:isom-L}(1),
we have a sequence of isometric embeddings
\[
\mathcal{L}_1(Y_1) \hookrightarrow \mathcal{L}_1(Y_2) \hookrightarrow \dots
\hookrightarrow \mathcal{L}_1(Y_m) \hookrightarrow \dots.
\]
Since $\{\mathcal{L}_1(Y_m)\}$ is $d_{GH}$-precompact,
it Gromov-Hausdorff converges to a compact metric space,
say $\mathcal{L}_1(\mathcal{P})$. \index{L1(P)@$\mathcal{L}_1(\mathcal{P})$}
Each $\mathcal{L}_1(Y_m)$ is embedded into $\mathcal{L}_1(\mathcal{P})$ isometrically.
\begin{lem} \label{lem:cL-prime}
Let $\mathcal{P}$ be a concentrated pyramid.
For any mm-space $Z \in \mathcal{P}$, there exists an isometric embedding
$\iota_Z : \mathcal{L}_1(Z) \hookrightarrow \mathcal{L}_1(\mathcal{P})$.
\end{lem}
\begin{proof}
Let $\{Y_m\}_{m=1}^\infty$ be as above.
Take any mm-space $Z \in \mathcal{P}$ and fix it.
There is a sequence of mm-spaces $Z_m \in \mathcal{P}_{Y_m}$, $m=1,2,\dots$,
that $\square$-converges to $Z$.
It follows from Lemma \ref{lem:dGH-dconc} and Proposition \ref{prop:dconc-box}
that $\mathcal{L}_1(Z_m)$ Gromov-Hausdorff converges to $\mathcal{L}_1(Z)$ as $m\to\infty$.
By Lemma \ref{lem:isom-L}(1),
each $\mathcal{L}_1(Z_m)$ is isometrically embedded in $\mathcal{L}_1(Y_m)$.
Since $\mathcal{L}_1(Y_m)$ Gromov-Hausdorff converges to $\mathcal{L}_1(\mathcal{P})$,
the space $\mathcal{L}_1(Z)$ is isometrically embedded in $\mathcal{L}_1(\mathcal{P})$.
This completes the proof.
\end{proof}
\begin{prop}
For a concentrated pyramid $\mathcal{P}$,
the metric space $\mathcal{L}_1(\mathcal{P})$ is {\rm(}up to isometry{\rm)}
independent of a defining sequence $\{Y_m\}$
of mm-spaces approximating $\mathcal{P}$.
\end{prop}
\begin{proof}
Let $\{Y_m\}$ and $\{Y_m'\}$ be two sequences of mm-spaces
that approximate $\mathcal{P}$, and assume that
$\mathcal{L}_1(Y_m)$ and $\mathcal{L}_1(Y_m')$ Gromov-Hausdorff converge
to $\mathcal{L}_1(\mathcal{P})$ and $\mathcal{L}_1'(\mathcal{P})$ respectively as $m\to\infty$.
By Lemma \ref{lem:cL-prime},
each $\mathcal{L}_1(Y_m)$ is isometrically embedded into $\mathcal{L}_1'(\mathcal{P})$,
so that $\mathcal{L}_1(\mathcal{P})$ is isometrically embedded into $\mathcal{L}_1'(\mathcal{P})$.
In the same way, $\mathcal{L}_1'(\mathcal{P})$ is isometrically embedded into $\mathcal{L}_1(\mathcal{P})$.
Therefore, $\mathcal{L}_1(\mathcal{P})$ and $\mathcal{L}_1'(\mathcal{P})$ are isometric to each other
(see \cite{BBI}*{1.6.14}).
This completes the proof.
\end{proof}
\begin{lem} \label{lem:KNcP}
Let $X$ be an mm-space and $\bar{X} \in \partial\mathcal{X}$ an ideal mm-space.
Then, for any natural number $N$, we have
\[
d_H(K_N(\mathcal{L}_1(X)),K_N(\mathcal{L}_1(\mathcal{P}_{\bar{X}})))
\le 2 d_H(\mathcal{M}(X;N),\mathcal{M}(\mathcal{P}_{\bar{X}};N)).
\]
In particular, if $\{X_n\}$ is a sequence of mm-spaces such that
$\mathcal{M}(X_n;N)$ Hausdorff converges to $\mathcal{M}(\mathcal{P}_{\bar{X}};N)$
as $n\to\infty$ for any $N$, then
$\mathcal{L}_1(X_n)$ Gromov-Hausdorff converges to $\mathcal{L}_1(\mathcal{P}_{\bar{X}})$.
\end{lem}
\begin{proof}
The proof is similar to that of Lemma \ref{lem:KN-Lip}.
Assume that $d_H(\mathcal{M}(X;N),\mathcal{M}(\mathcal{P}_{\bar{X}};N)) < \varepsilon$
for a real number $\varepsilon$ and for a natural number $N$.
Let us first prove that
$K_N(\mathcal{L}_1(X)) \subset B_{2\varepsilon}(K_N(\mathcal{L}_1(\mathcal{P}_{\bar{X}})))$.
We take any matrix $A = (\dKF([f_i],[f_j]))_{ij} \in K_N(\mathcal{L}_1(X))$,
where $f_i \in \mathcal{L}{\it ip}_1(X)$, $i=1,2,\dots,N$.
We set $F := (f_1,f_2,\dots,f_N) : X \to \field{R}^N$.
Since $F_*\mu_X \in \mathcal{M}(X;N)$,
there is a measure $\nu \in \mathcal{M}(\mathcal{P}_{\bar{X}};N)$
such that $d_P(F_*\mu_X,\nu) < \varepsilon$.
For $i=1,2,\dots,N$, we define
$g_i : \field{R}^N \to \field{R}$ by $g_i(x_1,x_2,\dots,x_N) := x_i$
for $(x_1,x_2,\dots,x_N) \in \field{R}^N$.
Since $(g_1,\dots,g_n) = \id_{\field{R}^N}$,
Lemma \ref{lem:me-diff-di}(2) implies that
\[
|\;\dKF([f_i],[f_j]) - d_{KF}^\nu([g_i],[g_j])\;|
\le 2\,d_P(F_*\mu_X,\nu) < 2\varepsilon
\]
for any $i,j = 1,\dots,N$.
Setting $B := (d_{KF}^\nu([g_i],[g_j]))_{ij}$ we have
$\|A-B\|_\infty < 2\varepsilon$.
$[g_i] \in \mathcal{L}_1(Y)$ implies $B \in K_N(\mathcal{L}_1(Y))$.
Let $Y := (\field{R}^N,\|\cdot\|_\infty,\nu)$.
Since $Y \in \mathcal{P}_{\bar{X}}$ and by Lemma \ref{lem:cL-prime},
we have $B \in K_N(\mathcal{L}_1(Y)) \subset K_N(\mathcal{L}_1(\mathcal{P}_{\bar{X}}))$
and therefore $A \in B_{2\varepsilon}(K_N(\mathcal{L}_1(\mathcal{P}_{\bar{X}})))$.
We obtain $K_N(\mathcal{L}_1(X)) \subset B_{2\varepsilon}(K_N(\mathcal{L}_1(\mathcal{P}_{\bar{X}})))$.
Let us next prove that
$K_N(\mathcal{L}_1(\mathcal{P}_{\bar{X}})) \subset B_{2\varepsilon}(K_N(\mathcal{L}_1(X)))$.
Take any matrix $B \in K_N(\mathcal{L}_1(\mathcal{P}_{\bar{X}}))$.
Let $\{Y_m\}$ be a sequence of mm-spaces approximating $\mathcal{P}_{\bar{X}}$.
Since $\mathcal{L}_1(Y_m)$ Gromov-Hausdorff converges to $\mathcal{L}_1(\mathcal{P}_{\bar{X}})$
as $m\to\infty$ and by Lemma \ref{lem:KN-GH},
$K_N(\mathcal{L}_1(Y_m))$ Hausdorff converges to $K_N(\mathcal{L}_1(\mathcal{P}_{\bar{X}}))$
as $m\to\infty$, so that there is a matrix $B_m \in K_N(\mathcal{L}_1(Y_m))$
such that $\lim_{m\to\infty} \|B_m-B\|_\infty = 0$.
We find functions $g_{mi} \in \mathcal{L}{\it ip}_1(Y_m)$, $i=1,2,\dots,N$,
with $B_m = (\dKF([g_{mi}],[g_{mj}]))_{ij}$.
Since $Y_m \in \mathcal{P}_{\bar{X}}$,
setting $G_m := (g_{m1},\dots,g_{mN}) : Y_m \to \field{R}^N$
we have $(G_m)_*\mu_{Y_m} \in \mathcal{M}(\mathcal{P}_{\bar{X}};N)$.
By the assumption, there is a $1$-Lipschitz map
$F_m : X \to (\field{R}^N,\|\cdot\|_\infty)$ such that
\[
d_P((F_m)_*\mu_X,(G_m)_*\mu_{Y_m}) < \varepsilon.
\]
Setting $(f_{m1},f_{m2},\dots,f_{mN}) := F_m$, we have,
by Lemma \ref{lem:me-diff-di},
\[
|\;\dKF([f_{mi}],[f_{mj}])-\dKF([g_{mi}],[g_{mj}])\;|
< 2\varepsilon.
\]
Let $A_m := (\dKF([f_{mi}],[f_{mj}]))_{ij}$.
Since $\|A_m-B_m\|_\infty < 2\varepsilon$ and
$A_m \in K_N(\mathcal{L}_1(X))$, we have
$B_m \in B_{2\varepsilon}(K_N(\mathcal{L}_1(X)))$
and therefore $B \in B_{2\varepsilon}(K_N(\mathcal{L}_1(X)))$.
We obtain the first part of the lemma.
The rest follows from Lemma \ref{lem:KN-GH-2}.
This completes the proof.
\end{proof}
\begin{prop} \label{prop:cL-barX}
Let $\bar{X} \in \partial\mathcal{X}$ be an ideal mm-space.
Then, $\mathcal{L}_1(\bar{X})$ and $\mathcal{L}_1(\mathcal{P}_{\bar{X}})$ are isometric to each other.
\end{prop}
\begin{proof}
Let $\bar{X} \in \partial\mathcal{X}$ be an ideal mm-space.
We take a sequence of mm-spaces $X_n$, $n=1,2,\dots$,
that asymptotically concentrates to $\bar{X}$.
It follows from the definition of $\mathcal{L}_1(\bar{X})$ that
$\mathcal{L}_1(X_n)$ Gromov-Hausdorff converges to $\mathcal{L}_1(\bar{X})$.
By `(1) $\implies$ (2)' of Theorem \ref{thm:cri-asymp-conc},
the measurement $\mathcal{M}(X_n;N)$ Hausdorff converges to $\mathcal{M}(\mathcal{P}_{\bar{X}};N)$
as $n\to\infty$ for any $N$.
By Lemma \ref{lem:KNcP},
$\mathcal{L}_1(X_n)$ Gromov-Hausdorff converges to $\mathcal{L}_1(\mathcal{P}_{\bar{X}})$
as $n\to\infty$.
This completes the proof.
\end{proof}
Since the following lemma is proved in the same way as in Lemma \ref{lem:L-Ln},
we omit the proof.
\begin{lem} \label{lem:L-dense}
Let $\mathcal{L}$, $\mathcal{L}_\varepsilon$, and $\mathcal{L}_\varepsilon'$ be compact metric spaces
and $q_\varepsilon : \mathcal{L}_\varepsilon \to \mathcal{L}_\varepsilon'$ be
an $\varepsilon$-isometric map for every positive number $\varepsilon$.
If $d_{GH}(\mathcal{L}_\varepsilon,\mathcal{L}) \le \varepsilon$ and if
$d_{GH}(\mathcal{L}_\varepsilon',\mathcal{L}) \le \varepsilon$,
then $q_\varepsilon(\mathcal{L}_\varepsilon)$
is $\varepsilon'$-dense in $\mathcal{L}_\varepsilon'$,
where $\varepsilon' = \varepsilon'(\varepsilon)$ is a function
with $\lim_{\varepsilon\to 0} \varepsilon' = 0$.
\end{lem}
\begin{lem} \label{lem:cri-asymp-conc}
Let $X_n$ and $Y_m$, $m,n=1,2,\dots$, be mm-spaces
such that $\{Y_m\}$ approximates a concentrated pyramid $\mathcal{P}$.
Assume that
\begin{enumerate}
\item $\mathcal{M}(X_n;N)$ Hausdorff converges to $\mathcal{M}(\mathcal{P};N)$ as $n\to\infty$ for any natural number $N$,
\item $\mathcal{L}_1(X_n)$ Gromov-Hausdorff converges to $\mathcal{L}_1(\mathcal{P})$
as $n\to\infty$.
\end{enumerate}
Then, as $m,n\to\infty$,
both $X_n$ and $Y_m$ asymptotically concentrate to
some common ideal mm-space $\bar{X} \in \partial\mathcal{X}$
with $\mathcal{P} = \mathcal{P}_{\bar{X}}$.
\end{lem}
\begin{proof}
Let $\varepsilon > 0$ be any number.
There is a number $m_0 = m_0(\varepsilon)$ such that
$d_{GH}(\mathcal{L}_1(Y_m),\mathcal{L}_1(\mathcal{P})) < \varepsilon$ for any $m \ge m_0$.
Let $m$ be any number with $m \ge m_0$.
Let $N = N(Y_m,\varepsilon)$ be as in Lemma \ref{lem:M-p}.
By the assumption, there is a number $n_0 = n_0(m,\varepsilon)$
such that $d_{GH}(\mathcal{L}_1(X_n),\mathcal{L}_1(\mathcal{P})) < \varepsilon$ and
$d_H(\mathcal{M}(X_n;N),\mathcal{M}(\mathcal{P};N)) < \varepsilon$ for any $n \ge n_0$.
We therefore have
\[
\mathcal{M}(Y_m;N) \subset \mathcal{M}(\mathcal{P};N) \subset B_\varepsilon(\mathcal{M}(X_n;N)).
\]
By Lemma \ref{lem:M-p},
there is a Borel map $p_{mn} : X_n \to Y_m$ that is $1$-Lipschitz
up to $5\varepsilon$ and satisfies
$d_P((p_{mn})_*\mu_{X_n},\mu_{Y_m}) \le 15\varepsilon$.
Using Lemma \ref{lem:1-Lip-up-to-Lip1}(2) we have
$p_{mn}^*\mathcal{L}{\it ip}_1(Y_m) \subset B_{5\varepsilon}(\mathcal{L}{\it ip}_1(X_n))$
and hence $p_{mn}^*\mathcal{L}_1(Y_m) \subset B_{5\varepsilon}(\mathcal{L}_1(X_n))$.
Lemma \ref{lem:pb-me-di}(2) tells us that
$p_{mn}^* : \mathcal{L}_1(Y_m) \to B_{5\varepsilon}(\mathcal{L}_1(X_n))$
is $30\varepsilon$-isometric.
Let $\pi : B_{5\varepsilon}(\mathcal{L}_1(X_n)) \to \mathcal{L}_1(X_n)$ be a nearest point projection.
Since this is $5\varepsilon$-isometric,
$\pi \circ p_{mn}^* : \mathcal{L}_1(Y_m) \to \mathcal{L}_1(X_n)$ is $35\varepsilon$-isometric.
We apply Lemma \ref{lem:L-dense}
for $\mathcal{L} := \mathcal{L}_1(\mathcal{P})$, $\mathcal{L}_\varepsilon := \mathcal{L}_1(Y_m)$,
$\mathcal{L}_\varepsilon' := \mathcal{L}_1(X_n)$, and $q_\varepsilon := \pi\circ p_{mn}^*$.
Then, $\pi \circ p_{mn}^*(\mathcal{L}_1(Y_m))$ is $\varepsilon'$-dense in $\mathcal{L}_1(X_n)$,
where $\varepsilon' \to 0$ as $\varepsilon \to 0$.
This implies that
\[
\lim_{m\to\infty}\limsup_{n\to\infty} d_H(p_{mn}^*\mathcal{L}_1(Y_m),\mathcal{L}_1(X_n)) = 0,
\]
which together with Lemma \ref{lem:enforce-conc} proves
\[
\lim_{m\to\infty}\limsup_{n\to\infty} \dconc(X_n,Y_m) = 0.
\]
The two sequences $\{X_n\}$ and $\{Y_m\}$ are both $\dconc$-Cauchy
and asymptotically concentrate to some common
ideal mm-space $\bar{X} \in \partial\mathcal{X}$.
The pyramid $\mathcal{P}_{Y_m}$ converges weakly to both $\mathcal{P}$ and $\mathcal{P}_{\bar{X}}$,
so that we have $\mathcal{P} = \mathcal{P}_{\bar{X}}$.
This completes the proof.
\end{proof}
\begin{cor} \label{cor:inj-pyramid}
If $\bar{X},\bar{X}' \in \partial\mathcal{X}$ are two ideal mm-spaces
with $\mathcal{P}_{\bar{X}} = \mathcal{P}_{\bar{X}'}$, then we have $\bar{X} = \bar{X}'$.
Namely, the map $\iota : \mathcal{X} \to \Pi$ is injective.
\end{cor}
\begin{proof}
Assume that $\mathcal{P}_{\bar{X}} = \mathcal{P}_{\bar{X}'}$.
There are mm-spaces $X_n$ and $X_n'$, $n=1,2,\dots$,
such that $\{X_n\}$ and $\{X_n'\}$ asymptotically concentrate to
$\bar{X}$ and $\bar{X}'$, respectively.
$\mathcal{L}_1(X_n)$ and $\mathcal{L}_1(X_n')$ Gromov-Hausdorff converges to
$\mathcal{L}_1(\bar{X})$ and $\mathcal{L}_1(\bar{X}')$, respectively.
Moreover, $\mathcal{L}_1(\bar{X})$ and $\mathcal{L}_1(\bar{X}')$ are isometric to
$\mathcal{L}_1(\mathcal{P}_{\bar{X}})$ and $\mathcal{L}_1(\mathcal{P}_{\bar{X}'})$ respectively
by Proposition \ref{prop:cL-barX}.
By `(1)$\implies$(2)' of Theorem \ref{thm:cri-asymp-conc},
$\mathcal{M}(X_n;N)$ and $\mathcal{M}(X_n';N)$ both Hausdorff converge
to $\mathcal{M}(\mathcal{P}_{\bar{X}};N) = \mathcal{M}(\mathcal{P}_{\bar{X}'};N)$ as $n\to\infty$
for any $N$.
Take a sequence of mm-spaces $Y_n$, $n=1,2,\dots$,
that approximates the pyramid $\mathcal{P}_{\bar{X}} = \mathcal{P}_{\bar{X}'}$.
It then follows from Lemma \ref{lem:cri-asymp-conc}
that $X_n$, $X_n'$, and $Y_m$ all asymptotically concentrate
to a common ideal mm-space as $m,n\to\infty$,
so that we obtain $\bar{X} = \bar{X}'$.
This completes the proof.
\end{proof}
\begin{proof}[Proof of `{\rm(2)} $\implies$ {\rm(1)}'
of Theorem \ref{thm:cri-asymp-conc}]
We assume (2).
It then follows from Lemma \ref{lem:KNcP}
that $\mathcal{L}_1(X_n)$ Gromov-Hausdorff converges to $\mathcal{L}_1(\mathcal{P}_{\bar{X}})$.
Taking a sequence of mm-spaces $Y_m$, $m=1,2,\dots$,
approximating $\mathcal{P}_{\bar{X}}$,
we apply Lemma \ref{lem:cri-asymp-conc} for $\mathcal{P} := \mathcal{P}_{\bar{X}}$
to prove that
$X_n$ and $Y_m$ both asymptotically concentrate to a common (ideal) mm-space
$\bar{X}' \in \bar{\mathcal{X}}$ with $\mathcal{P}_{\bar{X}} = \mathcal{P}_{\bar{X}'}$.
By Corollary \ref{cor:inj-pyramid}, we have $\bar{X} = \bar{X}'$.
This completes the proof of Theorem \ref{thm:cri-asymp-conc}.
\end{proof}
The following is clear.
\begin{cor} \label{cor:Ym}
If a sequence of mm-spaces approximates $\mathcal{P}_{\bar{X}}$
for an ideal mm-space $\bar{X}$, then
it asymptotically concentrates to $\bar{X}$.
\end{cor}
\begin{lem} \label{lem:conv-pyramid-MNR}
Let $\mathcal{P}$ and $\mathcal{P}_n$, $n=1,2,\dots$, be pyramids,
$N$ a natural number, and $R > 0$ a real number.
If $\mathcal{P}_n\cap\mathcal{X}(N,R)$ Hausdorff converges to $\mathcal{P}\cap\mathcal{X}(N,R)$
as $n\to\infty$ with respect to the box metric $\square$ on $\mathcal{X}$, then
$\mathcal{M}(\mathcal{P}_n;N,R)$ Hausdorff converges to $\mathcal{M}(\mathcal{P};N,R)$ as $n\to\infty$.
\end{lem}
\begin{proof}
Assume that $\mathcal{P}_n\cap\mathcal{X}(N,R)$ Hausdorff converges to $\mathcal{P}\cap\mathcal{X}(N,R)$
as $n\to\infty$.
By the compactness of $B^N_R$ and by Lemma \ref{lem:dH}(2),
there is a sequence $n_i \to \infty$ such that
$\mathcal{M}(\mathcal{P}_{n_i};N,R)$ Hausdorff converges to some compact subset
$\mathcal{A} \subset \mathcal{M}(N,R)$ as $i\to\infty$.
Let us prove $\mathcal{A} = \mathcal{M}(\mathcal{P};N,R)$.
By Proposition \ref{prop:box-di},
\begin{align*}
&d_H(\mathcal{P}_{n_i}\cap\mathcal{X}(N,R),\{(\field{R}^N,\|\cdot\|_\infty,\mu)\mid\mu\in\mathcal{A}\})\\
&\le 2\, d_H(\mathcal{M}(\mathcal{P}_{n_i};N,R),\mathcal{A}) \to 0 \quad\text{as $n\to\infty$},
\end{align*}
which together with the assumption implies that
\[
\{(\field{R}^N,\|\cdot\|_\infty,\mu)\mid\mu\in\mathcal{A}\} = \mathcal{P}\cap\mathcal{X}(N,R).
\]
Since Lemma \ref{lem:perfect-lim} implies the perfectness of $\mathcal{A}$
on $B^N_R$, we obtain $\mathcal{A} = \mathcal{M}(\mathcal{P};N,R)$.
Since this holds for any convergent subsequence $\{\mathcal{M}(\mathcal{P}_{n_i};N,R)\}$
of $\{\mathcal{M}(\mathcal{P}_n;N,R)\}$,
the measurement $\mathcal{M}(\mathcal{P}_n;N,R)$ Hausdorff converges to $\mathcal{M}(\mathcal{P};N,R)$
as $n\to\infty$. This completes the proof.
\end{proof}
\begin{cor} \label{cor:asymp-conc-pyramid}
Let $\{X_n\}$ be a sequence of mm-spaces
and $\bar{X} \in \partial\mathcal{X}$ an ideal mm-space.
Then, the following {\rm(1)} and {\rm(2)} are equivalent to each other.
\begin{enumerate}
\item $X_n$ asymptotically concentrates to $\bar{X}$ as $n\to\infty$.
\item $\mathcal{P}_{X_n}$ converges weakly to $\mathcal{P}_{\bar{X}}$ as $n\to\infty$.
\end{enumerate}
\end{cor}
\begin{proof}
`(1) $\implies$ (2)' is clear by the definition of $\mathcal{P}_{\bar{X}}$.
We prove `(2) $\implies$ (1)'.
If $\mathcal{P}_{X_n}$ converges weakly to $\mathcal{P}_{\bar{X}}$ as $n\to\infty$,
then, by Lemmas \ref{lem:pyramid-conv-dH} and \ref{lem:conv-pyramid-MNR},
$\mathcal{M}(X_n;N,R)$ Hausdorff converges to $\mathcal{M}(\mathcal{P}_{\bar{X}};N,R)$ as $n\to\infty$
for any $N$ and $R$.
By Theorem \ref{thm:cri-asymp-conc} we obtain (1).
\end{proof}
\begin{prop} \label{prop:concentrated}
For a given pyramid $\mathcal{P}$,
the following {\rm(1)} and {\rm(2)} are equivalent to each other.
\begin{enumerate}
\item $\mathcal{P}$ is concentrated.
\item There exists an {\rm(}ideal{\rm)} mm-space $\bar{X} \in \bar{\mathcal{X}}$ such that
$\mathcal{P} = \mathcal{P}_{\bar{X}}$.
\end{enumerate}
\end{prop}
\begin{proof}
`(2) $\implies$ (1)' follows from Corollary \ref{cor:concentrated}.
We prove `(1) $\implies$ (2)'.
Let $\{X_n\}$ be a sequence of mm-spaces that approximates $\mathcal{P}$.
$\mathcal{L}_1(X_n)$ Gromov-Hausdorff converges to $\mathcal{L}_1(\mathcal{P})$.
Since $\mathcal{P}_{X_n}$ converges weakly to $\mathcal{P}$ as $n\to\infty$,
and by Lemmas \ref{lem:pyramid-conv-dH} and \ref{lem:conv-pyramid-MNR},
$\mathcal{M}(X_n;N,R)$ Hausdorff converges to $\mathcal{M}(\mathcal{P};N,R)$ as $n\to\infty$
for any $N$ and $R$.
We apply Lemma \ref{lem:MR}, where the assumption of the lemma
follows from
Corollary \ref{cor:sup-diam-cP}.
Thus,
$\mathcal{M}(X_n;N)$ Hausdorff converges to $\mathcal{M}(\mathcal{P};N)$ as $n\to\infty$
for any $N$.
Lemma \ref{lem:cri-asymp-conc} proves (2).
\end{proof}
Proposition \ref{prop:conc-pyramid} and Corollary \ref{cor:asymp-conc-pyramid}
extend to the following
\begin{thm} \label{thm:conc-pyramid}
Let $\{\bar{X}_n\}_{n=1}^\infty \subset \bar{\mathcal{X}}$ and
$\bar{X} \in \bar\mathcal{X}$.
Then, the following {\rm(1)} and {\rm(2)} are equivalent to each other.
\begin{enumerate}
\item $\bar{X}_n$ $\dconc$-converges to $\bar{X}$ as $n\to\infty$.
\item $\mathcal{P}_{\bar{X}_n}$ converges weakly to $\mathcal{P}_{\bar{X}}$ as $n\to\infty$.
\end{enumerate}
\end{thm}
\begin{proof}
We prove `(1)$\implies$(2)'.
Suppose that
$\bar{X}_n$ $\dconc$-converges to $\bar{X}$
and $\mathcal{P}_{\bar{X}_n}$ does not converge weakly to $\mathcal{P}_{\bar{X}}$ as $n\to\infty$.
By replacing $\{\bar{X}_n\}$ with a subsequence,
we assume that $\mathcal{P}_{\bar{X}_n}$ converges weakly to a pyramid $\mathcal{P}$
with $\mathcal{P} \neq \mathcal{P}_{\bar{X}}$ as $n\to\infty$.
For each $n$ there is a sequence of mm-spaces
$X_{n,m} \in \mathcal{P}_{\bar{X}_n}$, $m=1,2,\dots$, that $\dconc$-converges
to $\bar{X}_n$.
Corollary \ref{cor:asymp-conc-pyramid} implies that
$\mathcal{P}_{X_{n,m}}$ converges weakly to $\mathcal{P}_{\bar{X}_n}$ as $m\to\infty$.
There is a sequence of numbers $m(n)$, $n=1,2,\dots$,
such that, as $n\to\infty$, $\mathcal{P}_{X_{n,m(n)}}$ converges weakly to $\mathcal{P}$
and $X_{n,m(n)}$ $\dconc$-converges to $\bar{X}$.
By Corollary \ref{cor:asymp-conc-pyramid}
and Proposition \ref{prop:conc-pyramid}, we have $\mathcal{P} = \mathcal{P}_{\bar{X}}$,
which is a contradiction.
We prove `(2)$\implies$(1)'.
Suppose that
$\mathcal{P}_{\bar{X}_n}$ converges weakly to $\mathcal{P}_{\bar{X}}$
and $\bar{X}_n$ does not $\dconc$-converges to $\bar{X}$ as $n\to\infty$.
Replacing $\{\bar{X}_n\}$ with a subsequence,
we assume that $\dconc(\bar{X}_n,\bar{X})$, $n=1,2,\dots$, are
bounded away from zero.
There is a natural number $n(k)$ for each $k$
such that $\rho(\mathcal{P}_{\bar{X}_{n(k)}},\mathcal{P}_{\bar{X}}) < 1/k$.
By Corollary \ref{cor:Ym},
there is an mm-space $Y_k \in \mathcal{P}_{\bar{X}_{n(k)}}$
for each $k$ such that
$\rho(\mathcal{P}_{\bar{X}_{n(k)}},\mathcal{P}_{Y_k}) \le \dconc(\bar{X}_{n(k)},Y_k) < 1/k$.
A triangle inequality yields that
$\rho(\mathcal{P}_{Y_k},\mathcal{P}_{\bar{X}}) < 2/k$ for any $k$
and $\mathcal{P}_{Y_k}$ converges weakly to $\mathcal{P}_{\bar{X}}$ as $k\to\infty$.
By Corollary \ref{cor:asymp-conc-pyramid},
as $k\to\infty$, $Y_k$ $\dconc$-converges to $\bar{X}$
and therefore $\bar{X}_{n(k)}$ $\dconc$-converges to $\bar{X}$,
which is a contradiction.
This completes the proof.
\end{proof}
We finally obtain the main theorem in this book.
\begin{thm}[Embedding theorem] \label{thm:emb-pyramid}
\index{embedding theorem}
The map
\[
\iota : \bar{\mathcal{X}} \ni \bar{X} \longmapsto \mathcal{P}_{\bar{X}} \in \Pi
\]
is a topological embedding map.
The space $\Pi$ of pyramids is a compactification of $\mathcal{X}$ and $\bar{\mathcal{X}}$.
\end{thm}
\begin{proof}
The first part follows from Corollary \ref{cor:inj-pyramid}
and Theorem \ref{thm:conc-pyramid}.
It follows from Lemma \ref{lem:approx-pyramid}
that $\iota(\mathcal{X})$ is dense in $\Pi$.
This completes the proof.
\end{proof}
\begin{rem}
Since $\mathcal{X}$ itself is a non-concentrated pyramid,
it does not belongs to $\iota(\bar{\mathcal{X}})$.
In particular, $\iota(\bar{\mathcal{X}})$ is a proper subset of $\Pi$.
As is seen in Example \ref{ex:prod-sph},
$\mathcal{X}$ is a proper subset of $\bar{\mathcal{X}}$.
We have
\[
\iota(\mathcal{X}) \subsetneq \iota(\bar{\mathcal{X}}) \subsetneq \Pi.
\]
and $\iota(\bar{\mathcal{X}})$ is not a closed subset of $\Pi$.
\end{rem}
\begin{prop}
The map $\mathcal{L}_1 : \bar{\mathcal{X}} \to \mathcal{H}$ is proper
with respect to $\dconc$ and $d_{GH}$.
\end{prop}
\begin{proof}
We take a sequence of (ideal) mm-spaces $\bar{X}_n \in \bar\mathcal{X}$,
$n=1,2,\dots$,
such that $\mathcal{L}_1(\bar{X}_n)$ Gromov-Hausdorff converges to
a compact metric space $\mathcal{L}$ as $n\to\infty$.
It suffices to prove that $\{\bar{X}_n\}$ has a subsequence
that is $\dconc$-convergent in $\bar{\mathcal{X}}$.
Theorem \ref{thm:cpt-pyramid} implies that,
by replacing $\{\bar{X}_n\}$ with a subsequence,
the pyramid $\mathcal{P}_{\bar{X}_n}$ converges weakly to a pyramid $\mathcal{P}$
as $n\to\infty$.
By Proposition \ref{prop:concentrated} and
Theorem \ref{thm:conc-pyramid},
it suffices to prove that $\mathcal{P}$ is concentrated.
Take any sequence $\{Y_i\}_{i=1}^\infty \subset \mathcal{P}$ and fix it.
For each $i$, there are a number $n(i)$ and an mm-space
$X_i' \in \mathcal{P}_{\bar{X}_{n(i)}}$ such that $\square(Y_i,X_i') \le 1/i$.
Lemma \ref{lem:isom-L} implies that
$\mathcal{L}_1(X_i')$ is isometrically embedded into $\mathcal{L}_1(\bar{X}_{n(i)})$.
Since $\mathcal{L}_1(\bar{X}_{n(i)})$ Gromov-Hausdorff converges to $\mathcal{L}$ as $i\to\infty$,
$\{\mathcal{L}_1(X_i')\}$ has a subsequence converging to a compact subset
$\mathcal{L}' \subset \mathcal{L}$. We replace $\{\mathcal{L}_1(X_i')\}$ with such a subsequence.
We have
$d_{GH}(\mathcal{L}_1(Y_i),\mathcal{L}_1(X_i')) \le \dconc(Y_i,X_i')
\le \square(Y_i,X_i') \le 1/i$ and hence
$\mathcal{L}_1(Y_i)$ Gromov-Hausdorff converges to $\mathcal{L}'$ as $i\to\infty$.
Thus, $\mathcal{P}$ is concentrated.
This completes the proof.
\end{proof}
\section{Infinite product, II}
In this section, we see asymptotically concentrating
product spaces.
\begin{defn}[L\'evy radius]
\index{Levy radius@L\'evy radius}
\index{LeRad@$\LeRad(X;-\kappa)$}
Let $X$ be an mm-space and $\kappa > 0$ a real number.
The \emph{L\'evy radius of $X$} is defined to be
\begin{align*}
\LeRad(X;-\kappa)
:= \inf\{\; & \rho > 0 \mid \mu_X(|f-m_f|> \rho) \le \kappa\\
&\text{for any $1$-Lipschitz function $f : X \to \field{R}$}\;\},
\end{align*}
where $m_f$ is the L\'evy mean of $f$.
\end{defn}
\begin{lem} \label{lem:LeRad-ObsDiam}
Let $X$ be an mm-space.
For any $\kappa$ with $0 < \kappa < 1/2$, we have
\[
\LeRad(X;-\kappa) \le \ObsDiam(X;-\kappa).
\]
\end{lem}
\begin{proof}
Assume that $\ObsDiam(X;-\kappa) < \varepsilon$ for a real number
$\varepsilon$.
For any $1$-Lipschitz function $f : X \to \field{R}$,
there is a closed interval $A \subset \field{R}$ such that
$f_*\mu_X(A) \ge 1-\kappa$ and $\diam A < \varepsilon$.
By $\kappa < 1/2$, we have $f_*\mu_X(A) > 1/2$,
so that any median of $f$ belongs to $A$ and
so does $m_f$.
Since $A \subset \{\;y\in\field{R} \mid |y-m_f| \le \varepsilon\;\}$,
we have
\[
1-\kappa \le f_*\mu_X(A) \le \mu_X(|f-m_f| \le \varepsilon)
\]
and so
\[
\mu_X(|f-m_f| > \varepsilon) \le \kappa,
\]
which implies $\LeRad(X;-\kappa) \le \varepsilon$.
This completes the proof.
\end{proof}
\begin{prop} \label{prop:prod-dconc}
Let $Y$ and $Z$ be two mm-spaces and let $1 \le p \le +\infty$.
We equip the product space $X := Y \times Z$ with the $d_{l_p}$ metric.
If $\ObsDiam(Z) < 1/2$, then we have
\[
\dconc(X,Y) \le \ObsDiam(Z).
\]
\end{prop}
\begin{proof}
Assume that $\ObsDiam(Z) < \varepsilon < 1/2$.
Letting $p : X \to Y$ be the projection,
we shall prove that
$p$ enforces $\varepsilon$-concentration of $X$ to $Y$.
Since $p$ is $1$-Lipschitz continuous,
we have $p^*\mathcal{L}{\it ip}_1(Y) \subset \mathcal{L}{\it ip}_1(X)$.
It suffices to prove that $\mathcal{L}{\it ip}_1(X) \subset B_\varepsilon(p^*\mathcal{L}{\it ip}_1(Y))$.
Take any function $f \in \mathcal{L}{\it ip}_1(X)$.
Since $f : Y \times Z \to \field{R}$ is $1$-Lipschitz with respect to $d_{l_p}$,
for any point $y \in Y$, the function
$f(y,\cdot) : Z \to \field{R}$ is $1$-Lipschitz continuous.
Denote by $\underline{m}(y)$ the minimum of medians of $f(y,\cdot)$.
Let us prove that $\underline{m}(y)$ is $1$-Lipschitz continuous
in $y \in Y$.
In fact, for any two points $y_1,y_2 \in Y$,
we have $f(\cdot,y_2) - d_Y(y_1,y_2) \le f(\cdot,y_1)$ and hence
\[
\mu_X\{f(\cdot,y_2) \le \underline{m}(y_1) + d_Y(y_1,y_2)\}
\ge \mu_X\{f(\cdot,y_1) \le \underline{m}(y_1)\} \ge \frac{1}{2},
\]
which together with the minimality of $\underline{m}(y_2)$
among all medians of $f(\cdot,y_2)$
proves that $\underline{m}(y_1) + d_Y(y_1,y_2) \ge \underline{m}(y_2)$.
Exchanging $y_1$ and $y_2$ we have
$\underline{m}(y_2) + d_Y(y_1,y_2) \ge \underline{m}(y_1)$.
Therefore, $\underline{m}$ is $1$-Lipschitz continuous.
In the same way, we see that the maximum of medians of $f(y,\cdot)$,
say $\overline{m}(y)$, is also $1$-Lipschitz continuous in $y \in Y$.
Define a map $\bar{f} : X = Y \times Z \to \field{R}$ by
\[
\bar{f}(y,z) := \frac{\overline{m}(y)-\underline{m}(y)}{2}
\]
for $(y,z) \in Y \times Z$.
$\bar{f}$ is $1$-Lipschitz with respect to $d_{l_p}$ and belongs to
$p^*\mathcal{L}{\it ip}_1(Y)$.
Since $\ObsDiam(Z) < \varepsilon < 1/2$ implies
$\ObsDiam(Z;-\varepsilon) < \varepsilon < 1/2$,
we have $\LeRad(Z;-\varepsilon) < \varepsilon$
by Lemma \ref{lem:LeRad-ObsDiam} and therefore
\[
\mu_Z(|f(y,\cdot)-\bar{f}(y,\cdot)| > \varepsilon)
\le \LeRad(Z;-\varepsilon) < \varepsilon.
\]
By Fubini's theorem,
\begin{align*}
\mu_X(|f-\bar{f}| > \varepsilon)
&= \int_Y \mu_Z(|f(y,\cdot)-\bar{f}(y,\cdot)| > \varepsilon)\;d\mu_Y(y)
\le \varepsilon,
\end{align*}
i.e., $\dKF(f,\bar{f}) \le \varepsilon$.
We obtain $\mathcal{L}{\it ip}_1(X) \subset B_\varepsilon(p^*\mathcal{L}{\it ip}_1(Y))$ and
$p$ enforces $\varepsilon$-concentration of $X$ to $Y$.
Since $p_*\mu_X = \mu_Y$, Lemma \ref{lem:enforce-conc} implies
$\dconc(X,Y) \le \varepsilon$.
This completes the proof.
\end{proof}
\begin{prop} \label{prop:prod-asympconc}
Let $F_n$, $n=1,2,\dots$, be mm-spaces and let $1 \le p \le +\infty$.
We define
\[
X_n := F_1 \times F_2 \times \dots \times F_n,
\qquad\Phi_{ij} := F_i \times F_{i+1} \times \dots \times F_j.
\]
and equip $X_n$ and $\Phi_{ij}$ with the product measure and
the $d_{l_p}$ metrics.
If $\ObsDiam(\Phi_{ij}) \to 0$ as $i,j \to +\infty$,
then $\{X_n\}$ asymptotically concentrates.
\end{prop}
\begin{proof}
Let $i < j$.
Since $X_j = X_i \times \Phi_{i+1,j}$,
Proposition \ref{prop:prod-dconc} proves that
$\dconc(X_i,X_j) \le \ObsDiam(\Phi_{ij})$
for any $i$ and $j$ large enough.
This completes the proof.
\end{proof}
Let $X$ and $Y$ be two compact Riemannian manifolds
and $X \times Y$ the Riemannian product of $X$ and $Y$.
Denote by $\sigma(\Delta_X)$ the spectrum of the Laplacian $\Delta_X$ on $X$.
It is well-known that
\[
\sigma(\Delta_{X\times Y}) = \{\;\lambda+\mu \mid
\lambda \in \sigma(\Delta_X),\;\mu \in \sigma(\Delta_Y)\;\},
\]
and, for any $i \ge 0$,
\[
m_i(X \times Y) = \sum_{j,k\, :\,\lambda_i(X \times Y) = \lambda_j(X) + \lambda_k(Y)}
(m_j(X) + m_k(Y) - 1),
\]
where $m_i(X)$ denotes the multiplicity of $\lambda_i(X)$, i.e.,
the number of $j$'s with $\lambda_i(X) = \lambda_j(X)$.
In particular we have
\[
\lambda_1(X \times Y) = \min\{\lambda_1(X),\lambda_1(Y)\}.
\]
\begin{cor} \label{cor:prod-asympconc}
Let $F_n$, $n=1,2,\dots$, be compact Riemannian manifolds
such that $\lambda_1(F_n)$ is divergent to infinity
as $n\to\infty$.
Let $X_n := F_1 \times F_2 \times \dots \times F_n$
be the Riemannian product space.
Then, $\{X_n\}$ asymptotically concentrates.
\end{cor}
\begin{proof}
Let $i < j$.
As $i,j \to \infty$,
$\lambda_1(\Phi_{ij}) = \min_{k=i}^j \lambda_1(F_k)$
is divergent to infinity.
Corollary \ref{cor:ObsDiam-spec} implies that
$\ObsDiam(\Phi_{ij}) \to 0$ as $i,j \to \infty$.
The corollary follows from Proposition \ref{prop:prod-asympconc}.
\end{proof}
\begin{ex} \label{ex:prod-sph}
Let
\[
X_n := S^1 \times S^2 \times \dots \times S^n
\]
be the Riemannian product of unit spheres in Euclidean spaces.
Since $\lambda_1(S^n) = n$, Corollary \ref{cor:prod-asympconc}
proves that $\{X_n\}$ asymptotically concentrates.
The infinite product space
\[
X_\infty := \prod_{n=1}^\infty S^n
\]
is not really an mm-space, but is considered to be an ideal mm-space.
\end{ex}
\begin{prop} \label{prop:prod-non-conc}
Let $F$ be an mm-space and let $1 \le p \le +\infty$.
Then, $\{(F^n,d_{l_p},\mu_F^{\otimes n})\}_{n=1}^\infty$ does not asymptotically
concentrate unless $F$ consists of a single point.
\end{prop}
Note that $\{(F^n,d_{l_p},\mu_F^{\otimes n})\}_{n=1}^\infty$ is asymptotic.
\begin{proof}
Assume that $F$ contains at least two different points.
We then find a non-constant $1$-Lipschitz function $\varphi : F \to \field{R}$.
Let $\varphi_{n,i} : F^n \to \field{R}$, $i=1,2,\dots,n$, be the functions defined by
\[
\varphi_{n,i}(x_1,x_2,\dots,x_n) := \varphi(x_i),
\qquad (x_1,x_2,\dots,x_n) \in F^n.
\]
Each $\varphi_{n,i}$ is $1$-Lipschitz continuous with respect to
$d_{l_p}$.
For any different $i$ and $j$,
\begin{align*}
d_{KF}^{\mu_F^{\otimes n}}([\varphi_{n,i}],[\varphi_{n,j}])
= d_{KF}^{\mu_F^{\otimes 2}}([\varphi_{2,1}],[\varphi_{2,2}])
= d_{KF}^{\mu_F^{\otimes 2}}(\varphi_{2,1},\varphi_{2,2}+c)
\end{align*}
for some real number $c$ (see Lemma \ref{lem:me-const}).
If $d_{KF}^{\mu_F^{\otimes 2}}(\varphi_{2,1},\varphi_{2,2}+c) = 0$ were to hold,
then $\varphi_{2,1} = \varphi_{2,2} + c$ $\mu_F^{\otimes 2}$-almost everywhere
and hence $\varphi(x_1) = \varphi(x_2) + c$ for all $x_1,x_2 \in F$,
which is a contradiction.
Thus, $d_{KF}^{\mu_F^{\otimes n}}([\varphi_{n,i}],[\varphi_{n,j}])$ is a positive
constant, say $\varepsilon_0$,
independent of $n$, $i$, and $j$ with $i \neq j$.
This implies that $\Cap_{\varepsilon_0/2}(\mathcal{L}_1(F^n,d_{l_p},\mu_F^{\otimes n})) \ge n$
and $\{\mathcal{L}_1(F^n,d_{l_p},\mu_F^{\otimes n})\}_{n=1}^\infty$
is not $d_{GH}$-precompact by Lemma \ref{lem:dGH-precpt}.
By Lemma \ref{lem:dGH-dconc},
$\{(F^n,d_{l_p},\mu_F^{\otimes n})\}_{n=1}^\infty$ does not asymptotically
concentrate.
The proof is completed.
\end{proof}
\section{Spheres and Gaussians}
\label{sec:Spheres-Gaussians}
\begin{defn}[Gaussian space]
\index{Gaussian space} \index{Gamma-n@$\Gamma^n$}
Let $\Gamma^n := (\field{R}^n,\|\cdot\|_2,\gamma^n)$.
We call the mm-space $\Gamma^n$
the \emph{$n$-dimensional} (\emph{standard}) \emph{Gaussian
space}.
\end{defn}
Recall that the $n$-dimensional Gaussian measure $\gamma^n$
coincides with the $n^{th}$ power of the one-dimensional
Gaussian measure $\gamma^1$, so that
we have the monotonicity of the Gaussian spaces
\[
\Gamma^1 \prec \Gamma^2 \prec \cdots \prec \Gamma^n \prec \cdots.
\]
Therefore, the sequence $\{\Gamma^n\}_{n=1}^\infty$ is asymptotic
and converges weakly to the $\square$-closure of the union of
$\mathcal{P}_{\Gamma^n}$, $n=1,2,\dots$, say $\mathcal{P}_{\Gamma^\infty}$.
Note that $\{\Gamma^n\}_{n=1}^\infty$ does not asymptotically
concentrate by Proposition \ref{prop:prod-non-conc}.
Although the infinite-dimensional Gaussian space $\Gamma^\infty
:= (\field{R}^\infty,\|\cdot\|_2,\gamma^\infty)$ is not an mm-space
(the infinite-dimensional Gaussian measure $\gamma^\infty$
is not a Borel measure with respect to $\|\cdot\|_2$;
cf.~\cite{Bog:Gm}*{\S 2.3}),
we consider the pyramid $\mathcal{P}_{\Gamma^\infty}$
as a substitute for $\Gamma^\infty$.
\begin{defn}[Virtual infinite-dimensional Gaussian space]
\index{virtual infinite-dimensional Gaussian space}
\index{P-Gamma-infty@$\mathcal{P}_{\Gamma^\infty}$}
We call $\mathcal{P}_{\Gamma^\infty}$ the \emph{virtual infinite-dimensional}
(\emph{standard}) \emph{Gaussian space}.
\end{defn}
In this section we prove
\begin{thm}
\label{thm:sphere-Gaussian}
The pyramid $\mathcal{P}_{S^n(\sqrt{n})}$ associated with $S^n(\sqrt{n})$
converges weakly to the virtual infinite-dimensional Gaussian space
$\mathcal{P}_{\Gamma^\infty}$ as $n\to\infty$,
where $S^n(\sqrt{n})$ is equipped with the Euclidean distance function.
\end{thm}
We need a lemma.
\begin{lem} \label{lem:sphere-Gaussian}
For any real number $\theta$ with $0 < \theta < 1$, we have
\[
\lim_{n\to\infty} \gamma^{n+1}\{\;x\in\field{R}^{n+1}
\mid \|x\|_2 \le \theta\sqrt{n}\;\} = 0.
\]
\end{lem}
\begin{proof}
Considering the poler coordinates on $\field{R}^n$,
we see that
\[
\gamma^{n+1}\{\;x\in\field{R}^{n+1} \mid \|x\|_2 \le r\;\}
= \frac{\int_0^r t^ne^{-t^2/2}\,dt}{\int_0^\infty t^ne^{-t^2/2}\,dt}.
\]
Integrating the both sides of $(\log(t^n e^{-t^2/2}))'' = -n/t^2-1 \le -1$
over $[\,t,\sqrt{n}\,]$ with $0 < t \le \sqrt{n}$ yields
\[
-(\log(t^n e^{-t^2/2}))'
= (\log(t^n e^{-t^2/2}))'|_{t=\sqrt{n}} - (\log(t^n e^{-t^2/2}))'
\le t-\sqrt{n}.
\]
Integrating this again over $[\,t,\sqrt{n}\,]$ implies
\[
\log(t^ne^{-t^2/2}) - \log(n^{n/2}e^{-n/2}) \le -\frac{(t-\sqrt{n})^2}{2},
\]
so that $t^ne^{-t^2/2} \le n^{n/2}e^{-n/2}e^{-(t-\sqrt{n})^2/2}$ and then,
for any $r$ with $0 \le r \le \sqrt{n}$,
\[
\int_0^{\sqrt{n}-r} t^ne^{-t^2/2}\,dt
\le n^{n/2}e^{-n/2} \int_r^{\sqrt{n}} e^{-t^2/2}\,dt
\le n^{n/2}e^{-n/2}e^{-r^2/2}.
\]
By setting
\[
I_n := \int_0^\infty t^ne^{-t^2/2}\,dt,
\]
Stirling's approximation implies
\[
I_n = 2^{\frac{n-1}{2}} \int_0^\infty s^{\frac{n-1}{2}} e^{-s}\,ds
\sim \sqrt{\pi} (n-1)^{\frac{n}{2}} e^{-\frac{n-1}{2}}.
\]
Therefore,
\[
\gamma^{n+1}\{\;x\in\field{R}^{n+1} \mid \|x\|_2 \le \theta\sqrt{n}\;\}
\le \frac{n^{n/2}e^{-n/2}e^{-(1-\theta)^2n/2}}{I_n} \to 0
\quad n\to\infty.
\]
This completes the proof.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:sphere-Gaussian}]
Suppose that $\mathcal{P}_{S^n(\sqrt{n})}$ does not converge weakly to
$\mathcal{P}_{\Gamma^\infty}$ as $n\to\infty$.
Then, by the compactness of $\Pi$, there is a subsequence
$\{\mathcal{P}_{S^{n_i}(\sqrt{n_i})}\}$ of $\{\mathcal{P}_{S^n(\sqrt{n})}\}$
that converges weakly to a pyramid $\mathcal{P}$ with $\mathcal{P} \neq \mathcal{P}_{\Gamma^\infty}$.
It follows from the Maxwell-Boltzmann distribution law
(Proposition \ref{prop:MB-law}) that
$\Gamma^k$ belongs to $\mathcal{P}$ for any $k$,
so that $\mathcal{P}_{\Gamma^\infty} \subset \mathcal{P}$.
We take any real number $\theta$ with $0 < \theta < 1$ and fix it.
Let $\theta\mathcal{P} := \{\;\theta X \mid X \in \mathcal{P}\;\}$.
We see that $\mathcal{P}_{S^{n_i}(\theta\sqrt{n_i})}$ converges weakly to $\theta\mathcal{P}$
as $i\to\infty$.
Define a function $f_{\theta,n} : \field{R}^{n+1} \to \field{R}^{n+1}$ by
\[
f_{\theta,n}(x) :=
\begin{cases}
\frac{\theta\sqrt{n}}{\|x\|_2} x & \text{if $\|x\|_2 > \theta\sqrt{n}$},\\
x & \text{if $\|x\|_2 \le \theta\sqrt{n}$},
\end{cases}
\]
for $x \in \field{R}^{n+1}$.
$f_{\theta,n}$ is $1$-Lipschitz continuous with respect to
the $l_2$ norm on $\field{R}^{n+1}$.
Let $\sigma_\theta^n$ be the normalized volume measure on
$S^n(\theta\sqrt{n})$. We consider $\sigma_\theta^n$
as a measure on $\field{R}^{n+1}$ via the natural embedding
$S^n(\theta\sqrt{n}) \subset \field{R}^{n+1}$.
From Lemma \ref{lem:sphere-Gaussian}, we have
\[
d_P((f_{\theta,n})_*\gamma^{n+1},\sigma_\theta^n)
\le \gamma^{n+1}\{\;x\in\field{R}^{n+1} \mid \|x\|_2 < \theta\sqrt{n}\;\}
\to 0 \ \text{as $n\to\infty$},
\]
so that the box distance between
$S_{\theta,n} := (\field{R}^{n+1},\|\cdot\|_2,(f_{\theta,n})_*\gamma^{n+1})$
and $S^n(\theta\sqrt{n})$
converges to zero as $n\to\infty$.
By Proposition \ref{prop:dconc-box} and Theorem \ref{thm:rho-dconc},
$\mathcal{P}_{S_{\theta,n_i}}$ converges weakly to $\theta\mathcal{P}$
as $i\to\infty$. Since $S_{\theta,n} \prec (\field{R}^{n+1},\|\cdot\|_2,\gamma^{n+1})$,
we have $\mathcal{P}_{S_{\theta,n}} \subset \mathcal{P}_{\Gamma^{n+1}} \subset \mathcal{P}_{\Gamma^\infty}$.
We thus obtain $\theta\mathcal{P} \subset \mathcal{P}_{\Gamma^\infty} \subset \mathcal{P}$
for any $\theta$ with $0 < \theta < 1$.
Since $\theta\mathcal{P}$ converges weakly to $\mathcal{P}$ as $\theta \to 1$,
we obtain $\mathcal{P} = \mathcal{P}_{\Gamma^\infty}$, which is a contradiction.
This completes the proof.
\end{proof}
\begin{cor} \label{cor:sphere-Gaussian}
The virtual infinite-dimensional Gaussian space $\mathcal{P}_{\Gamma^\infty}$
is non-concentrated,
or equivalently,
neither of $S^n(\sqrt{n})$ nor $\Gamma^n$
asymptotically concentrates as $n\to\infty$.
\end{cor}
\begin{proof}
Proposition \ref{prop:prod-non-conc} implies
that $\Gamma^n$ does not
asymptotically concentrates as $n\to\infty$.
By Proposition \ref{prop:concentrated},
$\mathcal{P}_{\Gamma^\infty}$ is non-concentrated.
This completes the proof.
\end{proof}
\section{Spectral concentration}
In this section, we prove that a spectrally compact sequence
of mm-spaces asymptotically concentrates
if the observable diameter is uniformly bounded above.
\begin{defn}[Gradient, energy, and $\lambda_1$]
\index{gradient} \index{energy} \index{lambda1X@$\lambda_1(X)$}
Let $X$ be an mm-space.
For a locally Lipschitz continuous function $f : X \to \field{R}$,
we define
\begin{align*}
|\grad f|(x) &:= \limsup_{\substack{y \to x\\ y\neq x}} \frac{|f(x)-f(y)|}{d_X(x,y)},
\quad x \in X,\\
\mathcal{E}(f) &:= \int_X |\grad f|^2 \; d\mu_X \quad (\le +\infty).
\end{align*}
\index{grad f@$\vert\grad f\vert$} \index{E f@$\mathcal{E}(f)$}
We also define
\[
\lambda_1(X) := \inf_f \frac{\mathcal{E}(f)}{\|f\|_{L_2}^2},
\]
where $f$ runs over all locally Lipschitz continuous functions on $X$
with $\int_X f\,d\mu_X = 0$.
\end{defn}
If $X$ is a compact Riemannian manifold, then
$\lambda_1(X)$ defined here coincides with that defined in
\S\ref{sec:spec-sep}.
\begin{defn}[Spectral compactness] \label{defn:spec-cpt}
\index{spectral compactness}
For an mm-space $X$, we denote by $\Dir_1(X)$
the set of locally Lipschitz continuous functions $f : X \to \field{R}$ with
$\mathcal{E}(f) \le 1$.
A subset $\mathcal{Y} \subset \mathcal{X}$ is said to be
\emph{spectrally compact} if
$(\Dir_1(X) \cap B^{L_2}_1(0),\|\cdot\|_{L_2})$ is compact
for each $X \in \mathcal{Y}$ and if
$\{(\Dir_1(X) \cap B^{L_2}_1(0),\|\cdot\|_{L_2})\}_{X \in \mathcal{Y}}$
is $d_{GH}$-precompact.
Here, $B^{L_2}_1(0)$ denotes the set of $L_2$ functions $f : X \to \field{R}$
with $\|f\|_{L_2} \le 1$.
We say that a pyramid is \emph{spectrally concentrated}
\index{spectrally concentrated}
if it has a $\square$-dense spectrally compact subfamily.
\end{defn}
Note that, in \cite{Gromov}*{\S 3$\frac{1}{2}$},
a spectrally concentrated pyramid is defined to be
a spectrally compact pyramid, which is slightly stronger than
Definition \ref{defn:spec-cpt}.
The reason for Definition \ref{defn:spec-cpt} is
that we assume the $\square$-closedness of a pyramid, namely
we always consider the $\square$-closure of a pyramid,
which is not spectrally compact in general even if the pyramid is
spectrally compact.
For a subset $\mathcal{Y} \subset \mathcal{X}$, we consider the condition:
\begin{equation}
\label{eq:fin-ObsDiam}
\sup_{X \in \mathcal{Y}} \ObsDiam(X;-\kappa) < +\infty
\qquad\text{for any $\kappa > 0$}.
\end{equation}
\begin{thm} \label{thm:spec-cpt}
If a subset $\mathcal{Y} \subset \mathcal{X}$ satisfies \eqref{eq:fin-ObsDiam}
and is spectrally compact,
then $\{\mathcal{L}_1(X)\}_{X\in\mathcal{Y}}$ is $d_{GH}$-precompact.
\end{thm}
We prove the theorem later.
\begin{cor} \label{cor:spec-cpt}
If a pyramid $\mathcal{P}$ spectrally concentrates and satisfies
\eqref{eq:fin-ObsDiam}, then $\mathcal{P}$ is concentrated.
\end{cor}
\begin{proof}
We have a $\square$-dense spectrally compact subfamily $\mathcal{P}' \subset \mathcal{P}$.
Theorem \ref{thm:spec-cpt} implies that
$\{\mathcal{L}_1(X)\}_{X\in\mathcal{P}'}$ is $d_{GH}$-precompact.
It follows from Lemma \ref{lem:dGH-dconc}
and Proposition \ref{prop:dconc-box} that
$\{\mathcal{L}_1(X)\}_{X\in\mathcal{P}}$ is contained in
the $d_{GH}$-closure of $\{\mathcal{L}_1(X)\}_{X\in\mathcal{P}'}$.
This completes the proof.
\end{proof}
\begin{rem}
The condition \eqref{eq:fin-ObsDiam} is necessary for
Theorem \ref{thm:spec-cpt} and Corollary \ref{cor:spec-cpt}.
In fact,
let $Y_0$ and $Y_1$ be two compact Riemannian manifolds
with diameter $\le 1$,
and $X_n$ the disjoint union of $Y_0$ and $Y_1$.
Define a metric $d_{X_n}$ on $X_n$ by
\[
d_{X_n}(x,y) :=
\begin{cases}
d_{Y_i}(x,y) &\text{for $x,y \in Y_i$, $i=0,1$,}\\
n &\text{for $x \in Y_i$ and $y \in Y_{1-i}$, $i=0,1$.}
\end{cases}
\]
and $\mu_{X_n} := (1/2)\mu_{Y_0} + (1/2)\mu_{Y_1}$.
Then each $X_n$ is an mm-space.
We see that $\{X_n\}$ is spectrally compact,
but $\{\mathcal{L}_1(X_n)\}$ is not $d_{GH}$-precompact.
It follows from Lemmas \ref{lem:dGH-precpt}, \ref{lem:Sep-CapLo},
and \ref{prop:ObsDiam-Sep}
that the $d_{GH}$-precompactness of $\{\mathcal{L}_1(X)\}_{X\in\mathcal{Y}}$ implies
\eqref{eq:fin-ObsDiam}.
\end{rem}
For the proof of Theorem \ref{thm:spec-cpt},
we need some lemmas.
\begin{lem} \label{lem:me-L2}
Let $X$ be an mm-space.
For any two $\mu_X$-measurable functions $f,g : X \to \field{R}$,
we have
\[
\dKF(f,g) \le \|f-g\|_{L_2}^{2/3}.
\]
\end{lem}
\begin{proof}
Setting $\varepsilon := \dKF(f,g)$,
we have $\mu_X(|f-g| \ge \varepsilon) \ge \varepsilon$.
(If otherwise, then we find $\varepsilon'$ such that
$0 < \varepsilon' < \varepsilon$ and
$\mu_X(|f-g| > \varepsilon') < \varepsilon$, which contradicts
$\varepsilon = \dKF(f,g)$.)
Therefore,
\begin{align*}
\|f-g\|_{L_2}^2
&\ge \int_{\{|f-g|\ge\varepsilon\}} |f-g|^2 \;d\mu_X\\
&\ge \varepsilon^2 \mu_X(|f-g| \ge \varepsilon) \ge \varepsilon^3
= \dKF(f,g)^3,
\end{align*}
which implies the lemma.
\end{proof}
\begin{lem} \label{lem:spec-cpt}
Let $X$ be an mm-space.
Then, for any real number $\varepsilon > 0$
there exists a subset $\mathcal{L}{\it ip}_1(X;\varepsilon) \subset \mathcal{L}{\it ip}_1(X)$
such that
\begin{enumerate}
\item the image of $\mathcal{L}{\it ip}_1(X;\varepsilon)$ by
the projection $\mathcal{L}{\it ip}_1(X) \to \mathcal{L}_1(X)$ coincides with $\mathcal{L}_1(X)$,
\item we have $\mathcal{L}{\it ip}_1(X;\varepsilon) \subset
B^{\dKF}_\varepsilon(\Dir_1(X) \cap B^{L_2}_{D_\varepsilon}(0))$,
where $B^{\dKF}_\varepsilon(A)$ is the set of $f \in A$ such that
$\dKF(f,A) \le \varepsilon$, and we set
$D_\varepsilon := \ObsDiam(X;-\varepsilon)$.
\end{enumerate}
\end{lem}
\begin{proof}
Take any $\varepsilon > 0$.
For any function $f \in \mathcal{L}{\it ip}_1(X)$
there are two real numbers $a_f$ and $b_f$ such that
$a_f \le b_f$, $f_*\mu_X[\,a_f,b_f\,] \ge 1-\varepsilon$,
and $b_f-a_f \le D_\varepsilon$.
We may assume that $a_{(f+c)} = a_f + c$ for any $f \in \mathcal{L}{\it ip}_1(X)$
and for any constant $c$.
Set
\[
\mathcal{L}{\it ip}_1(X;\varepsilon) := \{\;f\in\mathcal{L}{\it ip}_1(X) \mid a_f = 0\;\}.
\]
Then, (1) is clear. The rest is to prove (2).
For any $f \in \mathcal{L}{\it ip}_1(X;\varepsilon)$ we set
\[
\tilde{f}(x) := \max\{\min\{f(x),D_\varepsilon\},0\}, \quad x \in X.
\]
Since $\tilde{f} = f$ on
$f^{-1}[\,0,b_f\,]$ and $\mu_X(f^{-1}[\,0,b_f\,]) \ge 1-\varepsilon$,
we have $\dKF(\tilde{f},f) \le \varepsilon$.
Moreover, $0 \le \tilde{f} \le D_\varepsilon$ implies
$\|\tilde{f}\|_{L_2} \le D_\varepsilon$.
We thus have $\tilde{f} \in \Dir_1(X) \cap B^{L_2}_{D_\varepsilon}(0)$.
This completes the proof.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:spec-cpt}]
Take any $\varepsilon > 0$ and fix it.
We set
\[
D_\varepsilon := \max\{\sup_{X \in \mathcal{Y}} \ObsDiam(X;-\varepsilon),1\},
\]
which is finite by the condition \eqref{eq:fin-ObsDiam}.
By the spectral compactness of $\mathcal{Y}$, the family
$\{D_\varepsilon(\Dir_1(X) \cap B^{L_2}_1(0))\}_{X \in \mathcal{Y}}$
of $D_\varepsilon$-scaled sets
is $d_{GH}$-precompact with respect to the $L_2$ norm.
Since $D_\varepsilon(\Dir_1(X) \cap B^{L_2}_1(0))
= \Dir_{D_\varepsilon^2}(X) \cap B^{L_2}_{D_\varepsilon}(0)$ contains
$\Dir_1(X) \cap B^{L_2}_{D_\varepsilon}(0)$,
the family
$\{\Dir_1(X) \cap B^{L_2}_{D_\varepsilon}(0)\}_{X \in \mathcal{Y}}$
is $d_{GH}$-precompact. By Lemma \ref{lem:dGH-precpt},
there is a natural number $N_\varepsilon$ such that,
for any mm-space $X \in \mathcal{Y}$,
we find an $\varepsilon$-net $\mathcal{N} \subset
\Dir_1(X) \cap B^{L_2}_{D_\varepsilon}(0)$ with $\#\mathcal{N} \le N_\varepsilon$.
By Lemma \ref{lem:me-L2} we have
\[
\Dir_1(X) \cap B^{L_2}_{D_\varepsilon}(0) \subset B^{L_2}_\varepsilon(\mathcal{N})
\subset B^{\dKF}_{\varepsilon^{2/3}}(\mathcal{N}),
\]
which together with Lemma \ref{lem:spec-cpt} implies
\[
\mathcal{L}{\it ip}_1(X;\varepsilon) \subset B^{\dKF}_{\varepsilon+\varepsilon^{2/3}}(\mathcal{N}).
\]
Let $\mathcal{N}'$ be the image of $\mathcal{N}$ by a nearest point projection
to $\mathcal{L}{\it ip}_1(X;\varepsilon)$.
We see that $\mathcal{N}'$ is a $(2(\varepsilon+\varepsilon^{2/3}))$-net
of $\mathcal{L}{\it ip}_1(X;\varepsilon)$.
Since the projection $\mathcal{L}{\it ip}_1(X;\varepsilon) \to \mathcal{L}_1(X)$ is
$1$-Lipschitz continuous and surjective,
the image of $\mathcal{N}'$ by the projection
is a $(2(\varepsilon+\varepsilon^{2/3}))$-net of $\mathcal{L}_1(X)$.
This completes the proof.
\end{proof}
\begin{prop} \label{prop:spec-cpt}
Let $\mathcal{Y} \subset \mathcal{X}$ be a family of
compact Riemannian manifolds.
Then, the following {\rm(1)} and {\rm(2)} are equivalent to each other.
\begin{enumerate}
\item $I(\lambda) :=
\sup_{X \in \mathcal{Y}} \max\{\; i \mid \lambda_i(X) \le \lambda\;\}
< +\infty$ for any $\lambda > 0$.
\item $\mathcal{Y}$ is spectrally compact.
\end{enumerate}
\end{prop}
\begin{proof}
We prove (1) $\implies$ (2).
Let $X \in \mathcal{Y}$.
We take a complete orthonormal basis $\{\varphi_i\}_{i=0}^\infty$ on $L_2(X)$
such that $\Delta\varphi_i = \lambda_i(X)\varphi_i$ for any $i$.
For any function $u \in L_2(X)$,
we set $u_i := (u,\varphi_i)_{L_2}$.
We see that $u = \sum_{i=0}^\infty u_i\varphi_i$,
$\|u\|_{L_2}^2 = \sum_{i=0}^\infty u_i^2$,
and, by the Green formula,
\begin{align*}
\mathcal{E}(u) &= \int_X \langle du,du\rangle \; d\mu_X
= \int_X u \Delta u \; d\mu_X\\
&= \sum_{i,j=0}^\infty \int_X \lambda_i(X) u_i u_j \varphi_i \varphi_j \; d\mu_X
= \sum_{i=0}^\infty \lambda_i(X)u_i^2,
\end{align*}
where $\langle\cdot,\cdot\rangle$ is the Riemannian metric
on the cotangent space.
For a given $\varepsilon > 0$, we set
$i_\varepsilon := I(16/\varepsilon^2)$.
Note that, if $i > i_\varepsilon$, then $\lambda_i(X) > 16/\varepsilon^2$.
Define an orthogonal projection $\pi_\varepsilon : L_2(X) \to
L_\varepsilon := \langle\varphi_i\mid i=0,1,\dots,i_\varepsilon\rangle$
by
\[
\pi_\varepsilon(u) := \sum_{i=0}^{i_\varepsilon} u_i\varphi_i.
\]
For any $u \in \Dir_1(X)$,
\[
1 \ge \mathcal{E}(u) = \sum_{i=0}^\infty \lambda_i(X)u_i^2
\ge \sum_{i=i_\varepsilon + 1}^\infty \lambda_i(X)u_i^2
\ge \frac{16}{\varepsilon^2} \sum_{i=i_\varepsilon + 1}^\infty u_i^2
\]
and hence
\[
\| u - \pi_\varepsilon(u) \|_{L_2}
= \left(\sum_{i=i_\varepsilon + 1}^\infty u_i^2\right)^{1/2}
\le \frac{\varepsilon}{4}.
\]
This together with triangle inequalities proves that,
for any $u,v \in \Dir_1(X)$ with $\|u-v\|_{L_2} > \varepsilon$,
we have
\[
\|\pi_\varepsilon(u) - \pi_\varepsilon(v)\|_{L_2} > \frac{\varepsilon}{2}.
\]
Since $L_\varepsilon$ is isometric to $\field{R}^{i_\varepsilon+1}$,
\[
\Cap_\varepsilon(\Dir_1(X) \cap B^{L_2}_1(0))
\le \Cap_{\varepsilon/2}(B^{l_2}_1(o;\field{R}^{i_\varepsilon+1})),
\]
where $B^{l_2}_1(o;\field{R}^{i_\varepsilon})$ denotes the $i_\varepsilon$-dimensional
closed Euclidean unit ball.
Since the right-hand side of the above inequality
depends only on $\varepsilon$,
Lemma \ref{lem:dGH-precpt} proves (2).
We prove (2) $\implies$ (1).
Suppose that (1) does not hold.
Then, there are a number $\lambda > 0$ and a sequence
$\{X_n\} \subset \mathcal{Y}$ such that
$N_n := \max\{\; i \mid \lambda_i(X_n) \le \lambda\;\} \to +\infty$
as $n\to\infty$.
We may assume that $\lambda \ge 1$.
Let $\{\varphi^{(n)}_i\}_{i=0}^\infty$ be an complete orthonormal basis
of $L_2(X_n)$ such that
$\Delta\varphi^{(n)}_i = \lambda_i(X_n)\varphi^{(n)}_i$ for any $i$.
Set $\psi^{(n)}_i := \varphi^{(n)}_i/\sqrt{\lambda}$.
If $i \le N_n$, then $\|\psi^{(n)}_i\|_{L_2} = 1/\sqrt{\lambda} \le 1$
and
\begin{align*}
\mathcal{E}(\psi^{(n)}_i) &= \int_X \psi^{(n)}_i \Delta \psi^{(n)}_i \; d\mu_X
= \lambda_i(X) \int_X (\psi^{(n)}_i)^2 \; d\mu_X\\
&= \lambda_i(X) \|\psi^{(n)}_i\|_{L_2}^2
= \lambda_i(X_n)/\lambda \le 1,
\end{align*}
which imply $\psi^{(n)}_i \in \Dir_1(X) \cap B^{L_2}_1(0)$.
Moreover,
\[
\|\psi^{(n)}_i - \psi^{(n)}_j\|_{L_2} = \sqrt{2/\lambda}
\]
for all different $i$ and $j$.
We thus have
\[
\Cap_{\lambda^{-1/2}}(\Dir_1(X) \cap B^{L_2}_1(0)) \ge N_n \to \infty,
\]
so that $\{X_n\}$ is not spectrally compact because of
Lemma \ref{lem:dGH-precpt}.
$\mathcal{Y}$ is not spectrally compact either.
This completes the proof.
\end{proof}
\begin{lem} \label{lem:dom-Dir}
Let $X$ and $Y$ be two mm-spaces.
If $X$ dominates $Y$, then
$\Dir_1(Y)\cap B_1^{L_2}(0)$ is $L_2$-isometrically embedded into
$\Dir_1(X)\cap B_1^{L_2}(0)$.
\end{lem}
\begin{proof}
There is a $1$-Lipschitz map $f : X \to Y$ with $f_*\mu_X = \mu_Y$.
A required embedding map is defined to be
\[
f^* : L_2(Y) \ni u \mapsto u\circ f \in L_2(X),
\]
which is a linear isometric embedding since,
for any $u \in L_2(Y)$,
\[
\|f^*u\|_{L_2}^2 = \int_X (u\circ f)^2 \,d\mu_X
= \int_Y u^2 \,d\mu_Y = \|u\|_{L_2}^2.
\]
For any $u \in \Dir_1(Y)$ and any $x \in X$,
\begin{align*}
|\grad(f^*u)|(x)
&= \limsup_{y\to x} \frac{|u(f(x)) - u(f(y))|}{d_X(x,y)}\\
&\le \limsup_{y\to x} \frac{|u(f(x)) - u(f(y))|}{d_Y(f(x),f(y))}\\
&\le \limsup_{y'\to f(x)} \frac{|u(f(x)) - u(y')|}{d_Y(f(x),y')}
= |\grad u|(f(x)),
\end{align*}
which proves that $\mathcal{E}(f^*u) \le \mathcal{E}(u) \le 1$.
Therefore, $f^*(\Dir_1(Y)\cap B_1^{L_2}(0))
\subset \Dir_1(X)\cap B_1^{L_2}(0)$.
This completes the proof.
\end{proof}
\begin{thm} \label{thm:spec-conc}
Let $\{X_n\}_{n=1}^\infty$ be a spectrally compact and asymptotic sequence
of mm-spaces such that
\[
\sup_n \ObsDiam(X_n;-\kappa) < +\infty \quad\text{for any $\kappa > 0$}.
\]
Then, $\{X_n\}$ asymptotically concentrates and
the limit pyramid is spectrally concentrated.
\end{thm}
\begin{proof}
Theorem \ref{thm:spec-cpt} together with the assumption
implies that $\{\mathcal{L}_1(X_n)\}$ is $d_{GH}$-precompact,
so that $\{X_n\}$ asymptotically concentrates.
By Lemma \ref{lem:dom-Dir},
$\bigcup_{n=1}^\infty \mathcal{P}_{X_n}$ is spectrally compact.
Since the weak limit of $\mathcal{P}_{X_n}$ is contained in the $\square$-closure
of $\bigcup_{n=1}^\infty \mathcal{P}_{X_n}$, it is spectrally concentrated.
This completes the proof.
\end{proof}
\begin{defn}[Spectral concentration] \label{defn:spec-conc}
\index{spectral concentration} \index{spectral concentrate}
A sequence of mm-spaces is said to \emph{spectrally concentrates}
if it spectrally compact and asymptotically concentrates.
\end{defn}
If a sequence of mm-spaces spectrally concentrates,
then the limit pyramid is spectrally concentrated.
The following corollary is stronger than
Corollary \ref{cor:prod-asympconc}.
\begin{cor} \label{cor:prod-spec-conc}
Let $F_n$, $n=1,2,\dots$, be compact Riemannian manifolds
such that $\lambda_1(F_n)$ is divergent to infinity
as $n\to\infty$.
Let $X_n := F_1 \times F_2 \times \dots \times F_n$
be the Riemannian product space.
Then, $\{X_n\}$ spectrally concentrates.
\end{cor}
\begin{proof}
Since $X_n$ is monotone increasing in $n$ with respect to
the Lipschitz order, $\{X_n\}$ is asymptotic.
By Corollary \ref{cor:ObsDiam-spec},
$\ObsDiam(X_n;\kappa)$, $n=1,2,\dots$, are bounded from above
for any $\kappa > 0$.
It follows from $\lambda_1(F_n) \to\infty$ and Proposition \ref{prop:spec-cpt}
that $\{X_n\}$ is spectrally compact.
By Theorem \ref{thm:spec-conc}, $\{X_n\}$ is asymptotically
concentrates.
\end{proof}
\begin{rem}
In the proof of Corollary \ref{cor:prod-spec-conc},
we do not rely on Proposition \ref{prop:prod-asympconc}.
Namely, we have an alternative proof of
Corollary \ref{cor:prod-asympconc}.
\end{rem}
\begin{ex}[Compare Example \ref{ex:prod-sph}] \label{ex:prod-sph-spec}
Let
$X_n := S^1 \times S^2 \times \dots \times S^n$ (Riemannian product).
Since $\lambda_1(S^n) = n$, Corollary \ref{cor:prod-spec-conc}
proves that $\{X_n\}$ spectrally concentrates.
\end{ex}
\chapter{Dissipation}
\label{chap:dissipation}
\section{Basics for dissipation}
Dissipation is an opposite notion to concentration.
\begin{defn}[Dissipation]
\index{dissipation}
Let $X_n$, $n=1,2,\dots$, be mm-spaces and $\delta > 0$
a real number.
We say that $\{X_n\}$ \emph{$\delta$-dissipates}
\index{delta-dissipate@$\delta$-dissipate}
if for any real numbers $\kappa_0,\kappa_1,\dots,\kappa_N > 0$
with $\sum_{i=0}^N \kappa_i < 1$, we have
\[
\liminf_{n\to\infty} \Sep(X_n;\kappa_0,\kappa_1,\dots,\kappa_N) \ge \delta.
\]
We say that $\{X_n\}$ \emph{infinitely dissipates}
\index{infinitely dissipate}
if it $\delta$-dissipates for any $\delta > 0$.
\end{defn}
The following is obvious.
\begin{prop}
Let $X_n$ and $Y_n$, $n=1,2,\dots$, be mm-spaces such that
$X_n \prec Y_n$ for any $n$.
If $\{X_n\}$ $\delta$-dissipates for a positive real number $\delta$
{\rm(}resp.~infinitely dissipates{\rm)},
then so does $\{Y_n\}$.
\end{prop}
The following lemma is useful to detect dissipating families.
\begin{lem} \label{lem:diss}
Let $X_n$, $n=1,2,\dots$, be mm-spaces and let $\delta > 0$ be a real number.
Assume that for each $n$ there exists an at most countable sequence
$\{C_{ni}\}_{i=1}^{k_n}$, $k_n \le +\infty$, of mutually disjoint Borel subsets
of $X_n$ such that
\begin{align}
\tag{1} &\lim_{n\to\infty} \mu_X(\bigcup_{i=1}^{k_n} C_{ni}) = 1,\\
\tag{2} &\min_{i\neq j} d_{X_n}(C_{ni},C_{nj}) \ge \delta,\\
\tag{3} &\lim_{n\to\infty} \sup_{i=1}^{k_n} \mu_{X_n}(C_{ni}) = 0.
\end{align}
Then, $\{X_n\}$ $\delta$-dissipates.
\end{lem}
The proof of the lemma is easy and omitted.
\begin{ex}
\begin{enumerate}
\item Let $X$ be a compact Riemannian manifold and
$\{t_n\}_{n=1}^\infty$ a sequence of positive real numbers divergent to
infinity. Then, $\{t_n X\}$ infinitely dissipates, where $t_n X$
denotes the manifold $X$ with the Riemannian distance function
multiplied by $t_n$.
\item Let $T_n$ be a perfect rooted binary tree with root $o$
and depth $n$,
i.e., all vertices of $T_n$ are within distance at most $n$ from $o$,
and the number of vertices $v$ of $T_n$ with $d_{T_n}(o,v) = k$
is equal to $2^k$ for any $k \in \{1,2,\dots,n\}$,
where $d_{T_n}$ denotes the path metric on $T_n$.
Consider the equally distributed probability measure $\mu_{T_n}$ on
the set of vertices of $T_n$, i.e.,
\[
\mu_{T_n} = \sum_v \frac{1}{2^{n+1}-1} \delta_v,
\]
where $v$ runs over all vertices of $T_n$.
Then, $\{(T_n,d_{T_n},\mu_{T_n})\}$ infinitely dissipates,
and $\{(T_n,(1/n)d_{T_n},\mu_{T_n})\}$ $2$-dissipates.
\item Let $H$ be a complete simply connected hyperbolic space,
$\{t_n\}_{n=1}^\infty$ a sequence of positive numbers divergent to
infinity, and $B_n$ a metric ball of radius $t_n$ in $H$.
Then, $\{B_n\}$ infinitely dissipates, and
$\{(1/t_n)B_n\}$ $2$-dissipates.
\end{enumerate}
\end{ex}
\begin{prop} \label{prop:dissipate}
Let $X_n$, $n=1,2,\dots$, be mm-spaces and $\delta > 0$
a real number.
\begin{enumerate}
\item $\{X_n\}$ $\delta$-dissipates if and only if
the weak limit of any weakly convergent subsequence
of $\{\mathcal{P}_{X_n}\}$ contains all mm-spaces with diameter $\le \delta$.
\item $\{X_n\}$ infinitely dissipates if and only if
$\mathcal{P}_{X_n}$ converges weakly to $\mathcal{X}$ as $n\to\infty$.
\end{enumerate}
\end{prop}
\begin{proof}
We prove (1). Denote by $\mathcal{X}_{\delta}$ the set of mm-spaces with diameter
$\le \delta$.
Assume that $\{X_n\}$ $\delta$-dissipates and let $\mathcal{P}$ be the weak limit
of a weakly convergent subsequence of $\{\mathcal{P}_{X_n}\}$.
We are going to prove $\mathcal{X}_\delta \subset \mathcal{P}$.
Since the set of finite mm-spaces with diameter $< \delta$ is $\square$-dense
in $\mathcal{X}_\delta$, it suffices to prove that
any such mm-space belongs to $\mathcal{P}$.
Let $Y$ be any finite mm-space with diameter $< \delta$.
We take any number $\varepsilon$ with
$0 < \varepsilon < \delta-\diam Y$.
Let $\{y_0,y_1,\dots,y_N\} := Y$ and take
real numbers $\kappa_0,\kappa_1,\dots,\kappa_N$ such that
$0 < \kappa_i \le \mu_Y\{y_i\}$ for any $i$ and
\[
1-\varepsilon < \sum_{i=0}^N \kappa_i < 1.
\]
There is a natural number $n(\varepsilon)$ such that
$\Sep(X_n;\kappa_0,\dots,\kappa_N) > \delta-\varepsilon$
for any $n \ge n(\varepsilon)$.
Let $n \ge n(\varepsilon)$.
Then, there are Borel subsets $A_{n1},\dots A_{nN} \subset X_n$
such that $\mu_{X_n}(A_{ni}) \ge \kappa_i$,
$d_{X_n}(A_{ni},A_{nj}) > \delta-\varepsilon > \diam Y$
for any different $i$ and $j$.
Embedding $Y$ into $(\field{R}^M,\|\cdot\|_\infty)$ isometrically,
we assume that $Y$ is a subset of $\field{R}^M$.
We define a map $f_n : \bigcup_{i=0}^N A_{ni} \to \field{R}^M$ by
$f_n|_{A_{ni}} := y_i$ for any $i$.
$f_n$ is a $1$-Lipschitz map and extends
to a $1$-Lipschitz map $f_n : X_n \to \field{R}^M$.
We see that
$(f_n)_*\mu_{X_n}(\{y_i\}) = \mu_{X_n}(f_n^{-1}(y_i)) \ge \mu_{X_n}(A_{ni})
\ge \kappa_i$.
Setting $\nu := \sum_{i=0}^N \kappa_i \delta_{y_i}$,
we have
$\nu \le \mu_Y$, $\nu \le (f_n)_*\mu_{X_n}$, and
$\nu(\field{R}^M) = \sum_{i=0}^N \kappa_i > 1-\varepsilon$.
We apply Strassen's theorem (Theorem \ref{thm:di-tra})
to obtain $d_P((f_n)_*\mu_{X_n},\mu_Y) < \varepsilon$.
As $\varepsilon \to 0$,
$(f_{n(\varepsilon)})_*\mu_{X_n}$ converges weakly to $\mu_Y$.
Thus, $X_{n(\varepsilon)} \succ
(\field{R}^M,\|\cdot\|_\infty,(f_{n(\varepsilon)})_*\mu_{X_{n(\varepsilon)}})
\overset{\square}\to Y$
as $\varepsilon \to 0$, so that $Y$ belongs to $\mathcal{P}$.
We next prove the converse.
Assume that the weak limit of any weakly convergent subsequence
of $\{\mathcal{P}_{X_n}\}$ contains all mm-spaces with diameter $\le \delta$.
We take any real numbers $\kappa_0,\dots,\kappa_N > 0$ with
$\sum_{i=0}^N \kappa_i < 1$ and set
\[
\varepsilon_0 := \frac{1}{N+1}\left( 1-\sum_{i=0}^N \kappa_i \right).
\]
We see that $\sum_{i=0}^N(\kappa_i+\varepsilon_0) = 1$.
Let $Y = \{y_0,\dots,y_N\}$ be a finite metric space such that
$d_Y(y_i,y_j) = \delta$ for any different $i$ and $j$,
and let $\mu_Y := \sum_{i=0}^N (\kappa_i+\varepsilon_0)\delta_{y_i}$.
Then, $\mu_Y$ is a probability measure and $Y = (Y,d_Y,\mu_Y)$
is an mm-space.
Let $\mathcal{P}$ be the weak limit of a weakly convergent subsequence
of $\{\mathcal{P}_{X_n}\}$.
The assumption implies that $\mathcal{P}$ contains $Y$.
Thereby, there are mm-spaces $Y_n$, $n=1,2,\dots$,
such that $X_n \succ Y_n \overset{\square}\to Y$ as $n\to\infty$.
By Lemma \ref{lem:box-eps-mm-iso},
we have an $\varepsilon_n$-mm-isomorphism
$f_n : Y \to Y_n$, $\varepsilon_n \to 0$.
For each $n$ there is a Borel subset $Y_0 \subset Y$ such that
$d_P((f_n)_*\mu_Y,\mu_{Y_n}) \le \varepsilon_n$,
$\mu_Y(Y_0) \ge 1-\varepsilon_n$, and
\begin{equation}
\label{eq:diss}
|d_Y(y_i,y_j)-d_{Y_n}(f_n(y_i),f_n(y_j))| \le \varepsilon_n
\end{equation}
for any $y_i,y_j \in Y_0$.
We take $n$ so large that
$\varepsilon_n < \min\{\delta,\varepsilon_0\}$.
Then, since $\mu_Y(Y_0) \ge 1-\varepsilon_n > 1-\varepsilon_0$
and $\mu_Y\{y_i\} > \varepsilon_0$ for any $i$,
we have $Y_0 = Y$.
Since $d_Y(y_i,y_j) = \delta$ for $i\neq j$ and by \eqref{eq:diss},
we have $d_{Y_n}(f_n(y_i),f_n(y_j)) > 0$ for $i\neq j$,
i.e., $f_n$ is injective.
Let $\varepsilon_n'$, $n=1,2,\dots$, be numbers
such that $\lim_{n\to\infty} \varepsilon_n' = 0$
and $\varepsilon_n < \varepsilon_n' < \varepsilon_0$ for any $n$.
Set $A_{ni} := B_{\varepsilon_n'}(f_n(y_i))$.
Since $(f_n)_*\mu_Y = \sum_{i=0}^N (\kappa_i+\varepsilon_0)\delta_{f_n(y_i)}$,
the inequality $d_P((f_n)_*\mu_Y,\mu_{Y_n}) \le \varepsilon_n$
proves that
\[
\mu_{Y_n}(A_{ni}) \ge (f_n)_*\mu_Y\{f_n(y_i)\}-\varepsilon_n'
= \kappa_i+\varepsilon_0-\varepsilon_n' > \kappa_i.
\]
We also have, for $i \neq j$,
\begin{align*}
d_{Y_n}(A_{ni},A_{nj}) &\ge d_{Y_n}(f_n(y_i),f_n(y_j))-2\varepsilon_n'\\
&\ge d_Y(y_i,y_j)-\varepsilon_n-2\varepsilon_n'
= \delta-\varepsilon_n-2\varepsilon_n'.
\end{align*}
Therefore, $\Sep(Y_n;\kappa_0,\dots,\kappa_N)
\ge \delta-\varepsilon_n-2\varepsilon_n'$, so that,
by Lemma \ref{lem:Sep-prec}, we obtain
$\Sep(X_n;\kappa_0,\dots,\kappa_N)
\ge \delta-\varepsilon_n-2\varepsilon_n'$.
(1) has been proved.
We prove (2).
If $\{X_n\}$ infinitely dissipates, then
(1) implies that the weak limit of any subsequence of $\{\mathcal{P}_{X_n}\}$
contains any mm-spaces with finite diameter.
Since the set of mm-spaces with finite diameter is $\square$-dense
in $\mathcal{X}$, we have $\mathcal{P} = \mathcal{X}$.
The converse also follows from (1).
This completes the proof of the proposition.
\end{proof}
For an mm-space $F$, we denote by $F^n$ the $n^{th}$ power product space of $F$,
and by $\mu_F^{\otimes n}$ the product measure on $F^n$ of $\mu_F$.
Let $d_{l_p}$, $1 \le p \le +\infty$, be the $l_p$ metric on $F^n$
induced from $d_F$.
\begin{prop} \label{prop:disconn-dissipation}
If $F$ is a compact disconnected mm-space,
then $\{(F^n,d_{l_p},\mu_F^{\otimes n})\}_{n=1}^\infty$
$\delta$-dissipates for some $\delta > 0$ and
for any $p$ with $1 \le p \le +\infty$.
\end{prop}
\begin{proof}
Since $(F^n,d_{l_\infty},\mu_F^{\otimes n}) \prec (F^n,d_{l_p},\mu_F^{\otimes n})$,
it suffices to prove the proposition for $p = +\infty$.
Since $F$ is disconnected,
there are two disjoint subsets $F_1, F_2 \subset F$
with $F_1 \cup F_2 = F$ such that $F_1$ and $F_2$ are both open and closed.
The compactness of $F_1$ and $F_2$ proves that
$\delta := d_F(F_1,F_2) > 0$.
Set $a := \mu_F(F_1)$ and $b := \mu_F(F_2)$.
We have $0 < a,b < 1$ because $F_1$ and $F_2$ are open.
For $i_1,i_2,\dots,i_n = 1,2$, we set
\[
F_{i_1i_2\dots i_n} := F_{i_1} \times F_{i_2} \times \dots F_{i_n} \subset F^n.
\]
Since
\[
\mu_F^{\otimes n}(F_{i_1i_2\dots i_n}) = a^k b^{n-k},
\]
where $k$ is the number of $j$'s with $i_j = 1$,
we have
\[
\lim_{n\to\infty} \max_{i_1,i_2,\dots,i_n} \mu_F^{\otimes n}(F_{i_1i_2\dots i_n})
= 0.
\]
We also see
\[
d_{l_\infty}(F_{i_1i_2\dots i_n},F_{j_1j_2\dots j_n}) =
\begin{cases}
\delta &\text{if $(i_1,i_2,\dots,i_n) \neq (j_1,j_2,\dots,j_n)$},\\
0 &\text{if $(i_1,i_2,\dots,i_n) = (j_1,j_2,\dots,j_n)$}.
\end{cases}
\]
The proposition follows from Lemma \ref{lem:diss}.
\end{proof}
The following is another example of dissipation.
\begin{prop} \label{prop:Sn-dissipate}
Let $r_n$, $n=1,2,\dots$, be positive real numbers.
Then, $\{S^n(r_n)\}$ infinitely dissipates if and only if
$r_n/\sqrt{n} \to +\infty$ as $n\to\infty$.
\end{prop}
\begin{proof}
We first assume $r_n/\sqrt{n} \to +\infty$ as $n\to\infty$.
Take any finitely many positive real numbers
$\kappa_0,\kappa_1,\dots,\kappa_N$ with $\sum_{i=0}^N \kappa_i < 1$,
and fix them.
We find positive real numbers
$\kappa_0',\kappa_1',\dots,\kappa_N'$
in such a way that $\kappa_i < \kappa_i'$ for any $i$
and $\sum_{i=0}^N \kappa_i' < 1$.
For any $\varepsilon > 0$,
there are Borel subsets $A_0,A_1,\dots,A_N \subset \field{R}$ such that
$\gamma^1(A_i) \ge \kappa_i'$ for any $i$
and
\[
\min_{i\neq j} d_{\field{R}}(A_i,A_j)
> \Sep((\field{R},\gamma^1);\kappa_0',\dots,\kappa_N')-\varepsilon.
\]
We may assume that all $A_i$ are open subsets of $\field{R}$.
Denote by $\pi_n : S^n(\sqrt{n}) \to \field{R}$ the orthogonal projection
as in Section \ref{sec:obs-sphere}.
By the Maxwell-Boltzmann distribution law,
we have
$\liminf_{n\to\infty}(\pi_n)_*\sigma^n(A_i) \ge \gamma^1(A_i) \ge \kappa_i'$
for $i=0,1,\dots,N$.
There is a natural number $n_0$ such that
$\sigma^n(\pi_n^{-1}(A_i)) \ge \kappa_i$ for any $i$ and $n \ge n_0$.
The $1$-Lipschitz continuity of $\pi_n$ implies
that $d_{S^n(\sqrt{n})}(\pi_n^{-1}(A_i),\pi_n^{-1}(A_j)) \ge d_{\field{R}}(A_i,A_j)$.
Therefore, for any $n \ge n_0$,
\[
\Sep(S^n(\sqrt{n});\kappa_0,\dots,\kappa_N)
> \Sep((\field{R},\gamma^1);\kappa_0',\dots,\kappa_N')-\varepsilon,
\]
which proves
\[
\liminf_{n\to\infty} \Sep(S^n(\sqrt{n});\kappa_0,\dots,\kappa_N)
\ge \Sep((\field{R},\gamma^1);\kappa_0',\dots,\kappa_N') > 0.
\]
Since $r_n/\sqrt{n} \to +\infty$ as $n\to\infty$,
\begin{equation}
\label{eq:Levy-dissipate-sphere}
\Sep(S^n(r_n);\kappa_0,\dots,\kappa_N)
= \frac{r_n}{\sqrt{n}} \Sep(S^n(\sqrt{n});\kappa_0,\dots,\kappa_N)
\end{equation}
is divergent to infinity and so $\{S^n(r_n)\}$ infinitely dissipates.
We next prove the converse.
Assume that $\{S^n(r_n)\}$ infinitely dissipates
and $r_n/\sqrt{n}$ is not divergent to infinity.
Then, there is a subsequence $\{r_{n_i}\}$ of $\{r_n\}$
such that $r_{n_i}/\sqrt{n_i}$ is bounded for all $i$.
By \eqref{eq:Levy-dissipate-sphere},
$\{S^{n_i}(\sqrt{n_i})\}$ infinitely dissipates.
However, for each fixed $\kappa$ with $0 < \kappa < 1/2$,
$\ObsDiam(S^n(\sqrt{n});-\kappa)$ is bounded for all $n$
and, by Proposition \ref{prop:ObsDiam-Sep},
so is $\Sep(S^n(\sqrt{n});\kappa,\kappa)$,
which contradicts that $\{S^{n_i}(\sqrt{n_i})\}$ infinitely dissipates.
This completes the proof.
\end{proof}
\section{Obstruction for dissipation}
In this section, we study an obstruction for dissipation and prove
the following
\begin{thm}[Non-dissipation theorem] \label{thm:non-dissipation}
\index{non-dissipation theorem}
Let $F$ be a compact, connected, and locally connected mm-space.
Then, $\{(F^n,d_{l_\infty},\mu_F^{\otimes n})\}_{n=1}^\infty$ does not dissipate.
\end{thm}
Note that the connectivity of $F$ is necessary as is seen in
Proposition \ref{prop:disconn-dissipation}.
\begin{rem}
As is seen in Proposition \ref{prop:prod-non-conc},
$\{(F^n,d_{l_p},\mu_F^{\otimes n})\}_{n=1}^\infty$ for any $0 \le p \le +\infty$
does not asymptotically concentrate
unless $F$ consists of a single point.
In particular, $\{(F^n,d_{l_\infty},\mu_F^{\otimes n})\}$
neither asymptotically concentrate nor dissipate
for any connected and compact Riemannian manifold $F$
not consisting of a single point.
Since $\{(F^n,d_{l_\infty},\mu_F^{\otimes n})\}$ is a monotone increasing sequence
with respect to the Lipschitz order, this also holds for any
subsequence of $\{(F^n,d_{l_\infty},\mu_F^{\otimes n})\}$.
\end{rem}
\begin{defn}[Expansion coefficient]
\index{expansion coefficient} \index{exp@$\Exp(X;\kappa,\rho)$}
Let $X$ be an mm-space and let $\kappa,\rho > 0$.
The \emph{expansion coefficient $\Exp(X;\kappa,\rho)$ of $X$}
is defined to be the supremum of real numbers $\xi \ge 1$ such that,
if $\mu_X(A) \ge \kappa$ for a Borel subset $A \subset X$,
then $\mu_X(B_\rho(A)) \ge \xi\kappa$.
\end{defn}
It is clear that $\Exp(X;\kappa,\rho) \ge 1$.
\begin{lem} \label{lem:exp-diss}
Let $X_n$, $n=1,2,\dots$, be mm-spaces and let $\delta > 0$
be a real number.
If there exist two real numbers $\kappa$ and $\rho$
with $\kappa > 0$ and $0 < \rho < \delta$ such that
\[
\inf_n \Exp(X_n;\kappa,\rho) > 1,
\]
then $\{X_n\}$ does not $\delta$-dissipate.
\end{lem}
\begin{proof}
By the assumption, we find a constant $c > 1$
such that $\Exp(X_n;\kappa,\rho) > c$ for any $n$.
Take a number $\kappa'$ in such a way that
$1-c\kappa < \kappa' < 1-\kappa$.
Suppose that $\{X_n\}$ $\delta$-dissipate.
Then, we have
$\Sep(X_n;\kappa,\kappa') > \rho$ for every sufficiently large $n$.
For such a number $n$, we find two Borel subset $A,A' \subset X_n$
in such a way that $d_X(A,A') > \rho$, $\mu_{X_n}(A) \ge \kappa$,
and $\mu_{X_n}(A') \ge \kappa'$.
It follows from $\Exp(X_n;\kappa,\rho) > c$
that $\mu_{X_n}(B_\rho(A)) \ge c\kappa$.
Since $B_\rho(A)$ and $A'$ does not intersect to each other,
we have
\[
\mu_{X_n}(B_\rho(A) \cup A') \ge c\kappa + \kappa' > 1,
\]
which is a contradiction.
This completes the proof.
\end{proof}
\begin{prop} \label{prop:exp-lam1}
Let $X$ be an mm-space.
Then we have
\[
\Exp(X;\kappa,\rho) \ge \min\left\{1+\frac{\lambda_1(X)\rho^2}{4},2\right\}
\]
for any real numbers $\kappa$ and $\rho$ with $0 < \kappa \le 1/4$
and $\rho > 0$.
\end{prop}
\begin{proof}
If $\lambda_1(X) = 0$, then the proposition is trivial.
Assume that $\lambda_1(X) > 0$.
It suffices to prove that
\begin{equation}
\label{eq:exp-lam1}
\mu_X(B_\rho(A)) \ge e_\rho\kappa
\end{equation}
for any Borel subset $A \subset X$ with $\mu_X(A) \ge \kappa$,
where
\[
e_\rho := \min\left\{1+\frac{\lambda_1(X)\rho^2}{4},2\right\}.
\]
Let us first prove it in the case where $A$ is an open subset of $X$.
For two constants $c$ and $r > 0$,
we define a function $f_{c,r} : X \to \field{R}$ by
\[
f_{c,r}(x) :=
\begin{cases}
c + \frac{r}{\rho} d_X(x,A)
&\text{for $x \in B_\rho(A)$},\\
c+r &\text{for $x \in X \setminus B_\rho(A)$}.
\end{cases}
\]
$f_{c,r}$ is Lipschitz continuous with Lipschitz constant
$r/\rho$.
We have $f_{c,r} = c + f_{0,r}$ and so
\[
\int_X f_{c_r,r}\,d\mu_X = 0,\qquad
c_r := -\int_X f_{0,r}\,d\mu_X.
\]
Since $A \cup (X \setminus B_\rho(A))$ is open,
we have $|\grad f| = 0$ on $A \cup (X \setminus B_\rho(A))$,
so that
\[
\mathcal{E}(f_{c,r}) = \frac{r^2}{\rho^2}E,
\qquad E := \int_{B_\rho(A)\setminus A} |\grad d_X(\cdot,A)|^2\,d\mu_X.
\]
If $E = 0$, then the Rayleigh quotient satisfies $R(f_{c_r,r}) = 0$,
which contradicts $\lambda_1(X) > 0$.
We therefore have $E > 0$.
Setting $f := f_{c_r,r}$ for $r := \rho/E^{1/2}$,
we have $\mathcal{E}(f) = 1$ and $\int_X f\,d\mu_X = 0$.
Since $|\grad d_X(\cdot,A)| \le 1$,
\[
1 = \mathcal{E}(f) \le \frac{r^2}{\rho^2}
(\mu_X(B_\rho(A)) - \mu_X(A)),
\]
which implies
\[
r \ge \frac{\rho}{\sqrt{\mu_X(B_\rho(A)) - \mu_X(A)}}.
\]
Since $f \ge c_r$ on $X$ and $f = c_r+r$ on $X \setminus B_\rho(A)$,
\begin{align*}
0 &= \int_X f\,d\mu_X \ge c_r\,\mu_X(B_\rho(A)) + (c_r+r)\mu_X(X\setminus B_\rho(A))\\
&= c_r + r\,\mu_X(X\setminus B_\rho(A)).
\end{align*}
Therefore,
\[
-c_r \ge
\frac{(1-\mu_X(B_\rho(A)))\rho}
{\sqrt{\mu_X(B_\rho(A))-\mu_X(A)}}.
\]
This together with
\begin{align*}
\frac{1}{\lambda_1(X)} \ge \|f\|_{L_2}^2 \ge c_r^2\, \mu_X(A).
\end{align*}
implies
\begin{align} \label{eq:exp-lam1-2}
\mu_X(B_\rho(A))
\ge \{1 + \lambda_1(X)(1-\mu_X(B_\rho(A)))^2\rho^2\}\,
\mu_X(A)
\end{align}
If $\mu_X(B_\rho(A)) \ge 1/2$, then
$\mu_X(B_\rho(A)) \ge 1/2 \ge 2\kappa \ge e_\rho\kappa$.
If $\mu_X(B_\rho(A)) < 1/2$, then
\eqref{eq:exp-lam1-2} leads us to
\[
\mu_X(B_\rho(A))
\ge \left( 1 + \frac{\lambda_1(X)\rho^2}{4} \right)\kappa \ge e_\rho\kappa.
\]
We thus obtain \eqref{eq:exp-lam1} for any open subset $A \subset X$
with $\mu_X(A) \ge \kappa$.
Let $A$ be any Borel subset of $X$ with $\mu_X(A) \ge \kappa$.
We take any number $\delta$ with $0 < \delta < \rho$.
Applying \eqref{eq:exp-lam1} for $U_\delta(A)$ and $\rho-\delta$
yields that
\[
\mu_X(B_\rho(A)) \ge \mu_X(B_{\rho-\delta}(U_\delta(A)))
\ge e_{\rho-\delta}\,\kappa,
\]
so that $\Exp(X;\kappa,\rho) \ge e_{\rho-\delta}$
for any $\delta$ with $0 < \delta < \rho$.
This completes the proof.
\end{proof}
Lemma \ref{lem:exp-diss} and Proposition \ref{prop:exp-lam1}
together imply
\begin{thm} \label{thm:lam1-diss}
Let $X_n$, $n=1,2,\dots$, be mm-spaces
such that $\lambda_1(X_n)$ is bounded away from zero.
Then, $\{X_n\}$ does not dissipate, i.e.,
does not $\delta$-dissipate for any real number $\delta > 0$.
\end{thm}
\begin{cor} \label{cor:prod-diss}
Let $F$ be a compact connected Riemannian manifold.
Then, the sequence $\{F^n\}$ of Riemannian product spaces
does not dissipate.
\end{cor}
\begin{proof}
Since $\lambda_1(F^n) = \lambda_1(F) > 0$,
Theorem \ref{thm:lam1-diss} implies the corollary.
\end{proof}
\begin{defn}[Modulus of continuity]
\index{modulus of continuity}
Let $f : X \to Y$ be a map between two metric spaces $X$ and $Y$.
A function $\omega : [\,0,+\infty\,] \to [\,0,+\infty\,]$
is called a \emph{modulus of continuity of $f$}
if $\omega(0+) = \omega(0) = 0$ and if
\[
d_Y(f(x),f(y)) \le \omega(d_X(x,y))
\]
for any two points $x,y \in X$.
\end{defn}
A map $f : X \to Y$ has a modulus of continuity
if and only if $f$ is uniformly continuous.
\begin{rem} \label{rem:mod-cont}
If $\omega$ is a modulus of continuity of a map $f : X \to Y$,
then the function $\hat\omega(t) = \sup_{s \le t} \omega(s)$
is a monotone nondecreasing modulus of continuity of $f$.
We may assume without loss of generality that a modulus of continuity is
monotone nondecreasing.
\end{rem}
\begin{lem} \label{lem:pre-maj}
Let $f_n : X_n \to Y_n$, $n=1,2,\dots$, be uniformly continuous functions
between mm-spaces $X_n$ and $Y_n$
all which admit a common modulus of continuity
and satisfy $(f_n)_*\mu_{X_n} = \mu_{Y_n}$.
If $\{X_n\}$ does not dissipate, then so does $\{Y_n\}$.
\end{lem}
\begin{proof}
Assume that $\{Y_n\}$ $\delta$-dissipate for some number $\delta > 0$.
Let $\kappa_0,\dots,\kappa_N$ be any positive real numbers with
$\sum_{i=0}^N \kappa_i < 1$.
Then, there is a natural number $n_0$ such that
$\Sep(Y_n;\kappa_0,\dots,\kappa_N) > \delta/2$
for any $n \ge n_0$.
For each $n \ge n_0$, we find Borel subsets $A_{n1},\dots,A_{nN} \subset Y_n$
such that $\mu_{Y_n}(A_{ni}) \ge \kappa_i$ for any $i$
and $d_{Y_n}(A_{ni},A_{nj}) \ge \delta/2$
for any different $i$ and $j$.
Let $\omega$ be the common modulus of continuity of $f_n$.
We assume that $\omega$ is monotone nondecreasing
(see Remark \ref{rem:mod-cont}).
Then we have
\[
\omega(d_{X_n}(f_n^{-1}(A_{ni}),f_n^{-1}(A_{nj}))+0)
\ge d_{Y_n}(A_{ni},A_{nj}) \ge \delta/2,
\]
so that there is a number $\delta_0 > 0$ depending only
on $\delta$ and $\omega$ such that
\[
d_{X_n}(f_n^{-1}(A_{ni}),f_n^{-1}(A_{nj})) \ge \delta_0.
\]
We also have $\mu_{X_n}(f_n^{-1}(A_{ni})) = \mu_{Y_n}(A_{ni}) \ge \kappa_i$.
Therefore, $\Sep(X_n;\kappa_0,\dots,\kappa_N) \ge \delta_0$
for any $n \ge n_0$ and $\{X_n\}$ $\delta_0$-dissipates.
This completes the proof.
\end{proof}
\begin{lem}[Majorization lemma] \label{lem:maj}
\index{majorization lemma}
Let $f : F_0 \to F$ be a uniformly continuous map between
two mm-spaces $F_0$ and $F$ such that $f_*\mu_{F_0} = \mu_F$.
If $\{(F_0^n,d_{l_\infty},\mu_{F_0}^{\otimes n})\}$ does not dissipate,
then $\{(F^n,d_{l_\infty},\mu_F^{\otimes n})\}$ does not dissipate.
\end{lem}
\begin{proof}
Let $f_n : F_0^n \to F^n$ be the function defined by
\[
f_n(x_1,x_2,\dots,x_n) := (f(x_1),f(x_2),\dots,f(x_n))
\]
for $(x_1,x_2,\dots,x_n) \in F_0^n$.
We are going to apply Lemma \ref{lem:pre-maj}.
For any Borel subsets $A_i \subset F$, $i=1,2,\dots,n$,
we have
\begin{align*}
(f_n)_*\mu_{F_0^n}(A_1 \times \dots \times A_n)
&= \mu_{F_0^n}(f^{-1}(A_1) \times \dots \times f^{-1}(A_n))\\
&= \mu_{F^n}(A_1 \times \dots \times A_n).
\end{align*}
Since the family of the products of Borel subsets of $F$ generate
the Borel $\sigma$-algebra of $F^n$, we obtain
$(f_n)_*\mu_{F_0^n} = \mu_{F^n}$.
If $\omega$ is a monotone nondecreasing modulus of continuity
of $f$, then we have, for any $x = (x_1,x_2,\dots,x_n),
y = (y_1,y_2,\dots,y_n) \in F_0$,
\begin{align*}
d_{l_\infty}(f_n(x),f_n(y)) &= \max_i d_F(f(x_i),f(y_i))
\le \max_i \omega(d_{F_0}(x_i,y_i))\\
&\le \omega(\max_i d_{F_0}(x_i,y_i)) = \omega(d_{l_\infty}(x,y)),
\end{align*}
so that $\omega$ is a modulus of continuity of $f_n$
for any $n$.
Lemma \ref{lem:pre-maj} completes the proof.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:non-dissipation}]
It is known (see \cite{Bog}*{9.7.1, 9.7.2})
that there is a continuous map $f : [\,0,1\,] \to F$
such that $f_*\mathcal{L}^1 = \mu_F$.
By the compactness of $[\,0,1\,]$,
the map $f$ is uniformly continuous.
Denote by $\mathcal{L}^n$ the $n$-dimensional Lebesgue measure on $[\,0,1\,]^n$.
By Corollary \ref{cor:prod-diss},
$\{([\,0,1\,]^n,d_{l_2},\mathcal{L}^n)\}$ does not dissipate, which
together with $([\,0,1\,]^n,d_{l_\infty},\mathcal{L}^n) \prec ([\,0,1\,]^n,d_{l_2},\mathcal{L}^n)$
implies that $\{([\,0,1\,]^n,d_{l_\infty},\mathcal{L}^n)\}$ does not dissipate either.
The majorization lemma, Lemma \ref{lem:maj},
completes the proof.
\end{proof}
\chapter{Curvature and concentration}
\label{chap:curv-conc}
\section{Fibration theorem for concentration}
\label{sec:conc}
In this section, we see that a concentration $X_n \to Y$ yields
a Borel measurable map $p_n : X_n \to Y$ such that,
for every sufficiently large $n$,
(1) $p_n$ is $1$-Lipschitz up to a small additive error,
(2) every fiber of $p_n$ concentrates to a one-point mm-space,
and (3) all fibers of $p_n$ are almost parallel to each other
for $n$ large enough.
For a Borel subset $B$ of an mm-space $X$,
we equip $B$ with the restrictions of $d_X$ and $\mu_X$ on $B$,
so that $B$ becomes a (possibly incomplete) mm-space whose measure
is not necessarily probability.
For such a $B$, we define the \emph{observable diameter}
$\ObsDiam(B;-\kappa)$, $\kappa > 0$, by
\begin{align*}
\ObsDiam(B;-\kappa) &:= \sup\{\;\diam(f_*\mu_X;\mu_X(B)-\kappa) \mid\\
&\qquad\qquad\text{$f : B \to \field{R}$ $1$-Lipschitz}\;\}.
\end{align*}
\index{observable diameter}
\index{obsdiam@$\ObsDiam_Y(\cdots)$, $\ObsDiam(\cdots)$}
\begin{defn}[Effectuate relative concentration]
\index{effectuate relative concentration}
Let $X_n$ and $Y$ be mm-spaces
and let $p_n : X_n \to Y$ be Borel measurable maps, where $n = 1,2,\dots$.
We say that $\{p_n\}$ \emph{effectuates relative concentration} of $X_n$
over $Y$ if
\begin{align*}
\limsup_{n\to \infty} \ObsDiam(p_n^{-1}(B);-\kappa)\leq
\diam B
\end{align*}
for any $\kappa>0$ and any Borel subset $B\subset Y$.
\end{defn}
$\{p_n\}$ effectuating relative concentration of $X_n$ over $Y$
means that every fiber of $p_n$ concentrates to a one-point space
as $n\to\infty$.
\begin{lem} \label{lem:Lip1-ObsDiam}
Let $p : X \to Y$ be a Borel measurable map between
two mm-spaces $X$ and $Y$.
If we have
\begin{align*}
\mathcal{L}{\it ip}_1(X)\subset B_{\varepsilon}(p^*\mathcal{L}{\it ip}_1(Y))
\end{align*}
for a real number $\varepsilon > 0$,
then
\begin{align*}
\ObsDiam(p^{-1}(B);-\kappa)\leq
\diam B+ 2\varepsilon
\end{align*}
for any $\kappa\geq \varepsilon$ and any Borel subset
$B\subset Y$.
\end{lem}
\begin{proof}
Let $f : p^{-1}(B)\to \field{R}$ be an arbitrary $1$-Lipschitz function.
There is a $1$-Lipschitz extension $\tilde{f} : X \to \field{R}$ of $f$.
From the assumption, we have a $1$-Lipschitz function $g:Y\to \field{R}$ such that
$\dKF(\tilde{f}, g \circ p)\leq \varepsilon$.
Setting
\begin{align*}
A := \{\; x \in p^{-1}(B) \; ; \; |\tilde{f}(x)-(g \circ p)(x)|
\le \varepsilon\;\},
\end{align*}
we have $\mu_{X}(p^{-1}(B)) - \mu_X(A) \le \varepsilon \le \kappa$ and hence
\begin{align*}
f_*(\mu_{X}|_{p^{-1}(B)})(\overline{f(A)})
\ge \mu_{X}(A)
\ge \mu_{X}(p^{-1}(B))-\kappa.
\end{align*}
For any $x,x'\in A$ we have
\begin{align*}
|f(x)-f(x')| &=
|\tilde{f}(x)-\tilde{f}(x')|\\
&\leq |\tilde{f}(x)-(g\circ p)(x)|+|(g \circ
p)(x)-(g \circ p)(x')|\\
&\quad +|(g \circ p)(x')-\tilde{f}(x')|\\
&\le d_Y(p(x),p(x')) + 2\varepsilon
\le \diam B + 2\varepsilon,
\end{align*}
so that $\diam f(A) \leq \diam B + 2\varepsilon$.
This completes the proof.
\end{proof}
The following is a direct consequence of Lemma \ref{lem:Lip1-ObsDiam}.
\begin{cor} \label{cor:Lip1-ObsDiam}
Let $p_n:X_n\to Y$ be Borel measurable maps
that enforce $\varepsilon_n$-concentration of mm-spaces $X_n$
to an mm-space $Y$ for some $\varepsilon_n \to 0$.
Then, $\{p_n\}$ effectuates relative concentration of $X_n$ over $Y$.
\end{cor}
\begin{defn}[$\kappa$-distance]
\index{kappa-distance@$\kappa$-distance}
\index{dplus@$d_+(A_1,A_2;+\kappa)$}
Let $\kappa > 0$ and let $A_1$ and $A_2$ be two Borel subsets
in an mm-space $X$. We define the \emph{$\kappa$-distance}
\[
d_+(A_1,A_2;+\kappa)
\]
between $A_1$ and $A_2$ as the supremum of $d_X(B_1,B_2)$
over all Borel subsets $B_1 \subset A_1$ and $B_2 \subset A_2$
with $\mu_X(B_1) \geq \kappa$ and $\mu_X(B_2)\geq \kappa$.
If $\min\{\mu_X(A_1),\mu_X(A_2)\} < \kappa$,
then we set $d_+(A_1,A_2;+\kappa) := 0$.
\end{defn}
\begin{lem} \label{lem:kappa-dist}
Let $p : X \to Y$ be a Borel measurable
map between two mm-spaces $X$ and $Y$ such that
\begin{align*}
\mathcal{L}{\it ip}_1 (X)\subset B_{\varepsilon}(p^*\mathcal{L}{\it ip}_1(Y))
\end{align*}
for a real number $\varepsilon > 0$. Then, for any two Borel
subsets $A_1,A_2 \subset X$
and any real number $\kappa$ with $\varepsilon < \kappa$,
we have
\begin{align*}
d_+(A_1,A_2;+\kappa) \le d_Y(p(A_1),p(A_2))
+ \diam p(A_1) + \diam p(A_2) + 2\varepsilon.
\end{align*}
\end{lem}
\begin{proof}
Let $A_1'\subset A_1$ and $A_2'\subset A_2$ be
Borel subsets such that $\mu_X(A_1')\geq \kappa$ and
$\mu_X(A_2')\geq \kappa$.
We set
\[
f(x) := \min\{d_X(x,A_1'),d_X(A_1',A_2')\}, \qquad x \in X,
\]
which is a $1$-Lipschitz function on $X$.
By the assumption there is a $1$-Lipschitz
function $g : Y \to \field{R}$ such that $\dKF(f,g\circ p)\leq \varepsilon$.
Letting
\begin{align*}
B := \{\; x\in X \; ; \; |f(x)-(g\circ p)(x)|\leq \varepsilon \;\},
\end{align*}
we have $\mu_X(B)\geq 1-\varepsilon$.
Since $\mu_X(A_i') \ge \kappa > \varepsilon$,
the intersection of $A_i'$ and $B$ is nonempty for $i=1,2$.
We take a point $x\in A_1' \cap B$ and a point $x'\in A_2'\cap B$.
Since $f\equiv 0$ on $A_1'$ and $f\equiv d_X(A_1',A_2')$ on $A_2'$, we have
\begin{align*}
& |g(p(x))|=|(g\circ p)(x)-f(x)| \le \varepsilon,\\
& |g(p(x'))-d_X(A_1',A_2')|=|(g \circ p)(x')-f(x')| \le \varepsilon,
\end{align*}
and therefore
\begin{align*}
d_X(A_1',A_2') - 2\varepsilon &\le g(p(x'))-g(p(x))
\le d_Y(p(x),p(x'))\\
&\le d_Y(p(A_1),p(A_2)) + \diam p(A_1) + \diam p(A_2).
\end{align*}
This completes the proof.
\end{proof}
\begin{lem} \label{lem:lm-tra}
Let $\mu_1$ and $\mu_2$ be two Borel probability measures
on a metric space $X$. Assume that there exists a
$\rho$-transport plan $\pi$ between $\mu_1$ and $\mu_2$
with $\defi\pi < 1-2\kappa$ for two real numbers $\rho$ and $\kappa$
with $\rho > 0$ and $0 < \kappa < 1/2$.
Then, for any $1$-Lipschitz function $f : X \to \field{R}$, we have
\begin{align*}
&|\lm(f;\mu_1)- \lm(f;\mu_2)|\\
&\leq \rho + \ObsDiam(\mu_1;-\kappa) + \ObsDiam(\mu_2;-\kappa),
\end{align*}
where $\lm(f;\mu_i)$ is the L\'evy mean of $f$ with respect to $\mu_i$.
\end{lem}
\begin{proof}
Let $f : X \to \field{R}$ be a $1$-Lipschitz continuous function.
We set $D_i := \ObsDiam(\mu_i;-\kappa)$ for $i=1,2$.
Since Lemma \ref{lem:LeRad-ObsDiam} implies $\LeRad(\mu_i;-\kappa) \le D_i$, we have
\[
\mu_i(|f-\lm(f;\mu_i)| \le D_i) \ge 1-\kappa
\]
for $i = 1,2$.
Letting
\begin{align*}
I_i:=\{\;s\in \field{R}\mid |s-\lm(f;\mu_i)| \le D_i\;\},
\end{align*}
we have $\mu_i(f^{-1}(I_i)) \ge 1-\kappa$.
We prove that $\pi(f^{-1}(I_1)\times f^{-1}(I_2)) > 0$.
In fact, since
\begin{align*}
1-\kappa-\pi (f^{-1}(I_1)\times X) &\leq
\mu_1(f^{-1}(I_1))-\pi(f^{-1}(I_1)\times X)\\
&\leq
\mu_1(f^{-1}(I_1))-\pi (f^{-1}(I_1)\times X)\\
&\ + \mu_1(X\setminus f^{-1}(I_1))-\pi ((X \setminus f^{-1}(I_1))\times
X)\\
&= 1-\pi(X\times X),
\end{align*}
we have
\[
\pi(f^{-1}(I_1)\times X) \ge \pi(X\times X) - \kappa
\]
and, in the same way,
\[
\pi(X\times f^{-1}(I_2)) \ge \pi(X\times X) - \kappa.
\]
We therefore obtain
\begin{align*}
\pi(f^{-1}(I_1)\times f^{-1}(I_2))
&\ge \pi(f^{-1}(I_1)\times X) + \pi(X\times f^{-1}(I_2)) - \pi(X\times X)\\
&\ge \pi(X\times X) - 2\kappa = 1 - 2\kappa - \defi\pi > 0.
\end{align*}
There is a point $(x_1,x_2) \in (f^{-1}(I_1)\times f^{-1}(I_2)) \cap \supp\pi$.
By $\supp \pi \subset \{d_X \le \rho\}$,
we have $|f(x_1)-f(x_2)| \le d_X(x_1,x_2) \le \rho$.
Since $f(x_i)$ belongs to $I_i$ for $i=1,2$, we obtain
\[
|\lm(f;\mu_1)-\lm(f;\mu_2)| \le |f(x_1)-f(x_2)| + D_1 + D_2
\le \rho + D_1 + D_2.
\]
This completes the proof.
\end{proof}
\begin{defn}[$\lambda$-Prohorov metric]
\index{lambda-Prohorov distance/metric@$\lambda$-Prohorov distance/metric}
Let $X$ be a metric space.
Define the \emph{$\lambda$-Prohorov distance $d_P^{(\lambda)}(\mu,\nu)$
between two Borel probability measures $\mu$ and $\nu$ on $X$},
$\lambda > 0$, to be the infimum of $\varepsilon > 0$ such that
\[
\mu(B_\varepsilon(A)) \ge \nu(A) - \lambda\varepsilon
\]
for any Borel subset $A \subset X$.
\end{defn}
We see that $\lambda d_P^{(\lambda)}$ coincides with the Prohorov
metric with respect to the scaled metric $\lambda d_X$.
\begin{thm}[Fibration theorem for concentration] \label{thm:pn}
Let $p_n : X_n \to Y$ be Borel measurable maps
between mm-spaces $X_n$ and $Y$, $n=1,2,\dots$,
such that $(p_n)_*\mu_{X_n}$ converges weakly to $\mu_Y$ as $n\to\infty$.
Then, each $p_n$ enforces $\varepsilon_n$-concentration of $X_n$ to $Y$
for some sequence $\varepsilon_n \to 0$
if and only if we have the following {\rm(1)}, {\rm(2)}, and {\rm(3)}.
\begin{enumerate}
\item Each $p_n$ is $1$-Lipschitz up to some additive error $\varepsilon_n'$
such that $\varepsilon_n' \to 0$ as $n\to\infty$.
\item $\{p_n\}$ effectuates relative concentration of $X_n$ over $Y$.
\item For any two Borel subsets $A_1, A_2\subset Y$ and any $\kappa > 0$,
we have
\begin{align*}
\limsup_{n\to\infty} d_+(p_n^{-1}(A_1),p_n^{-1}(A_2);+\kappa)
\le d_Y(A_1,A_2)+\diam A_1 +\diam A_2.
\end{align*}
\end{enumerate}
\end{thm}
As is mentioned before, (2) means that every fiber of $p_n$ concentrates
to a one-point mm-space.
(3) means that all fibers of $p_n$ are almost parallel to each other
for $n$ large enough.
\begin{proof}
Let $p_n : X_n \to Y$ be as in the theorem.
If each $p_n$ enforces $\varepsilon_n$-concentration of $X_n$ to $Y$
for a sequence $\varepsilon_n \to 0$, then
Lemma \ref{lem:1-Lip-up-to-Lip1}, Corollary \ref{cor:Lip1-ObsDiam},
and Lemma \ref{lem:kappa-dist} respectively imply
(1), (2), and (3) of the theorem.
Conversely we assume (1), (2), and (3).
Lemma \ref{lem:1-Lip-up-to-Lip1} yields that
$p_n^*\mathcal{L}{\it ip}_1(Y) \subset B_{\varepsilon_n'}(\mathcal{L}{\it ip}_1(X_n))$
for some $\varepsilon_n' \to 0$.
Let $\varepsilon > 0$ be an arbitrary number.
It suffices to prove that
$\mathcal{L}{\it ip}_1(X_n)\subset B_{11\varepsilon}(p_n^*\mathcal{L}{\it ip}_1(Y))$
for all $n$ large enough.
Take any $1$-Lipschitz functions $f_n : X_n \to \field{R}$, $n=1,2,\dots$.
For the proof, it suffices to find $1$-Lipschitz functions
$g_n : Y\to \field{R}$ such that
\begin{align} \label{eq:pn2}
\dKF(f_n,g_n \circ p_n) \le 11\varepsilon
\end{align}
for every sufficiently large $n$.
Our idea to find such $g_n$ is
to take the L\'evy mean of $f_n$ along each fiber of $p_n$.
There are finitely many open subsets $B_1,B_2,\dots,B_N \subset Y$
such that $\mu_Y(\partial B_i)=0$, $\diam B_i < \varepsilon$, and
\begin{align*}
\mu_Y\Big(Y\setminus \bigcup_{i=1}^N B_i \Big)< \varepsilon.
\end{align*}
For each $i=1,2, \cdots , N$, we take a point $y_i\in B_i$.
We put $A_{in}:=p_n^{-1}(B_i)$ and
$\nu_{in}:=(1/\mu_{X_n}(A_{in}))\mu_{X_n}|_{A_{in}}$.
Note that $\mu_{X_n}(A_{in})$ converges to $\mu_Y(B_i)$ as $n\to\infty$.
Setting $\rho_{ij} := d_Y(y_i,y_j) + 5\varepsilon$,
we take a number $\lambda$ such that $0 < \lambda \rho_{ij} < 1/4$
for any $i,j$.
\begin{clm}\label{clm:di-rho}
For any $i,j$, and for every sufficiently large $n$, we have
\begin{align*}
d_P^{(\lambda)}(\nu_{in},\nu_{jn}) \leq \rho_{ij}.
\end{align*}
\end{clm}
\begin{proof}
We fix $i$ and $j$.
Let $C_n \subset X_n$ be any Borel subsets.
It suffices to prove that
\begin{align}\label{ccs14}
\nu_{jn}(B_{\rho_{ij}}(C_n))\geq
\nu_{in}(C_n)-\lambda \rho_{ij}
\end{align}
for every sufficiently large $n$.
Setting $C_n':=C_n\cap A_{in}$, we
are going to prove that
\begin{align}\label{ccs15}
\nu_{jn}(B_{\rho_{ij}}(C_n'))\geq
\nu_{in}(C_n')-\lambda \rho_{ij},
\end{align}
which is stronger than \eqref{ccs14}.
We take a number $\kappa$ such that
\[
0 < \kappa
\leq \lambda \rho_{ij} \inf_n\min\{\mu_{X_n}(A_{in}),\mu_{X_n}(A_{jn})\}.
\]
If $\mu_{X_n}(C_n')< \kappa$, then we have
$\nu_{in}(C_n') < \lambda \rho_{ij}$ and so
\eqref{ccs15} holds.
Assume that $\mu_{X_n}(C_n') \ge \kappa $.
We define a function $F_n:A_{jn}\to \field{R}$ by
$F_n(x):=d_{X_n}(x,C_n')$, $x \in A_{jn}$, and set
\begin{align*}
D_n := \{\; x \in A_{jn} \; ; \; |F_n(x)-\lm(F_n;\nu_{jn})|
\le \varepsilon\;\}.
\end{align*}
It follows from (2) and Lemma \ref{lem:LeRad-ObsDiam} that
$\LeRad(A_{jn};-\kappa') < \varepsilon$
for every sufficiently large $n$ and for any $\kappa' > 0$,
which implies that $\mu_{X_n}(A_{jn} \setminus D_n)$ converges to zero
as $n\to\infty$.
Remarking $\kappa < \mu_{X_n}(A_{jn})/4$,
we have
\begin{align*}
\mu_{X_n}(D_n) \ge \kappa \quad\text{and}\quad
\nu_{jn}(D_n) \ge 1 - \lambda\rho_{ij}
\end{align*}
for every sufficiently large $n$.
In the following, we assume $n$ to be large enough.
By $\diam B_i<\varepsilon$ and (3), we have
\begin{align*}
d_{X_n}(C_n', D_n)\leq
d_+(A_{in},A_{jn};+\kappa) < d_Y(y_i,y_j) + 2\varepsilon.
\end{align*}
We take a point $a \in D_n$ in such a way that
\begin{align*}
F_n(a) < d_{X_n}(C_n',D_n)+\varepsilon.
\end{align*}
Then,
\begin{align*}
\lm(F_n;\nu_{jn}) &\le F_n(a) + \varepsilon
< d_{X_n}(C_n',D_n) + 2\varepsilon\\
&< d_Y(y_i,y_j) + 4\varepsilon.
\end{align*}
For any $x \in D_n$,
\begin{align*}
d_{X_n}(x,C_n') = F_n(x) \le \lm(F_n;\nu_{jn})+\varepsilon < \rho_{ij},
\end{align*}
which implies $B_{\rho_{ij}}(C_n') \supset D_n$ and so
\begin{align*}
\nu_{jn}(B_{\rho_{ij}}(C_n')) \ge \nu_{jn}(D_n)
\ge 1-\lambda\rho_{ij}
\ge \nu_{in}(C_n')-\lambda \rho_{ij}.
\end{align*}
This completes the proof of the claim.
\end{proof}
Let us define a function $\tilde{g}_n: Y \to \field{R}$ by
\[
\tilde{g}_n(y) :=
\begin{cases}
\lm(f_n|_{A_{in}};\nu_{in}) & \text{if $y \in B_i$},\\
0 & \text{if $y \in Y \setminus \bigcup_{i=1}^N B_i$}.
\end{cases}
\]
We are going to prove that
$\tilde{g}_n$ is $1$-Lipschitz up to a small additive error.
Since $\lambda\, d_P^{(\lambda)}$ coincides with the Prohorov metric
with respect to $\lambda\, d_{X_n}$
and by Claim \ref{clm:di-rho},
Strassen's theorem proves that
there is a $\rho_{ij}$-transport plan $\pi_{ijn}$ between $\nu_{in}$ and
$\nu_{jn}$
such that $\pi_{ijn}(X_n\times X_n)\ge 1-\lambda \rho_{ij} > 3/4$.
Assume that $n$ is sufficiently large.
(2) together with $\diam B_l < \varepsilon$ implies that
$\ObsDiam(\nu_{ln};-1/4) < \varepsilon$ for every $l$.
Applying Lemma \ref{lem:lm-tra} we have,
for any $y\in B_i$ and $y'\in B_j$,
\begin{align*}
|\tilde{g}_n(y)-\tilde{g}_n(y')|
&= |\lm(f_n|_{A_{in}};\nu_{in} ) - \lm(f_n|_{A_{jn}};\nu_{jn} )|\\
&\le \rho_{ij}+ 2\varepsilon
= d_Y(y_i,y_j) + 7\varepsilon
< d_Y(y,y') + 9\varepsilon.
\end{align*}
Since
\begin{align*}
\lim_{n\to \infty} (p_n)_*\mu_{X_n}\Big(\bigcup_{i=1}^N B_i
\Big) = \mu_Y \Big( \bigcup_{i=1}^N B_i \Big)\geq 1-\varepsilon,
\end{align*}
the function $\tilde{g}_n$ is $1$-Lipschitz up to $9\varepsilon$
with respect to $(p_n)_*\mu_{X_n}$ for every sufficiently large $n$.
By Lemma \ref{lem:Lip-approx},
there is a $1$-Lipschitz function $g_n : Y \to \field{R}$ such that
\begin{align} \label{eq:pn3}
\dKF(g_n \circ p_n, \tilde{g}_n \circ p_n)\leq 9\varepsilon.
\end{align}
Let $\kappa := \varepsilon/N$.
For every sufficiently large $n$,
since $\LeRad(\nu_{in};-\kappa) \le \ObsDiam(\nu_{in};-\kappa)
< \varepsilon$, we have
\begin{align*}
&\mu_{X_n}(|f_n-\tilde{g}_n \circ p_n| > \varepsilon)\\
&\leq \sum_{i=1}^N\mu_{X_n}(
\{\; x \in A_{in} \mid |f_n(x)-\lm(f_n|_{A_{in}};\nu_{in})| >
\varepsilon \;\} )\\
&\quad + \mu_{X_n}\Big(X_n \setminus \bigcup_{i=1}^N A_{in}\Big)\\
&\leq N\kappa + \varepsilon = 2\varepsilon,
\end{align*}
so that $\dKF(f_n,\tilde{g}_n \circ p_n)\leq 2\varepsilon$.
Combining this with (\ref{eq:pn3}) implies (\ref{eq:pn2}). This
completes the proof of the theorem.
\end{proof}
\section{Wasserstein distance
and curvature-dimension condition}
This section is devoted to a quick overview on
the Wasserstein distance and the curvature-dimension condition.
Let $X$ be a complete separable metric space.
\begin{defn}[Wasserstein distance]
\index{Wasserstein distance/metric}
\index{Lp-Wasserstein distance/metric@$L_p$-Wasserstein distance/metric}
Let $\mu$ and $\nu$ be two Borel probability measures on $X$.
For $1 \le p < +\infty$,
the \emph{$L_p$-Wasserstein distance} between $\mu$ and $\nu$
is defined to be
\[
W_p(\mu,\nu) := \left( \inf_\pi \int_{X \times X} d_X(x,x')^p\;d\pi(x,x')
\right)^{\frac{1}{p}} \quad(\le +\infty),
\]
where $\pi$ runs over all transport plans between $\mu$ and $\nu$.
It is known that, if $W_p(\mu,\nu) < +\infty$, this infimum is achieved by some transport plan,
which we call an \emph{optimal transport plan for $W_p(\mu,\nu)$}.
\index{optimal transport plan}
\end{defn}
It follows from H\"older's inequality that
$W_p(\mu,\nu) \le W_q(\mu,\nu)$ for $1 \le p \le q < +\infty$.
\begin{defn}[$P_p(X)$]
\index{Pp(X)@$P_p(X)$}
Let $1 \le p < +\infty$.
Denote by $P_p(X)$ the set of Borel probability measures $\mu$ on $X$
with finite \emph{$p^{th}$ moment},
i.e.,
\[
W_p(\mu,\delta_{x_0})^p = \int_X d_X(x_0,x)^p \;d\mu_X(x) < +\infty
\]
for some point $x_0 \in X$.
\index{moment} \index{pth moment@$p^{th}$ moment}
The $L_p$-Wasserstein distance defines a metric on $P_p(X)$,
called the \emph{$L_p$-Wasserstein metric}.
\end{defn}
\begin{lem}[cf.~\cite{Villani:topics}*{Theorem 7.12}] \label{lem:Wp-weakconv}
Let $X$ be a complete separable metric space.
Let $\mu$ and $\mu_n$, $n=1,2,\dots$, be measures
in $P_p(X)$, $1 \le p < +\infty$.
Then the following {\rm(1)} and {\rm(2)} are equivalent to each other.
\begin{enumerate}
\item $\lim_{n\to\infty} W_p(\mu_n,\mu) = 0$.
\item $\mu_n$ converges weakly to $\mu$ as $n\to\infty$
and
\[
\lim_{R\to +\infty} \limsup_{n\to\infty} \int_{X \setminus B_R(x_0)} d_X(x_0,x)^p \;d\mu_n(x)
= 0
\]
for some {\rm(}and any{\rm)} point $x_0 \in X$.
\end{enumerate}
\end{lem}
\begin{thm}[Kantorovich-Rubinstein duality;
cf.~\cite{Villani:oldnew}*{Remark 6.5}]
\index{Kantorovich-Rubinstein duality}
\label{thm:KR-duality}
Let $X$ be a complete separable metric space.
For any measures $\mu,\nu \in P_1(X)$ we have
\[
W_1(\mu,\nu) = \sup_{f \in \mathcal{L}{\it ip}_1(X)} \left( \int_X f \;d\mu - \int_X f \;d\nu
\right).
\]
\end{thm}
\begin{defn}[Relative entropy]
\index{relative entropy} \index{Ent@$\Ent$}
Let $\mu$ and $\nu$ be two probability measures on a set.
The \emph{relative entropy} $\Ent(\nu|\mu)$ of $\nu$
with respect to $\mu$ is defined as follows.
If $\nu$ is absolutely continuous with respect to $\mu$, then
\[
\Ent(\nu|\mu) := \int_X U\left(\frac{d\nu}{d\mu}\right) \,d\mu
\quad (\le +\infty),
\]
otherwise $\Ent(\nu|\mu) := +\infty$,
where
\[
U(r) :=
\begin{cases}
0 &\text{if $r = 0$},\\
r\log r &\text{if $r > 0$}.
\end{cases}
\]
\index{Ur@$U(r)$}
\end{defn}
\begin{lem} \label{lem:Ent-push}
Let $p : X \to Y$ be a Borel measurable map between
two complete separable metric spaces $X$ and $Y$,
and let $\mu$ and $\nu$ be two Borel probability measures on $X$ such that
$\nu$ is absolutely continuous with respect to $\mu$.
Then, $p_*\nu$ is absolutely continuous
with respect to $p_*\mu$ and we have
\[
\Ent(p_*\nu|p_*\mu) \le \Ent(\nu|\mu).
\]
\end{lem}
\begin{proof}
Let $\{\mu_y\}_{y\in Y}$ be the disintegration of $\mu$ for $p : X \to Y$.
We set $\rho := \frac{d\nu}{d\mu}$ and
$\tilde\rho(y) := \int_{p^{-1}(y)} \rho \,d\mu_y$.
For any bounded continuous function $f : Y \to \field{R}$,
\begin{align*}
\int_Y f \,dp_*\nu &= \int_X f\circ p \,d\nu = \int_X (f\circ p) \rho\,d\mu\\
&= \int_Y \int_{p^{-1}(y)} (f\circ p)\rho\,d\mu_y dp_*\mu(y)\\
&= \int_Y f(y) \int_{p^{-1}(y)} \rho \,d\mu_y dp_*\mu(y)
= \int_Y f\tilde\rho \,dp_*\mu,
\end{align*}
which implies that $\frac{dp_*\nu}{dp_*\mu} = \tilde\rho$.
It follows from Jensen's inequality that
\begin{align*}
\Ent(p_*\nu|p_*\mu) &= \int_Y U(\tilde\rho) \,dp_*\mu
\le \int_X \int_{p^{-1}(y)} U(\rho) \,d\mu_y dp_*\mu(y)\\
&= \int_X U(\rho) \,d\mu = \Ent(\nu|\mu).
\end{align*}
This completes the proof.
\end{proof}
\begin{lem}[cf.~\cite{Villani:oldnew}*{Theorem 29.20(i)}]
\label{lem:Ent-lim}
Let $\mu$, $\nu$, $\mu_n$, and $\nu_n$, $n=1,2,\dots$, be
Borel probability measures on a compact metric space.
If $\mu_n$ and $\nu_n$ converge weakly to $\mu$ and $\nu$
respectively as $n\to\infty$, then we have
\[
\Ent(\nu|\mu) \le \liminf_{n\to\infty} \Ent(\nu_n|\mu_n).
\]
\end{lem}
\begin{defn}[Curvature-dimension condition] \label{defn:CD}
\index{curvature-dimension condition}
\index{CD(K,infty)@$\CD(K,\infty)$}
Let $K$ be a real number.
We say that an mm-space $X$ satisfies
the \emph{curvature-dimension condition $\CD(K,\infty)$}
if for any $\nu_0,\nu_1 \in P_2(X)$,
any $\varepsilon > 0$, and
any $t \in (\,0,1\,)$, there exists a measure $\nu_t \in P_2(X)$
such that
\begin{align}
\label{eq:CD1}
W_2(\nu_t,\nu_i) &\le t^{1-i}(1-t)^i W_2(\nu_0,\nu_1) + \varepsilon,
\quad i=0,1,\\
\label{eq:CD2}
\Ent(\nu_t|\mu_X) &\le (1-t)\Ent(\nu_0|\mu_X) + t\Ent(\nu_1|\mu_X)\\
&\quad -\frac{1}{2} K t(1-t)W_2(\nu_0,\nu_1)^2 + \varepsilon.\notag
\end{align}
\end{defn}
The following important theorem was first predicted by Otto \cite{Otto}.
\begin{thm}[\cites{CMS:PL,CMS:interp,vRS,Sturm:convex}]
Let $X$ be a complete Riemannian manifold $X$ and
$K$ a real number.
Then, $\CD(K,\infty)$ for $X$ is equivalent to $\Ric_X \ge K$.
\end{thm}
\begin{defn}[$P^{cb}(X)$]
\index{Pcb(X)@$P^{cb}(X)$}
For an mm-space $X$, we denote by $P^{cb}(X)$
the set of Borel probability measures $\nu$ on $X$ with compact support
such that
$\nu$ is absolutely continuous with respect to $\mu_X$
and that the Radon-Nikodym derivative $\frac{d\nu}{d\mu_X}$
is essentially bounded on $X$.
\end{defn}
Note that $P^{cb}(X) \subset P_p(X)$ for any $p$ with $1 \le p < +\infty$.
\begin{lem} \label{lem:WpEnt-approx}
Let $X$ be an mm-space and $\nu \in P_p(X)$ a measure
with $\Ent(\nu|\mu_X) < +\infty$, where $1 \le p < +\infty$.
Then, for any $\varepsilon > 0$
there exists a measure $\tilde\nu \in P^{cb}(X)$ such that
\[
W_p(\tilde\nu,\nu) < \varepsilon \quad\text{and}\quad
|\Ent(\tilde\nu|\mu_X) - \Ent(\nu|\mu_X)| < \varepsilon.
\]
\end{lem}
\begin{proof}
By the inner regularity of $\mu_X$,
there is a monotone nondecreasing sequence of compact subsets
$K_n \subset X$, $n=1,2,\dots$, such that
\[
\lim_{n\to\infty} \mu_X(X \setminus K_n) = 0.
\]
Let $\rho := \frac{d\nu}{d\mu_X}$ and
\[
\rho_n(x) :=
\begin{cases}
\frac{1}{c_n} \min\{\rho(x),n\} & \text{if $x \in K_n$},\\
0 & \text{if $x \in X \setminus K_n$},
\end{cases}
\]
where $c_n := \int_{K_n} \min\{\rho(x),n\} \,d\mu_X(x)$.
The measure $\nu_n := \rho_n \mu_X$ belongs to $P^{cb}(X)$.
It suffices to prove that
\begin{align}
\label{eq:WpEnt-approx1}
& \lim_{n\to\infty} W_p(\nu_n,\nu) = 0,\\
\label{eq:WpEnt-approx2}
& \lim_{n\to\infty} \Ent(\nu_n|\mu_X) = \Ent(\nu|\mu_X).
\end{align}
Since $\lim_{n\to\infty} c_n = 1$ and $\mu_X(\bigcup_{n=1}^\infty K_n) = 1$,
the function $\rho_n$ converges to $\rho$ $\mu_X$-a.e.~as $n\to\infty$.
It is easy to see that $\nu_n$ converges weakly to $\nu$ as $n\to\infty$.
Moreover, for a point $x_0 \in X$,
\begin{align*}
&\lim_{R\to\infty}\limsup_{n\to\infty} \int_{X\setminus B_R(x_0)} d_X(x_0,x)^p \;d\nu_n(x)\\
&\le \lim_{R\to\infty}\limsup_{n\to\infty} \frac{1}{c_n}
\int_{X\setminus B_R(x_0)} d_X(x_0,x)^p \;d\nu(x) = 0.
\end{align*}
By Lemma \ref{lem:Wp-weakconv} we obtain \eqref{eq:WpEnt-approx1}.
Letting
\[
I_n := \int_{\{0 < \rho \le 1\}} \rho_n\log\rho_n \;d\mu_X
\quad\text{and}\quad
J_n := \int_{\{\rho > 1\}} \rho_n\log\rho_n \;d\mu_X,
\]
we have $\Ent(\nu_n|\mu_X) = I_n + J_n$.
The dominated convergence theorem implies
\[
\lim_{n\to\infty} I_n = \int_{\{0 < \rho \le 1\}} \rho\log\rho \;d\mu_X.
\]
For every sufficiently large $n$ we have $c_n \ge 1/2$,
which implies $\rho_n \le 2\rho$ and then
$0 \le \rho_n\log\rho_n \le 2\rho\log(2\rho)$ on $\{\rho > 1\}$.
We see that $2\rho\log(2\rho)$ is $\mu_X$-integrable
on $\{\rho > 1\}$. By the dominated convergence theorem,
\[
\lim_{n\to\infty} J_n = \int_{\{\rho > 1\}} \rho\log\rho \;d\mu_X.
\]
We thus obtain \eqref{eq:WpEnt-approx2}.
This completes the proof.
\end{proof}
\begin{lem} \label{lem:CD-Pcb}
If we assume $\nu_0,\nu_1 \in P^{cb}(X)$ in the definition of
$\CD(K,\infty)$, then we still have $\CD(K,\infty)$.
\end{lem}
\begin{proof}
We assume the condition of $\CD(K,\infty)$ for any $\nu_0,\nu_1 \in P^{cb}(X)$.
Take any measures $\nu_0,\nu_1 \in P_2(X)$.
If $\Ent(\nu_i|\mu_X) = +\infty$ for $i=0$ or $1$,
then \eqref{eq:CD2} is trivial and \eqref{eq:CD1} follows from
the dense property of $P^{cb}(X)$ in $P_2(X)$.
Assume that $\Ent(\nu_i|\mu_X) < +\infty$ for $i=0,1$.
By Lemma \ref{lem:WpEnt-approx},
for any $\varepsilon > 0$ we find two measures
$\tilde\nu_0,\tilde\nu_1 \in P^{cb}(X)$
in such a way that
$W_2(\tilde\nu_i,\nu_i) < \varepsilon$
and $|\Ent(\tilde\nu_i|\mu_X)-\Ent(\nu_i|\mu_X)| < \varepsilon$
for $i=0,1$.
By the assumption, there is a measure $\nu_t \in P_2(X)$
for any $t \in (\,0,1\,)$
such that $\tilde\nu_0$, $\tilde\nu_1$, and $\nu_t$ together
satisfy \eqref{eq:CD1} and \eqref{eq:CD2}.
Remarking that $\varepsilon$ can be taken to be arbitrarily small
compared with $K$ and $W_2(\nu_0,\nu_1)$,
we obtain the lemma.
\end{proof}
\begin{defn}[Length of a curve]
\index{length}
For a continuous curve $c : [\,a,b\,] \to X$ on a metric space $X$,
we define the \emph{length $L(c)$ of $c$} by
\[
L(c) := \sup_{a = s_0 < s_1 < \dots < s_k = b} \sum_{i=1}^k d_X(c(s_{i-1}),c(s_i))
\quad (\le +\infty).
\]
A curve in a metric space is said to be \emph{rectifiable}
if the length of the curve is finite.
\index{rectifiable}
\end{defn}
\begin{defn}[Intrinsic metric space]
\index{intrinsic metric space} \index{length space}
An \emph{intrinsic metric space} (or \emph{length space})
is, by definition, a metric space such that,
for any given two points in the space,
the distance between them is equal to
the infimum of the lengths of curves joining them.
We assume that any two points in an intrinsic metric space
has finite distance unless otherwise stated.
In particular, any two points in an intrinsic metric space
can always be joined by a rectifiable curve.
\end{defn}
\begin{defn}[Minimal geodesic, geodesic space]
\index{minimal geodesic} \index{geodesic space}
A curve $\gamma : [\,a,b\,] \to X$ on a metric space $X$
is called a \emph{minimal geodesic} if
$d_X(\gamma(s),\gamma(t)) = |s-t|$ for all $s,t \in [\,a,b\,]$.
We say that a metric space $X$ is a \emph{geodesic space}
if for any two points $x,y \in X$ there exists a minimal geodesic
$\gamma$ joining $x$ and $y$ such that
$d_X(x,y) = L(\gamma)$.
\end{defn}
\begin{prop}[\cite{Sturm:geoI}*{Remark 4.6(iii)}]
If an mm-space $X$ satisfies $\CD(K,\infty)$ for some real number $K$,
then $(P_2(X),W_2)$ and $X$ are both intrinsic metric spaces.
\end{prop}
\begin{prop} \label{prop:CD-Sep}
If an mm-space $X$ satisfies $\CD(K,\infty)$ for a real number $K > 0$,
then
\begin{align}
\tag{1}
\Sep(X;\kappa_0,\kappa_1)
&\le \sqrt{\frac{4}{K} \log\frac{1}{\kappa_0\kappa_1}},\\
\tag{2}
\ObsDiam(X;-\kappa)
&\le \sqrt{\frac{8}{K} \log\frac{2}{\kappa}}
\end{align}
for any $\kappa,\kappa_0,\kappa_1 > 0$ with $\kappa_0+\kappa_1 < 1$.
\end{prop}
\begin{proof}
(2) follows from (1) and Proposition \ref{prop:ObsDiam-Sep}.
We prove (1).
Let $A_0,A_1 \subset X$ be any two Borel subsets with
$\mu_X(A_i) \ge \kappa_i$, $i=0,1$, and let
$\nu_i := \mu_X(A_i)^{-1} \mu_X|_{A_i}$.
By $\CD(K,\infty)$, for any $\varepsilon > 0$
there is a measure $\nu_{1/2} \in P_2(X)$ such that
\[
\Ent(\nu_{1/2}|\mu_X) \le \frac{1}{2}\Ent(\nu_0|\mu_X)
+ \frac{1}{2}\Ent(\nu_1|\mu_X)
- \frac{1}{8}K\,W_2(\nu_0,\nu_1)^2 + \varepsilon.
\]
We have $\Ent(\nu_i|\mu_X) = \log(1/\mu_X(A_i)) \le \log(1/\kappa_i)$
for $i=0,1$.
Jensen's inequality implies $\Ent(\nu_{1/2}|\mu_X) \ge 0$.
Therefore,
\[
0 \le \frac{1}{2}\log\frac{1}{\kappa_0} + \frac{1}{2}\log\frac{1}{\kappa_1}
- \frac{1}{8} K\,W_2(\nu_0,\nu_1)^2 + \varepsilon
\]
and, by the arbitrariness of $\varepsilon$,
\[
d_X(A_0,A_1)^2 \le W_2(\nu_0,\nu_1)^2
\le \frac{4}{K} \log\frac{1}{\kappa_0\kappa_1}.
\]
This completes the proof.
\end{proof}
The following corollary is a direct consequence of
Proposition \ref{prop:CD-Sep}.
\begin{cor} \label{cor:CD-Sep}
Let $X_n$, $n=1,2,\dots$, be mm-spaces.
If each $X_n$ satisfies $\CD(K_n,\infty)$ for some sequence
of real numbers $K_n \to +\infty$, then
$\{X_n\}$ is a L\'evy family.
\end{cor}
\section{Stability of curvature-dimension condition}
\label{sec:stab-CD}
A main purpose of this section is to prove that
the curvature-dimension condition is stable
under concentration.
\begin{lem} \label{lem:dH-me}
Let $p, q : X \to Y$ be two Borel measurable maps from an mm-space $X$
to a metric space $Y$.
Then we have
\[
d_H(p^*\mathcal{L}{\it ip}_1(Y),q^*\mathcal{L}{\it ip}_1(Y)) \le \dKF(p,q).
\]
\end{lem}
\begin{proof}
Let $f \in \mathcal{L}{\it ip}_1(Y)$ be any function.
Since $|f(p(x)) - f(q(x))| \le d_Y(p(x),q(x))$ for any $x \in X$,
we have
\[
\mu_X(|f\circ p - f\circ q| > \varepsilon)
\le \mu_X(d_Y(p,q) > \varepsilon)
\]
for any $\varepsilon \ge 0$, which implies that
\[
\dKF(f\circ p,f\circ q) \le \dKF(p,q).
\]
This proves the lemma.
\end{proof}
\begin{prop} \label{prop:pq}
Let $X_n$ and $Y$ be mm-spaces and let $p_n,q_n : X_n \to Y$
be Borel measurable maps, $n=1,2,\dots$, such that
$\dKF(p_n,q_n) \to 0$ as $n\to\infty$.
Then we have the following {\rm(1)} and {\rm(2)}.
\begin{enumerate}
\item If each $p_n$ enforces $\varepsilon_n$-concentration of $X_n$ to $Y$
for some $\varepsilon_n \to 0$,
then so does $q_n$.
\item If $(p_n)_*\mu_{X_n}$ converges weakly to $\mu_Y$ as $n\to\infty$,
then so does $(q_n)_*\mu_{X_n}$.
\end{enumerate}
\end{prop}
\begin{proof}
(1) follows from Lemma \ref{lem:dH-me}
and (2) from Lemma \ref{lem:di-me}.
\end{proof}
\begin{defn}[Bounded values on exceptional domains]
\index{bounded values on exceptional domains}
Let $X_n$, $n=1,2,\dots$, be mm-spaces and $Y$ a metric space.
Let $p_n : X_n \to Y$, $n=1,2,\dots$, be Borel measurable maps such that
each $p_n$ is $1$-Lipschitz up to an additive error $\varepsilon_n$
with $\varepsilon_n \to 0$.
We say that $\{p_n\}$ \emph{has bounded values on exceptional domains}
if we have
\[
\limsup_{n\to\infty} \sup_{x \in X \setminus \tilde X_n} d_Y(p_n(x),y_0) < +\infty
\]
for a point $y_0 \in Y$,
where $\tilde X_n$ is a non-exceptional domain of $p_n$
for the additive error $\varepsilon_n$.
\end{defn}
Note that if $\diam Y < +\infty$,
then $\{p_n\}$ always has bounded values on exceptional domains.
\begin{prop} \label{prop:bdd-exc-dom}
Assume that a sequence of mm-spaces $X_n$, $n=1,2,\dots$,
concentrates to an mm-space $Y$.
Then, there exist Borel measurable maps $p_n : X_n \to Y$, $n=1,2,\dots$,
such that
\begin{enumerate}
\item each $p_n$ enforces $\varepsilon_n$-concentration of $X_n$ to $Y$
for some $\varepsilon_n \to 0$,
\item $(p_n)_*\mu_{X_n}$ converges weakly to $\mu_Y$ as $n\to\infty$,
\item $\{p_n\}$ has bounded values on exceptional domains.
\end{enumerate}
\end{prop}
Note that (1) implies that $p_n$ is $1$-Lipschitz up to an
additive error $\varepsilon_n \to 0$, which defines
an exceptional domain of $p_n$ for (3).
\begin{proof}
By Corollary \ref{cor:enforce}, there is a sequence of
Borel measurable maps $p_n' : X_n \to Y$, $n=1,2,\dots$,
$\varepsilon_n'$-enforcing concentration of $X_n$ to $Y$ with
$\varepsilon_n' \to 0$ such that
$(p_n')_*\mu_{X_n}$ converges weakly to $\mu_Y$ as $n\to\infty$.
By Theorem \ref{thm:pn}, $p_n'$ is $1$-Lipschitz up to
some additive error $\varepsilon_n \to 0$ with a non-exceptional domain
$\tilde{X}_n \subset X_n$.
Define a map $p_n : X_n \to X$ by
\[
p_n(x) :=
\begin{cases}
p_n'(x) &\text{if $x \in \tilde{X}_n$},\\
y_0 &\text{if $x \in X_n \setminus \tilde{X}_n$},
\end{cases}
\]
for $x \in X$, where $y_0 \in Y$ is a fixed point.
Each $p_n$ is Borel measurable, $1$-Lipschitz up to $\varepsilon_n$
with the non-exceptional domain $\tilde{X}_n$, and satisfies
$\dKF(p_n,p_n') \le \varepsilon_n$.
(3) follows from the definition of $p_n$.
Proposition \ref{prop:pq} proves (1) and (2).
This completes the proof.
\end{proof}
\begin{defn}[$X^D$]
\index{XD@$X^D$}
For an mm-space $X = (X,d_X,\mu_X)$ and for a real number $D > 0$,
we define an mm-space $X^D$ to be $(X,d_{X^D},\mu_X)$,
where $d_{X^D}(x,x') := \min\{d_X(x,x'),D\}$ for $x,x' \in X$.
\end{defn}
For a Borel subset $B$ of an mm-space $X$ with $\mu_X(B) > 0$, we set
\[
\mu_B := \frac{\mu_X|_{B}}{\mu_X(B)}.
\]
\begin{lem} \label{lem:B0B1}
Let $p_n : X_n \to Y$ be Borel measurable maps
between mm-spaces $X_n$ and $Y$, $n=1,2,\dots$, such that
each $p_n$ enforces $\varepsilon_n'$-concentration of $X_n$ to $Y$
with $\varepsilon_n' \to 0$,
and that $(p_n)_*\mu_{X_n}$ converges weakly to $\mu_Y$ as $n \to \infty$.
For a real number $\delta > 0$,
we give two Borel subsets $B_0, B_1 \subset Y$ such that
\[
\diam B_i \le \delta, \quad
\mu_Y(B_i) > 0, \quad\text{and}\quad \mu_Y(\partial B_i) = 0
\]
for $i = 0,1$, and set
\[
\tilde{B}_i := p_n^{-1}(B_i) \cap \tilde{X}_n,
\]
where $\tilde{X}_n$ is a non-exceptional domain of $p_n$ for an additive error
$\varepsilon_n \to 0$ as $n\to\infty$.
Then, there exist Borel probability measures $\tilde\mu_0^n,\tilde\mu_1^n$
on $X_n$
and transport plans $\tilde\pi^n$ between $\tilde\mu_0^n$ and $\tilde\mu_1^n$,
$n = 1,2,\dots$, such that, for every sufficiently large $n$,
\begin{enumerate}
\item $\tilde\mu_i^n \le (1+O(\delta^{1/2}))\mu_{\tilde{B_i}}$,
where $O(\cdots)$ is a Landau symbol,
\item for any $x_i \in \tilde{B}_i$, $i=0,1$,
\[
d_{X_n}(x_0,x_1) \ge d_Y(B_0,B_1) - \varepsilon_n,
\]
\item $\supp\tilde\pi^n \subset \{d_{X_n} \le d_Y(B_0,B_1) + \delta^{1/2}\}$,
\item $-\varepsilon_n \le W_p(\tilde\mu_0^n,\tilde\mu_1^n)
- d_Y(B_0,B_1) \le \delta^{1/2}$ for any $p \ge 1$.
\end{enumerate}
\end{lem}
\begin{proof}
Since $p_n$ is $1$-Lipschitz up to $\varepsilon_n$ with
the non-exceptional domain $\tilde{X}_n \subset X_n$,
we have $\mu_{X_n}(X_n \setminus \tilde{X}_n) \le \varepsilon_n$ and
\[
d_Y(p_n(x),p_n(x')) \le d_{X_n}(x,x') + \varepsilon_n
\]
for any $x,x' \in \tilde{X}_n$.
This implies (2).
Put
\[
D := \sup_{y_0\in B_0,\; y_1\in B_1} d_Y(y_0,y_1) + \delta^{1/2}
\]
and let $f \in \mathcal{L}{\it ip}_1(X_n^D)$ be any function.
Note that $f \in \mathcal{L}{\it ip}_1(X_n)$.
We take any number $\kappa > 0$.
Since Theorem \ref{thm:pn} implies that
$\{p_n\}$ effectuates relative concentration of $X_n$ over $Y$,
we have
$\ObsDiam(p_n^{-1}(B_i);-\kappa) < 2\delta$
and so
\[
\diam(f_*(\mu_{X_n}|_{p_n^{-1}(B_i)});\mu_{X_n}(p_n^{-1}(B_i))-\kappa) < 2\delta
\]
for all sufficiently large $n$ and for $i=0,1$.
There is a number $c_i \in f(X_n)$ such that
\[
\mu_{X_n}(\{|f-c_i| \le \delta\} \cap p_n^{-1}(B_i))
\ge \mu_{X_n}(p_n^{-1}(B_i)) - \kappa.
\]
Note that $\mu_{X_n}(p_n^{-1}(B_i))$ and $\mu_{X_n}(\tilde{B}_i)$ both
converge to $\mu_Y(B_i)$ as $n\to\infty$.
Since $\mu_{X_n}(X_n \setminus \tilde{X}_n) \le \varepsilon_n$,
the above inequality leads us to
\[
\mu_{X_n}(\{|f-c_i| \le \delta\} \cap \tilde{B}_i)
\ge \mu_{X_n}(\tilde{B}_i) - \kappa - 2\varepsilon_n,
\]
so that, by chosing $\kappa$ small enough,
\[
\mu_{\tilde{B}_i}(|f-c_i| \le \delta)
\ge 1-\frac{\kappa+2\varepsilon_n}{\mu_{X_n}(\tilde{B}_i)}
> 1- \frac{\delta}{D}
\]
for all sufficiently large $n$ and for $i=0,1$.
It follows from $f \in \mathcal{L}{\it ip}_1(X_n^D)$ that
$|f-c_i| \le D$ on $X_n$. We have
\begin{align} \label{eq:int-ci}
\left| \int_{X_n} f \,d\mu_{\tilde{B}_i} - c_i \right|
&\le \int_{\{|f-c_i| \le \delta\}} |f-c_i| \,d\mu_{\tilde{B}_i}
+ \int_{\{\delta < |f-c_i| \le D\}} |f-c_i| \,d\mu_{\tilde{B}_i}\\
&\le \delta + D \cdot \frac{\delta}{D} = 2\delta. \notag
\end{align}
Since $p_n$ enforces concentration of $X_n$ to $Y$,
we may assume that $d_H(\mathcal{L}{\it ip}_1(X_n),p_n^*\mathcal{L}{\it ip}_1(Y)) < \varepsilon_n$.
For the $f$ there is $g \in \mathcal{L}{\it ip}_1(Y)$ such that
$\dKF(g\circ p_n,f) \le \varepsilon_n$.
Letting $A := \{|g\circ p_n - f| \le \varepsilon_n\}$
we have $\mu_{X_n}(A) \ge 1-\varepsilon_n$.
Since
$\mu_{X_n}(\{|f-c_i| \le \delta\} \cap \tilde{B}_i)
\ge \mu_{X_n}(\tilde{B}_i)(1-\delta/D) > \varepsilon_n$
for all sufficiently large $n$,
the intersection
$A \cap \{|f-c_i|\le\delta\} \cap \tilde{B}_i$ is nonempty
for $i = 0,1$.
We take a point $x_i$ in this set and put $y_i := p_n(x_i)$.
It then holds that
$|g(y_i)-f(x_i)| \le \varepsilon_n$, $|f(x_i)-c_i| \le \delta$,
and $y_i \in B_i$.
Therefore, setting $d_{01} := d_Y(B_0,B_1)$ we obtain
\begin{align*}
&\left| \int_{X_n} f \,d\mu_{\tilde{B}_0} - \int_{X_n} f \,d\mu_{\tilde{B}_1} \right|
\le |c_0-c_1| + 4\delta
\le |f(x_0)-f(x_1)| + 6\delta\\
&\le | g(y_0)-g(y_1) | + 2\varepsilon_n + 6\delta
\le d_Y(y_0,y_1) + 2\varepsilon_n + 6\delta\\
&\le d_{01} + 2\varepsilon_n + 8\delta < d_{01} + 9\delta
\end{align*}
for every sufficiently large $n$.
Note that the above estimate is uniform for all $f \in \mathcal{L}{\it ip}_1(X_n^D)$.
It follows from the Kantorovich-Rubinstein duality
(Theorem \ref{thm:KR-duality}) that
\begin{align}
\label{eq:W1-upper}
\hat W_1(\mu_{\tilde{B}_0},\mu_{\tilde{B}_1}) \le d_{01} + 9\delta,
\end{align}
where $\hat W_1$ denotes the $L_1$-Wasserstein metric on $P_1(X_n^D)$.
Take an optimal transport plan $\pi$ for
$\hat W_1(\mu_{\tilde{B}_0},\mu_{\tilde{B}_1})$.
Note that $d_{01} + \delta^{1/2} \le D$.
Setting $\xi := d_{X_n^D} - d_{01}$ we have, by \eqref{eq:W1-upper} and
(2),
\begin{align*}
9\delta &\ge \hat W_1(\mu_{\tilde{B}_0},\mu_{\tilde{B}_1}) - d_{01}
= \int_{X_n\times X_n} \xi\,d\pi\\
&= \int_{\{\xi < \delta^{1/2}\}} \xi\,d\pi
+ \int_{\{\xi \ge \delta^{1/2}\}} \xi\,d\pi
\ge -\varepsilon_n + \delta^{1/2} \pi(\xi \ge \delta^{1/2})
\end{align*}
and so $\pi(d_{X_n} \ge d_{01} + \delta^{1/2}) = \pi(\xi \ge \delta^{1/2})
\le 10\delta^{1/2}$ for every sufficiently large $n$.
We put
\begin{align*}
V_n &:= \pi(d_{X_n} \le d_{01} + \delta^{1/2}),\\
\tilde\pi^n &:= V_n^{-1} \pi|_{\{d_{X_n} \le d_{01} + \delta^{1/2}\}},\\
\tilde\mu_i^n &:= (\proj_i)_*\tilde\pi^n,
\end{align*}
where $\proj_0 : X_n \times X_n \to X_n$ is the first projection and
$\proj_1 : X_n \times X_n \to X_n$ the second.
(3) follows from the definition of $\tilde\pi^n$.
Since $\tilde\pi^n \le V_n^{-1}\pi$, we have
\[
\tilde\mu_i^n \le V_n^{-1}\mu_{\tilde{B}_i}
\le (1-10\delta^{1/2})^{-1} \mu_{\tilde{B}_i},
\]
which implies (1).
(4) is derived from (2) and (3).
This completes the proof.
\end{proof}
Let $\theta(\cdot) : \field{R} \to \field{R}$ be a function such that
$\theta(\varepsilon) \to 0$ as $\varepsilon \to 0$,
and $\theta(\cdot|\alpha_1,\alpha_2,\dots) : \field{R} \to \field{R}$ a function
depending on $\alpha_1,\alpha_2,\dots$
such that $\theta(\varepsilon|\alpha_1,\alpha_2,\dots) \to 0$
as $\varepsilon \to 0$.
We use $\theta(\cdots)$ like the Landau symbols.
\begin{lem} \label{lem:CD-conc}
Let $\{X_n\}$ be a sequence of mm-spaces satisfying $\CD(K,\infty)$
for a real number $K$, and $Y$ an mm-space.
Assume that a sequence of Borel measurable maps
$p_n : X_n \to Y$, $n=1,2,\dots$, satisfies {\rm(1)}, {\rm(2)},
and {\rm(3)} of Proposition \ref{prop:bdd-exc-dom}.
Then, for any $\nu_0,\nu_1 \in P^{cb}(Y)$
and any $t \in (\,0,1\,)$, there exist measures $\tilde\nu_t^n \in P(X_n)$,
$n=1,2,\dots$, such that
\begin{align}
\label{eq:lem-CD1}
\limsup_{n\to\infty} W_2((p_n)_*\tilde\nu_t^n,\nu_i)
&\le t^{1-i}(1-t)^i W_2(\nu_0,\nu_1),
\quad i=0,1,\\
\label{eq:lem-CD2}
\limsup_{n\to\infty} \Ent((p_n)_*\tilde\nu_t^n|(p_n)_*\mu_{X_n})
&\le (1-t)\Ent(\nu_0|\mu_Y) + t\Ent(\nu_1|\mu_Y)\\
&\quad -\frac{1}{2} K t(1-t)W_2(\nu_0,\nu_1)^2.\notag
\end{align}
\end{lem}
\begin{proof}
We take any $\nu_0,\nu_1 \in P^{cb}(Y)$ and fix them.
For any natural number $m$,
there are finitely many mutually disjoint Borel
subsets $B_j \subset Y$, $j=1,2,\dots,J$, such that
$\bigcup_{j=1}^J B_j = \supp\nu_0 \cup \supp\nu_1$,
$\diam B_j \le m^{-1}$, $\mu_Y(B_j) > 0$,
and $\mu_Y(\partial B_j) = 0$ for any $j$.
Take a point $y_j \in B_j$ for each $j$.
For each pair of $j,k = 1,2,\dots,J$,
we apply Lemma \ref{lem:B0B1} for $B_j$ and $B_k$ to obtain
measures $\tilde\mu_{jk}^{mn} \in P^{cb}(X_n)$, $n=1,2,\dots$, such that,
for every sufficiently large $n$,
\begin{align}
\label{eq:tildemujk1}
&\qquad\tilde\mu_{jk}^{mn} \le (1+\theta(m^{-1}))\mu_{\tilde{B}_j},\\
\label{eq:tildemujk2}
&|W_2(\tilde\mu_{jk}^{mn},\tilde\mu_{kj}^{mn})-d_Y(y_j,y_k)|
\le \theta(m^{-1}),
\end{align}
where $\tilde B_j := p_n^{-1}(B_j) \cap \tilde{X}_n$.
Replacing $\{n\}$ by a subsequence, we may assume that,
as $n\to\infty$, $(p_n)_*\tilde\mu_{jk}^{mn}$ converges weakly to
a measure, say $\tilde\mu_{jk}^m \in P^{cb}(Y)$,
because $(p_n)_*\tilde\mu_{jk}^{mn}$
is supported in the compact set $\supp\nu_0 \cup \supp\nu_1$.
Such the subsequence of $\{n\}$ is taken to be common
for all $j$, $k$, and $m$ by a diagonal argument.
Let $\pi$ be an optimal transport plan for $W_2(\nu_0,\nu_1)$
and let
\begin{alignat*}{2}
w_{jk} &:= \pi(B_j\times B_k), \\
\tilde\nu_0^{mn} &:= \sum_{j,k=1}^J w_{jk} \tilde\mu_{jk}^{mn}, & \qquad
\tilde\nu_1^{mn} &:= \sum_{j,k=1}^J w_{kj} \tilde\mu_{jk}^{mn},\\
\tilde\nu_0^m &:= \sum_{j,k=1}^J w_{jk} \tilde\mu_{jk}^m, & \qquad
\tilde\nu_1^m &:= \sum_{j,k=1}^J w_{kj} \tilde\mu_{jk}^m.
\end{alignat*}
It then holds that
$(p_n)_*\tilde\nu_i^{mn}$ converges weakly to
$\tilde\nu_i^m$ as $n \to \infty$ for any $m$ and $i=0,1$.
Since $(p_n)_*\mu_{\tilde{B}_j}$ converges weakly to $\mu_{B_j}$ as $n\to\infty$
and by \eqref{eq:tildemujk1}, we have
\[
\tilde\nu_0^m \le (1+\theta(m^{-1})) \sum_{j,k=1}^J w_{jk} \mu_{B_j}
= (1+\theta(m^{-1})) \sum_{j=1}^J \nu_0(B_j) \mu_{B_j}
\]
as well as
\[
\tilde\nu_1^m \le (1+\theta(m^{-1})) \sum_{j=1}^J \nu_1(B_j) \mu_{B_j}.
\]
Since $\sum_{j=1}^J \nu_i(B_j) \mu_{B_j} \to \nu_i$ as $m\to\infty$,
any weak limit of $\tilde\nu_i^m$ as $m\to\infty$ is less than
or equal to
$\nu_i$. Moreover, $\tilde\nu_i^m$ and $\nu_i$ are both
probability measures supported in the compact set
$\supp\nu_0 \cup \supp\nu_1$
and thus $\tilde\nu_i^m$ converges weakly to $\nu_i$
as $m\to\infty$ for $i=0,1$.
We next prove
\begin{align} \label{eq:Wnu01}
\lim_{m\to\infty}\liminf_{n\to\infty} W_2(\tilde\nu_0^{mn},\tilde\nu_1^{mn})
= \lim_{m\to\infty}\limsup_{n\to\infty} W_2(\tilde\nu_0^{mn},\tilde\nu_1^{mn})
= W_2(\nu_0,\nu_1)
\end{align}
in the following.
Take an optimal transport plan $\tilde\pi_{jk}$ for
$W_2(\tilde\mu_{jk}^{mn},\tilde\mu_{kj}^{mn})$ and set
$\tilde\pi' := \sum_{j,k} w_{jk} \tilde\pi_{jk}$.
Note that $\tilde\pi'$ is a (not necessarily optimal) transport
plan between $\tilde\nu_0^{mn}$ and $\tilde\nu_1^{mn}$.
By \eqref{eq:tildemujk2} we have, for every sufficiently large $n$,
\begin{align*}
W_2(\tilde\nu_0^{mn},\tilde\nu_1^{mn})^2
&\le \int_{X_n \times X_n} d_{X_n}^2 \,d\tilde\pi'
= \sum_{j,k} w_{jk} \int_{X_n \times X_n} d_{X_n}^2 \,d\tilde\pi_{jk}\\
&= \sum_{j,k} w_{jk} W_2(\tilde\mu_{jk}^{mn},\tilde\mu_{kj}^{mn})^2\\
&\le \sum_{j,k} w_{jk} (d_Y(y_j,y_k) + \theta(m^{-1}))^2\\
&\le \sum_{j,k} w_{jk} (d_Y(y_j,y_k)^2 + \theta(m^{-1})d_Y(y_j,y_k))
+ \theta(m^{-1})\\
&\le W_2(\nu_0,\nu_1)^2 + \theta(m^{-1}) W_2(\nu_0,\nu_1) + \theta(m^{-1}),
\end{align*}
where the last inequality follows from
$\diam B_j \le m^{-1}$, the definition of $w_{jk}$,
and the Schwartz inequality.
Let us give the opposite estimate.
We take an optimal transport plan $\tilde\pi$ for
$W_2(\tilde\nu_0^{mn},\tilde\nu_1^{mn})$.
Since $\tilde{\nu}_i^{mn}(X_n \setminus \tilde{X}_n) = 0$, we see that
\begin{align*}
W_2((p_n)_*\tilde\nu_0^{mn},(p_n)_*\tilde\nu_1^{mn})
&\le \int_{Y \times Y} d_Y^2 \,d(p_n\times p_n)_*\tilde\pi\\
&= \int_{X_n \times X_n} d_Y(p_n(x),p_n(x'))^2 \,d\tilde\pi(x,x')\\
&\le \int_{X_n \times X_n} (d_{X_n}(x,x') + \varepsilon_n)^2 \,d\tilde\pi(x,x')\\
&\le W_2(\tilde\nu_0^{mn},\tilde\nu_1^{mn})^2
+ 2\varepsilon_n W_2(\tilde\nu_0^{mn},\tilde\nu_1^{mn}) + \varepsilon_n^2.
\end{align*}
Here, $W_2(\tilde\nu_0^{mn},\tilde\nu_1^{mn})$
is uniformly bounded for all large $m$ and $n$.
Since $\lim_{m\to\infty}\lim_{n\to\infty} (p_n)_*\tilde\nu_i^{mn}
= \lim_{m\to\infty} \tilde\nu_i^m = \nu_i$,
we have
\[
\lim_{m\to\infty}\lim_{n\to\infty}
W_2((p_n)_*\tilde\nu_0^{mn},(p_n)_*\tilde\nu_1^{mn})
= W_2(\nu_0,\nu_1).
\]
This completes the proof of \eqref{eq:Wnu01}.
By \eqref{eq:tildemujk1} we have
\[
\tilde\nu_0^{mn}
= \sum_{j,k} w_{jk} \tilde\mu_{jk}^{mn}
\le (1+\theta(m^{-1})) \sum_j \nu_0(B_j)\mu_{\tilde{B}_j},
\]
which together with the monotonicity of $v(r) := U(r)/r \; (= \log r)$
implies that
\begin{align*}
&\Ent(\tilde\nu_0^{mn}|\mu_{X_n})
= \int_{X_n} v\left(\frac{d\tilde\nu_0^{mn}}{d\mu_{X_n}} \right)
d\tilde\nu_0^{mn}\\
&\le \int_{X_n} v\left((1+\theta(m^{-1})) \sum_j
\frac{\nu_0(B_j)}{\mu_{X_n}(\tilde{B}_j)}
1_{\tilde{B}_j} \right) d\tilde\nu_0^{mn}\\
&= \sum_j v\left((1+\theta(m^{-1}))
\frac{\nu_0(B_j)}{\mu_{X_n}(\tilde{B}_j)}
\right) \tilde\nu_0^{mn}(\tilde{B}_j)\\
&\le (1+\theta(m^{-1})) \sum_j v\left((1+\theta(m^{-1}))
\frac{\nu_0(B_j)}{\mu_{X_n}(\tilde{B}_j)}
\right) \nu_0(B_j)\\
&= (1+\theta(m^{-1})) \sum_j U\left(
(1+\theta(m^{-1})) \frac{\nu_0(B_j)}{\mu_{X_n}(\tilde{B}_j)}
\right) \mu_{X_n}(\tilde{B}_j),
\end{align*}
which converges to
\[
(1+\theta(m^{-1})) \sum_j U\left(
(1+\theta(m^{-1})) \frac{\nu_0(B_j)}{\mu_Y(B_j)}
\right) \mu_Y(B_j)
\]
as $n\to\infty$.
Since $\nu_0(B_j)/\mu_Y(B_j)$ is dominated by
the supremum of the density $\rho_0 := \frac{d\nu_0}{d\mu_Y}$ of $\nu_0$,
the above reduces to
\begin{align*}
&\sum_j U\left( \frac{\nu_0(B_j)}{\mu_Y(B_j)} \right) \mu_Y(B_j)
+ \theta(m^{-1}|\sup\rho_0)\\
&= \Ent(\bar\nu_0^m|\mu_Y) + \theta(m^{-1}|\sup\rho_0),
\end{align*}
where
\[
\bar\nu_i^m := \sum_j \nu_i(B_j) \mu_{B_j}.
\]
It follows from Jensen's inequality that
$\Ent(\bar\nu_0^m|\mu_Y) \le \Ent(\nu_0|\mu_Y)$.
In the same way, we obtain the estimate of $\Ent(\tilde\nu_1^{mn}|\mu_{X_n})$
and eventually, for $i=0,1$,
\begin{align} \label{eq:Ent-tildenui}
\limsup_{n\to\infty} \Ent(\tilde\nu_i^{mn}|\mu_{X_n})
\le \Ent(\nu_i|\mu_Y) + \theta(m^{-1}|\sup\rho_i).
\end{align}
The condition $\CD(K,\infty)$ implies that,
for any fixed $t \in (\,0,1\,)$,
there is a measure $\tilde\nu_t^{mn} \in P_2(X_n)$
such that
\begin{align}
\label{eq:CD-W2}
W_2(\tilde\nu_t^{mn},\tilde\nu_i^{mn})
&\le t^{1-i}(1-t)^i W_2(\tilde\nu_0^{mn},\tilde\nu_1^{mn}) + m^{-1},
\quad i=0,1,\\
\label{eq:CD-Ent}
\Ent(\tilde\nu_t^{mn}|\mu_{X_n})
&\le (1-t)\Ent(\tilde\nu_0^{mn}|\mu_{X_n})
+ t\Ent(\tilde\nu_1^{mn}|\mu_{X_n})\\
&\quad -\frac{1}{2} K t(1-t)W_2(\tilde\nu_0^{mn}\tilde\nu_1^{mn})^2
+ m^{-1}. \notag
\end{align}
Lemma \ref{lem:Ent-push} implies that
$\Ent((p_n)_*\tilde\nu_t^{mn}|(p_n)_*\mu_{X_n})
\le \Ent(\tilde\nu_t^{mn}|\mu_{X_n})$,
which together with
\eqref{eq:CD-Ent}, \eqref{eq:Ent-tildenui}, and \eqref{eq:Wnu01}
yields that
\begin{align}
\label{eq:pf-CD1}
&\limsup_{m\to\infty}\limsup_{n\to\infty}
\Ent((p_n)_*\tilde\nu_t^{mn}|(p_n)_*\mu_{X_n})\\
\notag &\le (1-t)\Ent(\nu_0|\mu_Y) + t\Ent(\nu_1|\mu_Y)
-\frac{1}{2} K t(1-t)W_2(\nu_0,\nu_1)^2.
\end{align}
We are going to estimate $W_2((p_n)_*\tilde\nu_t^{mn},\nu_i)$ for $i = 0,1$.
Taking an optimal transport plan $\pi$ for
$W_2(\tilde\nu_t^{mn},\tilde\nu_i^{mn})$, we see that
\begin{align*}
& W_2((p_n)_*\tilde\nu_t^{mn},(p_n)_*\tilde\nu_i^{mn})^2
\le \int_{Y \times Y} d_Y^2 \,d(p_n\times p_n)_*\pi\\
&= \int_{X_n \times X_n} d_Y(p_n(x),p_n(x'))^2 \,d\pi(x,x')
\intertext{and since $\tilde\nu_i^{mn}(X_n \setminus \tilde{X}_n) = 0$,}
&\le \int_{\tilde{X}_n \times \tilde{X}_n}
(d_{X_n}(x,x') + \varepsilon_n)^2 \,d\pi(x,x')\\
&\quad + \int_{(X_n \setminus \tilde{X}_n) \times \tilde{X}_n} d_Y(p_n(x),p_n(x'))^2 \;
d\pi(x,x')\\
&\le W_2(\tilde\nu_t^{mn},\tilde\nu_i^{mn})^2
+ 2\varepsilon_n W_2(\tilde\nu_t^{mn},\tilde\nu_i^{mn})
+ \varepsilon_n^2\\
&\quad + \int_{(X_n \setminus \tilde{X}_n) \times \tilde{X}_n} d_Y(p_n(x),p_n(x'))^2 \;
d\pi(x,x').
\end{align*}
Since $\{p_n\}$ has bounded values on exceptional domains
and since
\[
\tilde\nu_i^{mn}(X_n \setminus p_n^{-1}(\supp\nu_0 \cup \supp\nu_1)) = 0,
\]
there is a constant $D > 0$ such that
\[
d_Y(p_n(x),p_n(x'))^2 \le D
\]
for $\pi$-a.e. $(x,x') \in (X_n \setminus \tilde{X}_n) \times \tilde{X}_n$
and hence
\begin{align*}
&\int_{(X_n \setminus \tilde{X}_n) \times \tilde{X}_n} d_Y(p_n(x),p_n(x'))^2 \; d\pi(x,x')\\
&\le D \, \pi((X_n \setminus \tilde{X}_n) \times X_n) = D\,
\tilde\nu_t^{mn}(X_n \setminus \tilde{X}_n).
\end{align*}
Therefore,
\begin{align}
\label{eq:pn-tildenut}
&\limsup_{n\to\infty} W_2((p_n)_*\tilde\nu_t^{mn},(p_n)_*\tilde\nu_i^{mn})^2\\
\notag &\le \limsup_{n\to\infty} (W_2(\tilde\nu_t^{mn},\tilde\nu_i^{mn})^2
+ D \,\tilde\nu_t^{mn}(X_n \setminus \tilde{X}_n)).
\end{align}
Let us prove that
\begin{align}
\label{eq:nonexc0}
\lim_{n\to\infty} \tilde\nu_t^{mn}(X_n \setminus \tilde{X}_n) = 0.
\end{align}
\eqref{eq:CD-Ent} and \eqref{eq:Ent-tildenui} together imply
$\Ent(\tilde\nu_t^{mn}|\mu_{X_n}) \le C$, where $C$ is a constant
independent of large $m$ and $n$.
Put
$\tilde\rho_t := \frac{d\tilde\nu_t^{mn}}{d\mu_{X_n}}$.
Since $U(r)/r$ is monotone increasing in $r$, we have, for any $r > 0$,
\begin{align*}
\tilde\nu_t^{mn}(X_n \setminus \tilde{X}_n)
&= \int_{\{\tilde\rho_t \ge r\} \setminus \tilde{X}_n} \tilde\rho_t \,d\mu_{X_n}
+ \int_{\{\tilde\rho_t < r\} \setminus \tilde{X}_n} \tilde\rho_t \,d\mu_{X_n}\\
&\le \frac{r}{U(r)} \int_{\{\tilde\rho_t \ge r\} \setminus \tilde{X}_n}
U(\tilde\rho_t) \,d\mu_{X_n}
+ r \mu_{X_n}(X \setminus \tilde{X}_n).
\end{align*}
By $\int_{\{U(\tilde\rho_t) < 0\}} U(\tilde\rho_t)\;d\mu_{X_n} \ge \inf U$,
we have
\[
\int_{\{U(\tilde\rho_t) > 0\}} U(\tilde\rho_t)\;d\mu_{X_n} \le C - \inf U
\]
and thus
\[
\tilde\nu_t^{mn}(X_n \setminus \tilde{X}_n)
\le \frac{(C - \inf U)r}{U(r)} + r \mu_{X_n}(X_n \setminus \tilde X_n)
\]
for any $r > 0$ with $U(r) > 0$.
By remarking $\lim_{r\to +\infty} r/U(r) \to 0$,
the above inequality implies \eqref{eq:nonexc0}.
Combining \eqref{eq:nonexc0} with \eqref{eq:pn-tildenut} yields that
\[
\limsup_{n\to\infty} W_2((p_n)_*\tilde\nu_t^{mn},(p_n)_*\tilde\nu_i^{mn})
\le \limsup_{n\to\infty} W_2(\tilde\nu_t^{mn},\tilde\nu_i^{mn}).
\]
Since $(p_n)_*\tilde\nu_i^{mn} \overset{n\to\infty}\to
\tilde\nu_i^m \overset{m\to\infty}\to \nu_i$,
we have
\begin{align*}
&\limsup_{m\to\infty}\limsup_{n\to\infty} W_2((p_n)_*\tilde\nu_t^{mn},\nu_i)\\
&\le \limsup_{m\to\infty}\limsup_{n\to\infty}
W_2(\tilde\nu_t^{mn},\tilde\nu_i^{mn}),
\intertext{and by \eqref{eq:CD-W2} and \eqref{eq:Wnu01},}
&\le \limsup_{m\to\infty} \limsup_{n\to\infty}
t^{1-i}(1-t)^i W_2(\tilde\nu_0^{mn},\tilde\nu_1^{mn})\\
&= t^{1-i}(1-t)^i W_2(\nu_0,\nu_1).
\end{align*}
By this and \eqref{eq:pf-CD1}, there is a sequence $m(n) \to \infty$
as $n\to\infty$ such that $\tilde\nu_t^n := \tilde\nu_t^{m(n)n}$
satisfies \eqref{eq:lem-CD1} and \eqref{eq:lem-CD2}.
This completes the proof.
\end{proof}
\begin{lem} \label{lem:tight}
Let $Y$ be a proper metric space and $y_0 \in Y$ a point.
If a sequence of measures $\nu_n \in P_p(Y)$, $n=1,2,\dots$,
has uniformly bounded $p^{th}$ moment $W_p(\nu_n,\delta_{y_0})$
for some real number $p$ with $1 \le p < +\infty$, then $\{\nu_n\}$ is tight.
\end{lem}
\begin{proof}
Assume that a sequence of measures $\nu_n \in P_p(Y)$, $n=1,2,\dots$,
satisfies $W_p(\nu_n,\delta_{y_0}) \le C$ for any $n$, for some constant $C$,
and for some $p$ with $1 \le p < +\infty$.
By H\"older's inequality we have, for any $R > 0$,
\begin{align*}
C &\ge W_p(\nu_n,\delta_{y_0}) \ge W_1(\nu_n,\delta_{y_0})\\
&\ge \int_{Y \setminus B_R(y_0)} d_Y(y,y_0) \; d\nu_n(y)
\ge R \,\nu_n(Y \setminus B_R(y_0)).
\end{align*}
For any given $\varepsilon > 0$, setting $R := C/\varepsilon$
and $K_\varepsilon := B_R(y_0)$ yields that
$\nu_n(Y \setminus K_\varepsilon) \le \varepsilon$.
This completes the proof.
\end{proof}
The following is one of main theorems of this chapter.
\begin{thm} \label{thm:CD-conc}
Let $\{X_n\}_{n=1}^\infty$ be a sequence of mm-spaces satisfying $\CD(K,\infty)$
for a real number $K$.
If $X_n$ concentrates to an mm-space $Y$ as $n\to\infty$,
then we have the following {\rm(1)}, {\rm(2)}, and {\rm(3)}.
\begin{enumerate}
\item $Y$ is an intrinsic metric space.
\item $Y$ is a geodesic space satisfying $\CD(K,\infty)$
provided that $Y$ is proper.
\item $Y$ satisfies $\CD(K,\infty)$
provided that $(p_n)_*\mu_{X_n} = \mu_Y$ for every $n=1,2,\dots$,
where $\{p_n : X_n \to Y\}_{n=1}^\infty$ is a sequence of Borel measurable
maps enforcing concentration of $X_n$ to $Y$
and with bounded values on exceptional domains.
\end{enumerate}
\end{thm}
\begin{proof}
(1) follows from a standard discussion (cf.~\cite{Sturm:geoI}*{Remark 4.6})
using Lemma \ref{lem:CD-conc}.
(3) is derived from Lemmas \ref{lem:CD-conc} and \ref{lem:CD-Pcb}.
We prove (2).
Since $Y$ is proper, (1) implies that $Y$ is a geodesic space.
Take any $\nu_0,\nu_1 \in P^{cb}(Y)$ and fix them.
Fixing any $t \in (\,0,1\,)$,
we have measures $\nu_t^n := (p_n)_*\tilde\nu_t^n \in P_2(Y)$,
$n=1,2,\dots$, as in Lemma \ref{lem:CD-conc}.
Since $W_2(\nu_t^n,\nu_i)$ is uniformly bounded for all $n$,
Lemma \ref{lem:tight} together with Prohorov's theorem
(Theorem \ref{thm:Proh})
proves that $\{\nu_t^n\}$
has a subsequence converging weakly to a measure on $Y$,
say $\nu_t$. By \eqref{eq:lem-CD1}, we have
\[
W_2(\nu_t,\nu_i) = t^{1-i}(1-t)^i W_2(\nu_0,\nu_1)
\]
Let $C_0$ be the union of the images of all minimal geodesic segments
joining $y$ and $y'$, where $y$ and $y'$ run over all
points in $\supp\nu_0$ and $\supp\nu_1$ respectively.
Since $\supp\nu_i$, $i=0,1$, are compact and $Y$ is proper,
$C_0$ is a compact subset of $Y$.
Note that $\nu_t$ is supported in $C_0$.
We set $C_r := B_r(C_0)$ for $r > 0$, which is also compact.
Let $\rho_t^n := \frac{d\nu_t^n}{d\mu_Y}$.
Since
\begin{align*}
\int_{Y \setminus C_r} U(\rho_t^n) \; d(p_n)_*\mu_{X_n}
\ge (p_n)_*\mu_{X_n}(Y \setminus C_r) \inf U,
\end{align*}
we have
\[
\liminf_{n\to\infty} \int_{Y \setminus C_r} U(\rho_t^n) \; d(p_n)_*\mu_{X_n}
\ge \mu_Y(\overline{Y \setminus C_r}) \inf U,
\]
where we note that $\inf U < 0$.
By Lemma \ref{lem:Ent-lim},
\begin{align*}
\liminf_{n\to\infty} \int_{C_r} U(\rho_t^n) \; d(p_n)_*\mu_{X_n}
\ge \int_{C_r} U\Bigl(\frac{d\nu_t}{d\mu_Y}\Bigr) \; d\mu_Y
= \Ent(\nu_t|\mu_Y).
\end{align*}
We thus have
\begin{align*}
\liminf_{n\to\infty} \Ent(\nu_t^n|(p_n)_*\mu_{X_n})
&= \liminf_{n\to\infty} \int_Y U(\rho_t^n) \; d(p_n)_*\mu_{X_n}\\
&\ge \Ent(\nu_t|\mu_Y) + \mu_Y(\overline{Y \setminus C_r}) \inf U,
\end{align*}
where $\mu_Y(\overline{Y \setminus C_r}) \to 0$ as $r \to +\infty$.
This together with \eqref{eq:lem-CD2} implies
\[
\Ent(\nu_t|\mu_Y) \le (1-t)\Ent(\nu_0|\mu_Y) + t\Ent(\nu_1|\mu_Y)
-\frac{1}{2} K t(1-t)W_2(\nu_0,\nu_1)^2.
\]
The proof of the theorem is now completed.
\end{proof}
\section{$k$-L\'evy family}
\label{sec:k-Levy}
\begin{defn}[$k$-L\'evy family]
\index{k-Levy family@$k$-L\'evy family}
Let $k$ be a natural number.
A sequence $\{X_n\}_{n=1}^{\infty}$ of mm-spaces is called
a \emph{$k$-L\'evy family}
if
\[
\lim_{n\to\infty}\Sep(X_n;\kappa_0,\cdots,\kappa_k)=0
\]
for any $\kappa_0,\kappa_1,\cdots,\kappa_k >0$.
\end{defn}
It follows from Proposition \ref{prop:ObsDiam-Sep} that
a sequence of mm-spaces is a $1$-L\'evy family
if and only if it is a L\'evy family.
For $k \le k'$, any $k$-L\'evy family is a $k'$-L\'evy family.
A $k$-L\'evy family is a union of
$k$ number of L\'evy families.
The following is a direct consequence of Proposition \ref{prop:lamk-Sep}.
\begin{prop}\label{prop:k-Levy-lam}
Let $X_n$, $n=1,2,\dots$, be closed Riemannian manifolds.
If $\lambda_k(X_n)$ diverges to infinity as $n\to \infty$
for a natural number $k$,
then $\{X_n\}$ is a $k$-L\'evy family.
\end{prop}
\begin{thm} \label{thm:k-Levy-conc}
Let $\{X_n\}_{n=1}^{\infty}$ be a $k$-L\'evy family of mm-spaces
for a natural number $k$.
Then we have one of the following {\rm(1)} and {\rm(2)}.
\begin{enumerate}
\item $\{X_n\}$ is a L\'evy family.
\item There exist a subsequence $\{X_{n_i}\}_{i=1}^\infty$ of
$\{X_n\}_{n=1}^{\infty}$ and a sequence of numbers $t_i$ with $0 < t_i \le 1$
such that, as $i\to\infty$, $t_i X_{n_i}$
concentrates to a finite mm-space $Y$ with $2 \le \# Y \le k$.
\end{enumerate}
\end{thm}
Recall that $tX$ indicates the scaled mm-space $(X,td_X,\mu_X)$
for $t > 0$.
\begin{proof}
Assume that $\{X_n\}_{n=1}^{\infty}$ is a $k$-L\'evy family and
is not a L\'evy family.
We may assume that $k$ is the minimal number such that
$\{X_n\}$ is a $k$-L\'evy family.
We have $k \ge 2$.
Taking a subsequence of $\{X_n\}$, we have
\begin{align}
\Sep(X_n;\kappa_0,\cdots, \kappa_{k-1}) > c
\end{align}
for every $n$ and for some numbers
$c,\kappa_0,\kappa_1,\cdots,\kappa_{k-1} > 0$.
There are Borel subsets $A_{0n},\cdots , A_{k-1,n}\subset X_n$
such that $\mu_{X_n}(A_{in})\geq \kappa_i$ and
$d_{X_n}(A_{in},A_{jn}) > c$ for any different $i$ and $j$.
For simplicity we set
\begin{align*}
S_n(\kappa) &:= \Sep(X_n;\kappa_0,\cdots, \kappa_{k-1},\kappa)
\end{align*}
for $\kappa > 0$.
Since $\{X_n\}$ is a $k$-L\'evy family,
there is a sequence of positive numbers $\tilde\kappa_n \to 0$
such that $S_n(\tilde\kappa_n)$ converges to zero as $n\to\infty$.
Let
\begin{align*}
\quad r_n := \max\{S_n(\tilde{\kappa}_n),1/n\}
\quad\text{and}\quad B_{in}:=B_{2r_n}(A_{in}).
\end{align*}
Note that $S_n(\tilde{\kappa}_n)$ may be zero, but $r_n$ is positive
and still converges to zero as $n\to\infty$.
Since the distance between $X_n \setminus \bigcup_{i=0}^{k-1}B_{in}$
and $A_{jn}$, $j=0,\dots,k-1$, is strictly greater than
$S_n(\tilde{\kappa}_n)$,
we have
\begin{align} \label{eq:k-Levy-conc-1}
\mu_{X_n}\Big(X_n \setminus \bigcup_{i=0}^{k-1}B_{in}\Big) <
\tilde{\kappa}_n
\end{align}
for every $n$.
\begin{clm} \label{clm:Bin-Levy}
For each $i=0,1,\dots,k-1$, the sequence
$\{(B_{in},d_{X_n},\mu_{in})\}_{n=1}^{\infty}$ is a
L\'evy family, where $\mu_{in} := \mu_{X_n}|_{B_{in}}$.
\end{clm}
\begin{proof}
Take any numbers $\kappa,\kappa' > 0$ and fix them.
Since $\{X_n\}$ is a $k$-L\'evy family,
\begin{align*}
\alpha_n := \Sep(X_n;\kappa_0,\kappa_1, \cdots, \kappa_{i-1},
\kappa,\kappa', \kappa_{i+1}, \cdots,
\kappa_{k-1})
\end{align*}
converges to zero as $n\to\infty$, so that
\[
\alpha_n < c-4r_n \le \min_{j\neq j'}d_{X_n}(B_{jn},B_{j'n})
\]
for every sufficiently large $n$.
For the claim, it suffices to prove that
$\Sep(B_{in};\kappa,\kappa') \leq \alpha_n$.
In fact, if $\Sep(B_{in};\kappa,\kappa') > \alpha_n$,
then we have two Borel subsets $B_{in}', B_{in}'' \subset B_{in}$
such that $\mu_{X_n}(B_{in}') \ge \kappa$, $\mu_{X_n}(B_{in}'')
\ge \kappa'$, and $d_{X_n}(B_{in}',B_{in}'') > \alpha_n$.
By remarking that
$\mu_{X_n}(B_{jn})\geq \kappa_j$ for every $j$,
this is a contradiction to the definition of $\alpha_n$.
\end{proof}
We set, for $i,j = 0,1,\dots,k-1$,
\begin{align*}
d_{ijn} &:=
\sup\{\; |\lm (f;\mu_{in})- \lm (f;\mu_{jn})| ; f \in \mathcal{L}{\it ip}_1(X_n)\;\},\\
D_n &:= \max_{i,j=0,1,\dots,k-1} d_{ijn}.
\end{align*}
Since $d_{X_n}(B_{in},B_{jn})$ is bounded away from zero for $i \neq j$,
we have $\liminf_{n\to\infty} D_n > 0$.
Define
\[
t_n :=
\begin{cases}
1/D_n & \text{if $D_n > 1$},\\
1 & \text{if $D_n \le 1$}.
\end{cases}
\]
It is clear that $0 < t_n \le 1$ and $0 \le t_nd_{ijn} \le 1$.
Taking a subsequence of $\{n\}$, we see that, as $n\to\infty$,
$t_n d_{ijn}$ converges to a number, say $d_{ij} \in [\,0,1\,]$,
and $\mu_{X_n}(B_{in})$ to a number, say $w_i \in (\,0,1\,)$, for any $i,j$.
We have $d_{ii} = 0$, $d_{ij} \ge 0$, $d_{ji} = d_{ij}$ for any $i,j$,
and $\max_{i,j} d_{ij} > 0$.
We prove that $d_{ij} \le d_{il} + d_{lj}$ for any $i, j, l$.
In fact, for any $\varepsilon > 0$
there is a function $f \in \mathcal{L}{\it ip}_1(X_n)$ such that
\[
|\; |\lm(f;\mu_{in}) - \lm(f;\mu_{jn})| - d_{ijn} \;|
< \varepsilon.
\]
We have
\begin{align*}
d_{iln} &\ge |\lm(f;\mu_{in}) - \lm(f;\mu_{ln})|,\\
d_{ljn} &\ge |\lm(f;\mu_{ln}) - \lm(f;\mu_{jn})|,
\end{align*}
and therefore
$d_{ijn} -\varepsilon < d_{iln} + d_{ljn}$,
which implies $d_{ij} \le d_{il} + d_{lj}$.
Let $Y = \{y_0,y_1,\dots,y_{k-1}\}$ be a set consisting of
$k$ elements and define
$d_Y(y_i,y_j) := d_{ij}$.
Then, $(Y,d_Y)$ is a pseudometric space.
We define the measure $\mu_Y := \sum_{i=0}^{k-1} w_i \delta_{y_i}$ on $Y$.
It follows from \eqref{eq:k-Levy-conc-1} that $\mu_Y$ is a
probability measure.
Let $Y'$ be the quotient space of $Y$ by the equivalence relation
$d_Y = 0$, and let $[y_i] \in Y'$ denote
the equivalence class represented by $y_i \in Y$.
The pseudometric $d_Y$ induces a metric on $Y'$, say $d_{Y'}$.
Let $\mu_{Y'}$ be the push-forward of $\mu_Y$
by the projection $Y \to Y'$.
It then follows from $\max_{i,j}d_{ij} > 0$ that
$(Y',d_{Y'},\mu_{Y'})$ is an mm-space with $2 \le \# Y' \le k$.
Let us prove that $X_n$ concentrates $Y'$ as $n\to\infty$.
Define a map $p_n:X_n \to Y'$ by
\[
p_n(x) :=
\begin{cases}
[y_i] & \text{if $x \in B_{in}$},\\
[y_0] & \text{if $x \in X_n \setminus \bigcup_{i=0}^{k-1}B_{in}$}.
\end{cases}
\]
It is obvious that each $p_n$ is a Borel measurable map, and
that $(p_n)_*\mu_{X_n}$ converges weakly to $\mu_Y$ as $n\to\infty$.
It suffices to prove that
$\{p_n\}$ satisfies (1), (2), and (3) of Theorem \ref{thm:pn}.
We prove (1), i.e., $p_n$ is $1$-Lipschitz up to an additive error
tending to zero as $n\to\infty$.
There are $f_{ijn} \in \mathcal{L}{\it ip}_1(X_n)$ such that
\[
d_{ij} = \lim_{n\to\infty}
t_n |\lm(f_{ijn};\mu_{in})-\lm(f_{ijn};\mu_{jn})|.
\]
We find a sequence $\varepsilon_n \to 0$ such that for any $i$ and $j$,
\begin{align*}
|\;d_{ij} - t_n|\lm(f_{ijn};\mu_{in})-\lm(f_{ijn};\mu_{jn})|\;|
\le \varepsilon_n.
\end{align*}
It follows from Claim \ref{clm:Bin-Levy} and Lemma \ref{lem:LeRad-ObsDiam}
that there is a sequence $\varepsilon_n' \to 0$
in such a way that, for each $i$, the $\mu_{X_n}$-measure of
\begin{align*}
B_{in}' := \{\; x \in B_{in} \; ; \;
& |f_{ijn}(x)-\lm(f_{ijn};\mu_{in})| < \varepsilon_n' \\
&\ \text{for any $j$ different from $i$}\;\}
\end{align*}
converges to $w_i$ as $n\to\infty$.
For any $x \in B_{in}'$ and $x' \in B_{jn}'$ we have
\begin{align*}
d_{Y'}(p_n(x),p_n(x')) &= d_{Y'}([y_i],[y_j]) = d_{ij}\\
&\le t_n|\lm(f_{ijn};\mu_{in})-\lm(f_{ijn};\mu_{jn})| + \varepsilon_n\\
&\le t_n|f_{ijn}(x)-f_{ijn}(x')| + 2t_n\varepsilon_n' + \varepsilon_n\\
&\le t_n d_{X_n}(x,x') + 2t_n\varepsilon_n' + \varepsilon_n
\end{align*}
which implies (1).
We prove (2), i.e., $\{p_n\}$ effectuates relative concentration of $X_n$
over $Y'$. Let $f_n \in \mathcal{L}{\it ip}_1(X_n)$ and $B \subset Y'$.
Since $\ObsDiam(tX;-\kappa) = t\ObsDiam(X;-\kappa)$,
it suffices to prove that
\[
\limsup_{n\to\infty}
t_n \diam((f_n)_*(\mu_{X_n}|_{p_n^{-1}(B)});\mu_{X_n}(p_n^{-1}(X_n))-\kappa)
\le \diam B.
\]
Let $\tilde B$ be the preimage of $B$ by the projection $Y \to Y'$
and set $\tilde B = \{y_{i_1},\dots,y_{i_m}\}$.
We see that
\[
p_n^{-1}(B) =
\begin{cases}
\bigcup_{l=1}^m B_{i_ln} & \text{if $y_0 \notin \tilde{B}$},\\
\bigcup_{l=1}^m B_{i_ln} \cup (X_n \setminus \bigcup_{i=0}^k B_{in})
& \text{if $y_0 \in \tilde{B}$}.
\end{cases}
\]
Put $M_{ln} := \lm(f_n|B_{i_ln};\mu_{X_n}|_{B_{i_ln}})$.
Since
\[
\LeRad(\mu_{X_n}|_{B_{i_l}};-\kappa)
\le \ObsDiam(\mu_{X_n}|_{B_{i_l}};-\kappa) \to 0
\ \text{as $n\to\infty$},
\]
we have, for any $\varepsilon > 0$,
\begin{align*}
&(f_n)_*(\mu_{X_n}|_{B_{i_ln}})([\,M_{ln}-\varepsilon,M_{ln}+\varepsilon\,])\\
&= \mu_{X_n}(\{\; x \in B_{i_ln} \mid |f_n(x) - M_{ln}| \le \varepsilon\;\})
\to w_{i_l} \ \text{as $n\to\infty$}.
\end{align*}
Moreover, the total measure of $(f_n)_*(\mu_{X_n}|_{p_n^{-1}(B)})$
converges to $\sum_{l=1}^m w_{i_l}$.
Therefore,
\begin{align*}
&\limsup_{n\to\infty}
t_n \diam((f_n)_*(\mu_{X_n}|_{p_n^{-1}(B)});\mu_{X_n}(p_n^{-1}(X_n))-\kappa)\\
&\le \limsup_{n\to\infty} \max_{1 \le l,l' \le m} t_n |M_{ln}-M_{l'n}|
\le \limsup_{n\to\infty} \max_{1 \le l,l' \le m} t_n d_{i_l i_{l'} n}\\
&= \max_{1 \le l,l' \le m} d_{i_l i_{l'}} = \diam B,
\end{align*}
which completes the proof of (2).
For the proof of (3), we prove the following.
\begin{clm} \label{clm:d-plus}
For any $i,j=0,1,\dots,k-1$ and $\kappa > 0$ we have
\begin{align*}
\limsup_{n\to\infty} t_n d_+(B_{in},B_{jn};+\kappa) \le d_{Y'}([y_i],[y_j]).
\end{align*}
\end{clm}
\begin{proof}
We fix $i$, $j$, and $\kappa > 0$.
There are subsets $C_n \subset B_{in}$ and
$C_n' \subset B_{jn}$ such that
$\mu_{X_n}(C_n)\geq \kappa$, $\mu_{X_n}(C_n')\geq \kappa$,
and
\[
d_+ (B_{in},B_{jn};+\kappa) < d_{X_n}(C_n,C_n') + 1/n.
\]
Let $f_n : X_n \to \field{R}$ be the distance function from $C_n$,
i.e., $f_n(x) := d_{X_n}(x,C_n)$, $x \in X_n$.
By Claim \ref{clm:Bin-Levy},
there is a sequence $\varepsilon_n' \to 0$
such that, for each $l = 0,1,\dots,k-1$, the $\mu_{X_n}$-measure of
\begin{align*}
D_{ln} := \{\; x\in B_{ln} \mid
|f_n(x)-\lm (f_n; \mu_{ln})| \le \varepsilon_n' \; \}
\end{align*}
converges to $w_l$ as $n\to\infty$.
There is a natural number $N$ such that
the intersections $C_n \cap D_{in}$ and $C_n' \cap D_{jn}$ are both
nonempty for any $n \ge N$.
In what follows, let $n$ be any number with $n \ge N$.
Taking points $x_n \in C_n \cap D_{in}$ and $x_n' \in C_n' \cap D_{jn}$,
we have
\begin{align*}
|\lm (f_n;\mu_{in})|
&= |f_n(x_n)- \lm(f_n;\mu_{in}) | \le \varepsilon_n',\\
\lm(f_n;\mu_{jn}) &\ge f_n(x_n') - \varepsilon_n'
\geq d_{X_n}(C_n,C_n') - \varepsilon_n'.
\end{align*}
We therefore obtain
\begin{align*}
t_n d_+(B_{in},B_{jn};+\kappa)
&\le t_n d_{X_n}(C_n,C_n') + t_n/n\\
&\le t_n(\lm(f_n;\mu_{jn}) - \lm(f_n;\mu_{in}))
+ 2t_n\varepsilon_n' + t_n/n\\
&\le t_n d_{ijn} + \varepsilon_n + 2t_n\varepsilon_n' + t_n/n\\
&\to d_{ij} \quad\text{as $n\to\infty$},
\end{align*}
which implies the claim.
\end{proof}
Using Claim \ref{clm:d-plus} we now prove (3).
We take $A_1, A_2 \subset Y'$ and $\kappa > 0$.
Let $\{y_{i_1},\dots,y_{i_M}\}$ be the preimage of $A_1$ by
the projection $Y \to Y'$ and $\{y_{j_1},\dots,y_{j_N}\}$ the preimage
of $A_2$. We see that $A_1 = \{[y_{i_1}],\dots,[y_{i_M}]\}$
and $A_2 = \{[y_{j_1}],\dots,[y_{j_N}]\}$.
Assume that $n$ is large enough.
Since $\mu_{X_n}(X_n \setminus \bigcup_{i=0}^{k-1} B_{in}) < \kappa/2$,
it holds that
\begin{align*}
& d_+(p_n^{-1}(A_1),p_n^{-1}(A_2);+\kappa)\\
&\le d_+(B_{i_1 n} \cup \dots \cup B_{i_M n},
B_{j_1 n} \cup \dots \cup B_{j_N n};+\kappa/2)\\
&\le \max_{\alpha=1,\dots,M,\ \beta=1,\dots,N}
d_+(B_{i_\alpha n},B_{j_\beta n};+\kappa/(2\max\{M,N\}))
\end{align*}
and by Claim \ref{clm:d-plus},
\begin{align*}
&\limsup_{n\to\infty} t_n d_+(p_n^{-1}(A_1),p_n^{-1}(A_2);+\kappa)\\
&\le \max_{\alpha=1,\dots,M,\ \beta=1,\dots,N} d_{Y'}([y_{i_\alpha}],[y_{j_\beta}])\\
&\le d_{Y'}(A_1,A_2) + \diam A_1 + \diam A_2.
\end{align*}
(3) has been proved.
By Theorem \ref{thm:pn} and Corollary \ref{cor:enforce},
this completes the proof of Theorem \ref{thm:k-Levy-conc}.
\end{proof}
\begin{cor} \label{cor:k-Levy-conc-2}
Let $\{X_n\}_{n=1}^{\infty}$ be a $k$-L\'evy family of mm-spaces
for a natural number $k$ such that
\[
\limsup_{n\to\infty} \ObsDiam(X_n;-\kappa) < +\infty
\]
for any $\kappa > 0$.
Then, there exists a subsequence of $\{X_n\}$
that concentrates to a finite mm-space $Y$ with $\# Y \le k$.
\end{cor}
\begin{proof}
In the proof of the previous Theorem \ref{thm:k-Levy-conc},
the assumption for the observable diameter implies that
$d_{ijn}$ is bounded above uniformly for all $i$, $j$, and $n$,
so that $t_n$ is bounded away from zero.
Therefore, Theorem \ref{thm:k-Levy-conc} implies the corollary.
\end{proof}
The following is a direct consequence of Corollary \ref{cor:k-Levy-conc-2}
and Theorem \ref{thm:CD-conc}.
\begin{cor} \label{cor:lamk-conc-2}
Let $\{X_n\}_{n=1}^\infty$ be a sequence of closed
Riemannian manifolds with a lower bound of Ricci curvature
and with the property that
\[
\limsup_{n\to\infty} \ObsDiam(X_n;-\kappa) < +\infty
\]
for any $\kappa > 0$.
If $\lambda_k(X_n)$ diverges to infinity as $n \to \infty$
for some natural number $k$, then $\{X_n\}$ is a L\'evy family.
\end{cor}
\begin{cor} \label{cor:lamk-conc}
Let $\{X_n\}_{n=1}^\infty$ be a sequence of closed
Riemannian manifolds of nonnegative Ricci curvature.
If $\lambda_k(X_n)$ diverges to infinity as $n \to \infty$
for some natural number $k$, then $\{X_n\}$ is a L\'evy family.
\end{cor}
\begin{proof}
Let $\{X_n\}_{n=1}^\infty$ be a sequence of closed
Riemannian manifolds of nonnegative Ricci curvature
such that $\lambda_k(X_n)$ diverges to infinity as $n \to \infty$
for some natural number $k$.
By Proposition \ref{prop:k-Levy-lam}, $\{X_n\}$ is a $k$-L\'evy family.
Suppose that $\{X_n\}$ is not a L\'evy family.
By applying Theorem \ref{thm:k-Levy-conc}, there are
a subsequence $\{X_{n_i}\}$ of $\{X_n\}$ and a sequence of
numbers $t_i$ with $0 < t_i \le 1$ such that
$t_iX_{n_i}$ concentrates to a disconnected mm-space $Y$.
Since each $t_iX_{n_i}$ satisfies $\CD(0,\infty)$,
so does $Y$, which is a contradiction.
This completes the proof.
\end{proof}
It follows from Corollary \ref{cor:ObsDiam-spec} that
if $\lambda_1(X_n)$ diverges to infinity as $n\to\infty$
for a sequence of closed Riemannian manifolds $X_n$, $n=1,2,\dots$,
then it is a L\'evy family.
We have the converse under the nonnegativity of Ricci curvature.
\begin{thm}[E.~Milman; \cites{Emil:isop,Emil:role}]
\label{thm:Levy-lam-Ric}
If $\{X_n\}_{n=1}^{\infty}$ is a L\'evy family
of closed Riemannian manifolds of
nonnegative Ricci curvature,
then $\lambda_1(X_n)$ diverges to infinity as $n \to \infty$.
\end{thm}
Combining Corollary \ref{cor:lamk-conc}
and Theorem \ref{thm:Levy-lam-Ric} yields the following equivalence
\begin{align*}
&\text{$\{X_n\}$ is a L\'evy family}\\
\Longleftrightarrow\ &\lambda_1(X_n) \to +\infty\\
\Longleftrightarrow\ &\lambda_k(X_n) \to +\infty
\ \text{for some $k$}
\end{align*}
for a sequence of closed Riemannian manifolds $X_n$, $n=1,2,\dots$,
with nonnegative Ricci curvature.
Using Theorem \ref{thm:Levy-lam-Ric}, we prove
the following theorem.
\begin{thm} \label{thm:lamk1}
For any natural number $k$, there exists a positive constant
$C_k$ depending only on $k$
such that if $X$ is a closed Riemannian manifold
of nonnegative Ricci curvature,
then we have
\[
\lambda_k(X) \le C_k \lambda_1(X).
\]
\end{thm}
\begin{proof}
Suppose that Theorem \ref{thm:lamk1} is false.
Then, there are a natural number $k$ and
a sequence of closed Riemannian manifolds
$X_n$, $n=1,2,\dots$, of nonnegative Ricci curvature
such that $\lambda_k(X_n)/\lambda_1(X_n)$ diverges to infinity
as $n\to\infty$. Let $X_n'$ be the scale change of $X_n$ such that
$\lambda_1(X_n') = 1$.
Since
\[
\lambda_k(X_n')
= \frac{\lambda_k(X_n')}{\lambda_1(X_n')}
= \frac{\lambda_k(X_n)}{\lambda_1(X_n)}
\to +\infty \quad\text{as $n\to\infty$},
\]
and by Corollary \ref{cor:lamk-conc}, the sequence $\{X_n'\}$
is a L\'evy family.
By Theorem \ref{thm:Levy-lam-Ric}, $\lambda_1(X_n')$ must be divergent
to infinity, which is a contradiction.
This completes the proof.
\end{proof}
\begin{ex} \label{ex:k-spheres}
For any natural number $k \ge 2$, we give an example
of a sequence of closed Riemannian manifolds $X_n$, $n=1,2,\dots$,
such that, as $n\to\infty$,
\begin{enumerate}
\item $\lambda_{k-1}(X_n)$ converges to zero,
\item $\lambda_k(X_n)$ diverges to infinity,
\item $X_n$ concentrates to a finite mm-space $Y$ with $\# Y = k$
and, in particular, $\{X_n\}$ is not a L\'evy family.
\end{enumerate}
Such a sequence $\{X_n\}$ is constructed as follows.
Let $S_1^n,S_2^n,\dots,S_k^n$ be the $k$ copies of
an $n$-dimensional unit sphere in a Euclidean space,
and $p_i^n,q_i^n \in S_i^n$ two points that are antipodal to each other
for each $i$,
i.e., the geodesic distance between $p_i^n$ and $q_i^n$
is equal to $\pi$.
Let us consider the connected sum of $S_1^n,\dots,S_k^n$ with
small bridges.
Let $\delta_n$ be a small positive number.
For each $i=1,2,\dots,k-1$, we remove open geodesic metric balls
$U_{\delta_n}(q_i^n)$ and $U_{\delta_n}(p_{i+1}^n)$ of radius $\delta_n$
centered at $q_i^n$ and $p_{i+1}^n$ from $S_i^n$ and $S_{i+1}^n$,
and attach a copy of the Riemannian product space
$S^{n-1}(\sin\delta_n) \times [\,0,\delta_n\,]$ to the boundaries,
where $S^{n-1}(r)$ is an $(n-1)$-dimensional sphere of radius $r$
in a Euclidean space.
The boundary of $B_{\delta_n}(x)$, $x \in S_i^n$, is
isometric to $S^{n-1}(\sin\delta_n)$ and the attaching is by
an isometry.
The resulting manifold, say $X_n$, is homeomorphic to a sphere
and has a $C^0$ Riemannian metric.
We deform the metric of $X_n$ to a smooth one, so that
$X_n$ is a $C^\infty$ Riemannian manifold.
Let $\hat X_n$ be the disjoint union of $S_1^n,\dots,S_k^n$.
Take a sequence of positive numbers $C_n \to +\infty$.
We may assume that $\delta_n$ is so small that
the spectrum of the Laplacian of $X_n$ in $[\,0,C_n\,]$ is
very close to the spectrum of the Laplacian of $\hat X_n$ in $[\,0,C_n\,]$.
A precise proof of this follows from the same discussion as in
\cite{Fuk}*{\S 9}.
Since $\lambda_{k-1}(\hat X_n) = 0$ and $\lambda_k(\hat X_n) = n$,
the sequence $\{X_n\}$ satisfies (1) and (2).
It is easy to see that $X_n$ concentrates to a finite mm-space
$Y = \{y_1,\dots,y_k\}$ such that $d_Y(y_i,y_j) = \pi|i-j|$
and $\mu_Y(\{y_i\}) = 1/k$ for all $i,j = 1,2,\dots,k$.
The sequence $\{X_n\}$ do not satisfy the conclusions of
Theorem \ref{thm:lamk1} and Corollary \ref{cor:lamk-conc}.
It has no lower bound of Ricci curvature.
\end{ex}
\begin{rem}
The converse of Theorem \ref{thm:k-Levy-conc} also holds
in the following sense.
Let $X$ be an mm-space and $Y$ an mm-space consisting of
$k$ points. For any numbers $\kappa_0,\kappa_1,\cdots,\kappa_{k}$ with
$k\dconc(X,Y)< \min_i\kappa_i$, we have
\begin{align} \label{eq:Sep-dconc}
\Sep(X;\kappa_0,\kappa_1,\cdots,\kappa_k)\leq 2\dconc(X,Y).
\end{align}
In particular, if a sequence of mm-spaces concentrates to
a finite mm-space consisting of $k$ points, then
it is a $k$-L\'evy family.
\end{rem}
\begin{proof}[Proof of \eqref{eq:Sep-dconc}]
We take Borel subsets
$A_0,A_1,\cdots ,A_{k}\subset X$ such that
$\mu_X(A_i)\geq \kappa_{i}$ for $i=0,1,\cdots,k$.
Let $\alpha$ be an arbitrary positive number such that
\begin{align*}
\dconc(X,Y) < \alpha < \frac{1}{k}\min_i\kappa_i.
\end{align*}
There are two parameters $\varphi : I \to X$
and $\psi : I \to Y$ such that
\begin{align*}
d_H(\varphi^*\mathcal{L}{\it ip}_1(X),\psi^*\mathcal{L}{\it ip}_1(Y))<\alpha.
\end{align*}
We define functions $f_i :X\to \field{R} $
by $f_i(x):=d_X(x,A_i)$, $x \in X$, $i=0,1,\cdots,k$.
Since each $f_i$ is $1$-Lipschitz, there is $g_i \in \mathcal{L}{\it ip}_1(Y)$ such that
$\dKF(f_i \circ \varphi , g_i \circ \psi) < \alpha$.
This is equivalent to $\mathcal{L}^1(B_i)> 1-\alpha$,
where
\begin{align*}
B_i := \{\; s \in I \mid |(f_i \circ \varphi)(s) - (g_i \circ
\psi)(s)|<\alpha \;\}.
\end{align*}
For any $j=0,1,\cdots,k$, we have
\begin{align*}
\mathcal{L}^1\Big(\varphi^{-1}(A_j)\cap \bigcap_{i=0}^{k-1}B_i
\Big)\geq \mu_X(A_j)+ \mathcal{L}^1 \Big(\bigcap_{i=0}^{k-1}B_i
\Big) -1 \geq \kappa_j-k\alpha>0
\end{align*}
Take $s_j \in \varphi^{-1}(A_j) \cap \bigcap_{i=0}^{k-1}B_i$ for each $j$.
Since $\psi(s_j) \in Y$, $j=0,1,\dots,k$,
and $Y$ consists of $k$ points, it follows
from the pigeonhole principle that
$\psi(s_{i_1})=\psi(s_{i_2})$ for some $i_1$ and $i_2$
with $i_1<i_2$. Since $i_1\leq k-1$ and $s_{i_1}\in
\varphi^{-1}(A_{i_1})\cap B_{i_1}$, we have
\begin{align*}
|(g_{i_1} \circ \psi)(s_{i_2})|
&= |(g_{i_1} \circ \psi)(s_{i_1})|\\
&= |d_X(\varphi(s_{i_1}),A_{i_1})-(g_{i_1} \circ \psi)(s_{i_1})|
< \alpha.
\end{align*}
Combining this with $s_{i_2}\in \varphi^{-1}(A_{i_2})\cap B_{i_1}$,
we therefore obtain
\begin{align*}
d_X(A_{i_1},A_{i_2})\leq d_X(\varphi(s_{i_2}),
A_{i_1})<(g_{i_1}\circ \psi)(s_{i_2}) + \alpha <2\alpha,
\end{align*}
which implies \eqref{eq:Sep-dconc}.
\end{proof}
\begin{rem}
All the results in this section also hold for
weighted Riemannian manifolds with Bakry-\'Emery Ricci curvature
bounded from below, instead of Ricci curvature.
The proofs are the same.
\end{rem}
\section{Concentration of Alexandrov spaces}
\label{sec:Alex}
In this section, we prove the stability of a lower bound
of Alexandrov curvature under concentration.
We also prove a version of Corollary \ref{cor:lamk-conc}
for Alexandrov spaces.
\begin{defn}[$\tilde\angle x_1x_0x_2$]
\index{tildeangle@$\tilde\angle x_1x_0x_2$}
Take a real number $\kappa$ and fix it.
Let $X$ be a metric space and $M^2(\kappa)$ a complete simply connected
two-dimensional space form of constant curvature $\kappa$.
For different three points $x_0,x_1,x_2 \in X$,
we denote by $\tilde\angle x_1x_0x_2$
the angle between $\tilde x_0\tilde x_1$ and $\tilde x_0\tilde x_2$,
where $\tilde x_0$, $\tilde x_1$, $\tilde x_2$ are three points
in $M^2(\kappa)$ such that
$d_X(x_i,x_j) = d_{M^2(\kappa)}(\tilde{x}_i,\tilde{x}_j)$ for $i,j=0,1,2$,
and where $\tilde x_i\tilde x_j$
is a minimal geodesic joining $\tilde x_i$ to $\tilde x_j$.
$\tilde\angle x_1x_0x_2$ is defined
only if the following condition is satisfied:
\begin{align*}
(*)
\begin{cases}
&\kappa \le 0,\\
\text{or} & \kappa > 0
\ \text{and}\ \per(\triangle x_0x_1x_2) \le 2\pi/\sqrt{\kappa},
\end{cases}
\end{align*}
where $\per(\triangle x_0x_1x_2) := d_X(x_0,x_1) + d_X(x_1,x_2) + d_X(x_2,x_3)$.
We call ($*$) the \emph{perimeter condition for $x_0$, $x_1$, $x_2$}.
If the perimeter condition ($*$) is satisfied
for any different three points in $X$,
then we say that $X$ satisfies the \emph{perimeter condition for $\kappa$}.
\index{perimeter condition}
\end{defn}
Under the perimeter condition for $X$ and $\kappa > 0$,
we see that, if $d_X(x_0,x_i) = \pi/\sqrt{\kappa}$ for $i=1$ or $2$, then
$\tilde\angle x_1x_0x_2$ is not unique and can be taken to be
any real number between $0$ and $\pi$.
If $d_X(x_0,x_i) < \pi/\sqrt{\kappa}$ for $i=1,2$, then
$\tilde\angle x_1x_0x_2$ is uniquely determined and depends only on
$\kappa$ and $d_X(x_i,x_j)$, $i,j=0,1,2$.
\begin{defn}[Alexandrov space]
\index{Alexandrov space} \index{Alexandrov curvature}
A metric space $X$ is said to be of
\emph{Alexandrov curvature $\ge \kappa$}
if $X$ satisfies the perimeter condition for $\kappa$
and if for any different four points $x_0,x_1,x_2,x_3 \in X$ we have
\[
\tilde\angle x_1x_0x_2 + \tilde\angle x_2x_0x_3 + \tilde\angle x_3x_0x_1
\le 2\pi.
\]
An \emph{Alexandrov space of curvature $\ge \kappa$} is, by definition,
an intrinsic metric space of Alexandrov curvature $\ge \kappa$.
\end{defn}
This definition of Alexandrov space may be different from
those in the other literatures, but is equivalent to them.
For an Alexandrov space,
the Hausdorff dimension and covering dimension coincide to each other
and are called the \emph{dimension}.
Note that, for an Alexandrov space, the finiteness of dimension
implies the local compactness of the space, so that
a finite-dimensional Alexandrov space is a proper geodesic space.
We refer to \cites{BGP,Plaut} for the details for Alexandrov spaces.
\begin{lem}
\label{lem:dn}
Let $p_n : X_n \to Y$ be Borel measurable maps between
mm-spaces $X_n$ and $Y$, $n=1,2,\dots$. We assume that
each $p_n$ enforces $\varepsilon_n'$-concentration of $X_n$ to $Y$
with $\varepsilon_n' \to 0$,
and that $(p_n)_*\mu_{X_n}$ converges weakly to $\mu_Y$ as $n\to\infty$.
Let $B\subset Y$ be an open subset and let
\begin{align*}
d_n(x) := d_{X_n}(x,p_n^{-1}(B) \cap \tilde{X}_n) \quad\text{and}\quad
\underline{d}_n(x) := d_Y(p_n(x),B)
\end{align*}
for $x \in X_n$, where $\tilde{X}_n$ is a non-exceptional domain of $p_n$
for some additive error $\varepsilon_n \to 0$ as $n\to\infty$.
Then, for any $\varepsilon > 0$ we have
\begin{align}
\tag{1}
&\limsup_{n\to\infty}
\mu_{X_n}(d_n \ge \underline{d}_n + \diam B + \varepsilon) = 0,\\
\tag{2}
&\limsup_{n\to\infty} \mu_{X_n}(d_n \le \underline{d}_n - \varepsilon) = 0.
\end{align}
\end{lem}
\begin{proof}
We prove (1).
Suppose the contrary. Then, replacing $\{n\}$ with a subsequence,
we may assume that
\begin{align} \label{eq:dn3}
\mu_{X_n}(d_n \geq \underline{d}_n + D + \varepsilon) \ge \alpha
\end{align}
for all $n$ and for a constant $\alpha > 0$, where $D := \diam B$.
There are finitely many mutually disjoint Borel subsets
$Y_1,\dots,Y_N \subset Y$ such that
$\mu_Y((\bigcup_{i=1}^N Y_i)^\circ) > 1-\alpha/2$ and
$\diam Y_i \le \varepsilon/2$ for any $i=1,\dots,N$.
Since $(p_n)_*\mu_{X_n}$ converges weakly to $\mu_Y$, we have
\begin{align} \label{eq:dn4}
\mu_{X_n}\Bigl(p_n^{-1}\Bigl(\bigcup_{i=1}^N Y_i\Bigr)\Bigr) > 1-\alpha/2.
\end{align}
for all sufficiently large $n$.
By setting
\[
A_{in} := \{\;x \in p_n^{-1}(Y_i) \mid
d_n(x) \geq \underline{d}_n(x) + D + \varepsilon\;\},
\]
\eqref{eq:dn3} and \eqref{eq:dn4} together imply
\[
\mu_{X_n}\Bigl( \bigcup_{i=1}^N A_{in} \Bigr) > \alpha/2.
\]
There is a number $i_0$ such that $\mu_{X_n}(A_{i_0 n}) > \alpha/(2N)$
for infinitely many $n$.
Replacing $\{n\}$ with a subsequence we assume that
$\mu_{X_n}(A_{i_0 n}) > \alpha/(2N)$ for any $n$.
Letting $\kappa := \min\{\alpha /(2N),\mu_Y(B)/2\}$, we observe that
\[
\mu_{X_n}(A_{i_0 n}) > \kappa \quad\text{and}\quad
\mu_{X_n}(p_n^{-1}(B) \cap \tilde{X}_n) > \kappa
\]
for all sufficiently large $n$.
Applying Theorem \ref{thm:pn}(3), we obtain
\begin{align*}
\limsup_{n\to\infty} d_{X_n}(A_{i_0 n}, p_n^{-1}(B) \cap \tilde{X}_n)
&\le \limsup_{n\to\infty} d_+(p_n^{-1}(Y_{i_0}),p_n^{-1}(B);+\kappa)\\
&\le d_Y(Y_{i_0},B) + \varepsilon/2 + D,
\end{align*}
so that there is a point $x_n\in A_{i_0 n}$ for each $n$ such that
\[
\limsup_{n\to\infty} d_n(x_n) \le d_Y(Y_{i_0},B) + \varepsilon/2 + D
\le \inf_{p_n^{-1}(Y_{i_0})} \underline{d}_n + \varepsilon/2 + D,
\]
which contradicts $x_n\in A_{i_0 n}$.
(1) has been proved.
We prove (2).
Suppose that (2) does not hold.
Then, replacing with a subsequence we have
\[
\mu_{X_n}(d_n \le \underline{d}_n - \varepsilon) \ge \alpha
\]
for all $n$ and for some $\alpha,\varepsilon > 0$.
Since $\mu_{X_n}(\tilde{X}_n) \to 1$ as $n\to\infty$,
there is a point $x_n \in \tilde{X}_n$ for every sufficiently large $n$
such that $d_n(x_n) \le \underline{d}_n(x_n) -\varepsilon$.
We find a point $x_n' \in p_n^{-1}(B) \cap \tilde{X}_n$
such that $|d_n(x_n) - d_{X_n}(x_n,x_n')| \le \varepsilon_n$.
Since $p_n$ is $1$-Lipschitz up to the additive error $\varepsilon_n$,
we have
\[
\underline{d}_n(x_n) \le d_Y(p_n(x_n),p_n(x_n')) \le d_{X_n}(x_n,x_n') + \varepsilon_n
\le d_n(x_n) + 2\varepsilon_n,
\]
which is a contradiction.
This completes the proof.
\end{proof}
\begin{thm} \label{thm:Alexcurv-conc}
Let $X_n$, $n=1,2,\dots$, be mm-spaces of Alexandrov
curvature $\ge \kappa$ for a real number $\kappa$.
If $X_n$ concentrates to an mm-space $Y$ as $n\to\infty$,
then $Y$ is of Alexandrov curvature $\ge \kappa$.
\end{thm}
\begin{proof}
By Corollary \ref{cor:enforce},
there are Borel measurable maps $p_n : X_n \to Y$, $n=1,2,\dots$,
enforcing concentration of $X_n$ to $Y$ such that
$(p_n)_*\mu_{X_n}$ converges weakly to $\mu_Y$ as $n\to\infty$.
Applying Theorem \ref{thm:pn} yields that
$p_n$ is $1$-Lipschitz up to some additive error $\varepsilon_n \to 0$,
so that
\begin{equation}
\label{eq:Alex-curv}
d_Y(p_n(x),p_n(x')) \le d_{X_n}(x,x') + \epsilon_n
\end{equation}
for any $x,x' \in \tilde X_n$,
where $\tilde X_n$ is a non-exceptional domain of $p_n$ for $\varepsilon_n$.
We take any different four points $y_0,y_1,y_2,y_3 \in Y$ and fix them.
Let $o \in Y$ be a point different from $y_0,y_1,y_2,y_3$.
By Proposition \ref{prop:pq}, we may assume that
$p_n(x) = o$ for all $x \in X_n \setminus \tilde{X}_n$.
Let $\delta$ be any number such that $0 < \delta < d_Y(o,y_i)$
for $i=0,1,2,3$.
We see that $p_n^{-1}(B_\delta(y_i)) \subset \tilde{X}_n$
for any $i$ and $n$.
Set, for $x \in X_n$,
\[
d_{in}(x) := d_{X_n}(x,p_n^{-1}(B_\delta(y_i)),\quad
\underline{d}_{in}(x) := d_Y(p_n(x),B_\delta(y_i)).
\]
Lemma \ref{lem:dn} implies that
\[
\limsup_{n\to\infty} \mu_{X_n}(|d_{in}-\underline{d}_{in}| > 3\delta) = 0.
\]
Since $\liminf_{n\to\infty} \mu_{X_n}(p_n^{-1}(U_\delta(y_0)))
\ge \mu_Y(U_\delta(y_0)) > 0$,
there is a point $x_{0n} \in p_n^{-1}(U_\delta(y_0))$
for every sufficiently large $n$ such that
\[
|\,d_{in}(x_{0n})-\underline{d}_{in}(x_{0n})\,| \le 3\delta
\]
for $i=1,2,3$.
Since $d_Y(p_n(x_{0n}),y_0) < \delta$ and by a triangle inequality,
we have
\[
|\,\underline{d}_{in}(x_{0n})-d_Y(y_0,y_i)\,| < 2\delta.
\]
From the definition of $d_{in}$,
there is a point $x_{in} \in p_n^{-1}(B_\delta(y_i))$
such that
\[
|\,d_{in}(x_{0n})-d_{X_n}(x_{0n},x_{in})\,| < \delta.
\]
Combining these inequalities, we obtain
\[
|\,d_{X_n}(x_{0n},x_{in})-d_Y(y_0,y_i)\,| < 6\delta
\]
for $i=1,2,3$.
This holds for every $n$ large enough compared with $\delta$.
Besides, we have $d_Y(p_n(x_{in}),y_i) \le \delta$,
so that \eqref{eq:Alex-curv} and
$x_{in} \in p_n^{-1}(B_\delta(y_i)) \subset \tilde{X_n}$
together lead us to
\[
d_Y(y_i,y_j) - 2\delta
\le d_Y(p_n(x_{in}),p_n(x_{jn}))
\le d_{X_n}(x_{in},x_{jn}) + \epsilon_n
\]
for $i,j=1,2,3$.
By the arbitrariness of $\delta$,
we eventually obtain points $x_{in} \in X_n$ such that, for $i,j=1,2,3$,
\begin{align}
\label{eq:Alexcurv-conc1}
\lim_{n\to\infty} d_{X_n}(x_{0n},x_{in}) &= d_Y(y_0,y_i),\\
\label{eq:Alexcurv-conc2}
\liminf_{n\to\infty} d_{X_n}(x_{in},x_{jn}) &\ge d_Y(y_i,y_j).
\end{align}
Since $y_i$ are arbitrary points in $Y$,
formula \eqref{eq:Alexcurv-conc2}
and the perimeter condition for $X_n$ and $\kappa$ together
imply the perimeter condition for $Y$ and $\kappa$.
We also obtain, from \eqref{eq:Alexcurv-conc1} and
\eqref{eq:Alexcurv-conc2},
\[
\liminf_{n\to\infty} \tilde\angle x_{in}x_{0n}x_{jn}
\ge \tilde\angle y_iy_0y_j
\]
for $i,j=1,2,3$.
Since $X_n$ is of Alexandrov curvature $\ge \kappa$,
we have
\[
\tilde\angle x_{1n}x_{0n}x_{2n} + \tilde\angle x_{2n}x_{0n}x_{3n} +
\tilde\angle x_{3n}x_{0n}x_{1n} \le 2\pi,
\]
which together with the previous inequality yields
\[
\tilde\angle y_1y_0y_2 + \tilde\angle y_2y_0y_3 + \tilde\angle y_3y_0y_1
\le 2\pi.
\]
This completes the proof.
\end{proof}
\begin{thm}[Petrunin, Zhang, and Zhu; \cites{Petrunin,Zhang-Zhu}]
\label{thm:Alex-CD}
Let $X$ be an $n$-dimensional Alexandrov space of curvature $\ge \kappa$
for a constant $\kappa$.
Then, $X$ satisfies $\CD((n-1)\kappa,\infty)$.
\end{thm}
Note that they in fact proved that $X$ satisfies $\CD((n-1)\kappa,n)$
that is a stronger condition than $\CD((n-1)\kappa,\infty)$.
Let $X_n$, $n=1,2,\dots$, be compact Alexandrov spaces of
curvature $\ge \kappa$ for a real number $\kappa$.
We equip each $X_n$ with the $(\dim X_n)$-dimensional Hausdorff measure
normalized as the total measure to be one.
Assume that $X_n$ has an upper bound of diameter.
Let us discuss concentration of $X_n$ as $n\to\infty$.
In the case where the dimension of $X_n$ is bounded from above,
it is well-known (see \cite{Yamaguchi}*{Proposition A.4})
that each $X_n$ satisfies the doubling condition
and the doubling constant is bounded from above, so that
$\{X_n\}$ is a uniform family (see Remark \ref{rem:precpt}).
It then follows from Corollary \ref{cor:box-dconc-precpt} and
Remark \ref{rem:box-mGH} that
$X_n$ concentrates to an mm-space $Y$ as $n\to\infty$
if and only if $X_n$ converges to $Y$ in the sense of
measured Gromov-Hausdorff convergence.
It is more significant to consider the case where the dimension of $X_n$
diverges to infinity as $n\to\infty$.
Combining Theorem \ref{thm:Alex-CD} and Corollary \ref{cor:CD-Sep}
yields the following.
\begin{cor}
Let $X_n$, $n=1,2,\dots$, be compact finite-dimensional Alexandrov spaces of
curvature $\ge \kappa$ for a constant $\kappa$.
If $\kappa$ is positive
and if the dimension of $X_n$ diverges to infinity,
then $\{X_n\}$ is a L\'evy family.
\end{cor}
In the case where $X_n$ has nonnegative Alexandrov curvature,
we have the following theorem.
\begin{thm} \label{thm:Alex-conc}
Let $X_n$, $n=1,2,\dots$, be compact finite-dimensional Alexandrov spaces
of nonnegative curvature.
If $X_n$ concentrates to an mm-space $Y$ as $n\to\infty$,
then $Y$ is an Alexandrov space of nonnegative curvature.
\end{thm}
Note that $Y$ maybe infinite-dimensional.
\begin{proof}
By Theorem \ref{thm:Alexcurv-conc},
it suffices to prove that $Y$ is an intrinsic metric space.
Theorem \ref{thm:Alex-CD} says that each $X_n$ satisfies $\CD(0,\infty)$.
Applying Theorem \ref{thm:CD-conc} yields that $Y$ is an intrinsic metric
space.
This completes the proof.
\end{proof}
\begin{rem}
If $X_n$ have a negative lower curvature bound,
then the limit $Y$ is not necessarily an intrinsic metric space.
It is easy to construct such an example.
In fact, the manifold $(\,0,1\,) \times S^n$ with metric
$dt + f(t)\theta_n$ concentrates to a disconnected space
as $n\to\infty$ in general, where $\theta_n$ is the metric of
the unit sphere in $\field{R}^{n+1}$
and $f : (\,0,1\,) \to \field{R}$ is a function such that
the completion of the Riemannian manifold is diffeomorphic to a sphere,
i.e, $f(0+0) = f(1-0) = 0$, $f'(0+0) = f'(1-0) = 0$, etc.
If $-f''/f$ is bounded below, then the sectional curvature
of the manifold is bounded below.
The concentration limit is the closure of the subset of $(\,0,1\,)$ where
$f$ takes its maximum.
That is not necessarily connected and is not an intrinsic metric space
in general.
\end{rem}
We denote by $\lambda_k(X)$ the $k^{th}$ nonzero eigenvalue of the Laplacian
on a compact finite-dimensional Alexandrov space $X$,
where we refer to \cites{OS,KMS} for the Riemannian structure and
the Laplacian on an Alexandrov space.
We have the following proposition
in the same way as in the proof of Proposition \ref{prop:lamk-Sep}.
\begin{prop} \label{prop:lamk-sep}
Let $X$ be a compact finite-dimensional Alexandrov space.
Then we have
\[
\lambda_k(X) \Sep(X;\kappa_0,\dots,\kappa_k)^2
\le \frac{4}{\min_{i=0,1,\dots,k}\kappa_i}
\]
for any $\kappa_0,\dots,\kappa_k > 0$.
In particular, a sequence $\{X_n\}_{n=1}^\infty$ of compact
finite-dimensional Alexandrov spaces is a $k$-L\'evy family
if $\lambda_k(X_n)$ diverges to infinity as $n\to\infty$.
\end{prop}
\begin{cor} \label{cor:lamk-conc-alex}
Let $X_n$, $n=1,2,\dots$, be compact finite-dimensional Alexandrov spaces
of nonnegative curvature.
If $\lambda_k(X_n)$ diverges to infinity as $n \to \infty$
for some natural number $k$, then $\{X_n\}$ is a L\'evy family.
\end{cor}
\begin{proof}
Proposition \ref{prop:lamk-sep} says that
$\{X_n\}$ is a $k$-L\'evy family.
By Theorems \ref{thm:k-Levy-conc} and \ref{thm:Alex-conc},
$\{X_n\}$ is a L\'evy family.
This completes the proof.
\end{proof}
\begin{thm} \label{thm:lamk1-alex}
If $\{X_n\}_{n=1}^\infty$ is a L\'evy family of compact
Alexandrov spaces of nonnegative curvature,
then $\lambda_1(X_n)$ diverges to infinity as $n\to\infty$.
\end{thm}
\begin{proof}
The theorem follows from Theorem \ref{thm:Alex-CD},
\cite{Savare:self-imp}*{Corollary 4.3},
and the proof of \cite{Ldx:conc-isop}*{Theorem 1}.
\end{proof}
Using Theorem \ref{thm:lamk1-alex} and Corollary \ref{cor:lamk-conc-alex},
we obtain the following theorem
in the same way as in Theorem \ref{thm:lamk1}.
\begin{thm}
For any natural number $k$, there exists a positive constant
$C_k$ depending only on $k$
such that if $X$ is a compact Alexandrov space
of nonnegative curvature,
then we have
\[
\lambda_k(X) \le C_k \lambda_1(X).
\]
\end{thm}
\begin{bibdiv}
\begin{biblist}
\bib{AGZ}{book}{
author={Anderson, Greg W.},
author={Guionnet, Alice},
author={Zeitouni, Ofer},
title={An introduction to random matrices},
series={Cambridge Studies in Advanced Mathematics},
volume={118},
publisher={Cambridge University Press},
place={Cambridge},
date={2010},
pages={xiv+492},
isbn={978-0-521-19452-5},
}
\bib{Bil}{book}{
author={Billingsley, Patrick},
title={Convergence of probability measures},
series={Wiley Series in Probability and Statistics: Probability and
Statistics},
edition={2},
note={A Wiley-Interscience Publication},
publisher={John Wiley \& Sons Inc.},
place={New York},
date={1999},
pages={x+277},
isbn={0-471-19745-9},
}
\bib{Bog}{book}{
author={Bogachev, V. I.},
title={Measure theory. Vol. I, II},
publisher={Springer-Verlag},
place={Berlin},
date={2007},
pages={Vol. I: xviii+500 pp., Vol. II: xiv+575},
isbn={978-3-540-34513-8},
isbn={3-540-34513-2},
}
\bib{Bog:Gm}{book}{
author={Bogachev, Vladimir I.},
title={Gaussian measures},
series={Mathematical Surveys and Monographs},
volume={62},
publisher={American Mathematical Society},
place={Providence, RI},
date={1998},
pages={xii+433},
isbn={0-8218-1054-5},
}
\bib{Bon}{thesis}{
author={Bonciocat, A.-I.},
title={Curvature bounds and heat kernels: discrete versus continuous spaces},
type={Ph.D. Thesis},
organization={University of Bonn},
date={2008},
}
\bib{BBI}{book}{
author={Burago, Dmitri},
author={Burago, Yuri},
author={Ivanov, Sergei},
title={A course in metric geometry},
series={Graduate Studies in Mathematics},
volume={33},
publisher={American Mathematical Society},
place={Providence, RI},
date={2001},
pages={xiv+415},
isbn={0-8218-2129-6},
}
\bib{BGP}{article}{
author={Burago, Yu.},
author={Gromov, M.},
author={Perel{\cprime}man, G.},
title={A. D. Aleksandrov spaces with curvatures bounded below},
language={Russian, with Russian summary},
journal={Uspekhi Mat. Nauk},
volume={47},
date={1992},
number={2(284)},
pages={3--51, 222},
issn={0042-1316},
translation={
journal={Russian Math. Surveys},
volume={47},
date={1992},
number={2},
pages={1--58},
issn={0036-0279},
},
}
\bib{CGY}{article}{
author={Chung, F. R. K.},
author={Grigor{\cprime}yan, A.},
author={Yau, S.-T.},
title={Eigenvalues and diameters for manifolds and graphs},
conference={
title={Tsing Hua lectures on geometry \& analysis},
address={Hsinchu},
date={1990--1991},
},
book={
publisher={Int. Press, Cambridge, MA},
},
date={1997},
pages={79--105},
}
\bib{CS}{article}{
author={Colbois, Bruno},
author={Savo, Alessandro},
title={Large eigenvalues and concentration},
journal={Pacific J. Math.},
volume={249},
date={2011},
number={2},
pages={271--290},
issn={0030-8730},
}
\bib{CMS:PL}{article}{
author={Cordero-Erausquin, Dario},
author={McCann, Robert J.},
author={Schmuckenschl{\"a}ger, Michael},
title={Pr\'ekopa-Leindler type inequalities on Riemannian manifolds,
Jacobi fields, and optimal transport},
language={English, with English and French summaries},
journal={Ann. Fac. Sci. Toulouse Math. (6)},
volume={15},
date={2006},
number={4},
pages={613--635},
issn={0240-2963},
}
\bib{CMS:interp}{article}{
author={Cordero-Erausquin, Dario},
author={McCann, Robert J.},
author={Schmuckenschl{\"a}ger, Michael},
title={A Riemannian interpolation inequality \`a la Borell, Brascamp and
Lieb},
journal={Invent. Math.},
volume={146},
date={2001},
number={2},
pages={219--257},
issn={0020-9910},
}
\bib{DF}{article}{
author={Diaconis, Persi},
author={Freedman, David},
title={A dozen de Finetti-style results in search of a theory},
language={English, with French summary},
journal={Ann. Inst. H. Poincar\'e Probab. Statist.},
volume={23},
date={1987},
number={2, suppl.},
pages={397--423},
issn={0246-0203},
}
\bib{FLM}{article}{
author={Figiel, T.},
author={Lindenstrauss, J.},
author={Milman, V. D.},
title={The dimension of almost spherical sections of convex bodies},
journal={Acta Math.},
volume={139},
date={1977},
number={1-2},
pages={53--94},
issn={0001-5962},
}
\bib{Fuk}{article}{
author={Fukaya, Kenji},
title={Collapsing of Riemannian manifolds and eigenvalues of Laplace
operator},
journal={Invent. Math.},
volume={87},
date={1987},
number={3},
pages={517--547},
issn={0020-9910},
}
\bib{Funano:est-box}{article}{
author={Funano, Kei},
title={Estimates of Gromov's box distance},
journal={Proc. Amer. Math. Soc.},
volume={136},
date={2008},
number={8},
pages={2911--2920},
issn={0002-9939},
}
\bib{Funano:thesis}{article}{
author={Funano, Kei},
title={Asymptotic behavior of mm-spaces},
note={Doctoral Thesis, Tohoku University, 2009},
}
\bib{FS}{article}{
author={Funano, Kei},
author={Shioya, Takashi},
title={Concentration, Ricci curvature, and eigenvalues of Laplacian},
status={Geom. Funct. Anal. 23 (2013), Issue 3, 888-936.},
}
\bib{GroMil}{article}{
author={Gromov, M.},
author={Milman, V. D.},
title={A topological application of the isoperimetric inequality},
journal={Amer. J. Math.},
volume={105},
date={1983},
number={4},
pages={843--854},
issn={0002-9327},
}
\bib{Gromov}{book}{
author={Gromov, Misha},
title={Metric structures for Riemannian and non-Riemannian spaces},
series={Modern Birkh\"auser Classics},
edition={Reprint of the 2001 English edition},
note={Based on the 1981 French original;
With appendices by M. Katz, P. Pansu and S. Semmes;
Translated from the French by Sean Michael Bates},
publisher={Birkh\"auser Boston Inc.},
place={Boston, MA},
date={2007},
pages={xx+585},
isbn={978-0-8176-4582-3},
isbn={0-8176-4582-9},
}
\bib{Kechris}{book}{
author={Kechris, Alexander S.},
title={Classical descriptive set theory},
series={Graduate Texts in Mathematics},
volume={156},
publisher={Springer-Verlag},
place={New York},
date={1995},
pages={xviii+402},
isbn={0-387-94374-9},
}
\bib{Kondo}{article}{
author={Kondo, Takefumi},
title={Probability distribution of metric measure spaces},
journal={Differential Geom. Appl.},
volume={22},
date={2005},
number={2},
pages={121--130},
issn={0926-2245},
}
\bib{KMS}{article}{
author={Kuwae, Kazuhiro},
author={Machigashira, Yoshiroh},
author={Shioya, Takashi},
title={Sobolev spaces, Laplacian, and heat kernel on Alexandrov spaces},
journal={Math. Z.},
volume={238},
date={2001},
number={2},
pages={269--316},
issn={0025-5874},
}
\bib{KS}{article}{
author={Kuwae, Kazuhiro},
author={Shioya, Takashi},
title={Variational convergence over metric spaces},
journal={Trans. Amer. Math. Soc.},
volume={360},
date={2008},
number={1},
pages={35--75 (electronic)},
issn={0002-9947},
}
\bib{Ldx:book}{book}{
author={Ledoux, Michel},
title={The concentration of measure phenomenon},
series={Mathematical Surveys and Monographs},
volume={89},
publisher={American Mathematical Society},
place={Providence, RI},
date={2001},
pages={x+181},
isbn={0-8218-2864-9},
}
\bib{Ldx:conc-isop}{article}{
author={Ledoux, Michel},
title={From concentration to isoperimetry: semigroup proofs},
conference={
title={Concentration, functional inequalities and isoperimetry},
},
book={
series={Contemp. Math.},
volume={545},
publisher={Amer. Math. Soc., Providence, RI},
},
date={2011},
pages={155--166},
doi={10.1090/conm/545/10770},
}
\bib{Levy}{book}{
author={L{\'e}vy, Paul},
title={Probl\`emes concrets d'analyse fonctionnelle. Avec un compl\'ement
sur les fonctionnelles analytiques par F. Pellegrino},
language={French},
note={2d ed},
publisher={Gauthier-Villars},
place={Paris},
date={1951},
pages={xiv+484},
}
\bib{Lohr}{article}{
author={L\"ohr, Wolfgang},
title={Equivalence of Gromov-Prohorov- and Gromov's box-metric
on the space of metric measure spaces},
status={preprint},
}
\bib{LV}{article}{
author={Lott, John},
author={Villani, C{\'e}dric},
title={Ricci curvature for metric-measure spaces via optimal transport},
journal={Ann. of Math. (2)},
volume={169},
date={2009},
number={3},
pages={903--991},
issn={0003-486X},
}
\bib{Emil:isop}{article}{
author={Milman, Emanuel},
title={Isoperimetric and concentration inequalities: equivalence under
curvature lower bound},
journal={Duke Math. J.},
volume={154},
date={2010},
number={2},
pages={207--239},
issn={0012-7094},
}
\bib{Emil:role}{article}{
author={Milman, Emanuel},
title={On the role of convexity in isoperimetry, spectral gap and
concentration},
journal={Invent. Math.},
volume={177},
date={2009},
number={1},
pages={1--43},
issn={0020-9910},
}
\bib{Mil:Dvoretzky}{article}{
author={Milman, V. D.},
title={A new proof of A. Dvoretzky's theorem on cross-sections of convex
bodies},
language={Russian},
journal={Funkcional. Anal. i Prilo\v zen.},
volume={5},
date={1971},
number={4},
pages={28--37},
issn={0374-1990},
}
\bib{Mil:heritage}{article}{
author={Milman, V. D.},
title={The heritage of P.\ L\'evy in geometrical functional analysis},
note={Colloque Paul L\'evy sur les Processus Stochastiques (Palaiseau,
1987)},
journal={Ast\'erisque},
number={157-158},
date={1988},
pages={273--301},
issn={0303-1179},
}
\bib{Mil:inf-dim}{article}{
author={Milman, V. D.},
title={A certain property of functions defined on infinite-dimensional
manifolds},
language={Russian},
journal={Dokl. Akad. Nauk SSSR},
volume={200},
date={1971},
pages={781--784},
issn={0002-3264},
}
\bib{Mil:hom-sp}{article}{
author={Milman, V. D.},
title={Asymptotic properties of functions of several variables that are
defined on homogeneous spaces},
language={Russian},
journal={Dokl. Akad. Nauk SSSR},
volume={199},
date={1971},
pages={1247--1250},
translation={
journal={Soviet Math. Dokl.},
volume={12},
date={1971},
pages={1277--1281},
issn={0197-6788},
},
}
\bib{Ollivier:SnCPn}{article}{
author={Ollivier, Y.},
title={Diam\`etre observable des sous-vari\'et\'es de $S^n$ et $\field{C} P^n$},
note={m\'emoire de DEA, universit\'e d'Orsay},
date={1999},
}
\bib{Otto}{article}{
author={Otto, Felix},
title={The geometry of dissipative evolution equations: the porous medium
equation},
journal={Comm. Partial Differential Equations},
volume={26},
date={2001},
number={1-2},
pages={101--174},
issn={0360-5302},
}
\bib{OS}{article}{
author={Otsu, Yukio},
author={Shioya, Takashi},
title={The Riemannian structure of Alexandrov spaces},
journal={J. Differential Geom.},
volume={39},
date={1994},
number={3},
pages={629--658},
}
\bib{Petersen}{book}{
author={Petersen, Peter},
title={Riemannian geometry},
series={Graduate Texts in Mathematics},
volume={171},
edition={2},
publisher={Springer},
place={New York},
date={2006},
pages={xvi+401},
isbn={978-0387-29246-5},
isbn={0-387-29246-2},
}
\bib{Petrunin}{article}{
author={Petrunin, Anton},
title={Alexandrov meets Lott-Villani-Sturm},
journal={M\"unster J. Math.},
volume={4},
date={2011},
pages={53--64},
issn={1867-5778},
}
\bib{Plaut}{article}{
author={Plaut, Conrad},
title={Metric spaces of curvature $\geq k$},
conference={
title={Handbook of geometric topology},
},
book={
publisher={North-Holland},
place={Amsterdam},
},
date={2002},
pages={819--898},
}
\bib{vRS}{article}{
author={von Renesse, Max-K.},
author={Sturm, Karl-Theodor},
title={Transport inequalities, gradient estimates, entropy, and Ricci
curvature},
journal={Comm. Pure Appl. Math.},
volume={58},
date={2005},
number={7},
pages={923--940},
issn={0010-3640},
}
\bib{Savare:self-imp}{article}{
author={Savar{\'e}, Giuseppe},
title={Self-improvement of the Bakry-\'Emery condition and Wasserstein
contraction of the heat flow in ${\rm RCD}(K,\infty)$ metric measure
spaces},
journal={Discrete Contin. Dyn. Syst.},
volume={34},
date={2014},
number={4},
pages={1641--1661},
issn={1078-0947},
doi={10.3934/dcds.2014.34.1641},
}
\bib{Shioya:mmlim}{article}{
author={Shioya, Takashi},
title={Metric measure limits of spheres and complex projective spaces},
status={preprint},
}
\bib{Sturm:convex}{article}{
author={Sturm, Karl-Theodor},
title={Convex functionals of probability measures and nonlinear
diffusions on manifolds},
language={English, with English and French summaries},
journal={J. Math. Pures Appl. (9)},
volume={84},
date={2005},
number={2},
pages={149--168},
issn={0021-7824},
}
\bib{Sturm:geoII}{article}{
author={Sturm, Karl-Theodor},
title={On the geometry of metric measure spaces. II},
journal={Acta Math.},
volume={196},
date={2006},
number={1},
pages={133--177},
issn={0001-5962},
}
\bib{Sturm:geoI}{article}{
author={Sturm, Karl-Theodor},
title={On the geometry of metric measure spaces. I},
journal={Acta Math.},
volume={196},
date={2006},
number={1},
pages={65--131},
issn={0001-5962},
}
\bib{Vershik}{article}{
author={Vershik, A. M.},
title={The universal Uryson space, Gromov's metric triples, and random
metrics on the series of natural numbers},
language={Russian},
journal={Uspekhi Mat. Nauk},
volume={53},
date={1998},
number={5(323)},
pages={57--64},
issn={0042-1316},
translation={
journal={Russian Math. Surveys},
volume={53},
date={1998},
number={5},
pages={921--928},
issn={0036-0279},
},
}
\bib{Villani:topics}{book}{
author={Villani, C{\'e}dric},
title={Topics in optimal transportation},
series={Graduate Studies in Mathematics},
volume={58},
publisher={American Mathematical Society},
place={Providence, RI},
date={2003},
pages={xvi+370},
isbn={0-8218-3312-X},
}
\bib{Villani:oldnew}{book}{
author={Villani, C{\'e}dric},
title={Optimal transport},
series={Grundlehren der Mathematischen Wissenschaften [Fundamental
Principles of Mathematical Sciences]},
volume={338},
note={Old and new},
publisher={Springer-Verlag},
place={Berlin},
date={2009},
pages={xxii+973},
isbn={978-3-540-71049-3},
}
\bib{Yamaguchi}{article}{
author={Yamaguchi, Takao},
title={A convergence theorem in the geometry of Alexandrov spaces},
language={English, with English and French summaries},
conference={
title={Actes de la Table Ronde de G\'eom\'etrie Diff\'erentielle
(Luminy, 1992)},
},
book={
series={S\'emin. Congr.},
volume={1},
publisher={Soc. Math. France},
place={Paris},
},
date={1996},
pages={601--642},
}
\bib{Zhang-Zhu}{article}{
author={Zhang, Hui-Chun},
author={Zhu, Xi-Ping},
title={Ricci curvature on Alexandrov spaces and rigidity theorems},
journal={Comm. Anal. Geom.},
volume={18},
date={2010},
number={3},
pages={503--553},
issn={1019-8385},
}
\end{biblist}
\end{bibdiv}
\printindex
\end{document}
|
\section{Introduction}
\input{intro2}
\section{Correlation of the topological charge density}
\input{topcorr2}
\section{Applications of the BCNW formula}
\vspace*{-1mm}
\input{BCNW}
\section{Conclusions}
\input{conclu}
\vspace*{-2mm}
\input{acknow}
\vspace*{-1mm}
|
\section{Introduction}
Many systems of scientific and societal interest are composed of a large number of interacting elements, examples ranging from proteins interacting within each living cell to people interacting with one another within and across societies. These and many other systems can be conceptualized as networks, where network nodes represent the elements in a given system and network ties represent interactions between the elements. Network science and network analysis are used to analyze and model the structure of interactions in a network, an approach that is commonly motivated by the premise that network structure is associated with the dynamical behavior exhibited by the network, which in turn is expected to be associated with its function. In many cases, however, network structure is not static but instead evolves in time. This suggests that given a sequence of networks, it would be useful to determine points in time where the structure of the network changes in a non-trivial manner. Determining these points is known as the network change point detection problem. Given the connection between network structure and function, it seems reasonable to conjecture that a change in network structure may be coupled with a change in network function. Consequently, detecting structural change points for networks could be informative about functional change points as well.
In this paper, we consider the change point detection problem for correlation networks. These networks belong to a class of networks sometimes called similarity networks and they are obtained by defining the edges based on some form of similarity or correlation measure between each pair of nodes \citep{onnela2012taxonomies}.
Examples of correlation networks appear in many financial and biological contexts, such as stock market price and gene expression data \citep{onnela2004clustering,mizuno2006correlation,bhan2002duplication,kose2001visualizing,mantegna1999hierarchical}. In general, when evaluating correlation networks, the full data is used to estimate the correlations between the nodes. When using this approach for longitudinal data, it is sometimes implied that the network structure is the same over time. This assumption may however be inaccurate in some cases. For example, in \citet{onnela2004clustering} a stock market correlation network is created from almost two decades of stock prices. In reality the relationship between the stocks, and therefore the structure of the underlying network, likely change over such a long period of time, an issue that was addressed in \citet{onnela2004clustering} by dividing the data into shorter time windows. Similarly, in functional magnetic resonance imaging (fMRI) trials it is likely that the brain interacts differently during different tasks \citep{keightley2003fmri}, or possibly even within a given task, so it may be inaccurate to assume a constant brain activity correlation network in trials with multiple tasks.
Suppose that a network is constant or may be assumed so until a known point in time before undergoing sudden change. In this case the underlying data should be split up at the change point into two parts, and two separate correlation networks should be constructed from the two subsets of the data. In reality the location of the change point, or possibly several change points, is not known \textit{a priori} and must also be inferred from the data. This problem belongs to a wider class of so-called change point detection problems, which has been studied in the field of process control. When the observed node characteristics are independent and normally distributed, methods exist for general time series data to detect changes in the multivariate normal mean or covariance \citep{hawkins2005change,zamba2006multivariate,lowry1992multivariate}.
There have been some promising efforts at change point detection for structural networks, but in this case the actual network is observed over time rather than relying on correlations of node characteristics that are used to construct the network \citep{lindquist2007modeling,lindquist2008statistical,peel2014detecting,akoglu2010event,mcculloh2011detecting,tang2013attribute}. If a network is first inferred from correlations, then these methods could be applied. However, inferring networks from correlations is not trivial. When thresholding correlations to determine the adjacency matrix, which is a common approach, the inferred networks tend to be highly sensitive to the chosen threshold. For this reason, these methods cannot be directly applied to correlation networks without first solving the problem of inferring the correlation networks themselves. Therefore, despite this large body of methods developed for change point detection for both time series data and for networks, there is a need for a change point detection method specifically for correlation networks that is not hampered by stringent distributional assumptions.
In this paper we propose a computational framework for change point detection in correlation networks that is free from distributional assumptions. This framework offers a novel and flexible approach to change point detection. The change point detection method suggested by \citet{zamba2009multivariate,lowry1992multivariate} is adapted to our framework and its power to detect change points is compared to our method using simulation. Also, we investigate the general difficulty of change point detection near the boundaries of the data both analytically and through simulation. Finally, we apply our framework to both stock market and fMRI correlation network data and demonstrate its success and limitations for detecting functionally relevant change points.
\section{Method}
\subsection{Notation}
Assume that the system under investigation consists of a fixed number of $n$ nodes with characteristics observed at $T$ distinct time points, where the observed characteristics are $Y_{n \times T} = [Y_1,...,Y_T]$ where $Y_j = [Y_{1j},...,Y_{nj}]^T$ is the $j$th $n$-dimensional column vector of $Y$ and we assume that $Y_j \sim f(\cdot \mid \Sigma_j)$ is an unknown function with all columns of $Y$ i.i.d. (independent and identically distributed) and where $\mbox{cov}(Y_j)=\Sigma_j$. We also assume that the rows of $Y$, corresponding to observations at individual nodes, are centered to have temporal mean $0$ and scaled to have unit variance. Note that the centering and scaling, resulting in standardized observations for each node, can always be performed.
We define a set of diagonal matrices $D(i,j)_{T \times T}$ for $1 \leq i < j \leq T$ such that
$$D(i,j)_{kk} = \left\{\begin{array}{cc}
1/(j-i+1) & \text{if } i\leq k \leq j \\
0 & \mbox{otherwise} \\
\end{array}\right.$$
We define the covariance matrix $S(i,j)_{n \times n}$ on the subset of the data ranging from the $i$th column to the $j$th column, i.e., from time point $i$ to time point $j$ ($1 \leq i < j \leq T$), to be:
\begin{equation}\label{empcor}
S(i,j) = \sum_{k=i}^j Y_kY_k^T/(j-i+1) = YD(i,j)Y^T
\end{equation}
In order to detect a change point, we wish to find the value of $k$ in the range $(1+\Delta,T-\Delta)$ that maximizes the differences between $S(1,k)$ and $S(k+1,T)$, where $\Delta$ is picked large enough to avoid ill-conditioned covariance matrices ($\Delta > n$). The rationale for this approach is that if there were a change point in the data, the sample correlation matrices on each side of the change point ought to be different in structure. We choose the squared Frobenius norm as our metric for the distance between two matrices. Let our matrix distance metric be:
$$d(k) = ||S(1,k)-S(k+1,T)||_F^2 = \tr\{[S(1,k)-S(k+1,T)]^T[S(1,k)-S(k+1,T)]\},$$
where $\tr$ is the matrix trace operator.
We wish to test the hypotheses:
$$
\begin{array}{rl}
H_0 :& \Sigma_j = \Sigma \;\;\;\; \forall \; j \\
H_A :& \mbox{There exists $k$ such that } \Sigma_j = \left\{\begin{array}{cc}
\Sigma_1 & \mbox{ for } j \leq k \\
\Sigma_2 & \mbox{ for } j > k \\
\end{array} \right.
\end{array}
$$
\subsection{Existing methodology for change point detection}
Consider for a moment the case where the vector $Y_j$ is multivariate normal with expectation $\mu_1$ and variance-covariance matrix $\Sigma_1$ before the change point and expectation $\mu_2$ and variance-covariance $\Sigma_2$ after it. We denote this $Y_j \sim MVN(\mu_1,\Sigma_1)$ for $j\leq k$ and $Y_j \sim MVN(\mu_2,\Sigma_2)$ for $j>k$. A multivariate exponentially weighted moving average (EWMA) model has been developed for the detecting when $\mu_1$ changes to $\mu_2$ \citep{zamba2006multivariate,lowry1992multivariate}. A likelihood ratio test for detecting change points in the covariance matrix $\Sigma_1$ at a known fixed point $k$ was considered by \citet{zamba2009multivariate} and \citet{andersonw}. The likelihood ratio test statistic for detecting a change point at $k$ is
\begin{equation} \label{LRteststat}
\Lambda_k = \frac{|S(1,k)|^{\frac{k-1}{2}}\cdot |S(k+1,T)|^{\frac{T-k-1}{2}}}{|S(1,T)|^{\frac{T-1}{2}}},
\end{equation}
where $|\cdot|$ is the matrix determinant operator.
This approach makes the assumption that the location of the change point is known to be at $k$. In reality however the location of the change point is unknown, and the method can be extended to allow an unknown change point location by considering $\max_{1+\Delta\leq k \leq T-\Delta}\{\Lambda_k\}$. When the $Y_j$ are normally distributed then, for a fixed $k$, $-2\log(\Lambda_k)$ follows a chi-square distribution for large $T$ and for large $T-k$. Taking the maximum of $\Lambda_k$ over all possible $k$ results in a less tractable analytic distribution for the test statistic due to the necessity of correcting for multiple testing. For this reason, along with the fact that we do not wish to restrict ourselves to these distributional and asymptotic assumptions, we note that \eqref{LRteststat} can be easily adapted to the framework developed in Section \ref{ourmethod} by defining $d(k) = \Lambda_k$ and proceeding as usual. This suggests that different definitions of our matrix distance metric $d(k)$ can lead to substantially different results even in the same general framework. This idea is pursued further in Section \ref{normchoice}.
\subsection{Simulation based change point detection} \label{ourmethod}
It is of interest to establish a method of change point detection that does not require any distributional assumptions on $Y_j$, and we develop such a method in this section. Our approach is based on the bootstrap which offers a computational alternative that can well approximate the distribution of $Y_j$ through resampling. Under $H_0$, if the $Y_j$ are all independent and come from the same distribution $Y_j \sim f(\cdot \mid \Sigma)$ for all $1\leq j \leq T$, then bootstrapping the columns of $Y$ is appropriate. Though $f(\cdot \mid \Sigma)$ is unknown, we approximate it with the empirical distribution $\hat{f}(\cdot \mid \Sigma)$ which gives each observed column vector $Y_j$ an equal point mass of $1/T$. This is equivalent to resampling from the columns of $Y$ with replacement.
For many time series applications there may be autocorrelation present between the columns of $Y$. In this case resampling the columns of $Y$ would break the correlation structure and lead to bias in the approximation of the null distribution. To account for this autocorrelation in the resampling procedure, we use the sieve bootstrap \citep{buhlmann1997sieve}. In particular, for correlated data the $Y^{(b)}$ are generated for autocorrelation of order $s$ by fitting the model
\begin{equation}\label{sieveeq}
Y_j= \sum_{k=1}^s \hat{\phi}_k Y_{j-k} + \hat{\epsilon}_j
\end{equation}
for each $j>s$. The $\hat{\phi}_k$ are estimated from the Yule-Walker equations, and used to solve for the $\hat{\epsilon}_j$s through equation \eqref{sieveeq}. The bootstrap residuals , $\hat{\epsilon}_j^{(b)}$, are resampled from all the $\hat{\epsilon}_j$s with replacement. This generates the bootstrapped $Y_j^{(b)}$ according to $Y_j^{(b)}= \sum_{k=1}^s \hat{\phi}_k Y_{j-k} + \hat{\epsilon}_j^{(b)}$.
Let $Y^{(b)}$ be one of the bootstrap resamples from $Y$, where each $Y_j^{(b)}$ are generated by bootstrapping from $\hat{f}(\cdot,\Sigma)$ in the case of independence or from the sieve bootstrap for correlated data. This is repeated for $b \in \{1,\dots,B\}$ where $B$ is the total number of bootstrap samples. $\Delta$ is a ``buffer'' that limits the change point detection from searching too close to the boundaries of data. We recommend $\Delta \approx n$. In the case where a change point location $k$ is closer than $\Delta$ to either $1$ or $T$, the change point will not be detected but, as will be seen in Section \ref{boundarysection}, these cases are near impossible to detect regardless of how small we make $\Delta$. For each $k \in \{1+\Delta,\dots,T-\Delta\}$, $S^{(b)}(1,k)$, $S^{(b)}(k+1,T)$, and $d^{(b)}(k)$ are calculated where $S^{(b)}(i,j)=Y^{(b)}D(i,j)Y^{(b)T}$ and $d^{(b)}(k)=||S^{(b)}(1,k)-S^{(b)}(k+1,T)||_F$. Then $\hat{\mu}_{0}(k) = \frac{1}{B}\sum_{b=1}^B d^{(b)}(k)$ and $\hat{\sigma}_{0}^2(k)= \frac{1}{B-1}\sum_{b=1}^B (d^{(b)}(k)-\hat{\mu}_{0}(k))^2$ are calculated for each $k \in \{1+\Delta,\dots,T-\Delta\}$.
A z-score is then calculated for each potential change point $k \in \{1+\Delta,\dots,T-\Delta\}$ as
$$z^{(b)}(k) = \frac{d^{(b)}(k)-\hat{\mu}_{0}(k)}{\sqrt{\hat{\sigma}_{0}^2(k)}}$$
The change point occurs for the value of $k$ for which the z-score is largest, so we let $Z^{(b)} = \max_{k}\{z^{(b)}(k)\}$ for each bootstrap sample $b$. This is also performed on the observed data, with $z(k) = [d(k)-\hat{\mu}_{0}(k)]/\sqrt{\hat{\sigma}_{0}^2(k)}$ and $Z = \max_{k}\{z(k)\}$ being the test statistic. The corresponding p-value obtained from bootstrapping is
$$\mbox{p-value} = \frac{1}{B}\bigg|\left\{b: Z^{(b)} \geq Z\right\}\bigg|,$$
where $|\cdot|$ is the cardinality of the set. If the p-value is significant, i.e., if sufficiently few bootstrap replicates $Z^{(b)}$ exceed Z, then we reject $H_0$ and declare a change point exists for the value of $k$ with the highest z-score, i.e., at $\arg\max_k z(k)$.
It is also often the case that there exist more than one change point. In this case the data is split into two segments, one before the first change point and the other after it, and the bootstrap procedure is then repeated separately for each of the two segments. If a significant change point is found on a segment, then that segment is split in two again and this process is repeated until no more statistically significant change points are found. In practice, this procedure terminates after a small number of rounds because each iteration on average halves the amount of data which greatly reduces power to detect a change point after each subsequent iteration.
\subsection{Difficulty of detection near the boundary} \label{boundarysection}
We alluded above to the difficulty of detecting change points near the boundaries of the data, and will now investigate this issue in more detail. When a change point $k$ is very close $1$, then the empirical covariance matrix $S(1,k)$ is constructed using a very small amount of data and its estimate is unstable with high variance. Similarly, when $k$ is very close to $T$, $S(k+1,T)$ suffers from the same problem. This makes change point detection hard: if the empirical covariance matrix is highly variable, the noise from the estimation of the covariance matrices can make any possible differences between $\Sigma_1$ and $\Sigma_2$ statistically difficult to detect.
In an attempt to quantify just how difficult of a problem change point detection is near the boundary, we find the analytic form of $E[d(k)]$ under $H_0$ in the case of normally distributed $Y_{ij}$.
\newtheorem*{expval}{Theorem 1}
\begin{expval}
Let $Y_j \sim MVN(0,\Sigma)$ for $j \in \{1,...,T\}$ all i.i.d., then for $d(k) = ||S(1,k)-S(k+1,T)||_F^2$ for any $k \in \{2,...,T-1\}$ we have
\begin{equation} \label{expeq}
E[d(k)] = \left(\frac{1}{k} + \frac{1}{T-k}\right)(\tr(\Sigma^2)+\tr(\Sigma)^2)
\end{equation}
\end{expval}
\begin{proof}
We have that $d(k) = \tr\{[S(1,k)-S(k+1,T)]^2\} = \tr\{[Y^TY(D(1,k)-D(k+1,T))]^2\} = \sum_{i=1}^T\sum_{j=1}^T (Y_i^TY_j)^2C_{jj}C_{ii}$ where $C = D(1,k)-D(k+1,T)$.
From the variance of a Gaussian quadratic form we have that $E[(Y_i^TY_i)^2] = 2\tr(\Sigma^2)+\tr(\Sigma)^2$. Similarly for the covariance case when $i \neq j$, we have that $E[(Y_i^TY_j)^2)] = E[E[(Y_i^TY_j)^2|Y_j]] = E[Y_j^T\Sigma Y_j] = \tr(\Sigma^2)$. These combine to give us the result. For the more detailed algebra expanded upon, see Appendix \ref{expeqappendix}.
\end{proof}
The implication of the Theorem (Equation \eqref{expeq}) is that the expected difference asymptotes to infinity as $k$ approaches $0$ or $T$ and is minimized when $k=\lfloor T/2\rfloor$. Although this result assumes multivariate normal data, we expect that the qualitative nature of the result generalizes beyond the normal distribution. The increase in $E[d(k)]$ is confirmed through simulation under $H_0$ and demonstrated in Figure \ref{edktheoVSemp}. The implication is that the noise in the estimation of the covariance matrices on both sides of the change point is minimized when both $S(0,k)$ and $S(k+1,T)$ have sufficient data for their estimation. When $k$ is close to $0$, then even though $S(k+1,T)$ has low variability, the large increase in variability of $S(0,k)$ leads to an overall noisier outcome. This demonstrates that the strength of the method is only as strong as its weakest estimate. For the purposes of study design and data collection, if we suspect that a change point occurs at a certain location, perhaps for theoretical reasons or based on past studies, we need to ensure that there is sufficient data collected both before and after the suspected change point if we are to have any hope of detecting it.
\begin{figure}[!ht]
\centerline{\includegraphics[scale=.7,trim=0 0.6cm 0 1.5cm,clip=true]{Edk-theo-vs-empirical.pdf}}
\caption{Difficulty of change point detection near the boundaries of data. With $T=200$, $n=20$, and $\Sigma=I_n$ (the identity matrix of order $n$), for each potential change point $1< k< T$, we estimate $E[d(k)]$ by averaging $d(k)$ over $10000$ simulations under $H_0$ and show the location of expected values with markers. These empirical estimates are contrasted with the theoretical expectation, shown as a solid line, given by Theorem 1, Equation \eqref{expeq}.} \label{edktheoVSemp}
\end{figure}
\section{Simulation}
\subsection{The relationship between $T$ and $n$ for statistical power} \label{TVSn}
Estimation of the covariance matrix requires $T$ to be large relative to $n$ because the empirical covariance matrix has $n(n-1)/2$ elements that need to be estimated, so there is high variability in estimates if $T$ is small. If $T$ is too small, then even if a change point exists, the empirical covariance matrix may be so variable that the change point is undetectable. This problem is exacerbated when trying to detect change points near the boundary as discussed in Section \ref{boundarysection}.
While it is intuitive that $T$ needs to grow as some function of $n$ in order to maintain any reasonable statistical power to detect change points, it is unclear what that function of $n$ is. We investigate here further, using simulation, at what rate the number of longitudinal observations $T$ needs to grow with system size $n$ in order to maintain the same statistical power. We consider the case where a single change point occurs at the midpoint $\lfloor T/2\rfloor$, and $Y_j \sim MVN(0,\Sigma_1)$ for $j\leq T/2$ and $Y_j \sim MVN(0,\Sigma_2)$ for $j> T/2$ where:
$$
\Sigma_1=I_n \hspace{5pc}
\Sigma_2=\left[
\begin{array}{cc}
\underbrace{(1-\rho)I+\rho 11^T}_{4\times 4} & \underbrace{0}_{4 \times (n-4)} \\
\underbrace{0}_{(n-4)\times 4} & \underbrace{I}_{(n-4) \times (n-4)} \\
\end{array}
\right]
$$
where $\rho=0.9$. In other words, $\Sigma_2$ is a block or partitioned matrix with exchangeable correlation within the blocks on the diagonal, and 0s in the off-diagonal blocks. We simulate instances of $Y$ in this fashion 10000 times for each of $n=4,8,12$.
In Figure \ref{cpcombinedprobs} we compare the performance of the method, measured by the proportion of the 10000 iterations resulting in a statistically significant change point, described in Section \ref{ourmethod} for change point detection for the three different values of $n$. The asymmetry in Figure \ref{cpcombinedprobs} around the true change point is caused by having $\Sigma_1$ first followed by $\Sigma_2$. If the order of $\Sigma_1$ and $\Sigma_2$ is reversed, then the asymmetry will be reversed as well. We find that the probability of detecting the correct change point is the same for all $n$ if we increase $T$ by a quadratic rate in $n$ as $T(n) = n(n-1)+C$ for the constant $C$, a functional form we discovered by numerical exploration. In our simulations we considered the $\alpha=0.05$ significance level and $C=30$. The intuition behind a quadratic rate is that as $n$ increases, the number of entries in the empirical covariance matrix increases quadratically and therefore the noise in the Frobenius norm increases quadratically. Increasing $T$ quadratically with $n$ appears to balance out the added noise for increasing the dimensions of the correlation matrices, and stabilizes the statistical power to detect the change point. We would therefore recommend that if one wants to increase $n$, then there needs to be an associated increase in the number of observations that is quadratic in $n$ in order to retain the ability to detect a change point with the same power.
\begin{figure}[!ht]
\centerline{\includegraphics[scale=.9,trim=0 0.6cm 0 1.5cm,clip=true]{CPcombinedprobs.pdf}}
\caption{Change point detection statistical power as a function of $n$ and $T$. The $y$ axis represents the probability that a change point is detected at a particular time point. The $x$ axis is the distance of a time point in either direction from the true change point. These probabilities are estimated based on 10000 iterations for each $n$.} \label{cpcombinedprobs}
\end{figure}
\subsection{Comparison of different matrix norms} \label{normchoice}
Up until this point our proposed method has dealt with taking the Frobenius norm of the difference of empirical correlation matrices. The choice of the Frobenius norm was simply for algebraic simplicity of Theorem 1 (Equation \eqref{expeq}). Though it is more simple than many other matrix norms for such calculations, there is no reason to believe that the Frobenius norm is uniformly the best choice of matrix norm if the objective is to maximize the statistical power of change point detection. There may be some change points that the Frobenius norm is good at detecting, but there may be other change points for which a different matrix norm or distance metric would be more suitable. We investigate this question more closely in this Section.
Because the Frobenius norm sums the squared entries of a matrix, it is intuitive to expect that the Frobenius norm would be ideal for detecting change points in systems that demonstrate large-scale, network-wide changes in the correlation pattern. On the other hand, the Frobenius norm likely would not be very powerful in detecting small-scale local changes in correlation network structure. The rationale for this argument is that by summing over all the changes in the network structure, if there are very few changes relative to the entire network, then the Frobenius norm would be dominated by noise from the largely unchanged matrix elements.
We consider a different matrix norm, the Maximum norm, that is appealing for the case of small-scale, local changes. The Maximum norm of a matrix is simply the largest element of the matrix in absolute value. Intuitively, this norm would be ideal if there was just a single, but very large, change in the covariance matrix. If only one element of the covariance matrix changes, but the change is quite large, the Maximum norm would still be able to detect this change. Here the Frobenius norm would likely fail due to the sum of the all the changes being dominated by noise. The likelihood ratio test in Equation \eqref{LRteststat} is more similar to the Frobenius norm than the Maximum norm in that it utilizes all entries in the covariance matrix rather than using only one element. As a result, we may expect the likelihood-ratio distance metric to be more similar to the Frobenius norm in performance than it is to the Maximum norm.
We compare the Frobenius norm, the Maximum norm, and likelihood-ratio in Equation \eqref{LRteststat} through simulation with varying proportion of the network altered at the change point. To do this, we generated $Y_j$ from a multivariate normal distribution with $T=400$ and a single change point occurring at $t=200$. Prior to the change point $Y_j \sim MVN(0,\Sigma_1)$ for $j \leq t$ and after the change point $Y_j \sim MVN(0,\Sigma_2)$ for $j > t$, where we modify the dimension of the upper-left block of $\Sigma_2$ to change the proportion of the network that is altered at the change point. In each case $\rho$ is selected such that the change point is detected with $50\%$ power using the Frobenius norm. This provides a reference for how the Frobenius norm compares with the Maximum norm and the likelihood-ratio. The results are displayed in Figure \ref{comparenorms}, which confirms our intuition. When a small proportion of the network is altered, the Maximum norm is more powerful at change detection than the Frobenius norm and likelihood ratio metric. When a large proportion of the network is altered at the change point, then the Frobenius norm and likelihood ratio metric are more powerful. While the likelihood ratio metric is more similar to the Frobenius norm than it is to the Maximum norm, it is still less sensitive to wide-spread subtle network changes than the Frobenius norm.
\begin{figure}[!ht]
\centerline{\includegraphics[scale=.8,trim=0 0.6cm 0 1.5cm,clip=true]{PowerComparisonDifferentNorms.pdf}}
\caption{Power comparison for different matrix norms. For each point on the $x$-axis, a value of $\rho$ in the definition of $\Sigma_2$ is selected such that the Frobenius norm has $50\%$ power to detect a change point. As the proportion of the network altered at the change point increases, we adjust the value of $\rho$ correspondingly.} \label{comparenorms}
\end{figure}
These considerations naturally lead to the following question: which norm or metric should be used? The answer clearly depends on the anticipated nature of the change point, and is therefore difficult because often the nature of the change point is unknown. In fact, change point detection is used even when one is not sure a change point exists. If there is some \textit{a priori} knowledge of a type of change point perhaps specific to the problem at hand, then that information could be used to select an appropriate norm. For example, suppose we investigate a network constructed from stock return correlations and the time period under investigation happens to encompass a sudden economic recession. The moment the recession strikes, it is likely that there will be large-scale changes in the underlying network, and therefore the Frobenius norm might be a good choice.
One important issue that deserves emphasis is \textit{when} the choice of the norm to use should be made. It is very important that the choice of norm is made prior to looking at the data. If the analysis is performed multiple times repeatedly with different choices for the matrix norm, and the norm with the ``best'' results is selected, this would be deeply flawed and would invalidate the interpretation of the p-value. See \citet{gelman2013garden} for an informative discussion of the problem of inflated false positive rates that result when the choice of the specific statistical procedure to use, in this case the norm, is not made prior to all data analysis.
\subsection{Detecting multiple change points} \label{multiplesims}
It may be the case that more than one change point occurs in the data. In this case, the method described in Section \ref{ourmethod} can still be applied to search for additional change points by splitting the data into two segments at the first significant change point and then repeating the procedure on each segment separately. This process is repeated recursively on segments split around significant change points until no additional statistically significant change points remain. Each test is performed at the $\alpha=0.05$ level (or at another user-specifid level). Though multiple comparisons may seem like a potential problem here, in fact there is no problem because further tests are only performed conditional on the previous change points being elected as statistically significant. This prevents the false positive rate from being inflated. A sliding window approach can also be used to detect multiple change points as is done in \citet{hawkins2003changepoint} and \citet{peel2014detecting}.
We investigate the performance of our method for detecting multiple change points through simulation. Consider the case where $T=400$ time points are observed for $n=10$ nodes in a network. Data follows a multivariate normal distribution with mean 0 and covariance $\Sigma_1$ for $1\leq t\leq 100$ and for $201 \leq t \leq 300$, but has covariance $\Sigma_2$ for $101\leq t\leq 200$ and for $301 \leq t \leq 400$. We define $\Sigma_1$ and $\Sigma_2$ as in Section \ref{TVSn} with $\rho=0.9$, except that the upper left block of $\Sigma_2$ is $5 \times 5$ in this case. The probability that a change point is detected at each particular location is estimated from 10000 iterations and is shown in Figure \ref{multipleCPfig}. Statistical significance is assessed at the usual $\alpha=0.05$ significance level. We use the Frobenius norm here because changing between $\Sigma_1$ and $\Sigma_2$ constitutes large-scale change in the network.
\begin{figure}[!ht]
\centerline{\includegraphics[scale=.8,trim=4cm 3.48cm 4cm 4.3cm,clip=true]{CP-hist-05-n10-iters10000.pdf}}
\caption{Detection of multiple change points. The $y$ axis represents the probability a change point is detected in a bin, each containing five adjacent time points, based on 10000 simulations total. The true locations of the change points are marked with vertical blue lines. } \label{multipleCPfig}
\end{figure}
As expected, the closer a location is to a change point, the more likely that location is found to be a statistically significant change point. However, there is an asymmetry in the ability to detect the different change points. The change point at $t=200$ is more difficult to detect than the change points at $t=100$ and $t=300$. This is because if we consider all $400$ data points and split them around $t=200$, the two resulting covariance matrices on either side of $t=200$ are expected to be the same because for both sets of $200$ observations, half are from $\Sigma_1$ and half are from $\Sigma_2$. This means we have almost no statistical power to detect a change point in the first iteration at $t=200$. Instead, the change points at $t=100$ and $t=300$ are picked up first. The reason the change point at $t=100$ is easier to detect than the one at $t=300$, despite each being equally far from its respective boundary, is because the data is ordered with the first $100$ observations generated from $\Sigma_1$ and the last $100$ from $\Sigma_2$. If this is reversed, then $t=300$ becomes the change point most likely to be detected. After the first change point is detected, power is reduced for the remaining change points due to the reduction in sample size that occurs due to dividing the data into smaller segments.
\section{Data Analysis}
\subsection{Correlation networks of stock returns}
Our first data analysis example deals with networks constructed from correlations of stock returns. Networks constructed from correlations of stock returns have been used in the past to investigate the correlation structure of markets as well as to detect changes in their structure \citep{mantegna1999hierarchical,onnela2003dynamic,onnela2003dynamics}. Here we use a data set first analyzed in \citet{onnela2004clustering} and apply our change point detection methods to it.
A total of $n=114$ S\&P 500 stocks were followed from the beginning of 1982 to the end of 2000, keeping track of the stock price at closing for $T=4786$ trading days over that time period. This data is publicly available and had been gathered for analysis previously where correlation networks were constructed based on the correlation between log returns in moving time windows \citep{onnela2003dynamics}. If the price of the $i$th stock on the $j$th day is $P_{ij}$, then the corresponding log return is $R_{ij}=\log(P_{ij})-\log(P_{i,j-1)})$. The log returns did not demonstrate statistically significant autocorrelation (Durbin-Watson p-value of $0.11$) so the independent bootstrap was used.
Given that the stock market evolves constantly, and given the long time interval in the observed data, it may not be safe to assume that the correlation between the log returns of any two stocks stays fixed over time. If a correlation network were constructed by assuming edges between every two stocks with correlation greater than some threshold, and if all of the 19 years of data were used at once, the resulting network would likely be an inaccurate representation of the market if in fact the true underlying network changes with time \citep{onnela2006complex}. A more principled approach would be to first test for change points in the correlation network and then build multiple networks around those change points if necessary.
\begin{figure}[!ht]
\centerline{\includegraphics[scale=.9,trim=0 0 0 0,clip=true]{ChangePoint_StockPrice_sml.pdf}}
\caption{Change point detection in stock returns. With $n=116$ stocks tracked over $T=4786$ days ($\sim 19$ years), the blue line is the empirical z-score $z(k)$ while the clustered black lines are the z-scores simulated under $H_0$ using bootstrap as defined in Section \ref{ourmethod}. A significant change point is detected near the end of the year $1987$ corresponding to the well documented crash at the end of that year.} \label{StockPriceFig}
\end{figure}
Following the method proposed in Section \ref{ourmethod}, the stock price data is resampled 500 times assuming the null hypothesis of no change points, and the observed data is compared with these simulations to determine if and where a change point occurs. These simulation results are displayed in Figure \ref{StockPriceFig}. There is strong statistical evidence of a change point at the end of the year 1987 evidenced by a p-value $< 0.002$. The sieve bootstrap approach arrives at the same result, confirming the lack of temporal autocorrelation in log returns. We used the Frobenius norm as our goal was to find events that could lead to large-scale shocks to the correlation network. There were several other significant change points, but we focus on the first and most significant change point here. The stock market crash of October 1987, known as ``Black Monday'', coincides with the first detected change point \citep{onnela2003dynamic}. The stock market crash evidently drastically changed the relationship between many of the stocks leading to a stark change in the correlation network. For this reason it is advisable to consider the network of stocks before and after the stock market crash separately, as well as splitting the data further around potential additional significant change points, rather than lumping all of the data together to construct a single correlation network.
\subsection{Correlation networks of fMRI activity } \label{fmrisection}
Our second data analysis example deals with networks constructed from correlations in fMRI activity in the human brain. The Center for Cognitive Brain Imaging at Carnegie Mellon University collected fMRI data as part of the star/plus experiment for six individuals as they each completed a set of 40 trials \citep{mitchell2004learning}. Each trial took approximately 27 seconds to complete. The subjects were positioned inside an MRI scanner, and at the start of a trial, each subject was shown a picture for four seconds before it was replaced by a blank screen for another four seconds. Then a sentence making a statement about the picture just shown was displayed, such as ``The plus sign is above the star,'' and the subject had four seconds to press a button ``yes'' or ``no'' depending on whether or not the sentence was in agreement with the picture. After this the subject had an interstimulus period of no activity for 15 seconds until the end of the trial. We avoid referring to this as ``resting state'' due the reserved meaning of that label for extended periods of brain inactivity. Trials were repeated with different variations, such as the picture being presented first before the sentence, or with the sentence contradicting the picture. MRI images were recorded every $0.5$ seconds, for a total of about 54 images over the course of a trial, corresponding to a total of $40\times 54=2160$ images total. Each image was partitioned into 4698 voxels of width 3mm. The study data are publicly available \citep{cmufMRIwebsite}.
If we were to analyze a single trial, change point detection would be quite difficult for the data in its raw form for $n= 4698$ voxels which is very large compared with the number of data points $T=54$. Any empirical covariance matrix for these values of $n$ and $T$ would be too noisy to detect any statistically significant change point. We therefore combine our analysis on the eight trials where the picture is presented first and the sentence agrees with the picture for all six individuals. To accommodate the repeated trials in our correlation estimates, we define $S_{lk}(i,j)$ to be the covariance estimator in Equation \eqref{empcor} for the $l$th individual and $k$th trial only. The resulting covariance matrix averaged over all the trials and individuals is
$$S^*(i,j) = \sum_{l=1}^6\sum_{k=1}^8 S_k(i,j)/48$$
Change point detection is then performed as before except using the $S^*(i,j)$ instead of the usual $S(i,j)$. Even though we have effectively increased the amount of data 48-fold by combining multiple trials and individuals together, the number of observed data points is still far fewer than the $n=4698$ voxels. To reduce the number of nodes to a manageable size, we group the voxels into 24 distinct regions of interest (ROIs) in the brain following \citet{hutchinson2009modeling}, and we average the signals over all voxels within the same ROI. With 24 nodes and $54 \times 48 = 2592$ data points, empirical covariance matrices can be estimated with sufficient accuracy to detect change points in the network of ROIs so long as the change points occur sufficiently far from the beginning or end of the trial (see Section \ref{boundarysection}). Under the assumption that the network is drastically different when comparing the interstimulus state to the active state, we use the Frobenius norm. For each of the 6 individuals there was a significant presence of autocorrelation (Durbin-Watson p-value $<0.001$), so the sieve bootstrap is used for inference. First order autocorrelation ($s=1$) was used as it minimized mean squared error in cross-validation.
The most significant change point occurred $t=12$ seconds into the trial, though it was not statistically significant (p-value of $0.18$). This indicates that the network of interactions in the first part of the trial when the subject is actively reading, visualizing, responding, and connecting stimuli is most different from the interstimulus portion of the trial, though the difference is not statistically significant. If autocorrelation is ignored and instead independence is assumed, then the same change point appears significant with a p-value less than $0.001$. This example demonstrates how ignoring autocorrelation in the data can lead to an inflation of the false positive rate.
\begin{figure}[t!]
\centerline{\includegraphics[scale=1,trim=0 1cm 0 1cm ,clip=true]{BrainCPt12-ALL-edges20.pdf}}
\caption{Brain region of interest (ROI) networks before and after the most likely change point. The network transition around the most likely change point are displayed with two different layouts. In the top panel, nodes are positioned to best display the pre-change point network topology. Those same node positions are used in the post-change point network in the top right. The networks on the bottom have nodes positioned according to their Talaraich coordinates \citep{lancaster2000automated} that accurately represent their anatomical location in the brain.} \label{fMRInetworksfig}
\end{figure}
Though we fail to reject the null hypothesis and find no statistically significant change point at the $\alpha=0.05$ level, we examine the position of the most likely change point, if one exists, at $t=12$. We construct two networks between the ROIs, one from $S^*(1,24)$ corresponding to the first 12 seconds of the trial (recall that $i$ and $j$ in $S^*(i,j)$ index time points that are $0.5$ seconds apart) and one from $S^*(25,54)$ corresponding to the remaining 15 seconds of inactivity in the trial. The networks are constructed such that an edge is shown between two nodes if and only if their pairwise correlation is greater than 0.5 in absolute value. The two networks are displayed in Figure \ref{fMRInetworksfig}. The correlation threshold to determine if an edge is present was selected such that the network after the change point had $20$ edges, and the same threshold was used for both before and after networks. During the inactive period after $t=12$ there is an increase in connectivity in the network. An explanation for why the most likely change point occurs at $t=12$ could be because behavior after the change point corresponds with constant inactivity, whereas before the change point there is a mixture of inactivity, thinking, decisions, and other mental activity likely taking place. These different mental activities could dampen the observed correlations when averaged all together. Given the many different tasks that occur over the course of the trial, there almost certainly exist numerous changes in the correlation structure of brain activity. The lack of statistical significance in this example helps to illustrate the phenomenon discussed in section \ref{TVSn}; even if change points exist, as is almost certainly the case here due to the numerous mental tasks required over the course of a trial, there is no hope of discovering them unless the number of observations is far greater than the size of the network ($T \gg n$).
\section{Discussion}
In this paper, an existing change point detection method was adapted to correlation networks using a computational framework. Many past treatments of change point detection make a distributional assumption on the observed characteristics, but our framework utilizes the bootstrap in order to avoid this restriction. Traditional methods also assume independence between observations and upon first glance this assumption seems unreasonable. For instance, consider the stock market data. Stock prices are often modeled as a Markov process, which implies a strong autocorrelation between consecutive observations. For this reason the stock prices themselves cannot be used as input for our algorithm, but rather the log returns are used. Similar to the random noise in a Markov chain being independent, the assumption of the returns being independent is more reasonable, as was also demonstrated by the Watson-Durbin test. The fMRI voxel intensities, however, demonstrated significant autocorrelation and required the sieve bootstrap procedure.
We extended our framework to allow for multiple change points. If the first change point is found to be statistically significant, then the data is split into two parts on either side of the change point and the algorithm is repeated for each subset. This process of splitting the data around significant change points continues until there are no more significant change points. The fMRI data analysis in Section \ref{fmrisection} found no significant change points but, due to the many changes in stimuli, there are likely multiple points in time where the structure of interaction between regions of interest in the brain changes. This negative result could likely be remedied by collecting higher temporal frequency imaging data. With each split of the data, $T$ is approximately halved while $n$ remains the same and, as shown in Section \ref{TVSn}, this further lowers the power to detect change points. As long as the data have sufficient temporal resolution, it should be possible to use the proposed framework to detect multiple change points in correlation networks across different domains.
\section{Appendix: Proof of Theorem 1}
\subsection{Proof of Theorem 1} \label{expeqappendix}
The proof starts by writing the Frobenius norm as a sum over all pairs of observations, and then takes expectation, making use of what we know of the first two moments of a quadratic form of normally distributed variables. The following is the derivation of $E[d(k)]$ under $H_0$ for normally distributed observations:
\[\begin{array}{rl}
d(k) =& \tr\{[S(1,k)-S(k+1,T)]^T[S(1,k)-S(k+1,T)]\} \\
=& \tr\{[S(1,k)-S(k+1,T)]^2\} \\
=& \tr\{(Y^TY\underbrace{(D(1,k)-D(k+1,T))}_{C})^2\} \\
=& \sum_{i=1}^T\sum_{j=1}^T (Y_i^TY_j)^2C_{jj}C_{ii} \\
=& \sum_{i=1}^T\left\{(Y_i^TY_i)^2C_{ii}^2+\sum_{j \in \{1,...,i-1,i+1,...,T\}}(Y_i^TY_j)^2C_{ii}C_{jj}\right\} \\
\end{array}\]
Taking expectation gives
\[\begin{array}{rl}
E[d(k)] =& \sum_{i=1}^T\left\{E[(Y_i^TY_i)^2]C_{ii}^2+\sum_{j \in \{1,...,i-1,i+1,...,T\}}E[(Y_i^TY_j)^2]C_{ii}C_{jj}\right\} \\
=& \left(\frac{1}{k}+\frac{1}{T-k}\right)E[(Y_i^TY_i)^2] + \sum_{i=1}^k \left(\frac{k-1}{k^2}-\frac{1}{k}\right)E[(Y_i^TY_j)^2] \\
&+ \sum_{i=k+1}^T \left(\frac{T-k-1}{(T-k)^2}-\frac{1}{T-k}\right)E[(Y_i^TY_j)^2] \\
=& \left(\frac{1}{k}+\frac{1}{T-k}\right)E[(Y_i^TY_i)^2] + k \left(\frac{k-1}{k^2}-\frac{1}{k}\right)E[(Y_i^TY_j)^2] \\
& +(T-k)\left(\frac{T-k-1}{(T-k)^2}-\frac{1}{T-k}\right)E[(Y_i^TY_j)^2] \\
=& \left(\frac{1}{k}+\frac{1}{T-k}\right)E[(Y_i^TY_i)^2] + \left(\frac{k-1}{k}-1+\frac{T-k-1}{T-k}-1\right)E[(Y_i^TY_j)^2] \\
=& \left(\frac{1}{k}+\frac{1}{T-k}\right)(2\tr(\Sigma^2)+\tr(\Sigma)^2) + \left(\frac{k-1}{k}+\frac{T-k-1}{T-k}-2\right)\tr(\Sigma^2) \\
=& \left(\frac{1}{k} + \frac{1}{T-k}\right)(\tr(\Sigma^2)+\tr(\Sigma)^2) \\
\end{array}\]
The last line is the result of Theorem 1 in Equation \eqref{expeq}.
|
\section{Introduction}
The physics of perovskite compounds,
featuring $B$O$_6$ octahedra ($B$=cation) as building blocks,
is exceptionally rich.
The perovskites
can be driven through a variety of structural, electronic, and magnetic
phase transitions and
are hosts to such intriguing states of matter
as high-transition-temperature ({$T_{\text c}$}) superconductivity
[cuprates\cite{Bednorz86}, bismuth perovskites
$X$BiO$_3$ ($X$=Ba, Sr)
\cite{Cava88,Sleight75,Kazakov97}],
a pseudo-gap
state with strongly violated Fermi-liquid properties
(cuprates\cite{Timusk99}), and
a spin/charge density wave
[rare-earth nickelates $R$NiO$_3$ ($R$=rare-earth atom)
\cite{Medarde97}],
to name a few.
It is appealing to relate the diverse physical properties observed
across the perovskite family of materials with the individual
characteristics of the cation $B$. The latter can be magnetic (Cu
in the cuprates)
or not (Bi in the bismuth perovskites), orbitally active
(Mn in the manganites\cite{Tokura00})
or not (Cu, Bi),
prone to strong electronic correlations (transition-metal elements Cu and Ni)
or not (Bi).
With the due appreciation of the cation factor,
there are, however, many striking similarities
among different perovskite families suggesting
an equally important role of their common structural framework.
A vivid example is the transition
into a bond-disproportionated insulating phase found in both
the bismuthates
and the rare-earth
nickelates, in which oxygen plays an extremely important
role\cite{Mizokawa00,Park12,Lau13,Johnston14}.
Observations of this kind have given rise to
theories aimed at a unified description of the perovskites,
where the various competing phases emerge
from polaronic and bipolaronic excitations of
the polarizable oxygen sublattice.
While it is a common practice within this approach
to assume that the only effect
of electron-phonon coupling is variation of on-site energies,
in this Letter
we demonstrate that
the effects due to hybridization between
the oxygen-$p$ orbitals and cation orbitals
can be even more important.
For this purpose, we focus on the bismuth perovskites $X$BiO$_3$
where the analysis is greatly facilitated
by the fact that the Bi ion valence states are non-magnetic,
relatively weakly
correlated, orbitally non-degenerate, and, to a first approximation,
not affected
by spin-orbit coupling (the Bi-$6p$ states
are above the Fermi level).
Yet, we believe that our findings
have broader implications for the perovskites in general.
The bismuth perovskites {BaBiO$_3$} and {SrBiO$_3$}
are transition-metal-free high-{$T_{\text c}$} superconductors (when doped
with holes)
with intimately interlinked electronic and structural
phase transitions.
At low temperature, the parent compounds are
insulators with certain characteristic
distortions from an ideal cubic perovskite crystal structure.
Namely, the BiO$_6$ octahedra collapse
and expand alternately along all three cubic crystallographic
directions resulting
in {\it disproportionated Bi-O bond distances},
a so-called {\it breathing} distortion. Simultaneously, they rigidly rotate and tilt
to accommodate the Sr or Ba ions.
The insulating bond-disproportionated state
is often interpreted in terms of a charge-density wave.
It is speculated that the Bi ions, whose
nominal valency is $4+$, disproportionate
into Bi$^{3+}$ and Bi$^{5+}$\cite{Rice81,Cox76,Cox79,Varma88,Izumi07}
to avoid having a single electron in the $6s$ shell.
This picture, however,
is not supported experimentally\cite{deHair73,Orchard77,Wertheim82} and
is not consistent with the strongly covalent nature
of the Bi-O bonding on one hand and
the on-site repulsion effects on the other\cite{Harrison06,Mattheiss83}.
To stress the importance of these
two factors (covalency and on-site repulsion),
a useful analogy with the nickelates can be drawn.
There, the $3d^{7}$
configuration of Ni with one
electron in the $e_g$ doublet is similarly disfavored.
As was originally suggested by Mizokawa~{\it{et~al.}}\cite{Mizokawa00}
and later shown in Refs.~\onlinecite{Park12,Lau13,Johnston14},
electronic correlations,
strong nickel-oxygen hybridization,
and a negative charge-transfer gap
result in holes preferring to occupy
oxygen sites rather than nickel sites.
In a combination with electron-phonon coupling
to a lattice distortion of the breathing type,
this promotes a ($d^8$\underline{L}$^2$)$_{S=0}$($d^8$)$_{S=1}$
configuration instead of the classical
charge-disproportionated ($d^6$)$_{S=0}$($d^8$)$_{S=1}$ state
(\underline{L}=ligand hole, $S$=total on-site spin).
In the light of the above, here
we investigate the nature of
the insulating bond-disproportionated state
in the bismuthates from the perspective of strong Bi-$6s$/O-$2p$
hybridization
using density functional theory (DFT)
and local density approximation (LDA)\cite{SM}.
\begin{figure}[tb]
\begin{center}
\colorbox{white}
{\includegraphics[trim = 0mm 0mm 0mm 0mm, clip,width=\columnwidth]{./bands2}}
\end{center}
\caption{(a)~LDA electronic structure of {SrBiO$_3$}
projected onto the Bi-$6s$ orbital and
combinations of the O-$p_{\sigma}$ orbitals
of a collapsed (top) and expanded (bottom)
BiO$_6$ octahedron. For the doublet $E_g$
and the triplet $T_{1u}$, only one
projection is shown. The Fermi level is set to zero,
and PDOS stands for projected density of states
and is given in states/eV/cell.
(b)~An octahedron of
O-$p_{\sigma}$ orbitals coupled via
nearest-neighbor hopping integrals $-t$ and
its eigenstates.
}
\label{F.bands2}
\end{figure}
\begin{figure}[tb]
\begin{center}
\colorbox{white}
{\includegraphics[trim = 0mm 0mm 0mm 0mm, clip,width=\columnwidth]{./fig1}}
\end{center}
\caption{
(a), (b): LDA characterization of {SrBiO$_3$} model structures
with varying degrees of the BiO$_6$ octahedra's tilting, $t$, and breathing, $b$:
(a) total energy per formula unit (f. u.)
and (b) charge gap.
In (a), solid lines and filled circles (dashed lines
and open circles) represent model structures
with fixed (relaxed) Sr atoms. The horizontal dashed line
marks the energy of the experimental {SrBiO$_3$} structure.
(c), (d): The effect of tilting on (c) the half-filled band
and
on (d) the static susceptibility $\chi({\bf q},\omega=0)$,
at zero breathing.
In (d), solid (dashed) lines represent calculations
where non-linear effects due to tilting
are (are not) taken into account.
}
\label{F.str}
\end{figure}
\begin{figure*}[tb]
\begin{center}
\colorbox{white}
{\includegraphics[trim = 0mm 0mm 0mm 0mm, clip,width=2\columnwidth]{./bands3}}
\end{center}
\caption{
LDA electronic structure of {SrBiO$_3$}
as a function of breathing $b$ and tilting $t$.
Projections are made onto the Bi-$6s$ orbital and
the $A_{1g}$ combination of the O-$p_{\sigma}$ orbitals
of a collapsed BiO$_6$ octahedron,
as well as their bonding (``B'') and anti-bonding (``A'')
combinations.
}
\label{F.bands3}
\end{figure*}
We begin with highlighting basic features of the bismuthates' electronic
structure
\cite{Mattheiss83,Shirai90,Hamada89,
Liechtenstein91,Kunc91,Kunc94,Meregalli98,
Nourafkan12,Korotin12,Yin13,Korotin13}.
Since the LDA bandstructures of {BaBiO$_3$} and {SrBiO$_3$}
are almost identical\cite{SM},
in the following
we will concentrate on
{SrBiO$_3$}.
The Bi-$6s$
and O-$2p$ orbitals strongly hybridize to create
a broad band manifold near the Fermi level $E_F$ extending from
-14 to 3~eV [Fig.~\ref{F.bands2}~(a)].
The upper band of this manifold is
a mixture of the Bi-$6s$ orbitals and the $\sigma$ $2p$ orbitals of the oxygens,
{\it{i.~e.}}, the O-$2p$ orbitals with lobes pointing towards the central Bi\cite{comment}.
This upper band is half-filled as a result of {\it self-doping}:
the 18 $2p$ states of three oxygen ions and the two $6s$ states
of a Bi ion are short of one electron to be fully occupied.
Upon breathing, this band splits producing a charge gap.
We note that even in the bond-disproportionated
state, the bismuthates are very far from the ionic limit
of pure Bi$^{3+}$ and Bi$^{5+}$. The self-doped holes
reside predominantly on the oxygen-$p_{\sigma}$ orbitals,
a situation similar to that in the nickelates.
Below, we will infer the exact distribution of holes
by analysing the DFT data in more detail.
In our further analysis, we will be guided
by the following considerations.
First, cation valence states
of a given symmetry
will be strongly hybridized with only certain
combinations of surrounding O-$p$ orbitals, the Zhang-Rice singlet
in the cuprates
being an example\cite{Zhang88}.
Thus,
a spherically symmetric Bi-$s$ orbital
should couple only to the $A_{1g}$ ({\it{i.~e.}},
fully symmetric) combination
of the $p_{\sigma}$ orbitals of the surrounding
oxygen atoms [Fig.~\ref{F.bands2}~(b)]. Hoppings to any other combination
of the O-$p$ orbitals cancel out by symmetry.
Second, since the Bi-$6s$ orbitals are
quite extended,
the hybridization between the Bi-$s$ and O-$A_{1g}$
orbitals is very strong and even becomes a dominant effect.
Third, in the bond-modulated phase, hybridization
within a collapsed (expanded) BiO$_6$ octahedron
is strongly enhanced (suppressed), resulting
in a formation of {\it local, molecular-like orbitals}
on the collapsed octahedra.
We are able to observe all of these effects in our DFT results
by projecting the LDA single-particle eigenstates onto
the combinations of O-$p_{\sigma}$ orbitals.
Let us first focus of the {\it collapsed}
octahedra. In the top panel of Figure~\ref{F.bands2}~(a),
one finds the Bi-$6s$ and O-$A_{1g}$ characters
at $\sim2$~eV and at $\sim-10$~eV. The difference, $\sim12$~eV,
is mainly a result of the hybridization splitting
between the bonding and anti-bonding
combinations, which indeed turns out to be the dominant energy scale
for {SrBiO$_3$}.
The anti-bonding combination
is strongly peaked in the lowest two conductance
bands throughout the whole Brillouin zone.
In contrast, the location of the anti-bonding
$A_{1g}$ combinations of the {\it expanded} octahedra (bottom panel)
is rather diffuse, with some weight seen both below and above $E_F$
depending on the position in the Brillouin zone.
This indicates that the metal-insulator transition
with bond-disproportionation in the bismuthates
should be understood as a pairwise spatial condensation of
holes
into the anti-bonding
$A_{1g}$ molecular orbitals of the collapsed octahedra.
The small charge-disproportionation between the Bi ions
($\pm$0.15~$e$
inside the Bi muffin tin spheres)
appears to be a {\it marginal side effect} of such hole condensation.
Such a state,
which resembles the ($d^8$\underline{L}$^2$)$_{S=0}$ singlet
proposed for the nickelates,
was also hypothesized in Ref.~\onlinecite{Menushenkov01}.
It is interesting to trace the
evolution of the LDA projected density of states as a function of
breathing $b$ and tilting $t$.
For this,
we prepare a set of {SrBiO$_3$}
model structures with varying
degrees of $b$ and $t$\cite{SM}.
In Figures~\ref{F.str}~(a) and (b),
the model structures are
characterized in terms of, respectively, the total energy
and the charge gap calculated within
LDA.
Parameters $b$ and $t$ are given with respect to
the experimentally observed distortions $b_{\text{exp}}$
and $t_{\text{exp}}$.
We find that a finite $t$
is required to stabilize the breathing
distortion.
Although this result is known to be
sensitive to details of DFT calculations\cite{Kunc91},
the absence of the breathing instability
at zero tilting can be qualitatively
understood in terms of a missing linear component
in the elastic energy of an oxygen atom positioned between
two Bi atoms.
With increasing $t$, the equilibrium $b$ shifts to higher values
[Fig.~\ref{F.str}~(a)],
while the charge gap opens sooner as a function of $b$ [Fig.~\ref{F.str}~(b)].
We conclude, in agreement with previous DFT
studies\cite{Liechtenstein91,Kunc91,Korotin12},
that tilting enhances the breathing instability,
through a mechanism yet to be discussed.
Interestingly, letting
Sr atoms relax from the high-symmetry positions (while keeping $b$ and $t$ fixed)
can significantly lower
the total energy for structures with high $t$
[open circles in Fig.~\ref{F.str}~(a)].
For the $(b_{\text{exp}},t_{\text{exp}})$ structure, for instance,
the energy drop is 0.15~eV per formula unit.
Figure~\ref{F.bands3} presents
projections onto collapsed octahedra
for the $(0,0)$, $(b_{\text{exp}},0)$,
$(0,t_{\text{exp}})$, and $(b_{\text{exp}},t_{\text{exp}})$ structures.
Here, the bonding (``B'') and anti-bonding (``A'') combinations
of the Bi-$6s$ and O-$A_{1g}$ orbitals are shown explicitly.
An important observation is that
even with no breathing a major portion of the anti-bonding
orbital weight is above $E_F$. Thus, the system is {\it predisposed}
to hole condensation. The breathing distortion just
makes collapsed octahedra more preferable for the holes to go to.
On the other hand, the primary role of the tilting distortion
is to reduce the bandwidth.
This band-narrowing, however, is very {\it non-uniform}
leading to certain important consequences,
as will be discussed later.
\begin{figure}[tb]
\begin{center}
\colorbox{white}
{\includegraphics[trim = 0mm 0mm 0mm 0mm, clip,width=\columnwidth]{./bands4}}
\end{center}
\caption{
(a) LDA electronic structure
of the oxygen sublattice of {SrBiO$_3$}.
Projections are made onto
combinations of the O-$p_{\sigma}$ orbitals
of a collapsed O$_6$ octahedron.
(b) Model density of states
as a function of breathing $b$
and hybridization between $s$- and $p$-orbitals $h$.
CC (EC) stands for a collapsed (expanded) $p$-site cage.
The model states are 90\% filled,
{\it{i.~e.}}, there is one hole per $s$-orbital;
Fermi energy is set to zero
and marked with black dashed vertical lines.
}
\label{F.bands4}
\end{figure}
Formation of well-defined molecular O-$A_{1g}$ states on collapsed
octahedra is a property of the oxygen sublattice.
This is illustrated in Fig.~\ref{F.bands4}~(a)
which shows LDA calculations for an artificial
system consisting of only the oxygen sublattice of {SrBiO$_3$}.
Here, however, the $A_{1g}$ states are
the lowest in energy and fully occupied.
It is the strong hybridization with the nearly degenerate Bi-$6s$ states
that pushes them above $E_F$.
Upon hybridization, the 2~eV broad
oxygen band and the essentially flat Bi-$6s$ band will
acquire a bandwidth of $\sim15$~eV.
As a further illustration, we can consider a simple tight-binding (TB) model
for a two-dimensional perovskite-like lattice
of $p_{\sigma,\pi}$-orbital sites and $s$-orbital sites,
with nearest-neighbor $s-p$ and $p-p$ hoppings
as found in {SrBiO$_3$}:
$|t_{sp\sigma}|\sim2.3$, $|t_{pp\sigma}|\sim0.3$, and
$|t_{pp\pi}|\sim0.1$~eV\cite{SM}.
Switching on the breathing
distortion helps to form molecular $A_{1g}$ states
associated with collapsed $p_{\sigma}$-orbital cages
at the bottom of the $p$-band
[compare the left and central panels of Fig.~\ref{F.bands4}~(b)].
Subsequent hybridization with the flat $s$ orbitals,
located right below the $A_{1g}$ states,
has a much stronger effect (since $|t_{sp}|\gg|t_{pp}|$) and
pushes the anti-bonding $s-A_{1g}$ combination
above $E_F$
[right panel of Fig.~\ref{F.bands4}~(b)].
Finally, we demonstrate the
significance of the non-uniform nature of the
tilting-induced band-narrowing.
For this purpose, we calculate
the static susceptibility $\chi({\bf q},\omega=0)$
with the help of a single-band TB model
for structures with zero breathing and varying tilting.
Two kinds of TB models,
describing the band crossing the Fermi level,
are considered:
(1) realistic $t\neq0$ models closely following the LDA bands\cite{SM}
and (2) $t\neq0$ models uniformly rescaled
with respect to the $t=0$ case
such as to match the LDA bandwidth.
In Fig.~\ref{F.str}~(c), the band dispersions of the realistic models
are plotted in the Brillouin zone of a small cubic cell
so that the non-uniform nature of the $t$-induced band-narrowing
is particularly clear: the changes at $R$, for example, are much smaller than
at $\Gamma$. Using the realistic models and considering
all the energies spanned by the band, we
find a
dominating susceptibility
peak at ${\bf q}=(\pi,\pi,\pi)$ signaling breathing instability
that quickly grows with increasing tilting [Fig.~\ref{F.str}~(d), solid
lines].
This growth is much quicker
than it would be in the case of a uniform band-narrowing (compare
with the dashed lines). This indicates
that non-linear effects due to tilting,
such as, possibly, approaching perfect nesting conditions,
are very important for stabilizing the breathing distortion.
In summary, we have studied hybridization effects
in the bismuth perovskites and their
interplay with structural distortions such as breathing
and tilting. It is shown that strong hybridization between the Bi-$6s$
and O-$2p$ orbitals precludes purely ionic
charge disproportionation of a Bi$^{3+}$/Bi$^{5+}$ form.
Instead, the (self-doped) holes spatially condense into
molecular-orbital-like $A_{1g}$ combinations
of the Bi-$6s$ and O-$2p_{\sigma}$ orbitals
of {\it collapsed} BiO$_6$ octahedra,
with predominantly O-$2p_{\sigma}$ molecular orbital character.
The
tilting distortion is found to strongly enhance
the breathing instability through (at least in part)
an electronic mechanism,
as manifested by a ${\bf q}=(\pi,\pi,\pi)$ peak
in the static susceptibility $\chi({\bf q},\omega=0)$.
It is expected that similar processes
take place in other perovskites as well
and can involve localized molecular orbitals of symmetries
other than $A_{1g}$. Thus, in the rare-earth
nickelates, the orbitals of relevance
would be the $E_g$ combinations of the O-$p_{\sigma}$ orbitals
hybridized with the Ni-$e_g$ orbitals.
Our model calculations
indicate that two-dimensional systems, such as cuprates,
can also exhibit hybridization effects of this kind.
\bibliographystyle{apsrev}
|
\section{Introduction}
The mathematical theory of homogenization was introduced in order to describe the behavior of composite materials and reticulated structures.
These are characterized by the fact that they contain different constituents, finely mixed in a structured way which
bestows enhanced properties on the composite material.
Since heterogeneities are small compared with global dimensions, usually different scales are used to describe the material: a macroscopic scale describes the behavior of the bulk, while at least one microscopic scale describes the heterogeneities of the composite. In the context of Calculus of Variations the limiting (homogenized) behavior is usually captured through $\Gamma$-convergence techniques (see \cite{DGF}, \cite{DM93}, \cite{BRA}) or through the unfolding operator (see \cite{Dm1}, \cite{Dm2}, \cite{Dm3}). An important tool to address the case of two-scale homogenization was developed in the works of G.\@ Nguetseng \cite{NG} and G.\@ Allaire \cite{Allaire} (see also \cite {BA} and \cite{LNW}).
Though there is an extensive bibliography in the subject, here we mention some results closely connected to this work, namely homogenization results for different growth or coercivity conditions of the integrand, and for integrands depending on gradients, or more generally depending on vector fields in the kernel of a constant-rank, first-order linear differential operator $\cA$, i.e., in the context of $\mathcal{A}$-quasiconvexity.
The asymptotic behavior of functionals modeling elastic materials with fine microstructure (the fineness being accounted for by a small parameter $\eps>0$), without considering the possibility of fracture was developed in the seminal works of S.\@ M\"uller \cite{M} and A.\@ Braides \cite {B}. The case involving fracture was considered in \cite{BDV}.
In the first two cases, the models involve integrands depending on the derivative of a Sobolev function whereas in the third the integrands depend on the derivative of a special function of bounded variation.
The growth conditions assumed on the integrand function are usually of order $p>1$ and $p=1$, respectively, thus allowing for concentrations to occur in the latter case.
To ensure lower semicontinuity of the functionals, quasiconvexity is the property that the integrand functions are required to satisfy.
It is the natural generalization to higher dimension of the notion of convexity, and it is expressed by an optimality condition with respect to variations that are gradients.
This notion was further generalized by the introduction of $\cA$-quasiconvexity, and, as noted by Tartar \cite{T}, this allows to treat more general problems in continuum mechanics and electromagnetism, where constraints other than vanishing curl are considered.
We present here examples of such operators (for more examples see \cite{FM}):
\begin{enumerate}
\item ($\cA= \text{div}$) For $\mu \in {\cM}(\Omega; \R{N})$ we define
$$\cA \mu = \sum_{i=1}^N \frac {\partial \mu^i} {\partial x_i}.$$
\item ($\cA=\text{curl}$) For $\mu \in {\cal M}(\Omega; \R{d})$, with $d=m \times N$, we define
$$\cA \mu = {\left( \frac {\partial \mu_k^j} {\partial x_i} - \frac {\partial \mu_i^j} {\partial x_k} \right) }_{j=1,..,m;\, i,k=1,..,N}.$$
\item (Maxwell's Equations) For $\mu \in {\cM}(\R{3}; \R{3 \times 3})$ we define
$$ \cA \mu = \left( \text{div} (m+h), \text{curl} \,h \right)$$
where $\mu = (m,h).$
\end{enumerate}
The notion of $\mathcal{A}$-quasiconvexity and its implications for the lower semicontinuity of functionals was first investigated by B.\@ Dacorogna \cite{Dac} and later developed by I.\@ Fonseca and S.\@ M\"uller \cite{FM} for the study of lower semicontinuity of functionals on $\mathcal{A}$-free fields with growth $ p > 1.$
A.\@ Braides, I.\@ Fonseca, and G.\@ Leoni \cite{BFL} derived a homogenization result in the context of $\cA$-quasiconvexity for integrands with growth of order $ p >1$, with a microscopic scale.
In \cite {FK}, the authors use the unfolding operator to handle a homogenization problem with $p$-growth, but no coercivity, for integrands with two scales.
Allowing for linear growth, i.e., taking $p=1$, and coercivity implies working with sequences that are only bounded in $L^1$ and hence that can converge weakly-* (up to a subsequence) to some bounded Radon measure. In \cite{KriRin_CV_10} the relaxation of signed functionals with linear growth in the space BV of functions with bounded variation was studied. The generalization of this result in the context of $\mathcal{A}$-quasiconvexity was done in \cite{FM1}, \cite{rin1} and \cite{BCMS}.
In this work we derive a homogenization result in the context of $\cA$-quasiconvexity for integral functionals with linear growth.
This extends to the case $p=1$ the homogenization results derived in \cite{BFL}.
The linear growth condition implies that concentration effects may appear and they need to be treated by carefully applying homogenization techniques in the setting of weak-* convergence in measure.
\smallskip
\par In what follows let $\Omega\subset\R{N}$, $N\geq2$, be a bounded open set, and, for $d\geq1$,
let $f:\Omega{\times}\R{d}\to[0,+\infty)$ be a non-negative measurable function in the first variable and Lipschitz continuous in the second, that satisfies the following linear growth-coercivity condition: there exist $C_1,C_2>0$ such that
\be\label{200}
C_1|\zeta|\leq f(x,\zeta)\leq C_2(1+|\zeta|)\qquad\text{for all $(x,\zeta)\in\Omega{\times}\R{d}$}.
\ee
Moreover we assume that $x\mapsto f(x,\zeta)$ is $Q$-periodic for each $\zeta\in\R{d}$.
Here and in the following, $Q:=(-\frac12,\frac12)^N$ will denote the unit cube.
\par We will consider a linear first order partial differential operator $\cA:\cD'(\Omega;\R{d})\to\cD'(\Omega;\R{M})$ of the form
\be\label{operator}
\cA = \sum_{i=1}^N A^{(i)} \frac{\partial}{\partial x_i}, \qquad A^{(i)} \in \M{M\times d}, \; \; M \in \N{},
\ee
that we assume throughout to satisfy the following two conditions:
\begin{enumerate}
\item[$(H_1)$] Murat's condition of {\it constant rank} (introduced in \cite{Mur_84}; see also \cite{FM}) i.e., there exists $c\in \N{}$ such that
$
\mathrm{rank}\, \left(\sum_{i=1}^N A^{(i)} \xi_i\right)=c \quad\, \text{for all}\,\,\, \xi=(\xi_1,...,\xi_d)\in \cS^{d-1};
$
here and in the following $\cS^{d-1}$ denotes the unit sphere in $\R{d}$.
\item[$(H_2)$] $\mathrm{Span}(\cC)=\R{d}$ (where $\cC$ stands for the characteristic cone associated with the operator $\cA$; see Definition \eqref{356}).
\end{enumerate}
\par Let $\cM(\Omega;\R{d})$ denote the space of $\R{d}$-valued Radon measures defined on $\Omega$, and let $\mu\in\cM(\Omega;\R{d})\cap\ker\cA$.
The object of this work is to prove a representation theorem for the functional
$
\cF(\mu):=\inf\left\{\liminf_{n\to\infty} \int_\Omega f\left(\frac{x}{\eps_n},u_n(x)\right)\,\de x,\;\;\begin{array}{l}
\{u_n\}\subset\Lp1{}(\Omega;\R{d}),\;\;\cA u_n\stackrel{W^{-1,q}}{\longrightarrow}0, \\
u_n\wsto\mu,\,\, |u_n|\wsto\Lambda,\;\; \Lambda(\partial\Omega)=0
\end{array}\right\},
$
where $\cA$ is defined in \eqref{operator} and satisfies $(H_1)$ and $(H_2)$, and $q\in(1,\frac{N}{N-1})$.
\par Given a function $\phi:\R{d}\to\R{}$, we recall that its \emph{recession function} at infinity is defined by
\be\label{104}
\phi^\infty(b):=\limsup_{t\to+\infty}\frac{\phi(tb)}t,\qquad \text{for all $b\in\R{d}$.}
\ee
We also recall that for a measure $\mu\in\cM(\Omega;\R{d})$ we write $\mu=\mu^a+\mu^s$ for its Radon-Nikod\'ym decomposition, where $\mu^a$ is the absolutely continuous part with respect to the Lebesgue measure $\cL^N$, and $\mu^s$ is the singular part.
This means that there exists a density function $u^a\in L^1(\Omega;\R{d})$ such that $\mu^a=u^a\cL^N$.
In the following, we will denote by $\Lp1{Q-\per}(\R{N};\R{d})$ the space of $\R{d}$-valued $L^1$ functions defined on $Q$ and extended by $Q$-periodicity to the whole of $\R{N}$.
We now state the main result of this paper.
\begin{theorem}\label{main}
Let $\Omega\subset\R{N}$ be a bounded open set, and
let $f:\Omega{\times}\R{d}\to[0,+\infty)$ be a function which is measurable and $Q$-periodic in the first variable and Lipschitz continuous in the second, satisfying the growth condition \eqref{200}.
For any $b\in\R{d}$, define
\be\label{103}
f_{\cA-\hom}(b):=\inf_{R\in\N{}} \inf\left\{\ave_{RQ} f(x,b+w(x))\,\de x,\quad w\in\Lp1{RQ-\per}(\R{N};\R{d})\cap\ker\cA,\;\;\ave_{RQ} w=0\right\},
\ee
where $R\in\N{}$ and $f_{\cA-\hom}^\infty$ is the recession function of $f_{\cA-\hom}$ (see \eqref{104}).
For every $\mu\in\cM(\Omega;\R{d})\cap\ker\cA$, let $\mu=u^a\cL^N+\mu^s$ and let
$
\cF_{\cA-\hom}(\mu):=\int_\Omega f_{\cA-\hom}(u^a)\,\de x+\int_\Omega f_{\cA-\hom}^\infty\left(\frac{\de\mu^s}{\de|\mu^s|}\right)\,\de|\mu^s|.
$
Then, $\cF(\mu)=\cF_{\cA-\hom}(\mu)$.
\end{theorem}
\par The overall plan of this work in the ensuing sections is as
follows: in Section \ref{preliminaries} we set up the notation, concepts, and preliminary results that will be used throughout the paper.
Section \ref{mainproof} will be devoted to the proof of Theorem \ref{main}.
\section{Preliminaries}\label{preliminaries}
\subsection{Remarks on measure theory}
\par In this section we recall some notations and well known results in Measure Theory
(see, e.g., \cite{AmbrosioFuscoPallara00}, \cite{EG}, \cite{FonLeo}, and the
references therein).
\par Let $X$ be a locally compact metric space and let $C_{c}(X;\R{d})$, $d\geq 1,$
denote the set of continuous functions with compact support on
$X$. We denote by $C_{0}(X;\R{d})$ the completion of $C_{c}(X;\R{d})$ with
respect to the supremum norm. Let $\cB(X)$ be the Borel $\sigma$-algebra of $X.$ By
Riesz's Representation Theorem, the dual of the Banach space
$C_{0}(X;\R{d})$, denoted by $\cM (X; \R{d})$, is the space of finite $\R{d}$-valued Radon measures $\mu:
{\cB}(X)\to \R{d}$ under the pairing
$$<\mu,\varphi>:=\int_{X} \varphi \, d\mu\equiv \sum_{i=1}^{d} \int_{X}\varphi_i \, d\mu_i,\qquad \text{for every $\varphi\in C_0(X;\R{d})$}.$$
The space $\mathcal M (X; \R{d})$ will be endowed with the weak$^*$-topology deriving from this duality. In particular we say that
a sequence $\{\mu_n\}\subset \mathcal M (X; \R{d})$ converges weakly$^*$ to $\mu \in \mathcal M (X; \R{d})$ (indicated by $\mu_{n}
\overset{*}{\rightharpoonup} \mu$) if for all $\varphi\in
C_{0}(X;\R{d})$
$$\lim_{n\to \infty} \displaystyle\int_{X} \varphi\, d\mu_{n}=\displaystyle\int_{X} \varphi\, d\mu.$$
If $d=1$ we write by simplicity
$\mathcal M (X)$ and we denote by $\mathcal M^+ (X)$ its subset of positive measures.
Given $\mu\in \mathcal M (X; \R{d})$, let $|\mu|\in\cM^+(X)$ denote its {\it total
variation} and let $\supp\;\mu$ denote its {\it support}.
\begin{remark}\label{inclusions}
We recall that for an open set $\Omega\subset\R{N}$ we have $W_{0}^{1,p}(\Omega;\R{d}) \subset\subset C_{0}({\Omega};\R{d})$ for ${p}>N$.
Moreover, $\mathcal M (\Omega; \R{d})$ is compactly embedded in $W^{-1,q}(\Omega; \R{d})$,
$1<q<\frac{N}{N-1}$, where $W^{-1,q}(\Omega; \R{d})$ denotes the dual
space of $W_0^{1,q'}(\Omega; \R{d})$ with $q'$, the conjugate exponent of $q$, given by the relation
$\frac1{q} + \frac1{q'}=1$.
\end{remark}
\par We will need the following strong form of the Besicovitch's derivation theorem which is
due to Ambrosio and Dal Maso \cite{ADMaso_92}
(see also \cite[Theorem~2.22 and Theorem~5.52]{AmbrosioFuscoPallara00}
or \cite[Theorem~1.155]{FonLeo}).
\begin {theorem} \label{general}
Let $\mu\in \mathcal M^+(\Omega)$ and $\nu\in \mathcal M (\Omega; \R{d})$. Then
there exists a Borel set $N \subset \Omega$ with $\mu(N)=0$ such that
for every $x \in (\supp \mu) \backslash N$
$$\frac{d\nu}{d\mu}(x)
=\frac{d\nu^{a}}{d\mu}(x)
= \lim_{\epsilon \rightarrow 0}
\frac{\nu \big( D(x,\eps) \cap \Omega\big)}{\mu \big( D(x,\eps) \cap \Omega\big)}\in \R{}$$
and
$$\frac{d\nu^{s}}{d\mu}(x)
= \lim_{\epsilon \rightarrow 0}
\frac{\nu^s \big( D(x,\eps) \cap \Omega\big)}{\mu \big( D(x,\eps) \cap \Omega\big)}
=0,$$
where $D\subset\R{N}$ is any bounded, convex, open set containing the origin and $D(x,\eps):=x+\eps D$
(the exceptional set $N$ is independent of the choice of $D$).
\end{theorem}
\par The definition of {\it tangent measures} was originally introduced by Preiss \cite{Preiss} and is relevant for studying
the local behaviour (or blow-up) of a given
measure. Here we give an adaptation of this notion given by Rindler \cite{rin1} (see also \cite{AmbrosioFuscoPallara00}).
\begin{defin}[Tangent measures on convex sets] \label{tan}
Let $D\subset \R{N}$ be a bounded, convex, open set containing the origin.
Given $\mu\in \cM (\Omega; \R{d})$ and $x_0\in \Omega$ we define the measure $T_{*}^{(x_0,\delta)}\mu\in \cM (\overline{D}; \R{d})$, for any $\delta >0$, by
$$<T_{*}^{(x_0,\delta)}\mu,\varphi>= \int_{{D(x_0,\delta)}} \varphi \left(\frac{x-x_0}{\delta}\right)\, d\mu, \quad
\text{for any $\varphi\in C(\overline{D}; \R{d})$}.$$
The tangent space of $\mu$ at $x_0$, $\Tan_{D}(\mu,x_0)$, consists of all the $\R{d}$-values measures
$\nu \in \cM (\overline{D}; \R{d})$ which are the weak$^*$-limit of a rescaled sequence of measures
of the type
$$\frac{T_{*}^{(x_0,\delta_n)}\mu}{|\mu|(D(x_0,\delta_n))}$$
for some infinitesimal sequence $\{\delta_n\}_{n\in \N{}}$.
\end{defin}
\noindent In the conditions of Definition \ref{tan}, from an adaptation of Theorem 2.44 in \cite{AmbrosioFuscoPallara00} to convex sets, it follows that
$
\Tan_{D}(\mu,x_0)=\frac{d\mu}{d|\mu|}(x_0)\cdot \Tan_{D}(|\mu|,x_0),
$
whenever $\frac{d\mu}{d|\mu|}(x_0)$ exists and it is finite.
As noted in \cite{BCMS}, the following result holds
\begin{lemma}\label{lemmarin}
Let $\mu\in {\cM} (\Omega; \R{d})$. Then there exists $E\subset \Omega$
with $|\mu|(\Omega\setminus E)=0$ such that for all $x_0\in E$, given $D\subset \R{N}$ an open convex set containing the origin,
there exist $\tau\in \Tan_{D}(\mu,x_0)$ with $|\tau|(D)=1$ and $|\tau|(\partial D)=0.$
\end{lemma}
\begin{remark}\label{suc-front}
\begin{itemize}
\item[(i)] The measure $\tau$ in Lemma \ref{lemmarin} is defined in an open set $U$ containing $\overline{D}$, which will
be useful for regularization purposes.
Moreover we have that ${\cal A}\tau=0$ in $U$ whenever ${\cal A}\mu=0$ in $\Omega$.
\item[(ii)] Given $\Lambda \in \cM^{+}(\overline{D})$ it is possible to find a sequence $\delta_n$ with
$\Lambda (\partial D(x_0,\delta_n))=0$ such that
$
\frac{T_{*}^{(x_0,\delta_n)}\mu}{|\mu|(D(x_0,\delta_n))}\wsto \tau.
$
\end{itemize}
\end{remark}
We refer the reader to the proof of Lemma 3.1 in \cite{rin1}, from which Lemma \ref{lemmarin} and Remark \ref{suc-front} are derived.
\subsection{Remarks on $\cA$-quasiconvexity}
\par We start by recalling the notion and some properties of $\cA$-quasiconvex functions. This notion was introduced by Dacorogna \cite{Dac} following the works of Murat and Tartar
in compensated compacteness (see \cite{Mur_84} and \cite {T}) and was further
devoloped by Fonseca and M\"{u}ller \cite{FM} (see also Braides, Fonseca, and Leoni \cite{BFL}).
\par Let $\cA: \cD'(\Omega;\R{d}) \to \cD'(\Omega;\R{M})$ be the first order linear differential operator defined in \eqref{operator}.
\begin{defin}[$\cA$-quasiconvex function]\label{aquasi}
A locally bounded Borel function $f:\R{d}\to \R{}$ is said to be $\cA$-quasiconvex if
\[
f(v)\leq \int_Q f(v+w(x))\, \de x
\]
for all $v\in \R{d}$ and for all $w\in C_{Q-\per}^\infty(\R{N};\R{d})$
such that $\cA w=0$ in $\R{M}$ with $\int_Q w(x)\, \de x=0$.
\end{defin}
\begin{defin}[Characteristic cone of $\cA$]\label{355}
The characteristic cone of $\cA$ is defined by
\be\label{356}
\cC:=\left\{v\in\R{d}:\exists w\in\R{N}\setminus\{0\}, \left(\sum_{i=1}^N A^{(i)}w_i\right)v=0\right\}.
\ee
\end{defin}
Given $v\in\R{d}$, we define the linear subspace of $\R{N}$
$
\cV_v:=\left\{w\in\R{N}:\left(\sum_{i=1}^N A^{(i)}w_i\right)v=0\right\}.
$
If $v\notin\cC$, then $\cV_v=\{0\}$, otherwise $\cV_v$ is a non trivial subspace of $\R{N}$.
\par The following general result can be found in \cite[Lemma~2.14]{FM}.
We will use it in the sequel applied to smooth functions.
\begin{proposition}\label{projection}
Given $q>1$ there exists a bounded linear operator $\cP: L_{Q-\per}^q (\R{N};\R{d}) \to L_{Q-\per}^q (\R{N};\R{d})$ such that $\cA(\cP u)=0$.
Moreover, $\int_Q \cP u=0$ and there exists a constant $C>0$ such that
$$\left|\left| u- \cP u \right|\right|_{L^q} \leq C \|\cA u\|_{W^{-1,q}}$$
for every $u \in L_{Q-\per}^q (\R{N};\R{d})$ with $\int_Q u=0.$
\end{proposition}
\subsection{Other remarks}
Given a continuous function $f \in C(\Omega \times \R{d})$ we start by defining the linear mapping
$T:C(\Omega \times \R{d}) \to C(\Omega \times B)$ by
\begin{equation}\label{T}
Tf(x,\xi)=(1-|\xi|)f\left(x,\frac{\xi}{1-|\xi|}\right)
\end{equation}
and we introduce the set
$$E(\Omega \times \R d):=\left\{f \in C(\Omega \times \R d): Tf \text{ has a continuous extension to}
\,\, C(\overline{\Omega \times B}) \right\}.$$
For $f\in E(\Omega \times \R d)$, $f^{\infty}$ represents the continuous extension of $Tf$, precisely,
\begin{equation*
f^{\infty}(x,\xi):= \lim_{\tiny \begin{array}{c}
x^{'}\to x\\
\xi^{'}\to \xi\\
t \to \infty
\end{array}
} \frac{f(x',t\xi')}{t}
\end{equation*}
\begin{remark}
Note that definition \eqref{104} is a generalization of this definition of recession function for $\Phi: \R{d} \to \R{}$ with linear growth.
\end{remark}
\begin{theorem}[Reshetnyak's Continuity Theorem {\cite[Theorem~5]{KriRin_CV_10}}]\label{Reshetnyak}
Let $f\in E(\Omega;\R{d})$ and let $\mu,\,\mu_n\in \cM (\Omega; \R{d})$ be such that
$\mu_n
\wsto \mu$ in $\cM (\Omega; \R{d})$ and $\langle\mu_n\rangle(\Omega)
\to \langle\mu\rangle (\Omega),$ where
$$\langle\nu\rangle
:=\sqrt{1+|\nu^a|^2}\cL^N+|\nu^s|, \quad \nu=\nu^a \cL^N +\nu^s\in \cM (\Omega; \R{d}).$$
Then
$$\lim_{n\to \infty}\widehat{\cF}(\mu_n)= \widehat{\cF}(\mu)$$
where
\begin{equation}\label{Fhat}
\widehat\cF(\nu):= \int_\Omega f(x,\nu^a(x))\, dx
+\int_\Omega f^\infty\left(x,
\frac{d\nu^s}{d|\nu^s|}(x)\right)
\, d|\nu^s|,\quad \nu\in \cM (\Omega; \R{d}).
\end{equation}
\end{theorem}
The following upper semicontinuity result which is a consequence of Theorem \ref{Reshetnyak} will be useful in the proof of our main result Theorem \ref{main}. For a proof see \cite[Corollary 2.11]{BCMS}.
\begin{proposition} \label{usc}
Let $f: \Omega \times \R{d} \to \R{}$ be a continuous function such that there exists $C>0$ with
$$|f(x,\xi)|\leq C(1+|\xi|),\,\, \text{ for all $(x,\xi) \in \Omega \times \R{d}$.}$$
Let $\mu,\,\mu_n\in \mathcal M (\Omega; \R{d})$ be such that
$\mu_n \wsto \mu$ in $\mathcal M (\Omega; \R{d})$ and $\langle\mu_n\rangle(\Omega) \to \langle\mu\rangle (\Omega).$
Then
\begin{equation*
\widehat\cF(\mu)\geq \limsup_{n\to \infty} \widehat\cF(\mu_n)
\end{equation*}
where $\widehat\cF$ was defined in \eqref{Fhat}.
\end{proposition}
\subsection{Preliminary results}
Fix some sequence $\{ \eps_n \}$ and let $\cO(\Omega)$ denote the collection of all open subsets of $\Omega$.
For $A\in\cO(\Omega)$, the localized version of $\cF$, denoted by $\cF_{\{\eps_n\}}(\mu; A)$, is defined by
\be\label{202}
\cF_{\{\eps_n\}} (\mu;A) := \inf \left\{ \liminf_{n\to\infty} \int_A f\left(\frac{x}{\eps_n},u_n(x)\right)\,\de x,\begin{array}{l}
\{u_n\}\subset\Lp1{}(\Omega;\R{d}),\;\;\cA u_n\stackrel{W^{-1,q}}{\longrightarrow}0, \\
u_n\wsto\mu,\,\, |u_n|\wsto\Lambda,\;\; \Lambda(\partial\Omega)=0
\end{array}\right\}.
\ee
Note that using the regularization of $\mu$ as test sequence, that is $u_n = \rho_{\eps_n} * \mu$, by \eqref{200} we obtain the upper bound
$
\mathcal{F}_{\{\eps_n\}} (\mu;A) \leq C \left(|A|+ |\mu|(\overline{A})\right).
$
We begin by proving the following subadditivity result:
\begin{proposition}[Subadditivity for nested sets]\label{sub}
For any $A\subset\subset B\subset\subset C\subset\Omega$ there holds
\be\label{210}
\cF_{\{ \eps_n^1 \}}(\mu;C)\leq \cF_{\{ \eps_n^1 \}}(\mu;B)+\cF_{\{ \eps_n^1 \}}(\mu;C\setminus\cl A).
\ee
where $\{\eps_n^1\}\subset\{\eps_n\}$ is an appropriate subsequence.
\end{proposition}
\begin{proof}
Fix $\delta>0$. Consider a subsequence $\{\eps_n^1\}\subset\{\eps_n\}$ and an associated sequence $\{v_n^1\}\subset\Lp1{}(\Omega;\R{d})$ such that $v_n^1 \wsto \mu$, $\cA v_n^1 \to 0$ and
\be\label{211}
\cF_{\{ \eps_n \}}(u;B) + \delta \geq \lim_{n\to+\infty} \int_B f\left(\frac{x}{\eps_n^1},v_n^1\right)\de x,
\ee
By extracting another subsequence $\{\eps_n^2\}\subset\{\eps_n^1\}$ there exists a sequence $\{w_n^2\}\subset\Lp1{}(\Omega;\R{d})$ such that $w_n^2 \wsto \mu$, $\cA w_n^2 \to 0$ and
\be\label{211a}
\cF_{\{ \eps_n^1 \}}(u;C \setminus \cl A) + \delta \geq \lim_{n\to+\infty} \int_{C \setminus \cl A} f\left(\frac{x}{\eps_n^2},w_n^2\right)\de x.
\ee
We denote by $\{ v_n^2 \}$ the subsequence of $\{ v_n^1 \}$ corresponding to $\{ \eps_n^2 \}$.
Consider smooth cut-off functions $\phi_j$ such that, for all $j\in\N{}$, $\phi_j\equiv1$ on $A$, $\phi_j\equiv0$ on $C\setminus\cl B$, and such that
\be\label{212}
\lim_{j \to +\infty} \limsup_{n \to +\infty} \int_{ \{ 0 < \phi_j < 1 \}} \left(1+ |v_n^2|+|w_n^2| \right) \de x=0.
\ee
Property \eqref{212} ensures that $|\{0<\phi_j<1\}|\to0$ as $j\to\infty$ and that $\{v_n^2\}$ and $\{w_n^2\}$ do not have concentrations in the transition layer of $\phi_j$.
Define $ U_{n,j}:=(1-\phi_j)w_n^2+\phi_j v_n^2$. It is not difficult to show that $\{ U_{n,j}\}\subset L^1(\Omega;\R{d})$, and that $U_{n,j} \wsto \mu$, $\cA U_{n,j}\to0$ in $W^{-1,q}(\Omega;\R{M})$ as $n \to \infty$.
Let $A_j:=\{x\in\Omega:\phi_j(x)\equiv1\}$ and $B_j:=\{x\in\Omega:\phi_j(x)\equiv0\}$ and notice that $A\subset A_j\subset B$ and $C\setminus\cl B\subset B_j\subset C\setminus\cl A$.
We have
\begin{equation*
\begin{split}
\cF_{\{ \eps_n^1\}}(u;C) \leq & \liminf_{j \to +\infty} \liminf_{n\to+\infty} \int_C f\left(\frac{x}{\eps_n^2},U_{n,j}\right)\de x \\
\leq & \limsup_{j \to +\infty} \limsup_{n\to+\infty} \int_{A_j} f\left(\frac{x}{\eps_n^2},v_n^2\right)\de x + \limsup_{j \to +\infty} \limsup_{n\to+\infty} \int_{B_j} f\left(\frac{x}{\eps_n^2},w_n^2\right)\de x \\
& +\limsup_{j \to +\infty} \limsup_{n\to+\infty}\int_{\{0<\phi_j<1\}} f\left(\frac{x}{\eps_n^2},U_{n,j} \right)\de x\\
\leq & \lim_{n\to+\infty} \int_{B} f\left(\frac{x}{\eps_n^2},v_n^2\right)\de x + \lim_{n\to+\infty} \int_{C\setminus\cl A} f\left(\frac{x}{\eps_n^2},w_n^2\right)\de x \\
& +\limsup_{j \to +\infty} \limsup_{n\to+\infty} \int_{\{0<\phi_j<1\}} C(1+|U_{n,j}|)\de x\\
\leq & \cF_{\{ \eps_n^1 \}}(u;B)+\cF_{\{ \eps_n^1 \}}(u;C\setminus\cl A) + 2 \delta
\end{split}
\end{equation*
where we used that $f\geq0$, \eqref{211}, \eqref{211a}, and \eqref{212}.
Letting $\delta \to 0$ yields \eqref{210}.
\end{proof}
\begin{proposition}\label{measure}
Given $\mu \in \cM(\Omega;\R{d})$ and a sequence $\{\eps_n\}$, we can find a subsequence $\{ \eps_n^1 \} \subset \{ \eps_n \}$, a sequence $u_n^1 \wsto \mu$, $\cA u_n^1 \to 0$, and a bounded Radon measure $\Phi_{\mu}$ such that
\be\label{2061}
f \left( \frac {x}{\eps_n^1}, u_n^1(x) \right)\de x \wsto \Phi_\mu.
\ee
Moreover,
for every open set $A \subset \subset \Omega$, we can find a further subsequence $\{\eps_n^2\} \subset \{ \eps_n^1 \}$ such that
\be\label{206}
\cF_{\{\eps_n^2\}}(\mu;A)=\Phi_\mu(\overline A).
\ee
\end{proposition}
\begin{proof}
Fix $\mu\in\cM(\Omega;\R{d})$ and $\delta>0$. Then there exists a subsequence $\{\eps_n^1\}\subset\{\eps_n\}$ and a sequence $\{u_n^1\}\subset L^1(\Omega;\R{d})$ such that
\be\label{205}
\cF_{\{\eps_n\}}(\mu) + \delta \geq \lim \int_\Omega f\left(\frac{x}{\eps_n^1},u_n^1(x)\right)\,\de x
\ee
and therefore by \eqref{200} we have that
$
f\left(\frac{x}{\eps_n^1},u_n^1(x)\right)\,\de x\wsto\Phi_\mu,
$
for some bounded Radon measure $\Phi_\mu$.
We now claim that we can extract another subsequence $\{\eps_n^2\}\subset\{\eps_n^1\}$ for which \eqref{206} holds.
We start by noting that
\be\label{207}
\cF_{\{\eps_n^2\}}(\mu;A)\leq \liminf_{n\to\infty} \int_A f\left(\frac{x}{\eps_n^2},u_n^2(x)\right)\,\de x\leq \limsup_{n\to\infty} \int_{\cl A} f\left(\frac{x}{\eps_n^1},u_n^1(x)\right)\,\de x\leq \Phi_\mu(\cl A),
\ee
for every subsequence $\{(\eps_n^2,u_n^2)\}\subset\{(\eps_n^1,u_n^1)\}$, where the last inequality follows from known facts about measure theory (see, e.g., \cite{FonLeo}).
To prove the opposite inequality, fix $\delta>0$ and choose an open set $B\subset A$ such that $\Phi_\mu(\partial B)=0$ and $\Phi_\mu(\cl A\setminus B)<\delta$.
Now,
\be\label{208}
\begin{split}
\Phi_\mu(\cl A) & \leq \Phi_\mu(B)+\delta \\
& = \Phi_\mu(\Omega)-\Phi_\mu(\Omega\setminus\cl B) +\delta\\
& \leq \cF_{\{\eps_n^2\}}(\mu) + \delta -\cF_{\{\eps_n^2\}}(\mu;\Omega\setminus\cl B)+\delta,
\end{split}
\ee
where the last line follows by applying \eqref{207} to the closure of $\Omega\setminus\cl B$, and, by \eqref{205} and \eqref{210} the sequence $\{\eps_n^2\}$ is chosen in such a way that
\be\label{209}
\cF_{\{\eps_n^2\}}(\mu)\leq \cF_{\{\eps_n^2\}}(\mu;A)+\cF_{\{\eps_n^2\}}(\mu;\Omega\setminus\cl B).
\ee
By plugging \eqref{209} in \eqref{208}, and letting $\delta\to0$, we conclude the proof of \eqref{206}.
\end{proof}
\par In the proof of the main result we will use the following properties of $f_{\cA-\hom}$.
Recall its definition in \eqref{103}.
\begin{proposition}\label{Lipschitz}
Let $f:\Omega\times\R{d}\to\R{}$ be Lipschitz continuous in the second variable with Lipschitz constant $L>0$.
Then the function $f_{\cA-\hom}:\R{d}\to\R{}$ defined in \eqref{103} is Lipschitz continuous with the same constant.
\end{proposition}
\begin{proof}
Let $b_1,b_2\in\R{d}$ and let $\eps>0$.
From the definition of $f_{\cA-\hom}$, there exist $k_1\in\N{}$ and $w_1\in L^1_{kQ-\per}(\R{N};\R{d})\cap\ker\cA$ such that $\ave_{k_1Q} w_1=0$ and
$$\ave_{k_1Q} f(y,b_1+w_1(y))\,\de y\leq f_{\cA-\hom}(b_1)+\eps.$$
Again by the definition of $f_{\cA-\hom}$ and by the previous inequality we have
\begin{equation*}
\begin{split}
f_{\cA-\hom}(b_2)-f_{\cA-\hom}(b_1) \leq & \ave_{k_1Q} f(y,b_2+w_1(y))\,\de y-\ave_{k_1Q} f(y,b_1+w_1(y))+\eps \\
\leq & \ave_{k_1Q} |f(y,b_2+w_1(y))-f(y,b_1+w_1(y))|\,\de y +\eps \\
\leq & L|b_2-b_1|+\eps.
\end{split}
\end{equation*}
The result follows by exchanging the roles of $b_1$ and $b_2$ and by letting $\eps\to0$.
\end{proof}
\begin{proposition}[Invariance under translations]\label{translations}
Let $b \in \R{d}$ and $\gamma\in Q$. Define
\be\label{222}
f_{\cA-\hom}^\gamma(b):=\inf_{R\in\N{}}\inf\left\{\ave_{RQ} f(y+\gamma,b+w(y))\de y,\;\; w\in L^1_{RQ-\per}(\R{N};\R{d})\cap\ker\cA,\ave_{RQ} w=0\right\}.
\ee
Then we have that $f_{\cA-\hom}(b) = f_{\cA-\hom}^\gamma(b)$.
\end{proposition}
\begin{proof}
The proof is a simple computation.
Let $w\in L^1_{RQ-\per}(\R{N};\R{d})\cap\ker\cA$ with $\ave_{RQ} w=0$.
Then,
\be\label{223}
\begin{split}
\ave_{RQ} f(y+\gamma,b+w(y))\,\de y = & \ave_{RQ+\gamma} f(x,b+w(x-\gamma))\de x \\
= & \ave_{RQ} f(x,b+w(x-\gamma))\de x \\
& + \ave_{(RQ+\gamma)\setminus RQ} f(x,b+w(x-\gamma))\de x \\
& -\ave_{RQ\setminus(RQ+\gamma)} f(x,b+w(x-\gamma))\de x \\
=& \ave_{RQ} f(x,b+w_\gamma(x))\de x \\
\geq & f_{\cA-\hom}(b),
\end{split}
\ee
where we defined $w_\gamma(x):=w(x-\gamma)$.
We notice that $w_\gamma$ is $RQ$-periodic, $\ave_{RQ} w_\gamma=\ave_{RQ} w=0$, and $\cA w_\gamma=0$, which means that $w_\gamma$ is a competitor for \eqref{103}.
The third equality in \eqref{223} holds because the second and third integrals cancel out.
Indeed, by using the $RQ$-periodicity of $w$ and noticing that $f$ is also $RQ$-periodic in the first variable, by means of shifts by $R$ along the coordinate directions one can easily proof that cancellation occurs.
By taking the infimum over all $w\in L^1_{RQ-\per}(\R{N};\R{d})\cap\ker\cA$ with $\ave_{RQ} w=0$, and the infimum over $R\in\N{}$, we obtain
$
f_{\cA-\hom}^\gamma(b)\geq f_{\cA-\hom}(b).
$
The reverse inequality is obtained in the same way.
\end{proof}
The following result will be used in Section \ref{LBsp}.
It is an adaptation \cite[Lemma 2.20]{BCMS}.
\begin{lemma}\label{375}
Let $f_n:\Omega\times\R{d}\to\R{}$ be a family of Lipschitz continuous functions in the second variable with the same Lipschitz constant $L$ and let $\nu\in\cS^{N-1}$.
Let $\{u_n\},\{v_n\}\subset L^q(Q_\nu;\R{d})$ be sequences such that $u_n-v_n\wsto0$ in $\cM(Q_\nu;\R{d})$ and $|u_n|+|v_n|\wsto\Lambda$ in $\cM^+(\cl Q_\nu)$, with $\Lambda(\partial Q_\nu)=0$, and $\cA(u_n-v_n)\to0$ in $W^{-1,q}(Q_\nu;\R{M})$ for some $1<q<\frac{N}{N-1}$.
Then there exists a sequence $\{z_n\}\subset L_{Q_\nu-\per}^q(\R{N};\R{d})$ such that $\int_{Q_\nu} z_n=0$, $\cA z_n=0$, $z_n\wsto0$, and
\be\label{376}
\liminf_{n\to\infty} \int_{Q_\nu} f_n(x,u_n(x))\,\de x\geq \liminf_{n\to\infty} \int_{Q_\nu} f_n(x,z_n(x)+v_n(x))\,\de x.
\ee
\end{lemma}
\begin{proof}
The proof is the same as the proof of \cite[Lemma 2.20]{BCMS}.
It is achieved by proving \eqref{376} for the unit cube $Q$ and then for the rotated cube $Q_\nu$.
The proof relies on some Lipschitz estimates which involve only the second variable of $f$.
\end{proof}
\section{Main result}\label{mainproof}
\par In this section we prove Theorem \ref{main}.
We shall split the proof in several steps.
\subsection{Upper bound}
In this section we prove the upper bound inequality in Theorem \ref{main}.
We start with the estimate for regular measures $\mu = u\cL^N$, i.e., we prove that:
\be\label{214}
\frac{\de\Phi_u}{\de\cL^N}(x_0)\leq f_{\cA-\hom}(u(x_0)),\qquad\text{for $\cL^N$-a.e.\@ $x_0\in\Omega$}.
\ee
Let $x_0$ be a Lebesgue point for the function $u$.
Given a vanishing sequence of radii $\{r_j\}$ such that $\Phi_u(\partial Q(x_0,r_j))=0$ for all $j\in\N{}$, there exists a sequence $\{\eps_n^j\}$ such that
\be\label{218}
\Phi_u(Q(x_0,r_j))=\cF_{\{\eps_n^j\}}(u;Q(x_0,r_j))\qquad\text{for all $j\in\N{}$.}
\ee
We will contruct an appropriate sequence $\{u_n^j\}\subset L^1(Q(x_0,r_j);\R{d})$ such that
$
\left.
\begin{array}{l}
u_n^j\wsto u \\
\cA u_n^j=0
\end{array}
\right\}\qquad \text{as $n\to\infty$, for all $j\in\N{}$.}
$
To this end, consider $k\in\N{}$ and a function $w\in L^1_{kQ-\per}(\R{N};\R{d})\cap\ker\cA$, such that $\ave_{kQ} w=0$, and let $w_n^j(x):=w(h_nk(x-x_0)/r_j)$, with $h_n\in\N{}$ to be determined.
It is not hard to prove that
$
\cA w_n^j=0\;\text{for all $n,j\in\N{}$}\qquad\text{and that}\;\ave_{Q(x_0,r_j)} w_n^j=\ave_{kQ} w=0.
$
Moreover $u_n^j(x):=u(x)+w_n^j(x)$ satisfies
$
u_n^j\wsto u,\qquad \cA u_n^j=\cA u=0\; \text{for all $n,j\in\N{}$.}
$
We then have
\begin{equation*
\begin{split}
\frac{\de\Phi_u}{\de\cL^N}(x_0) = & \lim_{j\to\infty}\frac{\Phi_u(Q(x_0,r_j))}{r_j^N} \stackrel{\eqref{218}}{=} \lim_{j\to\infty} \frac{\cF_{\{\eps_n^j\}}(u;Q(x_0,r_j))}{r_j^N}\\
\leq & \limsup_{j\to\infty}\liminf_{n\to\infty}\frac1{r_j^N}\int_{Q(x_0,r_j)} f\left(\frac{x}{\eps_n^j},u_n^j(x)\right)\de x \\
= & \limsup_{j\to\infty}\liminf_{n\to\infty}\frac1{r_j^N}\int_{Q(x_0,r_j)} f\left(\frac{x}{\eps_n^j},u(x)+w\left(h_nk\frac{x-x_0}{r_j}\right)\right)\de x \\
\leq & \limsup_{j\to\infty}\left[ \limsup_{n\to\infty} \frac1{r_j^N}\int_{Q(x_0,r_j)} f\left(\frac{x}{\eps_n^j},u(x_0)+w\left(h_nk\frac{x-x_0}{r_j}\right)\right)\de x + A_j \right]\\
= & \limsup_{j\to\infty}\limsup_{n\to\infty}\frac1{(h_nk)^N}\int_{h_nkQ} f\left(\frac{r_jy}{h_nk\eps_n^j}+\frac{x_0}{\eps_n^j},u(x_0)+w(y)\right)\de y
\end{split}
\end{equation*
where in the fourth line we have used the Lipschitz continuity with respect to the second variable and the fact that $u$ has a Lebesgue point at $x_0$ to obtain that $A_j\to0$, and in the fifth line we changed variables $y:=h_nk\frac{x-x_0}{r_j}$.
We now choose $h_n:=\left[\frac{r_j}{k\eps_n^j}\right]$, and write
$
\frac{r_jy}{h_nk\eps_n^j}=y+\frac1{h_n}\left\langle\frac{r_j}{k\eps_n^j}\right\rangle y;
$
also note that, since $\left\langle\frac{r_j}{k\eps_n^j}\right\rangle\in Q$, then
\be\label{2201}
\lim_{n\to\infty} \frac1{h_n}\left\langle\frac{r_j}{k\eps_n^j}\right\rangle=0.
\ee
Let $\gamma_n^j:=\left\langle\frac{x_0}{\eps_n^j}\right\rangle$.
By using the $Q$-periodicity of $f$ in the first variable, the $kQ$-periodicity of $w$, and \eqref{2201}, we then have that
\begin{equation*
\begin{split}
\frac{\de\Phi_u}{\de\cL^N}(x_0)\leq &
\limsup_{j\to\infty}\limsup_{n\to\infty}\frac1{(h_nk)^N}\int_{h_nkQ} f\left(y+\frac1{h_n}\left\langle\frac{r_j}{k\eps_n^j}\right\rangle y+\gamma_n^j,u(x_0)+w(y)\right)\de y \\
= & \limsup_{j\to\infty}\limsup_{n\to\infty}\frac1{k^N}\int_{kQ} f\left(y+\frac1{h_n}\left\langle\frac{r_j}{k\eps_n^j}\right\rangle y+\gamma_n^j,u(x_0)+w(y)\right)\de y \\
= & \limsup_{j\to\infty}\frac1{k^N}\int_{kQ} f(y+\gamma_{n_j}^j,u(x_0)+w(y))\de y \\
\leq & \limsup_{j\to\infty}\left(f_{\cA-\hom}^{\gamma_{n_j}^j}(u(x_0))+\frac1j\right)= f_{\cA-\hom}(u(x_0))
\end{split}
\end{equation*
where in the last line we used the definition of $\inf$ and Proposition \ref{translations}.
This proves \eqref{214}.
\par Let now $\mu\in\cM$ be a general measure.
Then by defining $u_n:=\rho_n*\mu$ we construct a sequence of smooth functions such that
$
u_n\wsto\mu,\qquad\qquad\bracket{u_n}(\Omega)\to\bracket\mu(\Omega).
$
By Proposition \ref{usc} we the have that
\begin{equation*
\begin{split}
\cF_{\{\eps_n\}}(\mu)\leq & \liminf_{n\to+\infty} \cF_{\{\eps_n\}}(u_n)\leq\liminf_{n\to+\infty} \int_\Omega f_{\cA-\hom}(u_n(x))\,\de x \\
\leq & \limsup_{n\to+\infty} \int_\Omega f_{\cA-\hom}(u_n(x))\,\de x\leq \limsup_{n\to\infty} \cF_{\cA-\hom}(u_n) \leq \cF_{\cA-\hom}(\mu).
\end{split}
\end{equation*
This proves that
$
\cF(\mu)\leq\cF_{\cA-\hom}(\mu).
$
\subsection{Lower bound -- Absolutely continuous part}
Let $\mu\in\cM(\Omega;\R{d})$, let $\{\eps_n\}$ be a vanishing sequence and let $\Phi_\mu$ be the measure provided by the first statement in Proposition \ref{measure}, with the respective
${u_n}\subset L^1(\Omega;\R{d})$ (not relabeled) such that $u_n\wsto\mu$, and $\cA u_n\to0$ in $W^{-1,q}$.
We now prove the inequality
\be\label{311}
\frac{\de\Phi_\mu}{\de\cL^N}(x_0)\geq f_{\cA-\hom}(u(x_0)),
\ee
where we decomposed $\mu=\mu^a+\mu^s=u\cL^N+\mu^s$, and $x_0$ is a Lebesgue point for $\mu$. In particular, $\mu^s(\{x_0\})=0$.
If the Radon-Nikod\'ym derivative in the left-hand side of \eqref{311} is $\infty$ there is nothing to prove, so we shall assume that it is bounded.
By \eqref{2061} and Theorem \ref{general} we have that
$
\frac{\de\Phi_{\mu}}{\de\cL^N}(x_0)= \lim_{j\to\infty}\lim_{n\to\infty}\frac1{r_j^N}\int_{Q(x_0;r_j)} f\left(\frac{x}{\eps_n},u_{n}(x)\right)dx.
$
For any $x\in Q$, define $w_{n,j}(x):=u_{n}(r_jx+x_0)-u(x_0)$.
Then we have $w_{n,j}\wsto0$ since $x_0$ is a Lebesgue point, $w_{n,j}\in L^1$, $\cA w_{n,j}\to0$ in $W^{-1,q}$.
Notice that it is not restrictive to choose the sequence $\{r_j\}$ in such a way that
\be\label{321}
|w_{n,j}|\wsto\lambda\qquad\text{and} \qquad \lambda(\partial Q)=0.
\ee
So,
\begin{equation*
\begin{split}
\frac{\de\Phi_{\mu}}{\de\cL^N}(x_0)= & \lim_{j\to\infty}\lim_{n\to\infty}\frac1{r_j^N}\int_{Q(x_0;r_j)} f\left(\frac{x}{\eps_n},u(x_0)+w_{n,j}\left(\frac{x-x_0}{r_j}\right)\right)\,\de x,\\
=& \lim_{j\to\infty}\lim_{n\to\infty}\int_{Q} f\left(\frac{r_jy+x_0}{\eps_n},u(x_0)+w_{n,j}(y)\right)\,\de y,
\end{split}
\end{equation*
where we have changed variables $y:=(x-x_0)/r_j$.
Now we can diagonalize to obtain a sequence $\hat w_k\in L^1$, such that $\hat w_k\wsto0$, $\cA\hat w_k\to0$ in $W^{-1,q}$.
So, by invoking the $Q$-periodicity of $f$ in the first variable,
$
\frac{\de\Phi_{\mu}}{\de\cL^N}(x_0)= \lim_{k\to\infty}\int_{Q} f(s_ky+\gamma_{k},u(x_0)+\hat w_k(y))\,\de y,
$
with $s_k:=r_{j(k)}/\eps_{n(k)}$ in such a way that $\lim_{k\to\infty} s_k=\infty$, and $\gamma_{k}:=\langle x_0/\eps_{n(k)}\rangle$.
We claim we can find a sequence $w_k$ such that
\be\label{315}
\lim_{k\to\infty} \int_Q f( s_ky + \gamma_{k}, u(x_0) + \hat{w}_k(y)) \, \de y \geq \lim_{k\to\infty} \int_Q f(s_ky + \gamma_{k}, u(x_0) + w_k(y))\, \de y,
\ee
and $ w_k\wsto0, \; \cA w_k = 0, \; \ave_Q w_k = 0, \; w_k \in L^1_{Q-{\per}}(\R{N};\R{d}).$
To this end, consider a cut-off function $\theta_i$ such that $\theta_i\equiv1$ in $Q_i:= (1-1/i)Q$ and $\theta_i\equiv0$ in $\R{N}\setminus\cl Q$.
Then, by using Lispschitz continuity of $f$ we obtai
\be\label{316}
\begin{split}
\frac{\de\Phi_{\mu}}{\de\cL^N}(x_0)=&\lim_{k\to\infty} \int_Q f( s_ky + \gamma_{k}, u(x_0) + \hat{w}_k(y)) \,\de y \\
=&\lim_{k\to\infty} \int_Q f( s_ky + \gamma_{k}, u(x_0) + \theta_i(y)\hat{w}_k(y)+(1-\theta_i)\hat w_k(y)) \,\de y\\
\geq &\liminf_{i\to\infty}\liminf_{k\to\infty} \left[ \int_Q f( s_ky + \gamma_{k}, u(x_0) + \theta_i(y)\hat{w}_k(y)) \,\de y - I_{i,k} \right],
\end{split}
\ee
where
$
I_{i,k}:= L \int_{Q\setminus Q_i} ( 1 - \theta_i)|\hat{w}_k(y)| \,\de y.
$
Notice now that $\lim_{i\to\infty}\lim_{k\to\infty} I_{i,k}=0$.
Indeed, since $\hat w_k$ comes from the sequence $w_{n,j}$ by \eqref{321} we have that $|\hat w_k|\wsto\lambda$.
By definition of weak convergence, and again by \eqref{321}, we have
\be\label{318}
\lim_{i\to\infty}\lim_{k\to\infty} I_{i,k}\leq\lim_{i\to\infty} L\lambda\big(\cl{Q\setminus Q_i}\big)=L\lambda(\partial Q)=0.
\ee
Define now $\hat w_{i,k}:=\theta_i\hat w_k$, extended by $0$ on the whole $\R{N}$.
Notice that
\be\label{319}
\lim_{k\to\infty}\ave_Q \hat w_{i,k}
\lim_{k\to\infty}\ave_Q \theta_i\hat w_k=0.
\ee
From \eqref{316} and \eqref{318}
we obtain
\be\label{320}
\begin{split}
\frac{\de\Phi_{\mu}}{\de\cL^N}(x_0)=&\lim_{k\to\infty} \int_Q f( s_ky + \gamma_{k}, u(x_0) + \hat{w}_k(y)) \,\de y \\
\geq & \liminf_{i\to\infty}\liminf_{k\to\infty} \int_Q f(s_{k}y+\gamma_{k},u(x_0)+\hat w_{i,k}(y))\,\de y \\
= & \liminf_{i\to\infty}\liminf_{k\to\infty} \frac1{s_{k}^N} \int_{s_{k}Q} f\left(x+\gamma_{k},u(x_0)+\hat w_{i,k}\left(\frac{x}{s_{k}}\right)\right)\de x \\
\geq & \liminf_{i\to\infty}\liminf_{k\to\infty} \frac1{s_{k}^N} \int_{([s_{k}]+1)Q} f\left(x+\gamma_{k},u(x_0)+\hat w_{i,k}\left(\frac{x}{s_{k}}\right)\right)\de x \\
& -\limsup_{k\to\infty}\frac{C_1}{s_{k}^N} |([s_{k}]+1)Q\setminus s_{k}Q| |u(x_0)|,
\end{split}
\ee
where we have used the growth condition \eqref{200} together with the fact that $\hat w_{i,k}\left(\frac\cdot{s_{k}}\right)=0$ outside $s_{k}Q$.
It is easy to see that the last term in \eqref{320} vanishes.
Define $n_k:=[s_{k}]+1$
and notice that we can restrict $\hat w_{i,k}\left(\frac\cdot{s_{k}}\right)$ to $n_kQ$ and extend it by $n_kQ$-periodicity to the whole $\R{N}$. Let $U_{i,k}\left(\frac\cdot{n_k}\right)$ be the extended function.
Then \eqref{320} becomes
\begin{equation*
\begin{split}
\frac{\de\Phi_{\mu}}{\de\cL^N}(x_0)\geq & \liminf_{i\to\infty}\liminf_{k\to\infty} \frac1{n_k^N} \int_{n_kQ} f\left(x+\gamma_{k},u(x_0)+U_{i,k}\left(\frac{x}{n_{k}}\right)\right)\de x \\
=& \liminf_{i\to\infty} \liminf_{k\to\infty}\int_Q f(n_{k}y+\gamma_{k},u(x_0)+U_{i,k}(y))\,\de y.
\end{split}
\end{equation*
Notice that $U_{i,k}$ is a $Q$-periodic function such that $U_{i,k}\wsto0$ as $k\to\infty$, $\lim_{i\to\infty}\lim_{k\to\infty}\cA U_{i,k}=0$ in $W^{-1,q}$, and, by \eqref{319},
\be\label{323a}
\ave_Q U_{i,k}\to0\qquad \text{as $k\to\infty$}.
\ee
Therefore, we can apply Proposition \ref{projection} to the function $U_{i,k}-\ave_Q U_{i,k}$, to obtain a new function $V_{i,k}:=\cP(U_{i,k}-\ave_Q U_{i,k})$ such that
\be\label{324}
\begin{split}
\frac{\de\Phi_{\mu}}{\de\cL^N}(x_0)\geq & \liminf_{i\to\infty}\liminf_{k\to\infty} \int_Q f\left(n_{k}y+\gamma_{k},u(x_0)+V_{i,k}(y)+\ave_Q U_{i,k}\right)\de y \\
= & \liminf_{i\to\infty} \int_Q f(n_{k_i}y+\gamma_{k_i},u(x_0)+V_{i,k_i}(y))\de y, \\
\end{split}
\ee
where we have used the Lipschitz continuity of $f$ in the second variable, \eqref{323a}, and chosen an appropriate diagonalizing sequence.
By defining $w_i:=V_{i,k_i}$, formula \eqref{315} is proved.
Estimate \eqref{311} now follows upon changing variables and using Proposition \ref{translations}.
Indeed, from \eqref{324}
\begin{equation*
\begin{split}
\frac{\de\Phi_{\mu}}{\de\cL^N}(x_0)\geq & \liminf_{i\to\infty} \frac1{n_{k_i}^N} \int_{n_{k_i}Q} f\left(x+\gamma_{k_i},u(x_0)+w_i\left(\frac{x}{n_{k_i}}\right)\right) \de x \\
\geq & \liminf_{i\to\infty} f_{\cA-\hom}^{\gamma_{k_i}}(u(x_0))=\liminf_{i\to\infty} f_{\cA-\hom}(u(x_0))=f_{\cA-\hom}(u(x_0)).
\end{split}
\end{equation*
\begin{remark}[$\cA$-quasiconvexity of $f_{\cA-\hom}$]\label{fAhomAqc}
As a consequence of \eqref{214} and \eqref{311} we obtain that $f_{\cA-\hom}$ is $\cA$-quasiconvex.
This result will be used in Section \ref{LBsp}.
To prove it, let $b\in\R{d}$, let $w\in C^\infty_{Q-\per}(\R{N};\R{d})\cap\ker\cA$, and define $w_n(x):=w(nx)$.
Then it is easy to see that $\cA w_n=0$, and, by the Riemann-Lebesgue Lemma, $w_n\wsto0$.
Then,
\begin{equation*
\begin{split}
f_{\cA-\hom}(b)= & \int_Q f_{\cA-\hom}(b)\,\de x = \cF(b;Q) \leq \liminf_{n\to\infty} \cF(b+w_n;Q) \\
\leq & \liminf_{n\to\infty} \int_Q f_{\cA-\hom}(b+w_n(x))\,\de x= \liminf_{n\to\infty} \frac1{n^N}\int_{nQ} f_{\cA-\hom}(b+w(y))\,\de y \\
= & \int_Q f_{\cA-\hom}(b+w(y))\,\de y,
\end{split}
\end{equation*
where the second equality follows from \eqref{214} and \eqref{311}, and where we have used the lower semicontinuity of $\cF$ with respect to the weak-* convergence.
This proves that $f_{\cA-\hom}$ is $\cA$-quasiconvex.
\end{remark}
\subsection{Lower bound -- Singular part}\label{LBsp}
We now prove the inequality
\be\label{351}
\frac{\de\Phi_\mu}{\de|\mu^s|}(x_0)\geq f_{\cA-\hom}^\infty\left(\frac{\de\mu^s}{\de|\mu^s|}(x_0)\right)\qquad \text{for $|\mu^s|$-a.e.\@ $x_0\in\Omega$.}
\ee
Let $x_0\in\supp|\mu^s|\cap E$, where $E$ is the set given by Lemma \ref{lemmarin}.
Now call
$$v_{x_0}:=\frac{\de\mu^s}{\de|\mu^s|}(x_0),$$
that we suppose to be finite together with $\frac {\de \Lambda} {\de |\mu^s|}(x_0)$.
We distinguish two cases, namely $v_{x_0}\in\cC$ and $v_{x_0}\notin\cC$.
\textbf{Case 1: $v_{x_0}\in\cC$.}
Let $\{\omega_1,...,\omega_k\}$ be an orthonormal basis for ${\cal V}_{x_0}$ and let $\{\omega_{k+1},...,\omega_N\}$ be an orthogonal basis for the orthogonal complement of ${\cal V}_{x_0}$ in $\R{N}$. We denote by $Q_{0}$ the unitary cube with center at $0$ and faces orthogonal to $\omega_1,\ldots,\omega_k,\ldots,\omega_N$, that is
$$Q_{0}:= \left\{ x \in \R{N}: |x\cdot \omega_i| < \frac1 {2},\, i=1,\ldots,N \right\}.$$
Choose a decreasing sequence of positive radii $r_j \to 0$ as $j\to\infty$ such that $\Lambda(\partial Q_{0} (x_0,r_j))=0$, thus we also have $\Phi_\mu(\partial Q_{0} (x_0,r_j)) =0$.
We then have
\begin{equation*
\frac {\de \Phi_{\mu}} {\de |\mu^s|} (x_0) = \lim_{j \to \infty} \frac {\Phi_{\mu} (Q_{0} (x_0, r_j))}{|\mu^s|(Q_{0} (x_0,r_j))} = \lim_{j \to \infty} \lim_{n \to \infty} \frac{ \displaystyle\int_{Q_{0} (x_0,r_j)} f\left( \frac {x} {\eps_n}, u_n (x) \right)\de x } {|\mu^s|(Q_{0} (x_0,r_j))}.
\end{equation*}
Define
$$w_{j,n}(y):= \frac {u_n(x_0+r_jy)} {t_j}$$
for $y \in Q_{0}$, where
$$t_j := \frac{|\mu^s|(Q_{0} (x_0,r_j))} {r_j^N} .$$
By changing variables we obtain
\begin{equation*
\frac {\de \Phi_{\mu}} {\de |\mu^s|} (x_0) = \lim_{j \to \infty} \lim_{n \to \infty} \frac1 {t_j} \int_{Q_{0}} f \left( \frac {r_j y} {\eps_n} + \frac {x_0} {\eps_n}, t_j w_{j,n}(y) \right)\,\de y.
\end{equation*}
Note that
$$\frac {x_0} {\eps_n} = \left[ \frac {x_0} {\eps_n} \right] + \left< \frac {x_0} {\eps_n} \right>$$
and let $\gamma_n:=\langle x_0/\eps_n\rangle$.
Thus, using the periodicity of $f$ with respect to the first variable,
we have that
\begin{equation*
\frac {\de \Phi_{\mu}} {\de |\mu^s|} (x_0) = \lim_{j \to \infty} \lim_{n \to \infty}\frac1 {t_j} \int_{Q_{0}} f \left( \frac {r_j y} {\eps_n} + \gamma_n,
t_j w_{j,n}(y) \right)\,\de y.
\end{equation*}
We also have that
$$\lim_{j \to \infty} \lim_{n\to \infty} \int_{Q_{0}} w_{j,n} (y) \varphi(y) \,\de y = \langle \tau, \varphi \rangle$$
for every $\varphi \in C({\overline Q_{0}})$, where $\tau$ is a tangent measure to $\mu^s$ at $x_0$ (see Definition \ref{tan}).
Notice that we can choose the sequence of $r_j$ in such a way that all the properties above hold and we also have $|\tau|(\partial Q_{0})=0$ (see Lemma \ref{lemmarin} and Remark \ref{suc-front}).
We also have
$$\lim_{j \to \infty} \lim_{n\to \infty} \int_{Q_{0}} |w_{j,n} (y)| \,\de y = \frac {\de\Lambda} {\de |\mu^s|} (x_0).$$
Moreover, as ${\cal A}w_{j,n} \to 0$ in $W^{-1,q}$ for some $q \in \left(1, \frac{N}{N-1}\right)$, we can find a sequence ${\hat w}_j= w_{j,n_j}$ such that
$${\hat w}_j \weakst \tau, \qquad {\cal A}{\hat w}_j \to 0,$$
and
\begin{equation}\label{369}
\frac {\de \Phi_{\mu}} {\de |\mu^s|} (x_0) = \lim_{j \to \infty} \frac1 {t_j}\int_{Q_{0}} f \left( \frac {r_j y} {\eps_{n_j}} + \gamma_{n_j}, t_j {\hat w}_j (y) \right)\,\de y,
\end{equation}
where $\gamma_{n_j}$ is a subsequence of $\gamma_n$.
Note that ${\cal A} \tau=0$ and that we have
\begin{equation}\label{formamedtang}
\tau = \frac {\de \mu^s} {\de |\mu^s|} (x_0) |\tau|(x{\cdot}\omega_1,\ldots,x{\cdot}\omega_k),
\end{equation}
which means that $|\tau|$ is invariant for translations in the directions of the complement orthogonal of ${\cal V}_{x_0}$, that is, in the directions of $\mathrm{span} \{\omega_{k+1},\ldots,\omega_N\}$.
Indeed, we have
$${\cal A} \tau = \left( \sum_i A^{(i)} |\tau|_{x_i} \right) \frac {\de \mu^s} {\de |\mu^s|} (x_0)=0,$$
which implies that the vector $(|\tau|_{x_1},\ldots,|\tau|_{x_N})$ belongs to $\mathrm{span}\{\omega_1,\ldots,\omega_k\}$.
We regularize $\tau$ (see Remark \ref{suc-front}) and construct a sequence
$$v_j = \rho_{\eps_{n_j}} * \tau,$$
defined in $Q_{x_0}$.
As $v_j$ has the form as in \eqref{formamedtang} we can extend it by $Q_{0}$-periodicity to a function ${\tilde v}_j$ defined on $\R{N}$, and we still have
$${\cal A} {\tilde v}_j = 0$$
in $\R{N}$, that is, the jumps of $v_j$ are not penalized by ${\cal A}$ (see \cite{BCMS} for the details).
We now apply Lemma \ref{375} to the sequences $\{\hat w_j\}$ and $\{\tilde v_j\}$, with $f_n=f$ for all $n\in\N{}$, to obtain that there exists a sequence $\{z_i\}\subset L^q_{Q_{0}-\per}(\R{N};\R{d})$, $\int_{Q_{0}} z_j=0$, $\cA z_j=0$, $z_j\wsto0$ such that, from \eqref{376}, \eqref{369} becomes
\begin{equation*
\frac {\de \Phi_{\mu}} {\de |\mu^s|} (x_0) \geq \lim_{j \to \infty} \frac1 {t_j} \int_{Q_{0}} f \left( \frac {r_j y} {\eps_{n_j}} + \gamma_{n_j}, t_j \hat u_j (y) \right)\,\de y
\end{equation*}
where
$$\hat u_j (y) = z_j(y) + {\tilde v}_j(y)$$
is $L^q_{Q_{0}-\per} (\R{N}; \R{d})$, ${\cal A}u_j=0$, $u_j \weakst \tau$.
The objective is to modify $\{\tilde u_j\}$ in order to obtain periodicity in the directions of the coordinate axes. Consider cut-off functions $\theta_i \subset C_c^{\infty} (Q_{0}; [0,1])$ such that $\theta_i (x) \equiv 1$ in $\left( 1- \frac1 {i} \right) Q_{0}$. Thus, using the Lipschitz continuity and the condition $\tau (\partial Q_{0})=0$, we have
\begin{equation*}
\frac {d \Phi_\mu} {d |\mu^s|} (x_0) \geq \liminf_{i\to\infty} \liminf_{j\to\infty} \frac1 {t_j} \int_{Q_{0}} f \left(\frac {r_j y} {\eps_{n_j}} + \gamma_{n_j}, t_j \theta_i (y) \hat u_j(y) \right) \,dy.
\end{equation*}
After appropriate diagonalization we get a sequence $\{u_{j_i}\}$ such that
\begin{equation*}
\frac {d \Phi_\mu} {d |\mu^s|} (x_0) \geq \liminf_{i\to\infty} \frac1 {t_{j_i}} \int_{Q_{0}} f \left(\frac {r_{j_i} y} {\eps_{n_{j_i}}} + \gamma_{n_{j_i}}, t_{j_i} \hat u_{j_i} (y) \right) \,dy,
\end{equation*}
where $\hat u_{j_i} \wsto \tau$, ${\cal A}\hat u_{j_i} \to 0$ in $W^{-1,q}$.
By a change of variables we obtain
\begin{equation*}
\frac {d \Phi_\mu} {d |\mu^s|} (x_0) \geq \liminf_{i\to\infty} \frac1 {t_{j_i}} \ave_{T_{j_i} Q_{0}} f \left(x+\gamma_{n_{j_i}}, t_{j_i} U_i \left( \frac {x} {T_{j_i}} \right) \right) \,dx,
\end{equation*}
where $T_{j_i}:= \frac {r_{j_i}}{\epsilon_{n_{j_i}}} \to \infty$ is a sequence of no necessarily integers and $U_i := \hat u_{j_i}$.
Now choose a cube $T_iQ$ such that $T_{j_i} Q_0 \subset \subset T_iQ$ and extend $U_ i \left( \frac {x} {T_{j_i}} \right)$ by $T_i Q$-periodicity. We can choose, for instance,
\be\label{360}
T_i = [C_N T_{j_i} ] +1,
\ee
where $C_N = \sqrt{N}$ is the diagonal of unit cube in $\R{N}$.
We then have
\begin{equation*}
\frac {d \Phi_\mu} {d |\mu^s|} (x_0) \geq \liminf_{i\to\infty} \frac1 {t_{j_i}} \frac1 {T_{j_i}^N} \left [\int_{T_i Q} f \left(x+\gamma_{n_{j_i}}, t_{j_i} U_i \left( \frac {x} {T_{j_i}} \right) \right) \,dx - \int_{T_i Q \setminus T_{j_i} Q_0} f(x+\gamma_{n_{j_i}},0) \,dx\right],
\end{equation*}
and, as the second integral vanishes as $i\to\infty$, by using \eqref{360} we obtain
\begin{equation*}
\frac {d \Phi_\mu} {d |\mu^s|} (x_0) \geq \liminf_{i\to\infty} \frac{C_N^N}{t_{j_i}} \ave_{T_i Q} f \left(x+\gamma_{n_{j_i}}, t_{j_i} U_i \left( \frac {x} {T_{j_i}} \right) \right) \de x,
\end{equation*}
and by changing variables again we get
\begin{equation}\label{361}
\frac {d \Phi_\mu} {d |\mu^s|} (x_0) \geq \liminf_{i\to\infty} \frac{C_N^N}{t_{j_i}} \int_{Q} f \left(T_i y+\gamma_{n_{j_i}}, t_{j_i} U_i \left( \frac {T_i y} {T_{j_i}} \right) \right) \de y.
\end{equation}
As ${\cal A} U_i \left( \frac {T_i} {T_{j_i}}\cdot \right) \to 0$ and $U_i \left( \frac { T_i} {T_{j_i}} \cdot\right)$ is $Q$-periodic, $U_i \left( \frac {T_i} {T_{j_i}}\cdot \right)$ fulfills the hypotheses of Proposition \ref{projection}, and therefore \eqref{361} becomes
\begin{equation*}
\frac {d \Phi_\mu} {d |\mu^s|} (x_0) \geq \liminf_{i\to\infty} \frac{C_N^N}{t_{j_i}} \int_{Q} f \left(T_i y+\gamma_{n_{j_i}}, t_{j_i} \left( \int_Q U_i \left( \frac {T_i x} {T_{j_i}} \right) \de x + V_i(y) \right) \right) \de y,
\end{equation*}
where
$$V_i(y) := \cP \left(U_i \left( \frac {T_i y} {T_{j_i}} \right) - \int_Q U_i \left( \frac {T_i x} {T_{j_i}} \right) \de x \right).$$
It is trivial to verify that $\int_Q V_i(y)\de y=0$.
By changing variables back, we get
\begin{equation*}
\frac {d \Phi_\mu} {d |\mu^s|} (x_0) \geq \liminf_{i\to\infty} \frac{C_N^N}{t_{j_i}} \ave_{T_i Q} f \left(x+\gamma_{n_{j_i}}, t_{j_i} \left( \int_Q U_i \left( \frac {T_i x} {T_{j_i}} \right) \de x + V_i\left( \frac {x} {T_i} \right) \right) \right) \de x.
\end{equation*}
Now, since $\int_Q U_i \left( \frac {T_i x} {T_{j_i}} \right) \de x \to \frac1{C_N^N} \frac {\de \mu^s} {\de |\mu^s|} (x_0)$, by using the Lipschitz condition we get
\begin{equation*}
\begin{split}
\frac {d \Phi_\mu} {d |\mu^s|} (x_0) \geq & \liminf_{i\to\infty} \frac{C_N^N}{t_{j_i}}\ave_{T_i Q} f \left(x+\gamma_{n_{j_i}}, t_{j_i} \left( \frac1{C_N^N} \frac {\de \mu^s} {\de |\mu^s|} (x_0) + V_i\left( \frac {x} {T_i} \right) \right) \right) \de x \\
& -\limsup_{i\to\infty} \frac{LC_N^N}{t_{j_i}}\ave_{T_i Q} \left|\int_Q U_i \left( \frac {T_i x} {T_{j_i}} \right) \de x - \frac1{C_N^N} \frac {\de \mu^s} {\de |\mu^s|} (x_0) \right|\de x,
\end{split}
\end{equation*}
and the $\limsup$ vanishes.
Therefore, by \eqref{222}
we obtain
\be\label{365}
\begin{split}
\frac {\de \Phi_\mu} {\de |\mu^s|} (x_0) \geq & \liminf_{i\to\infty} \frac {f_{{\cal A}-\hom}^{\gamma_{n_{j_i}}} \left(\frac{t_{j_i}}{C_N^N}\frac {\de \mu^s} {\de |\mu^s|} (x_0)\right)} {t_{j_i}/ C_N^N} \\
= & \liminf_{i\to\infty} \frac {f_{{\cal A}-\hom} \left(\frac{t_{j_i}}{C_N^N}\frac {\de \mu^s} {\de |\mu^s|} (x_0)\right)} {t_{j_i}/ C_N^N} = {f^{\infty}_{{\cal A}-\hom}} \left( \frac {\de \mu^s} {\de |\mu^s|} (x_0)\right).
\end{split}
\ee
Here we have used Proposition \ref{translations} and, in the last equality, the fact that $\cA$-quasiconvex functions are convex in the directions of the characteristic cone $\cC$ (see \cite[Proposition 3.4]{FM}). $\cA$-quasiconvexity for $f_{\cA-\hom}$ was proved in Remark \ref{fAhomAqc}, and this implies that the $\limsup$ in \eqref{104} is actually a limit.
This concludes the proof of \eqref{351} in the case $v_{x_0}\in\cC$.
\textbf{Case 2: $v_{x_0}\notin\cC$.}
This case is analogous to Case 1, but simpler, since the condition $v_{x_0}\notin\cC$ implies that $\cV_{v_{x_0}}=\{0\}$, and in turn that the tangent measure $\tau$ is given by
\be\label{363}
\tau=v_{x_0}\cL^N.
\ee
In particular, we do not need to find a suitable rotated cube $Q_{0}$ to perform the homogenization and we do not need to regularize the tangent measure.
Since from Proposition \ref{Lipschitz} $f_{\cA-\hom}$ is Lipschitz continuous, by Lemma 4.2 in \cite{BCMS} there exists a sequence $\{t_j\}$ such that $t_j\to\infty$ as $j\to\infty$ and
\be\label{lim}
f_{\cA-\hom}^\infty(v_{x_0})=\lim_{j\to\infty}\frac{f_{\cA-\hom}(t_j v_{x_0})}{t_j}.
\ee
It is then possible to choose a sequence $\{\delta_j\}$ such that $\delta_j\to0$ as $j\to\infty$, $\Lambda(\partial Q(x_0,\delta_j)=0$, and
$
t_j=\frac{|\mu^s|(Q(x_0,\delta_j))}{\delta_j^N}.
$
With these sequences $\{t_j\}$ and $\{\delta_j\}$ we get the equivalent of equation \eqref{369} where the rotated cube $Q_{0}$ is replaced by the unit cube $Q$ and the sequence $\hat w_j$ converges weakly-* to the tangent measure $\tau$ in \eqref{363}.
The proof proceeds now as in Case 1, with the integers $T_i$ in \eqref{360} now replaced by
$
T_i=[T_{j_i}]+1
$
(in particular, there is no need to introduce the constant $C_N$).
To conclude, we observe that the last equality in \eqref{365} now follows from \eqref{lim}.
\begin{remark}
It easy to prove that under coercivity conditions any sequence of minimizers (or approximate minimizers in case the infimum is not attained) of the functional
$$I(u)= \int_{\Omega} f \left( \frac {x} {\epsilon_n}, u \right)\,dx$$
will converge (up to a subsequence) to the minimum points of the limit functional
$$\cF_{\cA-\hom}(\mu):=\int_\Omega f_{\cA-\hom}(u^a)\,\de x+\int_\Omega f_{\cA-\hom}^\infty\left(\frac{\de\mu^s}{\de|\mu^s|}\right)\,\de|\mu^s|.$$
In fact, in the particular case where the operator $\cal A$ admits an extension property, the condition ${\cal A} u_n \to 0$ in \eqref{202} can be replaced by ${\cal A} u_n = 0$, and the property above comes directly from known results in $\Gamma$-convergence on metric spaces.
\end{remark}
\noindent {\bf Acknowledgments.}
Partial support for this research was provided by the Funda\c{c}\~ao para a Ci\^encia e a Tecnologia (Portuguese Foundation for Science and Technology) through the Carnegie Mellon Portugal Program under Grant FCT-UTA\_CMU/MAT/0005/2009 ``Thin Structures, Homogenization, and Multiphase Problems''. The authors warmly thank the Centro de An\'alise Matem\'atica, Geometria e Sistemas Din\^amicos (CAMGSD) at the Departamento de Matem\'atica of the Instituto Superior T\'ecnico, Universidade de Lisboa, where the research was carried out.
|
\section{Introduction}
\label{sec:introd}
The overarching problem of variable selection is to choose the best
model out of a set of candidate models $\hmod_\smM$. Given measured data
$\mathcal{D}$, the Bayesian solution is to compute the posterior
probability for each model with Bayes' theorem,
\begin{align}
\label{ntcb}
p(\hmod_\smM\,|\,\mathcal{D})
&= \frac{p(\mathcal{D}\,|\,\hmod_\smM)\,p(\hmod_\smM)}{p(\mathcal{D})}
\ =\
\frac{p(\mathcal{D}\,|\,\hmod_\smM)\,p(\hmod_\smM)}{\sum_{\!\scriptscriptstyle M} p(\mathcal{D}\,|\,\hmod_\smM)p(\hmod_\smM)}\,.
\end{align}
As equal priors are usually assigned to the competing models, model
comparison becomes a task in finding the marginal likelihood or
evidence for each model, i.e.\ solving the integral over all $K$
model parameters ${\bm{\beta}}_{\!\scriptscriptstyle M} = (\theta_1,\ldots,\theta_{\!\scriptscriptstyle K})$ of
the likelihood $p(\mathcal{D}\,|\,{\bm{\beta}}_{\!\scriptscriptstyle M},\hmod_\smM)$ weighted by the
parameter prior $p({\bm{\beta}}_{\!\scriptscriptstyle M} \,|\, \hmod_\smM)$,
\begin{align}
\label{ntdb}
p(\mathcal{D}\,|\,\hmod_\smM)
&= \int p(\mathcal{D}\,|\,{\bm{\beta}}_{\!\scriptscriptstyle M},\hmod_\smM)\,
p({\bm{\beta}}_{\!\scriptscriptstyle M}\,|\,\hmod_\smM)\, d{\bm{\beta}}_{\!\scriptscriptstyle M}.
\end{align}
The preferred model will be the one with the largest evidence i.e.\
with the highest prior-weighted average over all parameters of the
likelihood.
Where computation of $p(\mathcal{D})$ over the entire model set is
impractical or even impossible, this is circumvented by taking ratios
of two model probabilities in the form of Bayes Factors, since
$p(\mathcal{D})$ cancels and under the equal-model-prior assumption, they
become ratios of the respective model evidences,
\begin{align}
\label{nte}
\mathrm{BF}(\hmod_\smM;\mathcal{M}_{{\!\scriptscriptstyle M}'}) %
&= \log \frac{p(\hmod_\smM\,|\,\mathcal{D})}{p(\mathcal{M}_{{\!\scriptscriptstyle M}'}\,|\,\mathcal{D})}
= \log \frac{p(\mathcal{D}\,|\,\hmod_\smM)}{p(\mathcal{D}\,|\,\mathcal{M}_{{\!\scriptscriptstyle M}'})}\,.
\end{align}
Finding the evidence can also be difficult since model parameter
spaces ${\mathcal{A}}({\bm{\beta}}_{\!\scriptscriptstyle M})$ differ widely in size and
dimension. While convenient at first sight, assigning uniform priors
to parameters results in the untenable situation of strong dependence
of each model's evidence and consequently of Bayes Factors on
arbitrarily chosen cutoff parameters introduced by the uniform priors.
In addition, the dimension $K$ of the parameter space often differs
from model to model, compounding the problems associated with uniform
parameter priors. Furthermore, improper priors must be excluded from
the start if they are model specific because they remain in the
evidence.
These problems appear even in the simplest case of ``canonical
regression'' in which the likelihood is Gaussian and the models are
restricted to linear function spaces as studied in the past by
\cite{Jeffreys1967}, \cite{Zellner1971}, \cite{Box1973} and many
others. %
The quest for robust and fair model comparison in this restricted
context dates back to \cite{Jeffreys1967} whose univariate Cauchy
prior was extended by \cite{Zellner1980} to multivariate form. A
simpler ``$g$-prior'' was subsequently invented by \cite{Zellner1986}
to facilitate ease of use by closed-form solutions and has found wide
application. The specific choice for $g$ and internal inconsistencies
have, however, dogged the simple $g$-prior, leading for example
\cite{Liang2008} to introduce mixtures of such $g$-priors. They showed
that $g$-mixtures resolved the inconsistencies of the simple $g$-prior
and could show it and the original Zellner-Siow prior to be special
cases within the mixture framework.
Common to all these efforts was the recognition, sometimes only
implicitly, of an underlying spherical symmetry in parameter
space. For example, the \cite{Zellner1986} prior was based on the use
of a Gaussian prior for parameters ${\bm{\beta}}$ with the same design
matrix ${\mathbb{X}}$ as the data and precision parameter $\phi = 1/\sigma^2$
but including an additional scale parameter $g$,
\begin{align}
\label{dce}
p({\bm{\beta}}\,|\, g,\sigma,\mathcal{H}_{\!\scriptscriptstyle Z})
&= \frac{\exp\bigl[-N {\bm{\beta}}^T {\mathbb{X}}^T{\mathbb{X}} {\bm{\beta}}
/2 \sigma^2 g \bigr] }
{(\det{\mathbb{X}}^T{\mathbb{X}})^{1/2}(2\pi\sigma^2 g)^{K/2}},
\end{align}
which as detailed in Section \ref{sec:nkrdiag} is easily
transformed into spherically symmetric form
\begin{align}
\label{dcf}
p({\bm{b}}\,|\, g,\sigma,\mathcal{H}_{\!\scriptscriptstyle Z})
&= \frac{e^{-N {\bm{b}}^2 / 2 \sigma^2 g}}{(2\pi\sigma^2 g)^{K/2}}.
\end{align}
As pointed out by \cite{Leamer1978}, the behaviour of the parameter
estimators is controlled by the symmetries of the prior. Often there
is no prior information which explicitly breaks the inherent spherical
symmetry of Gaussians, suggesting that spherical symmetry has been the
basis for many of the parameter priors in the canonical regression
literature all along.
In this paper, we take the underlying spherical symmetry to its
logical conclusion by introducing a radius variable $r$, common to all
models $\hmod_\smM$ and for arbitrary parameter space dimension $K$, and
explicitly enforcing spherical symmetry on the hypersphere of radius $r$
by means of a $r$-prior. The projection from ${\bm{b}}$ onto
$r$ is then carried out generally, thereby reducing the $K$-dimensional
problem to a one-dimensional integral.
The $r$-prior framework introduced here encompasses earlier work as
special cases, including the conjugate-prior results of
\cite{George1997}, \cite{Raftery1997}, \cite{Berger2001}, the
various Zellner priors and the $g$-prior mixtures of \cite{Liang2008}
and shares their computational efficiency.
Not surprisingly, we find significant mathematical correspondence
between our $r$-prior and the $g$-prior mixtures of \cite{Liang2008}
as the latter implicitly assumes the same spherical symmetry made
explicit by the $r$-prior. Unlike the $g$-prior mixtures, the
$r$-prior is however not limited to mixtures of conjugate (Gaussian)
priors.
In Section \ref{sec:nkr}, we first treat the case of a single unknown
dispersion parameter $\sigma$, using it by example to introduce the
radius $r$ of the parameter hypersphere. The central result in
Eq.~(\ref{prk}) is used both to show how $g$-priors and the prior of
\cite{Zellner1980} can be obtained with particular choices of
$r$-priors as well as to introduce a simpler yet equally powerful new
$r$-prior based on properties revealed by the Mellin transform.
In Section \ref{sec:fxr}, the single variable dispersion parameter
$\sigma$ is replaced by a set of fixed known error variances
$(\sigma_1^2,\ldots,\sigma_{\!\scriptscriptstyle N}^2)$, one for each data point. What we
have in mind here is the application of the $r$-prior formalism to
existing data with measured standard errors treating the
$\sigma_n$ not as likelihood variables but as constants.
In Section \ref{sec:bfc}, we test and compare our results to related
model comparison criteria, concluding with a discussion in Section
\ref{sec:dcn}.
\section{Single unknown dispersion parameter}
\label{sec:nkr}
\subsection{Definition and diagonalisation}
\label{sec:nkrdiag}
The generic model consists of a data set or response vector $\mathcal{D} =
{\bm{y}} = (y_1, \ldots, y_{\!\scriptscriptstyle N}) \in \mathbb{R}^N$ measured at fixed sampling
points ${\bm{c}} = (c_1, \ldots, c_{\!\scriptscriptstyle N} ) \in \mathbb{R}^N$. The set of
predictors is represented by $K$ column vectors ${\bm{X}}_k =
(X_k(c_1),\ldots,X_k(c_{\!\scriptscriptstyle N}))^T$ which together form the $N{\times}K$
design matrix ${\mathbb{X}} = ({\bm{X}}_1\, {\bm{X}}_2\, \cdots \, {\bm{X}}_{\!\scriptscriptstyle K})$.
While the information $\mathcal{H}_0 = \{{\bm{c}},N\}$ is the same for all
models, the design matrix ${\mathbb{X}}$ and the dimensionality of the
predictor space $K$ are model-specific, $\mathcal{H}_{\!\scriptscriptstyle M} = \{{\mathbb{X}}_{\!\scriptscriptstyle M},
K_{\!\scriptscriptstyle M}\}$. A given model $\hmod_\smM$ is specified by a prior $\mathcal{H}_p$ plus
$\{\mathcal{H}_0,\mathcal{H}_{\!\scriptscriptstyle M}\}$ and of course the assumption that the errors
between data and model are iid and Gaussian distributed.
From this point, we focus on developing a single model $\mathcal{M}$ and
hence drop the subscript $M$. We limit ourselves to linear regression
with coefficients ${\bm{\beta}} = (\beta_1, \ldots, \beta_{\!\scriptscriptstyle K}) \in
{\mathcal{A}}({\bm{\beta}}) = \mathbb{R}^K$ and errors ${\bm{\varepsilon}} = {\bm{y}} -
{\mathbb{X}}^T {\bm{\beta}}$ which are assumed to be iid and normally distributed
with a single unknown dispersion parameter, ${\bm{\varepsilon}} \sim
N(0,\sigma^2\bm{I_{\!\scriptscriptstyle N}})$, or
\begin{align}
\label{ntd}
p({\bm{\varepsilon}}\,|\, \sigma, \mathcal{H}_0) %
&= \prod_{n=1}^N
\frac{e^{-\varepsilon_n^2/2\sigma^2}}{\sigma\sqrt{2\pi}},
\end{align}
resulting in the joint likelihood
\begin{align}
\label{dgb}
p({\bm{y}}\,|\,{\bm{\beta}},\sigma,\mathcal{M}) &= (2\pi)^{-N/2} \sigma^{-N} e^{-NQ/2\sigma^2},
\end{align}
with
\begin{align}
\label{dgc}
Q({\bm{y}},{\bm{\beta}},\sigma \,|\, \mathcal{M})
&= \frac{1}{N}\tnorm{{\bm{y}} - {\mathbb{X}}{\bm{\beta}}}^2
= \frac{1}{N} ({\bm{y}} - {\mathbb{X}}{\bm{\beta}})^T ({\bm{y}} - {\mathbb{X}}{\bm{\beta}})
\nonumber\\
&= \frac{1}{N}\sum_{n=1}^N \left(y_n-{\textstyle\sum_{k=1}^K}
X_k(c_n)\,\beta_k\right)^2
\end{align}
related to the usual chisquared statistic by $NQ/\sigma^2 = \chi^2$.
Finding the maximum likelihood and the concomitant diagonalisation of
the parameters in ${\mathcal{A}}({\bm{\beta}})$ proceeds in the usual way, except
that we have extracted the explicit $N$-dependence in Eq.~(\ref{dgb})
and define $ {\langle \bmy^2 \rangle} = {\bm{y}}^T{\bm{y}} / N$, ${\mathbb{H}} = {\mathbb{X}}^T{\mathbb{X}} / N$ and
\begin{align}
\label{dgf}
{\bm{h}} &= {\mathbb{X}}^T{\bm{y}} / N,
\end{align}
in terms of which
\begin{align}
\label{dgh}
Q &= {\langle \bmy^2 \rangle} + {\bm{\beta}}^T {\mathbb{H}} {\bm{\beta}} - 2 {\bm{h}}^T{\bm{\beta}}.
\end{align}
The minimum of $Q$ occurs at the likelihood mode
\begin{align}
\label{dgj}
\bm{\hat{\beta}} &= {\mathbb{H}}^{-1}\,{\bm{h}} = ({\mathbb{X}}^T{\mathbb{X}})^{-1}{\mathbb{X}}^T {\bm{y}}.
\end{align}
The quadratic form in (\ref{dgh}) is standardised to the new parameter
set ${\bm{b}} \in {\mathcal{A}}({\bm{b}})$ via the eigenvalue equation ${\mathbb{H}}
\,{\bm{e}}_\ell = {\bm{e}}_\ell\lambda_\ell$ with eigenvalues $\lambda_\ell$
and column eigenvectors ${\bm{e}}_\ell$ which are orthonormalised,
${\bm{e}}^T{\bm{e}} = {\mathbb{I}}$, or using the diagonal eigenvalue matrix
${\mathbb{L}} = \mathrm{diag}(\lambda_1, \ldots, \lambda_{\!\scriptscriptstyle N})$ and
orthogonal eigenvector matrix ${\mathbb{S}} = ({\bm{e}}_1\, \cdots\, {\bm{e}}_{\!\scriptscriptstyle K})$,
\begin{align}
\label{dgk}
{\mathbb{H}} {\mathbb{S}} &= {\mathbb{S}} {\mathbb{L}}.
\end{align}
As in \cite{Bretthorst1988}, we transform from ${\bm{\beta}}$ to ${\bm{b}}$
by a rotation by ${\mathbb{S}}$ and a scale change by ${\mathbb{L}^{\scriptscriptstyle \! 1\! /\! 2}} =
\mathrm{diag}(\sqrt{\lambda_1}, \ldots, \sqrt{\lambda_{\!\scriptscriptstyle K}})$,
\begin{align}
\label{dgm}
{\bm{\beta}} &= {\mathbb{S}} {\mathbb{L}^{\scriptscriptstyle \! -1\! /\! 2}} \,{\bm{b}},\\
\label{dgl}
{\bm{b}} &= {\mathbb{L}^{\scriptscriptstyle \! 1\! /\! 2}}\, {\mathbb{S}}^T {\bm{\beta}},
\end{align}
so that the second and third terms of Eq.~(\ref{dgh})
become\footnote{Note that \cite{Bretthorst1988} uses row eigenvectors
rather than the column vectors used in the current literature.}
\begin{align}
\label{dgn}
{\bm{\beta}}^T{\mathbb{H}}{\bm{\beta}} &= {\bm{b}}^T{\bm{b}} = {\bm{b}}^2, \\
\label{dgo}
{\bm{\beta}}^T {\bm{h}} &= {\bm{b}}^T\bm{\hat{b}},\\
\label{dgp}
\bm{\hat{b}} &= {\mathbb{L}^{\scriptscriptstyle \! 1\! /\! 2}}\, {\mathbb{S}}^T \bm{\hat{\beta}},
\end{align}
and $Q$ is decomposed into a ${\bm{b}}$-independent minimum
(equivalent to minimum-$\chi^2$) and a quadratic around the mode,
\begin{align}
\label{dgq}
Q &= Q_0 + R_{\bm{\hat{b}}}^2, \\
\label{dgqb}
Q_0 &= \tfrac{1}{N} \tnorm{{\bm{y}} - {\mathbb{X}}\bm{\hat{\beta}}}^2
= {\langle \bmy^2 \rangle} - \bm{\hat{\beta}}^T {\mathbb{H}} \bm{\hat{\beta}}
= {\langle \bmy^2 \rangle} - \bm{\hat{b}}^2,
\\
\label{dgr}
R_{\bm{\hat{b}}}^2 &=
({\bm{\beta}} - \bm{\hat{\beta}})^T {\mathbb{H}} ({\bm{\beta}} - \bm{\hat{\beta}})
= \tnorm{{\bm{b}} - \bm{\hat{b}}}^2.
\end{align}
In terms of the standardised parameters, the likelihood is
\begin{align}
p({\bm{y}}\,|\, {\bm{b}},\sigma,\mathcal{M}) %
&= \frac{1}{(2\pi\sigma^2)^{N/2}}
\exp\left[ -\frac{N}{2\sigma^2}\left(Q_0 + R_{\bm{\hat{b}}}^2\right) \right]
\nonumber\\
\label{dgz}
&= F(\sigma) \exp
\left[ - \frac{N}{2\sigma^2} {\bm{b}}^2 + \frac{N}{\sigma^2} \bm{\hat{b}}^T {\bm{b}} \right],
\end{align}
with $F(\sigma) = (2\pi)^{-N/2} \sigma^{-N} e^{-N {\langle \bmy^2 \rangle} /2\sigma^2}$.
\subsection{Projection onto one dimension: the $r$-prior}
\label{sec:rpr}
For model comparison, we wish to calculate the evidence, which in the present
model family is the marginal likelihood
$ %
p({\bm{y}}\,|\,\mathcal{M})
= \int d{\bm{b}}\,d\sigma \,p({\bm{y}}\,|\, {\bm{b}},\sigma,\mathcal{M})
\,p({\bm{b}}\,|\, \sigma,\mathcal{M})\,p(\sigma\,|\, \mathcal{M})
$. %
Specification of the $K$-dimensio\-nal ${\bm{b}}$-prior and the integral
over ${\bm{b}}$ represents a significant challenge. In our view, the best
solution is to choose a prior for ${\bm{b}}$ which is explicitly
spherically symmetric in ${\mathcal{A}}({\bm{b}})$ by introducing a radius $r$,
\begin{align}
\label{prc}
p({\bm{b}}\,|\, r,\mathcal{M}) %
&= \frac{\Gamma(K/2)}{\pi^{K/2}} %
\frac{\delta(r - \tnorm{{\bm{b}}})}{2 \,r^{K-1}} %
\ =\ \frac{\Gamma(K/2)}{\pi^{K/2}} \frac{\delta(r^2 - {\bm{b}}^2)}{r^{K-2}},
\end{align}
where $\delta(x)$ is the Dirac delta function,\footnote{As set out in
the literature and motivated e.g.\ by \cite{Jaynes2003appb}, the
Dirac delta function is a limit of a sequence of probability density
functions, and transformation of its arguments follows the standard
rules for pdf transformation under change of variable.} plus an
intermediate $r$-prior $p(r\,|\, \sigma,\mathcal{M})$. This choice of prior
is equivalent to the assumption that the prior information available
to the observer is unchanged under rotation of ${\bm{b}}$ in
${\mathcal{A}}({\bm{b}})$. This rotational Principle of Indifference or
``information isotropy'' in parameter space implies that $p({\bm{b}}\,|\,
r,\mathcal{M})$ must be uniformly distributed over the surface of the
$K$-dimensional hypersphere of radius $r$, for every possible value of
$r$.
Specifically, the observer has no reason a priori to prefer, or give
nonuniform prior weight to, any one of the axial directions in
${\mathcal{A}}({\bm{b}})$ i.e.\ to any specific component of the transformed
design matrix ${\mathbb{X}}{\mathbb{S}}{\mathbb{L}}^{-1/2}$, and hence to any original
predictor $X_k(c)$, apart from the scales and covariances introduced
by the design matrix itself during the backtransformation from ${\bm{b}}$
to ${\bm{\beta}}$. The mathematical consequence of this argument is
Eq.~(\ref{prc}).
Once $r$ is included, the evidence is given by the $(K{+}2)$-fold
integral
\begin{align}
\label{ntb}
p({\bm{y}}\,|\,\mathcal{M}) %
&= \int_0^\infty \!d\sigma\, p(\sigma\,|\,\mathcal{M}) %
\int_0^\infty \!dr \, p(r\,|\,\sigma,\mathcal{M})%
\int_{\mathbb{R}^K} \! d{\bm{b}} \, p({\bm{y}},{\bm{b}} \,|\, r,\sigma,\mathcal{M}).
\end{align}
While at first sight the extra integral may seem unnecessary, the
symmetry of prior $p({\bm{b}}\,|\, r,\mathcal{M})$ significantly simplifies the
problem since
\begin{align}
\label{prig}
p({\bm{y}}\,|\, r,\sigma,\mathcal{M}) %
& = \int d{\bm{b}}\, p({\bm{y}},{\bm{b}} \,|\, r,\sigma,\mathcal{M})
\end{align}
can be calculated once and for all in terms of the likelihood
$p({\bm{y}}\,|\,{\bm{b}},\sigma,\mathcal{M})$ and $r$-prior $p({\bm{b}}\,|\, r,\mathcal{M})$,
leaving us with the comparatively simple task of a two-dimensional
integral over $dr$ and $d\sigma$.
We use the Laplace-type integral representation for the Dirac delta
function and an integral representation of the generalised confluent
hypergeometric function \cite{Watson1922}
\begin{align}
\label{prih}
\delta(r^2 - {\bm{b}}^2) &= \int_{\mathcal{C}}\frac{ds}{2\pi i}
\exp\left[ sr^2 - s {\bm{b}}^2 \right], \\
\label{prij}
{}_0F_1(b\,;\,z) &= \frac{\Gamma(b)}{2\pi i} %
\int_{\mathcal{C}} du\,u^{-b}\exp\left(u + \frac{z}{u}\right),
\end{align}
with $\mathcal{C}$ the contour integral along the imaginary line from
$(c{-}i\infty)$ to $(c{+}i \infty)$, to obtain
\begin{align}
p({\bm{y}}\,|\, r,\sigma,\mathcal{M})
&= \frac{F(\sigma)\,\Gamma(K/2)}{\pi^{K/2}\,r^{K-2}}
\int_{\mathcal{C}} \frac{ds}{2\pi i}\, e^{sr^2}
\int d{\bm{b}}\,\exp\left\{ - \left(\frac{N}{2\sigma^2} {+} s\right){\bm{b}}^2
+ \frac{N}{\sigma^2} \bm{\hat{b}}^T {\bm{b}}
\right\}
\nonumber\\
\label{prj}
&= \frac{F(\sigma)\,\Gamma(K/2)}{\pi^{K/2}\,r^{K-2}}
\int_{\mathcal{C}} \frac{ds}{2\pi i}
\left(\frac{2\pi\sigma^2}{N{+}2\sigma^2 s}\right)^{\tfrac{K}{2}}%
\exp\left\{ sr^2 + \frac{N^2\, \bm{\hat{b}}^2}{2\sigma^2 (N{+}2\sigma^2 s)} \right\},
\end{align}
with $\bm{\hat{b}}^2 = {\bm{h}}^T {\mathbb{H}}^{-1}{\bm{h}}$ a function of ${\bm{y}}$ through
Eq.~(\ref{dgf}), leading to a closed form in terms of the generalised
hypergeometric function,
\begin{equation}
\label{prk}
p({\bm{y}}\,|\, r,\sigma,\mathcal{M})
= \frac{e^{-N({\langle \bmy^2 \rangle}+r^2)/2\sigma^2}}{(2\pi\sigma^2)^{N/2}}
\;{}_0F_1\left(\frac{K}{2}\,;\, \frac{N^2 \bm{\hat{b}}^2 r^2}{4\sigma^4} \right).
\end{equation}
This result is central. It shows that the sufficient statistics are
$Q_0$ and $\bm{\hat{b}}^2$ or alternatively ${\langle \bmy^2 \rangle}$ and $\bm{\hat{b}}^2$, and
that the $K$-dimensional parameter spaces ${\mathcal{A}}({\bm{\beta}})$ and
${\mathcal{A}}({\bm{b}})$ can be reduced to the one-dimensional space
${\mathcal{A}}(r) = \mathbb{R}^+$.
The same result can be obtained via the Fourier transform
\begin{align}
\label{prr}
\Phi[{\bm{t}},{\bm{b}},p({\bm{y}},{\bm{b}}\,|\, r,\sigma,\mathcal{M})]
&= \int d{\bm{b}}\, e^{i{\bm{t}}^T {\bm{b}}}\,p({\bm{y}},{\bm{b}} \,|\, r,\sigma,\mathcal{M})
\end{align}
whose calculation proceeds exactly as above with the substitution of
$(N\bm{\hat{b}}/\sigma^2)$ by $(N\bm{\hat{b}}/\sigma^2) + i {\bm{t}}$, leading to
\begin{align}
\label{prs}
\Phi[{\bm{t}},{\bm{b}},p({\bm{y}},{\bm{b}}\,|\, r,\sigma,\mathcal{M})]
&= \frac{e^{-N({\langle \bmy^2 \rangle}+r^2)/2\sigma^2}}{(2\pi\sigma^2)^{N/2}}
\;{}_0F_1\left(\frac{K}{2}\,;\,
\frac{(N \bm{\hat{b}} + i\sigma^2 {\bm{t}})^2 r^2}{4\sigma^4} \right),
\end{align}
from which the evidence follows as $p({\bm{y}}\,|\, r,\sigma,\mathcal{M}) =
\Phi[{\bm{t}}{=}\bm{0},{\bm{b}},p({\bm{y}},{\bm{b}}\,|\, r,\sigma,\mathcal{M})]$.
\subsection{Connection of $r$-priors with the hyper-$g$ and Zellner-Siow priors}
\label{sec:zsg}
Before introducing a new $r$-prior, we first show that the $g$-prior
of \cite{Zellner1986}, the hyper-$g$ prior of \cite{Liang2008} and the
original Cauchy prior of \cite{Zellner1980} can all be written in
terms of suitable $r$-priors as follows. In the case of the simple
$g$-prior, the appropriate $r$-prior is gamma-distributed,
\begin{align} \label{dcg}
p(r\,|\, g,\sigma,\mathcal{H}_{\!\scriptscriptstyle Z}) &= \frac{2}{r\Gamma(K/2)}
\left(\frac{Nr^2}{2\sigma^2 g}\right)^{K/2}
e^{-N r^2 / 2\sigma^2 g},
\end{align}
leading to evidence
\begin{align}
p({\bm{y}}\,|\, g,\sigma,\mathcal{M})
&= \int dr\,p({\bm{y}}\,|\, r,\sigma,\mathcal{M}) \,p(r\,|\, g,\sigma,\mathcal{H}_{\!\scriptscriptstyle Z}) \\
\label{ygsh}
&= \frac{(1+g)^{-K/2}}{(2\pi\sigma^2)^{N/2}}
\exp\left[-\frac{N{\langle \bmy^2 \rangle}}{2\sigma^2}+\frac{g N\bm{\hat{b}}^2}{2(1+g)\sigma^2}\right]
\end{align}
whose $\sigma$-integrated version can be obtained on using a Jeffreys
prior $p(\sigma\,|\, H_{\!\scriptscriptstyle J})$.
Likewise, the evidence for the hyper-$g$ prior introduced by
\cite{Liang2008}, which according to \cite{Celeux2012} is
\begin{align}
p({\bm{y}}\,|\,\mathcal{H}_g,\mathcal{M})
&= \frac{(a-2)\Gamma(N/2)}{2(K+a-2)}
\left(N \pi {\langle \bmy^2 \rangle} \right)^{-N/2}
{}_2F_1\!\left( 1\,;\, \frac{N}{2} \,;\, \frac{K{+}a}{2} \,;\, \frac{\bm{\hat{b}}^2}{{\langle \bmy^2 \rangle}}
\right),
\end{align}
can be found either in terms of $g$ or $r$,
\begin{align}
p({\bm{y}}\,|\,\mathcal{H}_g,\mathcal{M})
&= \int dr\,d\sigma\,p({\bm{y}}\,|\, r,\sigma,\mathcal{M})\,p(r\,|\, \sigma,K,\mathcal{H}_g)\,
p(\sigma\,|\,\mathcal{H}_{\!\scriptscriptstyle J}) \\
&= \int dg\,d\sigma\,p({\bm{y}}\,|\, g,\sigma,\mathcal{M})\,p(g\,|\, \mathcal{H}_g)\,
p(\sigma\,|\,\mathcal{H}_{\!\scriptscriptstyle J})
\end{align}
by on the one hand again using Eq.~(\ref{prk}) and a $r$-prior based
on a confluent hypergeometric function,
\begin{align}
\label{ssd}
p(r\,|\,\sigma,K,\mathcal{H}_g)
&= \frac{\Gamma((a{+}K)/2-1)}{\Gamma(K/2)}\, \frac{(a{-}2)}{r}
\left(\frac{N r^2}{2\sigma^2}\right)^{K/2}
U\!\left(\frac{a{+}K{-}2}{2}\,;\,\frac{K}{2} \,;\, \frac{Nr^2}{2\sigma^2} \right),
\end{align}
while for the $g$-integral using Eq.~(\ref{ygsh}) and
\begin{align} \label{dci}
p(g\,|\, \mathcal{H}_g) &= \frac{a-2}{2(1+g)^{a/2}},
\quad a > 2.
\end{align}
Thirdly, the evidence for the \cite{Zellner1980} prior, which is a
complicated series of confluent hypergeometric functions
\begin{align}
\label{ssf}
p({\bm{y}}\,|\,\mathcal{H}_{{\!\scriptscriptstyle ZS}},\mathcal{M})
&=
\sum_{j=0}^\infty \left(\frac{N \bm{\hat{b}}^2}{2{\langle \bmy^2 \rangle}}\right)^j
\frac{\Gamma\left(\frac{1+K}{2}\right)\Gamma\left(j{+}\frac{N}{2}\right)}
{\left(N\pi{\langle \bmy^2 \rangle}\right)^{N/2} j!\,2\sqrt{\pi}}
\;
U\!\left(j{+}\frac{K}{2}\,;\,j{+}\frac{1}{2} \,;\, \frac{N}{2} \right),
\end{align}
can be found on the one hand in terms of $r$ using once again
Eq.~(\ref{prk}), a Jeffreys prior and a Zellner-Siow $r$-prior,
\begin{align}
\label{ssg}
p(r\,|\,\sigma,K,\mathcal{H}_{{\!\scriptscriptstyle ZS}})
&= \frac{\Gamma((K{+}1)/2)}{\Gamma(K/2)\,\Gamma(1/2)} %
\;\frac{2\sigma\,r^{K-1}}{\left(\sigma^2+r^2\right)^{(1+K)/2}}.
\end{align}
Taking the alternative $g$-route by integrating
\begin{align}
\label{ssgg}
p(g\,|\, \mathcal{H}_{zs}) &= \sqrt{\frac{N}{2\pi}} e^{-N/2g} g^{-3/2}
\end{align}
together with (\ref{dcg}) and a Jeffreys prior again yields (\ref{ssf}).
\subsection{A parabolic $r$-prior} \label{sec:gmp}
Beyond the special cases covered above, the choice of $p(r\,|\,\mathcal{M})$
leaves much room for new priors. In this section, we construct one
example $r$-prior, making use of the Mellin transform
\begin{align}
\mathcal{M}(f; s) &= \int_0^\infty f(r)\,r^{s-1} dr,
\end{align}
because of its useful property of immediately exhibiting both the
asymptotic and series behaviour of the function $f(r)$. Technically,
translating the contour of the inverse Mellin transform across the
poles left of the strip of analyticity results in a series expansion
in $r$, while translation across the poles to the right gives an
asymptotic expansion. These properties are useful for examining
functions and to construct a prior with the desirable properties.
We are looking for a prior with behaviour similar to the Zellner-Siow
$r$-prior of Eq.~(\ref{ssg}) but preferably with a closed-form
solution. The Zellner-Siow prior goes like $r^{K-1}$ close to zero
and like $r^{-2}$ for large $r$. The Mellin transform of the
Zellner-Siow $r$-prior
\begin{align}
\mathcal{M}(p(r\,|\, \sigma,K,\mathcal{H}_{{\!\scriptscriptstyle ZS}}); s)
&= \frac{\sigma^{s{-}1}}{\sqrt{\pi}}\,
\frac{\Gamma\left[1{-}(s/2)\right]\Gamma\left[(K{+}s{-}1)/2\right]}
{\Gamma\left[K/2\right]}
\end{align}
has a strip of convergence of $0<s<2$. Clearly,
$\Gamma\left[1-(s/2)\right]$ has poles at $s = 2,4,6,\ldots$, while
$\Gamma\left[(K{+}s{-}1)/2\right]$ has poles at $s = 1{-}K,
-1{-}K,\ldots$, which immediately gives the above desired series
expansions. This form leads, however, to a complicated evidence and so
cannot be used directly.
Taking the Mellin transform of the hyper-$g$ $r$-prior (\ref{ssd})
results in
\begin{align}
\mathcal{M}(p(r\,|\, \sigma,K,\mathcal{H}_{g}); s)
&= \frac{(a{-}2)}{2}\left(\frac{\sigma\sqrt{2}}{\sqrt{N}}\right)^{s{-}1}
\frac{\Gamma\left[(a{-}s{-}1)/2\right]\Gamma\left[(K{+}s{-}1)/2\right]\,\Gamma\left[1{+}(s/2)\right]}
{\Gamma\left[a/2\right]\Gamma\left[K/2\right]}.
\end{align}
The case $a=3$ is remarkably similar to the above Zellner-Siow case
and in a sense shows that the hyper-$g$ is trying to emulate the
Zellner-Siow behaviour. Based on the above considerations, we propose
to use an $r$-prior with a similar pole structure in its Mellin
transform
\begin{align}
\mathcal{M}(p(r\,|\, \sigma,K,\mathcal{H}_r); s) %
&= \left(\frac{\sigma}{\sqrt{2 N}}\right)^{s{-}1}
\frac{\Gamma\left[1{-}(s/2)\right]\,\Gamma\left[K{+}s{-}1\right]}{\sqrt{\pi}\,\Gamma\left[K\right]},
\end{align}
which on inversion gives us a prior in the form of a simple confluent
hypergeometric function,
\begin{align}
p(r \,|\, \sigma,K,\mathcal{H}_r) \label{parabolic}
&= \frac{K}{r\sqrt{\pi}}\left(\frac{Nr^2}{2\sigma^2}\right)^{K/2} %
U\left[\frac{K{+}1}{2};\frac{1}{2};\frac{Nr^2}{2\sigma^2}\right],
\end{align}
which can also be written as a parabolic cylinder function and which
we therefore call the parabolic $r$-prior. It is of the same family as
the hyper-$g$ prior and can be reproduced by using the $g$-prior
\begin{align}
p(g\,|\,\sigma,K,\mathcal{H}_r)
&= \frac{\Gamma\left[1+(K/2)\right]}{\sqrt{\pi}\,\Gamma\left[\frac{1+K}{2}\right]}\frac{g^{(K-1)/2}}{(1+g)^{K/2+1}}.
\end{align}
In both cases, the resulting evidence is
\begin{align}
p({\bm{y}}\,|\,\mathcal{H}_r,\mathcal{M})
= \frac{\Gamma(N/2)}{2^{K+1}}\left(N \pi {\langle \bmy^2 \rangle} \right)^{-N/2}
{}_2F_1\!\left( \frac{K+1}{2}\,;\, \frac{N}{2} \,;\, K{+}1 \,;\, \frac{\bm{\hat{b}}^2}{{\langle \bmy^2 \rangle}}\right).
\end{align}
The posterior and its characteristic function are easily derived,
given the closed forms for the evidence.
\section{Known error variance}
\label{sec:fxr}
\subsection{Definition and diagonalisation}
\label{sec:fxsdiag}
In this section, we change the information from a single variable
$\sigma$ to a set of widths ${\bm{\sigma}} = \{\sigma_n\}_{n=1}^N$ assumed
to be known constants, $\mathcal{H}_1 = \{{\bm{c}},{\bm{\sigma}},N\}$, so that
the Gaussian error distribution becomes
\begin{align}
\label{xrd}
p({\bm{\varepsilon}}\,|\, \mathcal{H}_1) %
&= \prod_{n=1}^N
\frac{e^{-\varepsilon_n^2/2\sigma_n^2}}{\sigma_n\sqrt{2\pi}}.
\end{align}
The data and predictors are now scaled individually by $\sigma_n$,
\begin{align}
\label{xrg}
{\bm{z}} &= \left(\frac{y_1}{\sigma_1},\ldots,\frac{y_{\!\scriptscriptstyle N}}{\sigma_{\!\scriptscriptstyle N}}\right)^T\\
{\bm{X}}_k &= \left(\frac{X_k(c_1)}{\sigma_1},\ldots,\frac{X_k(c_{\!\scriptscriptstyle N})}{\sigma_{\!\scriptscriptstyle N}}\right)^T\end{align}
with ${\mathbb{X}} = ({\bm{X}}_1\,\cdots\,{\bm{X}}_{\!\scriptscriptstyle K})$. The joint likelihood is
\begin{align}
\label{xre}
p({\bm{y}}\,|\,{\bm{\beta}},\mathcal{M})
&= C_\sigma\,e^{-NQ/2},
\end{align}
with $C_\sigma = \bigl[\prod_n 2\pi \sigma_n^2\bigr]^{-1/2}$ a
model-independent constant and $NQ = \chi^2$ given by
\begin{align}
\label{xrf}
Q({\bm{\beta}},{\bm{z}}\,|\, \mathcal{M})
&= \frac{1}{N}\tnorm{{\bm{z}} - {\mathbb{X}}{\bm{\beta}}}^2
= \frac{1}{N}\sum_n
\left(\frac{y_n-{\textstyle}\sum_k X_k(c_n)\beta_k}{\sigma_n}\right)^2.
\end{align}
Defining ${\langle \bmz^2 \rangle} = {\bm{z}}^T {\bm{z}} / N$, ${\mathbb{H}} = {\mathbb{X}}^T {\mathbb{X}} / N$ and
${\bm{h}} = {\mathbb{X}}^T{\bm{z}}/N$, we obtain
\begin{align}
\label{xrk}
Q &= {\langle \bmz^2 \rangle} + {\bm{\beta}}^T {\mathbb{H}} {\bm{\beta}} - 2 {\bm{h}}^T {\bm{\beta}}.
\end{align}
The likelihood mode in ${\mathcal{A}}({\bm{\beta}})$ is
\begin{align}
\label{xrl}
\bm{\hat{\beta}} &= {\mathbb{H}}^{-1}\,{\bm{h}} = ({\mathbb{X}}^T{\mathbb{X}})^{-1}{\mathbb{X}}^T {\bm{z}},
\end{align}
and $\bm{\hat{b}} = {\mathbb{L}^{\scriptscriptstyle \! 1\! /\! 2}}\, {\mathbb{S}}^T \bm{\hat{\beta}}$ as before but of course
with changed ${\mathbb{L}}$. Diagonalisation proceeds with the same
equations as in Section \ref{sec:nkrdiag} but subject to the above
changed definitions. We again end up with $Q = Q_0 + R_{\bm{\hat{b}}}^2$,
with minimum
\begin{align}
\label{xrn}
Q_0 &= {\langle \bmz^2 \rangle} - \bm{\hat{\beta}}^T {\mathbb{H}} \bm{\hat{\beta}}
= {\langle \bmz^2 \rangle} - \bm{\hat{b}}^2,
\end{align}
while $R_{\bm{\hat{b}}}^2 = ({\bm{\beta}} - \bm{\hat{\beta}})^T {\mathbb{H}} ({\bm{\beta}} -
\bm{\hat{\beta}})$ as before, and the likelihood itself is
\begin{align}
p({\bm{y}}\,|\, {\bm{b}},\mathcal{M}) %
&= C_\sigma \exp\left[ -\frac{N}{2}\left(Q_0 + R_{\bm{\hat{b}}}^2\right) \right]
\nonumber\\
\label{xrp}
&= C_\sigma\,e^{-N{\langle \bmz^2 \rangle}/2} \exp\left[ - \frac{N}{2} {\bm{b}}^2
+ N \bm{\hat{b}}^T {\bm{b}}\right]
\end{align}
and the evidence for fixed $r$ changes from Eq.~(\ref{prk}) to
\begin{align}
\label{xrc}
p({\bm{y}}\,|\, r,\mathcal{M}) &=
C_\sigma\, e^{-N({\langle \bmz^2 \rangle}+r^2)/2}
\;{}_0F_1\!\left(\frac{K}{2}\,;\,\frac{N^2\bm{\hat{b}}^2 r^2}{4} \right).
\end{align}
\subsection{Results for different $r$-priors}
\label{sec:gmpf}
Since ${\bm{\sigma}}$ is fixed, the parabolic $r$-prior becomes
\begin{align}
p(r\,|\, K,\mathcal{H}'_r)
&= \frac{K}{r\sqrt{\pi}}\left(\frac{Nr^2}{2}\right)^{K/2} U\left[\frac{K+1}{2};\frac{1}{2};\frac{Nr^2}{2}\right],
\end{align}
and the resulting evidence is
\begin{align}
\label{xrt}
p({\bm{y}}\,|\,\mathcal{H}'_r,\mathcal{M})
&= \int d{\bm{b}}\,dr\,p({\bm{y}}\,|\, {\bm{b}}, \mathcal{M})
\,p({\bm{b}}\,|\, r,\mathcal{M})\,p(r\,|\, \mathcal{H}'_r)
\\
&= C_{\sigma} 2^{-K}e^{-N{\langle \bmz^2 \rangle}/2} {}_1F_1\left(\frac{K+1}{2}\,;\,K+1\,;\, \frac{N \bm{\hat{b}}^2}{2} \right).
\nonumber
\end{align}
For comparison, the corresponding evidence expressions for the
hyper-$g$ and Zellner-Siow priors with their $\sigma$ set to 1 are,
respectively,
\begin{align}
\label{xrx}
p({\bm{y}}\,|\,\mathcal{H}_g,\mathcal{M}) &= C_\sigma \, e^{-N {\langle \bmz^2 \rangle} / 2} \,
\frac{(a-2)}{(K+a-2)}\;{}_1F_{1}\!\left( 1\,;\, \frac{K{+}a}{2} \,;\,
\frac{N \bm{\hat{b}}^2}{2} \right)
\\
\label{xry}
p({\bm{y}}\,|\,\mathcal{H}_{{\!\scriptscriptstyle ZS}},\mathcal{M})
&= \frac{C_\sigma \, e^{-N {\langle \bmz^2 \rangle} / 2}}{\sqrt{\pi}}
\Gamma\!\left(\frac{K{+}1}{2}\right) \sum_{j=0}^\infty\frac{1}{j!}
\left(\frac{N^2\bm{\hat{b}}^2}{4}\right)^{\!j}
\;U\!\left( \frac{K}{2}{+}j \,;\, \frac{1}{2}{+}j \,;\, \frac{N}{2}\right).
\end{align}
\subsection{Asymptotic forms}
\label{sec:asm}
As the argument $z$ of all the hypergeometric functions grows with
$N$, the asymptotic form for $z \gg 1$ according to
\cite{Bateman1953}
\begin{align}
\label{asmc}
{}_1F_{1}\!\left(a; c; z\right) &\simeq
\frac{\Gamma(c)}{\Gamma(a)}\, z^{a-c}\, e^z
\end{align}
will often suffice. The evidence based on the parabolic $r$-prior
Eq.~(\ref{parabolic}) becomes
\begin{align}
\label{xrsl}
p({\bm{y}}\,|\,\mathcal{H}'_r,\mathcal{M}) &\simeq \frac{C_\sigma}{\sqrt{\pi}}
\Gamma\!\left(\frac{K}{2}+1\right) \left(\frac{2}{N}\right)^{(K+1)/2}
\frac{e^{-NQ_0/2}}{\tnorm{\bm{\hat{b}}}^{K+1}}.
\end{align}
We also find the asymptotic form of the evidence for the hyper-$g$
prior (\ref{xrx}) to be
\begin{align}
\label{xrxl}
p({\bm{y}}\,|\,\mathcal{H}_g,\mathcal{M}) &=
\frac{C_\sigma (a{-}2)}{(K{+}a{-}2)} \;
\Gamma\!\left(\frac{K{+}a}{2}\right)
\left(\frac{2}{N}\right)^{(K+a)/2-1}
\frac{e^{-NQ_0/2}}{\tnorm{\bm{\hat{b}}}^{K+a-2}},
\end{align}
and with the help of
\begin{align}
U\!\left( \frac{K}{2}{+}j \,;\, \frac{1}{2}{+}j \,;\, \frac{N}{2}\right)
&= \left(\frac{2}{N}\right)^{K/2+j}
\;{}_2F_0\!\left(\frac{K}{2}{+}j \,;\, \frac{1}{2}{+}\frac{K}{2} \,;\,
\frac{-2}{N}\right)
\simeq \left(\frac{2}{N}\right)^{(K/2)+j},
\end{align}
approximate the Zellner-Siow evidence (\ref{xry}) by
\begin{align}
\label{xryl}
p({\bm{y}}\,|\,\mathcal{H}_{{\!\scriptscriptstyle ZS}},\mathcal{M})
&\simeq C_\sigma \;
\Gamma\!\left(\frac{K{+}1}{2}\right) \left(\frac{2}{N}\right)^{K/2}
e^{-NQ_0/2}.
\end{align}
Of course the asymptotic forms are not exactly normalised, so that we
can use them only for model comparison with information criteria or in
ratios such as Bayes Factors.
\section{Comparing model comparison schemes}
\label{sec:bfc}
Given the closed-form expressions for the evidence within each of the
different approaches, model comparison using Bayes Factors can, of
course, be effected simply by insertion of the relevant expression
into Eq.~(\ref{nte}). We shall not do so here, however, but rather
address by example the more general question as to which of the model
comparison schemes works best. In addition to the model schemes
$\mathcal{H}_r$, $\mathcal{H}_g$ and $\mathcal{H}_{{\!\scriptscriptstyle ZS}}$ considered so far,
we include several schemes that have been used in the literature, namely %
$\mathcal{H}_{\mathrm{AIC}}$, the Akaike Information Criterion of
\cite{Akaike1974}, %
$\mathcal{H}_{\mathrm{BIC}}$, the Bayesian Information Criterion of
\cite{Schwarz1978} and %
$\mathcal{H}_{\mathrm{AICc}}$, the Akaike Information Criterion as corrected
by \cite{Hurvich1989}. All of these can be shown to be equivalent to
$-2\,\log p({\bm{y}}\,|\,\mathcal{H})$ in our notation. For easier comparison,
we list in Table 1 the different schemes together with the
$-2\log p({\bm{y}}\,|\,\mathcal{H})$ versions of the asymptotic forms
(\ref{xrsl})--(\ref{xryl}). In the second part of Table 1, the
corresponding asymptotic forms for the evidences of Sections
\ref{sec:zsg} and \ref{sec:gmp} are shown using the relation $
{}_2F_1(a; b; c; z) \simeq \frac{\Gamma(c)}{\Gamma(a)} (bz)^{a-c}
e^{bz}$. $K$-independent constants have been omitted since they
cancel anyway once one does model comparison within any one scheme.
\begin{table}[htb]
\begin{center}
\begin{tabular}{|l|l|}
\hline
Scheme & \hspace*{15ex} $-2\log p({\bm{y}}\,|\,\mathcal{H})$ for fixed ${\bm{\sigma}}$ \\ \hline \hline
$\mathcal{H}'_r$ & $\displaystyle
N Q_0 + \left(K+1\right)\log \!\left(\frac{N\bm{\hat{b}}^2}{2}\right) -2\log \Gamma\!\left(\frac{K}{2}+1\right)$\\
$\mathcal{H}_g$ & $\displaystyle
N Q_0 + (K{+}a{-}2)\log\left(\frac{N}{2}\bm{\hat{b}}^2\right)
- 2 \log\Gamma\!\left(\frac{K{+}a{-}2}{2}\right)$ \\
$\mathcal{H}_{{\!\scriptscriptstyle ZS}}$ & $\displaystyle
N Q_0 + K\log\left(\frac{N}{2}\right) - 2 \log\Gamma\!\left(\frac{K{+}1}{2}\right)$ \\
$\mathcal{H}_{\mathrm{AIC}}$ & $N Q_0 + 2K$ \\
$\mathcal{H}_{\mathrm{AICc}}$ & $\displaystyle N Q_0 + 2 K + \frac{2K(K+1)}{N-K-1}$ \\
$\mathcal{H}_{\mathrm{BIC}}$ & $N Q_0 + K\log N $ \\
\hline
Scheme & \hspace*{15ex} $-2\log p({\bm{y}}\,|\,\mathcal{H})$ for variable $\sigma$\\ \hline \hline
$\mathcal{H}_r$ & $\displaystyle
-\frac{N\bm{\hat{b}}^2}{{\langle \bmy^2 \rangle}} + (K{+}1) \log \!\left(\frac{N\bm{\hat{b}}^2}{2{\langle \bmy^2 \rangle}}\right)
-2\log \Gamma\!\left(\frac{K}{2}+1\right)$\\[6pt]
$\mathcal{H}_g$ & $\displaystyle
- \frac{N\bm{\hat{b}}^2}{{\langle \bmy^2 \rangle}} + (K{+}a{-}2)\log\left(\frac{N\bm{\hat{b}}^2}{2{\langle \bmy^2 \rangle}}\right)
- 2 \log\Gamma\!\left(\frac{K{+}a{-}2}{2}\right)$ \\[6pt]
$\mathcal{H}_{{\!\scriptscriptstyle ZS}}$ & $\displaystyle
N\log\!\left( 1 - \frac{\bm{\hat{b}}^2}{{\langle \bmy^2 \rangle}} \right)
+ K\log\left(\frac{N}{2}\right) - 2 \log\Gamma\!\left(\frac{K{+}1}{2}\right)$ \\[6pt]
\hline
\end{tabular}
\caption{Summary of model comparison schemes for the fixed
${\bm{\sigma}}$ case of Section \ref{sec:fxr} (upper part) and for
the variable $\sigma$ case of Section \ref{sec:nkr} (lower
part). Constants that do not depend on $K$ are neglected.}
\end{center}
\end{table}
In order to test our results and to make a fair comparison between
different schemes, we generate data with fixed ${\bm{\sigma}}$ according to
\begin{align}
{\bm{y}} = {\mathbb{X}}{\bm{\beta}} + {\bm{\varepsilon}}
\end{align}
where ${\bm{\beta}}$ is drawn from a Cauchy distribution centered at 0 with
its dispersion parameter set to $1$ and $5$ respectively to mimick
weak and strong signal cases. The error ${\bm{\varepsilon}}$ is drawn from a
standardised Gaussian distribution with a sample size $N=100$. The
design matrix is taken as orthogonal, ${\mathbb{X}}^T{\mathbb{X}} = \mathbb{I}_{16}$. We
use the asymptotic form of the Zellner-Siow evidence as the full form
is too slow computationally. The model size ranges successively from
$1$ to $16$ by including the first $K$ coefficients of ${\bm{\beta}}$ to
generate data $\bm{y}$ while setting the rest of the coefficients to
zero. We then calculate the highest posterior probability model using
the different priors and mean squared error loss between the fitted
and true data
\begin{align}
\mathrm{MSE}(K) = \tnorm{{\mathbb{X}}{\bm{\beta}}-{\mathbb{X}}\hat{{\bm{\beta}}}^{(K)}}^2,
\end{align}
averaged over 1000 simulations. Figure 1 shows the average MSE as a
function of model size $K$ and the model comparison schemes listed
in Table 1, including the ``Oracle'' which is the least squares
solution for the true model. To facilitate comparison, the difference
between a given method and the Oracle is shown separately in the lower
panels, while in Tables 2 and 3 the MSE values are listed for the weak
and strong signal case respectively.
We note firstly that there are large differences in the behaviour of
the model schemes for the weak and strong signal cases. At one
extreme, the BIC is quite bad for weak signals but outperforms all
other schemes for strong signals. The corrected Akaike scheme does
well for weak signals but is in mid-field for the strong signal case.
As is already apparent from the close mathematical correspondence
between the hyper-$g$ and parabolic-$r$ schemes, they converge for
large $K$ as they must. For small $K$, however, the parabolic-$r$
scheme is far superior to the hyper-$g$ scheme.
\begin{figure}
\includegraphics[width=0.50\linewidth]{WeakSignal.pdf}
\includegraphics[width=0.50\linewidth]{StrongSignal.pdf}\\
\includegraphics[width=0.50\linewidth]{WeakSignalDiff.pdf}
\includegraphics[width=0.50\linewidth]{StrongSignalDiff.pdf}
\caption{Upper panels: Comparison of MSE values for different model
comparison schemes as a function of model size $K$ for weak
signal on the left and strong signal on the right. The lower
panels show corresponding differences between MSE($K$) and
MSE(Oracle).}
\end{figure}
\begin{table}[!h]
\begin{center}
\renewcommand{\arraystretch}{1.1}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
$K$ & Oracle & AIC & AICc & BIC & $\mathcal{H}_{\!\scriptscriptstyle ZS}$ & $\mathcal{H}_g$ & $\mathcal{H}_r$ \\ \hline \hline
1 & 1.01 & 3.37 & 2.85 & 1.31 & 1.53 & 4.44 & 2.48 \\
2 & 1.91 & 3.91 & 3.33 & 2.17 & 2.44 & 4.49 & 3.09 \\
3 & 3.05 & 4.94 & 4.46 & 3.58 & 3.77 & 5.32 & 4.21 \\
4 & 4.04 & 5.82 & 5.10 & 4.78 & 4.95 & 6.10 & 5.00 \\
5 & 4.94 & 6.60 & 6.07 & 6.06 & 6.00 & 6.73 & 6.19 \\
6 & 5.96 & 7.73 & 7.13 & 7.55 & 7.30 & 7.74 & 7.24 \\
7 & 6.96 & 8.61 & 8.15 & 8.83 & 8.64 & 8.54 & 8.19 \\
8 & 7.86 & 9.37 & 8.96 & 10.2 & 9.63 & 9.58 & 9.22 \\
9 & 9.11 & 10.6 & 10.0 & 11.3 & 10.7 & 10.7 & 10.4 \\
10 & 10.0 & 11.5 & 11.1 & 12.6 & 11.7 & 11.8 & 11.6 \\
11 & 11.0 & 12.5 & 12.2 & 14.2 & 12.9 & 12.9 & 12.8 \\
12 & 12.3 & 13.7 & 13.4 & 15.4 & 13.9 & 14.2 & 14.1 \\
13 & 13.0 & 14.1 & 13.9 & 16.2 & 14.5 & 15.0 & 14.9 \\
14 & 14.2 & 15.2 & 15.3 & 17.8 & 15.6 & 16.3 & 16.3 \\
15 & 14.8 & 15.4 & 15.8 & 19.0 & 15.9 & 16.8 & 17.0 \\
16 & 16.3 & 16.7 & 17.2 & 20.5 & 17.2 & 18.8 & 18.9 \\ \hline
\end{tabular}
\caption{Comparison of MSE values for different model comparison
schemes as a function of model size $K$ for the weak signal
case.}
\end{center}
\end{table}
\begin{table}[!h]
\begin{center}
\renewcommand{\arraystretch}{1.1}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
$K$ & Oracle & AIC & AICc & BIC & $\mathcal{H}_{\!\scriptscriptstyle ZS}$ & $\mathcal{H}_g$ & $\mathcal{H}_r$ \\ \hline \hline
1 & 1.03 & 3.44 & 2.90 & 1.41 & 1.72 & 2.77 & 2.16 \\
2 & 2.02 & 4.17 & 3.60 & 2.59 & 2.87 & 3.14 & 2.85 \\
3 & 2.80 & 4.97 & 4.40 & 3.43 & 3.77 & 3.75 & 3.68 \\
4 & 3.85 & 6.32 & 5.56 & 4.80 & 5.28 & 4.92 & 4.93 \\
5 & 5.14 & 7.27 & 6.58 & 6.02 & 6.47 & 6.36 & 6.31 \\
6 & 5.93 & 8.39 & 7.70 & 6.96 & 7.64 & 7.31 & 7.15 \\
7 & 6.87 & 9.09 & 8.44 & 7.70 & 8.53 & 7.98 & 7.93 \\
8 & 8.11 & 10.1 & 9.42 & 8.97 & 9.72 & 9.30 & 9.25 \\
9 & 8.87 & 10.9 & 10.2 & 9.73 & 10.6 & 10.1 & 10.1 \\
10 & 10.1 & 11.9 & 11.3 & 11.1 & 11.8 & 11.3 & 11.3 \\
11 & 10.7 & 12.4 & 11.8 & 11.6 & 12.4 & 11.8 & 11.8 \\
12 & 12.1 & 13.6 & 13.1 & 12.9 & 13.6 & 13.3 & 13.3 \\
13 & 13.2 & 14.4 & 14.1 & 14.0 & 14.4 & 14.5 & 14.4 \\
14 & 13.7 & 14.9 & 14.6 & 14.6 & 14.9 & 15.0 & 14.9 \\
15 & 14.9 & 15.6 & 15.5 & 15.8 & 15.6 & 16.1 & 16.1 \\
16 & 15.9 & 16.0 & 16.1 & 16.4 & 16.0 & 17.2 & 17.2 \\ \hline
\end{tabular}
\caption{Comparison of MSE values for different model comparison
schemes as a function of model size $K$ for the strong signal
case.}
\end{center}
\end{table}
\section{Discussion}
\label{sec:dcn}
We have introduced in this article the $r$-prior based on explicit
enforcement of spherical symmetry on the diagonalised parameter
space. The resulting formalism has been shown to encompass the
currently popular Zellner $g$-prior, Zellner-Siow Cauchy prior and the
hyper-$g$ prior as special cases. Beyond these, we have shown by
example of a new parabolic $r$-prior how different considerations such
as asymptotic behaviour may be incorporated. Other $r$-priors based on
further and different information can presumably be implemented in
future.
Conceptually, the $r$-priors appear to be a step towards a more formal
understanding of the symmetries on the hypersphere which are implicit
in canonical regression problems. The next step would be to understand
the scale symmetry governing $r$ itself.
The simulation shows that the $r$-prior gives good results, but also
that the detailed behaviour of it and other model schemes is quite
variable and poorly understood. Both the type of simulation and the
comparison criterion must in future be investigated in some detail.
\\
\noindent
\textbf{Acknowledgements}: This work is supported in part by a
Consolidoc fellowship of Stellenbosch University and by the National
Research Foundation of South Africa. We thank the referee for useful
comments and suggestions. Thanks also to the organisers of the
\textit{2014 ISBA--George Box Research Workshop on Frontiers of
Statistics} for support and the participants for helpful
discussions.
\\
|
\section{Introduction}
Deep learning is an aspect of machine learning that regards the question of
learning multiple levels of representation, associated with different
levels of abstraction~\citep{Bengio-2009-book}. These representations are
distributed~\citep{Hinton89b}, meaning that at each level there are many
variables or features, which together can take a very large number of
configurations.
An important conceptual challenge of deep learning is the following
question: {\em what is a good representation}? The question is most
challenging in the unsupervised learning setup. Whereas we understand
that features of an input $x$ that are predictive of some target $y$
constitute a good representation in a supervised learning setting,
the question is less obvious for unsupervised learning.
\subsection{Manifold Unfolding}
In this paper we explore
this question by following the geometrical inspiration introduced
by~\citet{Bengio-arxiv2014}, based on the notion of {\em manifold unfolding},
illustrated in Figure~\ref{fig:ddga}. It was already observed
by~\citet{Bengio-et-al-ICML2013} that representations obtained
by stacking denoising autoencoders or RBMs appear to yield ``flatter'' or
``unfolded'' manifolds:
if $x_1$ and $x_2$ are examples from the data generating distribution $Q(X)$ and $f$ is
the encoding function and $g$ the decoding function,
then points on the line $h_\alpha = \alpha f(x_1) + (1-\alpha) f(x_2)$ ($\alpha \in [0,1]$)
were experimentally found to correspond to probable input configurations, i.e., $g(h_\alpha)$ looks like
training examples (and quantitatively often comes close to one). This property
is not at all observed for $f$ and $g$ being the identity function: interpolating
in input space typically gives rise to non-natural looking inputs (we can immediately
recognize such inputs as the simple addition of two plausible examples).
This is illustrated in Figure~\ref{fig:flattening-manifold}.
It means that the input manifold (near which the distribution concentrates)
is highly twisted and curved and occupies a small volume in input space.
Instead, when mapped in the representation space of stacked autoencoders (the output of $f$),
we find that
the convex between high probability points (i.e., training examples) is often
also part of the high-probability manifold, i.e., the transformed manifold
is flatter, it has become closer to a convex set.
\begin{figure}[H]
\begin{center}
\centerline{\includegraphics[width=\columnwidth]{flattening-manifold.pdf}}
\caption{Illustration of the flattening effect observed in~\citet{Bengio-et-al-ICML2013}
by stacks of denoising autoencoders or RBMs, trained on MNIST digit images.
Whereas interpolating in pixel space ($X$-space)
between dataset examples (such as the 9 on the bottom
left and the 3 on the bottom right) gives rise to images unlike those in
the training set (on the bottom interpolation line), interpolating at the
first and second level of the stack of autoencoders ($H$-space) gives rise to images
(when projected back in input space, see text) that look like dataset examples.
These experiments suggest that the manifolds near which data concentrate, which are very
twisted and occupy a very small volume in pixel space, become flatter
and occupy more of the available volume in representation-space. Note
in particular how the two class manifolds have been brought closer to each other
(but interestingly there are also easier to separate, in representation space).
The manifolds associated with each class have become closer to a convex set, i.e., flatter.
}
\label{fig:flattening-manifold}
\end{center}
\end{figure}
\subsection{From Manifold Unfolding to Probability Modeling}
If it is possible to unfold the data manifold into a nearly
flat and convex manifold (or set of manifolds), then estimating
the probability distribution of the data becomes much easier.
Consider the situation illustrated in Figure~\ref{fig:ddga}: a highly
curved 1-dimensional low-dimensional manifold is unfolded so that it occupies
exactly one dimension in the transformed representation (the ``signal dimension''
in the figure). Moving on the manifold corresponds to changing the hidden unit corresponding to
that signal dimension in representation space. On the other hand, moving orthogonal
to the manifold in input space corresponds to changing the ``noise dimension''.
There is a value of the noise dimension that corresponds to being on the
manifold, while the other values correspond to the volume filled by
unlikely input configurations. With the manifold learning mental picture,
estimating the probability distribution of the data basically amounts
to distinguishing between ``off-manifold'' configurations, which should have
low probability, from ``on-manifold'' configurations, which should have
high probability. Once the manifold is unfolded, answering that question
becomes very easy.
\begin{figure}[ht]
\begin{center}
\centerline{\includegraphics[width=0.35\columnwidth]{flattening.pdf}}
\caption{Illustration of the work that the composed encoder $f = f_L \circ \ldots f_2 \ldots f_1$
should do: flatten the manifold (more generally the region where probability mass is
concentrated) in such a way that a simple (e.g. factorized) prior distribution $P(h)$ can well approximate
the distribution of the transformed data $Q(h)$ obtained by applying $f$ to the data $x\sim Q(x)$.
The red curve indicates the manifold, or region near which $Q(x)$ concentrates. The first
encoder $f_1$ is not powerful and non-linear enough to untwist the manifold and flatten
it, so applying a factorized prior at that level would yield poor samples, because most
generated samples from the prior (in orange) would fall far from the transformed data
manifold (in red). At the top level, if training is successful, the transformed data distribution $Q(h)$
concentrates in a small number of directions (the ``signal'' directions, in the figure),
making it easy to distinguish the inside of the manifold (signal) from the outside (noise),
i.e. to concentrate probability mass in the right places.
At that point, $Q(h)$ and $P(h)$ can match better, and less probability mass is wasted
outside of the manifold.
}
\label{fig:ddga}
\end{center}
\end{figure}
If probability mass is concentrated in a predictible and regular way
in representation space ($H$), it becomes easy to capture the data distribution $Q(X)$.
Let $H=f(X)$ be the random variable associated with the transformed data $X \sim Q(X)$,
with $Q(H)$ its marginal distribution.
Now, consider for example the extreme case where $Q(H)$ is almost factorized
and can be well approximated by a factorized distribution model $P(H)$.
It means that the probability distribution in $H$-space can be captured
with few parameters and generalization is easy. Instead, when
modeling data in $X$-space, it is very difficult to find a family of
probability distributions that will put mass where the examples are
but not elsewhere. What typically happens is that we end up choosing
parameters of our probability model such that it puts a lot more
probability mass outside of the manifold than we would like. The
probability function is thus ``smeared''. We clearly get that effect
with non-parametric kernel density estimation, where we see that
the kernel bandwidth controls the amount of smearing, or smoothing.
Unfortunately, in a high-dimensional space, this puts a lot more
probability mass outside of the manifold than inside. Hence what
we see the real challenge as finding an encoder $f$ that transforms
the data distribution $Q(X)$ from a complex and twisted one into
a flat one, $Q(H)$, where the elements $H_i$ of $H$ are easy
to model, e.g. they are independent.
\section{Directed Generative Autoencoder (DGA) for Discrete Data}
We mainly consider in this paper the case of a discrete variable, which is simpler
to handle.
In that case, a Directed Generative Autoencoder (DGA) is a model over the random variable $X$
whose training criterion is as follows
\begin{equation}
\label{eq:dga}
\log P(X=x | H=f(x)) + \log P(H=f(X))
\end{equation}
where the deterministic function $f(\cdot)$ is called the encoder
and the conditional distribution $P(x|h)$ is called the decoder
or decoding distribution. As shown below (Proposition~\ref{prop:exact-dga}),
this decomposition becomes
exact as the capacity of the decoder $P(x|h)$ increases sufficiently to capture
a conditional distribution that puts zero probability on any $x'$ for which $f(x')\neq h$.
The DGA is parametrized from two components:
\begin{enumerate}
\item $P(X=x | H=f(x))$ is an autoencoder with a {\em discrete}
representation $h=f(x)$. Its role in the log-likelihood
training objective is to make sure that $f$ preserves as much
information about $x$ as possible.
\item $P(H=h)$ is a probability model of the samples $H=f(X)$
obtained by sending $X$ through $f$. Its role in the log-likelihood
training objective is to make sure that $f$ transforms $X$
in a representation that has a distribution that can be well
modeled by the family of distributions $P(H)$.
\end{enumerate}
What can we say about this training criterion?
\begin{proposition}
\label{prop:exact-dga}
There exists a decomposition of the likelihood into \mbox{$P(X=x) = P(X=x | H=f(x)) P(H=f(X))$}
that is exact for discrete $X$ and $H$, and it is achieved when $P(x |h)$ is zero
unless $h=f(x)$. When $P(x|h)$ is trained with pairs $(X=x,H=f(x))$, it estimates
and converges (with enough capacity) to a conditional probability distribution which
satisfies this contraint. When $P(x|h)$ does not satisfy the condition, then
the unnormalized estimator \mbox{$P^*(X=x) = P(X=x|H=f(x))P(H=f(X))$} underestimates the true probability
\mbox{$P(X=x) = \sum_h P(x|H=h) P(H=h)$}.
\end{proposition}
\begin{proof}
We start by observing that because $f(\cdot)$ is deterministic, its value
$f(x)$ is perfectly predictible from the knowledge of $x$, i.e.,
\begin{equation}
P(H=f(x) | X=x) = 1.
\end{equation}
Therefore, we can multiply this value of 1 by $P(X)$ and obtain the joint $P(X,H)$:
\begin{equation}
P(X=x) = P(H=f(x) | X=x) P(X=x) = P(X=x, H=f(x))
\end{equation}
for any value of $x$.
Now it means that there exists a parametrization of the joint $P(X,H)$
into $P(H)P(X|H)$ that achieves
\begin{equation}
P(X=x) = P(X=x | H=f(X)) P(H=f(X))
\end{equation}
which is the first part of the claimed result. Furthermore, for exact relationship
with the likelihood is achieved
when $P(x|h)=0$ unless $h=f(x)$, since with that condition,
\[
P(X=x) = \sum_h P(X=x|H=h) P(H=h) = P(X=x|H=f(x)) P(H=f(x))
\]
because all the other terms of the sum vanish. Now, consider the case (which
is true for the DGA criterion) where the parameters
of $P(x|h)$ are only estimated to maximize the expected value of the
conditional likelihood $\log P(X=x | H=f(x))$. Because the maximum
likelihood estimator is consistent, if the family of distributions used
to estimate $P(x|h)$ has enough capacity to contain a solution
for which the condition of $P(x|h)=0$ for $h \neq f(x)$ is
satisfied, then we can see that with enough capacity (which may also
mean enough training time), $P(x|h)$ converges to a solution that satisfies
the condition, i.e., for which $P(x) = P(x|H=f(x))P(H=f(x))$.
In general, a learned decoder $P(x|h)$ will not achieve
the guarantee that $P(x|h)=0$ when $h \neq f(x)$. However, the correct
$P(x)$, for given $P(x|h)$ and $P(h)$ can always be written
\[
P(x) = \sum_h P(x|h) P(h) \geq P(x|h=f(x)) P(H=f(x)) = P^*(x)
\]
which proves the claim that $P^*(x)$ underestimates the true likelihood.
\end{proof}
What is particularly interesting about this bound is that {\em as training
progresses and capacity increases} the bound becomes tight.
However, if $P(x|h)$ is a parametric distribution (e.g. factorized Binomial,
in our experiments) whose parameters (e.g. the Binomial probabilities)
are the output of a neural net, increasing the capacity of the neural net may not be sufficient
to obtain a $P(x|h)$ that satisfies the desired condition and makes the bound tight.
However, if $f(x)$ does not lose information about $x$, then $P(x|f(x))$ should
be unimodal and in fact be just 1. In other words, we should penalize the
reconstruction error term strongly enough to make the reconstruction error
nearly zero. That will guarantee that $f(x)$ keeps all the information about $x$
(at least for $x$ in the training set) and that the optimal $P(x|h)$ fits our
parametric family.
\iffalse
We clearly do not have access that ${\cal P}$, but we can use the estimated $P(x|h)$
in at least two ways to recover a normalized estimator for $P(x)$. One option is
to consider
\[
P^*(x) = P(x|H=f(x)) P(H=f(x))
\]
as an unnormalized probability function, i.e., define
\[
P(x) = \frac{P^*(x)}{Z}.
\]
Then one has to estimate the associated
normalization constant or partition function $Z$, which could be done in various ways,
e.g. by importance sampling with a proposal distribution $\pi(x)$:
\[
Z = E_{\pi(x)}[ \frac{P*(x)}{\pi(x)} ].
\]
Another option is to take advantage of the knowledge of $f(x)$ to construct
a proposal distribution $q(h|x)$ which basically adds noise around $f(x)$
and estimates $P(h|x)$, and then use
another importance sampling estimator, this time for $P(x)$:
\[
P(x) = \sum_h P(x|h)P(h) = E_{h\sim q(h|x)}[ \frac{P(x|h)P(h)}{q(h|x)} ].
\]
\fi
\subsection{Parameters and Training}
The training objective is thus a lower bound on the true log-likelihood:
\begin{equation}
\label{eq:log-likelihood}
\log P(x) \geq \log P^*(x) = \log P(x|H=f(x)) + \log P(H=f(x))
\end{equation}
which can be seen to have three kinds of parameters:
\begin{enumerate}
\item The encoder parameters associated with $f$.
\item The decoder parameters associated with $P(x|h)$.
\item The prior parameters associated with $P(h)$.
\end{enumerate}
Let us consider how these three sets of parameters could be optimized
with respect to the log-likelihood bound (Eq.~\ref{eq:log-likelihood}).
The parameters of $P(h)$ can be learned by maximum likelihood (or any proxy for it),
with training examples that are the $h=f(x)$ when $x \sim Q(X)$
is from the given dataset. For example, if $P(H)$ is a factorized
Binomial, then the parameters are the probabilities \mbox{$p_i=P(H_i=1)$}
which can be learned by simple frequency counting of these events.
The parameters for $P(x|h)$ can be learned by maximizing the conditional
likelihood, like any neural network with a probabilistic output.
For example, if $X|H$ is a factorized Binomial, then we just have
a regular neural network taking $h=f(x)$ as input and $x$ as target,
with the cross-entropy loss function.
\subsection{Gradient Estimator for $f$}
One challenge is to deal with the training of the parameters of the
encoder $f$, because $f$ is discrete. We need to estimate a gradient
on these parameters in order to both optimize $f$ with respect to $\log P(h=f(x))$
(we want $f$ to produce outputs that can easily be modeled by the
family of distributions $P(H)$) and with respect to $\log P(x|h=f(x))$
(we want $f$ to keep all the information about $x$, so as to be able
to reconstruct $x$ from $h=f(x)$). Although the true gradient
is zero, we need to obtain an update direction (a pseudo-gradient)
to optimize the parameters of the encoder with respect to these
two costs. A similar question was raised in a different context
in~\citet{Bengio-arxiv2013,bengio2013estimating}, and a number
of possible update directions were compared. In our experiments
we considered the special case where
\[
f_i(x) = 1_{a_i(x)>0}
\]
where $a_i$ is the activation of the $i$-th output unit of the encoder before
the discretizing non-linearity is applied.
The actual encoder output is discretized, but we are interested in obtaining
a ``pseudo-gradient'' for $a_i$, which we will back-propagate inside
the encoder to update the encoder parameters.
What we did in the experiments is to compute the derivative of the reconstruction loss
and prior loss \begin{equation}
\label{eq:loss}
{\cal L}=-\log P(x|h=f(x)) - \log P(h=f(x)),
\end{equation}
with respect to $f(x)$,
as if $f(x)$ had been continuous-valued.
We then used the {\bf Straight-Through Pseudo-Gradient}:
the update direction (or pseudo-gradient)
for $a$ is just set to be equal to the gradient with respect to $f(x)$:
\[
\Delta a = \frac{\partial {\cal L}}{\partial f(x)}.
\]
The idea for this technique was proposed by~\citet{Hinton-Coursera2012} and
was used very succesfully in~\citet{bengio2013estimating}. It clearly has
the right sign (per value of $a_i(x)$, but not necessarily overall) but does not
take into account the magnitude of $a_i(x)$ explicitly.
Let us see how the prior negative log-likelihood bound can be written as a function of $f(x)$
in which we can pretend that $f(x)$ is continuous. For example, if
$P(H)$ is a factorized Binomial, we write this negative log-likelihood as the
usual cross-entropy:
\[
- \sum_i f_i(x) \log P(h_i=1) + (1-f_i(x)) \log (1-P(h_i=1))
\]
so that if $P(x|h)$ is also a factorized Binomial, the overall loss is
\begin{align}
\label{eq:binomial-loss}
{\cal L} = & - \sum_i f_i(x) \log P(h_i=1) + (1-f_i(x)) \log (1-P(h_i=1)) \nonumber \\
& - \sum_j x_j \log P(x_j=1|h=f(x)) + (1-x_j) \log (1-P(x_j=1|h=f(x)))
\end{align}
and we can compute $\frac{\partial {\cal L}}{\partial f_i(x)}$ as if $f_i(x)$ had been
a continuous-valued variable. All this is summarized in Algorithm~\ref{alg:dga}.
\begin{algorithm}[ht]
\caption{Training procedure for Directed Generative Autoencoder (DGA).}
\label{alg:dga}
\begin{algorithmic}
\STATE $\bullet$ Sample $x$ from training set.
\STATE $\bullet$ Encode it via feedforward network $a(x)$ and $h_i=f_i(x)=1_{a_i(x)>0}$.
\STATE $\bullet$ Update $P(h)$ with respect to training example $h$.
\STATE $\bullet$ Decode via decoder network estimating $P(x|h)$.
\STATE $\bullet$ Compute (and average) loss ${\cal L}$, as per Eq.~\ref{eq:loss},
e.g., in the case of factorized Binomials as per Eq.~\ref{eq:binomial-loss}.
\STATE $\bullet$ Update decoder in direction of gradient of $-\log P(x|h)$ w.r.t. the decoder parameters.
\STATE $\bullet$ Compute gradient $\frac{\partial \cal L}{\partial f(x)}$ as if $f(x)$ had been continuous.
\STATE $\bullet$ Compute pseudo-gradient w.r.t. $a$ as $\Delta a = \frac{\partial \cal L}{\partial f(x)}$.
\STATE $\bullet$ Back-propagate the above pseudo-gradients (as if they were true gradients of the loss on $a(x)$)
inside encoder and update encoder parameters accordingly.
\end{algorithmic}
\end{algorithm}
\section{Greedy Annealed Pre-Training for a Deep DGA}
For the encoder to twist the data into a form that fits the prior $P(H)$,
we expect that a very strongly non-linear transformation will be required. As deep
autoencoders are notoriously difficult to train~\citep{martens2010hessian}, adding the extra constraint
of making the output of the encoder fit $P(H)$, e.g., factorial, was found
experimentally (and without surprise) to be difficult.
What we propose here is to use a an annealing (continuation method) and a greedy pre-training strategy,
similar to that previously proposed to train Deep Belief Networks
(from a stack of RBMs)~\citep{Hinton06-small} or deep autoencoders (from a stack of shallow
autoencoders)~\citep{Bengio-nips-2006-small,Hinton+Salakhutdinov-2006}.
\subsection{Annealed Training}
Since the loss function of Eq.\ref{eq:dga} is hard to optimize directly, we consider a generalization
of the loss function by adding trade-off parameters to the two terms in the loss function corresponding
to the reconstruction and prior cost. A zero weight for the prior cost makes the loss function
same as that of a standard autoencoder, which is a considerably easier optimization problem.
\begin{align}
\label{eq:relaxed-binomial-loss}
{\cal L} = & - \beta \sum_i f_i(x) \log P(h_i=1) + (1-f_i(x)) \log (1-P(h_i=1)) \nonumber \\
& - \sum_j x_j \log P(x_j=1|h=f(x)) + (1-x_j) \log (1-P(x_j=1|h=f(x))).
\end{align}
\subsubsection{Gradient Descent on Annealed Loss Function}
Training DGAs with fixed trade-off parameters is sometimes difficult because it is
much easier for gradient descent to perfectly optimize the prior cost by making $f$ map
all $x$ to a constant $h$. This may be a local optimum or a saddle point, escaping from which is
difficult by gradient descent.
Thus, we use the tradeoff paramters to make the model first learn perfect reconstruction,
by setting zero weight for the prior cost. $\beta$ is then gradually increased to 1.
The gradual increasing schedule for $\beta$ is also important, as any rapid growth in $\beta$'s value
causes the system to `forget' the reconstruction and prioritize only the prior cost.
A slow schedule thus ensures that the model learns to reconstruct as well as to fit the prior.
\subsubsection{Annealing in Deep DGA}
Above, we describe the usefulness of annealed training of a shallow DGA. A similar trick
is also useful when pretraining a deep DGA. We use different values for these tradeoff parameters
to control the degree of difficulty for each
pretraining stage. Initial stages have a high weight for reconstruction, and a low weight for
prior fitting, while the final stage has $\beta$ set to unity, giving back the
original loss function. Note that the lower-level DGAs can sacrifice on prior
fitting, but must make sure that the reconstruction is near-perfect, so that no information is lost.
Otherwise, the upper-level DGAs can never recover from that loss, which will show up in
the high entropy of $P(x|h)$ (both for the low-level decoder and for the global decoder).
\section{Relation to the Variational Autoencoder (VAE) and Reweighted Wake-Sleep (RWS)}
The DGA can be seen as a special case of the Variational Autoencoder
(VAE), with various versions introduced
by~\citet{Kingma+Welling-ICLR2014,Gregor-et-al-ICML2014,Mnih+Gregor-ICML2014,Rezende-et-al-arxiv2014},
and of the Reweighted Wake-Sleep (RWS)
algorithm~\citep{Bornschein+Bengio-arxiv2014-small}.
The main difference between the DGA and these models is that with the latter the
encoder is stochastic, i.e., outputs a sample from an encoding (or approximate inference)
distribution $Q(h|x)$ instead of $h=f(x)$. This basically gives rise to a training
criterion that is not the log-likelihood but a variational lower bound on it,
\begin{equation}
\label{eq:vae-criterion}
\log p(x) \geq E_{Q(h|x)}[ \log P(h) + \log P(x|h) - \log Q(h|x)].
\end{equation}
Besides the fact that $h$ is now sampled, we observe that the training criterion
has exactly the same first two terms as the DGA log-likelihood, but it also has
an extra term that attempts to maximize the conditional entropy of the encoder output,
i.e., encouraging the encoder to introduce noise, to the extent that it does not
hurt the two other terms too much. It will hurt them, but it will also help
the marginal distribution $Q(H)$ (averaged over the data distribution $Q(X)$)
to be closer to the prior $P(H)$, thus encouraging the decoder to contract
the ``noisy'' samples that could equally arise from the injected noise in $Q(h|x)$
or from the broad (generally factorized) $P(H)$ distribution.
\section{Experiments}
In this section, we provide empirical evidence for the feasibility of the proposed model,
and analyze the influence of various techniques on the performance of the model.
We used the binarized MNIST handwritten digits dataset. We used the same binarized
version of MNIST as \cite{Uria+al-ICML2014}, and also used the same training-validation-test
split.
We trained shallow DGAs with 1, 2 and 3 hidden layers, and deep DGAs composed of
2 and 3 shallow DGAs. The dimension of $H$ was chosen to be 500, as this is
considered sufficient for coding binarized MNIST. $P(H)$ is modeled as
a factorized Binomial distribution.
Parameters of the model were learnt using minibatch gradient descent, with minibatch
size of 100. Learning rates were chosen from {10.0, 1.0, 0.1}, and halved whenever
the average cost over an epoch increased. We did not use momentum or L1, L2 regularizers.
We used tanh activation functions in the hidden layers and sigmoid outputs.
While training, a small salt-and-pepper noise is added to the decoder input to make it
robust to inevitable mismatch between the encoder output $f(X)$ and samples from the prior, $h \sim P(H)$. Each bit of decoder input is selected with 1\% probability and changed to $0$ or $1$ randomly.
\begin{figure}[t]
\begin{center}
\begin{tabular}{c c}
\includegraphics[width=6.5cm]{samples8x15.png} & \includegraphics[width=6.5cm]{shallowsamples8x15.png} \\
(a) & (b)
\end{tabular}
\caption{(a) Samples generated from a 5-layer deep DGA, composed of 3 shallow DGAs
of 3 layers (1 hidden layer) each. The model was trained greedily by training the
first shallow DGA on the raw data and training each subsequent shallow DGA
on the output code of the previous DGA. (b) Sample generated from a 3-layer shallow DGA.}
\label{fig:samples}
\end{center}
\end{figure}
\subsection{Loglikelihood Estimator}
When the autoencoder is not perfect, i.e. the encoder and decoder are not
perfect inverses of each other, the loglikelihood estimates using
Eq. \ref{eq:binomial-loss} are biased estimates of the true loglikehood. We
treat these estimates as an unnormalized probability distribution, where
the partition function would be one when the autoencoder is perfectly
trained. In practice, we found that the partition function is less than
one. Thus, to compute the loglikehood for comparison, we estimate the
partition function of the model, which allows us to compute normalized
loglikelihood estimates. If $P^*(X)$ is the unnormalized probability
distribution, and $\pi(X)$ is another tractable distribution on $X$ from
which we can sample, the
partition function can be estimated by importance sampling:
$$Z = \sum_x P^*(x) = \sum_x \pi(x) \frac{P^*(x)}{\pi(x)} = E_{x \sim
\pi(x)}[\frac{P^*(x)}{\pi(x)}]$$
As proposal distribution $\pi(x)$, We took $N$ expected values $\mu_j = E[X | H_j]$ under the decoder distribution,
for $H_j \sim P(H)$, and use them as centroids for a mixture model,
$$\pi(x) = \frac{1}{N} \sum_{j=1}^N FactorizedBinomial(x; \mu_j).$$
Therefore,
$$log(P(X)) = log(P^*(X)) - log(Z)$$ gives us the estimated normalized loglikelihood.
\subsection{Performance of Shallow vs Deep DGA }
The 1-hidden-layer shallow DGA gave a loglikelihood estimate of -118.12 on the test set. The 5-layer deep DGA,
composed of two 3-layer shallow DGAs trained greedily, gave a test set loglikelihood estimate of -114.29. We observed
that the shallow DGA had better reconstructions than the deep DGA. The deep DGA sacrificed on the reconstructibility, but
was more successful in fitting to the factorized Binomial prior.
Qualitatively, we can observe in Figure~\ref{fig:samples} that samples from the deep DGA are much better than those from the shallow DGA.
The samples from the shallow DGA can be described as a mixture of incoherent MNIST features: although all the decoder units
have the necessary MNIST features, the output of the encoder does not match well the factorized Binomial of $P(H)$,
and so the decoder is not correctly mapping these unusual inputs $H$ from the Binomial prior to data-like samples.
The samples from the deep DGA are of much better quality. However, we can see that some of the samples are non-digits. This is due to the fact that the autoencoder had to sacrifice some reconstructibility to fit $H$ to the prior. The reconstructions
also have a small fraction of samples which are non-digits.
\subsection{Entropy, Sparsity and Factorizability}
We also compared the entropy of the encoder output and the raw data, under a factorized
Binomial model. The entropy, reported in Table~\ref{tab:entropy},
is measured with the logarithm base 2, so it counts
the number of bits necessary to encode the data under the simple factorized Binomial distribution.
A lower entropy under the factorized distribution means that fewer independent units are necessary to
encode each sample. It means that the probability mass has been moved from a highly complex
manifold which is hard to capture under a factorized model and thus requires many dimensions
to characterize to a manifold that is aligned with a smaller set of dimensions,
as in the cartoon of Figure~\ref{fig:ddga}. Practically this happens when many
of the hidden units take on a nearly constant value, i.e., the representation
becomes extremely ``sparse'' (there is no explicit preference for 0 or 1 for $h_i$ but one could
easily flip the sign of the weights of some $h_i$ in the output layer to make sure that 0 is
the frequent value and 1 the rare one). Table~\ref{tab:entropy} also contains a measure of sparsity of the
representations based on such a bit flip (so as to make 0 the most frequent value).
This flipping allows to count the average number of 1's (of rare bits) necessary to
represent each example, in average (third column of the table).
We can see from Table~\ref{tab:entropy}
that only one autoencoder is not sufficient to reduce the entropy down
to the lowest possible value. Addition of the second autoencoder reduces
the entropy by a significant amount.
Since the prior distribution is factorized, the encoder has to map data
samples with highly correlated dimensions, to a code with independent
dimensions. To measure this, we computed the Frobenius norm of the
off-diagonal entries of the correlation matrix of the data represented in its
raw form (Data) or at the outputs of the different encoders. See the 3rd columns
of Table~\ref{tab:entropy}. We see
that each autoencoder removes correlations from the data representation,
making it easier for a factorized distribution to model.
\begin{table}[H]
\begin{center}
\begin{tabular}{|c| c| c|c|}
\hline
Samples & Entropy & Avg \# active bits & $||Corr-diag(Corr)||_{F}$ \\
\hline
Data ($X$) & 297.6 & 102.1 & 63.5 \\
Output of $1^{st}$ encoder ($f_1(X)$) & 56.9 & 20.1 & 11.2\\
Output of $2^{nd}$ encoder ($f_2(f_1(X))$) & 47.6 & 17.4 & 9.4\\
\hline
\end{tabular}
\end{center}
\caption{Entropy, average number of active
bits (the number of rarely active bits, or 1's if 0 is the most frequent bit value, i.e. a measure of non-sparsity)
and inter-dimension correlations,
decrease as we encode into higher levels of representation,
making it easier to model it.}
\label{tab:entropy}
\end{table}
\iffalse
\begin{figure}[t]
\begin{center}
\begin{tabular}{|c| c|}
\hline
Model & Lowerbound of Loglikelihood \\
DGA-1hl & -117 \\
DDGA-5hl & -111 \\
\hline
\end{tabular}
\end{center}
\caption{}
\end{figure}
\fi
\iffalse
\section{NOT-CLEAR-ENOUGH-YET - Extension to Continuous Random Variables}
How could the DGA likelihood criterion be extended to continuous-valued data?
To avoid the hard constraint that $f$ be a bijection (that is perfectly invertible),
we consider here a formulation of maximum likelihood that stays in the world
of probabilities instead of going to densities.
We consider a small ball $B_\epsilon(x)$ of radius $\epsilon$ around $x$ and the probability
of falling into it. Following Eq.~\ref{eq:dga},
\begin{equation}
\label{eq:continuous-dga}
\log P(x \in B_\epsilon(x)) = \log P(x \in B_\epsilon | H \in f(B_\epsilon(x)))
+ \log P(H \in f(B_\epsilon(x)))
\end{equation}
When $\epsilon$ is small with respect to the variations of $f$...
\fi
\section{Conclusion}
We have introduced a novel probabilistic interpretation for autoencoders as generative models,
for which the training criterion is similar to that of regularized (e.g. sparse) autoencoders
and the sampling procedure is very simple (ancestral sampling, no MCMC). We showed that this
training criterin is a lower bound on the likelihood and that the bound becomes tight as
the decoder capacity (as an estimator of the conditional probability $P(x|h)$) increases.
Furthermore, if the encoder keeps all the information about the input $x$, then the optimal
$P(x|h=f(x))$ is unimodal, i.e., a simple neural network with a factorial output suffices.
Our experiments showed that minimizing the proposed criterion yielded good generated
samples, and that even better samples could be obtained by pre-training a stack of
such autoencoders, so long as the lower ones are constrained to have very low reconstruction
error. We also found that a continuation method in which the weight of the prior term
is only gradually increased yielded better results.
These experiments are most interesting because they reveal a picture of the representation
learning process that is in line with the idea of manifold unfolding and transformation from
a complex twisted region of high probability into one that is more regular (factorized)
and occupies a much smaller volume (small number of active dimensions). They also help
to understand the respective roles of reconstruction error and representation prior
in the training criterion and training process of such regularized auto-encoders.
\subsection*{Acknowledgments}
The authors would like to thank Laurent Dinh, Guillaume Alain and Kush Bhatia for fruitful discussions, and also the developers of Theano~\citep{bergstra+al:2010-scipy,Bastien-Theano-2012}. We acknowledge the support of the following agencies for
research funding and computing support: NSERC, Calcul Qu\'{e}bec, Compute Canada,
the Canada Research Chairs and CIFAR.
|
\section{\label{}}
\section{Introduction\label{intro}}
Nonlinear delay dynamics have been a particularly prolific area of research
in the field of photonic devices during the last 30 years
\cite{erneux2010laser}. A large variety of setups exhibiting optical or
electro-optical delayed feedback loops have been explored for novel
applications, but also as experimental tools for delay systems in general.
They have stimulated fruitful interactions with researchers working in
different fields by emphasizing specific delay-induced phenomena
\cite{erneux,bala,Stepan28032009,atay2010complex,just2010delayed,KalmarNagy01062010,lakshmanan2011dynamics,smith2011introduction}%
.\ Examples include different forms of oscillatory instabilities,
stabilization techniques using a delayed feedback, and synchronization
mechanisms for delay-coupled systems. Most of the current lasers used in
applications are semiconductor lasers (SLs), which are highly sensitive to
optical feedback \cite{soriano2013complex}. Here, the light coming from the
laser is reflected back to the laser after a substantial delay.\ Another popular delay system is an optoelectronic oscillator (OEO) \cite{Larger28092013,devgan2013review} that consists of a laser injecting its light into an optoelectronic loop. For
OEOs, the feedback exhibits a large delay because of a long optical fiber line in the
OEO closed-loop configuration. An OEO is capable of generating, within the same
optoelectronic cavity, either an ultra-low-jitter single-tone microwave
oscillation, as used in radar applications \cite{yanne:09}, or a broadband
chaotic carrier typically intended for physical data encryption in high bit
rate optical communications \cite{gastaud04,callan}. The OEO is a particularly attractive system because it allows quantitative comparisons between experiments and theory
\cite{PhysRevE.79.026208,Levy:09,weicker:12b}.
For systems exhibiting a Hopf bifurcation in the absence of delay, a feedback
with a large delay may lead to the coexistence of stable periodic solutions in
the vicinity of the first Hopf bifurcation point.\ This multi-rhythmicity was
predicted theoretically using a Hopf normal-form equation with a delayed
feedback \cite{PhysRevLett.96.220201}, where the bifurcation scenario is
similar to Eckhaus instability in spatially extended systems
\cite{Tuckerman199057}. The bandpass OEO without its optical fiber line admits
a Hopf bifurcation.\ In this paper, we investigate the stabilization of nearby Hopf
bifurcation branches in this regime.
We conduct here a systematic experimental and numerical study of an OEO
exhibiting a large delay. We show that an OEO
admits coexisting stable periodic square-waves. Depending on the feedback
phase, they are characterized by frequencies close to either $(1+2n)/(2\tau
_{D})$ $(n=0,1,2,...)$ or $n/\tau_{D}$ $(n=1,2,...)$ where $\tau_{D}$ is the
delay of the feedback loop. In order to induce these periodic solutions, we
inject a periodic electrical signal into the oscillator during the
initialization phase of the experiment and then observe the resulting dynamics
after the injected signal is removed. In the simulations, we choose different
initial periodic functions in order to determine different periodic solutions.
Periodic regimes of an OEO showing frequencies that are multiple of
$1/\tau_{D}$ were found in the past. In \cite{Illing2005180,illing06}, the
authors progressively increased the delay and investigated the sequential jump
to stable oscillations of frequency $(2n+1)/(2\tau_{D})$ $(n=0,1,...)$. In
\cite{Rosin11,Rosin_master_thesis}, the authors found numerically periodic
solutions of frequency close to $n/\tau_{D}.$ In this paper, we demonstrate
the multi-rhythmicity phenomenon by exciting square-waves with a specific
frequency (specific $n$). Furthermore, we relate these periodic solutions to
nearby Hopf bifurcation points, a prerequisite for an Eckhaus bifurcation scenario.
\begin{figure}
\centering
\includegraphics{setup.eps}
\caption{Schematic of the experimental setup of an optoelectronic oscillator}
\label{setup_oeo}
\end{figure}
The experimental setup of an OEO is sketched in Fig.\ \ref{setup_oeo}. A semiconductor laser beam is injected into a Mach-Zehnder intensity modulator
(MZM). The MZM induces a nonlinear function of the applied voltage. The
modulated light passes through an optical fiber, which is used as a delay
line, and is injected into an inverting photodetector, which converts the
signal into the electrical domain.\ The voltage emitted from the
photodectector passes through a bandpass filter and then through a power
splitter. Half of the voltage, denoted by $V$, is amplified by an inverting
modulator driver (MD). This electric signal is then reinjected inside the MZM
via its radio frequency input port to close the feedback loop. The voltage
coming out of the other port of the power splitter is used to measure the
dynamical variable $V$ with a high-speed oscilloscope. The device used has an
8 GHz analog bandwidth and a 40 GS/s sampling rate. In our experiments, the
delay of the feedback loop is fixed at $\tau_{D}=22$ ns. The system described
is the same as in Refs.\ \cite{illing06,Rosin11,PhysRevLett.95.203903} except
that a pattern generator has been included to perturb the dynamics of the
system. An electrical switch is used to isolate this pattern generator from the rest of the system. The switch also allows a controllable electrical signal to be combined
with $V$ at the input of the MD.
Mathematically, we consider the evolution equations formulated in Refs.
\cite{callan,Rosin11} with time measured in units of the delay.\ They are
given by \footnote{From Eqs. (1) and (2) in \cite{callan}, we obtain our Eqs. (1) and (2) with $\varepsilon
= (T\Delta)^{-1} = (T(\omega_{+} - \omega_{-}))^{-1}$, and $\delta=T\omega_{0}^{2}/\Delta=T(\omega_{+}\omega_{-})/(\omega_{+}-\omega_{-})$ and $\beta=\gamma$. $T$ is the delay of the feedback loop, $\omega_{-}$ and $\omega_{+}$ represent the low and high frequency cut-off of the bandpass filter, respectively. Since $\omega_{+} \gg \omega_{-}$, $\varepsilon \simeq (T \omega_{+})^{-1}$ and $\delta \simeq T\omega_{-}$}%
\begin{align}
\varepsilon\frac{dx}{ds} & = -x-\delta y+\beta\left[ \cos^{2}\left(
m+\tanh\left( x\left( s-1\right) \right) \right) -\cos^{2}\left(
m\right) \right] ,\label{Eq6}\\
\frac{dy}{ds} & = x,\label{Eq7}%
\end{align}
where $s\equiv t/\tau_{D}$ and $x$ is the normalized voltage of the electrical
signal in the OEO. The feedback amplitude $\beta$ and the phase shift $m$ are
two control parameters. $\varepsilon\simeq(\tau_{D}\omega_{+})^{-1}=0.0157$ and $\delta\simeq\tau_{D}\omega_{-}=0.2042$ [26] are
dimensionless time constants fixed by the low and high cut-off frequencies of the bandpass filter denoted by $\omega_{-}$ and $\omega_{+}$, respectively.\ Equations (\ref{Eq6}) and (\ref{Eq7}) are the same equations studied
in \cite{PhysRevLett.95.203903} except of the hyperbolic tangent function in
Eq. (\ref{Eq6}) that accounts for the amplifier saturation.
Equations (\ref{Eq6}) and (\ref{Eq7}) admit a single steady state
$(x,y)=(0,0)$ and its linear stability has been analyzed in detail in Refs.
\cite{Illing2005180,PhysRevLett.95.203903}. Of particular interest are the
primary Hopf bifurcation points, which can be classified into two different
families. In the limit $\delta\rightarrow0$ and $\varepsilon\rightarrow0,$ the
critical feedback amplitudes and the Hopf bifurcation frequencies approach the
limits \footnote{From Eqs. (4) and (5) in \cite{PhysRevLett.95.203903},
introduce $\omega\longrightarrow\omega R,$ $\delta\longrightarrow\varepsilon
R,$ $\varepsilon\longrightarrow1/R.$ The new parameters $\varepsilon$ and
$\delta$ are small and their values are given in Section \ref{intro}.}%
\begin{align}
m >0 :\text{ }\beta_{n} & = 1/\sin(2m),\text{ and }\omega_{n}=(1+2n)\pi\text{
\ }(n=0,1,2,..),\label{OEO9} \\
m <0 :\text{ }\beta_{n} & = -1/\sin(2m),\text{ }\omega_{0}=\sqrt{\delta},\text{
and }\omega_{n}=2n\pi\text{ \ }(n=1,2,..).\label{OE010}%
\end{align}
If $m>0$, the frequencies are odd multiples of $\pi$, meaning that the
successive Hopf bifurcations lead to $2/(1+2n)-$periodic solutions [$2\tau
_{D}/(1+2n)-$periodic solutions in physical time]. If $m<0$ and $n=1,2,..,$
the frequencies are even multiples of $\pi$ and the successive Hopf
bifurcations lead to $1/n-$periodic solutions ($\tau_{D}/n-$periodic solutions
in physical time). In addition, there exists for $m<0$ a Hopf bifurcation
characterized by the low frequency $\omega_{0}=\sqrt{\delta}\ll1$. It leads to
oscillations with a large period compared to $1$ (large period compared to
$\tau_{D}$ in physical time).
The organization of the paper is as follows. In Section \ref{observations}, we
describe the experimental observations and numerical simulations for the two
families of Hopf bifurcations. In Section \ref{stability}, we propose a
partial stability analysis of the plateaus by associating their mean values to
stable fixed points of a map.\ Finally, we discuss our main results in Section
\ref{conclusions}.
\section{Experiments and simulations\label{observations}}
From the linear stability analysis of the zero solution discussed above, we
find that there exist two families of Hopf bifurcations depending on the sign
of $m.$\ For each case, we describe our experimental observations and compare
them to numerical simulations of Eqs. (\ref{Eq6}) and (\ref{Eq7}).
\subsection{Case $m>0$}
If $m>0$, oscillations of period close to $2\tau_{D}$ (corresponding to a
frequency of $22.7$ MHz) are observed experimentally [see Fig.
\ref{Fig2}(a)]. In order to find harmonic oscillations, we excite the system
with signals at different frequencies. To this end, the pattern generator
injects different periodic signals into the OEO loop during a few seconds.
Figure\ \ref{Fig2}(b) shows square-wave oscillations of period close to
$2\tau_{D}/5$ obtained by injecting a square-wave signal of frequency $114$
MHz. Similarly, by exciting the OEO with sine-wave signals of frequency $159$
MHz and of frequency $205$ MHz, we obtain $2\tau_{D}/7$ and $2\tau_{D}%
/9-$periodic oscillations, respectively [Fig. \ref{Fig2}(c) and Fig.
\ref{Fig2}(d), respectively]. $2\tau_{D}/3-$periodic oscillations are also
observed but are not shown for clarity.\ The observation of stable
oscillations characterized by higher frequencies ($n>4$) is not possible
because of the bandwidth limitation of the pattern generator.%
\begin{figure}
\centering
\includegraphics[width=10cm]{XP_phi_positive2.eps}
\caption{Experimental time series obtained after injecting different periodic
signals into the OEO loop during a few seconds and after the injected signal
is removed. The measured values of the parameters are $m=0.665$, $\beta=1.94$,
and $\tau_{D}=22$ ns.}
\label{Fig2}
\end{figure}
We next integrate numerically Eqs. (\ref{Eq6}) and (\ref{Eq7}) using the same
values of the parameters as for the experiments. Figure \ref{Fig3} shows four
different time series obtained using different initial functions described in
the caption. We note that the shape and the period of the oscillations are in
good agreement with the experimental observations. The plateaus of the
square-wave are slightly increasing or decreasing in time, which is an effect
of the small parameters $\varepsilon$ and $\delta$.\ If we decrease their
values, the plateaus become flatter. Another point raised by the numerical
simulations and by the experimental observations is that the mean values of
the plateaus are roughly the same for the main and harmonic periodic
solutions. From Fig. \ref{Fig3}(d), we evaluate theses values as%
\begin{equation}
x_{\max}\simeq0.9\text{ and }x_{\min}\simeq-0.95.\label{extrema}%
\end{equation}
\begin{figure}
\centering
\includegraphics[width=10cm]{phi_positive2.eps}
\caption{Numerical time series obtained from Eqs. (\ref{Eq6}) and (\ref{Eq7}).
(a) Oscillations of period close to $2$ with $x\left( s\right) =\cos\left(
\pi s\right) $ and $y\left( s\right) =0$ $(-1<s<0);$ (b) oscillations of
period close to $2/3$ with $x\left( s\right) =\cos\left( 3\pi s\right) $
and $y\left( s\right) =0$ $(-1<s<0);$ (c) oscillations of period close to
$2/5$ with $x\left( s\right) =\cos\left( 5\pi s\right) $ and $y\left(
s\right) =0$ $(-1<s<0)$; (d) oscillations of period close to $2/7$ with
$x\left( s\right) =\cos\left( 7\pi s\right) $ and $y\left( s\right) =0$
$(-1<s<0)$. The values of the control parameters are the same as in Fig.
\ref{Fig2}: $m=0.665$ and $\beta=1.94$. }
\label{Fig3}
\end{figure}
$2/9-$periodic oscillations are also found numerically but they are unstable
for long time. They are stable if we slightly decrease $\varepsilon$. A smaller $\varepsilon$ leads to sharper transition layers and contribute to the overal stability of the square-wave. The discrepancy between experimental and numerical solutions for the $2/9$-periodic regimes could be the result that the model slightly overestimated the effect of the amplifier saturation (value of $d$) which contributes to smooth the transition layers.
We also examine the effect of changing $m$. Figure \ref{Duke100}(a) shows the first Hopf bifurcation lines in the $(m,\beta)$ parameter plane. There is a stable steady state if
$\beta<1$.\ Increasing $\beta$ leads to a critical point $\beta_{H1}>1$ where
oscillations of period $2$ appear. The minimal value of $\beta_{H1}$ is
obtained if $m=\pi/4$. By progressively increasing $\beta$ from $\beta_{H1}$,
we may generate stable higher-order harmonic oscillations that become more
robust with respect to small perturbations.
\begin{figure}
\centering
\includegraphics[width=10cm]{duke100.eps}
\caption{Hopf bifurcation lines in the $(m,\beta)$ parameter plane. (a) corresponds to the case $m$ positive and (b) to the case $m$ negative. They have been determined numerically from the exact conditions with $\varepsilon=0.0157$ and $\delta=0.2042$. The numbers in the figures indicate the value of $n$ corresponding to a specific frequency defined in (\ref{OEO9}) and (\ref{OE010}). All curves are nearly parabolic with a minimum at $m=\pm\pi/4$. If $\varepsilon\rightarrow0$ and $\delta\rightarrow0$, all curves moves to a unique parabola with minimum located at $(m,\beta)=(\pm\pi/4,1)$.}
\label{Duke100}
\end{figure}
\subsection{Case $m<0$}
If $m<0$, we observe in the experiment stable square-wave oscillations of
period close to $\tau_{D}/n$. Figure \ref{Fig4}(b), (c), and (d) show
oscillations of period close to $\tau_{D}$, $\tau_{D}/2$, and $\tau_{D}/3$,
respectively. They are obtained by exciting the OEO\ with periodic signals of
different frequencies as described in the caption. Oscillations of period
close to $\tau_{D}/4$ are also observed. Moreover, we find stable
slowly-varying oscillations [See Fig. \ref{Fig4} (a)] in agreement with our
previous stability analysis that predicts a Hopf bifurcation for $m<0$ with a low
frequency. As for the case $m>0$, we do not find higher order harmonic
oscillations because of the bandwidth limitation of the pattern generator preventing us to initialize the system with frequencies above a certain threshold.%
\begin{figure}
\centering
\includegraphics[width=10cm]{xp_phi_negative22.eps}
\caption{Experimental time series obtained after injecting different periodic
signals into the OEO loop during a few seconds and after the injected signal
is removed. (a) Low-frequency oscillations of period close to $12\tau_{D}$
obtained by injecting a sine-wave signal of frequency $5$ MHz; (b)
oscillations of period close to $\tau_{D}$ obtained by injecting a sine-wave
signal of frequency $45.5$\ MHz; (c) oscillations of period close to $\tau
_{D}/2$ obtained by injecting a sine-wave signal of frequency $90.9$ MHz; (d)
oscillations of period close to $\tau_{D}/3$ obtained by injecting a sine-wave
signal of frequency $136$ MHz. The values of the delay is $\tau_{D}=22$
ns.\ The measured values of the control parameters are $m=-0.845$ and
$\beta=1.94$ for (a) and (b) and $m=-0.785$ and $\beta=2.2$ for (c) and (d).}
\label{Fig4}
\end{figure}
Integrating Eqs. (\ref{Eq6}) and (\ref{Eq7}) using different initial functions
leads to similar time-periodic regimes.\ The slowly-varying oscillations are
shown in Fig. \ref{Fig5}(a) and exhibit a period close to $T=17.2$.\ With the
Hopf bifurcation frequency $\omega_{0}$ given in (\ref{OE010}),\ we compute
$T_{0}=2\pi/\omega_{0}\simeq14$ which is of the same order of magnitude as
$T$. Figures \ref{Fig5}(b), (c), and (d) show oscillations of period close to
$1$, $1/2$, and $1/3$, respectively.
We investigate numerically the effect of changing $m<0$. The first Hopf bifurcation lines are shown in Fig. \ref{Duke100}(b). In contrast to the
case $m>0$ where the square-wave remains symmetric (same plateau
lengths), the shape of the square-wave depends here on $m$.\ If $m=-\pi/4$, we
observe symmetric square-wave oscillations with a period close to $\tau
_{D}/n.$ However, if $m+\pi/4\neq0,$ the square-wave becomes asymmetric with
different duty lengths for each plateau. The total period remains constant. As
for the case $m>0$, the square-wave oscillations become more robust if $\beta$
increases.\ The same properties are observed experimentally. Figure
\ref{Fig5}(b) and Figs.\ \ref{Fig5}(c) (d) are obtained using slightly
different values of the parameters $m$ and $\beta$. From Figs.\ \ref{Fig5}(c)
and (d), we find that the mean values of the plateaus are identical for the
two periodic regimes and are given by
\begin{equation}
x_{\max}\simeq1.1\text{ and }x_{\min}\simeq-1.1.\label{extrema1}%
\end{equation}
\begin{figure}
\centering
\includegraphics[width=10cm]{phi_negative2.eps}
\caption{Numerical time series obtained from Eqs. (\ref{Eq6}) and (\ref{Eq7}).
(a) Slowly varying solutions obtained with $x\left( s\right) =\cos\left(
0.5\pi s\right) $ and $y\left( s\right) =0$ ($-1<s<0);$ (b) oscillations of
period close to $1$ with $x\left( s\right) =\cos\left( 2\pi s\right) $ and
$y\left( s\right) =0$ ($-1<s<0)$; (c) oscillations of period close to $1/2$
obtained with $x\left( s\right) =\cos\left( 4\pi s\right) $ and $y\left(
s\right) =0$ ($-1<s<0);$ (d) oscillations of period close to $1/3$ with
$x\left( s\right) =\cos\left( 6\pi s\right) $ and $y\left( s\right) =0$
($-1<s<0)$ The values of the control parameters are the same as in Fig.
\ref{Fig4}: $m=-0.845$ and $\beta=1.94$ for (a) and (b) and $m=-0.785$ and
$\beta=2.2$ for (c) and (d).}
\label{Fig5}%
\end{figure}
\section{Linear stability of the plateaus \label{stability}}%
\begin{figure}
\centering
\includegraphics[width=10cm]{duke14.eps}
\caption{Stable fixed points of Eq.\ (\ref{map1}).\ (a) $m=0.665$. The diagram
shows branches of a Period 2 fixed point. The dashed line corresponds to the
experimental and numerical value of $\beta$. (b) $m=-0.785$.\ The diagram
shows two branches of Period 1 fixed points. The dashed lines corresponds to
the experimental and numerical value of $\beta$.}
\label{Fig6}
\end{figure}
In the limit $\delta\rightarrow0$ and $\varepsilon\rightarrow0$, Eqs.
(\ref{Eq6}) and (\ref{Eq7}) reduce to a single equation for a map given by
\begin{equation}
x_{n+1}=\beta\left[ \cos^{2}\left( m+\tanh(x_{n}\right) )-\cos^{2}\left(
m\right) \right] .\label{map1}%
\end{equation}
Here, we demonstrate that the plateaus of the square-waves can be partially
understood by considering the stable fixed points of this map. For the case
$m>0$, there is a Hopf bifurcation at $\beta_{c}=1/\sin(2m)$ to a stable
period-$2$ fixed point [see Fig. \ref{Fig6}(a)]. For the values of $m$ and
$\beta$ used in our experiments and simulations, the diagram in Fig.
\ref{Fig6}(a) indicates
\begin{equation}
x_{\max}=0.72\text{ and }x_{\min}=-1.04,\label{map2}%
\end{equation}
which agree qualitatively with the values (\ref{extrema}) estimated from the numerical
simulations.\
For the case $m<0,$ Eq. (\ref{map1}) admits two stable period-1 fixed points
that appear at $\beta_{c}=-1/\sin(2m)$ [$x_{n}>0$ \ and $x_{n}<0,$ see Fig.
\ref{Fig6}(b)]. For the values of the parameters used in our experiments and
simulations [Figs.\ \ref{Fig4}(c)(d) and for Figs. \ref{Fig5}(c)(d)], the
diagram in Fig.\ \ref{Fig6}(b) indicates
\begin{equation}
x_{\max}=1.10\text{ and }x_{\min}-1.10,\label{map3}%
\end{equation}
which agree quantitatively with the values (\ref{extrema1}) obtained from the
numerical simulations.
\section{Discussion\label{conclusions}}
Symmetric and asymmetric square-wave oscillations have previously been found
\cite{PhysRevE.79.026208,weicker:12b,Rosin11,Weicker:13} and were related to
the first Hopf bifurcation of a basic steady state. Here, we concentrate on
the next primary Hopf bifurcations and show that they quickly stabilize above
critical amplitudes.\ This is the bifurcation scenario related to the Eckhaus
instability known to exist in spatially extended systems.\ The Eckhaus
instability has been predicted to occur in a simple model equation in the
limit of large delays \cite{PhysRevLett.96.220201}. The idea is based on the observation that all Hopf bifurcation points move to a critical value in the limit of large delay (in our case, $\varepsilon\rightarrow0$ and $\delta\rightarrow0$). We may then apply the method of multiple time scales and formulate a partial differential equation for a small amplitude solution. In \cite{PhysRevLett.96.220201}, a single variable complex Ginzburg-Landau equation was derived for which the stability of the different periodic solutions can be demonstrated analytically. The mechanism responsible for the stabilization of each branches of periodic solutions is called the Eckhaus instability. Assuming $\beta-1=O(\varepsilon^2)$ and $\delta=O(\varepsilon^2)$, we have found that two coupled partial differential equations can be derived from Eqs. (\ref{Eq6}) and (\ref{Eq7}). By contrast to the case studied in \cite{PhysRevLett.96.220201}, these equations cannot be solved analytically. However, their similitude to the Ginzburg-Landau equation suggests that distinct stable periodic solutions may coexist through the same Eckhaus scenario. In this paper, we demonstrated both experimentally and numerically that this coexistence of square-waves with distinct periods is possible. Those regimes have already been referenced
in previous studies of OEOs but not as coexisting solutions. They were
obtained as sequencial jumps when varying either the delay
\cite{Illing2005180,illing06} or the low frequency cutoff
\cite{Rosin11,Rosin_master_thesis}. Here, we showed how to obtain these regimes systematically in an experiment without varying any parameters of the
OEO system. A remarkable property of our OEO is the possibility to compare
quantitatively experimental observations and numerical simulations. It
motivates asymptotic studies of the OEO equations based on the large delay
limit \cite{Weicker:13}.
We believe that this multi-rhythmicity of square-waves resulting from nearby
Hopf bifurcations is generic to a large class of delay systems exhibiting a
large delay.\ Recently, we studied a semiconductor laser subject to polarization-rotated
feedback \cite{Friart:14} and found this coexistence of harmonic periodic
regimes both experimentally and numerically.\ The laser rate equations are
completely different from the dynamical equations for an OEO and are
mathematically more complex to analyze.\ The common property is the presence
of nearby Hopf bifurcation points leading to square-waves with periods that
are close to $2\tau_{D}/\left( 2n+1\right) $, where $n=0,1,2,...$.
\begin{acknowledgments}
L.W. acknowledges the Belgian F.R.I.A., the Conseil Régional de Lorraine, the Agence Nationale de la Recherche (ANR) TINO project (ANR-12-JS03-005) and external funds from the F.N.R.S. T.E. acknowledges the support of the F.N.R.S. This work also benefited from the support of the Belgian Science Policy Office under Grant No IAP-7/35 “photonics@be.” D.P.R. and D.J.G. gratefully acknowledge the financial support by the U.S. Army Research Office, Grant No. W911NF-12-1-0099.
\end{acknowledgments}
|
\section{Introduction}
Galactic bars are believed to be directly related to the dynamical evolution of their host spiral galaxies.
Several studies of numerical simulation show that bars can efficiently transport gas from the outer regions of galaxies to the central
kiloparcec scale \citep{wei85,deb98,ath03}.
The gas clouds suffer shocks by interaction with the edges of the bars, producing angular momentum losses and allowing a flow of gas toward the innermost central regions of the barred galaxies \citep{SBF90}.
Given the high efficiency of gas inflow, the presence of bars plays an important role in enhancing the nuclear star formation
\citep{hec80,haw86,deve87,ars89,huang96,HFS97,MF97,emse01,kna02,jog05,hunt08}.
In addition, the bar-driven material toward the most inner regions has been considered an efficient mechanism for triggering activity in active galactic nuclei (AGN) \citep{cor03,comb93}.
\cite{SBF89} proposed a model that they called "bars within bars" as a mechanism for fueling AGN galaxies. In this scenario, the external bars transport gas on a scale of a few parsec and the material is accumulated.
It shows a gravitational instability producing
a secondary bar, which allows gas to approach closer
to a central black hole \citep{mul97,mac97,HFS97}.
Different observational studies, using the Hubble Space Telescope images and the spectroscopic data obtained from ISAAC/VLT and SAURON
\citep{emse01,malk98,laine02,caro02} have confirmed the presence of secondary bars, by observing nuclear bars inside large-scale external bars.
There is still a debate on the observational analysis that shows different results regarding the effects produced by bars on the central nuclear activity.
For instance, \cite{HFS97} found that bars have no significant effect on the AGN activity in a sample of 187 barred spiral galaxies.
\cite{lee12} also used a sample of AGN, selected with the \cite{kew01} classification,
from the Sloan Digital Sky Survey Data Release 7 (SDSS-DR7) volume limited sample of late-type galaxies ($b/a > 0.7$) concluding that bars do not have an important effect on the nuclear activity.
However, \cite{oh12} found evidence that bars enhance central activity in blue galaxies with low black hole masses
from a sample of late-type active galaxies ($b/a > 0.6$) selected from SDSS-DR7 with the \cite{kauff03} criteria.
In the analysis, \cite{lee12} do not include the composite galaxies within the sample of AGN
host and because these galaxies are in general, younger and less massive \citep{kew06} than the AGN
sample defined by the \cite{kew01} criteria. So, the different AGN selection criteria could
affect the results related to the effect of enhanced AGN activity by bars.
More recently, Alonso et al. (2013, hereafter A13) studied a sample of face-on barred AGN spiral galaxies, in comparison with a suitable control sample of unbarred active galaxies with similar redshift, magnitude, morphology, bulge sizes, and local environment distributions.
They found that barred AGN galaxies systematically show
a higher fraction of powerful activity with respect to unbarred AGN in the control sample. In addition, the analysis of the accretion rate onto the central black holes shows that barred AGN host galaxies show an excess of objects with high accretion rate values with respect to unbarred active ones.
Galaxy properties (star formation, gas content, morphology, etc.) show a strong dependence on the local environment in which galaxies reside.
Therefore, to have a better understanding of the formation and evolution of bars, an important point is to analyze, not only the internal structures of spiral galaxies, but also the external influence on their host environment.
Nevertheless, the dependence between the bar fraction and environment is a controversial point since different studies show contradictory results.
Several observational works found that the bar galaxy fraction
does not depend on the environment \citep{vandenB02,mendez10,martinez11,lee12}.
However, \cite{elme90} found a higher bar fraction in pair systems for a sample of spiral galaxies.
In addition, \cite{esk00} show that the bar fraction in the Fornax and Virgo clusters is slightly higher than the average value for fields. Moreover, from the analysis of galaxies in the Shapley-Ames Catalog, \cite{vandenB02} found that the fraction of barred galaxies in cluster environments is larger than in lower density regions.
Recently, \cite{skibba12} show that barred galaxies tend to reside in denser environments than their unbarred counterparts for a large sample obtained from the
Galaxy Zoo 2 project \citep{lintott11}, finding a clear environmental dependence on barred galaxies.
Motivated by these finds, here we analyze the effect of bars on active nuclei galaxies and the role played by high density environments.
We study active galaxies with and without bars within groups and clusters with the aim of assessing whether high density environments play a significant role in modifying the host properties and nuclear activity of barred active galaxies.
For this purpose, using data from the SDSS-DR7, we obtain samples of barred and unbarred AGN residing in groups and clusters of galaxies.
We analyze the spatial distribution, host galaxy properties, and nuclear activity of barred AGN in high density regions, and compare them to those of unbarred active galaxies from a suitable control sample, with the aim of obtaining additional clues about the interplay between the high density environments and barred AGN galaxies.
This paper is structured as follows. Section 2 describes the procedure used to construct the catalog of barred AGN group galaxies from SDSS-DR7, and the control sample selection criteria.
In Section 3, we study the barred and unbarred AGN relative position in groups/clusters.
Section 4 explores the bar dependence on colors of host active galaxies with respect to both group central distance and group virial masses and Section 5 analyses environmental properties.
In Section 6, we analyze the black hole activity in barred AGN in high density environments, and in Section 7 we summarize our main conclusions.
The adopted cosmology through this paper is
$\Omega = 0.3$, $\Omega_{\lambda} = 0.7$, and $H_0 = 100~ \kms \rm Mpc$.
\section{Selection of barred AGN galaxies in high density environments}
The analysis of this paper is based on the SDSS-DR7 \citep{abaz09} photometric and spectroscopic galaxy catalog; DR7 is the seventh data release corresponding to the survey SDSS-II, and comprises 11663 square degrees of sky imaged in five band widths ($u$, $g$, $r$, $i$, and $z$).
It provides imaging data for 357 million objects, as well as spectroscopy over $\simeq \pi$ steradians in the north Galactic cap and $250$ square degrees in the south Galactic cap.
The SDSS-DR7 main galaxy sample is essentially a magnitude limited spectroscopic sample with $r_{lim}<17.77$ covering a redshift range $0<z<0.25$, with a median redshift of 0.1 \citep{strauss02}.
For this work, several physical properties of galaxies were derived from the published SDSS-DR7 data:
stellar mass content, indicators of star-bursts, emission-line fluxes, concentration index, etc. \citep{brinch04, blanton05}, and were obtained from the \texttt{fits} files at the MPA/JHU \footnote{http://www.mpa-garching.mpg.de/SDSS/DR7/} and the NYU\footnote{http://sdss.physics.nyu.edu/vagc/} catalogs.
\begin{figure*}
\includegraphics[width=180mm,height=140mm ]{F1.ps}
\caption{Images of typical examples of barred active galaxies in high density environments.
The images are $\approx$ 3$'$.5 x 3$'$.5.
Right images show a zoom of the barred AGN galaxies.}
\label{Ej}
\end{figure*}
In our previous work (A13) we compiled a barred AGN catalog obtained from SDSS-DR7.
The active galaxy sample was selected by applying the \cite{kauff03} criteria, employing a standard diagnostic diagram proposed by \cite{BPT81} (hereafter BPT), and by using the publicly available emission-line fluxes (for details see A13).
By visual inspection of SDSS images
we classified 6772 hosts of face-on ($b/a>0.4$) AGN spiral galaxies brighter than $g-$mag$<$16.5 and with redshifts $z<0.1$ as barred or unbarred.
We found 1927 AGN hosted in barred galaxies, representing a fraction of 28.5$\%$ with respect to the full sample of AGN in spiral host galaxies, and this value agrees with the bar fraction found by visual inspection of optical galaxy samples in previous works
\citep{deVau91,nilson73,marinova09,master11,oh12}.
The remaining 4628 face-on spiral galaxies are unbarred. The authors excluded 207 objects that did not match the bar classification completely.
To analyze the properties of barred AGN galaxies in high density environments in detail, in comparison with AGN without bars, we identified barred and unbarred active objects that reside in groups and clusters by cross-correlating the total barred and unbarred AGN samples with the SDSS-DR7 group catalog constructed by \cite{zap09}.
These authors identified groups of galaxies using a friends-of-friends algorithm with
the same parameters corresponding to the values found by \cite{merch05}
to produce a reasonably complete sample (95$\%$) with low contamination ($<$ 8$\%$).
The group catalog has 122,962 galaxies within 12,630 groups that have a minimum number of four members
and $z<0.1$.
Moreover, it provides several useful galaxy group properties used throughout the paper, such as the geometric group center coordinates, virial mass, richness, line-of-sight velocity dispersion, and virial radius.
As a result of this cross-correlation, we obtained a sample of 319 barred and 519 unbarred
AGN galaxies in groups.
These values represent a fraction of 16.6$\%$ with respect to the full sample of 1927 barred AGN galaxies, and 11.2$\%$ with respect to a total of 4628 AGN without bars.
Table 1 lists the numbers and percentages of the different AGN galaxy samples used in this work, and
here we observed that the fraction of barred AGN galaxies in groups ($\approx$ 38$\%$) is higher than those in the total barred AGN sample ($\approx$ 28$\%$).
This finding shows a higher bar fraction in groups and clusters of galaxies, indicating that AGN spiral galaxies in groups are more
likely to be barred than those in the field, in agreement with previous work by \cite{skibba12} and \cite{esk00}.
This suggests that bars are stimulated in accordance with the local density environment and that the physical processes such as ram pressure stripping, starvation or strangulation, harassment and interactions in groups and clusters seem to be related to the bar phenomenon.
\cite{skibba12} found that bars are formed by secular evolution and that this process depends on the local environment of the host galaxies, suggesting that there is a dichotomy between internal secular mechanisms and external environment.
Recently, \cite{kor12} show that harassment may influence secular evolution of the bars by the cumulative effect of several rapid encounters with other cluster members.
In addition, starvation or strangulation, which are gas stripping processes of the hot diffuse
component of satellite galaxies, influences the star formation on longer timescales \citep{larson80} and could also stimulate the growing bars
\citep{bere07,master12}.
Figure 1 shows images of typical examples of barred AGN host galaxies in high density environments selected from our sample.
\begin{table}
\center
\caption{Results of classification.}
\begin{tabular}{|c c c| }
\hline
Classification & Number of galaxies & Percentages \\
\hline
\hline
Barred & 1927 & 28.5$\%$ \\
Unclear barred & 207 & 3.0$\%$ \\
Unbarred & 4638 & 68.5$\%$ \\
\hline
Total & 6772 & 100.0$\%$ \\
\hline
\hline
Barred in groups & 319 & 38.2$\%$ \\
Unbarred in groups & 519 & 61.8$\%$ \\
\hline
Total & 838 & 100.0$\%$ \\
\hline
\end{tabular}
{\small}
\end{table}
\begin{table}
\center
\caption{Percentages of barred and unbarred active galaxies with different color and stellar mass ranges.}
\begin{tabular}{|c c c| }
\hline
Color ranges & Barred AGN & Unbarred AGN \\
\hline
\hline
0.55$<(M_g-M_r)<$0.7 & 26.6$\%$ & 73.4$\%$ \\
0.7$<(M_g-M_r)<$0.85 & 45.4$\%$ & 54.6$\%$ \\
0.85$<(M_g-M_r)<$1.0 & 42.5$\%$ & 57.5$\%$ \\
\hline
\hline
Stellar mass ranges & & \\
\hline
10.2$<log(M_*/M_{\sun})<$10.6 & 31.2$\%$ & 68.8$\%$ \\
10.6$<log(M_*/M_{\sun})<$11.0 & 40.0$\%$ & 60.0$\%$ \\
11.0$<log(M_*/M_{\sun})<$11.4 & 40.5$\%$ & 59.5$\%$ \\
\hline
\end{tabular}
{\small}
\end{table}
In addition, we calculated the bar fraction as a function of colors and stellar masses of the host galaxies in our sample (see Table 2). We can observe that the bar fraction increases towards redder and more massive AGN host galaxies,
in accordance with previous work \citep{master11,oh12}.
\subsection{Control sample}
Some studies of the effect of bars from observational analysis show contradictory conclusions.
In several cases, authors have attempted to isolate the effects of bars by comparing barred galaxies with unbarred ones.
However, different authors have proposed different ways to build comparison samples, and so the discrepancy in the results could be mainly due to a biased selection of these control samples.
By using SDSS mock galaxy catalogs built from the Millennium Simulation, \cite{perez09} showed that a suitable control sample for pair galaxies should be selected that at least match, redshift, morphology, stellar masses, and local density environments.
This criteria is also applicable to the building of control samples of barred galaxies with the purpose of analysing different properties with respect to unbarred ones.
Our aim is to focus on the effects of bars on the active nuclei galaxies that reside in high density environments.
Therefore, we constructed a reliable control sample of unbarred AGN host galaxies that share similar environmental conditions in order to compare these with the barred host results.
Using a Monte Carlo algorithm we selected unbarred active galaxies in groups
with similar redshift and stellar mass distributions
(see panels $a$ and $b$ in Fig.~\ref{cont}).
We have also considered unbarred AGNs in the control sample with similar concentration indices\footnote{$C=r90/r50$ is the ratio of Petrosian 90 \%- 50\% r-band light radii}, $C$, as for the barred AGN host sample, with the purpose of obtaining a similar bulge-to-disk ratio (see panel $c$).
Thus, the differences in the results will be driven by the presence of bars and not by the difference in the morphology of the galaxies.
In addition, we have take into account active galaxies with similar bulge prominences in both AGN samples, selecting control galaxies
with similar distributions of the $fracdeV$ parameter to the
barred AGN host sample (panel $d$ in Fig.~\ref{cont}).
This parameter is a good indicator of bulge sizes in disk galaxies
\citep{ber10,kuehn05,master11,skibba12}.
The methodology followed to obtain the control sample of unbarred active galaxies in this work ensures that it will have the same selection effects as the barred AGN host catalog. It can then be used to estimate
the real difference between active galaxies with and without
bars in groups and clusters, unveiling the effect of bars on the nuclear activity and the role played by high density environments.
\begin{figure}
\includegraphics[width=87mm,height=130mm ]{F2.ps}
\caption{Distributions of redshift ($z$), stellar mass ($log(M_*)/M_{\sun}$), concentration index ($C$),
and bulge size indicator ($fracdeV$) (panels $a$, $b$, $c$, and $d$), for barred AGN host galaxies (solid lines) and active galaxies without bars in the control sample (dashed lines).}
\label{cont}
\end{figure}
\section{Analysis of the relative position in groups}
The study of the bar fraction in the extremely dense regions of the cluster core with respect to the lower density environments of the outer regions of the galaxy clusters can provide important clues to understanding the relative effects of internal and external physical mechanism on the evolution of the barred disk galaxies.
We may ask whether barred AGN galaxies have a particular location in groups and clusters with respect to the control sample.
Therefore, we analyze the distribution of projected radial distance with respect to group centers, $d_{CG}$, for barred AGN galaxies and AGN without bars in the control sample.
In order to calculate reliable group centers, we have only considered groups with a minimum number of seven members,
thereby avoiding large uncertainties in this determination.
Both the virial masses and virial radius distributions of the restricted sample are similar to the total group sample.
The resulting distributions are shown in Fig.~\ref{histdc}.
It can be clearly seen that barred AGN galaxies are more concentrated towards the group center
than the unbarred AGN in the control sample.
We also calculated the normalized group-centric distance, $d_{CG}/R_{Virial}$, finding the same
tendencies (see the inset).
In addition, we find that the fraction of barred AGN galaxies nearest the group-centric distance, $d_{CG}<300kpc$ $h^{-1}$, is higher ($\approx$ 49$\%$) with respect to those in the unbarred AGN sample ($\approx$ 38$\%$).
The pioneer work of \cite{thom81} also showed that the bar fraction is significantly larger in the core of the Coma cluster than in the outer region.
More recently, \cite{marinova12}, using high resolution images from the Hubble Space Telescope ACS Treasury survey of the Coma cluster at $z \approx 0.02$, studied bars in the dense core of this cluster.
They found a hint of an increase in the mean bar fraction toward the core of the Coma.
In addition, the study of the bars in lenticular galaxies (S0) in the rich Coma cluster showed that there is a radial increase in the bar fraction towards the cluster core with respect to the low density outer regions \citep{lans14}.
\cite{and96} found that the radial distribution of the barred spirals in the Virgo cluster is clearly different than that of the spirals without bars, showing that barred galaxies are more centrally condensed than the unbarred ones.
Moreover, observational studies in several galaxy clusters also found an increase in the bar fraction in the dense core of the clusters \citep{and96,bar09,marinova12}.
From numerical studies, \cite{byrd90} argued that the observed tendency for disk galaxies in the center of rich clusters to be barred is due to the strong tidal field. This can transform unbarred spiral galaxies into barred ones.
In this context, the tidal field has a constant direction and is symmetric about the spiral galaxy, as would be the case for cluster members approaching in an eccentric orbit towards the cluster center.
This symmetric perturbation can produce distortions in the galaxy disk and then form strong arms and bars.
In addition, other tidal effects of the cluster field have been studied, e.g., tidal stripping \citep{merritt84} and tidal disruption \citep{miller86}. These authors also found that galaxies in the core of the groups and clusters can survive over the long term without being destroyed.
Our results confirm these previous findings for statistically reliable observational data.
\begin{figure}
\includegraphics[width=87mm,height=87mm ]{F3.ps}
\caption{Distributions of group-centric galaxy distances, $d_{CG}$, for
barred AGN (solid lines) and unbarred AGN (full surface) in the control sample.
The inset shows to the distribution of the normalized group-centric distance, $d_{CG}/R_{Virial}$,
in the same samples.
}
\label{histdc}
\end{figure}
\section{Color in host AGN galaxies}
Galaxy colors, as is well known, are correlated with star formation, age, gas content, and environment, all of which may be related to the presence of disk instabilities such as bars, and with the central nuclear activity.
With the aim of exploring the colors of the barred active nuclei galaxies and the effects of denser regions, in this section we analyze colors in both samples of AGN with and without bars that inhabit groups and clusters.
In Fig.~\ref{histCol} we show the $(Mu-Mr)$ and $(Mg-Mr)$ color distributions of
barred and unbarred active galaxies residing in groups and clusters.
It can be clearly seen that there is a significant excess of barred AGN hosts
with red colors.
In Table 3 we quantify the excess of red color indexes ($(Mu-Mr)>2.4$ and $(Mg-Mr)>0.65$)
of barred AGN hosts with respect to the unbarred AGN in the control sample.
Several previous studies displayed an excess of barred galaxies with redder colors from different samples.
\cite{master11} found a high fraction of bars in the passive red spiral galaxies for a sample obtained from the Galaxy Zoo catalog.
In addition, \cite{oh12} showed that a significant number of barred
galaxies are redder than their counterparts of unbarred spiral galaxies.
Furthermore, in our previous work (A13) we found an excess of red colors
in a sample of isolated spiral barred AGN with respect to unbarred active galaxies in a suitable control sample.
The results obtained here from barred AGN hosts in high density environments are consistent with those of barred galaxies in low density regions.
In addition, we analyze the colors of barred AGN hosts in denser regions in comparison with those of active galaxies with bars in the field. For this purpose we
used isolated barred AGN hosts obtained from our previous work (see A13 for details). This sample has similar redshift, stellar mass, concentration index, and bulge size indicator distributions to the sample of barred active galaxies in high density environments.
From Fig.~\ref{histCol} we can see that there is an excess of barred AGN hosts with redder colors in groups and clusters with respect to isolated barred active galaxies.
Furthermore, we calculated the excess of red color indexes for the isolated sample, $(Mu-Mr)>2.4$ and $(Mg-Mr)>0.65$, finding 46.5$\%$ and 84.8$\%$, respectively.
This result also shows that isolated barred AGN hosts are redder than unbarred active galaxies in groups, suggesting that bars
determine changes in the host colors.
Moreover, to help us to understand the behavior of the colors in barred AGN hosts within high density environments, we measured the mean color values with respect to the group-centric distance.
In the left panels of Fig.~\ref{meancol}, we show the mean $<Mu-Mr>$ (upper panel) and $<Mg-Mr>$ (lower panel) colors as a function of group-centric distance for both samples
of barred and unbarred active galaxies.
As is well known, the central regions of the groups and clusters of galaxies are mainly populated by objects redder than those in the outer regions.
It can also be seen that the number of active galaxies with redder colors increases toward the group centers, and this tendency is more significant in barred AGN with respect to those in the control active galaxies without bars.
Figure~\ref{meancol} (right panels) shows the mean colors for barred AGN (solid lines) as a
function of the host group virial mass.
We also show the results for the control sample (dashed lines).
As can be seen, the number of redder AGN hosts increases towards higher group virial masses.
In addition, we can see that barred AGN have systematically redder colors than active galaxies without bars in the control sample.
Redder colors found in barred active galaxies, with respect to unbarred counterparts,
suggest that bar perturbations have a considerable effect on colors of AGN hosts.
\begin{table}
\center
\caption{Percentages of barred and unbarred AGN host galaxies in groups with red colors.}
\begin{tabular}{|c c c | }
\hline
Restrictions & $\%$ of barred & $\%$ of unbarred \\
\hline
\hline
$(Mu-Mr)>2.4$ & 56.2$\%$ & 40.3$\%$ \\
$(Mg-Mr)>0.65$ & 89.6$\%$ & 75.7$\%$ \\
\hline
\end{tabular}
{\small}
\end{table}
\begin{figure}
\includegraphics[width=80mm,height=90mm ]{F4.ps}
\caption{Distribution of colors, $(Mu-Mr)$ and $(Mg-Mr)$ (upper and lower panels),
for barred AGN host galaxies (solid lines) and AGN hosts without bars (shaded histograms)
within groups and clusters.
Dot-dashed lines represent barred AGN hosts in the field.
}
\label{histCol}
\end{figure}
\begin{figure}
\includegraphics[width=95mm,height=90mm ]{F5.ps}
\caption{Mean $<Mu-Mr>$ and $<Mg-Mr>$ galaxy colors (upper and lower panels)
as a function of the group-centric galaxy distances, $d_{CG}$ (left panels),
and virial group mass, $log(M_{Virial}/M_{\sun})$ (right panels), for barred and
unbarred AGN host galaxies (solid and dashed lines).
}
\label{meancol}
\end{figure}
\section{Local environmental properties}
\begin{figure}
\includegraphics[width=75mm,height=82mm ]{F6.ps}
\caption{Distributions of group virial masses, $log(M_{Virial}/M_{\sun})$ (upper panel),
and group colors, $(M_u-M_r)_{Group}$ (lower panel), for host groups of barred and unbarred AGN galaxies
(solid lines and full surface, respectively).
}
\label{Mvcol}
\end{figure}
\begin{figure}
\includegraphics[width=85mm,height=140mm ]{F7.eps}
\caption{Color-magnitude diagrams $(M_u-M_r)_{Group}$ versus $(M_r)_{Group}$.
The density map shows the host groups of barred AGN (color scale corresponding to the cumulative percentage of objects, as shown in the key).
For comparison, the black solid lines enclose 14$\%$, 28$\%$, 42$\%$, 57$\%$, 71$\%$, and 85$\%$ of the host groups of unbarred AGN in the control sample.
Middle and lower panels show host groups with different ranges of virial masses:
$M_{Virial}<10^{13.5} M_{\sun}$ and $M_{Virial}>10^{13.5} M_{\sun}$, respectively.
}
\label{colmr}
\end{figure}
In this section, we analysed the host group properties.
For this purpose we explore the virial masses and the group colors in both host groups
of barred and unbarred active galaxies.
We calculated the group color,
$(M_u-M_r)_{Group}$, which is measured by summing
the luminosities of the four brightest galaxy members in the $r$ and $u$ bands in each group.
Different authors \citep{Eke04,pad04} found that
this measure is a good parameter with which to quantify the global group luminosity.
The results are plotted in Fig.~\ref{Mvcol} where we show the distributions of the virial masses, $log(M_{Virial}/M_{\sun})$ (upper panel), and the group colors, $(M_u-M_r)_{Group}$ (lower panel), of the groups hosting, for barred AGN and their unbarred counterparts.
We find that both host groups of the barred and unbarred AGN show similar virial mass distributions.
Interestingly, the group color distributions of both samples show significant differences. The host groups of the corresponding barred active galaxies show a larger fraction of red colors with respect to the host groups of the unbarred galaxies in the control sample.
We calculate the fraction of active galaxies with and without bars that reside in red groups, defined as $(M_u-M_r)_{Group}>2.4$ and find a percentage of about 62.2 $\%$ and 51.8 $\%$, respectively.
Figure~\ref{colmr} shows color-magnitude diagrams for host groups of barred active galaxies and unbarred AGN in the control sample.
In addition, the analysis was performed for two ranges of group virial masses: $M_{Virial}<10^{13.5} M_{\sun}$ and $M_{Virial}>10^{13.5} M_{\sun}$ (middle and lower panels, respectively).
As can be observed, for the same $(M_r)_{Group}$, the host group colors of barred active galaxies show distributions spreading toward red populations with respect to the host group colors of the unbarred AGN objects.
In addition, this trend is more significant in less massive groups.
However, groups with $M_{Virial}>10^{13.5} M_{\sun}$ have a more similar color-magnitude distributions than that of host groups of active galaxies without bars.
\section{Nuclear activity in barred AGN galaxies}
With the aim of assessing the effect of bars on the nuclear activity and the role played by high density environments,
in this section we analyzed the central activity in the samples of AGN galaxies with and without bars in groups and clusters.
We focused on the dust-corrected luminosity of the [OIII]$\lambda$5007 line, $Lum[OIII]$, as a tracer of the AGN activity.
The [OIII] line is the strongest narrow emission
line in optically obscured AGN and has a low contamination from contributions of star formation lines in the host galaxy.
Furthermore, the $Lum[OIII]$ estimator has been widely used by different authors in several works \citep{mul94, kauff03, heck04, heck05, brinch04}.
From an inspection of the BPT diagrams, \cite{kauff03} found that high metallicity hosts show a low contamination from star formation in the $Lum[OIII]$.
In our sample, most of the galaxies have stellar masses $M^* >10^{10}$ $M_{\sun}$
(see panel $b$ in Fig.2), therefore their metallicities are expected to be high because of the mass-metallicity relation \citep{kauff03, tremonti04}, with low contamination of star formation lines.
In Fig.~\ref{histLOF} we show the distributions of $Lum[OIII]$ for AGN with and without bars within
groups and clusters.
We also plot isolated barred active galaxies residing in the field.
It is interesting that in high density regions barred AGN show a slight trend toward higher values of $Lum[OIII]$ than AGN without bars in the control sample.
Moreover, isolated active galaxies with bars show an excess of higher nuclear activity values with respect to those of barred AGN within groups and clusters.
Following A13, we divide the sample by using the $Lum[OIII] $ to study trends with luminosity. Then, we consider the luminosity $Lum[OIII] = 10^{6.4} L_{\odot}$ as a limit between the lower-luminosity and the higher-luminosity subsamples. Thus, we estimate that the $43.4\%$, $35.7\%$, and $31.5\%$ of isolated barred active objects, barred, and unbarred AGN hosts in denser regions, respectively, belong to the higher-luminosity subsample, because they are the strongest AGN.
These findings indicate that the central nuclear activity in barred AGN galaxies is more efficient in lower density regions than in higher density environments.
Figure~\ref{histLOIII} (upper and middle panels) shows the nuclear activity distributions, $log(Lum[OIII]/L_{\sun})$, of barred active galaxies compared with the unbarred AGN host counterparts.
The analysis was performed for the different ranges of group-centric distances
($d_{CG}<500kpc$ $h^{-1}$ and $d_{CG}>500kpc$ $h^{-1}$),
with the purpose of studying the efficiency of nuclear activity in AGN galaxies with and without bars located in the cluster cores and in the regions far from the group centers.
Different authors \citep{sar80, ken82} found that the values of the cluster core radius are typically between 300 $kpc$ and 600 $kpc$. \cite{dres80} also obtained similar values for the sample of 55 rich clusters. Moreover,
the X-ray emitting gas gives a suitable proxy to the mass distribution in the clusters, which is an important parameter for obtaining the core radius.
In this context, several X-ray observations \citep{for82, pat00, haa10} usually imply core radii of $\approx$ 500 $kpc$.
We found that towards the group center, the $Lum[OIII]$ distributions of barred AGN hosts are similar to those of AGN without bars in the control sample.
Interestingly, towards regions farthest from the group center,
the activity distributions of barred and unbarred AGN hosts exhibit significant differences, with barred objects having a higher fraction of powerful AGN.
We have also computed the mean $<log(Lum(OIII)/L_{\sun})>$ values for different bins of group-centric distance, $d_{GC}$,
for both AGN samples.
The results from the lower panel in Fig.~\ref{histLOIII} show clearly
that, in barred AGN for larger group-centric distance, the $Lum[OIII]$ values strongly increases, while in active galaxies without bars in the control sample remains approximately constant with a slight decline.
It could be interpreted in terms of the greater ability of bars to fuel the central black holes in host group member galaxies located far from the group center.
With the aim of understanding the behavior of the nuclear activity in AGN galaxies with and without bars as a function of host group colors, in Fig.~\ref{LODC} we show the nuclear activity distributions for different ranges of group colors:
$(M_u-M_r)_{Groups}>2.4$ and $(M_u-M_r)_{Groups}<2.4$ (upper and middle panels, respectively).
It can be clearly seen that barred AGN show a trend toward higher values of $Lum[OIII]$ than AGN without bars in the control sample. This tendency is more significant in active galaxies that inhabit bluer groups.
Moreover, in Table 4 we quantify the percentage of barred and unbarred active galaxies in denser regions with
$Lum[OIII] = 10^{6.4} L_{\odot}$, for different ranges of group-centric distance and host group colors.
In addition, in Fig.~\ref{LODC} (lower panel) we present the mean
$<log(Lum[OIII]/L_{\sun})>$ values as a function of the global group colors.
This relation was calculated for the corresponding host groups of the barred AGN and for active galaxies without bars in the control sample.
The results clearly show that generally in bluer groups, active galaxies show higher $Lum[OIII]$ values.
We can also see that barred AGN objects systematically show higher nuclear activity, irrespective of the group colors.
These findings show that the higher-luminosity subsample AGN reside preferentially in bluer groups, which could provide suitable conditions for the central black hole feeding.
In this context, the most efficient nuclear activity in barred AGN galaxies with respect to their unbarred counterparts reflect that bars have an important role in helping the gas infall towards the central regions in active galaxies.
Considering that the group colors, $(M_u-M_r)$, given by the four brightest members is a representative parameter of the
global color of the groups, this result is in agreement with that found by several authors.
\cite{cold09} show that the power of the AGN activity is strongly dependent on the environment, finding that the AGN with higher values of $<log(Lum[OIII]/L_{\sun})>$ are located in regions populated by bluer galaxies.
Similar results were found by \cite{Padilla10,cold14} suggesting that AGN very close to the high density regions need more available gas to effectively feed the central black hole. This is reflected in the higher probability of finding AGN in regions populated by bluer galaxies.
\begin{table}
\center
\caption{Percentages of barred and unbarred active galaxies with $Lum[OIII]> 10^{6.4} L_{\odot}$
in different ranges of group-centric distances and host group colors.
}
\begin{tabular}{|c c c| }
\hline
Ranges & $\%$ of barred AGN & $\%$ of unbarred AGN \\
\hline
\hline
$d_{CG}<500kpc$ $h^{-1}$ & 34.1$\%$ & 34.4$\%$ \\
$d_{CG}>500kpc$ $h^{-1}$ & 41.5$\%$ & 20.8$\%$ \\
\hline
$(M_u-M_r)_{Groups}>2.4$ & 31.1$\%$ & 26.3$\%$ \\
$(M_u-M_r)_{Groups}<2.4$ & 43.8$\%$ & 34.2$\%$ \\
\hline
\end{tabular}
{\small}
\end{table}
\begin{figure}
\includegraphics[width=70mm,height=60mm ]{F8.ps}
\caption{Distributions of $log(Lum[OIII]/L_{\sun})$ for barred AGN (solid line) and unbarred active galaxies (full surfaces) in high density environments.
Dot-dashed line represents isolated barred AGN galaxies.}
\label{histLOF}
\end{figure}
\begin{figure}
\includegraphics[width=100mm,height=115mm ]{F9.ps}
\caption{Distributions of $log(Lum[OIII]/L_{\sun})$ for barred AGN (solid lines) and unbarred active galaxies (full surfaces) with different group-centric distances:
$d_{CG}<500kpc$ $h^{-1}$ and $d_{CG}>500kpc$ $h^{-1}$ (upper and middle panels, respectively).
Lower panel shows mean $<log(Lum(OIII)/L_{\sun})>$ as a function of group-centric distance, $d_{CG}$,
for AGN galaxies with and without bars in groups (solid and dashed lines, respectively).
}
\label{histLOIII}
\end{figure}
\begin{figure}
\includegraphics[width=100mm,height=115mm ]{F10.ps}
\caption{Distributions of $log(Lum[OIII]/L_{\sun})$ for barred AGN (solid lines) and unbarred active galaxies (full surfaces) with different host group color ranges:
$(M_u-M_r)_{Groups}>2.4$ and $(M_u-M_r)_{Groups}<2.4$ (upper and middle panels, respectively).
Lower panel shows mean $log(Lum(OIII)/L_{\sun})$ as a function of the host group colors for
AGN galaxies with and without bars in groups (solid and dashed lines, respectively).
}
\label{LODC}
\end{figure}
\section{Summary and conclusions}
We have performed a statistical analysis of local environmental properties, host characteristics, and nuclear activity of the AGN spiral galaxies with and without bars within denser regions.
For this purpose, we have identified barred and unbarred objects that inhabit groups and clusters of galaxies, by cross-correlating the total barred and unbarred AGN catalogs obtained from our previous work (see Alonso et al. 2013 for details) with the SDSS-DR7 group catalog constructed by \cite{zap09}.
To obtain an appropriate quantification of the effects of bars on the active nuclei and the role played by high density environments, we constructed a suitable control sample of unbarred AGN galaxies within groups and clusters, with the same redshift, stellar mass, concentration index, and bulge size parameter distributions.
We now summarize the principal results of our analysis and its main conclusions.
{\em(i)} We found that the fraction of barred AGN galaxies in groups ($\approx$ 38$\%$) is higher than those in the total barred AGN sample ($\approx$ 28$\%$).
This result shows that AGN spiral galaxies in groups are more likely to be barred than those in the field,
indicating that galactic bars are stimulated in accordance with the local density environment, and also that the physical mechanisms produced within groups and clusters of galaxies (ram pressure, strangulation, harassment and interactions) seem to be related to the bar phenomenon
\citep{kor12,bere07,master12}.
In addition, we observed that the bar fraction increases towards redder and more massive AGN host galaxies.
{\em(ii)} We have found that barred AGN galaxies are more concentrated towards the group centers than the other unbarred AGN galaxy members. We quantified this trend and we found that nearest the group-centric distance, $d_{CG}<300kpc$ $h^{-1}$, the fraction of barred AGN galaxies is higher ($\approx$ 49$\%$) with respect to those in the unbarred AGN sample ($\approx$ 38$\%$).
This implies that barred phenomenon are more likely to occur in the dense core of the groups and clusters, where the strong tidal field can transform unbarred spiral galaxies into barred ones (Byrd \& Valtonen 1990).
{\em(iii)} We examined the color distributions of barred and unbarred AGN host galaxies within groups and clusters, and we found that barred AGN hosts have a clear excess of redder colors than
galaxies in the control sample.
In addition, we analyzed the colors of AGN with bars in the field, finding that isolated barred AGN hosts are redder than active galaxies without bars in groups.
This result suggests that bar perturbations can significantly affect galaxy colors in the hosts of AGN.
{\em(iv)} We also studied the mean colors in barred AGN galaxies and in the control sample as a function of the group-centric distance and group virial masses.
We found that the number of red galaxies increases toward the group center as expected, and
barred galaxies have a higher fraction of redder colors than active galaxies without bars within groups/clusters.
We also show that redder galaxies increases towards higher group virial masses and, in this context, barred AGN galaxies have systematically redder colors than active galaxies without bars in the control sample, for all virial mass ranges.
{\em(v)} We explored the host group properties of the AGN galaxies with and without bars, and we found that both host groups show similar virial masses. However, the group color distributions of both samples present significant differences; namely, the host groups of the barred AGN exhibit a larger fraction of red colors, with respect to the host groups of the corresponding unbarred active galaxies in the control sample.
We also examined the color-magnitude relation for both host groups and we found that at the same $(M_r)_{Group}$,
the host group colors of barred active galaxies display distributions spreading toward red populations with respect to the host groups of the unbarred AGN objects.
This trend is more significant in less massive groups than in
groups with $M_{Virial}>10^{13.5} M_{\sun}$.
{\em(vi)} We found that in high density regions, barred active galaxies show an excess of nuclear activity compared to galaxies without bars in the control sample.
We have also analyzed barred active galaxies in the field, and we found that isolated AGN galaxies with bars
show more efficient nuclear activity with respect to those of barred AGN within groups and clusters.
{\em(vii)} We also analyzed the relation between the central nuclear activity with the group-centric distances, and with the global group colors for barred active galaxies and unbarred AGN in the control sample.
From this study, we conclude that, in particular,
the nuclear activity has strongly increased in barred active galaxies located far of the group center, while the nuclear activity in AGN galaxies without bars remains approximately constant with the group-centric distance.
In addition, we found that in bluer host groups active galaxies show a higher $Lum[OIII]$ values for both AGN samples; however, barred active objects systematically show a higher nuclear activity, irrespective of the global group colors.
These results suggest that the efficiency of bars in transporting material towards the more central regions of the AGN galaxies in high density environments has a significant dependence on the localization of objects within the group and/or cluster and on the host group colors.
\begin{acknowledgements}
This work was partially supported by the Consejo Nacional de Investigaciones
Cient\'{\i}ficas y T\'ecnicas and the Secretar\'{\i}a de Ciencia y T\'ecnica
de la Universidad Nacional de San Juan.
Funding for the SDSS has been provided by the Alfred P. Sloan
Foundation, the Participating Institutions, the National Science Foundation,
the U.S. Department of Energy, the National Aeronautics and Space
Administration, the Japanese Monbukagakusho, the Max Planck Society, and the
Higher Education Funding Council for England. The SDSS Web Site is
http://www.sdss.org/.
The SDSS is managed by the Astrophysical Research Consortium for the
Participating Institutions. The participating institutions are the American
Museum of Natural History, Astrophysical Institute Potsdam, University of
Basel, University of Cambridge, Case Western Reserve University,
University of
Chicago, Drexel University, Fermilab, the Institute for Advanced Study, the
Japan Participation Group, Johns Hopkins University, the Joint Institute for
Nuclear Astrophysics, the Kavli Institute for Particle Astrophysics and
Cosmology, the Korean Scientist Group, the Chinese Academy of Sciences
(LAMOST), Los Alamos National Laboratory, the Max-Planck-Institute for
Astronomy (MPIA), the Max-Planck-Institute for Astrophysics (MPA), New Mexico
State University, Ohio State University, University of Pittsburgh, University
of Portsmouth, Princeton University, the United States Naval Observatory, and
the University of Washington.
\end{acknowledgements}
|
\section{Introduction}
Fundamental theories of everything and grand unification predict the coupling of fundamental constants to gravitational fields. In particular, variation of $\alpha$ presents itself as characteristic shifts in absorption features. The magnitude of this shift is dependent upon the transition being observed, the atomic number of the atom/ion, the ionisation stage, and the strength of the gravitational field in which the ion resides (proportional to the dimensionless potential, $GM/c^{2}r$). Heavy, highly ionised atoms are more sensitive to changes in $\alpha$. An attempt was made by \cite{berengut13a} to detect variations of $\alpha$ in the hot DA white dwarf G191-B2B ($T_{\mathrm{eff}}$=60,000K, log $g$=7.61) by measuring several features of Fe/Ni {\sc v} against their respective sensitivity indices, which is a measurement of how sensitive a transition is to a change in $\alpha$. The G191-B2B spectrum came from \cite{preval13a}, which was constructed by coadding 32 E140H Space Telescope Imaging Spectrometer (STIS) observations. The spectrum covers 1160-1645\AA, and has exceptionally high S/N, exceeding 100 in places. The authors found two conflicting values for possible $\alpha$ variation. The Fe {\sc v} lines gave $\Delta\alpha/\alpha=(4.2\pm{1.6})\times{10^{-5}}$ while the Ni {\sc v} lines gave $\Delta\alpha/\alpha=(-6.1\pm{5.8})\times{10^{-5}}$. It was suggested that the accuracy of the atomic data may have played a part in the discrepancy, however, it was also noted that there may have been a systematic effect in the wavelength calibration. One way to address both issues is to calibrate the Fe/Ni lines using the spectrum of another object with lower mass (and hence smaller gravitational field strength). BD+28$^{\circ}$4211 is a hot sdO ($T_{\mathrm{eff}}$=82,000K, log $g$=6.2), and has been observed extensively with STIS.
In this proceeding, we describe the spectroscopic survey we performed of BD+28$^{\circ}$4211, and using the newly calibrated Fe/Ni {\sc v} data, we also show that potential $\alpha$ variation can be probed down to sensitivities of $\sim{10^{-6}}$.
\section{Observations}
53 E140M and 24 E140H observations from STIS were coadded to construct two spectra, one being medium resolution with $\Delta\lambda$=0.01\AA\, spanning 1160-1720\AA, and one being high resolution with $\Delta\lambda$=0.075\AA\, spanning 1315-1515\AA. The S/N of both spectra are far in excess of the coadded G191-B2B spectrum utilised by \cite{preval13a}. In Figure \ref{fig:bd28g191}, a comparison between the G191-B2B and BD+28$^{\circ}$4211 spectrum is shown.
\begin{figure}
\begin{centering}
\includegraphics[width=100mm,clip=true,trim=0 0 0 5mm]{SPreval1.eps}
\caption{Comparison of coadded spectra of G191-B2B and BD+28$^{\circ}$4211. The flux of G191-B2B has been multiplied by a factor of 2 for clarity. The large features near 1334.5\AA\, is interstellar C {\sc ii}.}
\label{fig:bd28g191}
\end{centering}
\end{figure}
\section{Method}
To calibrate the Fe/Ni {\sc v} data, we first measure all of the absorption features present in the coadded BD+28$^{\circ}$4211 spectra, and identify their origin (photospheric or interstellar). Note that the photospheric velocity is a combination of the gravitational redshift of the star, and the space motion. As has been shown by \cite{ayers08a} in the Deep Lamp Project, the wavelength calibration of E140M spectra is of similar quality to the E140H spectra. Where possible, we use the highest resolution spectra to measure the absorption features, so the E140M spectrum is used in the wavelength ranges 1160-1315\AA\, and 1515-1720\AA, and the E140H spectrum is used in the wavelength range 1315-1515\AA. After identifying all of the photospheric lines with velocities $v_i$, the average photospheric velocity ($v_{\mathrm{BD+28}}$) is calculated, weighted by the inverse square error $\delta{v_i}$ of the individual lines:
\begin{equation}
v_{\mathrm{BD+28}}=\frac{\sum_{i=1}^{N}\frac{v_{i}}{\delta{v_i}^2}}{\sum_{i=1}^{N}\frac{1}{\delta{v_i}^2}}
\end{equation}
The error on $v_{\mathrm{BD+28}}$ is then:
\begin{equation}
\delta{v_{\mathrm{BD+28}}}=\sqrt{\frac{1}{\sum_{i=1}^{N}\frac{1}{\delta{v_i}^2}}}
\end{equation}
Next, to assess the accuracy of the Fe/Ni {\sc v} data, we work backwards, and use the determined photospheric velocity to calculate a "laboratory wavelength", $\lambda_{\mathrm{lab}}$, based on the observed wavelength of the feature. This is calculated as:
\begin{equation}
\lambda_{\mathrm{lab}}=\frac{\lambda_{\mathrm{obs}}}{1+\frac{v_{\mathrm{BD+28}}}{c}}
\end{equation}
The calculated $\lambda_{\mathrm{lab}}$ is compared with the tabulated wavelengths \citep{kurucz11a} use to calculate the photospheric velocity. If the difference between these two values <0.001\AA, then the tabulated laboratory wavelength is regarded as good, otherwise, it is regarded as poor, and is replaced by the calculated $\lambda_{\mathrm{lab}}$. After calculating the corrected wavelengths, this data is then used with the observations made in G191-B2B to repeat the analysis of \cite{berengut13a}.
\section{Results}
The spectroscopic survey of BD+28$^{\circ}$4211 yielded more than 1000 detections of absorption features, 671 of which were identified to originate from Fe/Ni {\sc iv-vii} transitions. 112 absorption features in the BD+28$^{\circ}$4211 spectrum were coincident with the G191-B2B spectrum, 82 of which are Fe {\sc v}, and 30 of which are Ni {\sc v}. The photospheric velocity of BD+28$^{\circ}$4211 was found to be 19.8kms$^{-1}$. Recalculating $\lambda_{\mathrm{BD+28}}$, it was found that >600 Fe/Ni features differed from the tabulated values by >1m\AA. In Figure \ref{fig:alplot}, we plot the measured redshift of the Fe/Ni {\sc v} lines using the corrected wavelengths against the sensitivity index $Q_{\alpha}$. We extracted $\alpha$ variations of $\Delta\alpha/\alpha=(3.99\pm{3.72})\times{10^{-6}}$ for the Fe lines, and $\Delta\alpha/\alpha=(-1.19\pm{0.10})\times{10^{-4}}$ for the Ni lines. While the Ni result implies a variation of $\alpha$ consistent with an $11.9\sigma$ detection, it must be stressed that the errors quoted are statistical only. The large scatter in the Ni redshifts is most likely due to misidentifications in the G191-B2B spectrum, as the Ni features are more difficult to differentiate from the Fe features due to line blending, and lower oscillator strengths in general.
\begin{figure}
\begin{centering}
\includegraphics[width=100mm]{SPreval2.eps}
\caption{Plot of measured redshift of Fe/Ni {\sc v} absorption features calculated with corrected wavelength data against sensitivity index $Q_{\alpha}$, where the blue triangles are Fe {\sc v}, and the red squares are Ni {\sc v}.}
\label{fig:alplot}
\end{centering}
\end{figure}
\section{Conclusion}
We performed a spectroscopic survey of BD+28$^{\circ}$4211 with the aim of refining the Fe/Ni atomic data for use in constant variation studies. This method had the advantage of removing any systematics present in data take with STIS. We calculated a revised constraint on any potential $\alpha$ variation in G191-B2B as $\Delta\alpha/\alpha=(3.99\pm{3.72})\times{10^{-6}}$ for the Fe {\sc v} lines, and $\Delta\alpha/\alpha=(-1.19\pm{0.10})\times{10^{-4}}$ for the Ni {\sc v} lines. The calculated errors are statistical only, and have yet to take into account data systematics. Combined, this is consistent with a null result, however, misidentified Ni {\sc v} features may have skewed the result. Our study has shown that with high quality atomic data, potential $\alpha$ variation can be measured down to this sensitivity. Our result has also shown the effectiveness of using standard stars as stellar laboratories in order to improve the accuracy of atomic databases.
\acknowledgements FUSE data was obtained from the Mikulski Archive for Space Telescopes (MAST). SPP and MAB acknowledge the support of the Science and Technology Facilities Council (STFC, UK). JBH acknowledges support from the NASA grant NNG056GC46G.
|
\subsection{Uniform measurements on the cocycle}\label{unif-meas}
\input{unif-meas.tex}
\subsection{The nearly almost invariance property}\label{nAI}
\input{nai.tex}
\subsection{The statement and proof of the LDT}\label{LDTsing-proof}
\input{ldt-sing-proof.tex}
\section{Introduction and statements}
\label{qp_intro}
We define the quasi-periodic cocycles and describe our assumptions on them. We then formulate the main statements and relate them to recent results for similar models.
\subsection{Description of the model}
\label{qp_model}
\input{model.tex}
\subsection{Brief literature review}
\label{qp_lit_review}
\input{lit_review.tex}
\section{Estimates on unbounded pluri-subharmonic functions}
\label{psh_functions}
In this section we derive certain uniform estimates on analytic functions of severable variables and on pluri-subharmonic functions. These estimates are of a general nature, and they will be applied in the next section to quantities related to iterates of analytic cocycles. Uniformity is understood relative to some measurements which are stable under small perturbations of the functions being measured. The main technical difficulties in establishing these estimates are related to the non-trivial nature of the zeros of an analytic function of several variables, and correspondingly, to the unboundedness of the pluri-subharmonic functions we study.
\subsection{The uniform {\L}ojasiewicz inequality}\label{uLojasiewicz}
\input{unif-lojasiewicz.tex}
\smallskip
\subsection{Uniform $L^2$-bounds on analytic functions}\label{unif-l2bounds}
\input{unif-l2bounds.tex}
\smallskip
\subsection{Estimates on unbounded subharmonic functions}\label{estimates_sh}
\input{estimates_sh.tex}
\subsection{Base LDT estimates for pluri-subharmonic observables}
\label{uqBET}
\input{base-ldt-sh.tex}
\section{The proof of the fiber LDT estimate}
\label{ldt_qp_proof}
\input{fiber-ldt-proof.tex}
\section{Deriving continuity of the Lyapunov exponents}
\label{cont_le_qp}
\input{cont_le_qp.tex}
\section{Refinements in the one-variable case}
\label{refinements_1var}
\input{refinements_1var.tex}
\bigskip
\subsection*{Acknowledgments}
The first author was supported by
Funda\c{c}\~{a}o para a Ci\^{e}ncia e a Tecnologia,
UID/MAT/04561/2013.
The second author was supported by the Norwegian Research Council project no. 213638, "Discrete Models in Mathematical Analysis".
The second author would like to thank C. Marx for suggesting the problem and M. Vod\u{a} for a useful conversation on this topic.
\bigskip
\bibliographystyle{amsplain}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Introduction.}
Strongly correlated electron materials have attracted interest as
candidate thermoelectric materials because they can exhibit values of
the Seebeck coefficient $S$ as large as 100 $\mu$V/K.\cite{zlatic09}
Understanding
and describing the temperature dependence of $S$ in strongly correlated
materials represents a significant theoretical challenge. Both the
magnitude and the temperature dependence of $S$ is distinctly
different than in elemental metals. At low temperatures $S$ increases
linearly with temperature, with a large slope, and reaches a maximum
value of order $k_B/e \simeq $ 86 $\mu$V/K ($k_B$ is Boltzmann’s
constant and $e$ is the charge of an electron). With increasing
temperature $S$ decreases and can even change sign.
These qualitative features are seen in diverse materials
including organic charge transfer salts \cite{yu91}, cuprates \cite{obertelli92,honma}, heavy fermion compounds \cite{mun12}, cobaltates \cite{wang03},
and iron pnictides \cite{hodovanets13}.
This is illustrated in Figure \ref{fig_1} with experimental results
for an organic metal.
Behnia, Jaccard, and Floquet showed that
for a wide range of materials that the slope of the temperature
dependence of $S$ and the specific heat capacity at low temperatures
were proportional to one another.\cite{behnia04} For heavy fermion
materials, this observation can be explained in terms of
a slave boson treatment of the Kondo lattice model.\cite{zlatic08}
Understanding the thermopower in strongly correlated electron
materials has recently attracted increasing theoretical interest.\cite{zlatic09,tomczak12}
Shastry and coworkers have argued \cite{peterson07,peterson07a,
shastry09} that the high frequency limit of the
Kubo formula for the thermopower
actually gives a good approximate value to the dc limit. This approach
has the
advantage that the thermopower (a transport property) can actually be
evaluated from an equal-time expectation value (an equilibrium
property). Peterson and Shastry \cite{peterson10} have shown that the
thermopower is approximately given by the Kelvin formula, the
derivative of the entropy with respect to the particle number; which
via a thermodynamic Maxwell identity equals the derivative of the
chemical potential with respect to temperature.
Recent Dynamical Mean-Field Theory (DMFT) calculations for the Hubbard
model\cite{arsenault} and the Falicov-Kimball\cite{zlatic14} model
show that in the bad metal regime the Kelvin formula is a reasonable
approximation.
The Kelvin formula has the significant advantage that a transport
property can be calculated from an equilibrium thermodynamic
property. It also illuminates the physical significance of work by
Jakli\v{c} and Prelov\v{s}ek who showed\cite{jaklic96} that for the
$t-J$ model the entropy as a
function of doping is a maximum close to optimal doping. This means
that the thermopower should change sign at optimal doping, as is
observed experimentally in the cuprates \cite{obertelli92,honma}.
\begin{figure}[htb]
\centering
\includegraphics[width =92mm,angle=-90] {eps_sa_sc_yu91.eps}
\caption{(color online) Temperature
dependence of the thermoelectric power in
the organic metal $\kappa$-(BEDT-TTF)$_2$Cu[N(CN)$_2$]Br.
The two different curves correspond to two different directions in the crystal.
Experimental data is taken from Ref. \onlinecite{yu91}.
Note the non-monotonic temperature dependence and that
the thermopower is comparable to $k_B/e = 86 \ \mu$V/K.
For temperatures below about 50 K the thermopower is approximately linear
in temperature,
as expected in a Fermi liquid.
The inset shows a schematic of the electron and hole Fermi surfaces
deduced from a tight-binding band structure \cite{merino00b}.
Transport in the $a$ and $c$ directions will be dominated by holes
and electrons, respectively.
}
\label{fig_1}
\end{figure}
Figure \ref{fig_1} shows the measured temperature dependence of
the thermopower of an organic metal.\cite{yu91}
The authors also calculated the thermopower using a Boltzmann equation and
a band structure
obtained with the Huckel approximation. They obtained values that
were about five times smaller than the experiment. However, they found that
if all the hopping integrals were reduced by about a factor of five that
the results were comparable to experiment.
Similar results were obtained earlier by Mori and Inokuchi \cite{mori88}.
Merino and McKenzie suggested that the non-monotonic temperature
dependence arose from a crossover with increasing temperature from
a renormalised Fermi liquid to a bad metal \cite{merino00}.
They showed this was consistent with results of calculations
for a Hubbard model
based on Dynamical Mean Field Theory.
\section{The Kelvin formula}
Starting from a Kubo formula Peterson and Shastry showed that if
one interchanges the thermodynamic and the static limits that the thermopower is given
by the temperature derivative of the chemical potential \cite{peterson10}
\begin{equation}
S_K =
-{1 \over e}\left({\partial \tilde S \over \partial N_{el}} \right)_{T,V}=
{1 \over e}\left({\partial \mu \over \partial T} \right)_{N,V} ,
\label{kelvin}
\end{equation}
where $e$ is the magnitude of the charge of an electron, $\tilde S$ is the entropy, and $N_{el}$ is the particle number.
Note that this result is independent of the direction of the thermal gradient
in the crystal.
Hence, it will be unable to explain the origin of the different
signs shown in Figure \ref{fig_1}.
As a result of the third law of thermodynamics, the entropy should vanish
as the temperature goes to zero for all $N_{el}$ and so $S_K(T) \to 0$ as $T \to 0$.
\section{Hubbard model}
For numerical calculations we consider a system at fixed temperature
$T$ and chemical potential $\mu$, and model it with a (grand
canonical) Hubbard model on the anisotropic triangular lattice,
\begin{equation}
\hat H_{el}=-\sum_{i,j,s} t_{i,j} c_{i,s}^\dagger c_{j,s} +U\sum_i
\hat n_{i,\uparrow} \hat n_{i,\downarrow} -\mu \hat N_{el}.
\label{eq_hel}
\end{equation}
This is a minimal effective Hamiltonian for several
classes of organic
charge transfer salts \cite{powell11} when at half filling.
$\hat N_{el} \equiv \sum_{i,s} \hat n_{i,s}$
is the total electron number operator,
$t_{i,j}=t$ for nearest neighbour bonds in two directions and
$t_{i,j}=t'$ for nearest neighbour bonds in the third
direction. Electronic spin is denoted with $s$ ($\uparrow$ or
$\downarrow$).
For fixed half filled system the chemical potential changes with
temperature and $\mu(T)$ is fixed by the constraint that
\begin{equation}
\langle \hat N_{el} \rangle = N ,
\end{equation}
where $N$ is the number of lattice sites, ensuring half-filling, and
$\langle \hat A \rangle$ denotes the grand canonical thermal average,
$\langle \hat A \rangle \equiv \textrm{Tr} [\hat A \exp(-\beta \hat H_{el})]/Z$ with $Z$
being the thermodynamic sum $Z=\textrm{Tr} [\exp(-\beta \hat H_{el})]$. Here we
have also used $\beta=1/(k_B
T)$.
Our numerical results for finite lattices were obtained by the
finite-temperature Lanczos method (FLTM) \cite{jaklic00}, which
we previously used to determine several thermodynamic quantities of
this Hubbard model \cite{kokalj13}. We showed that
there was a transition from a metal to a Mott insulator with increasing $U/t$,
with the critical value depending on the amount of frustration $t'/t$.
In the metallic phase as the temperature increased there is a crossover
from a Fermi liquid (with a specific heat and entropy that increased linearly with
temperature) to a bad metal, characterised by an entropy of order $k_B \ln (2)$.
The coherence temperature associated with this crossover, was substantially
reduced by strong correlations, having a value of order $t/10$.
\section{Results}
In Fig. \ref{fig_2} we show how the thermopower estimated with Kelvin
formula $S_K$ shows a large enhancement with increasing electronic
interactions $U$ at low $T$. In comparison to the non-interacting ($U=0$)
system the enhancement can be an order of magnitude and originates
in electronic correlations. The largest magnitude of $S_K$ is reached for
$T\sim T_\textrm{coh}\sim 0.1t$ which is much lower than the Fermi energy. Below
$T_\textrm{coh}$ one enters a coherent Fermi liquid regime in which one
expects a linear temperature dependence of $S_K$, extrapolating
to zero at zero temperature, in accordance with the third law
of thermodynamics. This regime is hard to
reach numerically and our results only indicate it with $S_K$ tending
to $0$ at $T\to 0$ for $T<T_\textrm{coh}$. In Fig. \ref{fig_2} we linearly
extrapolated $S_K$ to 0 for $T\to 0$ by hand to demonstrate the expected
behaviour.
\begin{figure}[htb]
\centering
\includegraphics[width
=92mm,angle=-90]{eps_dmudT_ne10_tp10_Udep_a.eps}
\caption{(color online) Enhancement of the thermopower by strong
correlations. The temperature dependence of the Kelvin thermopower
$S_K$ is shown for several different values of the Hubbard $U$. All
results are for the isotropic triangular lattice, $t'=t$. As the
Mott metal-insulator transition ($U_c \simeq 7.5 t$)\cite{kokalj13}
is approached the magnitude of the thermopower increases to values
that are an order of magnitude larger than for non-interacting
electrons ($U=0$) for temperatures of about $T \sim t/10$. The
maximum in $|S_K|$ at low temperatures corresponds to the crossover
from a Fermi liquid at low temperatures to a bad metal at higher
temperatures. This maximum is also seen in the specific heat
\cite{kokalj13} and the spin susceptibility. The curves have been
linearly extrapolated from their value at $T=0.06t$ to zero at
zero-temperature. Also shown is the linear temperature dependence
obtained by a Sommerfeld expansion for non-interacting electrons
\cite{Ashcroft}. }
\label{fig_2}
\end{figure}
\subsection{Non-interacting fermions}
In a non-interacting fermion system the chemical potential, at
temperatures much less than the Fermi temperature can be estimated via
Sommerfeld expansion leading to \cite{Ashcroft}
\begin{equation}
\mu(T) = E_F - {\pi^2 \over 6}(k_B T)^2 {g'(E_F) \over g(E_F)}.
\label{ashcroft-mu}
\end{equation}
Here $g(E_F)$ is the density of states (DOS) at the Fermi energy
($E_F$) and $g'(E_F)$ is its slope. Substituting
Eq. (\ref{ashcroft-mu}) in the Kelvin formula gives $S_K$ that is
linear in temperature, with a magnitude of order, $(k_B/e)(k_B
T/E_F)$, which for elemental metals will be very small. We show in
Fig. \ref{fig_2} that the Sommerfeld expansion,
Eq. \eqref{ashcroft-mu}, gives a good low $T$ estimate for
non-interacting electrons, up to about $T=0.3t$. The non-interacting
density of states is shown in Fig. \ref{fig_3}.
\begin{figure}[htb]
\centering
\includegraphics[width =92mm,angle=-90]{eps_NIE_DOS_tp10.eps}
\caption{(color online) Energy dependence of the density of states
$g(\epsilon)$ for the tight-binding band structure associated with
non-interacting electrons with $t'=t$. For half filling $\mu = E_F =
0.82t$ at zero temperature, and $g(E_F) = 0.14/t$ and $g'(E_F) =
0.056/t^2$. The latter determines the slope of the Kelvin
thermopower versus temperature for non-interacting
electrons. Reversing the sign of $t'$ or both $t$ and $t'$,
corresponds to a particle-hole transformation and reverses the sign
of this derivative ($g'(E_F)$). }
\label{fig_3}
\end{figure}
\subsection{Fermi liquid regime}
\label{sec_fl}
When using the Kelvin formula one should however be careful, since it
may not be a good approximation in some regimes. For example, its
weakness for $T<T_\textrm{coh}$ can be understood by starting with the Mott
formula \cite{shastry09,peterson10}
\eq{
S_{Mott}=-T\frac{\pi^2k_B^2}{3e}\frac{d}{d\mu}
\ln[
g(\mu)\overline{v_{k,x}^2 \tau_{k,\mu}}
]|_{\mu\to E_F}.
\label{eq_smott}
}
Here, $\overline{v_{k,x}^2 \tau_{k,\mu}}$ denotes the average of the
quasi-particle velocity at wave vector $k$ in the $x$ direction
($v_{k,x}$) times the quasi-particle lifetime ($\tau_{k,\mu}$) over a
surface in reciprocal space at energy equal to $\mu$. To obtain the
Kelvin formula from $S_{Mott}$ one needs to neglect the $\mu$
dependence of $\overline{v_{k,x}^2 \tau_{k,\mu}}$ in
Eq. \eqref{eq_smott} leading to
\eq{
S_{K}=-T\frac{\pi^2k_B^2}{3e}\frac{d}{d\mu}
\ln[
g(\mu)
]|_{\mu\to E_F}.
\label{eq_sklowt}
}
This is the same result as obtained for the non-interacting case via
Eq. \eqref{ashcroft-mu} and also represents the low-temperature Kelvin
formula in a coherent regime with well defined quasi-particles. The
problem with the Kelvin formula in a Fermi liquid regime is in
neglecting the $\mu$ dependence of the velocity in the term $\overline{v_{k,x}^2
\tau_{k,\mu}}$, while keeping the $\mu$ dependence of the density of states $g$, which is also
related to the velocity since $g\propto 1/v$. It is also unlikely that in a Fermi liquid regime that
$\tau$ would cancel the $\mu$ dependence of
$v_{k,x}^2$ in $\overline{v_{k,x}^2 \tau_{k,\mu}}$. That the Kelvin
formula is more appropriate for higher temperatures and in the incoherent
regime was already pointed out in Ref. \onlinecite{peterson10}, while in the
low temperature regime it only gives a rough approximation.
This is explicitly found in
recent Dynamical Mean-Field Theory (DMFT) calculations for the Hubbard
model\cite{arsenault} and the Falicov-Kimball\cite{zlatic14} model.
\subsection{Effect of dimerization}
\label{sec_dimer}
In Fig. \ref{fig_1} the measured thermopower of an organic metal in
two different directions is shown. The opposite signs for the
two directions was argued\cite{yu91} to originate in the finite
dimerization of the hopping (alternating hopping $t-\delta t$,
$t+\delta t$, \ldots) in two directions on the triangular
lattice. Such a dimerisation splits the band into two bands, one
electron and the other hole like \cite{ivanov97,merino00b}. Each band
dominates the thermopower in its own direction and leads to opposite
signs of the thermopower for the two directions. Due to the band
splitting the density of states is also split. However, it turns out
that just the density of states cannot capture the change of sign and
that $v^2$ term discussed in Sec. \ref{sec_fl} needs to be included to
reproduce the opposite signs. The Boltzmann transport equation
approach in Ref. \onlinecite{yu91} does take these terms into account
and captures the correct signs.
\subsection{Comparison to experiment}
For comparison of the experimental data shown in Fig. \ref{fig_1} and
our results shown in Fig. \ref{fig_2} we set the energy scale $t=50$
meV $\sim 580$ K as appropriate value obtained by Density Functional
Theory for organic charge transfer
salts \cite{kandpal09, nakamura09, jeschke12,scriven12}.
We note that with our definition of hopping parameters in
Eq.~\eqref{eq_hel} we should for organics either take both $t$ and
$t'$ negative\cite{kandpal09,jeschke12,scriven12} or positive $t$ and
negative $t'$\cite{nakamura09}, but both changes correspond at
half-filling to particle-hole transformation (with additional shift in
$k$ space for the later) and therefore only reverse the sign of $S_K$
shown in Fig. \ref{fig_2}.
Then we
estimate from Fig. \ref{fig_2} that the maximal thermopower would
appear at roughly $T_\textrm{coh}=60$ K, which is in agreement with
experiment. We also capture the qualitative $T$ dependence of the
thermopower. However, as already discussed in Sec. \ref{sec_dimer},
the Kelvin formula does not have the potential to describe the
orientational dependence shown in Fig. \ref{fig_1}, which originates
in the finite dimerization of the lattice.
\section{Conclusion}
In conclusion, we have shown with the Kelvin forumla, which is a good
approximation in the bad metallic regime, that the thermopower is
strongly enhanced by electronic correlations at low $T$, even by an
order of magnitude compared to the weak or non-interacting electron
limit. Comparing with experimental data for an organic charge
transfer salt, we capture qualitatively the temperature dependence and
overall magnitude of the thermopower. On the other hand, the Kelvin
formula can not capture the orientational dependence of $S$ observed
in experiment, for which one would need to employ a Kubo formula and
introduce dimerization of the lattice into the model. We leave this
as a future challenge.
\begin{acknowledgments}
We acknowledge helpful discussions with Nandan Pakhira, Peter
Prelov\v{s}ek, Sriram Shastry and Jernej Mravlje.
This work was supported by
Slovenian Research Agency grant Z1-5442 (J.K.) and an Australian
Research Council Discovery Project grant.
\end{acknowledgments}
\bibliographystyle{apsrev4-1}
|
\section{Introduction and Preliminaries}
\label{one}
Subgroup depth $d_0(H,G)$ of a subgroup $H$ in a finite group $G$ is introduced in \cite{BKK} as the minimum depth of the induction-restriction table of irreducible characters of $H$ and $G$, a matrix
of nonnegative integers with nonzero rows and columns. As the matrix of induction $K_0(\C H) \rightarrow K_0(\C G)$, the notion of depth also occurs in fields of topological algebra in various guises. In ring theory, the minimum depth $d(B,A)$ of a subring $B \subseteq A$ is introduced in \cite{BDK} in terms of the natural bimodule ${}_BA_B$ and its tensor powers (and for even depth, tensored one more time by
the bimodule ${}_BA_A$ or ${}_AA_B$). The depth $d_0(H,G)$ is recovered in \cite{BDK} by letting
$B$, $A$ be group algebras over a field of characteristic zero; indeed, \cite{BDK} shows that depth $d(kH,kG)$ depends only on the characteristic of the field $k$ \cite{BDK}.
The more general definition also allows one to consider depth of integral group ring extensions $\Z H \subseteq \Z G$ and their minimum depth $d_{\Z}(H,G)$.
In addition, a combinatorial depth $d_c(H,G)$ is introduced in \cite{BDK} by using G-set
analogues of balanced tensors and bimodules. The following string of inequalities is from \cite[4.5]{BDK}: $$d_0(H,G) \leq d_p(H,G) \leq d_{\Z}(H,G) \leq d_c(H,G) \leq 2|G: N_G(H) |.$$
In \cite{LK2011}, h-depth of a subring pair $B \subseteq A$ is introduced by the same process as in the definition of depth but focussing on the natural $A$-$A$-bimodules of the tensor powers of $A$ relative to $B$. The minimum h-depth $d_h(B,A)$
is closely related to $d(B,A)$ by the inequality $| d(B,A) - d_h(B,A) | \leq 2$. Its definition in~\ref{def-depth} suggests the notion of h-depth is natural and almost unavoidable when considering subgroup depth. In this paper, h-depth provides a natural transition (in Section~2) to depth of an $A$-coring, which is a notion of coalgebra generalized to $A$-bimodules
where the comultiplication and counits are $A$-bimodule morphisms.
In \cite{D} the subring depth of twisted group algebras of the permutation groups $\Sigma_n$
are computed as an intriguing contrast to the untwisted case in \cite{BKK}, where it was shown that $d_0(\Sigma_n, \Sigma_{n+1}) = 2n-1$.
In \cite{D} it was shown that with $\alpha$ the nontrivial $2$-cocycle (representing the nonzero element in $H^2(\Sigma_n, \C^{\times})$), the twisted complex group algebra extensions have minimum depth
\begin{equation}
\label{eq: Danz}
d(\C_{\alpha}\Sigma_n, \C_{\alpha}\Sigma_{n+1}) = 2(n - \lceil \frac{\sqrt{8n+1} - 1}{2} \rceil) + 1.
\end{equation}
Note that $d(\C_{\alpha}\Sigma_n,\C_{\alpha}\Sigma_{n+1}) \leq d_0(\Sigma_n, \Sigma_{n+1})$
with a difference that goes to infinity as $n \rightarrow \infty$. The same is true of the alternating group series $A_n$, where $d_0(A_n,A_{n+1}) = 2(n - \lceil \sqrt{n}\, \rceil) +1$ $\geq d(C_{\alpha}A_n, \C_{\alpha}A_{n+1})$, the last depth also equal to the right-hand side of Eq.~(\ref{eq: Danz}); cf.\ \cite[appendix]{BKK}.
Given a subgroup $H \leq G$, we show in this paper that the crossed product of a twisted $G$-algebra $A$ with subalgebra
$A \#_{\sigma} H$ has h-depth less than or equal to the h-depth of the corresponding group algebra extension:
i.e., we establish
$$d_h(A \#_{\sigma} H,A \#_{\sigma} G) \leq d_h(kH, kG)$$
in Eq.~(\ref{eq: hdepthineq}) in Section~4 below.
We will also extend an equality for h-depth of a Hopf subalgebra in terms of depth of its quotient module \cite{K2013} to the equalition
for h-depth of a left coideal subalgebra of a Hopf algebra $H$ in Corollary~\ref{cor-cue}.
Our method is to define depth of corings so that the depth of the Sweedler coring of a ring extension is its h-depth, and apply (Doi-Koppinen) entwining structures that are Galois corings.
In Corollary~\ref{cor-core} below we show that $d_h(H,G) = d_h(G/N,H/N)$ where $N$ is the core of a subgroup $H$
in a finite group $G$ (over any ground field). In Corollary~\ref{cor-cores} we note that subgroup depth behaves precisely like combinatorial depth with repect to this subgroup $N$, after proving
that subring depth is preserved by quotienting of relatively nice ideals (Theorem~\ref{prop-sigma}), or
better yet, by
relatively nice Hopf ideals (Proposition~\ref{prop-crop}). A final application is to a left coideal subalgebra $R$ of a finite-dimensional Hopf algebra $H$, which is left normal iff a nonzero right integral in $R$ is a normal element in $H$ (Theorem~\ref{th-leftrightleft}).
\subsection{Preliminaries on subalgebra depth} Let $A$ be a unital associative algebra over a field $k$. The category of modules over $A$ will be denoted
by $\M_A$. (For finite-dimensional $A$, the notation $\M_A$ denotes the category of finite-dimensional modules.)
Two modules $M_A$ and $N_A$ are \textit{similar} (or H-equivalent) if $M \oplus * \cong q \cdot N := N \oplus \cdots \oplus N$ ($q$ times) and $N \oplus * \cong r \cdot M$
for some $r,q \in \N$. This is briefly denoted by $M \| q \cdot N$ and $N \| r \cdot M$ for some $q,r \in \N$ $\Leftrightarrow$ $M \sim N$. Recall that similar modules have Morita equivalent endomorphism rings.
Let $B$ be a subalgebra of $A$ (always supposing $1_B = 1_A$). Consider the natural bimodules ${}_AA_A$, ${}_BA_A$, ${}_AA_B$ and
${}_BA_B$ where the last is a restriction of the preceding, and so forth. Denote the tensor powers
of ${}_BA_B$ by $A^{\otimes_B n} = A \otimes_B \cdots \otimes_B A$ ($n$ times $A$) for $n = 1,2,\ldots$, which is also a natural bimodule over $B$ and $A$ in any one of four ways; set $A^{\otimes_B 0} = B$ which is only a natural $B$-$B$-bimodule.
\begin{definition}
\label{def-depth}
Suppose that $A^{\otimes_B (n+1)}$ is similar to $A^{\otimes_B n}$ as natural $X$-$Y$-bimodules
for subrings $X$ and $Y$ of $A$ and $n \in \N$. One says that subalgebra $B \subseteq A$ has
\begin{itemize}
\item \textbf{depth} $2n+1$ if $X = B = Y$ for $n \geq 0$;
\item \textbf{left depth} $2n$ if $X = B$ and $Y = A$,
\item \textbf{right depth} $2n$ if $X = A$ and $Y = B$,
\item \textbf{h-depth} $2n-1$ if $X = A = Y$,
\end{itemize}
for $n \geq 1$.
Note that if $B \subseteq A$ has h-depth $2n-1$, the subalgebra has (left or right) depth $2n$ by restriction of modules. Similarly, if $B \subseteq A$ has depth $2n$, it has depth $2n+1$. If $B \subseteq A$ has depth $2n+1$, it has depth $2n+2$ by tensoring either $-\otimes_B A$ or $A \otimes_B -$ to $A^{\otimes_B (n+1)} \sim A^{\otimes_B n}$. Similarly, if $B \subseteq A$ has left or right depth $2n$, it has h-depth $2n+1$. Denote the \textbf{minimum depth}
of $B \subseteq A$ by $d(B,A)$ \cite{BDK}. Denote the \textbf{minimum h-depth} of $B \subseteq A$ by $d_h(B,A)$ \cite{LK2011}. Note that
$d(B,A) < \infty$ $\Leftrightarrow$ $d_h(B,A) < \infty$; if so, \cite{LK2011} shows that
\begin{equation}
-2 \leq d(B,A) - d_h(B,A) \leq 1.
\end{equation}
\end{definition}
For example, $B \subseteq A$ has depth $1$ iff ${}_BA_B$ and ${}_BB_B$ are similar \cite{BK2, LK2012}. In this case, one deduces the following algebra isomorphism,
\begin{equation}
\label{eq: d1eq}
A \cong B \otimes_{Z(B)} A^B,
\end{equation}
where
$Z(B), A^B$ denote the center of $B$ and centralizer of $B$ in $A$.
Another example is that $B \subset A$ has right depth $2$ iff ${}_AA_B$
and ${}_A A \otimes_B A_B$ are similar. If $A = \C G$ is a group algebra of a finite group $G$ and $B = \C H$ is a group algebra of a subgroup $H$ of $G$, then $B \subseteq A$ has right depth $2$ iff $H$ is a normal subgroup of $G$ iff $B \subseteq A$ has left depth $2$; a similar statement is true for a Hopf subalgebra $R \subseteq H$ of finite index and over any field \cite{BK}.
Now let $H$ be a Hopf algebra with counit $\eps: H \rightarrow k$, antipode $S: H \rightarrow H$ and
coproduct $\cop: H \rightarrow H \otimes H, \cop(h) = h\1 \otimes h\2$ (Sweedler notation suppressing summations). Let $A$ a \textit{right $H$-comodule algebra}, i.e., an $H$-comodule
with coaction $\rho: A \rightarrow A \otimes H, \rho(a) = a\0 \otimes a\1$ that is a unital algebra homomorphism. The
coinvariants $\{ b \in A \, | \, \rho(b) = b \otimes 1_H \}$ form a subalgebra
$B$; one says $A$ is an \textit{$H$-Galois extension} of $B$ if the Galois mapping $A \otimes_B A \rightarrow A \otimes H$ given by $x \otimes y \mapsto x y\0 \otimes y\1$ is bijective. Note that
the Galois mapping is an isomorphism of natural $A$-$B$-bimodules: if $\dim H = n$,
then $A \otimes_B A \cong n \cdot {}_AA_B$, and $A \supseteq B$ has depth $2$.
There is the following remarkable converse growing out of Ocneanu's ideas in subfactor theory with detailed papers by Szymanski, Longo, Nikshych-Vainerman and others:
\begin{theorem}
Let $A \supseteq B$ be a Frobenius algebra extension with $1$-dimensional centralizer and surjective Frobenius homomorphism. If $d(B,A) \leq 2$,
then $A$ is a Hopf-Galois extension of $B$.
\end{theorem}
The notion of Frobenius extension is defined in Example~\ref{example-Frobenius}; a short proof-with-references in \cite{LK2004}. The surjectivity condition ensures
that $A_B$ is a generator (and conversely \cite{LK2012}). This theorem requires particularly the use of the depth two condition for the construction of a Hopf algebra
structure on $\End {}_BA_B$ or its dual Hopf algebra structure on $\End {}_AA \otimes_B A_A \cong (A \otimes_B A)^B$ \cite{KS}. (However, if $d(B,A) = 1$, none of these difficulties arise, as Eq.~(\ref{eq: d1eq}) forces $A = B$ with one-dimensional center, the trivial Hopf algebra.) The stringent condition on the centralizer $A^B$ may be relaxed if
one considers more general Hopf algebras and their Galois coactions (such as weak Hopf algebras and
Hopf algebroids) \cite{KS}.
Note that one always has
\begin{equation}
\label{eq: divides}
A^{\otimes_B n} \| A^{\otimes_B (n+1)}
\end{equation} as natural $A$-bimodules for all $n \geq 2$: one applies the unit mapping and multiplication to obtain a split monic (or split epi). For $n =1$, though it holds for the other three of the bimodule structures, it
is not generally true as $A$-$A$-bimodules, $A \| A \otimes_B A$ being the separable extension condition on $B \subseteq A$.
Now $A \otimes_B A \| q \cdot A$ as $A$-$A$-bimodules for some $q \in \N$
is the H-separability condition and implies $A$ is a separable extension of $B$ by Hirata, cf.\ \cite[2.6]{NEFE}. Somewhat similarly, ${}_BA_B \| q \cdot {}_BB_B$ implies
${}_BB_B \| {}_BA_B$ \cite{LK2012}. It follows that subalgebra depth and h-depth may be equivalently defined by replacing the similarity bimodule conditions for depth and h-depth in Definition~\ref{def-depth} with the corresponding bimodules on
\begin{equation}
\label{eq: def}
A^{\otimes_B (n+1)} \| q \cdot A^{\otimes_B n}
\end{equation}
for some positive integer $q$ \cite{BDK, LK2011, LK2012}.
For example, for the permutation groups $\Sigma_n < \Sigma_{n+1}$
and their corresponding group algebras over any
commutative ring $K$, one has depth $d_K(\Sigma_n,\Sigma_{n+1}) = 2n-1$ \cite{BDK}. Depths of subgroups in $PGL(2,q)$, Suzuki groups, twisted group algebra extensions and Young subgroups of $\Sigma_n$ are computed in \cite{F, HHP, D, FKR}. If $B$ and $A$ are semisimple complex algebras, the minimum odd depth is computed from powers of an order $r$ symmetric matrix with nonnegative entries $\mathcal{S} := MM^T$ where $M$ is the inclusion matrix
$K_0(B) \rightarrow K_0(A)$ and $r$ is the number of irreducible representations of $B$ in a basic set of $K_0(B)$; the depth is $2n+1$ if $\mathcal{S}^n$ and $\mathcal{S}^{n+1}$ have an equal number of zero entries \cite{BKK}. It follows that the subalgebra pair of semisimple complex algebras $B \subseteq A$ always has finite depth.
Similarly, the minimum h-depth of $B\subseteq A$ is computed from
powers of an order $s$ symmetric matrix $\mathcal{T} = M^TM$, where $s$ is the rank of $K_0(A)$; the h-depth is $2n+1$ if $\mathcal{T}^n$ and $\mathcal{T}^{n+1}$ have an equal number of zero entries (equivalently, letting $\mathcal{T}^0 = I_{s \times s}$, one has
\begin{equation}
\label{matrixinequality}
\mathcal{T}^{n+1} \leq q\mathcal{T}^n
\end{equation}
for some $q \in \N$) \cite[Section 3]{LK2012}.
\subsection{Depth of Hopf subalgebras, coideal subalgebras and modules in a tensor category $\mathcal{M}_H$} Let $ H$ be a l Hopf algebra over an arbitrary field $k$. Let $R \subseteq H$ be a Hopf subalgebra, so that $\cop(R) \subseteq R \otimes R$ and the antipode satisfies $S(R) = R$. It was shown in \cite[Prop.\ 3.6]{K2013} that the tensor powers of $H$ over $R$, denoted by $H^{\otimes_R n}$, reduce to tensor powers of the generalized quotient $Q := H/ R^+H$ as follows:
\begin{equation}
\label{eq: diagaction}
H^{\otimes_R n} \stackrel{\cong}{\longrightarrow} H_{\cdot} \otimes Q_{\cdot}^{\otimes (n-1)}
\end{equation} which
for $n = 2$ is given by $x \otimes_R y \mapsto xy\1 \otimes \overline{y\2}$; see \cite[Eq. (21)]{HKY} for the
straightforward extension of this to all $n$.
The isomorphism is an $H$-$H$-bimodule isomorphism where the left $H$-module structures are the natural endpoint actions (as well as the right action to the left), and the right $H$-module structure on $H \otimes Q \otimes \cdots \otimes Q$ is given by the diagonal action of $H$:
$$h'(y \otimes q_1 \otimes \cdots \otimes q_{n-1}) \cdot h = h'yh\1 \otimes q_1 h\2 \otimes \cdots \otimes q_{n-1} h_{(n)}.$$
The following proposition directly makes use of the isomorphism for each $n \geq 2$. Given a subalgebra pair $U \supseteq T$, observe that the bimodule
${}_UU_T$ is projective iff the multiplication mapping $U \otimes T \rightarrow T$, $u \otimes t \mapsto ut$, is a split
epi of $U$-$T$-bimodules iff there is an element $e = e^1 \otimes e^2 \in U \otimes T$ such
that $e^1 e^2 = 1_U$ and $te = et$ for each $t \in T$ (a so-called ``right relative separable tower'' of algebras
$U \supseteq T \supseteq k1_U$).
\begin{prop}
Suppose $H$ is a finite-dimensional Hopf algebra with $R$ and $Q$ as above, with intermediate
Hopf subalgebras $H \supseteq U \supseteq T \supseteq R$.
If ${}_UU_T$ is a projective bimodule, then $d(R,H) < \infty$. In particular, the depth is finite if either $H$ or $R$ is semisimple \cite{LK2012, K2013}.
\end{prop}
\begin{proof}
In a finite tensor category, such as $\mathcal{M}_H$ with the diagonal action, $P \otimes_k X$ is projective if $P$ is projective and $X$ is any $H$-module \cite[Prop.\ 2.1]{EO} (its proof does not need $k$ to be algebraically closed).
The proof below will also require the notion of tensor product Hopf algebra $H \otimes K$ of two Hopf algebras $H$, $K$, as well as the Hopf opposite algebra (with antipode $S^{-1}$) \cite{Ma}. Then Eq.~(\ref{eq: diagaction}) shows (by restriction) that
each $H^{\otimes_R n}$ is a projective $U$-$T$-bimodule, equivalently projective $H \otimes T^{\rm op}$-module, for one extends $Q$ to an (left-sided trivial) $U$-$T$-bimodule via $uqt = \eps(u)qt$. Since $R^e = R \otimes R^{\rm op}$ is a Hopf subalgebra in the finite-dimensional $U \otimes T^{\rm op}$, and therefore a free extension, it follows that each $H^{\otimes_R n}$ is projective
as a natural $R$-$R$-bimodule.
The Krull-Schmidt Theorem, Eq.~(\ref{eq: divides}) and the fact that there are finitely many projective indecomposable isoclasses entail that $H^{\otimes_R (N+1)} \sim H^{\otimes_R (N+2)}$ for $N$ equal to this number (or the number of nonisomorphic simple $H$-$R$-bimodules). Thus, $d(R,H) \leq 2N + 3$.
A semisimple Hopf algebra $H$ (or $R$) is a separable algebra, separability being characterized by the condition ${}_{H^e}H$ (or ${}_RR_R$) is
projective. The finite depth follows by letting $U = T = H$ (or $U = T = R$).
\end{proof}
The bimodule isomorphism in Eq.~(\ref{eq: diagaction}) shows quite clearly
that the following definition applied to $Q$ will be of interest to computing $d(R,H)$. Let $W$ be a right $H$-module and
$T_n(W) :=
W \oplus W^{\otimes 2} \oplus \cdots \oplus W^{\otimes n}$.
\begin{definition}
A module $W$ over a Hopf algebra $H$ has \textbf{depth} $n$ if $T_{n+1}(W) \| t \cdot T_n(W) $
for some positive integer $t$,
and depth $0$ if $W$ is isomorphic to a direct sum of copies of $k_{\eps}$, where $\eps$ is the counit.
Note that this entails that $W$ also has depth $n+1$, $n+2$, $\ldots$. Let $d(W, \M_H)$ denote
its \textbf{minimum depth}. If $W$ has a finite depth, it is said to be an \textit{algebraic module}.
If $W$ is an $H$-module coalgebra, or $H$-module algebra, the condition of depth $n$ simplifies
to $W^{\otimes (n+1)} \| t \cdot W^{\otimes n}$ for some $t \in \Z_+$ and all $n \in \N$, where
$W^{\otimes 0}$ denotes $k_{\eps}$.
\end{definition}
\begin{lemma}
\label{lemma-modI}
Let $I$ be a Hopf ideal in a Hopf algebra $H$. Suppose $I$ is contained in the annihilator ideal of an $H$-module $W$.
Then depth of $W$ is the same over $H$ or $H/I$.
\end{lemma}
\begin{proof}
The lemma is proven by noting that a Hopf ideal $I$ in $\Ann_H W$ is contained in the annihilator ideal of each tensor power of $W$, since $I$ is a coideal. Additionally, split epis as in $T_{n+1}(W) \| t \cdot T_n(W) $ descend and lift along $H \rightarrow H/I$.
\end{proof}
Recall that the Green ring, or representation ring, of a finite-dimensional Hopf algebra $H$ over a field $k$, denoted by $A(H)$, is the free abelian group with basis consisting of indecomposable (finite-dimensional) $H$-module isoclasses,
with addition given by direct sum, and the multiplication in its ring structure given by the tensor product. For example,
$K_0(H)$ is a finite rank ideal in $A(H)$.
As shown in \cite{Fe}, a finite depth $H$-module $W$ satisfies a polynomial with integer coefficients in $A(H)$, and conversely. Thus, an algebraic module has isoclass an algebraic element in the Green ring,
which explains the terminology.
The main theorem in \cite[5.1]{K2013} proves that Hopf subalgebra (minimum) depth and depth of its generalized quotient $Q$ are closely related by
\begin{equation}
\label{eq: inequalityfordepth}
2d(Q,\M_R) + 1 \leq d(R,H) \leq 2d(Q, \M_R) + 2.
\end{equation}
Here one restricts $Q$ to an $R$-module, in order to obtain the better result on depth. If $R$ is a left coideal subalgebra of $H$, it is not itself a Hopf algebra, and this as a result is unavailable: in this case, we prove below (Corollary~\ref{cor-cue}) that the minimum h-depth satisfies
\begin{equation}
\label{eq: hdepthandmoduledepth}
d_h(R,H) = 2d(Q, \M_H) + 1.
\end{equation}
\begin{example}
\begin{rm}
For the reader who knows something about the Drinfeld double Hopf algebra $D(H)$ of a Hopf algebra $H$ (or group $G$),
we work an example using the notation of $h \in H$ and its dual Hopf subalgebra $f \in H^*$ as subalgebras of $fh \in D(H)$ subject to relations $hf = f(S^{-1}(h\3) ? h\1) h\2$ \cite{M,Ma}. We compute the quotient $Q$ of (coopposite) $H^*$ as a Hopf subalgebra of $D(H)$ simply by $$Q = H^* H/{H^*}^+H \cong H$$
via $\overline{fh} \mapsto f(1_H)h$ with right $H^*$-action $\overline{h}f = f(S^{-1}(h\3)h\1) h\2$. This gives $\overline{h} f = \overline{h} f(1_H)$ if $H$ is cocommutative.
Thus if $H$ is cocommutative, $d(Q, \M_{H^*}) = 0$, and by the inequality in (\ref{eq: inequalityfordepth}),
$$ d(H^*, D(H)) \leq 2,$$
i.e., $H^*$ is a normal Hopf subalgebra in $D(H)$ (recovering some folklore with a hands-off approach). Also by Eq.~(\ref{eq: hdepthandmoduledepth}), $d_h(H^*,D(H)) = 3$
since $Q_{D(H)} \cong H_H$ the regular representation, which has depth $1$ if $\dim H > 1$ \cite{K2013}. Minimum depth $d(G, D(G))$ is analyzed in \cite{HKY}.
\end{rm}
\end{example}
Finally we remark that if $H$ is semisimple, Eq.~(\ref{eq: hdepthandmoduledepth}) in principle computes the depth of the quotient module $Q$
in the finite tensor category $\M_H$ in terms of the symmetric matrix $\mathcal{T}$ in inequality~(\ref{matrixinequality}) (where $A = H$ and $B = R$ is semisimple
\cite[ch.\ 3]{M}): let
$d'(\mathcal{T})$ denote the least $n$ for which inequality~(\ref{matrixinequality}) holds, then
\begin{equation}
d(Q, \M_H) = d'(\mathcal{T})
\end{equation}
Thus if
$M$ denotes the inclusion matrix $K_0(R) \rightarrow K_0(H)$, then $2d'(\mathcal{T})+1 = d_{odd}(M^T)$ (cf.\ \cite[3.2]{LK2011}) in terms of the minimum (odd) depth of an irredundant nonnegative matrix $M$ (and its transpose $M^T$) defined in \cite[2.1]{BKK}.
\subsection{Depth of a group subalgebra pair compared with its corefree quotient pair}
Given a ring $A$ with subring $B$, say that an $A$-ideal is
\textit{relatively nice} if its intersection, the $B$-ideal $J = I \cap B$, satisfies $AJ = I$ or $JA = I$. Note that if both set equalities hold, such as when I is a relatively nice Hopf ideal in a finite-dimensional Hopf algebra with respect to a Hopf subalgebra, we are studying a type of normality condition related to the normality notion in \cite[Section 4]{BKK}.
Noting the canonical inclusion $B/J \into A/I$, we extend Theorem 3.6 in \cite{KY}, where $I$ is fully contained in $B$, to relatively nice ideals.
\begin{theorem}
\label{prop-sigma}
Let $I$ be a relatively nice ideal in $A$, where $J$ is its intersection with $B$. Minimum depth satisfies $d(B/J, A/I) \leq d(B, A)$.
\end{theorem}
\begin{proof}
If $d(B,A) = \infty$, there is nothing to prove.
Assume that there is a split monomorphism $\sigma: A^{\otimes_B (n+1)} \into q \cdot A^{\otimes_B n}$ of (natural $B$- or $A$-) bimodules for some $q, n \in \N$,
where we use balanced multilinear notation for the arguments $a_i \in A$ ($i = 1, \ldots, n+1$). Let $\pi$ denote the canonical surjection $A \rightarrow A/I$,
where $\pi(a_i) = \overline{a_i}$.
Define $$\overline{\sigma}(\overline{a_1},\ldots,\overline{a_{n+1}}) = q \cdot \pi^{\otimes n}\sigma(a_1,\ldots,a_{n+1}) = (\pi^{\otimes n}\sigma_i(a_1,\cdots,a_{n+1}))_{i=1}^q. $$
We must show that for any $i = 1,\ldots,q$ and $y \in I$, $$\pi^{\otimes n}\sigma_i(a_1, \cdots,y,\cdots,a_{n+1}) = 0$$ to see that $\overline{\sigma}$ is well-defined.
Suppose without loss of generality that $I = JA$, so there are finitely many $x_{j_{r+1}} \in J$ and $a_{j_{r+1}}\in A$ such that $y = \sum_{j_{r+1}} x_{j_{r+1}}a_{j_{r+1}}$, where we note $$\sigma_i(a_1, \cdots,a_r, y,\ldots,a_{n+1}) =
\sum_{j_{r+1}}\sigma(a_1,\ldots, a_rx_{j_{r+1}},a_{j_{r+1}},a_{r+1}, \ldots,a_{n+1}),$$
but $a_r x_{j_{r+1}} \in I = JA$. Then this process may be repeated $r$ times, until we have the terms
$$\sigma_i(a_1,\ldots,y,\ldots,a_{n+1}) = $$
$$\ \ \ \ \ \sum_{j_1,\cdots,j_{r+1}}^N x_{j_1 \cdots j_{r+1}}\sigma(a_{j_1 \cdots j_{r+1}},\ldots,a_{j_r j_{r+1}}, a_{j_{r+1}},a_{r+1},\ldots,a_{n+1})$$
where $x_{j_1 \cdots j_{r+1}} \in J$, all terms mapping to zero under $\pi^{\otimes n}$. Similarly a splitting for $\sigma$ descends to $q \cdot (A/I)^{\otimes_{B/J} n}
\rightarrow (A/I)^{\otimes_{B/J} (n+1)}$, a splitting for $\overline{\sigma}$. The inequality of minimum depths follows as a consequence of the characterization of depth in (\ref{eq: def}).
\end{proof}
\begin{remark}
\begin{rm}
The proof of Theorem~\ref{prop-sigma} may be adapted slightly to show that relative Hochschild cochain groups of a subring pair
$B \subseteq A$ with coefficients in an $A$-bimodule $M$ \cite{H} with left and right annihilator containing a relatively nice ideal $I$
in $A$ satisfies $C^n(A,B;M) \cong C^n(A/I,B/J; M)$ induced by the canonical algebra epi $A \rightarrow A/I$,
a cochain isomorphism for all $n \geq 0$.
\end{rm}
\end{remark}
For the following corollary, assume $I$ is a relatively nice Hopf ideal in
a Hopf algebra $H$.
\begin{prop}
\label{prop-crop}
The minimum depth satisfies $$d(R/J, H/I) \leq d(R,H) \leq d(R/J, H/I) + 1.$$ Moreover, if $d(R/J,H/I)$
is even, the two minimum depths are equal.
\end{prop}
\begin{proof}
Note that $J = R \cap I$ is a Hopf ideal in $R$ and satisfies both $HJ = I = JH$ by an application of the antipode.
Let $Q = H/R^+H$ be the quotient module of $R \subseteq H$, and
$Q'$ be the corresponding quotient module of $R/J \subseteq H/I$. Since $J \subseteq R^+$ and so $I \subseteq R^+H$,
it follows from a Noether isomorphism theorem that $Q' \cong Q$ as $H$-modules. Since $(H/R^+H)J = I/R^+H = 0$, it follows
from the lemma above that $d(Q, \M_R) = d(Q, \M_{R/J})$. From Eq.~(\ref{eq: inequalityfordepth})
it follows that $$2d(Q,\M_R) + 1 \leq d(R,H), d(R/J, H/I) \leq 2d(Q, \M_R) + 2. $$ The proposition now
follows from Theorem~\ref{prop-sigma}.
If $d(R/J,H/I)$ is even, then $d(R/J,H/I) = 2 d(Q, \M_{R/J}) + 2$ from Eq.~(\ref{eq: inequalityfordepth}), but the Theorem and Eq.~(\ref{eq: inequalityfordepth}) imply that
$d(R,H) = d(R/J, H/I)$.
\end{proof}
\begin{example}
\label{example-8dim}
\begin{rm}
Let $H_8$ be the $8$-dimensional small quantum group (at the root-of-unity $q = i$). This is generated as an algebra by $K,E,F$ such that $K^2 = 1$, $E^2 = 0 = F^2$,
$EF = FE$, $FK = - KF$, and $EK = - KE$. Let $R_4$ be the $4$-dimensional Hopf subalgebra
generated by $K,F$, which is isomorphic to the Sweedler algebra. (Please refer to \cite[Example 4.9]{K2013} for the coalgebra structure and details related to depth.)
Consider the relatively nice Hopf ideal $I$ with basis $\{ F, FK, EF, EFK \}$. Then $J$ is $\mbox{rad}\, R$ with basis $\{ F, FK \}$. The quotient Hopf algebras $H_8/I \cong R_4$
and $R_4/J \cong \C \Z_2$ have depth $d( \C \Z_2, R_4) = 3$ computed in \cite[4.1]{Y}. It follows from the proposition that
\begin{equation}
3 \leq d(R_4, H_8) \leq 4.
\end{equation}
\end{rm}
\end{example}
Recall that the core $\mbox{\rm Core}_G(H)$ of a subgroup pair of finite groups $H \leq G$ is the intersection of conjugate subgroups of $H$; equivalently, the largest normal subgroup contained in $H$.
Note that if $N = \mbox{\rm Core}_G(H)$, then $\mbox{\rm Core}_{G/N}(H/N)$ is the one-element group, i.e., $H/N \leq G/N$ is a \textit{corefree} subgroup.
\begin{cor}
\label{cor-cores}
Suppose $N$ is a normal subgroup of a finite group $G$ contained in a subgroup $H \leq G$. For any ground field $k$, ordinary depth satisfies the inequality,
$$ d_k(H/N, G/N) \leq d_k(H,G) \leq d_k(H/N, G/N) + 1. $$
Moreover, if $d_k(H/N, G/N)$ is even, the two minimum depths are equal.
\end{cor}
\begin{proof}
Note that $I = kGkN^+$ is a relatively nice Hopf ideal in $kG$ (generated by $\{ g - gn \| \, g \in G, n \in N \}$). (And conversely any Hopf ideal of the group algebra $kG$ is of this form \cite{PQ}.)
Also note $kG/I \cong k[G/N]$. The corollary
now follows from the previous proposition.
\end{proof}
This result improves \cite[Prop.\ 3.5]{HKY} to arbitrary characteristic. Note too that combinatorial depth satisfies a completely analogous property given in
\cite[Theorem 3.12(d)]{BDK}. The inequality is in a sense the best
result obtainable.
For example, let $G = \Sigma_4$, $H = D_8$, the dihedral group of
$8$ elements embedded in $G$, and $$N = \{ (1), (12)(34), (13)(24), (14)(23) \}.$$ It is noted after \cite[6.8]{BKK} that $G/N \cong \Sigma_3$,
$H/N \cong \Sigma_2$, and that minimum depth satisfies $d_0(G,H) = 4$ \cite[Example 2.6]{BKK}, $d_0(H/N, G/N) = 3$.
\section{Depth of corings}
\label{two}
Let $A$ be a ring and $(\mathcal{C}, A, \cop, \eps)$ be an $A$-coring. Recall that
$\mathcal{C}$ is an $A$-bimodule with coassociative coproduct $\cop: \mathcal{C} \rightarrow
\mathcal{C} \otimes_A \mathcal{C}$ and counit $\eps: \mathcal{C} \rightarrow A$, both $A$-bimodule morphisms satisfying $(\eps \otimes \id_{ \mathcal{C}})\cop = \id_{ \mathcal{C}} = (\id_{ \mathcal{C}} \otimes \eps) \cop$ \cite{BW}. It follows
that $\mathcal{C}^{\otimes_A n} \| \mathcal{C}^{\otimes_A (n+1)}$ as $A$-bimodules for each $n \geq 1$. For convenience in notation let $\mathcal{C}^{\otimes_A 0} = A$. The reader may consult \cite{BW} for more on corings, their morphisms, their comodules and useful examples, such as coalgebras over ground rings and others we bring up below.
\begin{definition}
An $A$-coring $\mathcal{C}$ has \textbf{depth} $2n+1$ if $\mathcal{C}^{\otimes_A n} \sim \mathcal{C}^{\otimes_A (n+1)} $, where $n$ is a nonnegative integer. If $n$ is a positive integer,
$\mathcal{C}$ has depth $2n+1$ if $\mathcal{C}^{\otimes_A (n+1)} \| m \cdot \mathcal{C}^{\otimes_A n}$ for some $m \in \N$. Note that $\mathcal{C}$ has depth
$2n+3$ if it has depth $2n+1$. If $\mathcal{C}$ has a finite depth, let $d(\mathcal{C}, A)$ denote
its \textbf{minimum depth}.
\end{definition}
\begin{example}
\begin{rm}
Given a ring extension $B \rightarrow A$, let $\mathcal{C}$ denote its Sweedler $A$-coring
$A \otimes_B A$ with coproduct $\cop: A \otimes_B A \rightarrow A \otimes_B A \otimes_B A$
($= A^{\otimes_B 3}$) given by $\cop(a \otimes_B a') = a \otimes 1 \otimes a'$
and counit $\mu: A \otimes_B A \rightarrow A$ by $\mu(a \otimes a') = aa'$ \cite{BW}.
Note that $\mathcal{C}^{\otimes_A n} \cong A^{\otimes_B (n+1)}$ for integers $n \geq 0$.
Therefore, comparing the tensor powers of $\mathcal{C}$ as natural $A$-bimodules is equivalent
to comparing the tensor powers $A^{\otimes_B n}$ as in the definition of h-depth in Definition~\ref{def-depth}. It follows that
\begin{equation}
\label{eq: coringdepth=hdepth}
d(A \otimes_B A, A) = d_h(B,A)
\end{equation}
\end{rm}
\end{example}
An $A$-coring $\mathcal{C}$ has a \textit{grouplike} element $g \in \mathcal{C}$ if $\cop(g) = g \otimes_A g$
and $\eps(g) = 1_A$. For example, the Sweedler $A$-coring $A \otimes_B A$ in the example
has a grouplike element $1_A \otimes_B 1_A$. An $A$-coring $\mathcal{C}$ with grouplike $g$
has invariant subalgebra $A_g^{\mbox{co}\, \mathcal{C}} := \{ b \in A \| \, bg = gb \} := B$ (notation refers to coinvariants of $\mathcal{C}$-comodule $A$ determined by grouplike \cite[p.\ 278]{BW}). Recall that
$\mathcal{C}$ is a \textit{Galois coring} if $A \otimes_B A \cong \mathcal{C}$ via $a \otimes_B a' \mapsto
aga'$, a coring isomorphism of $\mathcal{C}$ with the Sweedler coring of $B \subseteq A$.
For example, the Sweedler coring of a faithfully flat ring extension $A \supseteq B$ is Galois, since $B =
\{a \in A \| \, a \otimes_B 1_A = 1_A \otimes_B a \}$ follows from $A_B$ or ${}_BA$ being
faithfully flat (\cite[28.6]{BW}, or see Lemma~\ref{lemma-ff}).
\begin{example}
\label{example-Frobenius}
\begin{rm}
Recall that $A \supseteq B$ is a Frobenius extension with $F: A \rightarrow B$ a ``Frobenius'' homomorphism and $e := \sum_i x_i \otimes_B y_i$ a ``dual bases tensor,'' satisfying $ae= ea$ for every $a \in A$, if $A$
is a $B$-coring with coproduct $\cop:A \rightarrow A \otimes_B A, \ \cop(a) = ae$ and counit $F$, this coring being denoted by
$\mathcal{C}_{\mathrm{Frob}}$. The more familiar conditions of the counit equations
characterize Frobenius extensions \cite{NEFE}.
The tensor powers of this coring are now given
by natural $B$-bimodules $\mathcal{C}_{\mathrm{Frob}}^{\otimes_B n} = A^{\otimes_B n}$.
Using Definition~\ref{def-depth} for the definition of odd depth, we obtain
\begin{equation}
\label{eq: FrobeniusCoringDepth=OddDepth}
d(\mathcal{C}_{\mathrm{Frob}}, B) = d_{\mathrm{odd}}(B,A)
\end{equation}
in terms of the minimum odd depth.
\end{rm}
\end{example}
\section{Depth of Coalgebra-Galois Extensions}
\label{three}
In this section we define depth of a certain coalgebra-Galois extension and see that its minimum depth takes on at least as many interesting values as subgroup depth \cite{BK, BDK, BKK, D, F, FKR}. In contrast, the minimum depth of a Hopf-Galois extension is one or two.
Let $k$ be a field; all unlabeled tensors are over $k$. We begin with a review of the entwining
structure of an algebra $A$ and coalgebra $C$. The entwining (linear) mapping $\psi: C \otimes A \rightarrow A \otimes C$ satisfies two
commutative pentagons and two triangles (a bow-tie diagram on \cite[p.\ 324]{BW}). Equivalently,
$(A \otimes C, \id_A \otimes \cop_C, \id_A \otimes \eps_C)$ is an $A$-coring with respect
to the $A$-bimodule structure $a(a' \otimes c){a''} = aa' \psi(c \otimes a'')$
(or conversely defining $\psi(c \otimes a) = (1_A \otimes c) a$).
An entwining structure mapping $\psi: C \otimes A \rightarrow A \otimes C$ takes
values that may be denoted by $\psi(c \otimes a) = a_{\alpha} \otimes c^{\alpha} = a_{\beta} \otimes c^{\beta}$, suppressing linear sums of rank one tensors, and satisfies the axioms: (for all $a,b \in A, c \in C$)
\begin{enumerate}
\item $\psi(c \otimes ab) = a_{\alpha}b_{\beta} \otimes {c^{\alpha}}^{\beta}$;
\item $\psi(c \otimes 1_A) = 1_A \otimes c$;
\item $a_{\alpha} \otimes \cop_C(c^{\alpha}) = {a_{\alpha}}_{\beta} \otimes {c\1}^{\beta} \otimes {c\2}^{\alpha}$
\item $a_{\alpha} \eps_C(c^{\alpha}) = a \eps_C(c)$,
\end{enumerate}
which is equivalent to two commutative pentagons (for axioms 1 and 3) and two commutative
triangles (for axioms 2 and 4), in an exercise.
The following is \cite[32.6]{BW} or \cite[Theorem 2.8.1]{CMZ}.
\begin{prop}
\label{prop-1-1}
Entwining structures $\psi: C \otimes A \rightarrow A \otimes C$ are in one-to-one correspondence with $A$-coring structures \newline
$(A \otimes C, \id_A \otimes \cop_C, \id_A \otimes \eps_C)$.
\end{prop}
\begin{proof}
Given an entwining $\psi$, the obvious structure maps $(A \otimes C, \id_A \otimes \cop_C, \id_A \otimes \eps_C)$ form an $A$-coring with respect
to the $A$-bimodule structure $a(a' \otimes c){a''} = aa' \psi(c \otimes a'')$.
Conversely, given an $A$-coring $A \otimes C$ with
coproduct $\id_A \otimes \cop_C: A \otimes C \rightarrow A \otimes C \otimes C \cong
A \otimes C \otimes_A A \otimes C$ and counit $\id_A \otimes \eps_C: A \otimes C \rightarrow
A \otimes k \cong A$, one defines $\psi(c \otimes a) = (1_A \otimes c) a$ and checks
that $\psi$ is an entwining, and the other details, in an exercise.
\end{proof}
Our primary example in this section is $A = H$, a Hopf algebra with coproduct $\cop$, counit $\eps$
and antipode $S: H \rightarrow H$, and $C$ a right $H$-module coalgebra,
i.e. a coalgebra $(C, \cop_C, \eps_C)$ and module $C_H$ satisfying $\cop_C(ch) = c\1 h\1 \otimes c\2 h\2$
and $\eps_C (ch) = \eps_C(c) \eps(h)$ for each $c \in C, h \in H$. An entwining mapping $\psi: C \otimes H \rightarrow H \otimes C$ is defined by $\psi(c \otimes h) = h\1 \otimes c h\2$.
The entwining axioms are checked in a more general setup in Section~\ref{four}.
The associated $H$-coring $H \otimes C$ has coproduct $\id_H \otimes \cop_C$ and counit
$\id_H \otimes \eps_C$ with $H$-bimodule structure: ($x,y,h \in H, c \in C$)
\begin{equation}
\label{eq: coring}
x(h \otimes c) y = xh y\1 \otimes c y\2
\end{equation}
Notice that this is the diagonal action from the right.
\begin{prop}
\label{prop-entwinedepth}
The depth of the $H$-coring $H \otimes C$ and the depth of the $H$-module $C$ are related by
$d(H \otimes C, H) = 2d(C, \mathcal{M}_H) + 1$.
\end{prop}
\begin{proof}
The $n$-fold tensor product of $H \otimes C$ over $H$ reduces to the $H$-bimodule isomorphism
\begin{equation}
\label{eq: iso}
(H \otimes C)^{\otimes_H n} \cong H \otimes C^{\otimes n}
\end{equation}
via the mapping
$$ \otimes_{i=1}^n h_i \otimes c_i \longmapsto h_1 {h_2}\1 \cdots {h_n}\1 \otimes_{i=1}^{n-1} c_i {h_{i+1}}\i+1 \cdots {h_n}\i+1 \otimes c_n $$
with inverse $h \otimes c_1 \otimes \cdots \otimes c_n \mapsto h \otimes c_1 \otimes_H 1_H \otimes c_2 \otimes_H \cdots \otimes_H 1_H \otimes c_n$,
where $H \otimes C^{\otimes n}$ has right $H$-module structure from the diagonal action by $H$:
$(h \otimes c_1 \otimes \cdots \otimes c_n)x = hx\1 \otimes c_1 x\2 \otimes \cdots \otimes c_n x_{(n+1)}$. This follows from Eq.~(\ref{eq: coring}), cancellations of the type $M \otimes_H H \cong M$ for modules $M_H$, and an induction on $n$.
Suppose $d(C, \mathcal{M}_H) = n$, so that $C^{\otimes n} \sim C^{\otimes (n+1)}$ as right
$H$-modules (in the finite tensor category $\mathcal{M}_H$). Applying an additive functor,
it follows that $H \otimes C^{\otimes n} \sim H \otimes C^{\otimes (n+1)}$ as $H$-bimodules.
Applying the isomorphism~(\ref{eq: iso}) the coring depth satisfies $d(H \otimes C, H) \leq 2d(C, \mathcal{M}_H) + 1$.
Conversely, if $d(H \otimes C, H) = 2n+1$, so that $H \otimes C^{\otimes n} \sim H \otimes C^{\otimes (n+1)}$ from Eq.~(\ref{eq: iso}) again, we apply that additive functor $k \otimes_H -$ to the similarity
and obtain the similarity of right $H$-modules, $C^{\otimes n} \sim C^{\otimes (n+1)}$. Thus $2d(C, \mathcal{M}_H) +1 \leq d(H \otimes C, H)$.
\end{proof}
Now suppose $R \subseteq H$ is a left coideal subalgebra of a finite-dimensional Hopf algebra;
i.e., $\cop(R)\subseteq H \otimes R$. Let $R^+$ denote the kernel of the counit restricted to $R$.
Then $R^+H$ is a right $H$-submodule of $H$ and a coideal by a short computation given
in \cite[34.2]{BW}. Thus $Q := H/R^+H$ is a right $H$-module coalgebra (with a right $H$-module
coalgebra epimorphism $H \rightarrow Q$ given by $h \mapsto h + R^+H := \overline{h}$).
The $H$-coring $H \otimes Q$ has grouplike element $1_H \otimes \overline{1_H}$; in fact,
\cite[34.2]{BW} together with \cite{S} shows that this coring is Galois:
\begin{equation}
\label{eq: map}
H \otimes_R H \stackrel{\cong}{\longrightarrow} H \otimes Q
\end{equation}
via $x \otimes_R y \mapsto x y\1 \otimes \overline{y\2}$ (also noted in \cite{HKY} and in \cite{U}
for Hopf subalgebras). That $H_R$ is faithfully flat follows from the result that $R$ is a Frobenius algebra and $H_R$ is free \cite{S}. Note that an inverse to (\ref{eq: map}) is given
by $x \otimes \overline{z} \mapsto xS(z\1) \otimes_R z\2$ for all $x,z \in H$.
From Proposition~\ref{prop-entwinedepth}, Eqs.~(\ref{eq: map}) and~(\ref{eq: coringdepth=hdepth}) we note the following.
\begin{cor}
\label{cor-cue}
For a left coideal subalgebra $R$ in a Hopf algebra $H$, its h-depth
is related to the module depth of $Q$ by
\begin{equation}
\label{eq: cue}
d_h(R,H) = 2d(Q, \mathcal{M}_H) + 1.
\end{equation}
\end{cor}
\begin{proof}
The proof is sketched above for finite-dimensional $H$. For infinite-dimensional $H$, note
that the $H$-bimodule isomorphism in Eq.~(\ref{eq: map}) remains valid, as does Proposition~\ref{prop-entwinedepth} and Eq.~(\ref{eq: coringdepth=hdepth}).
\end{proof}
Suppose $R$ is a Hopf subalgebra of $H$. Then $Q$ is an $R$-module coalgebra by restriction.
A similar argument to the one above shows that $d(R,H)$ and $d(Q, \mathcal{M}_R)$
satisfy the inequalities in (\ref{eq: inequalityfordepth}).
Finally we recall that a \textit{$C$-Galois extension} $A \supseteq B$, where $C$ is a coalgebra and
$A$ a right $C$-comodule via coaction $\delta: A \rightarrow A \otimes C$, the subalgebra of coinvariants is characterized by satisfying $B =
\{ b \in A \| \, \forall a \in A, \delta(ba) = b\delta(a) \}$ and $\beta: A \otimes_B A \mapsto
A \otimes C$ given by $\beta(a \otimes a') = a \delta(a')$ is bijective. For example,
a left coideal subalgebra $R \subseteq H$ is a coalgebra-Galois extension with respect to
the ($H$-module) coalgebra $Q$, as sketched above (the details are in \cite[34.2]{BW}).
Of course, this applies to Hopf subalgebras and more particularly to finite group algebra extensions.
Then we see that coalgebra-Galois extensions have at least the range of values computed for
subgroup depth \cite{BDK, BKK, D, F, FKR, HHP}.
\subsection{A faithfully flat interlude} Let $C$ be a coalgebra and $H$-module quotient of $H$, with canonical epi $H \rightarrow C$ of right $H$-module coalgebras, and $A$ is a right $H$-comodule algebra, with the obvious $C$-comodule coaction $\delta: A \rightarrow A \otimes C$ given by $\delta(a) = a\0 \otimes \overline{a\1}$. Define the subalgebra $B = \{ b \in A \, | \, \delta(b) = b \otimes \overline{1_H} \}$. Now suppose that $D \subseteq B$ is a subalgebra for which the canonical (Galois) mapping $\beta: A \otimes_D A \stackrel{\cong}{\longrightarrow} A \otimes C$
given by $a \otimes_D a' \mapsto a\delta(a') = a{a'}\0 \otimes \overline{ {a'}\1}$, is an isomorphism.
If either of the natural modules $A_D$ or ${}_DA$ is faithfully flat, then $D = B$.
This follows from noting that for each $b \in B$, $\beta(1_A \otimes_D b) = b \otimes \overline{1_H} = \beta(b \otimes_D 1_A)$ and the next lemma.
\begin{lemma}
\label{lemma-ff}
If $b \otimes_D 1 = 1 \otimes_D b$ for $b \in A \supseteq D$ with $D$ a subring of $A$ such that
$A_D$ or ${}_DA$ is faithfully flat, then $b \in D$.
\end{lemma}
\begin{proof}
Recall that a flat module $A_D$ is faithfully flat iff for each module ${}_DN$ such that $A \otimes_D N = 0$,
it follows that $N = 0$. Now form the module $N = B'/D$, where $B' = \{ b \in A \, | \, b \otimes_D 1_A = 1_A \otimes b \}$. By an exercise with a commutative square, $A \otimes_D N = 0$, whence $D = B'$.
\end{proof}
\subsection{A normal element characterization of left ad-stabity for left coideal subalgebras}
Suppose $R$ is left coideal subalgebra of a finite-dimensional Hopf algebra $H$. Let $Q$ be its quotient
right $H$-module coalgebra defined above. By Skryabin's Freeness Theorem \cite{S} $\dim R$ divides $\dim H$,
and $(R, \eps)$ is an augmented Frobenius algebra, so $R$ has a nonzero right integral $t_R$ unique up to scalar multiplication \cite[Ch. 6]{NEFE}. Then one proves just as in \cite[Lemma 3.2]{K2013} that
\begin{equation}
\label{isomorph}
Q \stackrel{\cong}{\longrightarrow} t_R H
\end{equation}
via $h + R^+H \mapsto t_R h$.
Recall that a subalgebra $R$ in a Hopf algebra $H$ is stable under the left adjoint action of $H$
if $h\1 r S(h\2) \in R$ for all $r \in R$, $h \in H$. One briefly says that $R$ is left ad-stable
(or left normal \cite{B2012}). In section~3 of \cite{B2012}, Burciu shows how to prove the following.
\begin{lemma}(cf.\ \cite{B2012})
\label{lem-straightenedout}
A left coideal subalgebra $R$ in a finite-dimensional Hopf algebra $H$ is left ad-stable if and only if $HR^+ \subseteq R^+H$ if and only if the subalgebra $R \subseteq H$ has right depth $2$.
\end{lemma}
\begin{proof} For the convenience of the reader, we sketch the proof. If $R$ is left ad-stable in $H$,
$r \in R^+$ and $h \in H$, then $hr = h\1 r S(h\2) h\3 \in R^+H$.
If $HR^+ \subseteq R^+H$, the isomorphism $\beta: H \otimes_R H \stackrel{\cong}{\rightarrow}
H \otimes Q$ in Eq.~(\ref{eq: map}) is an $H$-$R$-bimodule morphism, since $Q_R$ is trivial.
Hence ${}_H H \otimes_R H_R \cong (\dim Q) \cdot {}_H H_R$ and $R$ has right depth $2$ in $H$.
If ${}_H H \otimes_R H \oplus * \cong n \cdot {}_H H_R$, then applying $k \otimes_H -$ yields
$Q_R \oplus * \cong n \cdot k_R$. Since $R^+ = \Ann\, k_R$, it follows easily that $HR^+ \subseteq R^+H$.
If $HR^+ \subseteq R^+H$, then $R^+ H$ is a bi-ideal in $H$, $Q$ is a bialgebra and has a natural $H$-bimodule structure from multiplication in $H$. Also $\beta$ defined above satisfies $\beta(1_H \otimes_R h) =
\beta(h \otimes_R 1_H)$ for $$h \in H^{\mbox{co}\, Q} = \{ x \in H | x\1 \otimes \overline{x\2} = x \otimes \overline{1} \}. $$
Of course, $R \subseteq H^{\mbox{co}\, Q}$.
Since $\beta$ is bijective, $h \otimes_R 1 = 1 \otimes_R h$, thus by Lemma~\ref{lemma-ff},
$R = H^{\mbox{co}\, Q}$. But $H^{\mbox{co}\, Q}$ is left ad-stable by an argument used in
\cite[3.4.2]{M} modified slightly by the remark in the first sentence of this paragraph.
\end{proof}
The following generalizes \cite[5.2]{K2013}.
\begin{theorem}
\label{th-leftrightleft}
A left coideal subalgebra $R$ in a finite-dimensional Hopf algebra $H$ is left ad-stable if and only if its right integral $t_R$ is a normal element in $H$.
\end{theorem}
\begin{proof}
($\Rightarrow$) One argues as in \cite[5.2]{K2013}, using Eq.~(\ref{isomorph}) and $Q$ is under the hypothesis a trivial right $R$-module, that $h\1 t_R Sh\2$
is a nonzero right integral in $R$ for any $h \in H$, so that $$ht_R = h\1 t_R S(h\2) h_3 \in t_R H.$$
Hence, $Ht_R \subseteq t_R H$. For the opposite inclusion, note that $$t_R S(h) = S(h\1) h\2 t_R S(h\3) \in Ht_R.$$
($\Leftarrow$) If $Ht_R = t_R H$, then $\Ann \, Q_R \supseteq R^+$. Thus $HR^+ \subseteq R^+H$.
We conclude that $R$ is left ad-stable in $H$ from Lemma~\ref{lem-straightenedout}.
\end{proof}
One says that a subalgebra $R$ of a Hopf algebra $H$ is right ad-stable if $S(h\1)rh\2 \in R$
for all $h \in H, r \in R$.
Applying the antipode anti-automorphism to the above theorem, one obtains (left as an exercise) the correct formulation and proof of the opposite (and equivalent) theorem.
\begin{cor}
\label{rightversion}
A right coideal
subalgebra is right ad-stable in a finite-dimensional Hopf algebra $H$ iff its left integral $t_L$ is a normal element in $H$.
\end{cor}
\begin{example}
\begin{rm}
Consider the $8$-dimensional small quantum group $H = H_8$ given as an algebra in Example~\ref{example-8dim}, with coalgebra
structure given by $\cop(K) = K \otimes K$, $\cop(E) = E \otimes 1 + K \otimes E$, and $\cop(F) = F \otimes K + 1 \otimes F$. It follows that $S(K) = K$, $S(E) = -KE$ and $S(F) = - FK$.
Consider the $2$-dimensional Frobenius subalgebra $R$ generated by $E$ (the ``ring of dual numbers''). Notice that $R$ is a left coideal subalgebra, since $\cop(E) \in H \otimes R$.
It is left ad-stable since $FEK + ES(F) = 0$. Also $t_R = E$ ($= t_L$) and $HE = EH$ is the $4$-dimensional
vector subspace of $H$ with basis $\{ E, EF, EK, EFK \} $: equivalently, $HR^+ = R^+ H$.
However, $R$ is not right ad-stable, since $S(F)EK + EF = 2EF \not \in R$. The Frobenius subalgebra
of dimension $2$ generated by $F$ in $H$ is however a right coideal subalgebra that is right ad-stable
with normal integral element, and provides an example of Corollary~\ref{rightversion}. This example does not
contradict the correct formulation of the opposite of Lemma~\ref{lem-straightenedout}:
\end{rm}
\end{example}
\begin{prop}[cf.\ \cite{B2012}]
A right coideal subalgebra $R$ of $H$
is right ad-stable iff $R^+H \subseteq HR^+$ iff $R$ has left depth $2$ in $H$.
\end{prop}
\section{The coring of an entwining structure of a comodule algebra and a module coalgebra}
\label{four}
Let $H$ be a Hopf algebra. Suppose $A$ is a right $H$-comodule algebra, i.e., there is a coaction
$\rho: A \rightarrow A \otimes H$, denoted by $\rho(a) = a\0 \otimes a\1$ that is an algebra homomorphism and $(A, \rho)$ is a right $H$-comodule \cite{BW, CMZ, M}. Moreover, let
$(C, \cop_C, \eps_C)$ be a right $H$-module coalgebra, i.e., a coalgebra in the tensor category
$\mathcal{M}_H$ introduced in more detail in Section~\ref{three}.
\begin{example}
\begin{rm}
The Hopf algebra $H$ is right $H$-comodule algebra over itself, where $\rho = \cop$. Given a Hopf subalgebra
$R \subseteq H$ the quotient module $Q$ defined in Section~1 as $Q = H/ R^+H$ is a right $H$-module
coalgebra.
Note that $(H, \cop, \eps)$
is also a right $H$-module coalgebra. The canonical epimorphism $H \rightarrow Q$ denoted by $h \mapsto \overline{h}$ is an epi of right $H$-module coalgebras, and module $Q_H$ has $\overline{1_H}$ as a cyclic generator.
Of course, if $H = k$ is the trivial one-dimensional Hopf algebra, $A$ may be any $k$-algebra and $C$
any $k$-coalgebra.
\end{rm}
\end{example}
The mapping $\psi: C \otimes A \rightarrow A \otimes C$ defined by
$\psi(c \otimes a) = a\0 \otimes c a\1$ is an entwining as the reader may easily check (the so-called Doi-Koppinen entwining
\cite[33.4]{BW}, \cite[2.1]{CMZ}, which includes the case considered in Section~\ref{three}).
From Proposition~\ref{prop-1-1} it follows that $A \otimes C$ has $A$-coring structure
\begin{equation}
\label{eq: comodulealgdiagonal}
a(a' \otimes c)a'' = aa' {a''}\0 \otimes c {a''}\1
\end{equation}
which defines the bimodule ${}_A(A \otimes C)_A$. The coproduct is given by $\id_A \otimes \cop_C$
and the counit by $\id_A \otimes \eps_C$.
Suppose in this paragraph that $H$ is a finite-dimensional Hopf algebra over an algebraically closed field, in which case $\mathcal{M}_H$
is a finite tensor category \cite{EO}.
Notice in the equation above that the right $A$-module is given by a version of the diagonal action
in which the category $\mathcal{M}_A$ is a module category over $\mathcal{M}_H$ \cite{AM, EO}.
\begin{prop}
\label{prop-comodalgentwinedepth}
The depth of the $A$-coring $A \otimes C$ and the depth of the $H$-module $C$ are related by
$d(A \otimes C, A) \leq 2d(C, \mathcal{M}_H) + 1$.
\end{prop}
\begin{proof}
The $n$-fold tensor product of $A \otimes C$ over $A$ reduces to the $A$-bimodule isomorphism
\begin{equation}
\label{eq: iso2}
(A \otimes C)^{\otimes_A n} \cong A \otimes C^{\otimes n}
\end{equation}
via the mapping $$a \otimes c_1 \otimes \cdots \otimes c_n \mapsto a \otimes c_1 \otimes_A 1_A \otimes c_2 \otimes_A \cdots \otimes_A 1_A \otimes c_n,$$
where $A \otimes C^{\otimes n}$ has right $A$-module structure from the diagonal action by $A$:
$(a \otimes c_1 \otimes \cdots \otimes c_n)b = ab\0 \otimes c_1 b\1 \otimes \cdots \otimes c_n b_{(n)}$. This follows from Eq.~(\ref{eq: comodulealgdiagonal}), cancellations of the type $M \otimes_A A \cong M$ for modules $M_A$, and an induction on $n$.
Suppose $d(C, \mathcal{M}_H) = n$, so that $C^{\otimes n} \sim C^{\otimes (n+1)}$ as right
$H$-modules (in the finite tensor category $\mathcal{M}_H$). Applying an additive functor,
it follows that $A \otimes C^{\otimes n} \sim A \otimes C^{\otimes (n+1)}$ as $A$-bimodules.
Then applying the isomorphism~(\ref{eq: iso2}) obtain $d(A \otimes C, A) \leq 2d(C, \mathcal{M}_H) + 1$.
\end{proof}
Suppose $C$ has grouplike $g'$. Then $g = 1_A \otimes g'$ is a grouplike
element in $A \otimes C$. Then we have from Eq.~(\ref{eq: coringdepth=hdepth}) and Proposition~\ref{prop-comodalgentwinedepth} the proof of the following.
\begin{lemma}
\label{lem-ineq}
Suppose $A \otimes C := \mathcal{C}$ is a Galois coring with $B = A_g^{\mbox{co}\, \mathcal{C}}$
(or just that $B$ is a subalgebra of $A$ such that $A \otimes_B A \cong \mathcal{C}$ as
natural and right-diagonal $A$-bimodules, respectively).
Then the h-depth of the subalgebra $B \subseteq A$ satisfies the inequality, $d_h(B,A) \leq 2d(C, \mathcal{M}_H) +1$.
\end{lemma}
\subsection{Crossed products as comodule algebras}
Now let $A$ be an associative crossed product $D \#_{\sigma} H$ for some Hopf algebra algebra $H$ and twisted $H$-module algebra $D$ with $2$-cocycle $\sigma: D \otimes D \rightarrow H$ that
is convolution-invertible \cite[chapter 7]{M}. Then $A$ is a right $H$-comodule algebra via
$\rho = \id_D \otimes \cop$, which is quite obvious from the formula for multiplication in
$D \#_{\sigma} H$: ($h,x,y \in H, d,d' \in D$)
\begin{equation}
\label{eq: crossprod}
(d \# h)(d' \# x) = d(h\1 \cdot d')\sigma(h\2, x\1) \# h\3 x\2
\end{equation}
Additionally, the formulas for $D$ a twisted right $H$-module and $\sigma$ a $2$-cocycle are useful:
\begin{equation}
\label{eq: twistedmodule}
h \cdot (x \cdot d) = \sigma(h\1, x\1) (h\2 x\2 \cdot d) \sigma^{-1}(h\3, x\3)
\end{equation}
\begin{equation}
\label{eq: twococycle}
(h\1 \cdot \sigma(x\1, y\1)) \sigma( h\2, x\2 y\2) = \sigma(h\1, x\1) \sigma( h\2 x\2, y)
\end{equation}
\begin{example}
\begin{rm}
For $H = kG$ a group algebra, one obtains from this the familiar conditions of the group crossed product, $A = D \# kG$,
where for $g,h, s \in G$ and $d, d' \in D$,
\begin{eqnarray}
g \cdot (h \cdot d) & = & \sigma(g, h) (gh \cdot d)\sigma(g,h)^{-1} \label{action} \\
(g \cdot \sigma(h,s)) \sigma(g, hs) & = & \sigma(g,h) \sigma(gh, s) \label{cocycle}\\
(d \# g) (d' \# h) & = & d(g \cdot d') \sigma(g,h) \# gh
\end{eqnarray}
Note that $\sigma^{-1}(g,h)$ is interchangeable with $\sigma(g,h)^{-1}$.
\end{rm}
\end{example}
\begin{example}
\begin{rm}
If $\sigma$ is trivial, $\sigma = 1_D$, then the crossed product reduces to the skew group algebra
$D \ast G$, where $D$ is a $G$-module and multiplication is given by $(d \# g)(d' \# h) = d (g \cdot d') \# gh$; this is a well-known setup in Galois theory of fields and commutative algebras. More generally,
$D \#_{\sigma} H$ is clearly isomorphic to the smash product $D \# H$ if $\sigma(x,y) = \eps(x)\eps(y)1_H$:
see Eq.~(\ref{eq: crossprod}). Then $D$ is a left $H$-module algebra. (If the action is Galois,
then the endomorphism algebra of $D$ over its invariant subalgebra is isomorphic to the smash
product $D \# H$ \cite[8.3.3]{M}.)
\end{rm}
\end{example}
\begin{example}
\begin{rm}
If $\sigma$ is instead nontrivial, but the action of $G$ on $D$ is trivial, i.e. $g \cdot d = d$
for each $g \in G$ and $d \in D$, then $D \#_{\sigma} kG = D_{\sigma}[G]$, the twisted group
algebra \cite{P}, with multiplication given by $(d \# g)(d' \# h) = dd' \sigma(g,h) \# gh$. For example,
the real quaternions are a twisted group algebra of $\R$ with $G \cong \Z_2 \oplus \Z_2$ and
$\sigma = \pm 1$.
\end{rm}
\end{example}
\begin{example}
\label{example-groupascrossedproduct}
\begin{rm}
For any group $G$ and normal subgroup $N$ of $G$, the group algebra is a crossed product of the quotient group algebra acting on
the subgroup $N$ as follows \cite[7.1.6]{M}. Let $Q$ denote the group $G/N$. For each coset $q \in Q$, let $\gamma(q)$ denote a coset representative, choosing $\gamma(\overline{1_G}) = 1_G$. It
is an exercise then to show that with $\sigma(\overline{x},\overline{y}) := \gamma(\overline{x}) \gamma(\overline{y}) \gamma(\overline{x}\overline{y})^{-1} \in N$ and action
of $Q$ on $kN$ given by $\overline{x}\cdot n = \gamma(\overline{x})n \gamma(\overline{x})^{-1}$, one has
$kG = kN \#_{\sigma} kQ$.
\end{rm}
\end{example}
Let $R \subseteq H$ be a Hopf subalgebra pair. Again form the quotient module $Q = H / R^+H$,
which is a right $H$-module coalgebra with canonical epi of right $H$-module coalgebras
$\pi: H \rightarrow Q$, $h \mapsto \overline{h}$. Then the crossed product
$A = D \#_{\sigma} H$, which is a right $H$-comodule algebra, is also a right $Q$-comodule
via $A \rightarrow A \otimes H \stackrel{A \otimes \pi}{\longrightarrow} A \otimes Q$, the composition $\rho$ being the coaction given by $\rho(d \# h) = d \# h\1 \otimes \overline{h\2}$.
Then one sees that $B : = D \#_{\sigma} R \subseteq A^{{\mbox co}\, Q} $, the $Q$-coinvariants defined as the subalgebra $\{ b \in A \| \, \forall a \in A, \rho(ba) = b\rho(a) \}$; the inclusion follows from noting $\overline{rh} = \eps(r) \overline{h}$ for all $r \in R, h \in H$.
Note that $Q$ has the grouplike element $\overline{1_H}$, so that $g = 1_A \otimes \overline{1_H}$
is a grouplike element in the $A$-coring $\mathcal{C} = A \otimes Q$. From the previous observation and
Eq.~(\ref{eq: comodulealgdiagonal}), it follows that $B = D \# R \subseteq A_g^{\mbox{co}\, \mathcal{C}}$.
Suppose
that $A \otimes_B A \cong A \otimes Q$ as $A$-bimodules, or more strongly $\mathcal{C}$
is a Galois $A$-coring with respect to $B = A_g^{\mbox{co}\, \mathcal{C}}$. Then putting together Lemma~\ref{lem-ineq} with Corollary~\ref{cor-cue} and Eq.~(\ref{eq: cue}), we arrive at the inequality of minimum
h-depths,
\begin{equation}
\label{eq: mainresult}
d_h(D \#_{\sigma} R, D \#_{\sigma} H) \leq d_h(R, H),
\end{equation}
the main aim of this paper: we next set about establishing this supposition for finite rank group algebra extensions and certain crossed product Hopf algebra extensions.
\begin{prop}
\label{prop-sch}
Consider the $H$-comodule algebra $A = D \#_{\sigma} H$. Let $B = D \#_{\sigma} R$ be
an $R$-comodule subalgebra of $A$, where $R$ is a Hopf subalgebra of $H$. Suppose the condition~(\ref{eq: cond}) below is satisfied. Then $A \otimes_B A \cong A \otimes Q$ as $A$-bimodules,
where $Q = H/ R^+H$.
If $A_B$ is faithfully flat, then $A$ is a coalgebra $Q$-Galois extension of $B$, i.e.,
$A \otimes Q$ is a Galois $A$-coring with $B$ the invariant subalgebra.
\end{prop}
\begin{proof}
We investigate if the canonical mapping $$\beta(a \otimes_B a') = aga' = a(d \# h\1) \otimes \overline{h\2}$$
(where $a' = d \# h \in D \# H = A$) is bijective, and the module $A_B$ is faithfully flat.
The map $\beta: A \otimes_B A \stackrel{\cong}{\longrightarrow} A \otimes Q$ is well-defined:
suppose $d' \in D, r \in R$, we check that $$\beta(a(d' \# r) \otimes_B (d \# h)) = \beta(a \otimes_B (d'(r\1 \cdot d)\sigma(r\2,h\1) \# r\3 h\2))$$
i.e., $ a(d' \# r)(d \# h\1) \otimes \overline{h\2} = ad'(r\1 \cdot d)\sigma(r\2,h\1) \# r\3 h\2 \otimes \overline{r\4 h\3}$, which is clear since $\overline{rh} = \eps(r) \overline{h}$ for each $r \in R, h \in H$.
We note that $\beta^{-1}: A \otimes Q \rightarrow A \otimes_B A$ is given by
$$\beta^{-1}(a \otimes \overline{h}) = a(\sigma^{-1}(S(h\2), h\3) \# S(h\1)) \otimes_B (1_D \# h\4).$$
The computation $\beta^{-1} \circ \beta = \id_{A \otimes_B A}$ and the computation
$\beta \circ \beta^{-1} = \id_{A \otimes Q}$ follow from $\gamma: H \rightarrow A$,
defined by $\gamma(h) = 1_D \# h$, having convolution-inverse
$\mu: H \rightarrow A$, defined by $$\mu(h) = \sigma^{-1}(S(h\2),h\3) \# S(h\1),$$
on \cite[p.\ 109]{M}. We are left with verifying that $\beta^{-1}$ is well-defined, i.e., vanishes when $h = rh'$
for some $r \in R^+, h' \in H$. Dropping the prime, this becomes the condition
\begin{equation}
\label{eq: cond}
\sigma^{-1}(S(h\2) S(r\2), r\3 h\3) \# S(h\1) S(r\1) \otimes_B (1_D \# r\4 h\4) = 0
\end{equation}
for all $r \in R, h \in H$ such that $\eps(r) = 0$.
Finally, suppose either natural module, $A_B$ or ${}_BA$ is faithfully flat. Then $A$ is a $Q$-Galois extension of $B$ (cf.\ \cite[34.10]{BW}, \cite[1.2, 3.1]{Sch}).
\end{proof}
\begin{cor}
Let $R \subseteq H$ be a finite-dimensional Hopf subalgebra pair, and $A$ a left $H$-module algebra. Then
h-depth satisfies
\begin{equation}
\label{eq: smashproddepth}
d_h(A \# R, A \# H) \leq d_h(R,H).
\end{equation}
\end{cor}
\begin{proof}
If $\sigma(x,y) = \eps(x)\eps(y)1_D$ in Proposition~\ref{prop-sch}, then the crossed products in the theorem are smash products. The corollary follows if Eq.~(\ref{eq: cond}) is satisfied with this choice of $\sigma$. But the left-hand side
reduces to $1 \# S(h\1) \otimes_B 1 \# S(r\1) r\2 h\2 $, indeed equal to zero for $r \in R^+$.
Finally, the natural module ${}_RH$ is free by Nichols-Zoeller, so that $A \# H$ is easily shown to be free as a natural left module
over $A \# R$.
\end{proof}
\begin{prop}
Suppose $G$ is a group and $S$ is a subgroup of $G$ such that $|G:S| < \infty$. Let $H = kG$, $R = kS$ and $A$,$B$
be crossed products of the group algebras $H$ and $R$ with left twisted $H$-module algebra $D$ as above. Then $A$ is a $Q$-Galois extension of $B$ (where $Q$ is the permutation module of right
cosets of $S$ in $G$).
\end{prop}
\begin{proof}
It suffice to check Eq.~(\ref{eq: cond}) for $h \in G$ and $r \in S$. Note that $1 - r \in R^+$
and such elements form a $k$-basis. Also note that $$\cop^{n-1}(1-r) = 1 \otimes \cdots \otimes 1
- r \otimes \cdots \otimes r$$ ($n$ $1$'s and $n$ $r$'s on the right-hand side of the equation). Eq.~(\ref{eq: cond}) becomes
\begin{eqnarray}
\label{eq: simp}
\sigma^{-1}(h^{-1}r^{-1}, rh) \# h^{-1} r^{-1} \otimes_B (1_D \# rh) && \nonumber \\ &= & \sigma^{-1}(h^{-1},h) \# h^{-1} \otimes_B (1_D \# h)
\end{eqnarray}
The inverse of the $2$-cocycle Equation~(\ref{cocycle}) is ($\forall x,y,z \in G$)
\begin{equation}
\label{eq: twococycleinverse}
\sigma^{-1}(xy,z) \sigma^{-1}(x,y) = \sigma^{-1}(x,yz) [ x \cdot \sigma^{-1}(y,z)]
\end{equation}
Letting $x = h^{-1}, y = r^{-1}, z = rh$, the left-hand side of Eq.~(\ref{eq: simp}) becomes
$$ \sigma^{-1}(h^{-1},h)[h^{-1} \cdot \sigma^{-1}(r^{-1}, rh)] \sigma(h^{-1}, r^{-1}) \# h^{-1} r^{-1} \otimes_B (1_D \# rh) $$
$$ = (\sigma^{-1}(h^{-1}, h) \# h^{-1}) ( \sigma^{-1}(r^{-1}, rh) \# r^{-1}) \otimes_B (1_D \# rh)$$
$$ = (\sigma^{-1}(h^{-1}, h) \# h^{-1}) \otimes_B (\sigma^{-1}(r^{-1}, rh) \sigma(r^{-1}, rh) \# h)$$
which equals the right-hand side of Eq.~(\ref{eq: simp}).
The natural module ${}_BA$ is free by a short argument using coset representatives $g_1,\ldots,g_q$
giving a left $R$-basis for $H$. I.e., $|G: S| = q$, $\sum_{i=1}^q Rg_i = H$ and $\sum_{i=1}^q r_i g_i = 0$
implies each $r_i = 0$. Then $\sum_{i=1}^q (D \#_{\sigma} R)(1_D \# g_i) = \sum_i D\sigma(R, g_i) \# Rh_i = D \# H$, since $\sigma(x,y)$ is invertible in $D$ for all $x,y \in G$.
Suppose $$0= \sum_{i=1}^q (d_i \# r_i)(1 \# g_i) = \sum_i d_i \sigma({r_i}\1, g_i) \# {r_i}\2 g_i $$
where each $r_i$ is a linear combination of group elements $s_{ij}$ in $S$. In this case each
$s_{ij} g_i \in Sg_i$ and it follows from the partition of $G$ into left cosets and the equation just above that each $d_i \# r_i = 0$.
Finally $Q = H / R^+H \cong k[S \setminus G]$ by noting that $R^+$ has $k$-basis $\{ 1 - s \, : \, s \in S \}$.
\end{proof}
\begin{theorem}
Let $G$ be a group with subgroup $H$ of finite index, $A$ a left twisted $G$-module $k$-algebra, and $A \#_{\sigma} G$ an (associative) crossed product. Then h-depth satisfies
\begin{equation}
\label{eq: hdepthineq}
d_h(A \#_{\sigma} H, A \#_{\sigma} G) \leq d_h(kH, kG).
\end{equation}
\begin{proof}
Follows from the previous proposition and inequality (\ref{eq: mainresult}).
\end{proof}
\end{theorem}
\subsection{General results for Hopf algebra $H$} We may extend Eq.~(\ref{eq: hdepthineq}) to a finite-dimensional Hopf algebra extension and crossed product algebra extension as follows.
\begin{theorem}
Suppose $H$ is a finite-dimensional Hopf algebra, with $R$ a Hopf subalgebra, $A = D \#_{\sigma} H$, $B = D \#_{\sigma} R$ and $Q = H / R^+H$ as above. Then
\begin{equation}
d_h(B,A) \leq d_h(R,H).
\end{equation}
\end{theorem}
\begin{proof}
The proof may be made to follow from \cite[3.6]{Sch} and Eq.~(\ref{eq: mainresult}), but we provide some more details as a convenience to the reader.
Note first that $\beta: A \otimes_B A \rightarrow A \otimes Q$, $x \otimes y \mapsto xy\0 \otimes \overline{y\1}$ in the proof of Proposition~\ref{prop-sch} is shown to be surjective from the formal
inverse $\beta^{-1}$ defined there and the equation $\beta \circ \beta^{-1} = \id_{A \otimes Q}$.
That $\beta$ is injective follows from \cite{Sch} in the following way using norms of augmented Frobenius algebras, freeness of $H$ over $R$ and the augmentation of the convolution algebra $Q^*$ induced from the grouplike element $\overline{1_H} \in Q$. Let $B$ denote $A^{\mbox{co}\, Q}$
for this argument. Let $\lambda_H, \lambda_R$ be nonzero left integrals on $H$ and $R$, respectively. Let $\Gamma' \in R$ satisfy $\lambda_R \Gamma' = \eps$, so that $\Gamma'$ is
a nonzero right integral in $R$. Define $\lambda: Q \rightarrow k$ by $\lambda(\overline{h}) = \lambda_H(\Gamma' h)$, and note the left integral property, $\overline{h\1} \lambda(\overline{h\2})$ for all $h \in H$
\cite[p.\ 307, (i)]{Sch}. By \cite[p.\ 307, (ii)]{Sch}, there is $\Lambda \in H$ such that $\eps_Q =
\Lambda \lambda$, which follows from expressing a right norm $\Gamma$ of $\lambda_H$
as $\Gamma = h \Gamma'$, then applying a Nakayama automorphism $\alpha$ of $H$
to express $\Lambda = \alpha(h)$. Next define as in \cite[p.\ 303, (1)]{Sch} $f: {}_BA \rightarrow {}_BB$ by $f(a) = a\0 \lambda(\overline{a\1})$ using the left integral property.
In order to continue, note that $\mbox{can}: A \otimes A \rightarrow A \otimes H$, given
by $x \otimes y \mapsto xy\0 \otimes y\1$ is surjective, since given $a \otimes h \in A \otimes H$,
$\mbox{can}(a(\sigma^{-1}(S(h\2), h\3) \# S(h\1)) \otimes 1_D \# h\4) = a \otimes h$
by the computation on \cite[p.\ 109]{M}. It follows from the bijectivity of the antipode
that $\mbox{can'}: A \otimes A \rightarrow A \otimes H$ given by $x \otimes y \mapsto x\0 y \otimes x\1$ is also surjective; cf. \cite[p.\ 124]{M}. Let $\mbox{can'}(\sum_i r_i \otimes \ell_i) = 1_A \otimes \Lambda$. Then \cite[p.\ 303, (2)]{Sch} shows that $a = \sum_i f(ar_i)\ell_i$ for each $a \in A$.
Applying this projectivity equation to $\beta(\sum_k x_k \otimes_B y_k) = 0$, \cite[p.\ 303, (3)]{Sch} shows that
$\sum_i x_k \otimes_B y_k = 0$.
Since $H$ is a free $R$-module, a faithfully flat descent along the crossed product extension shows
that $A \#_{\sigma} R = A^{\mbox{co}\, Q}$.
Since $\beta: A \otimes_B A \stackrel{\cong}{\longrightarrow} A \otimes Q$ as $A$-bimodules,
it follows from Eq.~(\ref{eq: mainresult}), and the proof preceding it, that the inequality in
the theorem holds.
\end{proof}
\begin{cor}
\label{cor-core}
Suppose $N$ is a normal subgroup of a finite group $G$ contained in a subgroup $H \leq G$. Then the h-depth satisfies the equality, $ d_h(kH,kG) = d_h(k[H/N], k[G/N])$.
\end{cor}
\begin{proof}
This follows from Corollary~\ref{cor-cue}, since in either case $Q \cong k[G \setminus H]$, while $N$ acts trivially
on this $G$-module. Note that $I = kGkN^+$ is a Hopf ideal in $kG$ (generated by $\{ g - gn \| \, g \in G, n \in N \}$), which annihilates $Q$, and we apply Lemma~\ref{lemma-modI}.
A second proof is to apply the observation in Example~\ref{example-groupascrossedproduct} that one has $kG = kN \#_{\sigma} k[G/N]$ and similarly $kH = kN \#_{\sigma} k[H/N]$. Apply now the inequality~(\ref{eq: mainresult}) to obtain $d_h(kH,kG) \leq d_h(k[H/N], k[G/N])$. The inequality $d_h(k[H/N], k[G/N]) \leq d_h(kH,kG)$ has a proof very similar to the proof of Theorem~\ref{prop-sigma} using the relatively nice Hopf ideal $I = kG kN^+$.
\end{proof}
\subsection{Acknowledgements} The authors thank Dr.\ C. Young for stimulating conversations about depth-preserving algebra homomorphisms. Research for this paper was funded by the European Regional Development Fund through the programme {\tiny COMPETE}
and by the Portuguese Government through the FCT under the project
\tiny{ PE-C/MAT/UI0144/2013.nts}.
|
\section{Introduction}
Linear state-space models (LSSM) are widely used in time-series analysis and
modelling of dynamical systems \cite{Bar-Shalom:2001,Shumway:2000}. They assume
that the observations are generated linearly from hidden states with a linear
dynamical model that does not change with time. The assumptions of linearity and
constant dynamics make the model easy to analyze and efficient to learn.
Most real-world processes cannot be accurately described by linear Gaussian
models, which motivates more complex nonlinear state-space models (see, e.g.,
\cite{Ghahramani:1999,valpola2002}). However, in many cases processes behave
approximately linearly in a local regime. For instance, an industrial process
may have a set of regimes with very distinct but linear dynamics. Such
processes can be modelled by switching linear state-space models
\cite{Ghahramani:1998:sssm,Pavlovic:2000} in which the transition between a set
of linear dynamical models is described with hidden Markov models. Thus, these
models have a small number of states defining their dynamics.
Instead of having a small number of possible states of the process dynamics,
some processes may have linear dynamics that change continuously in time. For
instance, physical processes may be characterized by linear stochastic partial
differential equations but the parameters of the equations may vary in time.
Simple climate models may use the advection-diffusion equation in which the
diffusion and the velocity field parameters define how the modelled quantity
mixes and moves in space and time. In a realistic scenario, these parameters
are time-dependent because, for instance, the wind modelled by the velocity
field changes with time.
This paper presents a Bayesian linear state-space model with time-varying
dynamics. The dynamics at each time is formed as a linear combination of a set
of state dynamics matrices, and the weights of the linear combination follow a
linear Gaussian dynamical model. The main difference to switching LSSMs is that
instead of having a small number of dynamical regimes, the proposed model allows
for an infinite number of them with a smooth transition between them. Thus, the
model can adapt to small changes in the system. This work is an extension of an
abstract \cite{raiko2010drifting} which presented the basic idea without the
Bayesian treatment. The model bears some similarity to relational feature
learning in modelling sequential data \cite{michalski2014}.
Posterior inference for the model is performed using variational Bayesian (VB)
approximation because the exact Bayesian inference is intractable
\cite{Beal:2003}. In order for the VB learning algorithm to converge fast, the
method uses a similar parameter expansion that was introduced in
\cite{Luttinen:2013}. This parameter expansion is based on finding the optimal
rotation in the latent subspace and it may improve the speed of the convergence
by several orders of magnitude.
The experimental section shows that the proposed LSSM with time-varying dynamics
is able to learn the varying dynamics of complex physical processes. The model
predicts the processes better than the classical LSSM and the LSSM with switching
dynamics. It finds latent processes that describe the changes in the dynamics
and is thus able to learn the dynamics at each time point accurately. These
experimental results are promising and suggest that the time-varying dynamics
may be a useful tool for statistical modelling of complex dynamical and physical
processes.
\section{Model}
\label{sec:model}
Linear state-space models assume that a sequence of $M$-dimensional observations
$(\mx{y}_1, \ldots, \mx{y}_N)$ is generated from latent $D$-dimensional states
$(\mx{x}_1, \ldots, \mx{x}_N)$ following a first-order Gaussian Markov process:
\begin{align}
\mathbf{y}_n &= \mathbf{C} \mathbf{x}_n + \mathrm{noise},
\\
\mathbf{x}_n &= \mathbf{W} \mathbf{x}_{n-1} + \mathrm{noise},
\label{eq:static_x}
\end{align}
where noise is Gaussian, $\mx{W}$ is the $D\times D$ state dynamics matrix and
$\mx{C}$ is the $M \times D$ loading matrix. Usually, the latent space
dimensionality $D$ is assumed to be much smaller than the observation space
dimensionality $M$ in order to model the dependencies of high-dimensional
observations efficiently. Because the state dynamics matrix is constant, the
model can perform badly if the dynamics of the modelled process changes in time.
In order to model changing dynamics, the constant dynamics in
\eqref{eq:static_x} can be replaced with a state dynamics matrix $\mx{W}_n$
which is time-dependent. Thus, \eqref{eq:static_x} is replaced with
\begin{align}
\mathbf{x}_n &= \mathbf{W}_n \mathbf{x}_{n-1} + \mathrm{noise}.
\end{align}
However, modelling the unknown time dependency of $\mathbf{W}_n$ is a
challenging task because for each $\mx{W}_n$ there is only one transition
$\mx{x}_{n-1} \rightarrow \mx{x}_n$ which gives information about each
$\mathbf{W}_n$.%
Previous work modelled the time-dependency using switching state dynamics
\cite{Pavlovic:2000}. It means having a small set of matrices $\mathbf{B}_1,
\ldots, \mathbf{B}_K$ and using one of them at each time step:
\begin{align}
\mathbf{W}_n &= \mathbf{B}_{z_n}, \label{eq:switching_W}
\end{align}
where $z_n\in \{1,\ldots,K\}$ is a time-dependent index. The indices $z_n$ then
follow a first-order Markov chain with an unknown state-transition matrix. The
model can be motivated by dynamical processes which have a few states with
different dynamics and the process jumps between these states.
This paper presents an approach for continuously changing time-dependent
dynamics. The state dynamics matrix is constructed as a linear combination of
$K$ matrices:
\begin{align}
\mathbf{W}_n &= \sum^K_{k=1} s_{kn} \mathbf{B}_k . \label{eq:varying_W}
\end{align}
The mixing weight vector $\mx{s}_n=\begin{bmatrix} s_{1n} & \ldots &
s_{Kn} \end{bmatrix}^{\operatorname{T}}$ varies in time and follows a first-order Gaussian
Markov process:
\begin{align}
\mathbf{s}_n &= \mathbf{A} \mathbf{s}_{n-1} + \mathrm{noise},
\end{align}
where $\mathbf{A}$ is the $K\times{}K$ state dynamics matrix of this latent
mixing-weight process. The model with switching dynamics in
\eqref{eq:switching_W} can be interpreted as a special case of
\eqref{eq:varying_W} by restricting the weight vector $\mx{s}_n$ to be a binary
vector with only one non-zero element. However, in the switching model,
$\mx{s}_n$ would follow a first-order Markov chain, which is different from the
first-order Gaussian Markov process used in the proposed model. Compared to
models with switching dynamics, the model with time-varying dynamics allows the
state dynamics matrix to change continuously and smoothly.
The model is motivated by physical processes which roughly follow stochastic
partial differential equations but the parameters of the equations change in
time. For instance, a temperature field may be modelled with a stochastic
advection-diffusion equation but the direction of the wind may change in time,
thus changing the velocity field parameter of the equation.
\subsection{Prior Probability Distributions}
We give the proposed model a Bayesian formulation by setting prior probability
distributions for the variables. It roughly follows the linear state-space
model formulation in \cite{Beal:2003,Luttinen:2013} and the principal component
analysis formulation in \cite{Bishop:1999}. The likelihood function is
\begin{align}
p(\mathbf{Y}|\mathbf{C},\mathbf{X},\tau) &= \prod^N_{n=1}
\mathcal{N}(\mathbf{y}_n | \mathbf{C}\mathbf{x}_n,
\diag(\boldsymbol\tau)^{-1}),
\end{align}
where $\mathcal{N}(y|m,v)$ is the Gaussian probability density function of $y$
with mean $m$ and covariance $v$, and $\diag(\boldsymbol\tau)$ is a diagonal
matrix with elements $\tau_1, \ldots, \tau_M$ on the diagonal. For simplicity,
we used isotropic noise ($\tau_m=\tau$) in our experiments.
The loading matrix $\mx{C}$ has the following prior, which is also known as an
automatic relevance determination (ARD) prior \cite{Bishop:1999}:
\begin{align}
p(\mx{C}|\mxg{\gamma}) &= \prod^D_{d=1} \mathcal{N} ( \mx{c}_d | \mx{0},
\diag(\mxg{\gamma})^{-1} ),
&
p(\mxg{\gamma}) &= \prod^D_{d=1} \mathcal{G}( \gamma_d | a_\gamma, b_\gamma),
\end{align}
where $\mx{c}_d$ is the $d$-th row vector of $\mx{C}$, the vector $\mxg{\gamma}
= \begin{bmatrix} \gamma_1 & \ldots & \gamma_D \end{bmatrix}^{\operatorname{T}}$ contains the ARD
parameters, and $\mathcal{G}( \gamma | a, b)$ is the gamma probability density
function of $\gamma$ with shape $a$ and rate $b$.
The latent states $\mx{X}=\begin{bmatrix} \mx{x}_0 & \ldots &
\mx{x}_N \end{bmatrix}$ follow a first-order Gaussian Markov process, which
can be written as
\begin{align}
p(\mathbf{X} | \mathbf{W}_n) &= \mathcal{N}(\mx{x}_0|\mxg{\mu}_0^{(x)},
\mx{\Lambda}_0^{-1}) \prod^N_{n=1} \mathcal{N}(\mx{x}_n | \mx{W}_n
\mx{x}_{n-1}, \mx{I}),
\end{align}
where $\mxg{\mu}_0^{(x)}$ and $\mx{\Lambda}_0$ are the mean and precision of the
auxiliary initial state $\mx{x}_0$. The process noise covariance matrix can be
an identity matrix without loss of generality because any rotation can be
compensated in $\mx{x}_n$ and $\mx{W}_n$. The initial state $\mx{x}_0$ can be
given a broad prior by setting, for instance, $\mxg{\mu}_0^{(x)}=\mx{0}$ and
$\mx{\Lambda}_0=10^{-6}\cdot\mx{I}$.
The state dynamics matrices $\mx{W}_n$ are a linear combination of matrices
$\mx{B}_k$ which have the following ARD prior:
\begin{align}
p(\mathbf{B}_k|\boldsymbol{\beta}_k) &= \prod^D_{c=1} \prod^D_{d=1}
\mathcal{N}(b_{kcd} | \mx{0}, \beta_{kd}^{-1}),
&
\!\!\!\! p(\beta_{dk}) &= \mathcal{G} (\beta_{kd}| a_\beta, b_\beta),
&
\!\!\!\! k &= 1, \ldots, K,
\end{align}
where $b_{kcd}=[\mx{B}_k]_{cd}$ is the element on the $c$-th row and $d$-th
column of $\mx{B}_k$. The ARD parameter $\beta_{kd}$ helps in pruning out
irrelevant components in each matrix.
In order to keep the formulas less cluttered, we use the following notation:
$\mx{B}$ is a $K\times{}D\times{}D$ tensor. When using subscripts, the first
index corresponds to the index of the state dynamics matrix, the second index to
the rows of the matrices and the third index to the columns of the matrices. A
colon is used to denote that all elements along that axis are taken. Thus, for
instance, $\mx{B}_{k::}$ is $\mx{B}_k$ and $\mx{B}_{:d:}$ is a $K\times{}D$
matrix obtained by stacking the $d$-th row vectors of $\mx{B}_k$ for each $k$.
The mixing weights $\mx{S}=\begin{bmatrix} \mx{s}_0 & \ldots &
\mx{s}_N \end{bmatrix}$ have first-order Gaussian Markov process prior
\begin{align}
p(\mathbf{S}|\mx{A}) &= \mathcal{N}(\mx{s}_0 | \mxg{\mu}^{(s)}_0,
\mx{V}_0^{-1}) \prod^N_{n=1} \mathcal{N} (\mx{s}_n | \mx{A}\mx{s}_{n-1}, \mx{I}),
\end{align}
where, similarly to the prior of $\mx{X}$, the parameters $\mxg{\mu}^{(s)}_0$
and $\mx{V}_0$ are the mean and precision of the auxiliary initial state
$\mx{s}_0$, and the noise covariance can be an identity matrix without loss of
generality. The initial state $\mx{s}_0$ can be given a broad prior by setting,
for instance, $\mxg{\mu}^{(s)}_0=\mx{0}$ and $\mx{V}_0=10^{-6}\cdot\mx{I}$.
The state dynamics matrix $\mx{A}$ of the latent mixing weights $\mx{s}_n$ is
given an ARD prior
\begin{align}
p(\mx{A}|\mxg{\alpha}) &= \prod^K_{k=1} \mathcal{N} \left( \mx{a}_k | \mx{0},
\diag(\mxg{\alpha})^{-1} \right),
&
p(\mxg{\alpha}) &= \prod^K_{k=1} \mathcal{G}( \alpha_k | a_\alpha, b_\alpha)
\end{align}
where $\mx{a}_k$ is the $k$-th row of $\mx{A}$, and $\mxg{\alpha}=\begin{bmatrix}
\alpha_1 & \ldots & \alpha_K \end{bmatrix}^{\operatorname{T}}$ contains the ARD parameters.
Finally, the noise parameter is given a gamma prior
\begin{align}
p(\boldsymbol{\tau}) &= \prod^M_{m=1} \mathcal{G}(\tau_m | a_\tau, b_\tau).
\end{align}
The hyperparameters of the model can be set, for instance, as $a_\alpha =
b_\alpha = a_\beta = b_\beta = a_\gamma = b_\gamma = a_\tau = b_\tau = 10^{-6}$
to obtain broad priors for the variables. Small values result in approximately
non-informative priors which are usually a good choice in a wide range of
problems.
\begin{figure}[tb]
\usetikzlibrary{shapes}
\usetikzlibrary{fit}
\usetikzlibrary{chains}
\usetikzlibrary{arrows}
\usetikzlibrary{bayesnet}
\centering
\begin{tikzpicture}
\tikzstyle{latent} += [minimum size=20pt];
\node[latent] (x0) {$\mathbf{x}_0$};
\node[latent, below right=of x0] (x1) {$\mathbf{x}_1$};
\node[right=of x1] (dots) {$\cdots$};
\node[latent, right=of dots] (xn) {$\mathbf{x}_N$};
\edge {x0}{x1};
\edge {x1}{dots};
\edge {dots}{xn};
%
\node[obs, below=1 of x1] (y1) {$\mathbf{y}_1$};
\node[right=of y1] (ydots) {$\cdots$};
\node[obs, right=of ydots] (yn) {$\mathbf{y}_N$};
%
\edge {x1} {y1} ;
\edge {dots} {ydots} ;
\edge {xn} {yn} ;
%
%
%
\node[latent, above=1 of x1] (s1) {$\mathbf{s}_1$};
\node[latent, above left=of s1] (s0) {$\mathbf{s}_0$};
\node[right=of s1] (sdots) {$\cdots$};
\node[latent, right=of sdots] (sn) {$\mathbf{s}_N$};
\node[latent, above right=of xn] (B) {$\mathbf{B}$};
\node[latent, right=1 of B] (beta) {$\boldsymbol{\beta}$};
\edge {beta} {B} ;
\edge {B} {x1, dots, xn};
\edge {s0} {s1} ;
\edge {s1} {x1};
\edge {s1} {sdots};
\edge {sdots} {dots};
\edge {sdots} {sn} ;
\edge {sn} {xn};
\node[latent, above right=of sn] (A) {$\mathbf{A}$} ;
\edge {A} {s1,sdots,sn};
\node[latent, right=1 of A] (alpha) {$\boldsymbol{\alpha}$};
\edge {alpha} {A} ;
\node[latent, above right=1 of yn] (C) {$\mathbf{C}$} ;
\node[latent, right=1 of C] (gamma) {$\boldsymbol{\gamma}$};
\edge {C} {y1,ydots,yn};
\edge {gamma} {C} ;
\node[latent, above left=1 of y1] (tau) {$\boldsymbol{\tau}$};
\edge {tau} {y1,ydots,yn} ;
%
\end{tikzpicture}
\caption{The graphical model of the linear state-space model with time-varying
dynamics}
\end{figure}
For the experimental section, we constructed the LSSM with switching dynamics by
using a hidden Markov model (HMM) for the state dynamics matrix $\mx{W}_n$. The
HMM had an unknown initial state and a state transition matrix with broad
conjugate priors. We used similar prior probability distributions in the
classical LSSM with constant dynamics, the proposed LSSM with time-varying
dynamics, and the LSSM with switching dynamics for the similar parts of the
models.
\section{Variational Bayesian Inference}
\label{sec:inference}
As the posterior distribution is analytically intractable, it is approximated
using variational Bayesian (VB) framework, which scales well to large
applications compared to Markov chain Monte Carlo (MCMC) methods
\cite{Bishop:2006}. The posterior approximation is assumed to factorize with
respect to the variables:
\begin{align}
p(\mx{X},\!\mx{C},\!\mxg{\gamma},\mx{B},\mxg{\beta},\mx{S},\mx{A},\mxg{\alpha},\mxg{\tau}|\mx{Y})
\!\approx\!
%
q(\mx{X})q(\mx{C})q(\mxg{\gamma})q(\mx{B})q(\mxg{\beta})q(\mx{S})
q(\mx{A})q(\mxg{\alpha})q(\mxg{\tau}).
\end{align}
This approximation is optimized by minimizing the Kullback-Leibler divergence
from the true posterior
by using the variational Bayesian expectation-maximization (VB-EM) algorithm
\cite{Beal:2003b}. In VB-EM, the posterior approximation is updated for the
variables one at a time and iterated until convergence.
\subsection{Update Equations}
The approximate posterior distributions have the following forms:
\begin{align}
q(\mx{X}) &= \mathcal{N}([\mx{X}]_:|\mxg{\mu}_{x}, \mx{\Sigma}_x),
&
q(\mxg{\tau}) &= \prod^M_{m=1} \mathcal{G} (\tau_m | \bar{a}_\tau^{(m)},
\bar{b}_\tau^{(m)}),
%
\\
q(\mx{C}) &= \prod^M_{m=1} \mathcal{N}( \mx{c}_{m} | \mxg{\mu}_c^{(m)}, \mx{\Sigma}_c^{(m)}),
&
q(\mxg{\gamma}) &= \prod^D_{d=1} \mathcal{G} (\gamma_d| \bar{a}_\gamma^{(d)},
\bar{b}_\gamma^{(d)}),
%
\\
q(\mx{B}) &= \prod^D_{d=1} \mathcal{N} ( [\mx{B}_{:d:}]_: | \mxg{\mu}_b^{(d)},
\mx{\Sigma}_b^{(d)}),
&
q(\mxg{\beta}) &= \prod^K_{k=1} \prod^D_{d=1} \mathcal{G} ( \beta_{kd} |
\bar{a}_\beta^{(kd)}, \bar{b}_\beta^{(kd)}),
\\
q(\mx{S}) &= \mathcal{N}( [\mx{S}]_: | \mxg{\mu}_s, \mx{\Sigma}_s ),
\\
q(\mx{A}) &= \prod^K_{k=1} \mathcal{N}(\mx{a}_{k} | \mxg{\mu}_a^{(k)}, \mx{\Sigma}_a^{(k)}),
&
q(\mxg{\alpha}) &= \prod^K_{k=1} \mathcal{G} (\alpha_k| \bar{a}_\alpha^{(k)},
\bar{b}_\alpha^{(k)}),
%
\end{align}
where $[\mx{X}]_:$ is a vector obtained by stacking the column vectors
$\mx{x}_n$. It is straightforward to derive the following update equations of
the variational parameters:
\begin{align}
\bar{a}_\tau^{(m)} &= a_\tau + \frac{N_m}{2},
%
&
%
\bar{b}_\tau^{(m)} &= b_\tau + \frac{1}{2} \sum_{n\in\mathcal{O}_{m:}} \xi_{mn},
%
%
\\
%
%
\mx{\Sigma}_c^{(m)} &= \left( \! \langle\diag(\mxg{\gamma})\rangle +
\!\!\! \sum_{n\in\mathcal{O}_{m:}} \!\!\langle\tau_m\rangle \langle\mx{x}_n\mx{x}_n^{\operatorname{T}}\rangle
\!\right)^{-1} \!\!\!\!\!\!\!,
%
&
%
\mxg{\mu}_c^{(m)} &= \mx{\Sigma}_c^{(m)} \!\!\!\sum_{n\in\mathcal{O}_{m:}} \!\!\! y_{mn}
\langle{\tau_m}\rangle \langle\mx{x}_n\rangle,
%
%
\\
%
%
\bar{a}_\gamma^{(d)} &= a_\gamma + \frac{M}{2},
%
&
%
\bar{b}_\gamma^{(d)} &= b_\gamma + \frac{1}{2} \sum^M_{m=1} \langle
c_{md}^2\rangle,
%
%
\end{align}
\begin{align}
%
\mx{\Sigma}_b^{(d)} &= \left( \langle\diag(\mxg{\beta})\rangle +
\sum^{N}_{n=1} \mx{\Omega}_n \right)^{-1} \!\!\!\!\!\!\!,
%
&
%
\mxg{\mu}_b^{(d)} &= \mx{\Sigma}_b^{(d)} \!\sum^N_{n=1} \!\left[ \langle
\mx{s}_{n}\rangle \langle x_{dn}\mx{x}_{n-1}^{\operatorname{T}} \rangle \right]_: ,
%
%
\\
%
%
\mx{\Sigma}_a^{(k)} &= \left( \! \langle\diag(\mxg{\alpha})\rangle +
\!\sum^{N}_{n=1} \langle\mx{s}_{n-1}\mx{s}_{n-1}^{\operatorname{T}}\rangle \!\right)^{-1}
\!\!\!\!\!\!\!,
%
&
%
\mxg{\mu}_a^{(k)} &= \mx{\Sigma}_a^{(k)} \sum^N_{n=1} \langle
s_{kn}\mx{s}_{n-1}\rangle,
%
%
\\
%
%
\bar{a}_\alpha^{(k)} &= a_\alpha + \frac{K}{2},
%
&
%
\bar{b}_\alpha^{(k)} &= b_\alpha + \frac{1}{2} \sum^K_{i=1} \langle a_{ik}^2
\rangle,
\end{align}
where $\mathcal{O}_{m:}$ is the set of time instances $n$ for which the observation
$y_{mn}$ is not missing, $N_m$ is the size of the set $\mathcal{O}_{m:}$, $\xi_{mn} =
\left\langle (y_{mn} - \mx{c}_m^{\operatorname{T}}\mx{x}_n)^2 \right\rangle$, $\mx{\Omega}_n =
\langle\mx{x}_{n-1}\mx{x}_{n-1}^{\operatorname{T}}\rangle \otimes
\langle\mx{s}_n\mx{s}_n^{\operatorname{T}}\rangle$, and $\otimes$ denotes the Kronecker product.
The computation of the posterior distribution of $\mx{X}$ and $\mx{S}$ is more
complicated and will be discussed next.
The time-series variables $\mx{X}$ and $\mx{S}$ can be updated using algorithms
similar to the Kalman filter and the Rauch-Tung-Striebel smoother. The
classical formulations of those algorithms do not work for VB learning because
of the uncertainty in the dynamics matrix \cite{Beal:2003,Barber:2007}. Thus,
we used a modified version of these algorithms as presented for the classical
LSSM in \cite{Luttinen:2013}. The algorithm performs a forward and a backward
pass in order to find the required posterior expectations.
The explicit update equations for $q(\mx{X})$ can be written as:
\begin{align}
\mx{\Sigma}_x^{-1} &=
\arraycolsep=6.0pt
\def1.5{1.5}
\begin{bmatrix}
\mx{\Lambda}_0 + \langle \mx{W}_1^{\operatorname{T}}\mx{W}_1 \rangle &
%
-\langle \mx{W}_1 \rangle^{\operatorname{T}} &
%
&
%
\\
-\langle \mx{W}_1 \rangle &
%
\mx{I} + \langle \mx{W}_2^{\operatorname{T}}\mx{W}_2 \rangle + \mx{\Psi}_1 &
%
\ddots
%
&
%
\\
%
&
%
\ddots &
%
\ddots &
%
-\langle \mx{W}_N \rangle^{\operatorname{T}}
%
\\
%
&
%
&
%
-\langle \mx{W}_N \rangle &
%
\mx{I} + \mx{\Psi}_N
\end{bmatrix},
\end{align}
\begin{align}
\mxg{\mu}_x &= \mx{\Sigma}_x
\def1.5{1.5}
\begin{bmatrix}
\mx{\Lambda}_0 \mxg{\mu}_0^{(x)}
\\
\sum_{m \in \mathcal{O}_{:1}} y_{m1} \langle\tau_m\rangle
\langle\mx{c}_m\rangle
\\
\vdots
\\
\sum_{m \in \mathcal{O}_{:N}} y_{mN} \langle\tau_m\rangle
\langle\mx{c}_m\rangle
\end{bmatrix},
\end{align}
where $\mathcal{O}_{:n}$ is the set of indices $m$ for which the observation
$y_{mn}$ is not missing, $\mx{\Psi}_n = \sum_{m \in \mathcal{O}_{:n}}
\langle\tau_m\rangle \langle\mx{c}_m\mx{c}_m^{\operatorname{T}}\rangle$,
$\langle\mx{W}_n\rangle=\sum_{k=1}^K\langle s_{kn} \rangle
\langle\mx{B}_k\rangle$, and $\langle\mx{W}_n^{\operatorname{T}}\mx{W}_n\rangle =
\sum_{k=1}^K\sum_{l=1}^K [\langle \mx{s}_{n}\mx{s}_{n}^{\operatorname{T}} \rangle]_{kl}
\langle\mx{B}_k^{\operatorname{T}}\mx{B}_l\rangle$. Instead of using standard matrix inversion,
one can utilize the block-banded structure of $\mx{\Sigma}_x^{-1}$ to compute the
required expectations $\langle\mx{x}_n\rangle$,
$\langle\mx{x}_n\mx{x}_n^{\operatorname{T}}\rangle$ and $\langle\mx{x}_n\mx{x}_{n-1}^{\operatorname{T}}\rangle$
efficiently. The algorithm for the computations is presented in
\cite{Luttinen:2013}.
Similarly for $\mx{S}$, the explicit update equations are
\begin{align}
\mx{\Sigma}_s^{-1} &=
\arraycolsep=6.0pt
\def1.5{1.5}
\begin{bmatrix}
\mx{V}_0 + \langle \mx{A}^{\operatorname{T}}\mx{A} \rangle &
%
-\langle \mx{A} \rangle^{\operatorname{T}} &
%
&
%
\\
-\langle \mx{A} \rangle &
%
\mx{I} + \langle \mx{A}^{\operatorname{T}}\mx{A} \rangle + \mx{\Theta}_1 &
%
\ddots
%
&
%
\\
%
&
%
\ddots &
%
\ddots &
%
-\langle \mx{A} \rangle^{\operatorname{T}}
%
\\
%
&
%
&
%
-\langle \mx{A} \rangle &
%
\mx{I} + \mx{\Theta}_N
\end{bmatrix},
\end{align}
\begin{align}
\mxg{\mu}_s &= \mx{\Sigma}_s
\def1.5{1.5}
\begin{bmatrix}
\mx{V}_0 \mxg{\mu}_0^{(s)}
\\
\sum_{d=1}^D \langle\mx{B}_{:d:}\rangle \langle x_{d1} \mx{x}_0^{\operatorname{T}} \rangle
\\
\vdots
\\
\sum_{d=1}^D \langle\mx{B}_{:d:}\rangle \langle x_{dN} \mx{x}_{N-1}^{\operatorname{T}} \rangle
\end{bmatrix},
\end{align}
where $\mx{\Theta}_n = \sum_{i=1}^D\sum_{j=1}^D
[\langle\mx{x}_{n-1}\mx{x}_{n-1}^{\operatorname{T}}\rangle]_{ij} \cdot
\langle\mx{B}_{::i}\mx{B}_{::j}^{\operatorname{T}}\rangle$. The required expectations
$\langle\mx{s}_n\rangle$, $\langle\mx{s}_n\mx{s}_n\rangle$ and
$\langle\mx{s}_n\mx{s}_{n-1}\rangle$ can be computed efficiently by using the
same algorithm as for $\mx{X}$ \cite{Luttinen:2013}.
The VB learning of the LSSM with switching dynamics is quite similar to the
equations presented above. The main difference is that the posterior
distribution of the discrete state variable $z_n$ is computed by using
alpha-beta recursion \cite{Bishop:2006}. The update equations for the state
transition probability matrix and the initial state probabilities are
straightforward because of the conjugacy. The expectations
$\langle\mx{W}_n\rangle$ and $\langle\mx{W}_n^{\operatorname{T}}\mx{W}_n\rangle$ are computed by
averaging $\langle\mx{B}_k\rangle$ and $\langle\mx{B}_k^{\operatorname{T}}\mx{B}_k\rangle$ over
the state probabilities $\mathrm{E}[z_n=k]$.
\subsection{Practical Issues}
The main practical issue with the proposed model is that the VB learning
algorithm may converge to bad local minima. As a solution, we found two ways of
improving the robustness of the method. The first improvement is related to the
updating of the posterior approximation and the second improvement is related to
the initialization of the approximate posterior distributions.
The first practical tip is that one may want to run the VB updates for the lower
layers of the model hierarchy first for a few times before starting to update
the upper layers. Otherwise, the hyperparameters may learn very bad values
because the child variables have not yet been well estimated. Thus, we updated
$\mx{X}$, $\mx{C}$, $\mx{B}$ and $\mxg{\tau}$ 5--10 times before updating the
hyperparameters and the upper layers. However, this procedure requires a
reasonable initialization.
We initialized $\mx{X}$ and $\mx{C}$ randomly but for $\mx{S}$ and $\mx{B}$ we
used a bit more complicated approach. One goal of the initialization was that
the model would be close to a model with constant dynamics. Thus, the first
component in $\mx{S}$ was set to a constant value and the corresponding matrix
$\mx{B}_k$ was initialized as an identity matrix. The other components in
$\mx{S}$ and $\mx{B}$ were random but their scale was a bit smaller so that the
time variation in the resulting state dynamics matrix $\mx{W}_n$ was small
initially. Obviously, this initialization leads to a bias towards a constant
component in $\mx{S}$ but this is often realistic as the system probably has
some average dynamics and deviations from it.
\subsection{Rotations for Faster Convergence}
One issue with the VB-EM algorithm for state-space models is that the algorithm
may converge extremely slowly. This happens if the variables are strongly
correlated because they are updated only one at a time causing zigzagging and
small updates. This effect can be reduced by the parameter expansion approach,
which finds a suitable auxiliary parameter connecting several variables and then
optimizes this auxiliary parameter \cite{Liu:1998,Qi:2007}. This corresponds to
a parameterized joint optimization of several variables.
A suitable parameter expansion for state-space models is related to the rotation
of the latent sub-space \cite{Luttinen:2010,Luttinen:2013}. It can be motivated
by noting that the latent variable $\mx{X}$ can be rotated arbitrarily by
compensating it in $\mx{C}$:
\begin{align}
\mx{y}_n &= \mx{C}\mx{x}_n = \mx{C}\mx{R}^{-1}\mx{R}\mx{x}_n =
\big(\mx{C}\mx{R}^{-1}\big) \big(\mx{R}\mx{x}_n\big) \quad \text{ for all
non-singular } \mx{R} \,.
\end{align}
The rotation of $\mx{X}$ must also be compensated in the dynamics $\mx{W}_n$ as
\begin{align}
\mx{R}\mx{x}_n &= \mx{R}\mx{W}_n\mx{R}^{-1} \mx{R}\mx{x}_{n-1} =
\big(\mx{R}\mx{W}_n\mx{R}^{-1}\big) \big(\mx{R}\mx{x}_{n-1}\big) \,.
\end{align}
The rotation $\mx{R}$ can be used to parameterize the posterior distributions
$q(\mx{X})$, $q(\mx{C})$ and $q(\mx{B})$. Optionally, the distributions of the
hyperparameters $q(\mxg{\gamma})$ and $q(\mxg{\beta})$ can also be
parameterized. Optimizing the posterior approximation with respect to $\mx{R}$
is efficient and leads to significant improvement in the speed of the VB
learning. Details for the procedure in the context of the classical LSSM can be
found in \cite{Luttinen:2013}.
Similarly to $\mx{X}$, the latent mixing weights $\mx{S}$ can also be rotated as
\begin{align}
[\mx{W}_n]_{d:} &= \mx{B}_{:d:}^{\operatorname{T}} \mx{s}_n = \mx{B}_{:d:}^{\operatorname{T}}\mx{R}^{-1}\mx{R}\mx{s}_n
= \Big(\mx{B}_{:d:}^{\operatorname{T}}\mx{R}^{-1}\Big) \Big(\mx{R}\mx{s}_n\Big) \,,
\end{align}
where $[\mx{W}_n]_{d:}$ is the $d$-th row vector of $\mx{W}_n$. The rotation
must also be compensated in the dynamics of $\mx{S}$ as
\begin{align}
\mx{R}\mx{s}_n &= \mx{R}\mx{A}\mx{R}^{-1} \mx{R}\mx{s}_{n-1} =
\big(\mx{R}\mx{A}\mx{R}^{-1}\big) \big(\mx{R}\mx{s}_{n-1}\big) \,.
\end{align}
Thus, the rotation corresponds to a parameterized joint optimization of
$q(\mx{S})$, $q(\mx{A})$, $q(\mx{B})$, and optionally also $q(\mxg{\alpha})$ and
$q(\mxg{\beta})$. Note that the optimal rotation of $\mx{S}$ can be computed
separately from the optimal rotation for $\mx{X}$.
The extra computational cost by the rotation speed up is small compared to the
computational cost of one VB update of all variables. Thus, the rotation can be
computed at each iteration after the variables have been updated. If for some
reason the computation of the optimal rotation is slow, one can use the
rotations less frequently, for instance, after every ten updates of all
variables, and still gain similar performance improvements. However, as was
shown in \cite{Luttinen:2013}, the rotation transformation is essential even for
the classical LSSM, thus ignoring it may lead to extremely slow convergence and
poor results. Thus, we used the rotation transformation for all methods in the
next section.
\section{Experiments}
\label{sec:experiments}
We compare the proposed linear state-space model with time-varying dynamics
(LSSM-TVD) to the classical linear-state space model (LSSM) and the linear
state-space model with switching dynamics (LSSM-SD) using three datasets: a
one-dimensional signal with changing frequency, a simulated physical process
with time-varying parameters, and real-world daily temperature measurements in
Europe. The methods are evaluated by their ability to predict missing values
and gaps in the observed processes.
\subsection{Signal with Changing Frequency}
We demonstrate the LSSM with time-varying dynamics using an artificial signal
with changing frequency. The signal is defined as
\begin{align}
f(n) = \sin ( a \cdot (n + c \sin(b \cdot 2\pi) ) \cdot 2\pi ), \quad
n=0,\ldots,999
\end{align}
where $a=0.1$, $b=0.01$ and $c=8$. The resulting signal is shown in
Fig.~\ref{fig:artificial}(a). The signal was corrupted with Gaussian noise
having zero mean and standard deviation $0.1$ to simulate noisy observations.
In order to see how well the different methods can learn the dynamics, we
created seven gaps in the signal by removing 15 consecutive observations to
produce each gap. In addition, 20\% of the remaining observations were randomly
removed. Each method (LSSM, LSSM-SD and LSSM-TVD) used $D=5$ dimensions for the
latent states $\mx{x}_n$. The LSSM-SD and LSSM-TVD used $K=4$ state dynamics
matrices $\mx{B}_k$.
\begin{figure}[tb]
\small
\begin{center}
\begin{tabular}{c}
\includegraphics[width=\linewidth]{fig_toy_data}
\\
(a) True signal
\\
\includegraphics[width=\linewidth]{fig_toy_lssm}
\\
(b) LSSM
\\
\includegraphics[width=\linewidth]{fig_toy_slssm}
\\
(c) LSSM-SD
\\
\includegraphics[width=\linewidth]{fig_toy_dlssm}
\\
(d) LSSM-TVD
\end{tabular}
\end{center}
\caption{Results for the signal with changing frequency: (a) the true signal,
(b) the classical LSSM, (c) the LSSM with switching dynamics, (d) the LSSM
with time-varying dynamics. In (b)-(d), the posterior mean is shown as
solid black line, two standard deviations are shown as a gray area, and the
true signal is shown in red for comparison. Vertical lines mark the seven
gaps that contain no observations.}
\label{fig:artificial}
\end{figure}
Figures~\ref{fig:artificial}(b)-(d) show the posterior distribution of the
latent noiseless function $f$ for each method. The classical LSSM is not able to
capture the dynamics and the reconstructions over the gaps are bad and have high
variance. The LSSM-SD learns two different states for the dynamics corresponding
to a lower and a higher frequency. The reconstructions over the gaps are better
than with the LSSM, but it still has quite a large variance and the fifth gap is
reconstructed using a wrong frequency. The gap reconstructions have large
variance because the two state dynamics matrices learned by the model do not fit
the process very well so the model assumes a larger innovation noise in the
latent process $\mx{X}$. In contrast to that, the LSSM-TVD learns the dynamics
practically perfectly and even learns the dynamics of the process which changes
the frequency. Thus, the LSSM-TVD is able to make nearly perfect predictions
over the gaps and the variance is small. It also prunes out one state dynamics
matrix, thus using effectively only three dimensions for the latent
mixing-weight process.
\subsection{Stochastic Advection-Diffusion Process}
We simulated a physical process with time-dependent parameters in order to
compare the considered approaches. The physical process is a stochastic
advection-diffusion process, which is defined by the following partial
differential equation:
\begin{align}
\frac{ \partial f}{ \partial t } &= \delta \nabla^2 f - \mathbf{v} \cdot
\nabla f + R ,
\label{eq:spde}
\end{align}
where $f$ is the variable of interest, $\delta$ is the diffusivity, $\mathbf{v}$
is the velocity field and $R$ is a stochastic source. We have assumed that the
diffusivity is a constant and the velocity field describes an incompressible
flow. The velocity field $\mathbf{v}$ changes in time. This equation could
describe, for instance, air temperature and the velocity field corresponds to
winds with changing directions. The spatial domain was a torus, a
two-dimensional manifold with periodic boundary conditions.
The partial differential equation \eqref{eq:spde} is discretized using the finite
difference method. This is used to generate a discretized realization of the
stochastic process by iterating over the time domain. The stochastic source $R$
is a realization from a spatial Gaussian process at each time step. The two
velocity field components are modelled as follows:
\begin{align}
\mx{v}(t+1) &= \sqrt{\rho}\cdot \mx{v}(t) + \sqrt{1-\rho} \cdot
\boldsymbol{\xi}(t+1),
\end{align}
where $\rho\in (0,1)$ controls how fast the velocity changes and
$\boldsymbol{\xi}(t+1)$ is Gaussian noise with zero mean and variance which was
chosen appropriately. Thus, there are actually two sources of randomness in the
stochastic process: the random source $R$ and the randomly changing velocity
field $\mathbf{v}$.
The data were generated from the simulated process as follows: Every 20-th
sample was kept in the time domain, which resulted in $N=2000$ time instances.
From the spatial discretization grid, $M=100$ locations were selected randomly
as the measurement locations (corresponding to weather stations). The simulated
values were corrupted with Gaussian noise to obtain noisy observations.
We used four methods in this comparison: LSSM, LSSM-SD and LSSM-TVD with $D=30$
dimensions for the latent states $\mx{x}_n$, and LSSM with $D=60$ to see if
adding more dimensions improves the performance of the classical LSSM. Both the
LSSM-SD and LSSM-TVD used $K=5$ state dynamics matrices $\mx{B}_k$.
For measuring the performance of the methods, we generated two test sets.
First, we created 18 gaps of 15 consecutive time points, that is, the
observations from all the spatial locations were removed over the gaps and the
corresponding values of the noiseless process $f$ formed the first test set.
Second, we randomly removed 20\% of the remaining observations and used the
corresponding values of the process $f$ as the second test set. The tests were
performed for five simulated processes. Figure~\ref{fig:spde} shows one process
at one time instance as an
example.\footnote{\url{http://users.ics.aalto.fi/jluttine/ecml2014/} contains a
video visualization of each of the simulated processes.}
\begin{figure}[tb]
\centering
\includegraphics[width=0.35\linewidth]{fig_spde_snapshot.pdf}
\caption{One of the simulated processes at one time instance. Crosses denote the
locations that were used to collect the observations. Note that the domain
is a torus, that is, a 2-dimensional manifold with periodic boundaries.}
\label{fig:spde}
\end{figure}
Table~\ref{tab:spde} shows the root-mean-square errors (RMSE) of the mean
reconstructions for both the generated gaps and the randomly removed values.
The results for each of the five process realizations are shown separately. It
appears that using $D=60$ components does not significantly change the
performance of the LSSM compared to using $D=30$ components. Also, the LSSM-SD
performs practically identically to the LSSM. The LSSM-SD used effectively two
or three state dynamics matrices. However, this does not seem to help in
modelling the variations in the dynamics and the learned model performs
similarly to the LSSM. In contrast to that, the proposed LSSM-TVD has the best
performance in each experiment. For the test set of random values, the
difference is not large because the reconstruction is mainly based on the
correlations between the locations rather than the dynamics. However, in order
to accurately reconstruct the gaps, the model needs to learn the changes in the
dynamics. The LSSM-TVD reconstructs the gaps more accurately than the other
methods, because it finds latent mixing weights $\mx{s}_n$ which model the
changes in the dynamics.
\begin{table}[b]
\centering
\caption{Results for five stochastic advection-diffusion experiments. The
root-mean-square errors (RMSE) have been multiplied by a factor of 1000 for
clarity.}
\small
%
\setlength{\tabcolsep}{6.5pt}
\begin{tabular}{c|ccccc|ccccc}
&
\multicolumn{5}{c|}{RMSE for gaps} &
\multicolumn{5}{c}{RMSE for random}
\\
Method & 1 & 2 & 3 & 4 & 5 & 1 & 2 & 3 & 4 & 5
\\
\hline
LSSM $D=30$ &
104 & 107 & 102 & 94 & 104 &
34 & 38 & 39 & 34 & 34
\\
LSSM $D=60$ &
105 & 107 & 110 & 98 & 108 &
35 & 39 & 40 & 35 & 35
\\
LSSM-SD $D=30$ &
106 & 117 & 113 & 94 & 102 &
35 & 37 & 39 & 34 & 34
\\
LSSM-TVD $D=30$ &
\bf 73 & \bf 81 & \bf 75 & \bf 67 & \bf 82 &
\bf 30 & \bf 34 & \bf 35 & \bf 31 & \bf 30
\end{tabular}
\label{tab:spde}
\end{table}
Figure~\ref{fig:spde_S}(a) shows the the posterior distribution of the $K=5$
mixing-weight signals $\mx{S}$ in one experiment: the first signal is
practically constant corresponding to the average dynamics, the third and the
fourth signals correspond to the changes in the two-dimensional velocity field,
and the second and the fifth signals have been pruned out as they are not
needed. Thus, the method was able to learn the effective dimensionality of the
latent mixing-weight process, which suggests that the method is not very
sensitive to the choice of $K$ as long as it is large enough. The results look
similar in all the experiments with the LSSM-TVD and in every experiment the
LSSM-TVD found one constant and two varying components. Thus, the posterior
distribution of $\mx{S}$ might give insight on some latent processes that affect
the dynamics of the observed process.
\begin{figure}[tb]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.49\linewidth]{fig_spde_s.pdf}
&
\includegraphics[width=0.49\linewidth]{fig_spde_z.pdf}
\\
(a) $q(\mathbf{S})$ in LSSM-TVD
&
(b) $q(\mathbf{Z})$ in LSSM-SD
\end{tabular}
\caption{Results for the advection-diffusion experiments. (a) The posterior
mean and two standard deviations of the latent mixing weights by the
LSSM-TVD. (b) The posterior probability of each state transition matrix as
a function of time in the LSSM-SD.}
\label{fig:spde_S}
\end{figure}
This experiment showed that the LSSM-SD is not good at modelling linear
combinations of the state dynamics matrices. Interestingly, although the VB
update formulas average the state dynamics matrices by their probabilities
resulting in a convex combination,
this mixing is not very prominent in the approximate posterior distribution as
seen in Fig.~\ref{fig:spde_S}(b).
Most of the time, only one state dynamics matrix is active with probability one.
This happens because the prior penalizes switching between the matrices and one
of the matrices is usually much better than the others on average over several
time steps.
\subsection{Daily Mean Temperature}
The third experiment used real-world temperature measurements in Europe. The
data were taken from the global surface summary of day product produced by the
National Climatic Data Center (NCDC) \cite{NCDC}. We studied daily mean
temperature measurements roughly from the European area\footnote{The longitude
of the studied region was in range $(-13, 33)$ and the latitude in range $(35,
72)$.} in 2000--2009. Stations that had more than 20\% of the measurements
missing were discarded. This resulted in $N=3653$ time instances and $M=1669$
stations for the analysis.
The three models were learned from the data. They used $D=80$ dimensions for
the latent states. The LSSM-SD and the LSSM-TVD used $K=6$ state dynamics
matrices. We formed two test sets similarly to the previous experiment. First,
we generated randomly 300 2-day gaps in the data, which means that measurements
from all the stations were removed during those periods of time. Second, 20\%
of the remaining data was used randomly to form another test set.
Table~\ref{tab:gsod} shows the results for five experiments using different test
sets. The methods reconstructed the randomly formed test sets equally well
suggesting that learning more complex dynamics did not help and the learned
correlations between the stations was sufficient. However, the reconstruction
of gaps is more interesting because it measures how well the method learns the
dynamical structure. This reconstruction shows consistent performance
differences between the methods: The LSSM-SD is slightly worse than the LSSM,
and the LSSM-TVD outperforms the other two. Because climate is a chaotic
process, the modelling is extremely challenging and predictions tend to be far
from perfect. However, these results suggest that the time-varying dynamics
might offer a promising improvement to the classical LSSM in statistical
modelling of physical processes.
\begin{table}[tb]
\centering
\caption{GSOD reconstruction errors of the test sets in degrees Celsius for five
runs }
\label{tab:gsod}
\small
\setlength{\tabcolsep}{3.1pt}
\begin{tabular}{c|ccccc|ccccc}
&
\multicolumn{5}{c|}{RMSE for gaps} &
\multicolumn{5}{c}{RMSE for randomly missing}
\\
Method & 1 & 2 & 3 & 4 & 5 & 1 & 2 & 3 & 4 & 5
\\
\hline
LSSM &
1.748 & 1.753 & 1.758 & 1.744 & 1.751 &
0.935 & 0.937 & 0.935 & 0.933 & 0.934
\\
LSSM-SD &
1.800 & 1.801 & 1.796 & 1.777 & 1.788 &
0.936 & 0.938 & 0.936 & 0.934 & 0.935
\\
LSSM-TVD &
\bf 1.661 & \bf 1.650 & \bf 1.659 & \bf 1.653 & \bf 1.660 &
0.935 & 0.937 & 0.935 & 0.932 & 0.934
\end{tabular}
\end{table}
\section{Conclusions}
\label{sec:conclusions}
This paper introduced a linear state-space model with time-varying dynamics. It
forms the state dynamics matrix as a time-varying linear combination of a set of
matrices. It uses another linear state-space model for the mixing weights in
the linear combination. This is different from previous time-dependent LSSMs
which use switching models to jump between a small set of states defining the
model dynamics.
Both the LSSM with switching dynamics and the proposed LSSM are useful but they
are suitable for slightly different problems. The switching dynamics is
realistic for processes which have a few possible states that can be quite
different from each other but each of them has approximately linear dynamics.
The proposed model, on the other hand, is realistic when the dynamics vary more
freely and continuously. It was largely motivated by physical processes based
on stochastic partial differential equations with time-varying parameters.
The experiments showed that the proposed LSSM with time-varying dynamics can
capture changes in the underlying dynamics of complex processes and
significantly improve over the classical LSSM. If these changes are continuous
rather than discrete jumps between a few states, it may achieve better modelling
performance than the LSSM with switching dynamics. The experiment on a
stochastic advection-diffusion process showed how the proposed model adapts to
the current dynamics at each time and finds the current velocity field which
defines the dynamics.
The proposed model could be further improved for challenging real-world
spatio-temporal modelling problems. First, the spatial structure could be taken
into account in the prior of the loading matrix using, for instance, Gaussian
processes \cite{Luttinen:2009:nips}. Second, outliers and badly corrupted
measurements could be modelled by replacing the Gaussian observation noise
distribution with a more heavy-tailed distribution, such as the Student-$t$
distribution \cite{Luttinen:2012:npl}.
The method was implemented in Python as a module for an open-source variational
Bayesian package called BayesPy \cite{BayesPy}. It is distributed under an open
license, thus making it easy for others to apply the method. In addition, the
scripts for reproducing all the experimental results are also
available.\footnote{Experiment scripts available at
\url{http://users.ics.aalto.fi/jluttine/ecml2014/}}
\subsubsection*{Acknowledgments.}
We would like to thank Harri Valpola, Natalia Korsakova, and Erkki Oja for
useful discussions. This work has been supported by the Academy of Finland
(project number 134935).
\bibliographystyle{splncs}
|
\section{Introduction}
The distribution of entanglement between macroscopically separated
parties constitutes a key ingredient of quantum information
networks~\cite{Nielsen,Kimble}. A quantum network is composed of
nodes, for processing and storing quantum states, and channels linking
the nodes. The implementation of quantum nodes is a major
challenge: different approaches are currently being pursued, most of them
involving single emitters, such as ions, atoms or nitrogen-vacancy centers~\cite{Olmschenk2,hofmann2012,Ritter,bernien2013}, even
though they are inherently probabilistic.
Photonic channels are especially advantageous, as optical photons can
carry information over long distances with almost negligible
decoherence.
In practice, there are two types of these channels:
optical fibers and free space.
Optical fibers are capable of transmitting single photons over large distances with high efficiency while suffering from effects like birefringence or dispersion.
The free space channel, however, does not suffer from these effects, but photon losses
due to beam wandering or beam broadening, for example, can play a prominent role. Thus both types of photonic channels have their own pros and cons~\cite{Gisin} and
distribution of entangled photonic qubits was
successfully demonstrated for both of them, over a distance of 200~km~\cite{Dynes} using
optical fibers and over 144~km~\cite{Ursin07} in free space.
The main issue with a free-space channel is the low
photon-collection efficiency. This can be improved by placing the
single emitter at the focus of a parabolic mirror~\cite{Maiwald},
which in addition enhances the atom-field interaction~\cite{leuchs2013o,
fischer2014}.
Here, we propose to use two opposing parabolic mirrors to
prepare maximally entangled states of two matter qubits at the
corresponding focal points. Our scheme involves an extreme multimode
scenario ,i.e., the atoms couple to a continuum of modes of the radiations field, due to the fact that the parabolic mirror is a half-open cavity. Thereby, we deal with intrinsic multimode effects like
spontaneous decay processes, which are usually considered as sources of
undesirable decoherence. Interestingly enough, we will be able to use
these effects as tools for entanglement generation, rather than
avoiding them.
In other multimode schemes~\cite{Moehring} each
deviation from the ideal situation, such as non perfect mode matching,
leads to a reduction of the fidelity of the generated state.
In contradistinction, our scheme is robust against the
vulnerabilities that arise in experimental implementations: they
reduce the success probability, but leave the fidelity unaltered (and,
accordingly, it can be very high). As outlined below, this is due to the use of photons originating from
circular-dipole transitions, a suitable choice of the quantization
axis and direct dispersive probing of the qubit
states.
This paper is organized as follows. In Sec.~\ref{setup} we
advance the basic ingredients of our scheme, which is fully
analyzed in Sec.~\ref{model} by resorting to a photon-path
representation~\cite{Alber2013,Milonni} especially germane for a
multimode description. To incorporate the boundary conditions for the
relevant solution of the Helmholtz equation, we apply a semiclassical
approximation~\cite{Berry1972,Maslov}. We discuss the results in
Sec.~\ref{entanglement} and their feasibility in
Sec. ~\ref{experiment}. Finally, our conclusions are briefly
summarized in Sec.~\ref{V. Conclusion}.
\section{Remote entanglement preparation}
\label{setup}
Our setup, as roughly schematized in Fig.~\ref{fig.post}, consists
of two parabolic mirrors opposing each other, so they direct any
electromagnetic field from one focal point to the other with great
efficiency.
\begin{figure}
\centerline{\includegraphics[width=.90\columnwidth]{fig1.eps}}
\caption{(Color online) Scheme of the setup, including the post
selection procedure: Two $^{171}$Yb$^+$ ions are trapped at the
foci of two parabolic mirrors. Entanglement between the two ions
is mediated by a circularly polarized photon ($\sigma$) emitted by
ion\,1 and absorbed by ion\,2. Successful entanglement is probed
by the dispersive interaction of weak linearly polarized coherent
states ($\pi$) with the ions. Only if an ion resides in one of the
desired entangled states, a phase shift is imprinted onto the
coherent state. Probe pulses are coupled into the parabolic
mirrors by means of beam splitters. For simplicity, the coherent
pulses used for dispersive state detection are indicated for only
one of the two ions.}
\label{fig.post}
\end{figure}
We consider a trapped $^{171}\text{Yb}^{+}$ ion at the focus of each
parabolic cavity. This ion has quite a suitable hyperfine electronic
structure due to its nuclear spin $I=1/2$. We concentrate on the
level scheme formed by the levels $6^{2}S_{1/2}$ and
$6^{2}P_{1/2}$ shown in Fig.~\ref{fig2}. The logical qubit is
defined by the levels $|6^{2}S_{1/2},F=1,m=-1\rangle$ and
$|6^{2}S_{1/2},F=1,m=1\rangle$ (note the different choice
in Ref.~\cite{Olmschenk1,Olmschenk2}). The corresponding
dipole matrix elements are denoted by $ \mathbf{d}_{ij} = \langle j |
\hat{\mathbf{d}} | i \rangle$, where $|i\rangle$ and $|j\rangle$ are
the wave functions of the different states.
The basic idea is to initially prepare ions 1 and 2 in the states
\begin{eqnarray}
|\psi^{(1)} (0) \rangle =
|6^{2}P_{1/2},F=1,m=0\rangle \, , \nonumber \\
\\
|\psi^{(2)} (0) \rangle =
|6^{2}S_{1/2},\; F=1,m=0\rangle\, , \nonumber
\end{eqnarray}
and use the time evolution to generate an entangled state. For that,
we notice that ion 1 can decay into three different states:
$|6^{2}S_{1/2},F=1,m=-1\rangle$, emitting a right-circularly
polarized ($\sigma_{+}$) photon, $|6^{2}S_{1/2},F=1,m=1\rangle$,
emitting a left-circularly polarized ($\sigma_{-}$) photon, and
$|6^{2}S_{1/2}, F=0,m=0\rangle$, emitting a linearly
polarized ($\Pi$) photon. Because we do not know which of the three
mentioned processes actually take place, the complete state of the
system is a linear superposition of the three corresponding
probability amplitudes. As a consequence, ion 1 and the radiation
field get entangled.
The geometry of the setup ensures that the photon wave packet
generated by the spontaneous decay of ion 1 propagates to the focus of
the second parabola, where it may excite ion 2. After the absorption, the
second ion is in the state $|6^{2}P_{1/2},F=1,m=1\rangle$,
if it absorbs a $\sigma_{+}$ polarized photon, and in the state
$|6^{2}P_{1/2},F=1,m=-1\rangle$, if it absorbs a
$\sigma_{-}$ polarized photon. These absorption processes map the field state onto the state
of the second ion and thereby generating entangled matter states.
The ion 2 being in an excited
state (in the $6^{2}P_{1/2},\; F=1$ manifold), is affected by
spontaneous decay. So, we have to perform a state transfer from the
manifold $6^{2}P_{1/2}$ to the manifold $6^{2}S_{1/2}$ that is
radiatively stable. One might think in using a single $\pi$-pulse,
but this is not a proper solution because the photon wave packet
radiated by ion 1 has a certain temporal width, which yields a
probabilistic determination for the time when the photon is absorbed
by ion 2. If ion 2 is still in the ground state when we apply the
$\pi$-pulse, the pulse does not have the desired effect. If we wait a
certain time to make sure that the absorption has already taken place
before applying the $\pi$-pulse, it is also likely that the
spontaneous decay process back to the $6^{2}S_{1/2}$ manifold may have
already occurred. We remind that unit excitation probability
can only be achieved with a time-reversed single-photon wave
packet~\cite{MSt}.
We suggest to use instead the spontaneous decay itself. To that end,
we have to take into account the different decay channels. For
example, consider that ion $2$ is in the state
$|6^{2}P_{1/2},F=1,m=1\rangle$: it can decay into the states
$|6^{2}S_{1/2},F=1,m=0\rangle$,
$|6^{2}S_{1/2},F=0,m=0\rangle$ and
$|6^{2}S_{1/2},F=1,m=1\rangle$. But only the last process
generates entanglement. In a similar
way, one can treat the case that ion $2$ is in the state
$|6^{2}P_{1/2},F=1,m=-1\rangle$.
\begin{figure}
\centering
\includegraphics[height=0.35\columnwidth]{fig2.eps}
\caption{(Color online) Hyperfine level scheme of a
$^{171}\text{Yb}^{+}$ ion: The states of the logical qubit,
depicted with lighter colors, are defined by the electronic levels
$|6^{2}S_{1/2}, F=1, m=-1\rangle$ and $|6^{2}S_{1/2}, F=1,
m=1\rangle$.}
\label{fig2}
\end{figure}
\begin{figure*}
\centering
\includegraphics[height=0.45\columnwidth]{fig3.eps}
\caption{
(Color online) Sequence of the processes which are the building
blocks for the remote entanglement preparation: The states of the
logical qubit, depicted with lighter colors, are defined by the
electronic levels $|6^{2}S_{1/2}, F=1, m=-1\rangle$ and
$|6^{2}S_{1/2}, F=1, m=1\rangle$. Optical transitions which are
necessary for the entanglement generation are indicated by solid
arrows, whereas the undesired transitions are indicated by dashed
arrows. The three columns correspond to three phases. The first
column shows the possible decay channels of the first ion's
initially prepared state $|6^{2}P_{1/2},F=1,m=0\rangle$; the
second column shows the possible excitation procedures of the
second ion which was initially prepared in the state
$|6^{2}S_{1/2},F=1,m=0\rangle$ followed by the spontaneous decay
processes used to accomplish the state transfer from the
$6^{2}P_{1/2}$ manifold to the radiatively stable $6^{2}S_{1/2}$
manifold; the last column shows the optical transitions used to
perform the postselection procedure based on hyperfine splitting
and off-resonant matter field interactions. }
\label{fig3}
\end{figure*}
To discard the undesired decay processes, we have to perform a
postselection. Since only in case of successful entanglement
generation both ions end up in the qubit state, by probing the
occupation of the qubit states we discard the undesired decay
processes. This can be performed with negligible loss of entanglement
by using dispersive state detection. It suffices to couple weak off-resonant
coherent pulses to the $\Pi$ transitions from ${S_{1/2},F=1
,m=\pm 1}$ to ${P_{1/2},F=1 ,m=\pm 1}$: population
is then detected by the phase shifts imprinted onto the coherent
states. This procedure allows us to check the population of the qubit
states while preserving the possible linear superpositions and thus
does not disturb the entangled state.
Furthermore, this postselection also detects photon losses, so that the
scheme is loss tolerant. This is due to the fact that upon photon loss ion\,2 remains in
$|S_{1/2},F=1,m=0\rangle$ and post selection is probed on
$\Pi$-transitions, which are for $|S_{1/2},F=1,m=0\rangle$
either forbidden or detuned so strongly that no phase shift of the
probe pulse occurs. Of course, losses reduce the success
probability, but the fidelity after a successful postselection is not
affected. The low success probability can be overcome with a high
repetition rate.
All the steps for generating entangled states described above are depicted in Fig.~\ref{fig3}.
\section{Theoretical analysis}
\label{model}
\subsection{System Hamiltonian}
\label{sub2.2}
In the rotating-wave and dipole approximations, the Hamiltonian of the
foregoing system can be written as
\begin{equation}
\hat{H} = \hat{H}_{A}+ \hat{H}_{R} + \hat{H}_{AR} \, ,
\end{equation}
where
\begin{eqnarray}
\hat{H}_{A} &=& \sum_{i\in S_{e}\cup S_{g}} \sum_{j\in S_{e}\cup S_{g}}
\hbar (\omega_{i} + \omega_{j}) \,
| i^{(1)}\rangle\,\langle i^{(1)}| \otimes|j^{(2)}\rangle\,\langle j^{(2)}|,
\nonumber \\
\hat{H}_{R} & = & \sum_{r} \hbar \omega_{r} \,
\hat{a}_{r}^{\dagger} \hat{a}_{r} , \\
\hat{H}_{AR} & = & -\sum_{\alpha\in \{ 1,2\}}
\hat{\mathbf{E}}^{+}(\mathbf{x}_{\alpha}) \cdot
\hat{\mathbf{d}}_{\alpha}^{-}+
{\rm H.\, c.} \nonumber
\label{Hamilton}
\end{eqnarray}
Here, $\hat{H}_{A}$ describes the dynamics of the matter. We indicate
by $S_{e}$ (excited) the set of states in the manifold
$6^{2}P_{1/2}$ and by $S_{g}$ (ground) the set of states in
the manifold $6^{2}S_{1/2}$. The vectors $|i^{(1)} \rangle$,
$|j^{(2)} \rangle$ represent states of the ion 1 and ion 2 living in
$i,j \in S_{e} \cup S_{g}$, with energies $\hbar \omega_{i}$ and
$\hbar \omega_{j}$, respectively. $\hat{H}_{R}$ gives the dynamics of
the field, characterized by the annihilation ($\hat{a}_{r}$) and
creation ($\hat{a}_{r}^{\dagger}$) operators of the modes (of
frequency $\omega_{r}$) that couple to the ions (they depend on the
boundary conditions). Finally, the interaction between the ions and
the field is given by $\hat{H}_{AR}$, wherein H.c. stands for the
Hermitian conjugate and
\begin{eqnarray}
\hat{\mathbf{E}}^{+} (\mathbf{r} ) & = & - i
\sum_{r}\sqrt{\frac{\hbar\omega_{r}}{2\epsilon_{0}}}
\mathbf{g}_{r} ( \mathbf{r} ) \, \hat{a}_{r}^{\dagger} , \nonumber \\
& &\\
\hat{\mathbf{d}}_{\alpha}^{-} & = & \sum_{i \in S_{e}}
\sum_{j \in S_{g}} \mathbf{d}_{ij} \,
|j ^{(\alpha)} \rangle \, \langle i^{(\alpha)} | , \nonumber
\end{eqnarray}
$\mathbf{x}_{1}$ and $\mathbf{x}_{2}$ being the position of the first
and second ion, respectively. The orthonormal mode functions
$\mathbf{g}_{r} (\mathbf{r})$ are solutions of the Helmholtz equation
with the proper boundary conditions, fulfilling, in addition, the
transversality condition $\nabla \cdot \mathbf{g}_{r} (\mathbf{r}
)=0$.
\subsection{Photon-path-representation}
\label{sub2.3}
Since only one excitation is available in our initial state, and the
Hamiltonian \eqref{Hamilton} preserves the number of excitations, the
state of the system at time $t$ can be written as
\begin{eqnarray}
\label{eq:ans}
| \psi(t) \rangle & = & \sum_{i \in S_{e}}
\sum_{j \in S_{g}} b_{ij}^{(1)} (t) \, |i^{(1)} \rangle
| j^{(2)} \rangle | \{ 0 \} \rangle \nonumber \\
& + & \sum_{i \in S_{g}} \sum_{j\in S_{e}}
b_{ji}^{(2)}(t) \, | i^{(1)} \rangle |j^{(2)} \rangle
| \{ 0 \} \rangle \nonumber \\
& + & \sum_{r} \sum_{i \in S_{g}}
\sum_{i \in S_{g}} f_{ij}^{(r)} (t) \, | i^{(1)} \rangle
| j^{(2)} \rangle | 1_{r} \rangle ,
\end{eqnarray}
where $| \{0 \} \rangle$ is the vacuum state
and $ | 1_{r} \rangle = \hat{a}^{\dagger}_{r} | \{ 0 \} \rangle$ a
single-photon state of the radiation field. The
amplitude $b_{ij}^{(1)}(t)$ describes the evolution when the field is
in the vacuum, the first ion is in one of the excited levels $i\in
S_{e}$ and the second ion is in one of the ground levels $j\in S_{g}$
and an analogous interpretation for $b_{ji}^{(2)}(t)$. The amplitude
$f_{ij}^{(r)}(t)$ is related to the evolution when there is an
excitation in the field mode $r$ and both ions are in one of the
ground electronic levels $i,j\in S_{g}$.
Now, we can solve the time-dependent Schr\"{o}dinger equation, with
the \emph{ansatz} \eqref{eq:ans}. If we assume that the
field is initially in a vacuum state and we use the Laplace transform, we
get, after eliminating the transforms of the probability amplitudes
for photonic excitations $\widetilde{f}_{ij}^{(r)}(s)$,
\begin{align}
s & \, \widetilde{b}_{ij}^{(\alpha)} (s) - b_{ij}^{(\alpha)} (0) =
- i (\omega_{i} + \omega_{j} ) \, \widetilde{b}_{ij}^{(\alpha)} (s) \nonumber \\
+ & \sum_{\beta \in \{ 1,2 \} }
\sum_{k \in S_{e}} \sum_{\ell \in S_{g}}
T_{\beta;k\ell}^{\alpha;ij} (s) \, \widetilde{b}_{k\ell}^{(\beta)}(s) \, ,
\label{eq:compact_scalar_equation-laplace_transform}
\end{align}
where $ \alpha \in \{ 1,2 \}$ indexes the ions, $i\in S_{e},\; j\in S_{g}$ and
\begin{equation}
T_{\beta;k\ell}^{\alpha;ij} (s ) =
\left \{
\begin{array}{ll}
\displaystyle
\delta_{j \ell} \sum_{m \in S_{g}}
A_{\alpha;im}^{\beta;km} (s + i \omega_{m}+i \omega_{j} ), & \alpha =\beta,\\
A_{\alpha;i\ell}^{\beta;kj} (s+i \omega_{j}+i \omega_{\ell} ), &
\alpha\neq \beta.
\end{array}
\right .
\end{equation}
The function
\begin{equation}
A_{\beta; k\ell}^{\alpha;ij} (s )
= \sum_{r}\frac{\kappa_{r}^{\alpha;ij}
\left(\kappa_{r}^{\beta; k\ell}\right)^{\ast}}{s+i \omega_{r}} ,
\end{equation}
with
\begin{equation}
\kappa_{r}^{\alpha;ij} =
\sqrt{\frac{\omega_{r}}{2\epsilon_{0}\hbar}}
\mathbf{d}_{ij}^{\dagger} \cdot \mathbf{g}_{r}
(\mathbf{x}_{\alpha} ) \, ,
\end{equation}
describes all possible photon emission and absorption processes and
encodes the whole geometry of the setup through the modes
$\mathbf{g}_{r}(\mathbf{r})$. Its explicit calculation turns to be a
difficult task, although in Appendix~\ref{sub2.4} we sketch a
semiclassical method.
Equation~\eqref{eq:compact_scalar_equation-laplace_transform} can be
recast in a suggestive vectorial form
\begin{equation}
(s+i \omega+ T) \, \widetilde{\mathbf{b}} (s) =\mathbf{b}(0)\;,
\label{Tsystem}
\end{equation}
where the functions $\widetilde{b}_{ij}^{(a)}$ are arranged in a
vector $\widetilde{\mathbf{b}}$ and $i \omega$ is a diagonal matrix
which represents the term $i (\omega_{1} +\omega_{2} )$. The
contribution $T$ can be split as
\begin{equation}
T=T_{0}+T_{1}\;,
\end{equation}
where $T_{0}$ is equal to $\frac{3}{2}\Gamma$, with $\Gamma$ being the
spontaneous decay rate in free space, and $T_{1}$ embodies
exponentially decaying terms of the form $e^{-s t}$. In principle, one
could try to perform an inverse Laplace transform to
solve~\eqref{Tsystem}. However, this involves finding the poles of the
integrand, which is a formidable task because $T$ depends itself on
$s$ in a highly nontrivial way.
To determine $\widetilde{\mathbf{b}}(s)$ we use instead an alternative
route based on the Neumann expansion
\begin{eqnarray}
(\openone -K)^{-1} = \openone + K + K^{2} +\dots+ K^{n} +\dots
\end{eqnarray}
If we take $K:=-T_{1}(s+i \omega+\frac{3}{2}\Gamma)^{-1}$ we get
\begin{equation}
\widetilde{\mathbf{b}} (s) =
\sum_{n=0}^{\infty} (s+i \omega+ 3\Gamma /2 )^{-1} K^{n}
\mathbf{b} (0) \, ,
\label{eq:Neumann_expansion}
\end{equation}
and each term in the series can be immediately Laplace inverted. The
price we pay is that we have to deal with an infinite series.
In most circumstances, only a few terms contribute to
Eq.~\eqref{eq:Neumann_expansion}: since
each summand is damped by an exponential of the form $e^{-s t}$, and
when applying the inverse Laplace transform, each term
leads to a Heaviside step function, i.e., a retardation, shifted into
the positive direction by time $t$. If we are looking at the evolution
in a finite time interval, we can neglect terms which are so far
retarded that they do not contribute.
The time interval of interest in our setup is of the order of $\tau =
(4f+d)/c$, which is the typical travel time of a photon to go
from the first to the second ion and $f$ being the focal length of the mirrors and $d$ the
separation between foci. In consequence, we can neglect all terms
$n>1$ in the sum; as shown in Ref.~\cite{Alber2013}, the higher-order
terms are relevant when the focal length is comparable with the
wavelength, which is not the case for our actual mirrors~\cite{Maiwald}
($f = $2.1 mm and wavelength $\lambda =369$ nm).
\section{Results}
\label{entanglement}
If one uses the method of the preceding section in the time interval
$t \in[0, 2\tau)$, it turns out that only four of the atomic
probability amplitudes $b_{ij}^{(\alpha)} $ ($ \alpha \in \{1, 2\} $) are of
relevance:
\begin{equation}
\begin{array}{c}
b_{6^{2}P_{1/2},\; F=1\; m=0,6^{2}S_{1/2},\; F=1\; m=0}^{(1)}(t),\\
b_{6^{2}P_{1/2},\; F=1\; m=1,6^{2}S_{1/2},\; F=1\; m=-1}^{(2)}(t),\\
b_{6^{2}P_{1/2},\; F=1\; m=-1,6^{2}S_{1/2},\; F=1\; m=1}^{(2)}(t),\\
b_{6^{2}P_{1/2},\; F=0\; m=0,6^{2}S_{1/2},\; F=0\; m=0}^{(2)}(t).
\end{array}
\end{equation}
If the hyperfine splitting is large in comparison to the spontaneous
decay rate $\Gamma$, as it happens for $^{171}\text{Yb}^{+}$ in the
time window of interest, we can also neglect
$b_{6^{2}P_{1/2},\; F=0\; m=0,6^{2}S_{1/2},\;
F=0\; m=0}^{(2)}(t)$.
\begin{figure}[t]
\begin{centering}
\includegraphics[width=0.95\columnwidth]{fig4.eps}
\end{centering}
\caption{(Color online) Time evolution of the excitation probability
$P$ in the case of ion\, 1 (dashed line) and of ion\, 2 (solid
line): The interaction time $t$ is plotted in units of the time
$\tau=(4f+d)/c$ which a photon needs to travel from the first ion
to the second ion. $f$ is the focal length of the parabolas and
$d$ is the distance between the foci. We set the spontaneous decay
$\Gamma\tau=3$ and the Zeeman splitting was neglected.}
\label{fig6}
\end{figure}
In Fig.~\ref{fig6} we plot the excitation probabilities of the two
ions for vanishing Zeeman splitting. As discussed in
Sec.~\ref{setup}, the process generates an entangled state if one uses the states
$|6^{2}P_{1/2},\; F=1\; m=1\rangle$ and
$|6^{2}P_{1/2},\; F=1\; m=-1\rangle$ of the second ion as
qubit. But these states do not form a stable qubit; spontaneous decay
transfers them to the ground states $|6^{2}S_{1/2},\; F=1\;
m=1\rangle$ and $|6^{2}S_{1/2},\; F=1\; m=-1\rangle$
by emitting a single photon. By detecting whether or not both ions
are in one of the states $|6^{2}S_{1/2},\; F=1\;
m=\pm1\rangle$ , we check if the entanglement generation was
successful.
The postselection is equivalent to a von Neumann measurement described
by the projection operator
\begin{equation}
\hat{P}=\ket{00}\bra{00}+\ket{01}\bra{01}+
\ket{10}\bra{10}+\ket{11}\bra{11},
\end{equation}
where $\ket{q_{1}q_{2}}$ ($q_1,q_2\in\left\{ 0,1\right\}$) correspond
to the states of the logical qubit. In addition, we also have to deal
with the photon emitted in the transfer from the excited to the ground
states. This photon, which carries information about the
state of the ions, might cause decoherence and therefore destroy the
entangled state generated by the time evolution. To certify that this
is not the case, we have to trace out the uncontrolled photonic
degrees of freedom, which amounts to know
\begin{equation}
\hat{\varrho} (t) =
\Tr_{R} ( | \psi(t) \rangle \langle \psi(t) | ) \, .
\end{equation}
This density matrix is evaluated in Appendix ~\ref{subTracing_out}. In the limit $\Gamma (t-\tau)\rightarrow\infty$ with
$\tau<t<2\tau$, we have that
\begin{small}
\begin{align}
\hat{P} & \hat{\varrho}(t) \hat{P} = \frac{2}{3 (9+\delta^{2})} \,
| 01 \rangle \langle 01| + \frac{2}{3\left(9+\delta^{2}\right)} |
10 \rangle \langle 10|
\nonumber \\
+ & \frac{2}{3[-9+\delta(-9i+2\delta)]} | 01 \rangle \langle 10| +
\frac{2}{3[ -9+\delta(9i+2\delta)]} | 10 \rangle \langle 01| \; .
\end{align}
\end{small}
The parameter $\delta= ( \Delta_{1}-\Delta_{2})/\Gamma$ characterizes
the Zeeman splitting of the energy levels: $\Delta_{1} m$ is the
splitting in the $6^{2}P_{1/2},\; F=1$ manifold and
$\Delta_{2} m$ the splitting in the $6^{2}S_{1/2},\; F=1$
manifold. Note that the magnetic field has the same orientation and
strength for both ions.
For $| \delta | \ll1$, which is justified in our experimental
setup~\cite{Maiwald}, we get
\begin{equation}
\hat{P}\hat{\varrho}(t)\hat{P} =
\frac{2}{27} ( | 01 \rangle-| 10 \rangle) (\langle 01|-\langle 10|) \, ,
\end{equation}
which corresponds to a maximal entangled state with a success
probability $4/27 \approx 15~\%$. Of course, in a real experiment, one
has to take additional effects into account. As we explore in
the next Section, it should be possible to achieve free-space
communication over several kilometers.
\section{Experimental feasibility}
\label{experiment}
\subsection{Realistic parabolic mirrors}
\label{sub:Non_perfect_boundary_conditions}
In a real setup, the parabolic mirror does not cover the full solid
angle. Actually, in our parabolic
mirror~\cite{Maiwald} we have
\begin{equation}
\Omega = \left\{ (\varphi,\theta) \, : \,
\varphi\in ( 0, 360^{\circ} ) ,
\theta \in (20^{\circ}, 135^{\circ} ) \right\} \, ,
\end{equation}
whereby the angle $135^{\circ}$ gives the front opening of the
parabola and the angle $20^{\circ}$ accounts for the hole on
the backside for inserting the ion trap. This has to be taken into
account in integrations as in Eq.~\eqref{eq:Gamma_rel}.
Furthermore, our mirrors are made out of aluminum, which has a finite
electrical conductivity. The properties of the material are well
described by introducing a frequency dependent dielectric constant
$\epsilon(\omega)$. In our case
$\epsilon(\omega)=-18.74+i3.37$~\cite{refractiveindex}. Now, we have
to split the field in a transverse electric (TE) and a magnetic (TM)
part and apply Fresnel equations to deal with the boundary
conditions. But these equations are different for the two basic
polarizations and give angle-dependent phase shifts and
reflectivities, which leads to a further reduction of the efficiency
for entanglement generation. One might think that this effect could
also reduce the fidelity of the entangled state but this is not the
case. Such a reduction could occur if a $\sigma_{+}$($\sigma_{-}$)
decay of the first ion could drive a $\sigma_{-}$($\sigma_{+})$
transition of the second ion, but due to the symmetry
this does not occur.
This is obvious from the following reasoning: After collimation by the
parabolic mirror, the polarization vector of the electric field in the
exit pupil of the parabolic mirror reads~\cite{sondermann2008}
\begin{align}
\label{eq.psigma}
\mathbf{\sigma}_\pm \simeq&
(r^2 - 4)(\cos{\phi}\pm i\sin{\phi})\cdot \mathbf{e}_r\\
& + (r^2 + 4)(\sin{\phi}\mp i\cos{\phi})\cdot \mathbf{e}_\phi \nonumber
\end{align}
with $r$ the distance to the optical axis in units of the mirror's
focal length, $\phi$ the azimuthal angle, and $\mathbf{e}_r$ and
$\mathbf{e}_\phi$ the unit vectors in radial and azimuthal
direction, respectively.
Upon reflection on the parabolic surface, these vectors correspond to
TM- and TE-components.
The influence of the metallic mirror can be accounted for by
additional complex pre-factors which depend on $r$ only.
It is straightforward to show that the overlap $\int
\tilde{\mathbf{\sigma}}_\pm\cdot\mathbf{\sigma}^\star_\mp$ of this
modified field $\tilde{\mathbf{\sigma}}_\pm$ with the state of
opposite helicity vanishes.
We can sum all the above effects in a factor $\eta$ which has to be
multiplied with the probability for a successful entanglement
creation to take the more realistic mirrors into account: $\eta=1$
corresponds to perfectly conducting parabolic mirrors that cover the
full solid angle. In the specific case treated here, we have
$\eta\approx 0.47$.
\subsection{Free-space versus fiber-based transmission}
\label{Freespace_vs_fiber}
Our scheme is designed to be compatible with free-space
communication by photonic qubits, for it does not rely on the
strong coupling regime, but on intrinsic multimode effects like
spontaneous emission.
There are other multimode schemes, such as the
one in Ref.~\cite{Moehring}, which might be adapted to free-space
communication, but our proposal offers considerable advantages.
The scheme in Ref.~\cite{Moehring} heavily relies on fibers as mode filters
to achieve almost perfect mode matching on a beam splitter and,
besides, the fidelity is mainly limited by the fact that the postselection
is performed on the radiation field and is sensitive to dark counts of
the detectors. In contrast, in our proposal, postselection is
performed on the ions, which circumvents detector dark
counts.
Of course, we have to take into account experimental
imperfections, mainly connected with the free-space transmission of
the one-photon wave packet. This gives rise to beam wandering and
phase-front distortions due to atmospheric
turbulences~\cite{Ursin07}. In both cases, the intensity at the focus
is reduced~\cite{lieb2001,leuchs2008,april2011}, affecting the success
probability. Once the distance between the two parabolic mirrors
becomes large enough, beam broadening plays a crucial role, which also
results in a lower success probability. All these effects,
however, diminish the success probability but seem to leave the
fidelity rather untouched, which is of big importance for practical
applications.
One could also think about the transmission of the photon from ion 1
to ion~2 via an optical fiber. This would circumvent all problems
related to atmospheric transmission, but, due to their complex
polarization pattern, cf. Eq.~\eqref{eq.psigma}, the photons collimated by the parabolic mirror
have subunit overlap with a fundamental Gaussian mode with circular
polarization. Hence the efficiency in coupling these photons to a
single-mode fiber is limited to a maximum of
49~\%~\cite{sondermann2008}. Therefore, fiber transmission alone would
limit the success probability
of our entanglement scheme to 24~\%. Moreover, the strong
attenuation of ultraviolet radiation in standard optical fibers
reduces the success probability by orders of magnitude, even for
distances about 1~km. Finally, fibers are not well suited to
perform communication via polarization coding \cite{Gisin}, as in
our scheme.
\subsection{Postselection}
\label{Post_selection}
As advanced in Sec.~\ref{setup}, the best way to perform postselection
seems to probe qubit states directly by dispersive state
detection. This can be implemented by coupling weak coherent-state
pulses to the $\Pi$ transitions from $S_{1/2},F=1 ,m=\pm
1$ to ${P_{1/2},F=1 ,m=\pm 1}$. Population in the
$S_{1/2},m=\pm 1$ states is then detected by the phase shifts
imprinted onto the coherent states. The detuning and pulse amplitudes
can be chosen such that one is far from saturating the respective
transitions. For example, choosing an on-resonance saturation
parameter of $s_{0}=0.01$ and a detuning of two linewidths, the
excitation probability is about $10^{-5}$, while the phase of the
coherent pulse is shifted by $25^{\circ}$,
according to the formalism of Ref.~\cite{sondermann2013p}.
One has to balance the amplitude and the detuning of the incident coherent
state carefully. Larger amplitudes and smaller detunings result in
lower error probabilities for detecting the phase of the coherent
state, but also enforce a stronger excitation of the ion. The latter
might lead to transferring the ion out of the $m=\pm 1$ state during
state detection, hindering the phase shift of the coherent state and
hence resulting in erroneous postselection. Furthermore, the
reflectivity of the beam splitters in front of the parabolic mirrors
not only affects the success probability of our entangling scheme, but
also influences the error in measuring the phase of the coherent
state.
\begin{figure}
\centerline{\includegraphics{fig5.eps}}
\caption{\label{fig.detect} (Color online) Error probability $P$ in
determining the phase of a coherent pulse probing the
$\Pi$-transitions from $S_{1/2}$ to $P_{1/2}$ plotted over the beam splitter reflectivity $R$. Solid line: ion~1,
dashed line: ion~2. In both cases, the relative detuning respect
to the resonance is two linewidths, corresponding to a phase shift
of $0.14\pi$. The length of the pulse is $10^4$ upper-state
lifetimes or 81~$\mu$s, respectively. The amplitude of the
coherent state incident onto the ion is chosen such that the
probability to excite the respective upper state is $5\times 10^{-4}$, as marked by the
thin dotted line. The calculation for ion~2 accounts for the
threshold reflectivity found for ion~1, which is marked by the
crossing of the solid and the dotted line.}
\end{figure}
We compute the corresponding error probabilities according to the
Helstrom bound~\cite{bergou2010}. The \emph{a priori} probabilities
in this calculation are obtained from all relevant branching ratios,
excitation probabilities, and reflectivities. The amplitude of the
coherent state is chosen such that the probability to excite the ions
with the probe pulse is $5\times 10^{-4}$. We also choose this value
as an upper bound for the acceptable error probability. This is
motivated by the fact that postselection schemes probing the
$m=0$ states are limited in fidelity to values $\le 0.995$ by the
branching ratio of the $P_{1/2}$ state to the $D_{3/2}$ state of
0.5\%. Keeping all errors in our postselection scheme an order of
magnitude below this value is reasonable and desirable.
The minimum error probabilities as a function of the reflectivity of
the beam splitters coupling the coherent states into the parabolic
mirrors is plotted in Fig.~\ref{fig.detect}. First, we
determined the reflectivity for the beam splitter in front of ion~1
that ensures being below the error threshold for a set of suitable
parameters, yielding $R_1=0.5$. Next, this result was used in the
calculations for ion~2, leading to $R_2=0.22$. From these
reflectivity values one would obtain a reduction of the success
probability for entanglement generation by 61~\%.
In practice the Helstrom bound will not be reached entirely, with the
actually obtainable error probability depending on the method employed
for measuring the phase of the probe pulse. Nevertheless the error
threshold marked in Fig.~\ref{fig.detect} can be reached. This may be
achieved at the cost of using beam splitters with larger
reflectivities and thus accepting lower success probabilities.
To guarantee that the entangled state is not destroyed, we have to
ensure that no information about the state of the qubit is extracted
by our postselection. The latter condition is fulfilled if the
magnetic field fixing the quantization axis is sufficiently small (,i.e., the frequency shifts caused by the Zeeman effect are small compared to the spontaneous decay rate), so that the
phase shift imprinted by an ion in the $m=-1$ Zeeman state will be
practically the same as for the other ion in the $m=+1$ state.
Therefore, probing the qubit dispersively will not project the ions
into one of these states and entanglement is preserved. The
parameter set in Fig.~\ref{fig.detect} yields a fidelity of 0.998
when postselecting. Even higher fidelities can be reached by larger
beam splitter reflectivities (accompanied by decreasing success
probabilities), lower pulse amplitudes or longer pulse lengths. A
lower pulse amplitude has to be compensated for by larger beam
splitter reflectivities or longer pulse lengths. The latter in turn
affects the repetition rate.
\subsection{Repetition rate}
We finally estimate the achievable repetition rate. Typically, an
experimental cycle starts with Doppler cooling the ion, which takes
about 200~$\mu$s for the ions treated
here~\cite{PhysRevA.80.022502}. For the trap frequencies inherent to
the parabolic mirror trap, 500\,kHz in radial direction and 1\,MHz
along the optical axis, the average number of motional quanta
according to the Doppler limit is 20 and 10, respectively. This
corresponds to widths of the ion wave function in position space about
0.13 and 0.07 wavelengths. With these numbers we estimate that the
ions experience 78~\% of the focal intensity obtained by diffraction
limited focusing. Applying only Doppler cooling the success
probability of our entanglement scheme would be reduced accordingly.
One could additionally apply resolved side-band cooling, but the
increase of the success rate is obviously accompanied by a lower
repetition rate due to the elongated cooling procedure. Furthermore,
as soon as there is a broadened focus due to incompletely compensated
atmospheric aberrations etc. the above spread of the ion's wave
function is negligible.
After cooling, both ions have to be prepared in the state
${S_{1/2},F=0}$ which takes less than 1~$\mu$s~\cite{Olmschenk1}.
Additionally, ion~2 has to be flipped to the state
${S_{1/2},F=1,m=0}$. This can be accomplished in 6~$\mu$s using
microwaves~\cite{Olmschenk1} or in 100~ps applying Raman
transitions~\cite{campbell2010}. Likewise, ion~1 is brought to the
${P_{1/2},F=1,m =0}$ state by an optical $\pi$-pulse on a time
scale smaller than a nanosecond. The postselection
requires around 80~$\mu$s, as it was outlined in
Sec.~\ref{Post_selection}. The photon traveling time from
ion~1 to ion~2 is of the order of 10~$\mu$s for distances of a few
kilometers. At least, the same time has to be spent in communicating
the postselection via a classical communication channel. Thus,
the time spent for state preparation, attempting entanglement of the
ions and postselection is on the order of 100\,$\mu$s.
From the numbers given above, one would estimate a repetition rate of
3.3\,kHz if Doppler cooling is applied after each entanglement
attempt. One could increase the repetition rate if Doppler cooling is
performed regularly after a certain number of entanglement trials.
Since one entanglement trial takes about 100\,$\mu$s, a repetition
rate in excess of 10\,kHz is not feasible, unless one accepts a
reduced fidelity and/or success probability. Assuming a realistic
heating rate of 10 quanta per ms~\cite{mcloughlin2011}, the
spread of the ion wave function would roughly double in radial
direction within 8~ms. Accepting the accompanying, continuously
increasing loss of success probability, one could enhance the
repetition rate towards 9.8~kHz, which is close to the inverse of the
duration of one entanglement trial. Anyhow, in every
experimental realization, the repetition rate is dictated by the
specific requirements on fidelity, success probability and inter-ion
distance.
\section{Concluding remarks}
\label{V. Conclusion}
In summary, we have presented a scheme for preparing maximally
entangled states of two matter qubits with high fidelity by using a
free-space channel. The qubits are encoded in the level
structure of two distant $^{171}\text{Yb}^{+}$ ions located
at the foci of two parabolic mirrors. The theoretical
description of the setup involves an extreme multimode
scenario to model the radiation field and a level structure
far more complicated than a simple two level atom.
We have used a semiclassical photon-path representation to deal
with the boundary conditions at the two parabolic mirrors, which leads
to a intuitive representation of the quantum dynamics of the two ions
and the radiation field.
To obtain a more realistic description, we have focused on the
experimental details in Ref.~\cite{Maiwald} and on more
realistic boundary conditions. Our results confirm the feasibility
of the scheme to achieve reasonable success probabilities, which in
combination with a relatively high repetition rate leads to a proper
rate for preparing entangled matter qubits. Indeed, we expect an
entanglement rate of 54 per second under diffraction-limited
focusing.
One of the main issues is the fidelity of these states. Our scheme is
robust against imperfections arising in the experimental
implementation. All these effects reduce the success
probability of entanglement generation, but leave the fidelity
untouched.
We hope that our work is a step towards an experimental realization of
remote entangled matter qubits in free space, which is a key building block for
future quantum technologies.
\begin{acknowledgments}
N.~T., J.~Z.~B., and G.~A. acknowledge support by the BMBF
Project Q.com and CASED III. M.~S. and G.~L. are grateful for the
financial support of the European Research Council under the
Advanced Grant PACART. Finally, L.~L.~S.~S. acknowledges the
support from the Spanish MINECO (Grant No. FIS2011-26786).
\end{acknowledgments}
|
\section{Introduction}
We consider here a classical fixed design regression model \[\forall i\in \{1,\ldots, n\},\, Y_i = f_0(x_i) + W_i \] with $f_0$ an unknown function, $x_i$ the fixed design points and $W$ a centered sub-Gaussian noise. Our aim is to estimate the function $f_0$ at the grid points. We study a strategy in which a collection of smoothed projection $\{\hat{f}_t(Y)=P_t Y|P_t \in \mathcal{S}^+_n(\mathbb{R}), t \in \mathcal{T}\}$ is aggregated into a single adaptive estimator using a PAC-Bayesian aggregation.
Aggregation procedures have been introduced by \citet{Vovk1990,Littlestone1994212,
Cesa-Bianchi1997, MR1765620}. They are a core ingredient of
bagging~\citep{raey}, boosting \citep{MR1348530,Schapire1990} or
random forest (\citet{Amit1997} or \citet{randomforests}; or
more recently \citet{MR2447310,MR2719877,MR2930634,Genuer2011}).
The general aggregation framework is detailed in \citet{MR1775640}
and studied in \citet{MR2163920,MR2483528} through a PAC-Baysian
framework as well as in \citet{MR1804557, MR1790617, MR1762904,
MR1946426, MR1997174, MR2028357, MR2044592}. See for instance
\citet{MR2503001} for a survey. Optimal rates of aggregation in
regression and density estimation are studied by
\citet{Tsybakov2003,MR2356820, MR2356821, Rigollet2006} and \citet{MR2364224}
We follow the exponentially weighted aggregation strategy, in which
the weight of each element in the collection is proportional to
$\exp\left( \frac{\widetilde{r_t}}{\beta}\right) \pi(t)$ where
$\widetilde{r_t}$ is a, possibly penalized, estimate of the risk of
$\hat{f}_t$ , $\beta$ is a positive parameter, called the temperature, that
has to be calibrated and $\pi$ is a prior measure over
$\mathcal{T}$. Our aim is to give sufficient conditions on the
penalized risk estimate and the temperature to obtain an oracle
inequality for the risk of our estimate.
This scheme here has been used first by \citet{MR2242356} that
have obtained the first exact oracle inequality in this setting by
aggregation projection with exponential weights in a Gaussian
regression framework. Those results have been extended to several
setting \citet{Dalalyan2007, DT08,
MR2926142,MR2543587,Dalalyan_Hebiri_Meziani_Salmon13,
MR2860324, MR2981354, MR3025131, MR3059085, MR2999166,MR3025134} under a
\emph{frozen} estimator assumption: they should not depend on the
observed sample. This restriction, not present in the work by
\citet{MR2242356}, has been removed by \citet{MR3059085} within
the context of affine estimator and exponentially weighted
aggregation. However, \citet{MR3015047} have shown the sub-optimality in deviation of exponential weighting, not allowing to obtain a sharp oracle inequality in probability. Nevertheless, penalizing the risk in the weights and taking a temperature at least 20 times greater than the noise variance allows to upper bound the risk of the aggregate in probability \citep{MR3192554}. Furthermore, the corresponding oracle inequality is not sharp.
Our contribution is twofold. First, we propose the first extension to general sub-Gaussian noise. Second, we conduct a fine analysis of the relationship between the choice of the penalty and the temperature. In particular, we are able to take into account the signal to noise ratio to provide sharp oracle inequalities for bounded functions.
Not that our results are similar to the one obtained for a
slightly different aggregation scheme by
\cite{bellec:_concen} in a preprint written while the authors
were working independently on this one.
\section{Framework and estimate}
Recall that we observe \[ \forall i \in \{1,\ldots, n\},\, Y_i = f_0(x_i) + W_i \] with $f_0$ an unknown function and $x_i$ the fixed grid points. Our main assumption on the noise is that $W \in \mathbb{R}^n$ is a centered sub-Gaussian variable, i.e. $\mathbb{E}(W)=0$ and there exists $\sigma^2 \in \mathbb{R}^+$ such that \[ \forall \alpha \in \mathbb{R}^n,\,
\mathbb{E} \left[ \exp \left(\alpha^\top W \right) \right] \leq \exp \left(\frac{\sigma^2}{2} \|\alpha\|^2_2 \right), \] where $\|.\|_2$ is the usual euclidean norm in $\mathbb{R}^n$.
If $W$ is a centered Gaussian vector with covariance matrix $\Sigma$ then $\sigma^2$ is nothing but the largest eigenvalue of $\Sigma$.
The quality of our estimate will be measured through its error at the design point points. More precisely, we will consider the classical euclidean loss, related to the squared norm
\[ \|g\|^2_2= \sum_{i=1}^n g(x_i)^2.\]
Thus, our unknown is the vector $(f_0(x_i))_{i=1}^n$ rather than the function $f_0$.
Assume that we have at hand a collection of data dependent smoothed projection estimates
\[ \hat{f}_t(Y) = \sum_{i=1}^n \rho_{t,i} \langle Y, b_{t,i} \rangle b_{t,i} \]
where $(b_{t,i})_{i=1}^n$ is an orthonormal basis and $(\rho_{t,i})_{i=1}^n$ a sequence of non-negative real numbers. For such an estimate, it exists a symmetric positive semi-definite real matrix of size $n,$ $P_t$ such that $\hat{f}_t(Y) = P_t Y$. For the sake of simplicity, we use this representation of our estimators. Note that we depart from the affine estimator studied by \citet{MR3059085} and \citet{MR3192554} because we consider only linear estimate. This choice was made to simplify our exposition but similar results as the one we obtain for linear estimates hold for affine ones.
To define our estimate from the collection $\{\hat{f}_t(Y)=P_t Y|P_t \in \mathcal{S}^+_n(\mathbb{R}), t \in \mathcal{T}\}$, we specify the estimate $\widetilde{r_t}$ of the risk of the estimator $\hat{f}_t(Y)$, choose a prior probability measure $\pi$ over $\mathcal{T}$ and a temperature $\beta>0$. We define $f_{EWA}$ by $f_{EWA}=\int \hat{f}_t \, d\rho(t),$ with
\[
d\rho(t)=\frac{\exp\left(-\frac{1}{\beta}\widetilde{r_t} \right)}
{\int \exp\left(-\frac{1}{\beta}\widetilde{r_{t'}} \right)
d\pi(t')} d\pi(t)
\]
a probability measure over $\mathcal{T}$. The intuition behind this construction to favor low risk estimates.
When the temperature goes to $0$ this estimator becomes very similar to the one minimizing the risk estimates while it becomes an indiscriminate average when $\beta$ grows to infinity. The choice of the temperature appears thus to be crucial and a low temperature seems to be desirable.
Our choice for the risk estimate $\widetilde{r_t}$ is to use the classical Stein unbiased estimate
\[ r_t=\|Y-\hat{f}_t(Y)\|^2_2+ 2 \sigma^2 tr(P_t)- n \sigma^2 \]
to which a penalty $\mathop{\mathrm{pen}}(t)$ is added. We will consider simultaneously the case of a penalty that depends on $f_0$ through an upper bound of a kind of sup norm and the case of a penalty that does not depend on $f_0$.
More precisely, we allow the use, at least in the analysis, of an upper bound $\widetilde{\|f_0\|_{\infty}}$ which can be thought as the supremum of the sup norm of the coefficients of $f_0$ in any basis appearing in
$\mathcal{T}$. Indeed, we define $\widetilde{\|f_0\|_{\infty}}$ as the smallest non negative real number $C$ such that for any $t \in \mathcal{T}$,
\begin{align*}
\|P_t f_0\|_2^2 \leq C^2 tr(P_t^2).
\end{align*}
By construction, $\widetilde{\|f_0\|_{\infty}}$ is indeed smaller than the sup norm of any coefficients of $f_0$ in any basis appearing in the $\mathcal{T}$. Note that $\widetilde{\|f_0\|_{\infty}}$ can also be upper bounded by $\|f_0\|_1$, $\|f_0\|_2$ or $\sqrt{n}\|f_0\|_{\infty}$ where the $\ell_1$ and sup norm can be taken in any basis.
Our aim is to obtain sufficient conditions on the penalty $\mathop{\mathrm{pen}}(t)$ and the temperature $\beta$ so that an oracle inequality of type
\begin{multline*}
\|f_0 - f_{EWA}\|^2_2 \leq \inf_{\mu \in \mathcal{M}_+^1(\mathcal{T})}
(1 +\epsilon) \int \|f_0 - \hat{f}_t\|^2_2 d\mu(t)
+ (1 + \epsilon') \int (\mathop{\mathrm{pen}}(t) + \mathop{\mathrm{price}}(t)) d\mu(t) \\
+ (1 + \epsilon') \beta \left( 2KL(\mu,\pi) + \ln\frac{1}{\eta} \right)
\end{multline*}
holds, with $\epsilon$ and $\epsilon'$ small non-negative numbers possibly equal to $0$ and $\mathop{\mathrm{price}}(t)$ a loss depending on the choice of $\mathop{\mathrm{pen}}(t)$ and $\beta$. Such an oracle proves that the risk of our aggregate estimate is of the same order as the one of the best estimate in the collection up to some controlled cost.
\section{A general oracle inequality}
Our main result is the following:
\begin{thm}
\label{thm:general_PAC}
Assume $W$ is a centered sub-Gaussian noise with parameter
$\sigma^2$. Assume $\{\hat{f}_t(Y)=P_t Y|P_t \in \mathcal{S}^+_n(\mathbb{R}), t \in
\mathcal{T}\}$ are such that there exists $V >0$ satisfying $\sup_{t \in \mathcal{T}} \| P_t \|_2 \leq V$.
Let $\pi$ be a arbitrary prior measure on $\mathcal{T}$, $\beta >4 \sigma^2 V$ an
arbitrary temperature and $\mathop{\mathrm{pen}}(t)$ a penalty so that
$f_{EWA}=\int \hat{f}_t \, d\rho(t)$ with
\[
d\rho(t)=\frac{\exp\left(-\frac{1}{\beta}\left[r_t + \mathop{\mathrm{pen}}(t)\right] \right)}
{\int \exp\left(-\frac{1}{\beta}\left[ r_{t'} + \mathop{\mathrm{pen}}(t')\right] \right)
d\pi(t')} d\pi(t)
\]
For any $\delta \in [0,1]$, if $\beta \geq 4 \sigma^2 V (1+4\delta ),$
let \[ \gamma=\frac{\beta -4 \sigma^2 V (1+2\delta)-\sqrt{\beta -4 \sigma^2 V}\sqrt{\beta -4 \sigma^2 V (1+4\delta)}}{16 \sigma^2 \delta V^2} \mathds{1}_{\delta>0}. \]
If for any $t \in \mathcal{T},$
\begin{align*}
\mathop{\mathrm{pen}}(t) \geq
\frac{4\sigma^2}{\beta -4\sigma^2 V}\left(1+(1-\delta)(1+2\gamma V)^2
\frac{\widetilde{\|f_0\|_\infty^2}}{\sigma^2} \right) tr(P_t^2) \sigma^2,
\end{align*}
then
\begin{itemize}
\item for any $\eta \in (0,1]$, with probability at least $1-\eta,$
\begin{multline*}
\|f_0-f_{EWA}\|_2^2 \leq \inf_{\nu \in N} \inf_{\mu \in \mathcal{M}_+^1(\mathcal{T})}
\left(1+\epsilon(\nu) \right) \int \|f_0-\hat{f}_t\|_2^2 d\mu(t)\\
+(1+\epsilon'(\nu)) \int \left( \mathop{\mathrm{pen}}(t)+ \mathop{\mathrm{price}}(t) \right) d\mu(t)
+\beta (1+\epsilon'(\nu)) \left(2KL(\mu,\pi)+\ln\frac{1}{\eta} \right)
\end{multline*}
\item Furthermore
\begin{multline*}
\mathbb{E} \|f_0-f_{EWA}\|_2^2 \leq \inf_{\nu \in N} \inf_{\mu \in
\mathcal{M}_+^1(\mathcal{T})} \left(1+\epsilon(\nu) \right) \int \mathbb{E}
\|f_0-\hat{f}_t\|_2^2 d\mu(t)\\
+(1+\epsilon'(\nu)) \int \left( \mathop{\mathrm{pen}}(t)+ \mathop{\mathrm{price}}(t) \right)
d\mu(t)
+2\beta (1+\epsilon'(\nu))KL(\mu,\pi)
\end{multline*}
\end{itemize}
with
\begin{align*}
\mathop{\mathrm{price}}(t)& = 2\sigma^2 \left(tr(P_t)+\frac{2\sigma^2(1-\delta) (1+2\gamma
V)^2}{\beta -4 \sigma^2 V} \frac{\widetilde{\|f_0\|_\infty^2}}{\sigma^2} tr(P_t^2) \right)\\
\epsilon'(\nu)&=\frac{1}{1-(1+\nu)\gamma}-1\\
\epsilon(\nu)&=\frac{(1+\nu)^2 \gamma}{\nu(1-(1+\nu)\gamma)}=\frac{(1+\nu)^2}{\nu} \gamma (1+\epsilon'(\nu))
\end{align*}
and $N=\{\nu>0 | (1+\nu)\gamma<1\}$.
\end{thm}
This theorem is similar to the one obtained by \citet{MR3192554}. It yields a sufficient condition on the penalty for oracle inequalities to hold both in probability and in expectation. It holds however under a milder sub-Gaussianity assumption on the noise and allows to take into account a sup norm information in the penalty as used for instance in~\citet{MR3020421} . Note that the result in expectation requires a penalty that is not necessary, at least in the Gaussian case, as shown by \citet{MR3059085}.
If we are authorized to use the upper bound of the sup norm, we may ensure that the penalty satisfies the lower bound condition with $\delta=0$. In that case, $\gamma=0$ and $\epsilon'(\nu)=0, \epsilon(\nu)=0$. Thus, there is no need to optimize $\nu$ and it suffices to notice that $N=(0,+\infty)$ to obtain that
\begin{coro}
\label{thm:general_PAC2} Under the assumptions of Theorem~\ref{thm:general_PAC},
if $\beta > 4 \sigma^2 V,$
if for any $t \in \mathcal{T},$
\begin{align*}
\mathop{\mathrm{pen}}(t) \geq
\frac{4\sigma^4}{\beta -4\sigma^2 V}\left(1+
\frac{\widetilde{\|f_0\|_\infty^2}}{\sigma^2} \right) tr(P_t^2),
\end{align*}
then
\begin{itemize}
\item for any $\eta \in (0,1]$, with probability at least $1-\eta,$
\begin{multline*}
\|f_0-f_{EWA}\|_2^2 \leq \inf_{\mu \in \mathcal{M}_+^1(\mathcal{T})}
\int \|f_0-\hat{f}_t\|_2^2 d\mu(t)\\
+ \int \left( \mathop{\mathrm{pen}}(t)+ \mathop{\mathrm{price}}(t) \right) d\mu(t)
+\beta \left(2KL(\mu,\pi)+\ln\frac{1}{\eta} \right)
\end{multline*}
\item Furthermore
\begin{multline*}
\mathbb{E} \|f_0-f_{EWA}\|_2^2 \leq \inf_{\mu \in
\mathcal{M}_+^1(\mathcal{T})} \int \mathbb{E}
\|f_0-\hat{f}_t\|_2^2 d\mu(t)\\
+ \int \left( \mathop{\mathrm{pen}}(t)+ \mathop{\mathrm{price}}(t) \right)
d\mu(t)
+2\beta KL(\mu,\pi)
\end{multline*}
\end{itemize}
with
\begin{align*}
\mathop{\mathrm{price}}(t)& = 2\sigma^2 \left(tr(P_t)+\frac{2}{\beta -4 \sigma^2 V} \widetilde{\|f_0\|_\infty^2} tr(P_t^2) \right).
\end{align*}
\end{coro}
As soon as $\delta>0$, a simple calculation yields that for any $\beta \geq 4 \sigma^2 V(1+4\delta ),$
$0<2\gamma V\leq 1$. As a result, if $V>0.5$, $(0, 2V-1) \subseteq N$. Furthermore if $\beta > 4 \sigma^2 V (1+4\delta )$ and $V>0.5$, then $2V-1 \in N$.
Else, if $\delta>0$ and $0<V \leq 0.5$, $N$ is non-empty if and only if $\beta > 4 \sigma^2 V + 2 \sigma^2 \delta (1+2V)^2$, which is a stronger condition than $\beta \geq 4 \sigma^2 V(1+4\delta )$. Since $V$ is an upper bound, it may be chosen greater than $0.5$.
If $\delta=1$, we obtain weak oracle inequalities that do not require the use of any side information:
\begin{coro}
\label{thm:general_PAC3} Under the assumptions of Theorem~\ref{thm:general_PAC}, if $\beta \geq 20 \sigma^2 V,$ let \[ \gamma=\frac{\beta -12 \sigma^2 V -\sqrt{\beta -4 \sigma^2 V}\sqrt{\beta -20 \sigma^2 V}}{16 \sigma^2 V^2}. \]
If for any $t \in \mathcal{T},$
\begin{align*}
\mathop{\mathrm{pen}}(t) \geq
\frac{4\sigma^4}{\beta -4\sigma^2 V} tr(P_t^2),
\end{align*}
then
\begin{itemize}
\item for any $\eta \in (0,1]$, with probability at least $1-\eta,$
\begin{multline*}
\|f_0-f_{EWA}\|_2^2 \leq \inf_{\nu \in N} \inf_{\mu \in \mathcal{M}_+^1(\mathcal{T})}
\left(1+\epsilon(\nu) \right) \int \|f_0-\hat{f}_t\|_2^2 d\mu(t)\\
+(1+\epsilon'(\nu)) \int \left( \mathop{\mathrm{pen}}(t)+ \mathop{\mathrm{price}}(t) \right) d\mu(t)
+\beta (1+\epsilon'(\nu)) \left(2KL(\mu,\pi)+\ln\frac{1}{\eta} \right)
\end{multline*}
\item Furthermore
\begin{multline*}
\mathbb{E} \|f_0-f_{EWA}\|_2^2 \leq \inf_{\nu \in N} \inf_{\mu \in
\mathcal{M}_+^1(\mathcal{T})} \left(1+\epsilon(\nu) \right) \int \mathbb{E}
\|f_0-\hat{f}_t\|_2^2 d\mu(t)\\
+(1+\epsilon'(\nu)) \int \left( \mathop{\mathrm{pen}}(t)+ \mathop{\mathrm{price}}(t) \right)
d\mu(t)
+2\beta (1+\epsilon'(\nu))KL(\mu,\pi)
\end{multline*}
\end{itemize}
with
\begin{align*}
\mathop{\mathrm{price}}(t)& = 2\sigma^2 tr(P_t)\\
\epsilon'(\nu)&=\frac{1}{1-(1+\nu)\gamma}-1\\
\epsilon(\nu)&=\frac{(1+\nu)^2 \gamma}{\nu(1-(1+\nu)\gamma)}=\frac{(1+\nu)^2}{\nu} \gamma (1+\epsilon'(\nu))
\end{align*}
and $N=\{\nu>0 | (1+\nu)\gamma<1\}$.
\end{coro}
Note that the parameter $\gamma$ allows us to obtain a weak oracle inequality. It links $\|(P_t-P_u)f_0\|_2^2$ to $\|(P_t-P_u)Y\|_2^2$.
Finally, assume that we let
\begin{align*}
\mathop{\mathrm{pen}}(t) \geq \kappa tr(P_t^2) \sigma^2.
\end{align*}
The previous corollary implies that a weak oracle inequality holds for any temperature greater than $20\sigma^2 V$ as soon as $\kappa \geq \frac{4 \sigma^2}{\beta -4\sigma^2 V}$.
Corollary~\ref{thm:general_PAC2} implies that an exact oracle inequality holds for any vector $f_0$ and any temperature $\beta$ greater than $4\sigma^2V$ as soon as
\begin{align*}
\frac{\beta -4\sigma^2 V}{4\sigma^2
}\kappa - 1 \geq
(1+2\gamma V)^2 \frac{\widetilde{\|f_0\|_\infty^2}}{\sigma^2}.
\end{align*}
For fixed $\kappa$ and $\beta$, this corresponds to a low peak signal to noise ratio $\frac{\widetilde{\|f_0\|_\infty^2}}{\sigma^2}$.
Theorem~\ref{thm:general_PAC} shows that there is a continuum between those two cases as weak oracle inequalities, with smaller leading constant than the one of Corollary~\ref{thm:general_PAC3}, hold as soon as
there exists
$\delta \in [0,1]$ such that $\beta \geq 4 \sigma^2 (1+4\delta) V$ and
\begin{align*}
\frac{\beta -4\sigma^2 V}{4\sigma^2
}\kappa - 1 \geq
(1-\delta)(1+2\gamma V)^2 \frac{\widetilde{\|f_0\|_\infty^2}}{\sigma^2},
\end{align*}
where the signal to noise ratio guides the transition.
The temperature required remains nevertheless
always above $4 \sigma^2 V$.
The minimal temperature of $4\sigma^2 V(1+4\delta)$ can be replaced by some smaller values if one further restrict the smoothed projections used. As it appears in the proof, the temperature can be replaced by $4\sigma^2 (1+\delta)$ or even $2\sigma^2(2+\delta)$ when the smoothed projections are respectively classical projections (see Theorem~\ref{thm:Gaussian}) and projections in the same basis. The question of the minimality of such temperature is still open. Note that in this proof, there is no loss due to the sub-Gaussianity assumption, since the same upper bound on the exponential moment of the deviation as in the Gaussian case are found, providing the same penalty and bound on temperature.
The proof of this result is quite long and thus postponed in Appendix~\ref{sec:Appendix_Proofs_subGaussian}. We provide first the generic proof of the oracle inequalities, highlighting the role of Gibbs measure and of some control in deviation. Then, we focus on the aggregation of projection estimators in the Gaussian model. This example already conveys all the ideas used in the complete proof of the deviation lemma~: exponential moments inequalities for Gaussian quadratic form
and the control of the bias $\|f_0-P_t f\|_2^2$ by $\widetilde{\|f_0\|^2}$ on the one hand, to obtain an exact oracle inequality, and by $\|f_0-P_t Y\|_2^2$ on the other hand, giving a weak inequality.
The extension to the general case is obtained by showing that similar exponential moments inequalities can be obtained for quadratic form of sub-Gaussian random variables, working along the fact that the systematic bias $\|f_0-P_t f\|_2^2$ is no longer always smaller than $\|f_0-P_t Y\|_2^2$ and providing a fine tuning optimization allowing the equality in the constraint on $\beta$ and an optimization on the parameters $\epsilon$.
\section{Proof of the oracle inequalities}
\label{sec:Proof_oracle}
Theorem~\ref{thm:general_PAC} relies on the characterization of Gibbs measure (Lemma~\ref{Gibbs}) and a control of deviation of the empirical risk of any aggregate around its true risk, allowed by Lemma~\ref{lem:Dev_Gaussian} or Lemma~\ref{lem:Dev_General}.
$\rho$ is a Gibbs measure. Therefore it maximizes the entropy for a given expected energy.
That is the subject of Lemma 1.1.3 in~\citet{MR2483528}:
\begin{lem}
\label{Gibbs}
For any bounded measurable function $h:\mathcal{T} \rightarrow \mathbb{R},$ and any probability distribution $\rho \in \mathcal{M}_+^1\left(\mathcal{T}\right)$ such that $KL(\rho,\pi)<\infty,$
\[ \log \left( \int \exp(h) d\pi \right)= \int h d\rho-KL(\rho,\pi)+KL(\rho,\pi_{\exp(h)}),
\]
where by definition $\frac{d\pi_{\exp(h)}}{d\pi}=\frac{\exp[h(t)]}{\int \exp(h) d\pi }.$
Consequently,
\[ \log \left( \int \exp(h) d\pi \right)=\sup_{\rho \in \mathcal{M}_+^1\left(\mathcal{T}\right)} \int h d\rho-KL(\rho,\pi). \]
\end{lem}
With $h(t)=-\frac{1}{\beta} [r_t+\mathop{\mathrm{pen}}(t)],$ this lemma states that for any probability distribution $\mu \in \mathcal{M}_+^1\left(\mathcal{T}\right)$ such that $KL(\mu,\pi)<\infty,$
\[ \int h d\rho -KL(\rho,\pi) \geq \int h d\mu -KL(\mu,\pi). \]
Equivalently,
\begin{align*}
&\int \|f_0-\hat{f}_t\|_2^2 d\rho(t) +\int \left(r_t-\|f_0-\hat{f}_t\|_2^2 +\mathop{\mathrm{pen}}(t) \right) d\rho(t) +\beta KL(\rho,\pi)\\
& \qquad \leq
\int \|f_0-\hat{f}_t\|_2^2 d\mu(t)
+\int \left(r_t-\|f_0-\hat{f}_t\|_2^2 +\mathop{\mathrm{pen}}(t) \right) d\mu(t) +\beta KL(\mu,\pi)
\\
\Leftrightarrow
&\int \|f_0-\hat{f}_t\|_2^2 d\rho(t) -\int \|f_0-\hat{f}_t\|_2^2 d\mu(t) \leq \int \left(\|f_0-\hat{f}_t\|_2^2-r_t \right) d\rho(t)-\beta KL(\rho,\pi) \\
&-\int \left(\|f_0-\hat{f}_t\|_2^2-r_t \right) d\mu(t)
-\int \mathop{\mathrm{pen}}(t) d\rho(t)+\int \mathop{\mathrm{pen}}(t)d\mu(t)
+\beta \mathrm{KL}(\mu,\pi).
\end{align*}
The key is to upper bound the right-hand side with terms that may depend on $\rho,$ but only through $\int \|f_0-\hat{f}_t\|_2^2 d\rho(t)$ and Kullback-Leibler distance. This is the purpose of Lemma~\ref{lem:Dev_Gaussian} in the case of Gaussian noise with projections estimators and Lemma~\ref{lem:Dev_General} in the sub-Gaussian case.
Under mild assumptions, they provide upper bounds in probability (and in expectation) of type:
\begin{multline*}
\int\left(\|f_0-\hat{f}_t\|^2_2-r_t\right)d\rho(t)-\int\left(\|f_0-\hat{f}_u\|^2_2-r_u \right) d\mu(u)
\\
\leq
C_1 \int \|f_0-\hat{f}_t\|_2^2 d\rho(t) + C_2 \int \|f_0-\hat{f}_u\|_2^2 d\mu(u)
\\
+C_3 \int tr(P_t^2) d\rho(t)
+ C_4 \int tr(P_u) d\mu(u) + C_5 \int tr(P_u^2) d\mu(u)
\\
+\beta \left( KL(\rho,\pi)+ KL(\mu,\pi)+ \ln\frac{1}{\eta} \right)
\end{multline*}
where $C_1$ to $C_6$ are known functions.
Combining with the previous inequality and taking $\mathop{\mathrm{pen}}(t) \geq C_3\, tr(P_t^2)$ gives
\begin{multline*}
(1-C_1)\int \|f_0-\hat{f}_t\|_2^2 d\rho(t) -(1+C_2)\int \|f_0-\hat{f}_t\|_2^2 d\mu(t) \\
\leq
C_4 \int tr(P_u) d\mu(u) + C_5 \int tr(P_u^2) d\mu(u) +\int \mathop{\mathrm{pen}}(t)d\mu(t)
\\
+\beta \left( 2 KL(\mu,\pi)+ \ln\frac{1}{\eta} \right).
\end{multline*}
The additional condition $C_1<1$ allows to conclude.
It is now clear that the whole work lies in the obtention of the lemma.
\section{The expository case of Gaussian noise and projection estimates}
In this section, to provide a simplified proof, we assume that $P_t$ are the matrices of orthogonal projections and the noise $W$ is a centered Gaussian random variable with variance $\sigma^2 I$. The previous theorem becomes:
\begin{thm}
\label{thm:Gaussian}
Let $\pi$ be an arbitrary prior measure over $\mathcal{T}$. For any $\delta \in [0,1]$, any $\beta>4\sigma^2(\delta+1)$, the aggregate estimator $f_{EWA}$ defined with \[ \mathop{\mathrm{pen}}(t) \geq \frac{2\sigma^4}{\beta -4\sigma^2} \left(1+2(1-\delta) \frac{\widetilde{\|f_0\|_\infty^2}}{\sigma^2} \right) tr(P_t) \] satisfies
\begin{itemize}
\item for any $\eta \in (0,1]$, with probability at least $1-\eta$,
\begin{multline*}
\|f_0-f_{EWA}\|_2^2 \leq \inf_{\mu \in \mathcal{M}_+^1(\mathcal{T})}
\left(1+2\epsilon \right) \int \|f_0-\hat{f}_t\|_2^2 d\mu(t)\\
+(1+\epsilon) \int \left( \mathop{\mathrm{pen}}(t)+ \mathop{\mathrm{price}}(t) \right) d\mu(t)
+\beta (1+\epsilon) \left(2KL(\mu,\pi)+\ln\frac{1}{\eta} \right).
\end{multline*}
\item Furthermore,
\begin{multline*}
\mathbb{E} \|f_0-f_{EWA}\|_2^2 \leq \inf_{\mu \in \mathcal{M}_+^1(\mathcal{T})}
\left(1+2\epsilon \right) \int \mathbb{E} \|f_0-\hat{f}_t\|_2^2 d\mu(t)\\
+(1+\epsilon) \int \left( \mathop{\mathrm{pen}}(t)+ \mathop{\mathrm{price}}(t) \right) d\mu(t)
+ 2 \beta (1+\epsilon) 2KL(\mu,\pi),
\end{multline*}
\end{itemize}
with
\[
\mathop{\mathrm{price}}(t) =2\left(1+\frac{2(1-\delta)\sigma^2}{\beta -4 \sigma^2} \frac{\widetilde{\|f_0\|_\infty^2}}{\sigma^2} \right) tr(P_t)\sigma^2
\quad \text{ and }\quad
\epsilon=\frac{4\sigma^2 \delta}{\beta -4\sigma^2 (\delta+1)}.
\]
\end{thm}
Note that $\mathop{\mathrm{pen}}(t) \geq \mathop{\mathrm{price}}(t)+ 2 \sigma^2
\left(\frac{\sigma^2}{\beta - 4\sigma^2}-1 \right) tr(P_t)$, and the
result may be further simplified using $\mathop{\mathrm{pen}}(t)+\mathop{\mathrm{price}}(t) \leq 2
\left( \mathop{\mathrm{pen}}(t)+ \sigma^2 tr(P_t)\right)$.
As announced in the scheme of proof of the oracle inequalities (section~\ref{sec:Proof_oracle}), the key is
a control of the deviation of the empirical risk of any aggregate around its true risk. It is allowed by Lemma~\ref{lem:Dev_Gaussian} in this case.
\begin{lem}
\label{lem:Dev_Gaussian}
For any prior probability distribution $\pi,$ any $\delta \in [0,1]$ and any $\beta>4\sigma^2,$ for any probability distributions $\rho$ and $\mu$,
\begin{itemize}
\item For any $\eta>0,$ with probability at least $1-\eta,$
\begin{multline*}
\int\left(\|f_0-\hat{f}_t\|^2_2-r_t\right)d\rho(t)-\int\left(\|f_0-\hat{f}_u\|^2_2-r_u \right) d\mu(u)
\\
\leq
\frac{4\delta\sigma^2}{\beta -4\sigma^2} \left( \int \|f_0-\hat{f}_t\|_2^2 d\rho(t) + \int \|f_0-\hat{f}_u\|_2^2 d\mu(u)\right)
\\
+\frac{4\sigma^2}{\beta -4\sigma^2} \left(\sigma^2+(1-\delta) \widetilde{\|f_0\|_\infty^2} \right) \int tr(P_t) d\rho(t)
\\
+ 2\sigma^2 \left(1+\frac{2(1-\delta)\widetilde{\|f_0\|_\infty^2}}{\beta -4\sigma^2} \right) \int tr(P_u) d\mu(u)
\\
+\beta \left( KL(\rho,\pi)+ KL(\mu,\pi)+ \ln\frac{1}{\eta} \right)
\end{multline*}
\item Moreover,
\begin{multline*}
\mathbb{E} \left[ \int\left(\|f_0-\hat{f}_t\|^2_2-r_t\right)d\rho(t)-\int\left(\|f_0-\hat{f}_u\|^2_2-r_u \right) d\mu(u) \right]
\\
\leq \mathbb{E} \left[
\frac{4\delta\sigma^2}{\beta -4\sigma^2} \left( \int \|f_0-\hat{f}_t\|_2^2 d\rho(t) + \int \|f_0-\hat{f}_u\|_2^2 d\mu(u)\right)
\right. \\
+\frac{4\sigma^2}{\beta -4\sigma^2} \left(\sigma^2+(1-\delta) \widetilde{\|f_0\|_\infty^2} \right) \int tr(P_t) d\rho(t)
\\
+ 2\sigma^2 \left(1+\frac{2(1-\delta)\widetilde{\|f_0\|_\infty^2}}{\beta -4\sigma^2} \right) \int tr(P_u) d\mu(u)
\\
+\beta \left( KL(\rho,\pi)+ KL(\mu,\pi) \right) \Big].
\end{multline*}
\end{itemize}
\end{lem}
The use of this lemma is detailed in section~\ref{sec:Proof_thm_Gaussian}. We focus now on its proof mixing control of exponential moments of a quadratic form of a Gaussian random variable with basic inequalities like Jensen, Fubini, and the important link between $\|f_0-P_t f_0\|_2^2$ and $\|f_0-P_t Y\|_2^2$. Note that this link is obvious in the case of orthogonal projections and need to be established differently in the general case, leading to technicalities (the introduction of $\gamma$).
\subsection{Proof of Lemma~\ref{lem:Dev_Gaussian}}
\begin{proof}
For the sake of clarity, for any $t, u \in \mathcal{T},$ let \[\Delta_{t,u}=\|f_0-\hat{f}_t\|^2_2-r_t-\|f_0-\hat{f}_u\|^2_2+r_u.\]
A simple calculation yields
\[
\Delta_{t,u}=2\left(W^\top(P_t-P_u)W +W^\top(P_t-P_u)f_0-\sigma^2 tr(P_t-P_u) \right).
\]
Since $(P_t)_{t \in \mathcal{T}}$ are positive semi-definite matrices, $W^\top(P_t-P_u)W \leq W^\top P_t W,$ and there exist an orthogonal matrix $U$ and a diagonal matrix $D$ such that $P_t=U^\top D U.$
For any $\beta>0,$
\[
\mathbb{E}\left[\exp \frac{\Delta_{t,u}}{\beta}\right]
\leq
\mathbb{E}\left[\exp \frac{2}{\beta}\left( (UW)^\top D(UW)
+(UW)^\top U(P_t-P_u)f-\sigma^2 tr(P_t-P_u)\right) \right] .
\]
Following lemma 2.4 of~\citet{MR2994877}, if $\beta>4 \sigma^2,$
\begin{align}
\label{eq:Delta_t,u}
\mathbb{E}\left[\exp \frac{\Delta_{t,u}}{\beta}\right]
\leq
\exp \frac{2\sigma^2}{\beta}\left(tr(P_u)+\frac{2\sigma^2 tr(P_t)+ \|(P_t-P_u)f_0\|_2^2 }{\beta-4\sigma^2}\right).
\end{align}
Note that \[
\|(P_t-P_u)f_0\|_2^2 \leq 2 \left(\|f_0-P_t f_0\|_2^2+\|f_0-P_u f_0\|_2^2 \right)
\leq 2 \left(\|f_0-P_t Y\|_2^2+\|f_0-P_u Y\|_2^2 \right)
\] and
\[
\|(P_t-P_u)f_0\|_2^2 \leq 2 \left(\|P_t f_0\|_2^2+\|P_u f_0\|_2^2 \right)
\leq 2 \widetilde{\|f_0\|_\infty^2} \left( tr(P_t)+tr(P_u)\right).
\]
Thus, for any $\beta>4\sigma^2,$ for any $\delta \in [0,1],$
\begin{multline*}
\mathbb{E} \exp \left[\frac{\Delta_{t,u}}{\beta}
- \frac{2\sigma^2}{\beta}\left(tr(P_u)+\frac{2\sigma^2 tr(P_t)}{\beta -4\sigma^2} \right)
-\frac{4\sigma^2 \delta}{\beta (\beta -4\sigma^2)}
\left(\|f_0-\hat{f}_t\|_2^2+\|f_0-\hat{f}_u\|_2^2 \right)
\right. \\
\left.
-\frac{4\sigma^2}{\beta (\beta -4\sigma^2)}
(1-\delta) \widetilde{\|f_0\|_\infty^2} \left( tr(P_t)+tr(P_u)\right)
\right]\leq 1.
\end{multline*}
Along the same lines as~\citet{MR2786484},
we first integrate according to the prior $\pi$ and use Fubini's theorem,
\begin{multline*}
\mathbb{E}\int \int \exp \frac{1}{\beta} \left[\Delta_{t,u}
- 2\sigma^2\left(tr(P_u)+\frac{2\sigma^2 tr(P_t)}{\beta -4\sigma^2} \right)
-\frac{4\sigma^2 \delta}{\beta -4\sigma^2}
\left(\|f_0-\hat{f}_t\|_2^2+\|f_0-\hat{f}_u\|_2^2 \right)
\right.\\
\left.
-\frac{4\sigma^2}{\beta -4\sigma^2}
(1-\delta) \widetilde{\|f_0\|_\infty^2} \left( tr(P_t)+tr(P_u)\right)
\right]d\pi(t) d\pi(u) \leq 1,
\end{multline*}
then introduce the probability distributions $\rho$ and $\mu$, and $\eta>0$
\begin{multline*}
\mathbb{E}\int \int \exp \frac{1}{\beta} \left[\Delta_{t,u}
- 2\sigma^2\left(tr(P_u)+\frac{2\sigma^2 tr(P_t)}{\beta-4\sigma^2} \right)
-\frac{4\sigma^2 \delta}{\beta-4\sigma^2}
\left(\|f_0-\hat{f}_t\|_2^2+\|f_0-\hat{f}_u\|_2^2 \right)
\right.\\
\left.
-\frac{4\sigma^2 (1-\delta)}{\beta-4\sigma^2}
\widetilde{\|f_0\|_\infty^2} \left( tr(P_t)+tr(P_u)\right)
-\beta \left(\ln \frac{d\rho}{d\pi}(t)+\ln\frac{d\mu}{d\pi}(u)+\ln\frac{1}{\eta} \right)
\right]d\rho(t) d\mu(u) \leq \eta,
\end{multline*}
before applying Jensen's inequality
\begin{multline}
\label{findeviation}
\mathbb{E} \exp \frac{1}{\beta} \left[\int \int \Delta_{t,u} d\rho(t)d\mu(u)
-\frac{4\delta\sigma^2}{\beta -4\sigma^2} \left( \int \|f_0-\hat{f}_t\|_2^2 d\rho(t) + \int \|f_0-\hat{f}_u\|_2^2 d\mu(u)\right)
\right.\\ \left.
-\frac{4\sigma^2}{\beta -4\sigma^2} \left(\sigma^2+(1-\delta) \widetilde{\|f_0\|_\infty^2} \right) \int tr(P_t) d\rho(t)
-\beta \left( KL(\rho,\pi)+ KL(\mu,\pi)+\ln\frac{1}{\eta} \right)
\right.\\
\left.
- 2\sigma^2 \left(1+\frac{2(1-\delta) \widetilde{\|f_0\|_\infty^2}}{\beta-4\sigma^2} \right) \int tr(P_u) d\mu(u)
\right] \leq \eta.
\end{multline}
Finally, using the basic inequality $\exp(x)\geq \mathds{1}_{\mathbb{R}_+}(x),$
\begin{multline*}
\mathbb{P}\left[ \int \int \Delta_{t,u} d\rho(t)d\mu(u)
\leq
\frac{4\delta\sigma^2}{\beta -4\sigma^2} \left( \int \|f_0-\hat{f}_t\|_2^2 d\rho(t) + \int \|f_0-\hat{f}_u\|_2^2 d\mu(u)\right)
\right.\\ \left.
+\frac{4\sigma^2}{\beta-4\sigma^2} \left(\sigma^2+(1-\delta) \widetilde{\|f_0\|_\infty^2} \right) \int tr(P_t) d\rho(t)
+\beta \left( KL(\rho,\pi)+KL(\mu,\pi)+ \ln\frac{1}{\eta}\right)
\right.\\ \left.
+ \frac{2\sigma^2}{n} \left(1+\frac{2(1-\delta)n \widetilde{\|f_0\|_\infty^2}}{\beta n-4\sigma^2} \right) \int tr(P_u) d\mu(u)
\right]\geq 1-\eta.
\end{multline*}
The result in expectation is obtained by Equation~\eqref{findeviation} with $\eta=1$:
\begin{multline*}
\mathbb{E} \exp \frac{1}{\beta} \left[\int \int \Delta_{t,u} d\rho(t)d\mu(u)
-\frac{4\delta\sigma^2}{\beta-4\sigma^2} \left( \int \|f_0-\hat{f}_t\|_2^2 d\rho(t) + \int \|f_0-\hat{f}_u\|_2^2 d\mu(u)\right)
\right.\\ \left.
-\frac{4\sigma^2}{\beta-4\sigma^2} \left(\sigma^2+(1-\delta) \widetilde{\|f_0\|_\infty^2} \right) \int tr(P_t) d\rho(t)
-\beta \left( KL(\rho,\pi)+ KL(\mu,\pi)\right)
\right.\\
\left.
- 2\sigma^2 \left(1+\frac{2(1-\delta)\widetilde{\|f_0\|_\infty^2}}{\beta -4\sigma^2} \right) \int tr(P_u) d\mu(u)
\right] \leq 1,
\end{multline*}
combined with the inequality $t \leq \exp(t)-1$.
\end{proof}
\subsection{Proof of Theorem~\ref{thm:Gaussian}}
\label{sec:Proof_thm_Gaussian}
We follow the scheme of proof given in section~\ref{sec:Proof_oracle} and use Lemma~\ref{lem:Dev_Gaussian}, leading to the following result:
for any $\eta>0,$ any prior probability distribution $\pi,$ any $\delta \in [0,1]$ and any $\beta>4\sigma^2 (1+\delta),$ with probability at least $1-\eta,$ for any probability distribution $\mu,$
\begin{multline*}
\int \|f_0-\hat{f}_t\|_2^2 d\rho(t)-\int \|f_0-\hat{f}_t\|_2^2 d\mu(t)\\
\leq
\frac{4\delta\sigma^2}{\beta-4\sigma^2} \left( \int \|f_0-\hat{f}_t\|_2^2 d\rho(t) + \int \|f_0-\hat{f}_u\|_2^2 d\mu(u)\right)
\\
+\frac{4\sigma^2}{\beta-4\sigma^2} \left(\sigma^2+(1-\delta) \widetilde{\|f_0\|_\infty^2} \right) \int tr(P_t) d\rho(t)-\int \mathop{\mathrm{pen}}(t) d\rho(t)
\\
+ 2\sigma^2 \left(1+\frac{2(1-\delta) \widetilde{\|f_0\|_\infty^2}}{\beta -4\sigma^2} \right) \int tr(P_t) d\mu(t) +\int \mathop{\mathrm{pen}}(t) d\mu(t)
\\
+\beta \left(2 KL(\mu,\pi)+ \ln\frac{1}{\eta} \right).
\end{multline*}
With $\mathop{\mathrm{pen}}(t)\geq\frac{4\sigma^2}{\beta -4\sigma^2} \left(\sigma^2+(1-\delta) \widetilde{\|f_0\|_\infty^2} \right) tr(P_t),$ the previous inequality becomes
\begin{align*}
&\left(1-\frac{4\delta\sigma^2}{\beta -4\sigma^2}\right) \int \|f_0-\hat{f}_t\|_2^2 d\rho(t)-\left(1+\frac{4\delta\sigma^2}{\beta -4\sigma^2}\right) \int \|f_0-\hat{f}_t\|_2^2 d\mu(t)\\
&\leq
2\sigma^2 \left(1+\frac{2(1-\delta)\widetilde{\|f_0\|_\infty^2}}{\beta-4\sigma^2} \right) \int tr(P_t) d\mu(t) +\int \mathop{\mathrm{pen}}(t) d\mu(t)
+\beta \left(2 KL(\mu,\pi)+ \ln\frac{1}{\eta} \right).
\end{align*}
Furthermore, using \[\|f_0-f_{EWA}\|_2^2 \leq \int \|f_0-\hat{f}_t\|_2^2 d\rho(t), \] if $\beta>4\sigma^2(\delta+1),$ we obtain
\begin{multline*
\|f_0-f_{EWA}\|_2^2 \leq \inf_{\mu \in \mathcal{M}_+^1(\mathcal{T})} \left(1+\frac{8\sigma^2\delta}{\beta -4\sigma^2(\delta+1)}\right) \int \|f_0-\hat{f}_t\|_2^2 d\mu(t)\\
+\left(1+\frac{4\sigma^2\delta}{\beta -4\sigma^2(\delta+1)}\right)
2\sigma^2 \left(1+\frac{2(1-\delta)\widetilde{\|f_0\|_\infty^2}}{\beta-4\sigma^2} \right) \int tr(P_t) d\mu(t) \\
+\left(1+\frac{4\sigma^2\delta}{\beta -4\sigma^2(\delta+1)}\right) \int \mathop{\mathrm{pen}}(t)d\mu(t)
+\beta \left(1+\frac{4\sigma^2\delta}{\beta -4\sigma^2(\delta+1)}\right)\left(2KL(\mu,\pi)+\ln\frac{1}{\eta} \right).
\end{multline*}
In addition, taking $\epsilon= \frac{4\sigma^2\delta}{\beta -4\sigma^2(\delta+1)},$ gives
\begin{multline*}
\|f_0-f_{EWA}\|_2^2 \leq \inf_{\mu \in \mathcal{M}_+^1(\mathcal{T})} \left(1+2\epsilon\right) \int \|f_0-\hat{f}_t\|_2^2 d\mu(t)\\
+ 2\sigma^2 \left(1+\epsilon\right) \left(1+\frac{2(1-\delta)\widetilde{\|f_0\|_\infty^2}}{\beta -4\sigma^2} \right) \int tr(P_t) d\mu(t)\\
+\left(1+\epsilon\right)\left(\int \mathop{\mathrm{pen}}(t)d\mu(t)+2\beta KL(\mu,\pi)+\beta\ln\frac{1}{\eta} \right).
\end{multline*}
\section{Appendix : Proofs in the sub-Gaussian case}
\label{sec:Appendix_Proofs_subGaussian}
\subsection{Proof of Theorem~\ref{thm:general_PAC}}
The proof follows from the scheme described in section~\ref{sec:Proof_oracle}. The main point is still to control \[\int \left(\|f_0-\hat{f}_t\|_2^2-r_t \right) d\rho(t)-\int \left(\|f_0-\hat{f}_t\|_2^2-r_t \right) d\mu(t). \] We recall that $P_t$ is a symmetric positive semi-definite matrix, there exists $V>0$ such that $\sup_{t \in \mathcal{T}} \|P_t\|_2 \leq V$ and $W$ is a centered sub-Gaussian noise.
For any $t, u \in \mathcal{T},$ we still denote $\Delta_{t,u}=\|f_0-\hat{f}_t\|^2_2-r_t-\|f_0-\hat{f}_u\|^2_2+r_u.$
\begin{lem}
\label{lem:Dev_General}
Let $\pi$ be an arbitrary prior probability. For any $\delta \in [0,1]$, any $\beta >4\sigma^2 V$ and $\beta \geq 4\sigma^2 V (1+4\delta)$, let
\[ \gamma= \frac{1}{16\sigma^2 \delta V^2} \left(\beta -4\sigma^2 V (1+2\delta)-\sqrt{\beta -4\sigma^2 V}\sqrt{\beta -4\sigma^2 V(1+4\delta)} \right) \mathds{1}_{\delta>0}. \]
Then, for any probability distributions $\rho$ and $\mu$, for any $\nu>0$,
\begin{itemize}
\item for any $\eta \in (0,1]$, with probability at least $1-\eta$,
\begin{align*}
&\int \int \Delta_{t,u} d\rho(t) d\mu(u)
\leq
(1+\nu)\gamma \int \|P_t Y-f_0\|_2^2 d\rho(t)\\
&+\frac{4\sigma^2}{\beta-4\sigma^2 V} \left(\sigma^2+(1-\delta)(1+2\gamma V)^2 \widetilde{\|f_0\|_\infty^2} \right) \int tr(P_t^2) d\rho(t)\\
&+2\sigma^2 \left( \int tr(P_u) d\mu(u) +\frac{2(1-\delta)(1+2\gamma V)^2}{\beta -4 \sigma^2 V} \widetilde{\|f_0\|_\infty^2} \int tr(P_u^2) d\mu(u) \right)\\
&+\left(1+\frac{1}{\nu}\right)\gamma \int \|P_u Y-f_0\|_2^2 d\mu(u)
+\beta \left( KL(\rho,\pi)+KL(\mu,\pi)+\ln\frac{1}{\eta}\right)
\end{align*}
\item Moreover,
\begin{align*}
&\mathbb{E} \left[ \int \int \Delta_{t,u} d\rho(t) d\mu(u) \right]
\leq \mathbb{E}\Big[
(1+\nu)\gamma \int \|P_t Y-f_0\|_2^2 d\rho(t)\\
&+\frac{4\sigma^2}{\beta-4\sigma^2 V} \left(\sigma^2+(1-\delta)(1+2\gamma V)^2 \widetilde{\|f_0\|_\infty^2} \right) \int tr(P_t^2) d\rho(t)\\
&+2\sigma^2 \left( \int tr(P_u) d\mu(u) +\frac{2(1-\delta)(1+2\gamma V)^2}{\beta -4 \sigma^2 V} \widetilde{\|f_0\|_\infty^2} \int tr(P_u^2) d\mu(u) \right)\\
&+\left(1+\frac{1}{\nu}\right)\gamma \int \|P_u Y-f_0\|_2^2 d\mu(u)
+\beta \left( KL(\rho,\pi)+KL(\mu,\pi) \right)\Big]
\end{align*}
\end{itemize}
\end{lem}
Under the assumptions of the previous lemma, with probability at least $1-\eta$,
\begin{multline*}
\int \|f_0-\hat{f}_t\|_2^2 d\rho(t) -\int \|f_0-\hat{f}_t\|_2^2 d\mu(t)
\leq
(1+\nu)\gamma \int \|\hat{f}_t-f_0\|_2^2 d\rho(t)\\
+\frac{4\sigma^2}{\beta-4\sigma^2 V} \left(\sigma^2+(1-\delta)(1+2\gamma V)^2 \widetilde{\|f_0\|_\infty^2} \right) \int tr(P_t^2) d\rho(t)-\int \mathop{\mathrm{pen}}(t) d\rho(t)\\
+2\sigma^2 \left( \int tr(P_t) d\mu(t) +\frac{2(1-\delta)(1+2\gamma V)^2}{\beta -4 \sigma^2 V} \widetilde{\|f_0\|_\infty^2} \int tr(P_t^2) d\mu(t) \right)+\int \mathop{\mathrm{pen}}(t)d\mu(t)\\
+\left(1+\frac{1}{\nu}\right)\gamma \int \|\hat{f}_t-f_0\|_2^2 d\mu(t)
+\beta \left(2 KL(\mu,\pi)+\ln\frac{1}{\eta}\right).
\end{multline*}
Taking $\mathop{\mathrm{pen}}(t) \geq \frac{4\sigma^2}{\beta-4\sigma^2 V} \left(\sigma^2+(1-\delta)(1+2\gamma V)^2 \widetilde{\|f_0\|_\infty^2} \right) tr(P_t^2)$ and $\nu \in N=\{\nu>0|(1+\nu)\gamma<1\}$, such that the inequality stays informative,
\begin{multline*}
\left(1-(1+\nu)\gamma\right) \int \|f_0-\hat{f}_t\|_2^2 d\rho(t)
\leq \left(1+\left(1+\frac{1}{\nu}\right)\gamma \right) \int \|f_0-\hat{f}_t\|_2^2 d\mu(t) \\
+2\sigma^2 \left( \int tr(P_t) d\mu(t) +\frac{2(1-\delta)(1+2\gamma V)^2}{\beta -4 \sigma^2 V} \widetilde{\|f_0\|_\infty^2} \int tr(P_t^2) d\mu(t) \right)\\
+\int \mathop{\mathrm{pen}}(t)d\mu(t)+\beta \left(2 KL(\mu,\pi)+\ln\frac{1}{\eta}\right).
\end{multline*}
Finally, since $\|f_0-f_{EWA}\|_2^2 \leq \int \|f_0-\hat{f}_t\|_2^2 d\rho(t)$,
\begin{multline*}
\|f_0-f_{EWA}\|_2^2
\leq \left(1+\frac{(1+\nu)^2\gamma}{\nu(1-(1+\nu)\gamma)} \right) \int \|f_0-\hat{f}_t\|_2^2 d\mu(t) \\
+\frac{2\sigma^2}{1-(1+\nu)\gamma} \left( \int tr(P_t) d\mu(t) +\frac{2(1-\delta)(1+2\gamma V)^2}{\beta -4 \sigma^2 V} \widetilde{\|f_0\|_\infty^2} \int tr(P_t^2) d\mu(t) \right)\\
+\frac{1}{1-(1+\nu)\gamma} \left(\int \mathop{\mathrm{pen}}(t)d\mu(t)+\beta \left(2 KL(\mu,\pi)+\ln\frac{1}{\eta}\right) \right).
\end{multline*}
The result in expectation is obtained in the same fashion.
\subsection{Proof of Lemma~\ref{lem:Dev_General}}
The exponential moment of $\Delta_{t,u}$ is easily controlled by a term involving $\|P_tf_0-f_0\|_2^2$ (see Equation~\eqref{eq:Delta_t,u}). Since $P_t$ are not projections, $\|P_tf_0-f_0\|_2^2 \leq \|P_t Y-f_0\|_2^2$ does not hold any more. The presence of $\|P_t Y-f_0\|_2^2$ allows us to obtain a weak oracle inequality. To overcome this difficulty, $\|(P_t-P_u)Y\|_2^2$ is introduced and for an arbitrary $\gamma \geq 0$, we try to control $\Delta_{t,u}-\gamma \|(P_t-P_u)Y\|_2^2$.
\begin{proof}
A simple calculation yields
\begin{multline*}
\Delta_{t,u}-\gamma \|(P_t-P_u)Y\|_2^2=
W^\top (2I-\gamma (P_t-P_u)^\top)(P_t-P_u)W \\
+ 2 W^\top (I-\gamma (P_t-P_u)^\top)(P_t-P_u)f_0- 2\sigma^2 tr(P_t-P_u)-\gamma \|(P_t-P_u)f_0\|_2^2 .
\end{multline*}
Noting that $W^\top (2I-\gamma (P_t-P_u)^\top)(P_t-P_u)W \leq 2 W^\top (P_t-P_u)W$ and since $(P_t)_{t \in \mathcal{T}}$ are positive semi-definite matrices, $2 W^\top(P_t-P_u)W \leq 2 W^\top P_t W$. Thus, for any $\beta>0,$ any $\gamma \geq 0$,
\begin{multline*}
\mathbb{E}\exp \left(\frac{\Delta_{t,u}}{\beta}-\frac{\gamma}{\beta} \|(P_t-P_u)Y\|_2^2\right)\\
\leq
\mathbb{E}\left[\exp \frac{2}{\beta}\left( W^\top P_t W +W^\top (I-\gamma (P_t-P_u))(P_t-P_u)f_0 \right)\right]\\
\times \exp \frac{-1}{\beta}\left( 2 \sigma^2 tr(P_t-P_u)+ \gamma \|(P_t-P_u)f_0\|_2^2 \right) .
\end{multline*}
The first step is to bring us back to the Gaussian case, using $W$'s sub-Gaussianity and an idea of \citet{MR2994877}. Let $Z$ be a standard Gaussian random variable, independent of $W$. Then,
\begin{align*}
&\mathbb{E} \exp \left( \frac{2}{\sqrt{\beta}} W^\top \sqrt{P_t} Z+\frac{2}{\beta} W^\top (I-\gamma (P_t-P_u))(P_t-P_u)f_0 \right) \\
& \qquad= \mathbb{E} \left[ \mathbb{E} \left[ \exp \left( \frac{2}{\sqrt{\beta}} W^\top \sqrt{P_t} Z+\frac{2}{\beta} W^\top (I-\gamma (P_t-P_u))(P_t-P_u)f_0 \right) \big|W \right] \right]\\
& \qquad=\mathbb{E} \left[ \mathbb{E} \left[ \exp \left( \frac{2}{\sqrt{\beta}} W^\top \sqrt{P_t} Z \right) \big|W \right] \exp \left(\frac{2}{\beta} W^\top (I-\gamma (P_t-P_u))(P_t-P_u)f_0 \right) \right]\\
& \qquad=\mathbb{E} \exp \frac{2}{\beta} \left(W^\top P_t W+ W^\top (I-\gamma (P_t-P_u))(P_t-P_u)f_0 \right).
\end{align*}
On the other hand,
\begin{multline*}
\mathbb{E}\left[\exp \frac{2}{\beta}\left( W^\top P_t W +W^\top (I-\gamma (P_t-P_u))(P_t-P_u)f_0 \right)\right]\\
= \mathbb{E} \left[ \mathbb{E} \left[ \exp \left( \frac{2}{\sqrt{\beta}} W^\top \sqrt{P_t} Z+\frac{2}{\beta} W^\top (I-\gamma (P_t-P_u))(P_t-P_u)f_0 \right) \big|Z \right] \right].
\end{multline*}
Since $W$ is sub-Gaussian with parameter $\sigma$,
\begin{multline*}
\mathbb{E}\left[\exp \frac{2}{\beta}\left( W^\top P_t W +W^\top (I-\gamma (P_t-P_u))(P_t-P_u)f_0 \right)\right]\\
\leq \mathbb{E} \exp \left( \frac{\sigma^2}{2} \left \| \frac{2}{\sqrt{\beta}} \left(\sqrt{P_t}Z +\frac{1}{\sqrt{\beta}} (I-\gamma (P_t-P_u))(P_t-P_u)f_0 \right)\right \|_2^2 \right)
\end{multline*}
Hence,
\begin{multline*}
\mathbb{E}\exp \left(\frac{\Delta_{t,u}}{\beta}-\frac{\gamma}{\beta} \|(P_t-P_u)Y\|_2^2\right)\\
\leq
\mathbb{E}\left[\exp \frac{2\sigma^2}{\beta}\left( Z^\top P_t Z+\frac{2}{\sqrt{\beta}} Z^\top \sqrt{P_t} (I-\gamma (P_t-P_u))(P_t-P_u)f_0 \right)\right]\\
\times \exp \left( \frac{2\sigma^2}{\beta^2} \left \|(I-\gamma (P_t-P_u))(P_t-P_u)f_0 \right \|_2^2-\frac{2 \sigma^2}{\beta} tr(P_t-P_u)-\frac{\gamma}{\beta} \|(P_t-P_u)f_0\|_2^2 \right) .
\end{multline*}
The expectation is similar to the one obtained in the Gaussian case: the exponential of some quadratic form. The same recipe is applied. Since $P_t$ is positive semi-definite, there exist an orthogonal matrix $U$ and a diagonal matrix $D$ such that $P_t=U^\top D U.$ Note that $UZ$ is a standard Gaussian variable.
This diagonalization step and the non-negativity of the eigenvalues allow to apply Lemma 2.4 of~\citet{MR2994877}. Then, for any $\beta > 4\sigma^2 V$, any $\gamma \geq 0$,
\begin{multline*}
\mathbb{E}\exp \left(\frac{\Delta_{t,u}}{\beta}-\frac{\gamma}{\beta} \|(P_t-P_u)Y\|_2^2\right)\\
\leq
\exp \frac{2\sigma^2}{\beta}\left( tr(P_t)+\frac{2\sigma^2}{\beta (\beta-4\sigma^2 V)} \left(\beta tr(P_t^2)+2 \left \| \sqrt{P_t} (I-\gamma (P_t-P_u))(P_t-P_u)f_0 \right \|_2^2 \right) \right)\\
\times \exp \left( \frac{2\sigma^2}{\beta^2 } \left \|(I-\gamma (P_t-P_u))(P_t-P_u)f_0 \right \|_2^2-\frac{2 \sigma^2}{\beta} tr(P_t-P_u)-\frac{\gamma}{\beta} \|(P_t-P_u)f_0\|_2^2 \right) .
\end{multline*}
Consequently,
\begin{align*}
&\mathbb{E}\exp \left(\frac{\Delta_{t,u}}{\beta}+\frac{\gamma}{\beta} \left(\|(P_t-P_u)f_0\|_2^2-\|(P_t-P_u)Y\|_2^2 \right)\right)\\
&\leq
\exp \frac{2\sigma^2}{\beta}\left( tr(P_u)+\frac{2\sigma^2}{\beta-4\sigma^2 V} tr(P_t^2) \right)\\
&\quad \times \exp \left( \frac{2\sigma^2}{\beta^2} \left(\frac{4\sigma^2 V}{\beta-4 \sigma^2 V}(1+2\gamma V)^2+(1+2\gamma V)^2 \right) \|(P_t-P_u)f_0\|_2^2\right).\\
&\leq \exp \frac{2\sigma^2}{\beta}\left( tr(P_u)+\frac{2\sigma^2}{\beta-4\sigma^2 V} tr(P_t^2) +\frac{(1+2\gamma V)^2}{\beta -4 \sigma^2 V} \|(P_t-P_u)f_0\|_2^2\right) .
\end{align*}
If an exact oracle inequality is wished, $\|(P_t-P_u)f_0\|_2^2$ should be upper bounded by some constant and $\gamma$ should be set to zero. Else, $\gamma$ is used to \emph{replace} the terms in $\|(P_t-P_u)f_0\|_2^2$ by $\|(P_t-P_u)Y\|_2^2$.
Thus, the terms depending on $f_0$ will be upper bounded in two ways: \begin{itemize}
\item on the one hand, using $\widetilde{\|f_0\|_\infty^2}$
\begin{align*}
\|(P_t-P_u)f_0\|_2^2
\leq 2 \left(\|P_t f_0\|_2^2+\|P_u f_0\|_2^2 \right)
&\leq 2 \left( tr(P_t^2)+ tr(P_u^2) \right) \widetilde{\|f_0\|_\infty^2}
\end{align*}
For any $\delta \in [0,1]$,
\begin{align*}
&\mathbb{E}\exp \left(\frac{\Delta_{t,u}}{\beta}+\frac{\gamma}{\beta} \left(\|(P_t-P_u)f_0\|_2^2-\|(P_t-P_u)Y\|_2^2 \right)\right)\\
&\leq \exp \frac{2\sigma^2}{\beta}\left( tr(P_u)+\frac{2\sigma^2}{\beta-4\sigma^2 V} tr(P_t^2) +\frac{ (1+2\gamma V)^2 (1-\delta)}{\beta -4 \sigma^2 V} \|(P_t-P_u)f_0\|_2^2\right)\\
&\times \exp \left(\frac{2\sigma^2 (1+2\gamma V)^2 \delta}{\beta(\beta -4 \sigma^2 V)} \|(P_t-P_u)f_0\|_n^2 \right)\\
&\leq \exp \frac{2\sigma^2}{\beta}\left( tr(P_u)+\frac{2\sigma^2}{\beta-4\sigma^2 V} tr(P_t^2)+\frac{ (1+2\gamma V)^2 \delta}{\beta -4 \sigma^2 V} \|(P_t-P_u)f_0\|_2^2 \right)\\
&\times \exp\left(\frac{4\sigma^2 (1+2\gamma V)^2 (1-\delta)}{\beta(\beta-4 \sigma^2 V)} \left(tr(P_t^2)+tr(P_u^2)\right) \widetilde{\|f_0\|_\infty^2}\right).
\end{align*}
\item on the other hand, introducing $\|P_t Y-f_0\|_2^2$ to obtain a weak oracle inequality: conditions should be found on $\gamma$ such that
\begin{multline*}
\frac{2\sigma^2 (1+2\gamma V)^2 \delta}{\beta-4 \sigma^2 V} \|(P_t-P_u)f_0\|_2^2 -\gamma \left(\|(P_t-P_u)f_0\|_2^2-\|(P_t-P_u)Y\|_2^2\right)\\
\leq C_1 \|P_t Y-f_0\|_2^2+ C_2 \|P_u Y-f_0\|_2^2
\end{multline*}
for some non-negative constants $C_1$ and $C_2$ and with $\delta>0$.
Since for any $\nu>0$, $\|(P_t-P_u)Y\|_2^2 \leq (1+\nu) \|P_t Y-f_0\|_2^2+ \left(1+\frac{1}{\nu}\right) \|P_u Y-f_0\|_2^2$, it suffices that \[
\frac{2\sigma^2 (1+2\gamma V)^2 \delta}{\beta -4 \sigma^2 V} \|(P_t-P_u)f_0\|_2^2 -\gamma \|(P_t-P_u)f_0\|_2^2 \leq 0.
\]
This condition may be fulfilled if $\beta \geq 4\sigma^2 V(1+4\delta)$. The smallest $\gamma\geq 0$ among all the possible ones is chosen~:
\[ \gamma= \frac{1}{16\sigma^2 \delta V^2} \left(\beta -4\sigma^2 V (1+2\delta)-\sqrt{\beta -4\sigma^2 V}\sqrt{\beta -4\sigma^2 V(1+4\delta)} \right) \mathds{1}_{\delta>0}. \]
\end{itemize}
This leads to the following inequality~: for any $\delta \in [0,1]$,
for any $\beta >4 \sigma^2 V$ and $\beta \geq 4 \sigma^2 V (1+4\delta)$, with $\gamma$ previously defined, for any $\nu>0$,
\begin{multline*}
\mathbb{E}\exp \left(\frac{\Delta_{t,u}}{\beta}
-\frac{\gamma}{\beta}\left((1+\nu)\|P_t Y-f_0\|_2^2 +\left(1+\frac{1}{\nu}\right)\|P_u Y-f_0\|_2^2 \right)\right)\\
\leq \exp \frac{2\sigma^2}{\beta}\left( tr(P_u)+\frac{2\sigma^2}{\beta-4\sigma^2 V} tr(P_t^2) +\frac{2 (1+2 \gamma V)^2 (1-\delta)}{\beta -4 \sigma^2 V} \left(tr(P_t^2)+tr(P_u^2)\right) \widetilde{\|f_0\|_\infty^2}\right).
\end{multline*}
The rest of the proof follows the same steps as in the Gaussian case: we first integrate according to the prior $\pi$, use Fubini's theorem, introduce the probability measures $\rho$ and $\mu$ and apply Jensen's inequality to obtain that for any $\eta \in (0,1]$,
\begin{multline}
\label{eq:Delta_t,u_General}
\mathbb{E} \exp \frac{1}{\beta} \Big[ \int \int \Delta_{t,u} d\rho(t) d\mu(u)
-(1+\nu)\gamma \int \|P_t Y-f_0\|_2^2 d\rho(t)\\
-\frac{4\sigma^2}{\beta-4\sigma^2 V} \left(\sigma^2+(1-\delta)(1+2\gamma V)^2 \widetilde{\|f_0\|_\infty^2} \right) \int tr(P_t^2) d\rho(t)\\
-2\sigma^2 \left( \int tr(P_u) d\mu(u) +\frac{2(1-\delta)(1+2 \gamma V)^2}{\beta -4 \sigma^2 V} \widetilde{\|f_0\|_\infty^2} \int tr(P_u^2) d\mu(u) \right)\\
-\left(1+\frac{1}{\nu}\right)\gamma \int \|P_u Y-f_0\|_2^2 d\mu(u)\\
-\beta \left( KL(\rho,\pi)+KL(\mu,\pi)+\ln\frac{1}{\eta}\right)
\Big] \leq \eta.
\end{multline}
Finally, using $\exp(x)\geq \mathds{1}_{\mathbb{R}_+}(x),$ for any $\delta \in [0,1]$, any $\beta>4\sigma^2 V$ and $\beta \geq 4\sigma^2 V (1+4\delta)$, with $\gamma$ previously defined,
for any $\eta \in (0,1]$, for any $\nu >0$,
\begin{align*}
&\mathbb{P}\Big[
\int \int \Delta_{t,u} d\rho(t) d\mu(u)
\leq
(1+\nu)\gamma \int \|P_t Y-f_0\|_2^2 d\rho(t)\\
&+\frac{4\sigma^2}{\beta-4\sigma^2 V} \left(\sigma^2+(1-\delta)(1+2\gamma V)^2 \widetilde{\|f_0\|_\infty^2} \right) \int tr(P_t^2) d\rho(t)\\
&+2\sigma^2 \left( \int tr(P_u) d\mu(u) +\frac{2(1-\delta)(1+2 \gamma V)^2}{\beta -4 \sigma^2 V} \widetilde{\|f_0\|_\infty^2} \int tr(P_u^2) d\mu(u) \right)\\
&+\left(1+\frac{1}{\nu}\right)\gamma \int \|P_u Y-f_0\|_2^2 d\mu(u)
+\beta \left( KL(\rho,\pi)+KL(\mu,\pi)+\ln\frac{1}{\eta}\right)
\Big] \geq 1-\eta.
\end{align*}
The result in expectation comes from Equation~\eqref{eq:Delta_t,u_General} with $\eta=1$, combined with the inequality $t \leq \exp(t)-1$.
\end{proof}
\bibliographystyle{plainnat}
|
\section{Introduction}
In 1994 Maxim Kontsevich interpreted a duality coming from physics in a
consistent, powerful mathematical framework called Homological Mirror
Symmetry (HMS).
HMS is now the foundation of a wide range of contemporary mathematical
research.
Numerous works by many authors have demonstrated the interaction of
mirror symmetry and HMS with a wide range of new and subtle mathematical
structures.
One of this structures is the moduli space of stability conditions.
The study of stability in triangulated categories was initiated by M.
Douglas
and mathematically by T. Bridgeland.
The majority of the activity since then
has focused on categories of algebro-geometric origin.
Signicant work in this direction is due to T. Bridgeland, A. King, E. Macr\'i, S. Okada, Y. Toda, A.
Bayer, J. Woolf, J. Collins, A. Polishchuck et. al.
In previous works \cite{DK1}, \cite[joint with F. Haiden and M. Kontsevich]{DHKK} we developed results and ideas by T. Bridgeland \cite{Bridg1}, A. King \cite{King}, E. Macr\'i \cite{Macri}, J. Collins and A. Polishchuck \cite{CP}.
Recently in \cite{Woolf} J. Woolf showed classes of categories with contractible component in the space of stability conditions. His paper generalizes and unifies various known results for stability spaces of specific categories, and settles some conjectures about the stability spaces associated to Dynkin quivers, and to their Calabi-Yau-N Ginzburg algebras. However the results in \cite{Woolf} do not cover tame representation type quivers, these quivers are beyond the scope of \cite{Woolf}.
In the present paper we give a new example of a tame representation type quiver with contractible space of stability conditions. This paper is a natural consequence of our previous paper \cite{DK1}. Both are based on ideas of E. Macr\'i, which he gave in \cite{Macri} studying $ \st(D^b(K(l))$, where $K(l)$ is the $l$-Kronecker quiver.
\vspace{3mm}
1.1. T. Bridgeland defined in \cite{Bridg1} the space of stability conditions on a triangulated category $\mc T$, denoted by $\st(\mc T)$, and proved that it is a a complex manifold on which act the groups $\widetilde{GL}^+(2,\RR)$ and ${\rm Aut}(\mc T)$. To any bounded t-structure of $\mc T$ he assigned a family of stability conditions.
E. Macri constructed in \cite{Macri} stability conditions using exceptional collections and the action of $\widetilde{GL}^+(2,\RR)$ on $\st(\mc T)$. Applying results in \cite{BBD}, he showed that the extension closure of a full Ext-exceptional collection\footnote{An exceptional collection $\mc E$ is said to be \textit{Ext-exceptional} if $\Hom^{\leq 0}(E_i,E_j)=0$ for $0\leq j<j\leq n$.} $\mc E=(E_0,E_1,\dots, E_n)$ in $\mc T$ is a bounded t-structure.
The stability conditions obtained from this t-structure together with their translations by the right action of $\widetilde{GL}^+(2,\RR)$ will be referred to as \textit{generated by $\mc E$}.
E. Macr\`i, studying $ \st(D^b(K(l))$ in \cite{Macri}, gave an idea for producing an exceptional pair generating a given stability condition $\sigma$ on $D^b(K(l))$, where $K(l)$ is the $l$-Kronecker quiver.
We defined in \cite{DK1} the notion of a \textit{$\sigma$-exceptional collection} (\cite[Definition 3.19]{DK1}), so that the full $\sigma$-exceptional collections are exactly the exceptional collections which generate $\sigma$, and we focused on constructing $\sigma$-exceptional collections from a given $\sigma \in \st(D^b(\mc A))$, where $\mc A$ is a hereditary, $\hom$-finite, abelian category. We
developed tools for constructing $\sigma$-exceptional collections of length at least three in $D^b(\mc A)$. These tools are based on
the notion of \textit{regularity-preserving hereditary category}, introduced in \cite{DK1} to avoid difficulties related to the Ext-nontrivial couples (couples of exceptional objects in $\mc A$ with ${\rm Ext}^1(X,Y)\neq 0$ and ${\rm Ext}^1(Y,X)\neq 0$).
After a detailed study of the exceptional objects of the affine quiver $Q$ (see figure \eqref{Q1} below) it was shown in \cite{DK1} that $Rep_k(Q)$ is regularity preserving and the newly obtained methods for constructing $\sigma$-triples were applied to the case $\mc A = Rep_k(Q)$. As a result we obtained the following theorem:
\begin{theorem}[\cite{DK1}] \label{main theorem for Q in intro} Let $k$ be an algebraically closed field.
For each $\sigma \in \st(D^b(Rep_k(Q)))$ there exists a full
$\sigma$-exceptional collection.
\end{theorem}
In other words, all stability conditions on $D^b(Q)$ are generated by exceptional collections (in this case exceptional triples). This theorem implies
that $\st(D^b(Q))$ is connected \cite[Corollary 10.2]{DK1}.
Using Theorem \ref{main theorem for Q in intro} and the data about the exceptional collections
given in \cite[Section 2]{DK1}, we prove here the following:
\begin{theorem} \label{main theo} Let $k$ be an algebraically closed field. Let $Q $ be the following quiver:
\be \label{Q1} Q = \begin{diagram}[1em]
& & \circ & & \\
& \ruTo & & \luTo & \\
\circ & \rTo & & & \circ
\end{diagram}. \ee The space of Bridgeland stability conditions $\st(D^b(Rep_k(Q))$ is a contractible (and connected) manifold, where $D^b(Rep_k(Q))$ is the derived category of representations of $Q$.
\end{theorem}
1.2. We give now more details about the structure of $\st(D^b(Rep_k(Q)))$ and about the proof of Theorem \ref{main theo}.
We call an exceptional pair $(E,F)$ in $D^b(Rep_k(Q))$ a 2-Kronecker pair if $\hom^{\leq 0}(E,F)=0$, and $\hom^{1}(E,F)=2$.
Recall that the Braid group on two strings $B_2\cong \ZZ$ acts on the set of equivalence classes of exceptional pairs in $\mc T$. \footnote{Here we take the equivalence $\sim$ explained in \textbf{Some notations} and it is clear when a given equivalence class w.r. $\sim$ will be called a 2-Kronecker pair} The set of equivalence classes of 2-Kronecker pairs is invariant under the action of $B_2$. In Subsection \ref{the two orbits} are described the orbits of this action on the 2-Kronecker pairs (using \cite[Corollary 2.9]{DK1}). There are two such orbits and
in terms of our notations they are $\{(a^m,a^{m+1}[-1])\}_{m\in \ZZ}$ and $\{(b^m,b^{m+1}[-1])\}_{m\in \ZZ}$ (see Remark \ref{T_1234}).
It turns out that the exceptional objects of
$D^b(Rep_k(Q))$ can be grouped as follows $\{a^m\}_{m\in \ZZ} \cup \{M,M'\}\cup \{b^m\}_{m\in \ZZ}$, where $\{M,M'\}\subset \ Rep_k(Q)$ is the unique Ext-nontrivial couple of $D^b(Rep_k(Q))$.
Let $\mk{T}_a^{st}$ and $\mk{T}_b^{st}$ be the stability conditions generated by the exceptional triples containing a subsequence of the from $(a^m[p],a^{m+1}[q])$ and $(b^m[p],b^{m+1}[q])$ for some $m,p,q \in \ZZ$, respectively. Using Theorem \ref{main theorem for Q in intro} we show in Section \ref{the union} that $\st(D^b(Rep_k(Q)))=\mk{T}_a^{st}\cup (\_,M,\_) \cup (\_,M',\_)\cup \mk{T}_b^{st}$, where $(\_,M,\_) \cup (\_,M',\_)$ denotes the set of stability conditions generated by triples of the form $(A,M[p],C)$ or $(A,M'[p],C)$ with $p\in \ZZ$ (these turn out to be the triples $(A,B,C)$ for which $\dim(\Hom^i(A,B))\leq 1 $, $\dim(\Hom^i(A,C))\leq 1 $, $\dim(\Hom^i(B,C))\leq 1 $ for all $i\in \ZZ$).
The main steps are as follows.
In Section \ref{mk T_12 cap mk T_43} we how that $\mk{T}_a^{st} \cap \mk{T}_b^{st}=\emptyset$. In Section \ref{T_a T_b cont} we show that $\mk{T}_a^{st}$ and $\mk{T}_b^{st}$ are contractible. In Section \ref{connecting} we connect $\mk{T}_a^{st}$ and $\mk{T}_b^{st}$ by $(\_,M,\_)\cup (\_,M',\_)$ and show that in this procedure the contractibility is preserved.
The theorem from topology which we use to glue stability conditions generated by different exceptional triples is the Seifert-van Kampen theorem, modified about contractile subsets in manifolds (see Remark \ref{VK}). In Section \ref{general remarks} are given several important tools, which we use throughout to analyze the intersection of the sets of stability conditions generated by different exceptional collections. These tools are extensions of results and ideas in \cite{King}, \cite{Macri}, \cite{DHKK}, \cite{DK1}. In the final step (Section \ref{connecting}) we utilize as such a tool also the relation $\bd R & \rDotsto & (S,E) \ed $ between a $\sigma$-regular object $R$ and an exceptional pair generated by it (introduced in \cite{DK1}).
In Section \ref{the exceptional objects} we organize in a better way the obtained in \cite[Section 2]{DK1} data about $\Hom(X,Y)$ and ${\rm Ext}^1(X,Y)$, where $X,Y$ vary throughout the exceptional objects of $Rep_k(Q)$, and we add some observations about the behavior of the central charges of the exceptional objects, which are very essential for the proof of Theorem \ref{main theo} as well.
Today, in view of the parallel between dynamical systems and categories \cite{DHKK}, \cite{BS} and in view of the Motivic Donaldson Thomas invariants \cite{KS} the importance of studying the topology of the space of Bridgeland stability conditions is even bigger.
We still do not understand the meaning of the obtained picture about $\st(D^b(Rep_k(Q)))$. We hope that an understanding of this meaning will open a way to analyzing more cases.
\textit{{\bf Acknowledgements:}}
The authors wish to express their gratitude to Tom Bridgeland, Dragos Deliu, Fabian Haiden, Umut Isik, Maxim Kontsevich, Alexander Noll, Tony Pantev, Pranav Pandit for their interest in this paper.
Both the authors were funded by NSF DMS 0854977 FRG, NSF DMS 0600800, NSF DMS 0652633
FRG, NSF DMS 0854977, NSF DMS 0901330, FWF P 24572 N25, by FWF P20778 and by an
ERC Grant.
The second author was funded by DMS-1265230 Wall Crossings in Geometry and Physics,
DMS-1201475 Spectra Gaps Degenerations and Cycles, and by
OISE-1242272 PASI On Wall Crossings Stability Hodge Structures \& TQFT Grant.
\textit{\textbf{Some notations.}} In these notes the letters ${\mathcal T}$ and $\mc A$ denote always a triangulated category and an abelian category, respectively, linear over an algebraically closed field $k$. T
he shift functor in ${\mathcal T}$ is designated by $[1]$. We write $\Hom^i(X,Y)$ for $\Hom(X,Y[i])$ and $\hom^i(X,Y)$ for $\dim_k(\Hom(X,Y[i]))$, where $X,Y\in \mc T$. For $X,Y\in\mc A$, writing $\Hom^i(X,Y)$, we consider $X,Y$ as elements in $\mc T=D^b(\mc A)$, i.e. $\Hom^i(X,Y)={\rm Ext}^i(X,Y)$.
We write $\langle S \rangle \subset \mc T$ for the triangulated subcategory of $\mc T$
generated by $S$, when $S \subset Ob(\mc T)$.
An \textit{exceptional object} is an object $E\in \mc T$ satisfying $\Hom^i(E,E)=0$ for $i\neq 0$ and $\Hom(E,E)=k $. We denote by ${\mc A}_{exc}$, resp. $D^b(\mc A)_{exc}$, the set of all
exceptional objects of $\mc A$, resp. of $D^b(\mc A)$.
An \textit{exceptional collection} is a sequence $\mc E = (E_0,E_1,\dots,E_n)\subset \mc T_{exc}$ satisfying $\hom^*(E_i,E_j)=0$ for $i>j$. If in addition we have $\langle \mc E \rangle = \mc T$, then $\mc E$ will be called a full exceptional collection. For a vector $\textbf{p}=(p_0,p_1,\dots,p_n)\in \ZZ^{n+1}$ we denote $\mc E[\textbf{p}]=(E_0[p_0], E_1[p_1],\dots, E_n[p_n])$. Obviously $\mc E[\textbf{p}]$ is also an exceptional collection. The exceptional collections of the form $\{\mc E[\textbf{p}]: \textbf{p} \in \ZZ^{n+1} \}$ will be said to be shifts of $\mc E$.
For two exceptional collections $\mc E_1$, $\mc E_2$ of equal length we write $\mc E_1 \sim \mc E_2$ if $\mc E_2 \cong \mc E_1[\textbf{p}]$ for some $\textbf{p} \in \ZZ^{n+1}$.
An abelian category $\mc A$ is said to be hereditary, if ${\rm Ext}^i(X,Y)=0$ for any $X,Y \in \mc A$ and $i\geq 2$, it is said to be of finite length, if it is Artinian and Noterian.
For any quiver $Q$ we denote by $D^b(Rep_k(Q))$ or just by $D^b(Q)$ the derived category of the category of representations of $Q$.
For any $a\in \RR$ and any complex number $z \in {\rm e}^{\ri \pi a} \cdot (\RR + \ri \RR_{>0})$, respectively $z \in {\rm e}^{\ri \pi a} \cdot \left (\RR_{<0} \cup (\RR + \ri \RR_{>0}) \right )$, we denote by $\arg_{(a,a+1)}(z)$, resp. $\arg_{(a,a+1]}(z)$, the unique $\phi \in (a,a+1)$, resp. $\phi \in (a,a+1]$, satisfying $z=\abs{z} \exp(\ri \pi \phi)$.
For a non-zero complex number $v \in \CC$ we denote the two connected components of $\CC \setminus \RR v $ by:
\begin{gather} \label{complement of a line} v^c_+ = v \cdot (\RR + \ri \RR_{>0}) \qquad v^c_- = v \cdot (\RR - \ri \RR_{>0}) \qquad \qquad v \in \CC \setminus \{0\}.\end{gather}
For $b\in (a,a+1)$, $c\in(a-1,a)$ $r_1>0$, $r_2>0$ we have
\begin{gather}
\arg_{(a,a+1)}(r_1 \exp(\ri \pi a)+r_2 \exp(\ri \pi b))=a+\arg_{(0,1)}(r_1+r_2 \exp(\ri \pi (b-a))) \nonumber\\[-3mm]
\label{arg1} \\[-3mm] \arg_{(a-1,a)}(r_1 \exp(\ri \pi a)+r_2 \exp(\ri \pi c))=a+\arg_{(-1,0)}(r_1+r_2 \exp(\ri \pi (c-a))). \nonumber \end{gather}
These formulas imply that for $c\in(a-1,a)$, $r_1>0$, $r_2>0$ we have
\begin{gather} \label{arg2}
\arg_{(a-1,a)} \left ( r_1 \exp(\ri \pi a) + r_2 \exp(\ri \pi c)\right )=-\arg_{(-a,-a+1)} \left ( r_1 \exp(-\ri \pi a) + r_2 \exp(-\ri \pi c)\right ).
\end{gather}
\section{Some general remarks} \label{general remarks}
Here we give tools which will be used throughout to analyze the intersection of the sets of stability conditions generated by different exceptional collections (Propositions \ref{all exc semistable}, \ref{phi_1>phi_2}, \ref{mutations} and Lemmas \ref{three comp factors}, \ref{two comp factors}).
A description of the set of stability conditions generated by all shifts of a fixed exceptional triple is given in Proposition \ref{lemma for f_E(Theta_E)}, which is also important for the rest of the paper. Due to Remark \ref{for n geq 3} it seems that Proposition \ref{lemma for f_E(Theta_E)} can not be generalized straightforwardly to the case of exceptional collections of length bigger than $3$.
\subsection{Basic facts and notations related to Bridgeland stability conditions} We use freely the axioms and notations on stability conditions introduced by Bridgeland in \cite{Bridg1} and some additional notations used in \cite[Subsection 3.2]{DK1}. In particular, for $\sigma = ({\mathcal P}, Z) \in \st(\mathcal T)$ we denote by
$\sigma^{ss}$ the set of $\sigma$-semistable
objects, i. e. \be
\label{sigma^{ss}} \sigma^{ss}=\cup_{t \in \RR} {\mathcal P}(t)\setminus \{0\}.
\ee
For any interval $I\subset \RR$ the extension closure of the slices $\{\mc P(x)\}_{x\in I}$ is denoted by $\mc P(I)$ in \cite{Bridg1}. The nonzero objects in the subcategory $\mc P(I)$ are exactly those $X\in \mc T\setminus \{0\}$, which satisfy $\phi_\pm(X)\in I$, i. e. whose HN factors have phases in $I$. In particular, if $X \in \mc P(a-1,a]\setminus \{0\}$ then $Z(X) \in \exp(\ri \pi a)_-^c\cup \RR_{>0} \exp(\ri \pi a) $.
From \cite{Bridg1} we know that for any $\sigma = (\mc P, Z) \in \st(\mc T)$ and any $t \in \RR$ the subcategory $\mc P(t,t+1]$ is a heart of a bounded t-structure. In particular $\mc P(t,t+1]$ is an abelian category, whose short exact sequences are exactly these sequences
$\bd A & \rTo ^{\alpha}& B & \rTo^{\beta} & C \ed $ with $A,B,C \in \mc P(t,t+1]$, s. t. for some $\gamma : C \rightarrow A[1]$ the sequence
$\bd A & \rTo ^{\alpha}& B & \rTo^{\beta} & C & \rTo^{\gamma} A[1]\ed $ is a triangle in $\mc T$. Using these remarks, the HH filtration and by drawing pictures one easily shows the following properties:
\begin{remark} \label{arg remark}
Let $t\in \RR$ and $X \in \mc P(a-1,a]$. Then:
\begin{itemize}
\item[(a)] If $ X \not \in \sigma^{ss}$ then $\phi_-(X) < \arg_{(a-1,a]}(Z(X)) < \phi_+(X) $.
\item[(b)] $X \not \in \sigma^{ss} $ iff there exists a monic arrow $X'\rightarrow X$ in the abelian category $ P(a-1,a]$ satisfying $\arg_{(a-1,a]}(Z(X')) >\arg_{(a-1,a]}(Z(X))$.
\item[(c)] If $Z(X)\in v_+^c$ for some $v\in \CC^*$ with $v=\abs{v}\exp(\ri \pi t)$ and $a-1 \in(t,t+1)$ or $a \in(t,t+1)$, then $\arg_{(a-1,a]}(Z(X))=\arg_{(t,t+1)}(Z(X))$. In particular, when $X \in \sigma^{ss}$, we have: \\ $\phi(X)=\arg_{(t,t+1)}(Z(X))$.
\end{itemize}
\end{remark}
\subsection{ Some remarks on \texorpdfstring{$\sigma$}{\space}-exceptional collections}
E. Macr\`i proved in \cite[Lemma 3.14]{Macri} that the extension closure ${\mc A}_{\mc
E}$ of a full Ext-exceptional collection ${\mc
E}=(E_0,E_1,\dots,E_n)$ in $\mc T$ is
a heart of a bounded t-structure. Furthermore,
${\mc A}_{\mc E}$ is of finite length and $E_0, E_1,\dots, E_n$ are the simple objects in it.
By Bridgeland's \cite[Proposition 5.3]{Bridg1} from the bounded t-structure ${\mc A}_{\mc E}$ is produced a family of stability conditions, which we denote by $\HH^{\mc A_{\mc E}}\subset \st(\mc T)$ or sometimes just $\HH^{\mc E}\subset \st(\mc T)$.
For a given $\sigma \in \st(\mc T)$ we define a
\textit{$\sigma$-exceptional collection} (\cite[Definition 3.19]{DK1}) as an Ext-exceptional collection ${\mc E } = (E_0,E_1,\dots,E_n)$, s. t. the
objects $\{E_i\}_{i=0}^n$ are $\sigma$-semistable, and $\{\phi(E_i)\}_{i=0}^n \subset (t,t+1)$
for some $t\in \RR$. The following Proposition is basic for this paper:
\begin{prop} \label{all exc semistable} Let $\mc T$ be a $k$-linear triangulated category and $\sigma=(\mc P, Z) \in \st(\mc T)$. Let $\mc E=(E_0,E_1,\dots, E_n)$ be a full $\sigma$-exceptional collection such that $\phi(E_i)\geq \phi(E_{i+1})$ and $\hom^1(E_i,E_{i+1})$ $\neq $ $ 0$ for some $i\in \{0,1,\dots,n-1\}$. Let $\mc A_{i,i+1}$ be the extension closure of $E_i, E_{i+1}$ in $\mc T$. Then each element in ${\mc T}_{exc}\cap \mc A_{i,i+1}$ is semistable.
\end{prop}
\bpr If $\phi(E_i)=\phi(E_{i+1})=t$, then $\mc A_{i,i+1}\subset \mc P(t)$ and hence all non-zero objects in $\mc A_{i,i+1}$ are semistable, therefore we can assume that $\phi(E_i)>\phi(E_{i+1})$.
By \cite[Corollary 3.20]{DK1} we have $\sigma \in \Theta_{\mc E}'=\HH^{\mc E}\cdot\widetilde{GL}^+(2,\RR)$. Since the action of $\widetilde{GL}^+(2,\RR)$ does not change the order of the phases, we can assume that $\sigma=(\mc P, Z) \in \HH^{\mc E}$, which means that the extension closure of $\mc E$ is the t-structure $\mc P(0,1]$ and \be \label{all exc semistable1} \phi(E_j)=\arg_{(0,1]}(Z(E_j)) \qquad j=1,\dots,n. \ee
Let us denote $ \mc T_{i,i+1} =\left \langle E_i,E_{i+1}\right \rangle $. From \cite[Proposition 3.17]{DK1} we have a projection map $\HH^{\mc E} \rightarrow \HH^{{\mc A}_{i,i+1}}\subset \st(\mc T_{i,i+1})$ and it maps $\sigma =(\mc P,Z)$ to a stability condition $\sigma' =(\mc P',Z')\in \HH^{{\mc A}_{i,i+1}}$ with $Z'(E_i)=Z(E_i)$, $Z'(E_{i+1})=Z(E_{i+1})$ and $\{{\mc P}'(t)=\mc P(t)\cap \mc T_{i,i+1}\}_{t\in\RR}$. Therefore it remains to show that the objects in ${\mc T}_{exc}\cap \mc A_{i,i+1}$ are $\sigma'$-semistable.
From \cite[Lemma 3.22]{DHKK} we have that $\mc A_{i,i+1}$ is a bounded t-structure in $\mc T_{i,i+1}$ and an equivalence of abelian categories $F:\mc A_{i,i+1} \rightarrow Rep_k(K(l))$ with $F(E_i)=s_1$, $F(E_{i+1})=s_2$, where $l=\hom^1(E_i,E_{i+1})$ and $s_1$, $s_2$ are the simple representations of $K(l)$ with $k$ at the source, sink, respectively. This equivalence maps $\sigma'\in \HH^{\mc A_{i,i+1}}$ to a stability condition
\begin{gather} \sigma'' = (\mc P'', Z'')\in \HH^{Rep_k(K(l))} \subset \st(D^b(K(l)))
\quad Z(E_i)=Z''(s_1), Z(E_{i+1})=Z''(s_2). \nonumber \end{gather} If $E\in \mc T_{exc}\cap \mc A_{i,i+1}$, then by the fact that $F$ is an equivalence of abelian categories it follows that $F(E)\in Rep_k(K(l))$ is an exceptional representation. Since $\{F(\mc P'(t))=\mc P''(t) \}_{t\in (0,1]}$, it remains to show that each exceptional representation of $Rep_k(K(l))$ is $\sigma''$-semistable.
Let $\rho\in Rep_k(K(l))_{exc}$. Then the dimension vector $\ul{\dim}(\rho)=(n,m)\in (n,m)$ is a real root of $K(l)$, furthermore it is a Schur root. From \eqref{all exc semistable1} we have $\arg(Z''(s_1))>\arg(Z''(s_2))$. By the arguments in the proof of \cite[Lemma 3.19]{DHKK} using a theorem by King (\cite[Proposition 4.4]{King} ) and $\arg(Z''(s_1))>\arg(Z''(s_2))$ we obtain a $\sigma''$-stable representation $X\in Rep_k(K(l))$ with $\ul{\dim}(X)=(n,m)$. Since $X$ is stable, it is simple in $\mc P''(t)$, where $t=\phi''(X)$, in particular it is indecomposable in $\mc P''(t)$. Since $\mc P''(t)$ is a thick subcategory (see \cite[Lemma 3.7]{DK1}), it follows that $X$ is indecomposable in $Rep_k(K(l))$. Since $\ul{\dim}(\rho)$ is a real root and both $X$, $\rho$ are indecompsable representations, the equality $\ul{\dim}(\rho)=\ul{\dim}(X)$ implies $\rho \cong X$(see \cite[Theorem 2, c)]{Kac}). The proposition follows.
\epr
Other statements, which will be widely used in the next sections are Propositions \ref{lemma for f_E(Theta_E)}, \ref{phi_1>phi_2} and \ref{mutations}. For the proof of Proposition \ref{lemma for f_E(Theta_E)} it is useful to define:
\begin{df} \label{S(I)} Let $n\geq 1$ be an integer. Let $\mc I = \{I_{ij}=(l_{ij},r_{ij})\subset \RR\}_{ 0 \leq i <j \leq n}$ be a family of non-empty open intervals, and let $ {\mk l}=\{ l_{ij} \in \{-\infty\}\cup \RR \}_{0 \leq i <j \leq n}$, ${\mk r}= \{ r_{ij} \in \RR \cup \{+\infty\} \}_{0 \leq i <j \leq n}$ be the corresponding families of left and right endpoints.
We will denote the following open convex set $ \{(y_0,y_1,\dots,y_n)\in \RR^{n+1} : y_i - y_j \in I_{ij} \ i<j\}\subset \RR^{n+1}$
by $S^n(\mc I)$ or $S^n({\mk l}, {\mk r})$.
\end{df}
For a full Ext-exceptional collection $\mc E=(E_0, E_1, \dots, E_n)$ in $\mc T$ we denote $\Theta_{\mc E}'=\HH^{\mc E}\cdot\widetilde{GL}^+(2,\RR)$. If $\mc E$ is a full Ext-exceptional collection, then we have (see \cite[Remark 3.21]{DK1}):
\begin{gather} \label{Theta_mc E'} \Theta_{\mc E}'=\HH^{\mc E}\cdot\widetilde{GL}^+(2,\RR)=\left \{\sigma : {\mc E}\subset \sigma^{ss} \ \mbox{and} \ \abs{\phi^\sigma(E_i)-\phi^\sigma(E_j)}<1 \ \mbox{for} \ i < j \right \} \end{gather}
and the assignment:
\begin{gather} \label{the map} \bd \{ \sigma \in \st(\mc T): \mc E \subset \sigma^{ss}\} \ni (\mc P,Z) & \rMapsto^{f_{\mc E}} & \left (\{\abs{Z(E_i)} \}_{i=0}^n, \{ \phi^\sigma(E_i) \}_{i=0}^n\right ) \in \RR^{2(n+1)} \ed \end{gather}
restricted to $\Theta_{\mc E}'$ defines a homeomorphism between $\Theta_{\mc E}'$ and $\RR_{>0}^{n+1}\times S^n(-\textbf{1},+\textbf{1})$ (as defined in Definition \ref{S(I)}).
Assume now that $\mc E=(E_0, E_1,\dots, E_n)$ is any full exceptional collection in $\mc T$ (not restricted to be Ext). If $\mc T$ is a triangulated category of finite type, then there are infinitely many choices of $\textbf{p}\in \ZZ^{n+1}$ such that $\mc E[\textbf{p}]=(E_0[p_0], E_1[p_1],\dots, E_n[p_n])$ is an Ext-exceptional collection. \cite[Lemma 3.19]{Macri} says that the following open subset of stability conditions is connected and simply conected:
\begin{gather}\label{theta_{mc E}} \Theta_{{\mc E}}= \bigcup_{\left \{\textbf{p} \in \ZZ^{n+1} : {\mc E}[\textbf{p}] \ \mbox{is Ext} \right \}} \Theta_{{\mc E}[\textbf{p}]}'\subset \st(\mc T). \end{gather}
For the sake of completeness we will comment on this set as well (compare with \cite[proof of Lemma 3.19]{Macri}).
By \cite[Corollary 3.20]{DK1} $\Theta_{{\mc E}}$ is the set of stability conditions $\sigma \in \st(\mc T)$ for which a shift of ${\mc E}$ is a $\sigma$-exceptional collection, in particular for each $\sigma \in \Theta_{{\mc E}}$ we have $\mc E \subset \sigma^{ss}$. Hence the assignment \eqref{the map} is well defined on $\Theta_{\mc E}$. Furthermore, this defines a homeomorphism between $\Theta_{\mc E}$ and $f_{\mc E}(\Theta_{\mc E})$. Indeed, if $\mc E[\textbf{p}]$ is an Ext-collection for some $\textbf{p}\in \ZZ^{n+1}$, then $f_{\mc E[\textbf{p}]}$ maps $\Theta_{{\mc E}[\textbf{p}]}'$ homeomorphically to $\RR_{>0}^{n+1}\times S^n(-\textbf{1},+\textbf{1})$ (see after \eqref{the map}) and due to $f_{\mc E[\textbf{p}]}-(0,\textbf{p})=f_{\mc E}$ we see that ${f_{\mc E}}_{\vert \Theta_{{\mc E}[\textbf{p}]}'}$ is homeomorphism onto its image $\RR_{>0}^{n+1}\times( S^n(-\textbf{1},+\textbf{1}) - \textbf{p})$. Therefore, provided that $f_{\mc E}$ is injective on $\Theta_{\mc E}$, \textbf{ the following restriction is a homeomorphism:}
\begin{gather} \label{A and homeo} {f_{\mc E}}_{\vert \Theta_{\mc E}}: \Theta_{{\mc E}} \ \ \rightarrow \ \ \RR_{>0}^{n+1} \times \left ( \bigcup_{ \textbf{p} \in A } S^n(-\textbf{1},+\textbf{1}) - \textbf{p} \right ), \ \ \mbox{where} \ \ A=\left \{\textbf{p} \in \ZZ^{n+1} : {\mc E}[\textbf{p}] \ \mbox{is Ext} \right \}. \end{gather}
To show that the obtained function is injective, assume that $\sigma_i=(\mc P_i, Z_i) \in \Theta_{\mc E}$, $i=1,2$ and $f_{\mc E}(\sigma_1)=f_{\mc E}(\sigma_2)$, i. e. $ \abs{Z_1(E_j)}=\abs{Z_2(E_j)}, \phi^{\sigma_1}(E_j)=\phi^{\sigma_2}(E_j)$ for all $j$, then by \eqref{Theta_mc E'} and the axiom $\phi^{\sigma}(E_j[p_j])=\phi^{\sigma}(E_j)+p_j$ we see that for any $\textbf{p}$ the incidence $\sigma_1 \in \Theta_{\mc E[\textbf{p}]}'$ is equivaaent to $\sigma_2 \in \Theta_{\mc E[\textbf{p}]}'$, hence by the injectivity of $f_{\mc E[\textbf{p}]}$ and $f_{\mc E[\textbf{p}]}-(0,\textbf{p})=f_{\mc E}$ we obtain $\sigma_1=\sigma_2$. Thus, we see that \eqref{A and homeo} is a homeomorphism.
Finally, note that by \eqref{Theta_mc E'}, \eqref{the map}, \eqref{theta_{mc E}}, and $f_{\mc E[\textbf{p}]}=f_{\mc E}+(0,\textbf{p})$ one easily shows that
\begin{gather} \label{Theta_E} \Theta_{\mc E} =\left \{\sigma\in \st(\mc T): \mc E \subset \sigma^{ss} \ \mbox{and} \ \phi^\sigma(\mc E) \in \bigcup_{ \textbf{p} \in A } S^n(-\textbf{1},+\textbf{1}) - \textbf{p} \right \}. \end{gather}
\subsection{The set \texorpdfstring{$f_{\mc E}(\Theta_{\mc E})$}{\space} when \texorpdfstring{$n=2$}{\space}} \label{f_E(Theta_E)}
Here is given an explicit representation of $f_{\mc E}(\Theta_{\mc E})$, when $n=2$. Remark \ref{for n geq 3} shows that the case $n\geq 3$ is not completely analogous. The only statement of this subsection, which will be used later is Proposition \ref{lemma for f_E(Theta_E)}, the rest is its proof.
Let us denote first: \be B^n=\{(0,q_1 ,q_2,\dots, q_n)\subset \NN^{n+1} : 0\leq q_1 \leq q_2 \leq\dots\leq q_n \}\ee
The following properties are clear from the definitions of $S^n(\mc J)$(Definition \ref{S(I)}) and of $A\subset \ZZ^{n+1}$(formula \eqref{A and homeo})
\begin{gather} \label{prop1} \forall \textbf{v}\in diag(\RR^{n+1}) \ \ \ \ S^n(\mc J) - \textbf{v} = S^n(\mc J) \\
\forall \textbf{v}\in diag(\ZZ^{n+1}) \ \ \ \ A-\textbf{v} = A \\
\forall \textbf{v} \in B^n \ \ \ \ A-\textbf{v} \subset A.
\end{gather}
Any $\textbf{p}=(p_0,p_1,\dots,p_n)\in A$ can be represented as $\textbf{p}-(p_0,p_0,\dots,p_0) + (p_0,p_0,\dots,p_0)$, hence if we denote
\be \label{A_0} A_0=\left \{\textbf{p} \in \ZZ^{n+1} : p_0=0, {\mc E}[\textbf{p}] \ \mbox{is Ext} \right \} \ee
by the properties above we can write
\begin{gather} \label{f_E(Theta_E)1} \bigcup_{ \textbf{p} \in A } S^n(-\textbf{1},+\textbf{1}) - \textbf{p}= \bigcup_{ \textbf{p} \in A_0 } S^n(-\textbf{1},+\textbf{1}) - \textbf{p} = \bigcup_{ \textbf{p} \in A_0 } \left ( \bigcup_{ \textbf{v} \in B^n } S^n(-\textbf{1},+\textbf{1}) + \textbf{v} \right ) - \textbf{p}.
\end{gather}
For the cases $n=1,2$ we have the following simple form of the expression in the brackets:
\begin{lemma} \label{lemma about S^2(-1,1)} The following equalities hold:
\begin{gather} \label{lemma about S^2(-1,1)eq} \bigcup_{ \textbf{v} \in B^1 } S^1(-1,+1) + \textbf{v} = S^1(-\infty,1) \qquad \bigcup_{ \textbf{v} \in B^2 } S^2(-\textbf{1},+\textbf{1}) + \textbf{v} = S^2(-\infty,\textbf{1}).\end{gather}
Recall that $S^n(-\infty,\textbf{1})=\{(y_0,y_1,\dots,y_n)\in \RR^{n+1} : y_i-y_j<1, i<j \}$ (see Definition \ref{S(I)}).
\end{lemma}
\begin{remark} \label{for n geq 3} For $n\geq 3$ we have not such an equality. For example, we have $(0,-\frac{1}{2},\frac{1}{2}, 0,\dots,0) \in S^n(-\infty,\textbf{1})$ but $(0,-\frac{1}{2},\frac{1}{2}, 0,\dots,0) \not \in \bigcup_{ \textbf{v} \in B^n } S^n(-\textbf{1},+\textbf{1}) + \textbf{v} $ for $n\geq 3$.
More precisely, it holds
$\bigcup_{ \textbf{v} \in B^n } S^n(-\textbf{1},+\textbf{1}) + \textbf{v} \subset \not = S^n(-\infty,\textbf{1})$ for $n\geq 3$.
\end{remark}
\bpr(of Lemma \ref{lemma about S^2(-1,1)}) Note first that for any $\mc I = \{I_{ij}: i <j \}$ as in Definition \ref{S(I)} and any $\textbf{p} \in \ZZ^{n+1}$ we have \be \label{shift of S(I)} S^n(\{I_{ij}: i <j \})-\textbf{p}=S^n(\{I_{ij}-(p_i-p_j): i <j \}). \ee
In particular for $n=1$ we have(now the index set of $\mc I$ has only one element: $(0,1)$): \begin{gather} \bigcup_{ \textbf{v} \in B^1 } S^1(-1,+1) + \textbf{v} = \bigcup_{ (0,k) \in \NN^2 } S^1(-1,+1) + (0,k) = \bigcup_{k \in \NN } S^1(-1-k,1-k) \nonumber \\ = \bigcup_{k \in \NN } \{-1-k < y_0-y_1 < 1-k\}= \{ y_0-y_1 < 1\}= S^1(-\infty ,+1) \nonumber \end{gather}
Using \eqref{prop1} and \eqref{shift of S(I)} one easily shows that:
\begin{gather} \bigcup_{ \textbf{v} \in B^2 } S^2(-\textbf{1},+\textbf{1}) + \textbf{v} = diag(\RR^{n+1}) \oplus \{y_2=0\} \cap \left (\bigcup_{ \textbf{v} \in B^2 } S^2(-\textbf{1},+\textbf{1}) + \textbf{v} \right ) \nonumber \\
S^2(-\infty,\textbf{1}) = diag(\RR^{n+1}) \oplus \{y_2=0\} \cap S^2(-\infty,\textbf{1}). \nonumber \end{gather}
Obviously we have
\begin{gather} \{y_2=0\} \cap S^2(-\infty,\textbf{1}) = \{y_2=0\}\cap \left \{ \begin{array}{c} y_0 - y_1 <1 \\ y_0 - y_2 <1\\ y_1 - y_2 <1\end{array} \right \} = \left \{ \begin{array}{c} y_0 - y_1 <1 \\ y_0 <1\\ y_1 <1\end{array} \right \}.\nonumber \end{gather}
We will prove the second equality in \eqref{lemma about S^2(-1,1)eq} by showing that:
\begin{gather}\label{lemma about S^2(-1,1)eq1} \{y_2=0\} \cap \left (\bigcup_{ \textbf{v} \in B^2 } S^2(-\textbf{1},+\textbf{1}) + \textbf{v} \right )= \left \{ \begin{array}{c} y_0 - y_1 <1 \\ y_0 <1\\ y_1 <1\end{array} \right \}. \end{gather}
Let $(0,k,k+l) \in B^2$, $k$, $l\in \NN$ be a vector in $B^2$. By \eqref{shift of S(I)} we have:
\begin{gather} S^2(-\textbf{1},\textbf{1})+(0,k,k+l)=\left \{ \begin{array}{c}-1-k < y_0-y_1 < 1-k \\
-1-k-l < y_0-y_2 < 1-k - l \\ -1-l < y_1-y_2 < 1- l \end{array} \right \}\subset \left \{ \begin{array}{c} y_0-y_1 < 1 \\
y_0-y_2 < 1 \\ y_1-y_2 < 1 \end{array} \right \}. \end{gather}
Denoting the unit open square by $C(-1,+1)=\{\abs{y_i}<1;i=0,1\}\subset \RR^2$, we can write:
\begin{gather} \{y_2=0 \}\cap \left (S^2(-\textbf{1},\textbf{1})+(0,k,k+l) \right )=\left \{ \begin{array}{c}-1-k < y_0-y_1 < 1-k\nonumber \\
-1-k-l < y_0 < 1-k - l \\ -1-l < y_1 < 1- l \end{array} \right \}\nonumber \\
=S^1(-1-k,+1-k)\cap \left ( C(-1,+1) -(k+l,l)\right )\nonumber \\
=\left ( S^1(-1,+1)+(0,k) \right ) \cap \left ( C(-1,+1) -(k+l,l)\right ) \nonumber \\
=\left ( S^1(-1,+1)-(k+l,k+l)+(0,k) \right ) \cap \left ( C(-1,+1) -(k+l,l)\right ) \nonumber \\
= \left ( S^1(-1,+1) \cap C(-1,+1)\right )-(k+l,l). \nonumber\end{gather}
Therefore:
\begin{gather}\label{lemma about S^2(-1,1)1} \{y_2=0 \}\cap \left (\bigcup_{\textbf{v}\in B^2}S^2(-\textbf{1},\textbf{1})+\textbf{v} \right )= \bigcup_{k\in \NN} \left ( \bigcup_{l\in \NN} \left ( S^1(-1,1) \cap C(-1,1)\right )-(l,l) \right ) - (k,0).\end{gather}
Before we continue with the proof of Lemma \ref{lemma about S^2(-1,1)}, we prove:
\begin{lemma} For any $k\in \ZZ\cup\{+\infty\}$ we have the following equality: \begin{gather} \label{eq for S(-1,1)}
\bigcup_{l\leq k} S^1(-1,+1) \cap C(-1,+1) +(l,l) = S^1(-1,+1) \cap \left \{\begin{array}{c} y_0 <1+k \\ y_1 <1+k \end{array}\right \} . \end{gather}
\end{lemma}
\bpr We show first the equality for $k=+\infty$. Let $(a_0, a_1) \in S^1(-1,+1)$, i. e. $\abs{a_0-a_1}<1$. Since $\RR=\bigcup_{l\in \ZZ} [2l-1,2l+1)$, there exists $l\in \ZZ$ such that $a_0+a_1 \in [2 l -1, 2 l +1)$, i. e. $-1 \leq a_0+a_1 - 2 l \leq 1$. We have also $-1<a_0-a_1<+1$ and due to the equalities:
\begin{gather} a_0 -l =\frac{ a_0+a_1-2l }{2} + \frac{ a_0-a_1}{2}; \ \ a_1 -l =\frac{ a_0+a_1-2l }{2} + \frac{ a_1-a_0}{2}\nonumber\end{gather}
we obtain $-1 =-\frac{1}{2}-\frac{1}{2}< a_i -l < \frac{1}{2}+\frac{1}{2}=1$ for $i=0,1$. Hence $(a_0,a_1) -(l,l)\in C(-1,+1) \cap S(-1,+1)$, and we proved the equality \eqref{eq for S(-1,1)} with $k=+\infty$. By \eqref{prop1} and since the translation in $\RR^2$ is bijective we rewrite this equality as follows $ S^1(-1,+1)= \bigcup_{l\in \ZZ} S^1(-1,+1) \cap C(-1,+1) +(l,l) = \bigcup \left ( S^1(-1,+1) +(l,l)\right ) \cap\left ( C(-1,+1) +(l,l) \right )= S^1(-1,+1) \cap \left ( \bigcup_{l\in \ZZ} C(-1,+1) +(l,l) \right )$. Hence
\begin{gather} \label{some shifts of square} S^1(-1,1) \cap \left \{\begin{array}{c} y_0 <1+k \\ y_1 <1+k \end{array}\right \} = S^1(-1,1) \cap \left ( \bigcup_{l\in \ZZ} C(-1,1) +(l,l) \right ) \cap \left \{\begin{array}{c} y_0 <1+k \\ y_1 <1+k \end{array}\right \}. \end{gather}
Due to the equalities
\begin{gather} \left \{\begin{array}{c} y_0 <1+k \\ y_1 <1+k \end{array}\right \} \cap \left ( C(-1,1) +(l,l) \right ) = \left \{ \begin{array}{c c} \emptyset & \ \mbox{if} \ \ l\geq k+2 \\
\left \{\begin{array}{c} k< y_0 <1+k \\ k < y_1 <1+k \end{array}\right \} \subset C(-1,1) +(k,k) & \ \mbox{if} \ \ l= k+1 \\
C(-1,+1) +(l,l) & \ \mbox{if} \ \ l\leq k \end{array} \right. \nonumber \end{gather}
we obtain $ \left ( \bigcup_{l\in \ZZ} C(-1,1) +(l,l) \right ) \cap \left \{\begin{array}{c} y_0 <1+k \\ y_1 <1+k \end{array}\right \} = \bigcup_{l\leq k} C(-1,1) +(l,l) $. By \eqref{some shifts of square} and applying again \eqref{prop1} we obtain the equality \eqref{eq for S(-1,1)} for $k\in \ZZ$.
\epr
Now we put \eqref{eq for S(-1,1)} with $k=0$ in \eqref{lemma about S^2(-1,1)1} and obtain
\begin{gather}\label{lemma about S^2(-1,1)2} \{y_2=0 \}\cap \left (\bigcup_{\textbf{v}\in B^2}S^2(-\textbf{1},\textbf{1})+\textbf{v} \right )= \bigcup_{k\in \NN} \left ( S^1(-1,+1) \cap \left \{\begin{array}{c} y_0 <1 \\ y_1 <1 \end{array}\right \} \right )- (k,0).\end{gather}
The next step is to show that
\begin{gather}\label{lemma about S^2(-1,1)3} \bigcup_{k\in \NN} \left ( S^1(-1,+1) \cap \left \{\begin{array}{c} y_0 <1 \\ y_1 <1 \end{array}\right \} \right )- (k,0)= \bigcup_{k\in \NN} \left ( S^1(-1,+1) - (k,0) \right ) \cap \left \{\begin{array}{c} y_0 <1 \\ y_1 <1 \end{array}\right \}.\end{gather}
The inclusion $\subset$ is clear. Assume now that $a_0, a_1 \in \RR$, $k\in \NN$ and $\abs{a_0-a_1} <1$ and $a_0-k <1$, $a_1 <1$. We have to find $a'_0 \in \RR$, and $k' \in \NN$ such that \be \abs{a'_0-a_1}<1 \qquad a'_0 <1 \qquad a'_0 - k'= a_0-k. \ee
First note that $a_0 = a_0- a_1 + a_1 < \abs{a_0- a_1 } + a_1 < 2$.
If $k=0$ or $a_0 <1$, then we put $a'_0 = a_0$, $k'=k$.
Thus, we can assume that $k\geq 1$ and $1 \leq a_0 < 2$. Now $a_1 <1$ and $\abs{a_0-a_1}<1$ imply $0\leq a_1 < 1$. It follows that $ -1 < -a_1\leq a_0-1-a_1 <1 $, therefore we can put $a'_0=a_0-1$, $k'=k-1$. Hence we obtain \eqref{lemma about S^2(-1,1)3}.
On the other hand by \eqref{prop1} and the already proven first equality in \eqref{lemma about S^2(-1,1)eq} we have
$$\bigcup_{k\in \NN} S^1(-1,1)-(k,0) = \bigcup_{k\in \NN}S^1(-1,1)+(0,k) =S^1(-\infty,1).$$ The latter equality and equalities \eqref{lemma about S^2(-1,1)2}, \eqref{lemma about S^2(-1,1)3} imply \eqref{lemma about S^2(-1,1)eq1} and the lemma follows.
\epr
Putting \eqref{lemma about S^2(-1,1)eq} in \eqref{f_E(Theta_E)1} and then using \eqref{shift of S(I)} we obtain for the case $n=2$:
\begin{gather} \label{f_E(Theta_E)2} \bigcup_{ \textbf{p} \in A } S^2(-\textbf{1},+\textbf{1}) - \textbf{p}= \bigcup_{ \textbf{p} \in A_0 } S^2(-\infty, \textbf{1}) - \textbf{p}= \bigcup_{ (0,p_1,p_2) \in A_0 } \left \{ \begin{array}{c} y_0 - y_1 < 1+p_1 \\ y_0 - y_2 < 1+p_2 \\ y_1 - y_2 < 1+p_2-p_1\end{array}\right \}.
\end{gather}
Using the equality \eqref{f_E(Theta_E)2}, the homeomorphism \eqref{A and homeo}, and \eqref{Theta_E} we will prove the main result of this subsection:
\begin{prop} \label{lemma for f_E(Theta_E)} Let $\mc T$ be a $k$-linear triangulated category. Let $\mc E = (A_0,A_1,A_2)$ be a full exceptional collection, such that: \begin{gather} 1+\alpha = \min \{i:\hom^i(A_0,A_1)\neq 0\}\in \ZZ \nonumber \\ \label{alpha,beta,gamma} 1+\beta = \min \{i:\hom^i(A_0,A_2)\neq 0\}\in \ZZ \\ 1+\gamma = \min \{i:\hom^i(A_1,A_2)\neq 0\}\in \ZZ . \nonumber \end{gather}
Then the subset $ \Theta_{\mc E}\subset \st(\mc T)$ defined in \eqref{theta_{mc E}} has the following description: \begin{gather} \label{Theta_E n=2} \Theta_{\mc E} =\left \{\sigma\in \st(\mc T): \mc E \subset \sigma^{ss} \ \mbox{and} \ \begin{array}{l} \phi^\sigma(A_0) -\phi^\sigma(A_1) < 1+\alpha \\ \phi^\sigma(A_0) -\phi^\sigma(A_2) < 1+\min\{\beta, \alpha+\gamma\} \\\phi^\sigma(A_1) -\phi^\sigma(A_2) < 1+\gamma\end{array} \right \} \end{gather}
and $\Theta_{\mc E}$ is homeomorphic with the set $ \RR_{>0}^3 \times \left \{ \begin{array}{c} y_0 - y_1 < 1+\alpha \\ y_0 - y_2 < 1+\min\{\beta, \alpha+\gamma\} \\ y_1 - y_2 < 1+\gamma\end{array} \right \}$ by the map $f_{\mc E}$ in \eqref{the map} restricted to $\Theta_{\mc E}$. In particular $\Theta_{\mc E}$ is contractible.
\end{prop}
\bpr Given a family $\mc I$ of the form: $\mc I$ $=$ $ \{ I_{01}=(-\infty,u), I_{02}=(-\infty,v), I_{12}=(-\infty,w) \}$, we write $S\left (\begin{array}{c} -\infty,u \\ -\infty,v \\ -\infty,w \end{array} \right ) $ for $S^2(\mc I) $ throughout the proof. By \eqref{f_E(Theta_E)2}, \eqref{Theta_E}, and \eqref{A and homeo} the proof is reduced to showing that:
\begin{gather} \label{lemma for f_E(Theta_E)1} \bigcup_{ (0,p_1,p_2) \in A_0 }S\left (\begin{array}{c} -\infty,1+p_1 \\ -\infty,1+p_2 \\ -\infty,1+p_2-p_1 \end{array} \right ) = S\left (\begin{array}{c} -\infty,1+\alpha \\ -\infty,1+\min\{\beta, \alpha+\gamma\} \\ -\infty,1+\gamma \end{array} \right ) . \end{gather}
From the definition of $A_0$ in \eqref{A_0} and the definition of $\alpha$, $\beta$, $\gamma$ one easily obtains:
\begin{gather} \label{p_1 leq alpha} (0,p_1,p_2) \in A_0 \ \ \Rightarrow \ \ p_1 \leq \alpha, p_2 \leq \min\{\beta,\alpha+\gamma \}; \qquad (0,\alpha, \min\{\beta, \alpha+\gamma\}) \in A_0. \end{gather}
If $u\leq u'$, $v\leq v'$, $w\leq w'$, then $S\left ( -\infty, (u,v ,w ) \right ) \subset S\left ( -\infty, (u',v' ,w' ) \right )$, hence by \eqref{p_1 leq alpha} we have:
\begin{gather} \bigcup_{ (0,p_1,p_2) \in A_0 }S\left (\begin{array}{c} -\infty,1+p_1 \\ -\infty,1+p_2 \\ -\infty,1+p_2-p_1 \end{array} \right )=S\left (\begin{array}{c} -\infty,1+\alpha \\ -\infty,1+\min\{\beta, \alpha+\gamma\} \\ -\infty,1+\min\{\beta, \alpha+\gamma\}-\alpha \end{array} \right ) \cup \nonumber \\[-2mm] \label{lemma for f_E(Theta_E)2} \\[-2mm] \bigcup_{\left \{ \begin{array}{c} (0,p_1,p_2)\in A_0: \\ p_2-p_1 > \min\{\beta, \alpha+\gamma\}-\alpha \end{array} \right \}} S\left (\begin{array}{c} -\infty,1+p_1 \\ -\infty,1+p_2 \\ -\infty,1+p_2-p_1 \end{array} \right ). \nonumber \end{gather}
Now we consider two cases.
\ul{If $\min\{\beta, \alpha+\gamma\}= \alpha + \gamma$,} then $\min\{\beta, \alpha+\gamma\}-\alpha=\gamma$ and $(A_0,A_1[p_1], A_2[p_2])$ is not an Ext-collection for $p_2-p_1 >\gamma$ (since $\hom^{p_1+\gamma+1-p_2}(A_1[p_1],A_2[p_2]) \neq 0 $, $p_2-p_1-\gamma-1\geq 0$), hence the equality \eqref{lemma for f_E(Theta_E)2} reduces to \eqref{lemma for f_E(Theta_E)1}.
\ul{If $\min\{\beta, \alpha+\gamma\}= \beta < \alpha + \gamma$,} then $\beta \leq \alpha -i + \gamma$ for $i \leq \alpha + \gamma - \beta$ and hence
\be \{(0,\alpha-i, \beta): 0\leq i \leq \alpha + \gamma - \beta \} \subset A_0. \ee
Furthermore, we claim that the equality \eqref{lemma for f_E(Theta_E)2} reduces to
\begin{gather} \label{lemma for f_E(Theta_E)3} \bigcup_{ (0,p_1,p_2) \in A_0 }S\left (\begin{array}{c} -\infty,1+p_1 \\ -\infty,1+p_2 \\ -\infty,1+p_2-p_1 \end{array} \right )= \bigcup_{i=0}^{\alpha + \gamma - \beta} S\left (\begin{array}{c} -\infty,1+\alpha -i \\ -\infty,1+\beta \\ -\infty,1+\beta-\alpha+i \end{array} \right ). \end{gather}
Indeed, the first set of the union in \eqref{lemma for f_E(Theta_E)2} is the same as the first set of the union \eqref{lemma for f_E(Theta_E)3}. Now assume that $(0,p_1,p_2) \in A_0$ and $p_2-p_1>\beta -\alpha$, then $\beta -\alpha < p_2-p_1 \leq \gamma$. Therefore for some $1\leq i \leq \gamma+\alpha-\beta$ we have $ p_2-p_1 =\beta -\alpha + i $.
From \eqref{p_1 leq alpha} we have also $ p_2 \leq \beta $, therefore $ p_1 =p_2-\beta +\alpha - i\leq \alpha - i$, and then $S\left (\begin{array}{c} -\infty,1+p_1 \\ -\infty,1+p_2 \\ -\infty,1+p_2-p_1 \end{array} \right )\subset S\left (\begin{array}{c} -\infty,1+\alpha -i \\ -\infty,1+\beta \\ -\infty,1+\beta-\alpha+i \end{array} \right )$ and we showed \eqref{lemma for f_E(Theta_E)3}. The last step of the proof is to show that
\begin{gather} \label{lemma for f_E(Theta_E)4} \bigcup_{i=0}^{\alpha + \gamma - \beta}S\left (\begin{array}{c} -\infty,1+\alpha -i \\ -\infty,1+\beta \\ -\infty,1+\beta-\alpha+i \end{array} \right )= S\left (\begin{array}{c} -\infty,1+\alpha \\ -\infty,1+\beta \\ -\infty,1+\gamma \end{array} \right ). \end{gather}
The inclusion $\subset $ is clear. To show the inclusion $\supset$, assume that $(a_0,a_1,a_2) \in \RR^3$ and \\ $a_0 - a_1 < 1 +\alpha$, $a_0 - a_2 < 1 +\beta$, $a_1 - a_2 < 1 +\gamma$.
If $a_0 - a_1 < 1+\alpha -(\alpha + \gamma -\beta) = 1+\beta - \gamma$, then by $a_1 - a_2 < 1 +\gamma$ it follows that $(a_0,a_1,a_2)$ is in the set with index $i=\alpha + \gamma -\beta$ on the right-hand side.
It remains to consider the case, when $ 1+\alpha -i > a_0-a_1 \geq 1 + \alpha -i -1 $ for some $0\leq i<\alpha + \gamma -\beta$. Now $(a_0,a_1,a_2)$ is in the set indexed by the given $i$. Indeed, now $ a_1-a_0 \leq i-\alpha $ and by $a_0-a_2 < 1 + \beta $ we have $a_1-a_2 = a_1-a_0 + a_0-a_2<1+\beta +i -\alpha$.
\epr
\subsection{More propositions used for gluing} Since we will often use the notion of a $\sigma$-triple, for the sake of completeness we rewrite here \cite[Definition 3.19]{DK1} for triples (see also \cite[Remark 3.31]{DK1}):
\begin{df} \label{sigma triple} An exceptional triple $(A_0, A_1,A_2)$ is a $\sigma$-triple iff the following conditions hold: {\rm \textbf{(a)}} $\hom^{\leq 0}(A_i, A_j)=0$ for $i\neq j$; {\rm \textbf{(b)}} $ \{A_i\}_{i=0}^2\subset \sigma^{ss} $ ; {\rm \textbf{(c)}} $ \{\phi(A_i)\}_{i=0}^2\subset (t,t+1) $ for some $t\in \RR$.
\end{df}
We enhance now Proposition \ref{all exc semistable} for the case $n=2$:
\begin{prop} \label{phi_1>phi_2} Let $\mc T$ be a $k$-linear triangulated category. Let $\mc E = (A_0,A_1,A_2)$, $\alpha$, $\beta$, $\gamma$ be as in Proposition \ref{lemma for f_E(Theta_E)}. Let $\sigma \in \Theta_{\mc E}$ (hence we have the inequalities in \eqref{Theta_E n=2}).
{\rm (a)} If $\phi^\sigma(A_0)\geq \phi^\sigma(A_1[\alpha])$, then $\mc A \cap \mc T_{exc}\subset \sigma^{ss}$, where $\mc A$ is the extension closure of $(A_0,A_1[\alpha])$.
{\rm (b) } If $\phi^\sigma(A_1)\geq \phi^\sigma(A_2[\gamma])$, then $\mc A \cap \mc T_{exc}\subset \sigma^{ss}$, where $\mc A$ is the extension closure of $(A_1,A_2[\gamma])$.
\end{prop}
\bpr If an equality holds in (a) or (b), then we have $\mc A \subset \mc P(t)$ for some $t\in \RR$ and the Proposition follows. Hence we can assume that we have a proper inequality in both the cases.
(a) By the definition of $\Theta_{\mc E}$ in \eqref{theta_{mc E}} and \cite[Corollary 3.20]{DK1} we see that $(A_0[l], A_1[i], A_2[j])$ is a $\sigma$-triple for some $l,i,j \in \ZZ$. We can assume\footnote{note that $(A_0, A_1[i], A_2[j])$ is a $\sigma$-triple iff $(A_0[k], A_1[i+k], A_2[j+k])$ is a $\sigma$-triple} $l=0$ and then $\hom^{\leq 0}(A_0,A_1[i])=0$ and $\abs{\phi(A_0)- \phi(A_1[i])}<1$. From the definition of $\alpha$ we see that $i\leq \alpha$. Actually we must have $i=\alpha$, otherwise the given inequality $\phi(A_0)- \phi(A_1[\alpha])>0$ implies $\phi(A_0)- \phi(A_1[i])>1$, which is a contradiction. Thus $(A_0,A_1[\alpha],A_2[j])$ is a $\sigma$-triple for some $j\in \ZZ$. Now we apply Proposition \ref{all exc semistable}.
(b) In this case we shift the given triple to a $\sigma$-triple of the form $(A_0[l], A_1, A_2[j])$ for some $l,j \in \ZZ$, in particular we have $\hom^{\leq 0}(A_1,A_2[j])=0$ and $\abs{\phi(A_1)- \phi(A_2[j])}<1$. From the definition of $\gamma$ and the given inequality $\phi(A_1)- \phi(A_2[\gamma])>0$ it follows that $j= \gamma$. Thus $(A_0[l],A_1,A_2[\gamma])$ is a $\sigma$-triple for some $l\in \ZZ$. Now we apply Proposition \ref{all exc semistable}.
\epr
\begin{prop} \label{mutations} Let $\mc T$ has the property that for each exceptional triple $(A_0,A_1,A_2)$ and any two $0\leq i<j\leq 2$ there exists unique $k\in \ZZ$ satisfying $\hom^k(A_i,A_j)\neq 0$. Let $\mc E = (A_0,A_1,A_2)$ be a full exceptional collection in $\mk{\mc T}$.
Let $R_0(\mc E) = (A_1,R_{A_1}(A_0),A_2)$, $L_0(\mc E) = (L_{A_0}(A_1), A_0, A_2)$, $R_1(\mc E) = (A_0, A_2, R_{A_2}(A_1))$, $L_1(\mc E) = (A_0, L_{A_1}(A_2), A_1)$ be the triples obtained by a single mutation applied to $\mc E$.\footnote{ Recall that for any exceptional pair $(A,B)$ the exceptional objects $L_A(B)$ and $R_B(A)$ are determined by the triangles $\bd L_A(B)&\rTo &\Hom^*(A,B)\otimes A & \rTo^{ev^*_{A,B}} & B \ed$; \ $\bd A &\rTo^{coev^*_{A,B}} &\Hom^*(A,B)^{\check{}}\otimes B &\rTo & R_B(A) \ed$ and that $(L_A(B),A)$, $(B,R_B(A))$ are exceptional pairs.}
Then the four intersections $\Theta_{\mc E}\cap \Theta_{R_0(\mc E)}$, $\Theta_{\mc E}\cap \Theta_{L_0(\mc E)}$, $\Theta_{\mc E}\cap \Theta_{R_1(\mc E)}$, $\Theta_{\mc E}\cap \Theta_{L_1(\mc E)}$ are all contractible and non-empty.
\end{prop}
\bpr Since $\mc E \sim \mc E'$ implies $\Theta_{\mc E}= \Theta_{\mc E'}$, $R_i(\mc E)\sim R_i(\mc E')$, $L_i(\mc E)\sim L_i(\mc E')$, we can assume that $l=\hom^1(A_0,A_1)> 0$, $p = \hom^1(A_1,A_2)> 0$. By the assumptions on $\mc T$ the other degrees are zero and it follows that the integers $\alpha$, $\gamma$ defined in \eqref{alpha,beta,gamma} vanish and from Proposition \ref{lemma for f_E(Theta_E)} we get:
\begin{gather} \label{Theta_E n=2 1} \Theta_{\mc E} =\left \{\sigma\in \st(\mc T): \mc E \subset \sigma^{ss} \ \mbox{and} \ \begin{array}{l} \phi^\sigma(A_0) -\phi^\sigma(A_1) < 1 \\ \phi^\sigma(A_0) -\phi^\sigma(A_2) < 1+\min\{\beta,0\} \\\phi^\sigma(A_1) -\phi^\sigma(A_2) < 1 \end{array} \right \}. \end{gather}
We start with the intersection $\Theta_{\mc E}\cap \Theta_{R_0(\mc E)}$.
Let us denote $X = R_{A_1}(A_0)[-1]$. Let $\alpha'$, $\beta'$, $\gamma'$ be the integers corresponding to the triple $(A_1,X,A_2)$ used in Proposition \ref{lemma for f_E(Theta_E)}. We have $1+\beta'=\min\{k : \hom^k(A_1,A_2)\neq 0\}=1+\gamma=1$, hence $\beta'=0$. On the other hand from the definition of $R_{A_1}(A_0)$ we have a triangle
\begin{gather} \label{one triangle} A_1^{\oplus l} \rightarrow X \rightarrow A_0 \rightarrow A_1^{\oplus l}[1]\end{gather}
and it follows that $\hom(A_1,X)\neq 0$, hence $\alpha'=-1$. We apply Proposition \ref{lemma for f_E(Theta_E)} to the triple $(A_1,X,A_2)$ and obtain (note that $1+\min\{\beta', \alpha'+\gamma'\}=1+\min\{0, \gamma'-1\}=\min\{1, \gamma'\}$) \begin{gather} \label{Theta_E n=2 2} \Theta_{R_0(\mc E)}=\Theta_{(A_1,X,A_2)} =\left \{\sigma\in \st(\mc T): \begin{array}{c}A_1\in \sigma^{ss} \\ X \in \sigma^{ss}\\ A_2\in \sigma^{ss}\end{array} \ \mbox{and} \ \begin{array}{l} \phi^\sigma(A_1) -\phi^\sigma(X) < 0 \\ \phi^\sigma(A_1) -\phi^\sigma(A_2) < \min\{1, \gamma'\}\\ \phi^\sigma(X) -\phi^\sigma(A_2) < 1+\gamma'\end{array} \right \}. \end{gather}
From the defintion of $\beta, \gamma$ we have $0=\hom^{\leq \min\{\beta,\gamma\}}(A_0,A_2)=\hom^{\leq \min\{\beta,\gamma\}}(A_1,A_2)$, and then the triangle \eqref{one triangle} implies that $\hom^{\leq \min\{\beta,\gamma\}}(X,A_2)=0$, it follows that
\begin{gather} \label{min beta, gamma leq gamma'} \min\{\beta, \gamma\}= \min\{\beta, 0\} \leq \gamma'. \end{gather}
Assume that $\sigma\in \Theta_{(A_1,X,A_2)}\cap \Theta_{\mc E}$. Then $A_0,A_1,A_2,X$ are all semi-stable and $\phi(A_1)<\phi(X)$.\footnote{We omit sometimes the superscript $\sigma$ in expressions like $\phi^\sigma(X)$ and write just $\phi(X)$.} It is easy to show that $\hom(X,A_0)\neq 0$ (using the triangle \eqref{one triangle}), hence $\phi(X)\leq \phi(A_0)$ and therefore $\phi(A_1)<\phi(A_0)$, and we obtain the inclusion $\subset$ in the following formula (the third inequality in this formula is the second in \eqref{Theta_E n=2 2}, the other inequalities are in \eqref{Theta_E n=2 1} together with $\phi(A_1)<\phi(A_0)$)
\begin{gather} \label{Theta_E n=2 3} \Theta_{\mc E}\cap \Theta_{R_0(\mc E)} =\left \{\sigma\in \st(\mc T): \mc E \subset \sigma^{ss} \ \mbox{and} \ \begin{array}{l} 0<\phi^\sigma(A_0) -\phi^\sigma(A_1) < 1 \\ \phi^\sigma(A_0) -\phi^\sigma(A_2) < 1+\min\{\beta, 0\} \\ \phi^\sigma(A_1) -\phi^\sigma(A_2) < \min\{\gamma',1\}\end{array} \right \}. \end{gather}
We show now the inclusion $\supset$. Assume that $\mc E \subset \sigma^{ss}$ and that the inequalities on the right hand side of \eqref{Theta_E n=2 3} hold. In particular the inequalities in \eqref{Theta_E n=2 1} hold, hence we have $\sigma \in \Theta_{\mc E}$ and $\phi^\sigma(A_0) >\phi^\sigma(A_1)$. Proposition \ref{phi_1>phi_2} (a) ensures $X\in \sigma^{ss}$ and by \eqref{one triangle} we get $\hom(A_1,X)\neq 0$, $\hom(X,A_2)\neq 0$, hence
\be \label{Z(X)=l Z(A_1) + Z(A_0)2} X \in \sigma^{ss} \quad \phi(A_1)\leq \phi(X)\leq \phi(A_0) \quad Z(X)=l Z(A_1) + Z(A_0). \ee
Using \eqref{phase formula} and $ 0<\phi^\sigma(A_0) -\phi^\sigma(A_1) < 1$ we see that $Z(A_1), Z(A_0)$ are not collinear(see Definition \ref{collinear}), therefore $Z(X)=l Z(A_1) + Z(A_0)$ is collinear neither with $ Z(A_1)$ nor with $Z(A_0)$. Now we apply \eqref{phase formula} again and by \eqref{Z(X)=l Z(A_1) + Z(A_0)2} we obtain $ \phi(A_1)< \phi(X)< \phi(A_0) $. In particular, we obtain the first inequality in \eqref{Theta_E n=2 2}. The second inequality in \eqref{Theta_E n=2 2} is the same as the third inequality of \eqref{Theta_E n=2 3}. From $\phi^\sigma(A_0) -\phi^\sigma(A_2) < 1+\min\{\beta, 0\} $ and \eqref{min beta, gamma leq gamma'} we get $\phi(X)-\phi^\sigma(A_2) < \phi^\sigma(A_0) -\phi^\sigma(A_2) < 1+\gamma' $, hence the third inequality in \eqref{Theta_E n=2 2} is verified also. Thus we showed \eqref{Theta_E n=2 3}. This equality implies that the set $\Theta_{\mc E}\cap \Theta_{R_0(\mc E)}$ is contractible. Indeed,we have a homeomorphism ${f_{\mc E}}_{\vert \Theta_{\mc E}}:\Theta_{\mc E} \rightarrow f_{\mc E}(\Theta_{\mc E})$ (see \eqref{A and homeo}, \eqref{the map}). The proved equality \eqref{Theta_E n=2 3} shows that:
$ f_{\mc E}\left ( \Theta_{\mc E}\cap \Theta_{R_0(\mc E)}\right )= \RR_{>0}^3 \times \left \{ \begin{array}{c} 0 <\phi_0 - \phi_1 <1 \\ \phi_0 -\phi_2 < 1+\min\{\beta, 0\} \\ \phi_1 -\phi_2 < \min\{\gamma',1\}\end{array}\right\} $,
hence $\Theta_{\mc E}\cap \Theta_{R_0(\mc E)}$ is contractible.
Next, we consider the intersection $\Theta_{\mc E}\cap \Theta_{L_1(\mc E)}$, where $L_1(\mc E)=(A_0, L_{A_1}(A_2), A_1)$.
Let us denote $Y = L_{A_1}(A_2)[1]$. Let $\alpha'$, $\beta'$, $\gamma'$ be the integers corresponding to the triple $(A_0,Y,A_1)$. Obviously $\beta'=\alpha=0$. From the definition of $L_{A_1}(A_2)$ we have a triangle
\begin{gather} \label{another triangle} A_2 \rightarrow Y \rightarrow A_1^{\oplus p} \rightarrow A_2[1]\end{gather}
and it follows that $\hom(Y,A_1)\neq 0$, hence $\gamma'=-1$. Proposition \ref{lemma for f_E(Theta_E)} applied to the triple $(A_0,Y,A_1)$ results in the equality(note that $1+\min\{0',\alpha'-1\}=\min\{1,\alpha'\}$) \begin{gather} \label{Theta_E n=2 5} \Theta_{L_1(\mc E)}=\Theta_{(A_0,Y,A_1)} =\left \{\sigma\in \st(\mc T): \begin{array}{c}A_0\in \sigma^{ss} \\ Y \in \sigma^{ss}\\ A_1\in \sigma^{ss}\end{array} \ \mbox{and} \ \begin{array}{l} \phi^\sigma(A_0) -\phi^\sigma(Y) < 1+\alpha' \\ \phi^\sigma(A_0) -\phi^\sigma(A_1) < \min\{1,\alpha'\} \\ \phi^\sigma(Y) -\phi^\sigma(A_1) < 0 \end{array} \right \}. \end{gather}
From the defintion of $\alpha, \beta$ for the initial sequence $\mc E$ we have $0=\hom^{\leq \min\{\alpha,\beta\}}(A_0,A_1)$,\\ $0=\hom^{\leq \min\{\alpha,\beta\}}(A_0,A_2)$, and then the triangle \eqref{another triangle} implies that $\hom^{\leq \min\{\alpha,\beta\}}(A_0,Y)=0$, it follows that
\begin{gather} \label{min beta, gamma leq gamma'1} \min\{\alpha,\beta\}= \min\{0,\beta\} \leq \alpha'. \end{gather}
Assume that $\sigma\in \Theta_{(A_0,Y,A_1)}\cap \Theta_{\mc E}$. Then $A_0,A_1,A_2,Y$ are all semi-stable and by \eqref{Theta_E n=2 5} $\phi(Y)<\phi(A_1)$. The triangle \eqref{another triangle} implies $\hom(A_2,Y)\neq 0$, hence $\phi(A_2)\leq \phi(Y) <\phi(A_1)$. Combining this inequality with the inequalities in \eqref{Theta_E n=2 5}, \eqref{Theta_E n=2 1} we obtain the inclusion $\subset$ in the following formula:
\begin{gather} \label{Theta_E n=2 6} \Theta_{\mc E}\cap \Theta_{L_1(\mc E)} =\left \{\sigma\in \st(\mc T): \mc E \subset \sigma^{ss} \ \mbox{and} \ \begin{array}{l} \phi^\sigma(A_0) -\phi^\sigma(A_1) < \min\{1,\alpha'\} \\ \phi^\sigma(A_0) -\phi^\sigma(A_2) < 1+\min\{\beta,0\} \\0<\phi^\sigma(A_1) -\phi^\sigma(A_2) < 1 \end{array} \right \}. \end{gather}
To show the inclusion $\supset$, assume that $\mc E \subset \sigma^{ss}$ and that the inequalities on the right hand side of \eqref{Theta_E n=2 6} hold. In particular, we have $\sigma \in \Theta_{\mc E}$(see \eqref{Theta_E n=2 1}). It remains to show that $Y\in\sigma^{ss}$ and that the inequalities in \eqref{Theta_E n=2 5} hold. From Proposition \ref{phi_1>phi_2} (b) and $\sigma \in \Theta_{\mc E}$, $\phi^\sigma(A_1) >\phi^\sigma(A_2)$ we obtain $Y \in \sigma^{ss}$. The triangle \eqref{another triangle} implies
\be \label{Z(Y)= Z(A_1) +p Z(A_2)} Y \in \sigma^{ss} \quad \phi(A_2)\leq \phi(Y)\leq \phi(A_1) \quad Z(Y)=p Z(A_1) + Z(A_2). \ee
By similar arguments as in the previous case, using \eqref{phase formula}, $ 0<\phi^\sigma(A_1) -\phi^\sigma(A_2) < 1$ and \eqref{Z(Y)= Z(A_1) +p Z(A_2)} one shows that $ \phi(A_2) < \phi(Y) < \phi(A_1) $. In particular, we obtain the third inequality in \eqref{Theta_E n=2 5}. The second inequality in \eqref{Theta_E n=2 5} is the same as the first inequality of \eqref{Theta_E n=2 6}. From $\phi^\sigma(A_0) -\phi^\sigma(A_2) < 1+\min\{\beta, 0\} $ and \eqref{min beta, gamma leq gamma'1} we get $\phi(A_0)-\phi^\sigma(Y) < \phi^\sigma(A_0) -\phi^\sigma(A_2) < 1+\alpha' $ and the first inequality in \eqref{Theta_E n=2 6} is verified also. Thus we showed \eqref{Theta_E n=2 6}. As in the previous case this implies that $ \Theta_{\mc E}\cap \Theta_{L_1(\mc E)}$ is contractible.
Finally, recall that $\mc E \sim R_0(L_0(\mc E))$, therefore $\Theta_{\mc E}\cap \Theta_{L_0(\mc E)}$ = $\Theta_{R_0(L_0(\mc E))}\cap \Theta_{L_0(\mc E)}$ and by the already proved first case we see that $\Theta_{\mc E}\cap \Theta_{L_0(\mc E)}$ is contractible. For the case $\Theta_{\mc E}\cap \Theta_{R_1(\mc E)}$ we have $\Theta_{\mc E}\cap \Theta_{R_1(\mc E)}=\Theta_{L_1(R_1\mc E)}\cap \Theta_{R_1(\mc E)}$ and contractibillity follows from a previous case. The Proposition is proved.
\epr
Propositions \ref{phi_1>phi_2} and \ref{all exc semistable} ensure semi-stability of certain exceptional objects. The following two lemmas are similar in that respect and will be used later, when we analyze the intersections of the form $\Theta_{\mc E_1}\cap \Theta_{\mc E_2}$, when $\mc E_2$ is obtained from $\mc E_1$ by more than one and different mutations.
\begin{lemma}\label{three comp factors} Let $\mc T=D^b(\mc A)$, where $\mc A$ is a hereditary abelian hom-finite category, and let for any two exceptional objects $E,F \in \mc T_{exc}$ there exists at most one $k\in \ZZ$ satisfying $\hom^k(E,F)\neq 0$.
Let $(A_0,A_1,A_2)$ be a full Ext-exceptional (``Ext-'' means that it satisfies (a) in Definition \ref{sigma triple}) collection in $\mc T$, such that $\hom^1(A_0,A_2)= 0$ and $A_0, A_1, A_2$ are semistable. Let $X,Y$ be exceptional objects in $\mc T$ for which we have a diagram of distinguished triangles, where all arrows are non-zero:
\be
\begin{diagram}[size=1em] \label{JH filtration}
0 & \rTo & & & A_2 & \rTo & & & X & \rTo & & & Y \\
& \luDashto & & \ldTo & &\luDashto & & \ldTo & & \luDashto & & \ldTo & \\
& & A_2 & & & & A_1 & & & & A_0 &
\end{diagram}.
\ee
{\rm (a) } If we have the following system of inequalities:
$$ \begin{array}{c} \phi(A_0)-1 < \phi(A_1) < \phi(A_0), \ \ \ \phi(A_0)-1 < \phi(A_2) < \phi(A_0) \\
\arg_{(\phi(A_0)-1,\phi(A_0))}(Z(A_0)+Z(A_1)) > \phi(A_2)
\end{array}, $$
then $Y \in \sigma^{ss}$ and $\phi(Y)<\phi(A_0)$.
{\rm (b) } If we have
$ \phi(A_2)< \phi(A_1)\leq \phi(A_0) < \phi(A_2)+1$,
then $Y \in \sigma^{ss}$ and $\phi(Y)<\phi(A_0)$.
\end{lemma}
\bpr We note first some vanishings. From the given diagram it follows that $\hom(Y,A_0)\neq 0$ and $\hom(X,A_1)\neq 0$. Since $X,Y$ are also exceptional objects, from \cite[Lemma 9.1]{DK1} and the hereditariness of $\mc A$ it follows that $\hom(A_0,Y)=\hom(A_1,X)= 0$. On the other hand $\hom(A_1,Y)=\hom(A_1,X)$ (follows by applying $\hom(A_1,\_)$ to the last triangle and using $\hom^*(A_1,A_0)=0$).
Thus, we obtain
\begin{gather} \label{three comp factors1}\hom(A_0,Y)=\hom(A_1,Y)= 0. \end{gather}
Since $(A_0,A_1,A_2)$ is an Ext-exceptional collection, its extension closure is a heart of a bounded t-structure(\cite[Lemma 3.14]{Macri}), furthermore this heart is of finite length and $(A_0,A_1,A_2)$ are the simple objects in it. Let us denote for simplicity $t=\phi(A_0)$(in case (a)) or $t-1=\phi(A_2)$(in case (b)). In both the cases from the given inequalities and since $\mc P(t-1,t]$ is also a heart, it follows that the extension closure of $(A_0,A_1,A_2)$ is exactly $\mc P(t-1,t]$. Now \eqref{JH filtration} can be considered as the Jordan-H\"older filtration of $Y$ in the abelian category $\mc P(t-1,t]$ and the composition factors of $Y$ are $\{A_0,A_1,A_2\}$.
Suppose that $Y\not \in \sigma^{ss}$. From Remark \ref{arg remark} (b) there exists $Y'\in \mc P(t-1,t]$ and a non-trivial monic arrow $Y'\rightarrow Y$, s. t. $\arg_{(t-1,t]}(Z(Y'))>\arg_{(t-1,t]}(Z(Y))$.
We have $ Z(Y)=Z(A_0)+Z(A_1)+Z(A_2) $ and one can show that the given inequalities in either case (a) or (b) imply that \begin{gather} \label{arg condition1} \arg_{(t-1,t]}(Z(Y)) > \arg_{(t-1,t]}(Z(A_2)), \quad \arg_{(t-1,t]}(Z(Y)) > \arg_{(t-1,t]}(Z(A_2)+Z(A_1)). \end{gather}
Since $Y'$ is a subobject of $Y$, the composition factors of $Y'$ in $\mc P(t-1,t]$ are subset of $\{A_0,A_1,A_2\}$. The cases $Y'\cong A_0$, $Y'\cong A_1$ are excluded by \eqref{three comp factors1}. The case $Y'\cong A_2 $ is excluded by the first inequality in \eqref{arg condition1} and the condition $\arg_{(t-1,t]}(Z(Y'))>\arg_{(t-1,t]}(Z(Y))$. Since $Y'$ is a proper subobject of $Y$ we reduce to the case when $Y'$ has two composition factors (two different elements of the set $\{A_0,A_1,A_2\}$). Using \eqref{three comp factors1} again we reduce to the following options for a Jordan H\"older filtration \begin{gather} \bd[1em] 0 &\rTo & A_2 &\rTo & Y' &\rTo & A_1 &\rTo & 0 \ed \qquad \bd[1em] 0 &\rTo & A_2 &\rTo & Y' &\rTo & A_0 &\rTo & 0 \ed. \end{gather}
In the first case we have $Z(Y')=Z(A_2)+Z(A_1)$ which contradicts the second inequality on \eqref{arg condition1}. In the second case we have a distinguished triangle$ \bd[1em] A_2 &\rTo & Y' &\rTo & A_0 &\rTo & A_2[1] \ed $ in $\mc T$, and form the given vanishing $\hom^1(A_0,A_2)=0$ it follows $Y'\cong A_0\oplus A_2$, which contradicts \eqref{three comp factors1}.
So we proved $Y\in \sigma^{ss}$.
The inequality $\phi(Y)<\phi(A_0)$(in either case (a) or (b)) follows from the given inequalities and $ Z(Y)=Z(A_0)+Z(A_1)+Z(A_2) $.
The lemma is proved.
\epr
\begin{lemma}\label{two comp factors} Let $\mc T=D^b(\mc A)$ be as in Lemma \ref{three comp factors}.
Let $(A_0,A_1,A_2)$ be a full Ext-exceptional collection\footnote{ ``Ext-'' means that it satisfies (a) in Definition \ref{sigma triple}} in $\mc T$ such that $A_0, A_1, A_2$ are semistable. Let $Y$ be an exceptional object in $\mc T$ for which we have a triangle, where all arrows are non-zero:
\begin{gather}
\begin{diagram}[size=1em] \label{JH filtration1}
A_2 & \rTo & & & Y \\
& \luDashto & & \ldTo & \\
& & A_0 &
\end{diagram}.
\end{gather}
If one of the two systems:
$ \begin{array}{c} \phi(A_2) < \phi(A_0) < \phi(A_2)+1 \\
\phi(A_2) < \phi(A_1) < \phi(A_2)+1
\end{array}$ or $ \begin{array}{c} \phi(A_0)-1 < \phi(A_1) < \phi(A_0) \\
\phi(A_0)-1 < \phi(A_2) < \phi(A_0)
\end{array}$ holds,
then we have:
$Y \in \sigma^{ss}$, $\phi(Y)=\arg_{(\phi(A_2), \phi(A_2)+1)}(Z(A_0)+Z(A_2))=\arg_{( \phi(A_0)-1 , \phi(A_0))}(Z(A_0)+Z(A_2)) $ and $\phi(A_2) < \phi(Y)<\phi(A_0)$.
\end{lemma}
\bpr Since $Y, A_0$ are exceptional objects and $\hom(Y,A_0)\neq 0$, from \cite[Lemma 9.1]{DK1} and the hereditarines of $\mc A$ it follows that $\hom(A_0,Y)= 0$. Due to the given inequalities, in both the cases we can choose $t$ so that $\phi(A_0), \phi(A_1), \phi(A_2) \in (t-1,t]$. By the same arguments as in the previous lemma, one sees that the extension closure of $(A_0,A_1,A_2)$ is $\mc P(t-1,t]$ and that this is an abelian category of finite length with simple objects $A_0,A_1,A_2 $. Now \eqref{JH filtration1} can be considered as the Jordan-H\"older filtration of $Y$ in the abelian category $\mc P(t-1,t]$ and the composition factors of $Y$ are $\{A_0,A_2\}$.
We have $ Z(Y)=Z(A_0)+Z(A_2) $ and the given inequalities (in either case)
imply that: \begin{gather} \phi(A_2)=\arg_{(t-1,t]}(Z(A_2)) < \arg_{(t-1,t]}(Z(Y))=\arg_{( \phi(A_0)-1 , \phi(A_0))}(Z(A_0)+Z(A_2)) \nonumber \\[-2mm] \label{arg condition2} \\[-2mm]=\arg_{(\phi(A_2), \phi(A_2)+1)}(Z(A_0)+Z(A_2)) < \arg_{(t-1,t]}(Z(A_0))= \phi(A_0). \nonumber \end{gather}
Suppose that $Y\not \in \sigma^{ss}$. From Remark \ref{arg remark} (b) it follows that there exists $Y'\in \mc P(t-1,t]$ and a non-trivial monic arrow $Y'\rightarrow Y$ in $\mc P(t-1,t]$, s. t. $\arg_{(t-1,t]}(Z(Y'))>\arg_{(t-1,t]}(Z(Y))$.
Since the composition factors of $Y$ are $\{A_0,A_2\}$ and $Y'$ is a non-zero proper sub-object of $Y$, then we have $Y'\cong A_2$ or $Y'\cong A_0$. The case $Y'\cong A_0$, is excluded by $\hom(A_0,Y)= 0$. The case $Y'\cong A_2 $ is excluded by \eqref{arg condition2} and the condition $\arg_{(t-1,t]}(Z(Y'))>\arg_{(t-1,t]}(Z(Y))$.
So we proved $Y\in \sigma^{ss}$.
By $Y\in \mc P(t,t+1]$ it follows that $\phi(Y)=\arg_{(t-1,t]}(Z(Y))$. Now the lemma follows from \eqref{arg condition2}. \epr
\section{The exceptional objects in \texorpdfstring{$D^b(Q )$}{\space}} \label{the exceptional objects}
From now on we fix $\mc T=D^b(Rep_k(Q ))$, where $Q$ is the affine quiver in figure \eqref{Q1}.
In this Section we organize in a better way the data about $\{ \Hom(X,Y), {\rm Ext}^1(X,Y)\}_{X\in \mc T_{exc}}$ obtained in \cite[Section 2]{DK1}. In Subsection \ref{central charges} are given some observations about the behavior of the vectors $\{ Z(X)\}_{X\in \mc T_{exc}}\subset \CC$, which will be helpful when we analyze the intersections of the form $\Theta_{\mc E_1}\cap \Theta_{\mc E_2}$ in the next sections.
We start by recalling the classification of the exceptional objects in $Rep_k(Q )$ obtained in \cite{DK1}.
Let us denote for any $m\geq 1$:
\begin{gather} \nonumber \pi_+^m: k^{m+1} \rightarrow k^{m}, \quad \pi_-^m: k^{m+1} \rightarrow k^{m}, \quad j_+^m: k^{m} \rightarrow k^{m+1}, \quad j_-^m: k^{m} \rightarrow k^{m+1} \\
\nonumber \pi_+^m(a_1,a_2,\dots, a_m, a_{m+1}) =(a_1,a_2,\dots, a_m) \qquad \pi_-^m(a_1,a_2,\dots, a_m, a_{m+1})=(a_2,\dots, a_m, a_{m+1}) \\
\nonumber j_+^m(a_1,a_2,\dots, a_m) =(a_1,a_2,\dots, a_m,0) \qquad
j_-^m(a_1,a_2,\dots, a_m)=(0,a_1,\dots, a_m).
\end{gather}
In \cite{DK1} was shown that:
\begin{prop}\cite[Proposition 2.2]{DK1} \label{exceptional objects in Q1} The exceptional objects up to isomorphism in $ Rep_{k}(Q ) $ are ($m=0,1,2,\dots$)
\begin{gather} E_1^{m} = \begin{diagram}[1em]
& & k^m & & \\
& \ruTo^{\pi_+^m} & & \luTo^{Id} & \\
k^{m+1} & \rTo^{\pi_-^m} & & & k^m
\end{diagram} \ \ \ \ E_2^m = \begin{diagram}[1em]
& & k^{m+1} & & \\
& \ruTo^{j_+^m} & & \luTo^{Id} & \\
k^{m} & \rTo^{j_-^m} & & & k^{m+1}
\end{diagram} \ \ \ \ E_3^m = \begin{diagram}[1em]
& & k^{m+1} & & \\
& \ruTo^{j_+^m} & & \luTo^{j_-^m} & \\
k^{m} & \rTo^{Id} & & & k^{m}
\end{diagram} \nonumber \\
E_4^m = \begin{diagram}[1em]
& & k^m & & \\
& \ruTo^{\pi_+^m} & & \luTo^{\pi_-^m} & \\
k^{m+1} & \rTo^{Id} & & & k^{m+1}
\end{diagram} \ \ \ \ M = \begin{diagram}[1em]
& & 0 & & \\
& \ruTo & & \luTo & \\
0 & \rTo & & & k
\end{diagram} \ \ \ \ M'= \begin{diagram}[1em]
& & k & & \\
& \ruTo^{Id} & & \luTo & \\
k & \rTo & & & 0
\end{diagram}.\nonumber \end{gather}
\end{prop}
We denote by $K(\mc T)$ the Grothendieck group of $\mc T$. For $X\in \mc T$ we denote by $[X]\in K(T)$ the corresponding equivalence class in $K(\mc T)$. From Proposition \ref{exceptional objects in Q1} it follows:
\begin{coro} \label{[X]} Let us denote $\delta =[E_1^0]+[E_3^0]+[M]\in K(\mc T)$. We have the following equalities in $K(\mc T)$:
\begin{gather} \label{[delta]} \delta =[E_1^0]+[E_3^0]+[M] = [E_1^0]+[E_2^0] =[E_3^0]+[E_4^0] = [M]+ [M']\\
\label{[E_1^m]} [E_1^m]= m \delta + [E_1^0] = (m+1) \delta - [E_2^0] \qquad [E_2^m]= m \delta + [E_2^0] = (m+1) \delta - [E_1^0] \\ \label{[E_3^m]}
[E_3^m]= m \delta + [E_3^0] = (m+1) \delta - [E_4^0] \qquad [E_4^m]= m \delta + [E_4^0] = (m+1) \delta - [E_3^0] \\
\label{E M M'} [E_1^m]+[M]=[E_4^m] \quad [E_3^m]+[M]=[E_2^m] \quad [E_4^m]+[M']=[E_1^{m+1}] \quad [E_2^m]+[M']=[E_3^{m+1}].
\end{gather}
\end{coro}
\subsection{The two orbits of 2-Kronecker pairs in \texorpdfstring{$D^b(Q)$}{\space}} \label{the two orbits}\mbox{}\\
In Propositions \ref{all exc semistable} and \ref{phi_1>phi_2} were discussed exceptional pairs $(E,F)$ with $\hom^{\leq 0}(E,F)=0$ and $\hom^{1}(E,F)=l\neq 0$, and their extension closures. We call such a pair \textit{$l$-Kronecker pair}. Kronecker pairs were used in \cite{DHKK} for studying the density of the set of phases of Bridgeland stability conditions. In \cite[Corollary 3.31]{DHKK} was shown that for any affine acyclic quiver $A$ (like the quiver $Q$ in figure \eqref{Q1}) only 1- and 2-Kronecker pairs can appear in $D^b(A)$. In this subsection we give some comments on the 1- and 2-Kronecker pairs in $D^b(Q)$, which will be useful later when we apply Propositions \ref{all exc semistable}, \ref{phi_1>phi_2} and Lemmas \ref{three comp factors}, \ref{two comp factors}.
From \cite[Remark 2.11]{DK1} we see that the 2-Kronecker pairs in $D^b(Q )$ up to shifts are:
\begin{gather} \label{Kron12}\mk{P}_{12}=\{ (E_1^{m+1},E_1^m[-1]), (E_1^0,E_2^0), (E_2^m,E_2^{m+1}[-1]): m\in \NN \} \\
\label{Kron34} \mk{P}_{43}=\{ (E_4^{m+1},E_4^{m}[-1]), (E_4^0,E_3^0), (E_3^m,E_3^{m+1}[-1]): m\in \NN \}.
\end{gather}
Recall that the Braid group on two strings $B_2\cong \ZZ$ acts on the set of equivalence classes of exceptional pairs in $\mc T$ (here we take the equivalence $\sim$ explained in \textbf{Some notations} and it is clear when a given equivalence class w.r. $\sim$ will be called a 2-Kronecker pair). Using \cite[Corollary 2.10]{DK1} and the list of triples in \cite[Corollary 2.9]{DK1} one can show that the set of 2-Kronecker pairs is invariant under this action of $B_2$ and this action on the 2-Kronecker pairs has two orbits. They are \eqref{Kron12} and \eqref{Kron34}.
We will describe now the sets ${\mc T}_{exc}\cap \mc A$, up to isomorphism, where $\mc A$ is the extension closure in $\mc T$ of a 2-Kronecker pair. This will be helpful later (e. g. when we apply Propositions \ref{all exc semistable} and \ref{phi_1>phi_2}).
We note first a simple lemma (in which $Rep_k(Q)$ can be any hereditary category):
\begin{lemma} \label{simple lemma} Let $A$, $B \in Rep_k(Q)$, let $\mc C$ be the extension closure of $A, B[-1]$ in $D^b(Rep_k(Q))=\mc T$. Then any $X \in \mc C \cap \mc T_{exc}$ has the form $X'[i]$, where $X'\in Rep_k(Q )_{exc}$ and $i\in \{0,-1\}$.
\end{lemma}
\bpr Since $Rep_k(Q)$ is hereditary, any object $X \in D^b(Rep_k(Q))$ decomposes as follows
$X \cong \bigoplus_{i \in \ZZ} H^i(X)[-i]$, where $H^i: \mc T \rightarrow Rep_k(Q)$ are the cohomology functors.
Since $A,B\in Rep_k(Q)$, it follows that $H^i(A) = H^i(B[-1]) = 0$ for each $i \neq \{0,1\}$. The functors $H^i: \mc T \rightarrow Rep_k(Q)$ map triangles to short exact sequences (see e.g. \cite{GM}), therefore $H^i(X)=0$ for any $X \in \mc C$ and any $i \neq \{0,1\}$. By the first paragraph of the proof we see that each $X \in \mc C$ has the form $X' \oplus X''[-1]$ with $X'$, $X'' \in Rep_k(Q)$. Finally, if $X \in \mc C \cap \mc T_{exc}$ , then $X$ is indecomposable in $\mc T$, hence either $X\cong X'$ or $X\cong X'[-1]$ for some $X'\in Rep_k(Q)$ and obviously $X'$ is also exceptional, i. e. $X'\in Rep_k(Q)_{exc}$.
\epr
\begin{lemma} \label{closures of 2-Kronecker} Let $(U,V)$ be one of the 2-Kronecker pairs given in \eqref{Kron12} or \eqref{Kron34}. Let $\mc A$ be its extension closure in $\mc T$. Then representatives of the iso-classes of objects in $\mc A \cap \mc T_{exc}$ are:
\begin{gather} \begin{array}{| c | c | c | c |}
\hline
(U,V) = & (E_{1/4}^{m+1},E_{1/4}^{m}[-1] ) & (E_{1/4}^0,E_{2/3}^0 ) & (E_{2/3}^{m},E_{2/3}^{m+1}[-1] ) \\ \hline
\mc A \cap {\mc T}_{exc}= & \left \{\begin{array}{c c} E_{1/4}^n[-1] & 0\leq n\leq m \\
E_{1/4}^n & n\geq m+1 \\
E_{2/3}^n & n\in \NN \end{array}\right \} & \left \{\begin{array}{c c} E_{1/4}^n & n\in \NN \\ E_{2/3}^n & n\in \NN \end{array} \right \} & \left \{\begin{array}{c c} E_{2/3}^n & 0\leq n\leq m \\
E_{2/3}^n[-1] & n\geq m+1 \\
E_{1/4}^n[-1] & n\in \NN \end{array}\right \} \\ \hline
\end{array} \nonumber
\end{gather}
where the subscript in the table is either everywhere the first or everywhere the second.
\end{lemma}
\bpr We show the case when the subscript is everywhere the first (i. e. the pairs in \eqref{Kron12}), the other case is analogous. From \cite[Lemma 3.22]{DHKK} we have that $\mc A$ is a bounded t-structure in $\scal{U,V}$ and we have also an equivalence of abelian categories \begin{gather} \label{closures of 2-Kronecker1} F:\mc A \rightarrow Rep_k(K(2)) \qquad F(U)=\bd[1em]k & \pile{\rTo \\ \rTo }& 0 \ed, \quad F(V)=\bd[1em]0 & \pile{\rTo \\ \rTo }& k \ed. \end{gather}
Using the facts that $\mc A$ is a bounded t-structure and that $F$ is equivalence, one can show that if $X \in \mc A \cap \mc T_{exc}$, then $F(X) \in Rep_k(K(2))_{exc}$. Furthermore, since $\mc T = D^b(Rep_k(Q ))$ and $Rep_k(Q )$ is a hereditary abelian category, it is easy to show that:
\begin{gather} \label{closures of 2-Kronecker2} X \in \mc A \cap {\mc T}_{exc} \ \ \Leftrightarrow \ \ F(X) \in Rep_k(K(2))_{exc}.\end{gather}
As in the proof of \cite[Proposition 2.2]{DK1} one can classify $ Rep_k(K(2))_{exc} $ and the result is:
\begin{gather}\label{closures of 2-Kronecker3} \forall X \in Rep_k(K(2))_{exc} \qquad \quad X \cong \bd k^{n+1} & \pile{\rTo^{\pi_+^n} \\ \rTo_{\pi_-^n}} & k^n \ed \ \ \mbox{or} \ \ X \cong \bd k^{n} & \pile{\rTo^{j_+^n} \\ \rTo_{j_-^n}} & k^{n+1} \ed \ \ \mbox{for some } \ n \in \NN. \end{gather}
Since $\mc A$ is a bounded t-structure in $\scal{U,V}$, the inclusion functor $\mc A \rightarrow \mc T$ induces an embedding of groups $K(\mc A)\rightarrow K(\mc T)$. Now from \eqref{closures of 2-Kronecker1}, \eqref{closures of 2-Kronecker2}, \eqref{closures of 2-Kronecker3} it follows that:
\begin{gather} \label{closures of 2-Kronecker4} \left \{[X] \in K(\mc T): X \in \mc A \cap \mc T_{exc}\right \}=\left \{(n+1) [U]+n [V], \ \ n [U]+(n+1) [V]: n\in \NN\ \right \}.\end{gather}
\ul{If $(U,V)=(E_1^{m+1},E_1^{m}[-1] )$,} then using \eqref{[E_1^m]} we obtain:
\begin{gather} (n+1) [U]+n [V]=(n+1)\left [E_1^{m+1}\right ]-n \left [E_1^{m}\right ]=(n+1) \left ((m+1)\delta + \left [E_1^0 \right ] \right )-n \left (m \delta +\left [E_1^0 \right ] \right ) \nonumber \\
= (n +m+1) \delta +\left [E_1^0\right ] =\left [E_1^{n+m+1} \right ] \nonumber \\
n [U]+(n+1) [V]=n\left ((m+1)\delta + \left [E_1^0\right ]\right )-(n+1) \left (m \delta +\left [E_1^0\right ] \right )
= (n -m) \delta -\left [E_1^0\right ] \nonumber \\ = \left \{ \begin{array}{c c} \left [E_2^{n-m-1} \right ] \ & \ n \geq m+1 \\ -\left [ E_1^{m-n} \right ] =\left [ E_1^{m-n}[-1] \right ] \ & \ n \leq m \end{array}
\right. . \nonumber
\end{gather}
Hence \eqref{closures of 2-Kronecker4} in this case is
$ \left \{[X] \in K(\mc T): X \in \mc A \cap \mc T_{exc}\right \}= \left \{\begin{array}{c c}\left [ E_1^n[-1] \right ] & 0\leq n\leq m \\
\left [ E_1^n \right ] & n\geq m+1 \\
\left [ E_2^n \right ] & n\in \NN \end{array}\right \}
$.
Now the second column in the table follows easily from Lemma \ref{simple lemma} and the fact that there is at most one, up to isomorphism, exceptional representation in $Rep_k(Q)$ of a given dimension vector (\cite[p. 13]{WCB2}).
\ul{If $(U,V)=(E_1^0,E_2^0 )$,} then using \eqref{[E_1^m]} and \eqref{[delta]} we reduce \eqref{closures of 2-Kronecker4} to $\left \{[X] \in K(\mc T): X \in \mc A \cap \mc T_{exc}\right \}=\{[E_1^n],[E_2^n]: n\in \NN \}$ and the third column of the table follows.
\ul{If $(U,V)=(E_2^{m},E_2^{m+1}[-1] )$,} then using \eqref{[E_1^m]} we obtain:
\begin{gather} (n+1) [U]+n [V]=(n+1) \left (m\delta + \left [E_2^0 \right ] \right )-n \left ((m+1) \delta +\left [E_2^0 \right ] \right )
= (m-n) \delta +\left [E_2^0\right ] \nonumber \\
= \left \{ \begin{array}{c c} \left [E_2^{m-n} \right ] \ & \ n \leq m \\ -\left [ E_1^{n-m-1} \right ] =\left [ E_1^{n-m-1}[-1] \right ] \ & \ n \geq m+1 \end{array} \right. \nonumber \\[2mm]
n [U]+(n+1) [V]=n\left (m\delta + \left [E_2^0\right ]\right )-(n+1) \left ((m+1)\delta +\left [E_2^0\right ] \right )
\nonumber \\= -(n +m+1) \delta -\left [E_2^0 \right ] = \left [ E_2^{n+m+1}[-1] \right ]
\nonumber
\end{gather}
and now \eqref{closures of 2-Kronecker4} and similar arguments as in the first case give the fourth column of the table.
The case when the subscript is everywhere the second (i. e. the pairs in \eqref{Kron34})
is obtained by substituting $E_1$ with $E_4$, $E_2$ with $E_3$, and using \eqref{[E_3^m]} instead of \eqref{[E_1^m]}.
\epr
Some 1-Kronecker pairs in $D^b(Q)$ are (see \cite[table (4)]{DK1}):
\begin{gather} \label{1-kronecker} (M', E_1^m[-1]), (M', E_2^m), (M, E_3^m), (M, E_4^m[-1]).\end{gather}
In the following lemma are listed several short exact sequences in $Rep_k(Q)$. On one hand, these sequences determine the set $\mc A\cap \mc T_{exc}$, where $\mc A$ is the extension closure of some of the 1-Kronecker pairs in \eqref{1-kronecker}, so they will be helpful when we apply Propositions \ref{all exc semistable} and \ref{phi_1>phi_2}. On the other hand, they (and their combinations) will play the role of the triangles \eqref{JH filtration} and \eqref{JH filtration1} when we apply Lemmas \ref{three comp factors} and \ref{two comp factors}.
\begin{lemma} There exist arrows in $Rep_k(Q )$ as shown below, so that the resulting sequences are exact($m\in \NN$):
\begin{gather
\label{ses2} \bd 0 & \rTo & E_3^{m} & \rTo & E_2^m & \rTo & M & \rTo & 0 \ed \\
\label{ses4} \bd 0 & \rTo & M & \rTo & E_4^m & \rTo & E_1^{m} & \rTo & 0 \ed\\
\label{ses7} \bd 0 & \rTo & M' & \rTo &E_1^{m+1} & \rTo & E_4^m & \rTo & 0 \ed \\
\label{ses8} \bd 0 & \rTo & E_2^m & \rTo & E_3^{m+1} & \rTo & M' & \rTo & 0 \ed\\
\label{ses10} \bd 0 & \rTo & E_3^0 & \rTo & M' & \rTo & E_1^{0} & \rTo & 0. \ed
\end{gather}
\end{lemma}
\bpr The proof is an exercise using Proposition \ref{exceptional objects in Q1} .\epr
\subsection{Recollection of some results of \cite{DK1} with new notations} \mbox{}\\
It is useful to introduce some notations (see Proposition \ref{exceptional objects in Q1} for the notations $E_i^j$, $M$, $M'$):
\begin{gather} \label{a^m b^m} a^m=\left \{ \begin{array}{c c} E_1^{-m} & m\leq 0 \\ E_2^{m-1}[1] & m\geq 1 \end{array} \right. ; \qquad \qquad \qquad b^m= \left \{ \begin{array}{c c} E_4^{-m} & m\leq 0 \\ E_3^{m-1}[1] & m\geq 1 \end{array} \right . . \end{gather}
\begin{remark} \label{exceptional objects} The objects in $\mc T_{exc}$ up to isomorphism are $\{a^j[k], b^j[k], M[k],M'[k]: j\in \ZZ, k\in \ZZ\}$.
\end{remark}
Using \cite[table (4) in Proposition 2.4]{DK1}), one verifies that:
\begin{coro}(of \cite[Proposition 2.4]{DK1})\label{nonvanishings} For each $m\in \ZZ$ we have:
\begin{gather} \label{nonvanishing1} \hom(M',a^m)\neq 0; \quad \hom(M,b^m)\neq 0; \quad \hom^*(a^m,M')= 0; \\
\label{nonvanishing2} \hom^1(a^m,M)\neq 0; \quad \hom^1(b^m,M')\neq 0; \quad \hom^*(b^m, M)=0 \\
\label{nonvanishing3} \hom^1(b^{m+1},a^n) \neq 0 \ \mbox{for} \ m> n; \quad \hom(b^{m},a^{n}) \neq 0 \ \mbox{for} \ m\leq n; \quad \hom^*(b^{m+1},a^m) = 0 \\
\label{nonvanishing4} \hom^1(a^{m},b^{n}) \neq 0 \ \mbox{for} \ m>n; \quad \hom(a^m,b^{n+1}) \neq 0 \ \mbox{for} \ m\leq n; \quad \hom^*(a^m,b^{m}) = 0; \\
\label{nonvanishing5} \hom(a^{m},a^n) \neq 0 \ \mbox{for} \ m\leq n; \quad \hom^1(a^{m},a^{n}) \neq 0 \ \mbox{for} \ m> n+1; \quad \hom^*(a^{m},a^{m-1})=0\\
\label{nonvanishing6} \hom(b^{m},b^n) \neq 0 \ \mbox{for} \ m\leq n; \quad \hom^1(b^{m},b^{n}) \neq 0 \ \mbox{for} \ m> n+1; \quad \hom^*(b^{m},b^{m-1})=0 \\
\label{nonvanishing7} \hom^1(M,M') \neq 0 \qquad \hom^1(M',M)\neq 0 .\end{gather}
\end{coro}
It is useful to keep in mind the following remarks:
\begin{remark} \label{inc dec seq} Recall that $\phi_-(A)> \phi_+(B)$ implies $\hom(A,B) = 0$ and in particular $\hom(A,B) \neq 0$ implies $\phi_-(A)\leq \phi_+(B)$ (for each stability condition).
Let $\{x^i\}_{i\in \ZZ}$ be either $\{a^i\}_{i\in \ZZ}$ or $\{b^i\}_{i\in \ZZ}$. From \eqref{nonvanishing5} and \eqref{nonvanishing6} we see that:
{\rm (a)} For $m\leq n$ we have $\hom(x^{m}, x^n) \neq 0 $. In particular, if $x^{m}, x^n \in \sigma^{ss}$ and $m\leq n$ then $\phi(x^m)\leq \phi(x^n)$.
{\rm (b)} For $m+1<n$ we have $\hom^1(x^n, x^{m}) \neq 0$. In particular, if $x^{m}, x^n \in \sigma^{ss}$ and $m+1<n$ then $\phi(x^n)\leq \phi(x^{m})+1$.
\end{remark}
\begin{remark} \label{ext closure ab} Let $\{x^i\}_{i\in \ZZ}$ be either $\{a^i\}_{i\in \ZZ}$ or $\{b^i\}_{i\in \ZZ}$. Lemma \ref{closures of 2-Kronecker} in terms of the notations \eqref{a^m b^m} is equivalent to saying that for any three integers $i\leq p$, $ p+1\leq j$ we have that $x^i$ and $x^j[-1]$ are in the extension closure of $\{x^p,x^{p+1}[-1]\}$.
\end{remark}
Keeping these remarks in mind one easily proves:
\begin{lemma}\label{from phases of big distance} Let $\{x^i\}_{i\in \ZZ}$ be either $\{a^i\}_{i\in \ZZ}$ or $\{b^i\}_{i\in \ZZ}$.
If there exists $m\in \ZZ$ such that $\{x^m, x^{m+1}\}\subset \sigma^{ss}$ and $\phi(x^m)+1<\phi(x^{m+1})$, then for $i\not \in \{m,m+1\}$ we have $x^i\not \in \sigma^{ss}$.
\end{lemma}
\bpr
Suppose $ x^{i} \in \sigma^{ss}$ with $i< m$, then by Remark (a) we have \ref{inc dec seq} $\phi(x^{i})+1\leq \phi(x^{m})+1<\phi(x^{m+1})$, hence $\hom^1(x^{m+1},x^{i})=0$, which contradicts the second part of Remark \ref{inc dec seq} (b). If $ x^{i} \in \sigma^{ss}$ with $i> m+1$, then by Remark \ref{inc dec seq} (a) we obtain $\hom^1(x^i,x^m)=0$, which again contradicts Remark \ref{inc dec seq} (b).
\epr
We will use often the following result obtained in \cite{DK1}:
\begin{coro} \label{one nonvanishing degree} \cite[Corollary 2.6 (b)]{DK1} For any two exceptional objects $X, Y \in D^b(Q )$ at most one element of the family $\{ \hom^p(X,Y) \}_{p\in \ZZ}$ is nonzero.
\end{coro}
Due to Corollary \ref{one nonvanishing degree} we can apply Lemmas \ref{three comp factors}, \ref{two comp factors} to $D^b(Q)$. Furthermore, we have:
\begin{coro}(\cite{DK1}) \label{exceptional colleections} The full exceptional collections in $D^b(Q )$ up to isomorphism and schifts are in the set of triples $\mk T$ given below. Propositions \ref{lemma for f_E(Theta_E)}, \ref{phi_1>phi_2}, \ref{mutations} can be applied to any of these triples.
\ben {\mk T} = \left \{ \begin{array}{c c c} (M',a^{m},a^{m+1}) & (a^m, b^{m+1},a^{m+1}) & (a^m,a^{m+1},M) \\
(M,b^{m},b^{m+1}) &(b^m,a^m,b^{m+1}) & (b^m,b^{m+1},M')\\
(b^{m},M',a^{m}) &(a^m , M, b^{m+1}) & .\end{array}: m\in \NN \right \}.
\een
\end{coro}
\bpr The list $\mk T$ follows straightforwardly from \cite[Corollary 2.9]{DK1}. By Corollary \ref{nonvanishings} $\hom^*(X,Y)\neq 0$ for any exceptional pair $(X,Y)$, therefore Propositions \ref{lemma for f_E(Theta_E)}, \ref{phi_1>phi_2} can be applied to any of the triples. Proposition \ref{mutations} can be applied due to Corollary \ref{one nonvanishing degree}. \epr
\begin{remark} It is known \cite{WCB1} that the Braid group on three strings $B_3$ acts transitively on the exceptional triples of $Rep_k(Q)$. This action is not free (see \cite[Remark 2.12]{DK1}).
\end{remark}
\begin{remark} \label{T_1234} With the notations \eqref{a^m b^m} the two orbits of 2-Kronecker pairs (see \eqref{Kron12} and \eqref{Kron34}) are $\{(a^m,a^{m+1}[-1])\}_{m\in \ZZ}$ and $\{(b^m,b^{m+1}[-1])\}_{m\in \ZZ}$. Each of these pairs can be extended to three non-equivalent triples, so we obtain two sets of triples. Having the list $\mk{T}$ above, we see that these two sets of triples are:
\begin{gather} \label{mk{T}_12} \mk{T}_{a}=\{ (M',a^{m},a^{m+1}) , (a^m, b^{m+1},a^{m+1}), (a^m,a^{m+1},M): m\in \ZZ\} \\
\label{mk{T}_43} \mk{T}_{b}=\{(M,b^{m},b^{m+1}), (b^m,a^m,b^{m+1}), (b^m,b^{m+1},M'): m \in \ZZ \}.
\end{gather}
Furthermore we have:
\begin{gather} \label{one union for mkT} \mk{T} =\mk{T}_{a} \cup \{ (b^{m},M',a^{m}), (a^m , M, b^{m+1}):m\in \ZZ \} \cup \mk{T}_{b} \qquad \mk{T}_{a}\cap \mk{T}_{b}=\emptyset. \end{gather}
\end{remark}
\begin{remark} \label{ses ab}
The short exact sequences \eqref{ses7}, \eqref{ses8}, \eqref{ses10} in terms of the notations \eqref{a^m b^m} become a sequence of distinguished triangles (for each $p$):
\be \label{short filtration 1}
\begin{diagram}[size=1em]
b^{p+1}[-1] & \rTo & & & M' \\
& \luDashto & & \ldTo & \\
& & a^p & &
\end{diagram}
\ee
The short exact sequences \eqref{ses2} and \eqref{ses4} become the following distinguished triangles ($q \in \ZZ$):
\be \label{short filtration 2}
\begin{diagram}[size=1em]
a^{q}[-1] & \rTo & & & M \\
& \luDashto & & \ldTo & \\
& & b^q & &
\end{diagram}.
\ee
\end{remark}
\subsection{Comments on the vectors \texorpdfstring{$\{Z(X):X \in \mc T_{exc}\}$}{\space}.} \label{central charges} \mbox{} \\
Corollary \ref{[X]} shows that for each $\sigma=(\mc P, Z) \in \st(\mc T)$ we have $ Z(\delta) = Z(E_1^0)+ Z(E_2^0)= Z(E_4^0)+Z(E_3^0)$ and $ Z(E_k^m)= m Z(\delta) + Z(E_k^0) $ for each $m\in \NN$ and each $ k=1,2,3,4$. Due to these equalities, with the notations \eqref{a^m b^m} we can write (recall that $Z(X[j]) = (-1)^j Z(X)$ for any $j\in \ZZ$, $X \in \mc T$):
\begin{gather}
\label{Z(delta)}\forall j \in \ZZ \qquad Z(a^{j+1})=Z(a^{j})-Z(\delta) \ \ \mbox{and} \ \ Z(b^{j+1})=Z(b^{j})-Z(\delta).\end{gather}
Therefore for any two integers $m,n$ we have:
\begin{gather} \label{Z(E_k^m)} Z(a^m)=Z(a^n)-(m-n) Z(\delta) \qquad \mbox{and} \qquad Z(b^m)=Z(b^n)-(m-n) Z(\delta) . \end{gather}
Next we discuss collinear vectors among $\{Z(a^j)\}_{j\in \ZZ}$ and $\{Z(b^j)\}_{j\in \ZZ}$. We fix first the meaning of ``collinear'':
\begin{df} \label{collinear} We say that a family $\{A_i\}_{i \in I}$ of complex numbers is collinear if $\{A_i\}_{i \in I} \subset \RR c$ for some $c \in \CC \setminus \{0\}$. In particular, $0 \in \CC$ is collinear to any $a\in \CC$. \end{df}
With this definition we have:
\begin{lemma}\label{vectors Z(E_k^m)} Let $\sigma=(\mc P, Z) \in \st(\mc T)$. Let $\{x^i\}_{i\in \ZZ}$ be either $\{a^i\}_{i\in \ZZ}$ or $\{b^i\}_{i\in \ZZ}$. Recall that $\delta$ is defined in \eqref{[delta]} and consider a sequence in $\CC$(infinite in both directions) of the form: \begin{gather} \label{the sequence of vectors} \dots, Z(x^{-i}),\dots, Z(x^{-2}),Z(x^{-1}), Z(x^0),Z(\delta), Z(x^1), Z(x^2), Z(x^3), \dots, Z(x^j), \dots. \end{gather}
Then the following conditions are equivalent: (a) Two of the vectors in this sequence are collinear; (b) The entire sequence is collinear.
\end{lemma} \bpr
Recall that formula \eqref{Z(E_k^m)} holds for any $m,n \in \ZZ$.
If $Z(x^i)$ and $Z(\delta)$ are collinear for some $i\in \ZZ$, then $ Z(\delta)$ and $Z(x^0) $ are collinear by $Z(x^0) =Z(x^i)+i Z(\delta) $ and (b) follows from the equalities $Z(x^j)=Z(x^0) -j Z(\delta)$, $j\in \ZZ$.
If $Z(x^i)$ and $Z(x^j)$ are collinear for some $i\neq j$, then by the equality$Z(\delta)=\frac{1}{j-i} (Z(x^i)-Z(x^j))$ we see that $ Z(\delta)$ and $Z(x^i) $ are collinear and (b) follows from the considered case above
\epr
\begin{coro} \label{from equal phases to} Let $\{x^i\}_{i\in \ZZ}$ be either $\{a^i\}_{i\in \ZZ}$ or $\{b^i\}_{i\in \ZZ}$. Let two of the vectors in the sequence \eqref{the sequence of vectors} be non-collinear.
Then:
{\rm (a) } All the vectors in this sequence are non-zero and no two of them are collinear.
{\rm (b) } If for two integers $n\neq m$ holds $\{x^n, x^m\}\subset \sigma^{ss}$, then we have $\phi(x^n)\not \in \phi(x^n)+\ZZ$.
{\rm (c) } The numbers $\{Z(x^j)\}_{j\in \ZZ} $ are contained in a common connected component of $\CC \setminus \RR Z(\delta)$.
{\rm (d) } If for two integers $n< m$ we have $\{x^n, x^m\}\subset \sigma^{ss}$ and $\phi(x^m)<\phi(x^n)+1$, then:\footnote{See \eqref{complement of a line} for the notations $Z(\delta)^c_\pm$.} $$\{Z(x^j)\}_{j\in \NN} \subset Z(\delta)^c_+.$$
\end{coro}
\bpr (a) and (b) follow from Lemma \ref{vectors Z(E_k^m)}, and the following axiom in \cite{Bridg1}:
\be
\label{phase formula} X\in \sigma^{ss} \qquad
\Rightarrow \qquad Z(X)=r(X)\ \exp(\ri \pi \phi(X)), \ r(X) >0.
\ee
Since $Z(x^0)$ and $Z(\delta)$ are non-colinear, it follows that either $Z(x^0)\in Z(\delta)^c_+$ or $Z(x^0)\in Z(\delta)^c_-$. From formula \eqref{Z(E_k^m)} we have $Z(x^j)=Z(x^0) -j Z(\delta)$ for any $j\in \ZZ$ therefore either $\{Z(x^j)\}_{j\in\ZZ}\subset Z(\delta)^c_+$ or $\{Z(x^j)\}_{j\in\ZZ}\subset Z(\delta)^c_-$. Therefore we obtain (c). Now to show (d), it is enough to show that $Z(x^m)\in Z(\delta)^c_+$.
From (b) and Remark \ref{inc dec seq} (a) we get the inequalities $\phi(x^n)<\phi(x^m)<\phi(x^n)+1$. By drawing a picture and taking into account formula \eqref{phase formula} and the equality $Z(x^n)=Z(x^m)+(m-n) Z(\delta)$, one sees that $\phi(x^n)<\phi(x^m)<\phi(x^n)+1$ is impossible if $Z(x^m)\in Z(\delta)^c_-$.
\epr
\begin{coro} \label{noncolinear ab} Let $\{x^i\}_{i\in \ZZ}$ be either the sequence $\{a^i\}_{i\in \ZZ}$ or the sequence $\{b^i\}_{i\in \ZZ}$.
If $Z(\delta) \neq 0$ and $Z(x^q) \in Z(\delta)_+^c$ for some $q\in \ZZ$, then $\{Z(x^i)\}_{i\in \ZZ}\subset Z(\delta)_+^c$ and for any $t\in \RR$ with $Z(\delta)=\abs{Z(\delta)}\exp(\ri \pi t)$ we have: \begin{gather} \label{noncolinear ab1} \forall p \in \ZZ \ \ \ \arg_{(t,t+1)}(Z(x^p))<\arg_{(t,t+1)}(Z(x^{p+1})) \\
\label{noncolinear ab2} \lim_{p\rightarrow -\infty } \arg_{(t,t+1)}(Z(x^p))=t; \quad \lim_{p\rightarrow+ \infty } \arg_{(t,t+1)}(Z(x^p))=t+1. \end{gather}
\end{coro}
\bpr
Since $Z(\delta)$ and $Z(x^q)$ are not collinear by Corollary \ref{from equal phases to} (c) and $Z(x^q) \in Z(\delta)_+^c$ it follows that $\{Z(x^i)\}_{i\in \ZZ}\subset Z(\delta)_+^c$. The inequalities \eqref{noncolinear ab1} follow from $Z(x^{p+1}), Z(x^p) \in Z(\delta)_+^c$ and $ Z(x^{p+1})=Z(x^p)-Z(\delta)$ (see \eqref{Z(E_k^m)} ). The formulas in \eqref{noncolinear ab2} follow also from \eqref{Z(E_k^m)} and $\{Z(x^i)\}_{i\in \ZZ}\subset Z(\delta)_+^c$.
\epr
\begin{coro} \label{noncolinear ab full1} Let $Z(M)$ and $Z(M')$ be non-zero and have the same direction.\footnote{We mean that $Z(M)= y Z(M')$ for some $y\in \RR_{>0}$. In particular, by $Z(\delta)=Z(M)+Z(M')$ (recall \eqref{[delta]}) we see that $Z(\delta)$ is non-zero.} Let $Z(a^q), Z(b^p) \in Z(\delta)_+^c$ for some $p,q\in \ZZ$.
Then $\{Z(a^j), Z(b^j)\}_{j\in \ZZ}\subset Z(\delta)^c_+$ and for any $t\in \RR$ with $Z(\delta)=\abs{Z(\delta)}\exp(\ri \pi t)$ the formulas \eqref{noncolinear ab2}, \eqref{noncolinear ab1} hold for both the sequences $\{Z(a^j)\}_{j\in \ZZ}$ and $\{Z(b^j)\}_{j\in \ZZ}$.
Furthermore, for any three integers $i,j,m$ we have:
\begin{gather} \label{j<mleqi} j<m\leq i \ \ \Rightarrow \ \ \arg_{(t,t+1)}(Z(a^j))< \arg_{(t,t+1)}(Z(b^m))<\arg_{(t,t+1)}(Z(a^i)).\end{gather}
\end{coro}
\bpr Corollary \ref{noncolinear ab} shows the first part of the conclusion.
To show \eqref{j<mleqi} we note first that the equalities \eqref{E M M'} with the notations \eqref{a^m b^m} become the following (for any $m\in \ZZ$):
\begin{gather} \label{a b M M'} Z(b^m)-Z(M)=Z(a^m) \qquad Z(b^m)+Z(M')=Z(a^{m-1}). \end{gather}
Since $Z(a^{m-1}), Z(a^{m}), Z(b^{m}) \in Z(\delta)^c_+$ for any $m\in \ZZ$ and $Z(M)$, $Z(M')$ have the same direction as $Z(\delta)$ (recall \eqref{[delta]}) the equalities \eqref{a b M M'} imply that $\arg_{(t,t+1)}(Z(a^{m-1}))< \arg_{(t,t+1)}(Z(b^m))<\arg_{(t,t+1)}(Z(a^m))$ for any $m\in \ZZ$. Now \eqref{j<mleqi} follows from \eqref{noncolinear ab1} (applied to the case $\{x^i\}_{i\in \ZZ }=\{a^i\}_{i\in \ZZ }$).
\epr
\section{The union \texorpdfstring{$\st(D^b(Q )) = \mk{T}_{a}^{st} \cup (\_,M,\_) \cup (\_,M',\_) \cup \mk{T}_{b}^{st}$}{\space}} \label{the union}
In this Section we distinguish some building blocks of $\st(D^b(Q ))$ and organize them in a manner consistent with the order in which we will glue these blocks in the next sections.
Theorem \ref{main theorem for Q in intro} says that for each $\sigma \in \st(D^b(Q ))$ there exists a $ \sigma$-triple. This means that (see \cite[Corollary 3.20]{DK1}) for each $\sigma$ there exists an Ext-exceptional triple $\mc E$ with $\sigma \in \Theta_{\mc E}'$. From Corollary \ref{exceptional colleections} we see that $\mc E$ is a shift of some of the triples in $\mk{T}$. Recalling the notation \eqref{theta_{mc E}} we get
$ \st(D^b(Q )) = \bigcup_{{\mc E} \in \mk{T}} \Theta_{\mc E}.$ Our basic building blocks are $\{ \Theta_{\mc E}\}_{\mc E \in \mk{T}}$ and by Proposition \ref{lemma for f_E(Theta_E)} they are contractible.
\textit{For a given triple $(A,B,C)\in \mk{T}$ we will denote the open subset $\Theta_{(A,B,C)} \subset \st(D^b(Q ))$ by $(A,B,C)$, when (we believe that) no confusion may arise.} With this convention we can write
\begin{gather} \label{st} \st(D^b(Q )) = \bigcup_{(A,B,C) \in \mk{T}} (A,B,C).
\end{gather}
\textit{For a given $X\in \{M,M'\}$ we denote by $ (X,\_,\_)$ the following open subset of $\st(D^b(Q )) $:}
\begin{gather} \label{(A,_,_)} \st(D^b(Q )) \supset (X,\_,\_) = \bigcup_{\{(B_0,B_1,B_2) \in \mk{T} : B_0 = X \}} (X,B_1,B_2).
\end{gather}
\textit{Similarly we define $(\_,X,\_)$ and $(\_,\_,X)$.} Looking at the list $\mk{T}$ and denoting (see \eqref{mk{T}_12}, \eqref{mk{T}_43}):
\begin{gather} \label{mk{T}} \mk{T}_{a}^{st}=\bigcup_{(A,B,C)\in \mk{T}_{a}} (A,B,C) \subset \st(D^b(Q )); \qquad \mk{T}_{b}^{st}=\bigcup_{(A,B,C)\in \mk{T}_{b}} (A,B,C) \subset \st(D^b(Q ))\end{gather} we can regroup the union \eqref{st} using \eqref{one union for mkT} as follows:
\begin{gather} \label{st with T} \st(D^b(Q )) = \mk{T}_{a}^{st} \cup (\_,M,\_) \cup (\_,M',\_) \cup \mk{T}_{b}^{st}. \end{gather}
\begin{remark} \label{remark about shifts} From the very definition \eqref{theta_{mc E}} of $\Theta_{\mc E}$ it is clear that $\Theta_{\mc E[\textbf{p}]} = \Theta_{\mc E}$ for any triple $\mc E=(A,B,C)\in \mk{T}$ and any $\textbf{p}\in \ZZ^3$. Using the notations explained here, we have $(A,B,C)=(A[p_0],B[p_1],C[p_2])\subset \st(D^b(Q ))$ for any $p_0,p_1,p_2 \in \ZZ$.
\end{remark}
\section{Some contractible subsets of \texorpdfstring{$\mk{T}_{a}^{st} \ \mbox{and}\ \mk{T}_{b}^{st}$}{\space}. Proof that \texorpdfstring{$\mk{T}_{a}^{st}\cap \mk{T}_{b}^{st}=\emptyset$}{\space}} \label{mk T_12 cap mk T_43}
In this Section is shown that $(X,\_,\_)$ and $(\_,\_,X)$ are contractible subsets of $\st(\mc T)$ for any $X\in \{M,M'\}$ and that {$\mk{T}_{a}^{st}\cap \mk{T}_{b}^{st}=\emptyset$.
We will refer often
to some of the formulas in Corollary \ref{nonvanishings}.
Whenever we discuss $\hom(A,B)$ or $\hom^1(A,B)$ with $A,B$ varying in the symbols $M,M', a^m, b^m$, $m\in \ZZ$, we refer to
Corollary \ref{nonvanishings}.
Putting \eqref{mk{T}_12}, \eqref{mk{T}_43} in \eqref{mk{T}} we obtain: \begin{gather} \label{T_12new} \mk{T}_{a}^{st} = (M',\_,\_) \cup (\_,\_,M) \cup \bigcup_{p\in \ZZ}(a^p,b^{p+1},a^{p+1})\\ \label{T_43new} \mk{T}_{b}^{st} = (M,\_,\_) \cup (\_,\_,M') \cup\bigcup_{q\in \ZZ}(b^q,a^q,b^{q+1}) \\ \label{(M',_,_)} (M',\_,\_) = \bigcup_{m \in \ZZ} (M',a^m,a^{m+1}) ; \qquad (M,\_,\_) = \bigcup_{m \in \ZZ} (M, b^m,b^{m+1})\\ \label{(_,_,M)} (\_,\_,M) = \bigcup_{m \in \ZZ} (a^m,a^{m+1},M) ; \qquad (\_,\_,M') = \bigcup_{m \in \ZZ} (b^{m},b^{m+1},M').
\end{gather}
We apply Proposition \ref{lemma for f_E(Theta_E)} to the triples $(a^p,b^{p+1},a^{p+1})$ and $(b^q,a^q,b^{q+1})$. Using Corollary \ref{one nonvanishing degree} and the formulas in Corollary \ref{nonvanishings} we see that in both the cases the coefficients $\alpha, \beta, \gamma$ defined in \eqref{alpha,beta,gamma} are $\alpha=\beta=\gamma=-1$. Thus, we obtain the following formulas for the sets $(a^p,b^{p+1},a^{p+1})\subset \st(D^b(\mc T))$ and $(b^q,a^q,b^{q+1})\subset \st(D^b(\mc T))$ in the first and the second column, respectively:
\begin{gather} \label{no M} \begin{array}{| c | c |} \hline
(a^p,b^{p+1},a^{p+1}) & (b^q,a^q,b^{q+1}) \\
\hline
\left \{a^p,b^{p+1},a^{p+1} \in \sigma^{ss} : \begin{array}{c} \phi\left ( a^{p}\right) < \phi\left (b^{p+1}\right) \\ \phi\left ( a^{p}\right)+1 < \phi\left (a^{p+1}\right) \\ \phi\left ( b^{p+1}\right) < \phi\left (a^{p+1}\right)\end{array} \right \} & \left \{ b^q,a^q,b^{q+1} \in \sigma^{ss} : \begin{array}{c} \phi\left ( b^{q}\right) < \phi\left (a^{q}\right) \\ \phi\left ( b^{q}\right)+1 < \phi\left (b^{q+1}\right) \\ \phi\left ( a^{q}\right) < \phi\left (b^{q+1}\right)\end{array} \right \} \\ \hline
\end{array}
\end{gather}
Similarly, applying Proposition \ref{lemma for f_E(Theta_E)} to the triples in the unions \eqref{(_,_,M)}, \eqref{(M',_,_)} (with the help of Corollary \ref{nonvanishings} and Corollary \ref{one nonvanishing degree}) we see that $ (M',\_,\_) \cup (\_,\_,M) $ and $(M,\_,\_)\cup (\_,\_,M') $ are the unions of the sets in the first and the second column of the following table, respectively (where $m,n,i,j$ vary throughout $\ZZ$):
\begin{gather} \label{left right M} \begin{array}{| c | c |} \hline
(M',\_,\_) \cup (\_,\_,M) & (M,\_,\_)\cup (\_,\_,M') \\
\hline
\left \{ M',a^{j},a^{j+1} \in \sigma^{ss} : \begin{array}{c} \phi\left (M'\right) < \phi\left (a^{j}\right) \\ \phi\left ( M'\right)+1 < \phi\left (a^{j+1}\right) \\ \phi\left ( a^{j}\right) < \phi\left (a^{j+1} \right)\end{array}\right \} & \left \{ M,b^{n},b^{n+1} \in \sigma^{ss} : \begin{array}{c} \phi\left (M\right) < \phi\left (b^{n}\right) \\ \phi\left ( M\right)+1 < \phi\left (b^{n+1}\right) \\ \phi\left ( b^{n}\right) < \phi\left (b^{n+1} \right)\end{array} \right \} \\ \hline
\left \{ a^{m},a^{m+1}, M \in \sigma^{ss} : \begin{array}{c} \phi\left ( a^{m}\right) < \phi\left (a^{m+1}\right) \\ \phi\left ( a^{m}\right) < \phi\left (M\right) \\ \phi\left ( a^{m+1}\right) < \phi\left (M\right)+1\end{array} \right \} & \left \{ b^{i},b^{i+1}, M' \in \sigma^{ss}: \begin{array}{c} \phi\left ( b^{i}\right) < \phi\left (b^{i+1}\right) \\ \phi\left ( b^{i}\right) < \phi\left (M'\right) \\ \phi\left ( b^{i+1}\right) < \phi\left (M'\right)+1\end{array} \right \} . \\ \hline
\end{array}
\end{gather}
For the triples on the first row of table \eqref{left right M} we have $\alpha=\beta=\gamma-1$ and for the triples on the second row we have $\alpha=-1$, $\beta=\gamma=0$ (one shows this using Corollaries \ref{nonvanishings} and \ref{one nonvanishing degree}).
\subsection{Proof that \texorpdfstring{$\mk{T}_{a}^{st}\cap \mk{T}_{b}^{st}=\emptyset$}{\space}} \label{empty intersection} \mbox{}\\
\textit{Recall the axioms of Bridgeland \cite{Bridg1}, that $\phi(A[1])=\phi(A)+1$ for any $A\in \sigma^{ss}$, and that $A, B \in \sigma^{ss}$ and $\phi(A)>\phi(B)$ imply $\hom(A,B)=0$.} We will use these axiom often implicitely.
We start with:
\begin{lemma} \label{mk T_12 cap mk T_43 1} $\left ( (M',\_,\_) \cup (\_,\_,M) \right ) \cap \left ( (M,\_,\_)\cup (\_,\_,M') \right )=\emptyset $. \end{lemma}
\bpr
Suppose $\sigma \in (a^{m},a^{m+1}, M) \cap ( M,b^{n},b^{n+1})$, then by the table \eqref{left right M} we obtain $\hom^1(b^{n+1},a^m)=0$ and $\hom(b^{n+1},a^{m+1})=0$, which contradicts \eqref{nonvanishing3}.
Suppose $\sigma \in (a^{m},a^{m+1}, M) \cap (b^{i},b^{i+1}, M')$, then by $\hom(M',a^{m})\neq 0$ (see \eqref{nonvanishing1}) and table \eqref{left right M} we obtain $ \phi\left ( b^{i}\right) <\phi(M')\leq\phi\left ( a^{m}\right) <\phi(M)$, which contradicts $\hom(M,b^i)\neq 0$ (see \eqref{nonvanishing1}).
Suppose $\sigma \in (M',a^{j},a^{j+1} ) \cap ( M,b^{n},b^{n+1})$, then by $\hom^1(a^{j+1},M)\neq 0$ (see \eqref{nonvanishing2}) and table \eqref{left right M} we obtain $ \phi(M')+1 <\phi\left ( a^{j+1}\right)\leq \phi(M)+1<\phi\left ( b^{n+1}\right) $, which contradicts $\hom^1(b^{n+1},M')\neq 0$.
Suppose $\sigma \in (M',a^{j},a^{j+1} ) \cap (b^{i},b^{i+1}, M')$, then by the table we have $\hom^1(a^{j+1},b^i)=0$, $\hom(a^{j+1},b^{i+1})=0$, which contradicts \eqref{nonvanishing4}. The lemma is proved.
\epr
\begin{lemma} \label{mk T_12 cap mk T_43 3} For any $p,q\in \ZZ$ we have $ (a^p,b^{p+1},a^{p+1}) \cap (b^q,a^q,b^{q+1})=\emptyset$.
\end{lemma}
\bpr Let $\sigma \in (a^p,b^{p+1},a^{p+1})$, then in table \eqref{no M} we see that $a^p ,a^{p+1} \in \sigma^{ss}$ and $\phi\left ( a^{p}\right)+1 < \phi\left (a^{p+1}\right)$. Now by Lemma \ref{from phases of big distance} we see that $a^q \not \in \sigma^{ss}$ for $q \not \in \{p,p+1\}$, and therefore $\sigma \not \in (b^q,a^q,b^{q+1})$ for $q \not \in \{p,p+1\}$.
Suppose that $\sigma \in (b^p,a^p,b^{p+1})$, then from table \eqref{no M} we obtain $\phi\left ( b^{p}\right)+1 <\phi\left ( a^{p}\right)+1 < \phi\left (a^{p+1}\right)$, hence $\hom^1(a^{p+1},b^p)=0$, which contradicts \eqref{nonvanishing4}.
Suppose that $\sigma \in (b^{p+1},a^{p+1},b^{p+2})$, then from table \eqref{no M} we obtain $\phi\left ( a^{p}\right)+1 < \phi\left (a^{p+1}\right)<\phi\left (b^{p+2}\right)$, hence $\hom^1(b^{p+2},a^p)=0$, which contradicts \eqref{nonvanishing3}. The lemma is proved.
\epr
\begin{lemma} \label{mk T_12 cap mk T_43 4} For any $p,q \in \ZZ$ we have:
$
\left ( (M',\_,\_) \cup (\_,\_,M)\right ) \cap (b^q,a^q,b^{q+1})=\emptyset $ and \\
$ \left ( (M,\_,\_) \cup (\_,\_,M')\right ) \cap (a^p,b^{p+1},a^{p+1}) =\emptyset $.
\end{lemma}
\bpr Assume first that $\sigma \in (b^q,a^q,b^{q+1})$, then table \eqref{no M} shows that:
\begin{gather}\label{big distance1} \phi(b^q)+1 <\phi(b^{q+1}).\end{gather}
Suppose that $\sigma \in (\_,\_,M)$, then using \eqref{big distance1}, $\hom(M,b^q)\neq 0$ and table \eqref{left right M} we see that $\phi(a^m)+1<\phi(b^{q+1})$ and $\phi(a^{m+1})<\phi(b^{q+1})$ for some $m\in \ZZ$, hence $\hom^1(b^{q+1},a^m)=\hom(b^{q+1},a^{m+1})=0$, which contradicts \eqref{nonvanishing3}.
Suppose that $\sigma \in (M',\_,\_)$, then using \eqref{big distance1}, $\hom^1(b^{q+1},M')\neq 0$ and table \eqref{left right M} we see that $\phi(b^q)+1<\phi(a^{j+1})$ and $\phi(b^{q})<\phi(a^{j})$ for some $j\in \ZZ$, hence $\hom^1(a^{j+1},b^q)=\hom(a^{j},b^{q})=0$, which contradicts \eqref{nonvanishing4}. So far we proved that $
\left ( (\_,\_,M) \cup (M',\_,\_) \right ) \cap (b^q,a^q,b^{q+1})=\emptyset $.
Assume now that $\sigma \in (a^p,b^{p+1},a^{p+1})$, then table \eqref{no M} shows that:
\begin{gather}\label{big distance2} \phi(a^p)+1 <\phi(a^{p+1}).\end{gather}
Suppose that $\sigma \in (\_,\_,M')$, then using \eqref{big distance2}, $\hom(M',a^p)\neq 0$ and table \eqref{left right M} we see that $\phi(b^i)+1<\phi(a^{p+1})$ and $\phi(b^{i+1})<\phi(a^{p+1})$ for some $i\in \ZZ$, hence $\hom^1(a^{p+1},b^i)=\hom(a^{p+1},b^{i+1})=0$, which contradicts \eqref{nonvanishing4}.
Suppose that $\sigma \in (M,\_,\_)$, then using \eqref{big distance2}, $\hom^1(a^{p+1},M)\neq 0$ and table \eqref{left right M} we see that $\phi(a^p)+1<\phi(b^{n+1})$ and $\phi(a^{p})<\phi(b^{n})$ for some $n\in \ZZ$, hence $\hom^1(b^{n+1},a^p)=\hom(b^{n},a^{p})=0$, which contradicts \eqref{nonvanishing3}. Thus, we proved the second equality as well.
\epr
Lemmas \ref{mk T_12 cap mk T_43 1}, \ref{mk T_12 cap mk T_43 3}, \ref{mk T_12 cap mk T_43 4}, and formulas \eqref{T_12new}, \eqref{T_43new} imply that $\mk{T}_{a}^{st}\cap \mk{T}_{b}^{st}=\emptyset$.
\subsection{The subsets \texorpdfstring{$(\_,\_,M), (\_,\_,M')$}{\space}, \texorpdfstring{$(M,\_,\_) \ \mbox{and} \ (M',\_,\_)$}{\space} are contractible
} \mbox{}\\
We start with:
\begin{lemma} \label{(_,_,X)0} Let $\{x^i\}_{i\in \ZZ}$ be either the sequence $\{a^i\}_{i\in \ZZ}$ or the sequence $\{b^i\}_{i\in \ZZ}$. If $m>j$ then: \begin{gather}(x^m, x^{m+1},X)\cap (x^j, x^{j+1},X) =\left \{ \sigma: \begin{array}{c} x^m \in \sigma^{ss}\\ x^{m+1} \in \sigma^{ss} \\ X \in \sigma^{ss}\end{array} , \ \begin{array}{c} 0 <\phi(x^{m+1}) - \phi(x^m) <1 \\ \phi(x^m) < \phi(X) \\ \phi(x^{m+1}) < \phi(X)+1 \end{array} \right \}, \nonumber \end{gather}
where $X=M$ if $\{x^i\}_{i\in \ZZ}=\{a^i\}_{i\in \ZZ}$, and $X=M'$ if $\{x^i\}_{i\in \ZZ}=\{b^i\}_{i\in \ZZ}$.
In particular, $(x^m, x^{m+1},X)\cap (x^j, x^{j+1},X)$ and $(x^m, x^{m+1},X)\cup (x^j, x^{j+1},X)$ are contractible. \end{lemma}\bpr
We show first the inclusion $\subset$. Assume that $\sigma \in (x^m, x^{m+1},X)\cap (x^j, x^{j+1},X)$ and $m>j$. Then $X,x^{m+1},x^m, $ $x^{j+1},x^j$ are all semistable and by table \eqref{left right M} we have
\begin{gather} \begin{array}{c} \phi ( x^m) < \phi (x^{m+1}) \\ \phi ( x^m) < \phi\left (X\right) \\ \phi ( x^{m+1}) < \phi\left (X\right)+1\end{array} \quad \begin{array}{c} \phi( x^j) < \phi (x^{j+1}) \\ \phi ( x^j) < \phi (X) \\ \phi ( x^{j+1} ) < \phi\left (X\right)+1\end{array}. \end{gather}
By $m>j$ it follows $\hom^1(x^{m+1}, x^j)\neq 0$, hence $\phi(x^{m+1})\leq \phi(x^j)+1$ (see Remark \ref{inc dec seq} (b)). On the other hand from the inequalities above we have $\phi(x^j)+1<\phi(x^{j+1})+1$ and by Remark \ref{inc dec seq} (a) we obtain $\phi(x^{j+1})+1\leq \phi(x^m)+1$. Thus we obtain $\phi(x^{m+1})< \phi(x^m)+1$ and $\subset$ follows.
We show now $\supset$. The condition defining the set on the right-hand side is the same as $\sigma \in (x^m, x^{m+1},X)$ and $ \phi(x^m) > \phi(x^{m+1}[-1])$ (see table \eqref{left right M}). From Proposition \ref{phi_1>phi_2} (a) it follows that $\mc A \cap \mc T_{exc}\subset \sigma^{ss}$, where $\mc A$ is the extension closure of $(x^m, x^{m+1}[-1])$, hence By Remark \ref{ext closure ab} we have $\{x^{j+1},x^j\}\subset \sigma^{ss}$. The inequality $0 <\phi(x^{m+1}) - \phi(x^m) <1$ and \eqref{phase formula} show that $Z(x^{m+1})$, $Z(x^m)$ are not collinear, hence by Corollary \ref{from equal phases to} (b) we get $\phi(x^{j+1}) \neq \phi(x^j)$. Now by Ramark \ref{inc dec seq} (a) and the incidence $\sigma \in (x^m, x^{m+1},X)$ we get: $ \begin{array}{c} \phi ( x^j) < \phi(x^{j+1}) \\ \phi ( x^j) \leq \phi ( x^m)< \phi (X) \\ \phi ( x^{j+1})\leq \phi ( x^{m+1}) < \phi (X)+1\end{array} $. In table \eqref{left right M} we see that $\sigma \in (x^j, x^{j+1}, X)$ and the inclusion $\supset$ is shown.
The proved equality implies that $(x^m, x^{m+1}, X) \cap (x^j, x^{j+1}, X)$ is contractible (see the arguments for the proof that \eqref{Theta_E n=2 3} is contractible in Proposition \ref{mutations}). Since $(x^m, x^{m+1}, X)$ and $ (x^j, x^{j+1}, X)$ are contractible, by Remark \ref{VK} we see that $(x^m, x^{m+1}, X) \cup (x^j, x^{j+1}, X)$ is contractible as well.
\epr
\begin{coro} \label{(_,_,X)} The subsets $(\_,\_,M)$ and $(\_,\_,M')$ of $\st(D^b(Q ))$ are contractible.
\end{coro}
\bpr Recalling \eqref{(_,_,M)} and using the notations of the previous lemma, we see that we have to show that $ \bigcup_{j\in \ZZ} (x^{j}, x^{j+1},X)$ is contractible. It is shown in Lemma \ref{(_,_,X)0}, that for a given $m\in \ZZ$ the intersection $(x^m, x^{m+1},X)\cap (x^j, x^{j+1},X)$ is contractible and it is the same for all $j<m$. Now by induction and using Remark \ref{VK} one shows that $\bigcup_{k=0}^n(x^{m-k}, x^{m-k+1},X)$ is contractible for any $n\in \NN$ and any $m\in \ZZ$. Using again Remark \ref{VK} we deduce that $ \bigcup_{j\in \ZZ} (x^{j}, x^{j+1},X)$ is contractible. The corollary follows.
\epr
The steps in the proof that $(M,\_,\_)$ and $(M',\_,\_)$ are analogous. We show first:
\begin{lemma} \label{(X,_,_)0} Let $\{x^i\}_{i\in \ZZ}$ be either the sequence $\{a^i\}_{i\in \ZZ}$ or the sequence $\{b^i\}_{i\in \ZZ}$ . If $m<j$, then: \begin{gather}\label{(X,_,_)a} (X,x^m, x^{m+1})\cap (X,x^j, x^{j+1}) =\left \{ \sigma:\begin{array}{c} X\in \sigma^{ss}\\ x^m\in \sigma^{ss} \\ x^{m+1} \in \sigma^{ss}\end{array}, \ \begin{array}{c} \phi(X)<\phi(x^m) \\ \phi(X) +1 < \phi(x^{m+1}) \\ 0 <\phi(x^{m+1}) - \phi(x^m) <1 \end{array} \right \} \end{gather}
where $X=M'$ if $\{x^i\}_{i\in \ZZ}=\{a^i\}_{i\in \ZZ}$, and $X=M$ if $\{x^i\}_{i\in \ZZ}=\{b^i\}_{i\in \ZZ}$.
In particular, $ (X,x^m, x^{m+1})\cap (X,x^j, x^{j+1})$ and $(X,x^m, x^{m+1})\cup (X,x^j, x^{j+1})$ are contractible. \end{lemma}
\bpr By table \eqref{left right M} we see that the condition defining the set on the right-hand side of \eqref{(X,_,_)a} is the same as $\sigma \in (X,x^m, x^{m+1})$ and $\phi(x^m)>\phi(x^{m+1}[-1])$.
The inclusion $\subset$ follows from table \eqref{left right M}, $\hom^1(x^{j+1},x^m) \neq 0$ and Remark \ref{inc dec seq} (a) as follows $\phi(x^{m+1})\leq \phi(x^j)<\phi(x^{j+1})\leq \phi(x^m)+1$.
To show the converse inclusion $\supset$ in \eqref{(X,_,_)a}, assume that $\sigma \in (X,x^m, x^{m+1})$ and $\phi(x^m)>\phi(x^{m+1}[-1])$. From Proposition \ref{phi_1>phi_2} (b) and Remark \ref{ext closure ab} it follows that $x^j, x^{j+1} \in \sigma^{ss}$. Since we have $0 <\phi(x^{m+1}) - \phi(x^m) <1$, it follows that $Z(x^m)$, $Z(x^{m+1})$ are not collinear, therefore by Corollary \ref{from equal phases to} (b) and Remark \ref{inc dec seq} (a) we obtain $\phi(x^j)<\phi(x^{j+1})$. Since $j>m$, by Remark \ref{inc dec seq} (a) we obtain also $\phi(X)<\phi(x^m)\leq \phi(x^j)$, $ \phi(X) +1 < \phi(x^{m+1}) \leq \phi(x^{j+1})$, hence $\sigma \in (X,x^j, x^{j+1})$.
The proved equality implies that $ (X,x^m, x^{m+1})\cap (X,x^j, x^{j+1})$ is contractible (see the arguments for the proof that \eqref{Theta_E n=2 3} is contractible in Proposition \ref{mutations}). Since $ (X,x^m, x^{m+1})$ and $(X,x^j, x^{j+1})$ are contractible, by Remark \ref{VK} we see that $ (X,x^m, x^{m+1})\cup (X,x^j, x^{j+1})$ is contractible as well.
\epr
\begin{coro} \label{(X,_,_)} The subsets $(M,\_,\_), (M',\_,\_)\subset \st(D^b(Q ))$ are contractible.
\end{coro}
\bpr Recalling \eqref{(M',_,_)} and using the notations of the previous lemma, we see that we have to show that $ \bigcup_{j\in \ZZ} (X,x^{j}, x^{j+1})$ is contractible. From Lemma \ref{(X,_,_)0} we know that for a given $m\in \ZZ$ the intersection $(X,x^m, x^{m+1})\cap (X,x^j, x^{j+1})$ is contractible and it is the same for all $j>m$. Now by induction and using Remark \ref{VK} one shows that $\bigcup_{k=0}^n(X,x^{m+k}, x^{m+k+1})$ is contractible for any $n\in \NN$ and any $m\in \ZZ$. Using again Remark \ref{VK} we deduce that $ \bigcup_{j\in \ZZ} (X,x^{j}, x^{j+1})$ is contractible. The corollary follows.
\epr
\section{The subsets \texorpdfstring{$\mk{T}_{a}^{st}$}{\space} and \texorpdfstring{$\mk{T}_{b}^{st}$}{\space} are contractible} \label{T_a T_b cont}
We start by showing some empty intersections:
\begin{lemma} \label{T12lemma1} The unions $\bigcup_{p\in \ZZ}(a^p,b^{p+1},a^{p+1})$ and $\bigcup_{p\in \ZZ}(b^p,a^p,b^{p+1})$ are disjoint. Furthermore, we have:
\begin{gather} \label{T_12c} p\neq q \ \Rightarrow \ (a^p,b^{p+1},a^{p+1}) \cap (a^q,a^{q+1},M)=(a^p,b^{p+1},a^{p+1}) \cap (M',a^q,a^{q+1})=\emptyset\\ \label{T_43c} p\neq q \ \Rightarrow \ (b^p,a^p,b^{p+1}) \cap (b^q,b^{q+1},M')=(b^p,a^p,b^{p+1}) \cap (M,b^q,b^{q+1})=\emptyset. \end{gather} \end{lemma}
\bpr
If $\sigma \in (a^p,b^{p+1},a^{p+1})$, then these exceptional objects are semistable and by table \eqref{no M} we have $\phi(a^p)+1 < \phi(a^{p+1})$. Now by Lemma \ref{from phases of big distance} we see that $a^{j}$ with $j\not \in \{p,p+1\}$ can not be semistable, therefore $\sigma \not \in (a^q,b^{q+1},a^{q+1})$, $\sigma \not \in (a^q,a^{q+1},M)$, and $\sigma \not \in (M',a^q,a^{q+1})$ for $q\neq p$.
If $\sigma \in (b^p,a^p,b^{p+1}) $, then $b^p,a^p,b^{p+1}$ are semistable and by table \eqref{no M} we have $\phi(b^p)+1 < \phi(b^{p+1})$. Now by Lemma \ref{from phases of big distance} we see that $b^{j}$ with $j\not \in \{p,p+1\}$ can not be semistable, therefore $\sigma \not \in (b^q,a^q,b^{q+1})$, $\sigma \not \in (b^q,b^{q+1},M')$, and $\sigma \not \in (M,b^q,b^{q+1})$ for $q\neq p$.
\epr
Now we attach the pairwise non-intersecting contractible blocks $\{ (a^p,b^{p+1},a^{p+1}) \}_{p\in\ZZ}$ to $(\_,\_,M)$ and to $(M',\_,\_)$
\begin{lemma} \label{T12lemma3} For any $p\in \ZZ$ the sets $ (a^p,b^{p+1},a^{p+1}) \cap (\_,\_,M)$; $ (a^p,b^{p+1},a^{p+1}) \cap (M',\_,\_)$; $ (b^p,a^p,b^{p+1}) \cap (\_,\_,M')$; and $ (b^p,a^p,b^{p+1}) \cap (M,\_,\_)$ are non-empty and contractible.
\end{lemma}
\bpr From \eqref{(M',_,_)}, \eqref{(_,_,M)} and \eqref{T_12c} it follows that: $(a^p,b^{p+1},a^{p+1}) \cap (\_,\_,M) = (a^p,b^{p+1},a^{p+1})\cap (a^{p}, a^{p+1}, M) $ and $(a^p,b^{p+1},a^{p+1})\cap (M',\_,\_) =(a^p,b^{p+1},a^{p+1})\cap (M',a^p,a^{p+1})$.
From \eqref{(M',_,_)}, \eqref{(_,_,M)} and \eqref{T_43c} it follows that: $(b^p,a^p,b^{p+1}) \cap (\_,\_,M') = (b^p,a^p,b^{p+1})\cap (b^{p}, b^{p+1}, M') $ and $ (b^p,a^p,b^{p+1}) \cap (M,\_,\_) =(b^p,a^p,b^{p+1})\cap (M,b^p,b^{p+1})$.
From Proposition \ref{mutations} it follows that $(a^p,b^{p+1},a^{p+1})\cap (a^{p}, a^{p+1}, M) $, $(a^p,b^{p+1},a^{p+1})\cap (M',a^p,a^{p+1})$, $(b^p,a^p,b^{p+1})\cap (b^{p}, b^{p+1}, M') $, and $(b^p,a^p,b^{p+1})\cap (M,b^p,b^{p+1})$ are contractible.
The lemma follows.
\epr
Let us denote:
\begin{gather} \label{T12Z} Z= (M',\_,\_) \cup \bigcup_{p\in \ZZ}(a^p,b^{p+1},a^{p+1}). \end{gather}
Corollary \ref{(X,_,_)} and Lemmas \ref{T12lemma1}, \ref{T12lemma3} imply (recall Remark \ref{VK}) that $Z$ is contractible. From \eqref{T_12new} and \eqref{(_,_,M)} we see that: \begin{gather} \label{T12=Zcup} \mk{T}_{a}^{st} =Z \cup (\_,\_,M)= Z \cup \bigcup_{m\in \ZZ} (a^{m}, a^{m+1}, M).\ \end{gather}
We start to glue the contractible summands in formula \eqref{T12=Zcup}. The first step is:
\begin{lemma} \label{T12Zcap(E_1)} The set $ (a^{m}, a^{m+1}, M) \cap Z$ consists of the stability conditions $\sigma$ for which $ a^{m},a^{m+1}, M $ are semistable and: \begin{gather} \label{T12Zcap(E_1)ineq} \begin{array}{c} \phi(a^{m})-1<\phi(a^{m+1}[-1])< \phi(a^{m})\\ \phi(a^{m})-1<\phi(M[-1])< \phi(a^{m}) \\ \arg_{(\phi(a^{m})-1,\phi(a^{m}))}(Z(a^{m})- Z(a^{m+1}))> \phi(M)-1 \end{array} \ \mbox{or} \ \begin{array}{c}\phi(M)<\phi(a^{m+1}) \\ \phi\left ( a^{m}\right) < \phi\left (a^{m+1}\right) \\ \phi\left ( a^{m}\right) < \phi\left (M\right) \\ \phi\left ( a^{m+1}\right) < \phi\left (M\right)+1\end{array}. \end{gather}
It follows that $ (a^{m}, a^{m+1}, M) \cap Z$ and $ (a^{m}, a^{m+1}, M) \cup Z$ are contractible.
\end{lemma}
\bpr In \eqref{T_12c} we have that $(a^{m}, a^{m+1}, M)\cap (a^{j}, b^{j+1} , a^{j+1})=\emptyset$ for $j\neq m$. Therefore (recall \eqref{T12Z})
\begin{gather}\label{T12Zcap(E_1)0} (a^{m}, a^{m+1}, M) \cap Z = (a^{m}, a^{m+1}, M) \cap ( (a^{m},b^{m+1} , a^{m+1})\cup (M',\_,\_)).\end{gather}
We show first the inclusion $\subset$. Assume that $\sigma \in (a^{m}, a^{m+1}, M) \cap Z$. Then $a^{m}, a^{m+1}, M$ are semistable and from table \eqref{left right M} we see that
\begin{gather}\label{T12Zcap(E_1)1} \begin{array}{c} \phi\left ( a^{m}\right) < \phi\left (a^{m+1}\right) \\ \phi\left ( a^{m}\right) < \phi\left (M\right) \\ \phi\left ( a^{m+1}\right) < \phi\left (M\right)+1\end{array}. \end{gather} Taking into account \eqref{T12Zcap(E_1)0} we consider two cases.
\ul{If $\sigma \in (a^{m},b^{m+1} , a^{m+1})$,} then $b^{m+1} \in \sigma^{ss}$ and in table \eqref{no M} we see that $\phi(b^{m+1})<\phi(a^{m+1})$. From $\hom(M,b^{m+1})\neq 0$ (see \eqref{nonvanishing1}) it follows that $\phi(M)\leq \phi(b^{m+1})<\phi(a^{m+1})$ and we obtain the second system of inequalities in \eqref{T12Zcap(E_1)ineq}.
\ul{If $\sigma \in (M', a^{j}, a^{j+1} )$,} then $M', a^{j}, a^{j+1} \in \sigma^{ss}$ and in table \eqref{left right M} we see that $\phi(M')+1<\phi(a^{j+1})$ and $\phi(a^{j})<\phi(a^{j+1})$. From $\hom^1(M,M')\neq 0$ it follows that $\phi(M)\leq \phi(M')+1<\phi(a^{j+1})$. Since $\phi(a^{m})<\phi(M)$, it follows from Remark \ref{inc dec seq} (a) that $m\leq j$.
If $m=j$, then $\phi(M)<\phi(a^{m+1})$ and we obtain the second system of inequalities.
If $m<j$, then we show that the first system of inequalities in \eqref{T12Zcap(E_1)ineq} holds. Now $\phi(M)<\phi(a^{j+1})$ and $\hom^1(a^{j+1}, a^{m})\neq 0$, hence $\phi(a^{m+1})\leq \phi(a^{j})<\phi(a^{j+1})\leq \phi(a^{m})+1$ and $\phi(M)<\phi(a^{j+1})\leq \phi(a^{m})+1$. We have also $M'\in \sigma^{ss}$ and by $\hom^1(M,M')\neq 0$ it follows that $\phi(M)\leq \phi(M')+1$.
From $\hom(M',a^{m})\neq 0$ it follows $\phi(M')\leq a^{m}$. These arguments together with \eqref{T12Zcap(E_1)1} imply:
\begin{gather} \label{T12Zcap(E_1)3} \begin{array}{c} \phi(a^{m})-1<\phi(a^{m+1}[-1])< \phi(a^{m});\\ \phi(a^{m})-1<\phi(M[-1])\leq \phi(M') \leq \phi(a^{m}) ; \ \ \ \ \ \phi(M[-1]) < \phi(a^{m}). \end{array} \end{gather}
In \eqref{Z(delta)} we have $Z(a^{m})-Z(a^{m+1})=Z(\delta)$, therefore it remains to show that:
\begin{gather} \label{T12Zcap(E_1)4} \arg_{(\phi(a^{m})-1,\phi(a^{m}))}(Z(\delta))> \phi(M)-1. \end{gather}
From the second row of \eqref{T12Zcap(E_1)3} and \eqref{phase formula} we see that $Z(\delta)$ and $Z(M[-1])$ both lie in the half-plane\footnote{The notation $Z(a^{m})^c_-$ is explained in \eqref{complement of a line}.} $Z(a^{m})^c_-$. In \eqref{[delta]} we have aslo $Z(M')=Z(\delta) + Z(M[-1])$, therefore the vector $Z(M')$ is in $Z(a^{m})^c_-$ as well, hence by $Z(M')=Z(\delta) + Z(M[-1])$ it follows that the inequality \eqref{T12Zcap(E_1)4} is equivalent to $\phi(M')>\phi(M[-1]) $. Therefore it remains to show that $\phi(M') \neq \phi(M[-1]) $. Indeed, on one hand $\phi(M[-1]) = \phi(M')$ implies $ \arg_{(\phi(a^{m})-1,\phi(a^{m}))}(Z(\delta))= \phi(M')$. On the other hand,
$\sigma \in (M', a^{j}, a^{j+1} )$, $m<j$ and \eqref{T12Zcap(E_1)1} imply $\phi(M')+1< \phi(a^{j+1})\leq \phi(a^{m})+1 < \phi(M)+1\leq\phi(M')+2$. Thus, we see that $\phi(M[-1]) = \phi(M')$ implies $Z(a^{j+1}) \in Z(\delta)^c_-$. However, from the first inequality in \eqref{T12Zcap(E_1)3} and Corollary \ref{from equal phases to} (d) it follows that $Z(a^{j+1}) \in Z(\delta)^c_+$, which is a contradiction, and \eqref{T12Zcap(E_1)4} follows.
So far we showed that $\sigma \in (a^{m}, a^{m+1}, M) \cap Z$ implies \eqref{T12Zcap(E_1)ineq}. We show now converse inclusion.
We assume first that the second system of inequalities in \eqref{T12Zcap(E_1)ineq} holds. In particular $\sigma \in (a^{m}, a^{m+1}, M)$. By the inequality $\phi(M)<\phi(a^{m+1}) $ we can apply Proposition \ref{phi_1>phi_2} (b), hence the triangle \eqref{short filtration 2} implies that $b^{m+1} \in \sigma^{ss}$, $\phi(M)\leq \phi(b^{m+1})\leq \phi(a^{m+1})$, and $Z(M)+Z(a^{m+1})=Z(b^{m+1})$. We have in \eqref{T12Zcap(E_1)ineq} also $\phi(a^{m+1})-1<\phi(M)<\phi(a^{m+1})$ and it follows that $\phi(M)<\phi(b^{m+1})<\phi(a^{m+1})$. If the inequality $\phi(a^{m+1})> \phi(a^{m})+1$ holds, then due to $\phi(M)<\phi(b^{m+1})<\phi(a^{m+1})$ and $\phi(a^{m})<\phi(M)$ hold we obtain that $\sigma \in (a^{m},b^{m+1},a^{m+1})\subset Z$ (see table \eqref{no M}).
Thus, we can assume that $\phi(a^{m+1}[-1])\leq \phi(a^{m})$ and combining with the inequalities $\phi(M[-1])<\phi(a^{m+1}[-1])<\phi(M)$, $\phi(a^{m})<\phi(M)$ (given in \eqref{T12Zcap(E_1)ineq}) we get $\phi(M[-1])< \phi(a^{m+1}[-1])\leq \phi(a^{m})<\phi(M)$. Now it is easy to show with the help of Corollaries \ref{nonvanishings} and \ref{one nonvanishing degree} that $(a^{m}, a^{m+1}[-1], M[-1])$ is a $\sigma$-triple (see Definition \ref{sigma triple}).
Combining the triangles \eqref{short filtration 1} and \eqref{short filtration 2} we get the following sequence: \be \label{T12Zcap(E_1)the filtation}
\begin{diagram}[size=1em]
0 & \rTo & & & M[-1] & \rTo & & & b^{m+1}[-1] & \rTo & & & M' \\
& \luDashto & & \ldTo & &\luDashto & & \ldTo & & \luDashto & & \ldTo & \\
& & M[-1] & & & & a^{m+1}[-1] & & & & a^{m} &
\end{diagram}.
\ee
The conditions of Lemma \ref{three comp factors} (b) are satisfied with the triple $(a^{m}, a^{m+1}[-1], M[-1])$ and the diagram above. Therefore $M'\in \sigma^{ss}$ and $\phi(M')<\phi(a^{m})$.
If $\phi(a^{m+1}[-1])=\phi(a^{m})$, then it follows that $\phi(M')+1<\phi(a^{m+1})$, and recalling that we have also $\phi(a^{m})<\phi(a^{m+1})$ we see that $\sigma \in (M', a^{m}, a^{m+1} )\subset Z$ (see table \eqref{left right M}).
Therefore we can assume that $\phi(M[-1])< \phi(a^{m+1}[-1]) < \phi(a^{m})<\phi(M)$. We will show in this case that $\sigma \in (M',a^{j}, a^{j+1})$ for some big enough $j$. From Proposition \ref{all exc semistable} and Remark \ref{ext closure ab} it follows that $\{a^{j+1}\}_{j\in \ZZ}\subset \sigma^{ss}$. From $\phi(a^{m})<\phi(a^{m+1}) < \phi(a^{m})+1 $ and Corollary \ref{from equal phases to} (b) and (d) we see that $\phi(a^{j})<\phi(a^{j+1})$ and $Z(a^j)\in Z(\delta)_+^c$ for each $j$ (recall also Remark \ref{inc dec seq} (a)).
We will show that for big enough $j$ we have $\phi(M')+1< \phi(a^j) $ and then from table \eqref{left right M} we obtain $\sigma \in (M',a^{j}, a^{j+1})\subset Z$.
Now we have $ \phi(a^{m}[-1])<\phi(M[-1])< \phi(a^{m+1}[-1]) < \phi(a^{m})$. Recalling that $Z(\delta)=Z(a^{m})+Z(a^{m+1}[-1])$, we see that we can choose $t\in \RR$ so that $t< \phi(a^{m})<\phi(M)< \phi(a^{m+1})<t+1 $ and $Z(\delta)=\abs{Z(\delta)} \exp(\ri \pi t)$. Since $\hom^1(a^j,a^{m})\neq 0$, $\hom(a^{m},a^j)\neq 0$ for $j>m+1$ and by Corollary \ref{from equal phases to} (b), we have $\phi(a^{m})<\phi(a^j)< \phi(a^{m})+1 $ for $j>m+1$. These inequalities together with the incidence $Z(a^j)\in Z(\delta)_+^c$ imply that $\arg_{(t,t+1)}(Z(a^j))=\phi(a^j)$ for $j>m+1$ (see Remark \ref{arg remark} (c)). Now the formula \eqref{noncolinear ab2} in Corollary \ref{noncolinear ab} gives us the following equality:
\begin{gather}\label{T12Zcap(E_1)the limit} \lim_{j\rightarrow \infty} \phi(a^j)= t+1.\end{gather}
We showed that $\phi(M')<\phi(a^{m})$ (see below \eqref{T12Zcap(E_1)the filtation}) and we have also $\phi(a^{m})<\phi(M)$. Using $\hom^1(M,M')\neq 0$ we see that $\phi(a^{m}[-1])<\phi(M')<\phi(a^{m})$. We showed also that $t< \phi(a^{m})<\phi(M)< \phi(a^{m+1})<t+1 $. Since $Z(M)+Z(M')=Z(\delta)=\abs{Z(\delta)} \exp(\ri \pi t)$, it follows that $Z(M')\in Z(\delta)^c_-$ and $\phi(M')<t$. By \eqref{T12Zcap(E_1)the limit} we get the desired $\phi(a^j)>\phi(M')+1$ for big $j$.
So, we showed that the second system of inequalities in \eqref{T12Zcap(E_1)ineq} implies that $\sigma\in (a^{m}, a^{m+1}, M) \cap Z$.
We show now that the first system in \eqref{T12Zcap(E_1)ineq} implies $\sigma\in (a^{m}, a^{m+1}, M) \cap Z$ as well.
Assume that $a^{m}, a^{m+1}, M \in \sigma^{ss}$ and that these inequalities hold. They contain the inequalities defining $(a^{m}, a^{m+1}, M)$ (see table \eqref{left right M}), therefore we obtain $\sigma \in (a^{m}, a^{m+1}, M)$ immediately. Furthermore the first two inequalities show that $(a^{m}, a^{m+1}[-1], M[-1])$ is a $\sigma$-triple.
The conditions of Lemma \ref{three comp factors} (a) are satisfied with the triple $(a^{m}, a^{m+1}[-1], M[-1])$ and the diagram \eqref{T12Zcap(E_1)the filtation}. Therefore $M'\in \sigma^{ss}$ and $\phi(M')<\phi(a^{m})$. By $\hom^1(M,M')\neq 0$ we can write
$\phi(a^{m})-1< \phi(M[-1])\leq \phi(M') < \phi(a^{m})$, hence by \eqref{phase formula} we see that $Z(\delta), Z(M[-1]), Z(M') \in Z(a^{m})^c_-$. Let us denote $t=\arg_{(\phi(a^{m})-1,\phi(a^{m}))}(Z(\delta))$. The third inequality in \eqref{T12Zcap(E_1)ineq} is the same as $t> \phi(M)-1 $. Combining these arguments with the equality $Z(M')=Z(\delta)+Z(M[-1])$ we write:
\begin{gather} \label{T12Zcap(E_1) help} \phi(a^{m}[-1])< \phi(M[-1])< \phi(M') <t< \phi(a^{m}). \end{gather}
We will show that $\sigma \in (M', a^{j}, a^{j+1} )$ for some big enough $j$. We have $\phi(a^{m})<\phi(a^{m+1}) < \phi(a^{m})+1 $ (the first inequality in \eqref{T12Zcap(E_1)ineq}), which by Proposition \ref{all exc semistable} and Remark \ref{ext closure ab} implies that $\{a^{j+1}\}_{j\in \ZZ}\subset \sigma^{ss}$, and by Corollary \ref{from equal phases to} (b), (d) implies that $\phi(a^{j})<\phi(a^{j+1})$ and $Z(a^j)\in Z(\delta)_+^c$ for each $j$. Since $\hom^1(a^j,a^{m})\neq 0$, $\hom(a^{m},a^j)\neq 0$ for $j>m+1$, we have $\phi(a^{m})<\phi(a^j)< \phi(a^{m})+1 $ for $j>m+1$. These inequalities together with the incidences $Z(a^j)\in Z(\delta)_+^c$, $\phi(a^m)\in(t,t+1)$ imply that $\arg_{(t,t+1)}(Z(a^j))=\phi(a^j)$ for $j>m+1$ (see Remark \ref{arg remark} (c)). Now the formula \eqref{noncolinear ab2} in Corollary \ref{noncolinear ab} leads to \eqref{T12Zcap(E_1)the limit} again. Therefore by \eqref{T12Zcap(E_1) help} we see that $\phi(a^j)>\phi(M')+1$ for big enough $j$. It follows that $\sigma \in (M', a^{j}, a^{j+1} )\subset Z$ (see table \eqref{left right M}) .
The first part of the lemma is shown.
It is easy now to show that the intersection is contractible. The intersection in quiestion is the same as $ (a^{m}, a^{m+1}[-1], M[-1]) \cap Z$. Let us denote $\mc E= (a^{m}, a^{m+1}[-1], M[-1])$. We have a homeomorphism ${f_{\mc E}}_{\vert \Theta_{\mc E}}:\Theta_{\mc E} \rightarrow f_{\mc E}(\Theta_{\mc E})$ (see \eqref{A and homeo}, \eqref{the map}). The proved description of $Z\cap \Theta_{\mc E}$ by the inequalities \eqref{T12Zcap(E_1)ineq} shows that $f_{\mc E}(Z\cap \Theta_{\mc E}) $ is union of two sets. The first set after permutation of the coordinates in $\RR^6$ is the same as the set considered in Corollary \ref{contractible 1b}, hence it is also contractible. The second is obviously contractible. Furthermore, one easily shows that the intersection of these two sets is
$ \RR_{>0}^3 \times \left \{ \phi_0-1 < \phi_2 <\phi_1 < \phi_0 \right \},$ which is contractible as well. Now by Remark \ref{VK} it follows that $f_{\mc E}(Z\cap \Theta_{\mc E}) $ is contractible, therefore $Z\cap \Theta_{\mc E}$ is contractible as well. Recalling that $Z$ and $ (a^{m}, a^{m+1}, M) $ are contractible and applying again Remark \ref{VK} we deduce that $ (a^{m}, a^{m+1}, M) \cup Z$ is contractible.
The lemma is proved.
\epr
\begin{coro} \label{T12Zcap union(E_1)} The set $\mk{T}_{a}^{st}$ is contractible.
\end{coro}
\bpr Recall that $\mk{T}_{a}^{st}= Z \cup \bigcup_{j\in \NN} (a^{j}, a^{j+1}, M) $ (see \eqref{T12=Zcup}). We will show that $Z \cup \bigcup_{j=0}^n (a^{m-j+1}, a^{m-j}, M) $ is contractible for each $m\in \ZZ$ and each $n\in \NN$. Then the corollary follows from Remark \ref{VK}.
Assume that for some $n\in \NN$ the set $Z \cup \bigcup_{j=0}^n (a^{m-j+1}, a^{m-j}, M) $ is contractible for each $m\in \ZZ$. We have shown this statement for $n=0$ in Lemma \ref{T12Zcap(E_1)}, and now we make induction assumption.
Take any $m\in \NN$ and consider $Z \cup \bigcup_{j=0}^{n+1} (a^{m-j+1}, a^{m-j}, M)$ $= \left (Z \cup \bigcup_{j=1}^{n+1} (a^{m-j+1}, a^{m-j}, M)\right ) \cup (a^{m}, a^{m+1}, M)$. By the induction assumption $Z \cup \bigcup_{j=1}^{n+1} (a^{m-j+1}, a^{m-j}, M) $ and $(a^{m}, a^{m+1}, M)$ are contractible. We will show now that the intersection of these sets is contractible as well and then by Remark \ref{VK} we obtain that the union $Z \cup \bigcup_{j=0}^{n+1} (a^{m-j+1}, a^{m-j}, M)$ is contractible. Indeed, we have
\begin{gather} \left (Z \cup \bigcup_{j=1}^{n+1} (a^{m-j+1}, a^{m-j}, M) \right ) \cap (a^{m}, a^{m+1}, M) = \nonumber \\[-2mm] \label{T12Zcap union(E_1)1} \\[-2mm]\left ( (a^{m}, a^{m+1}, M) \cap Z \right ) \cup \left ( (a^{m}, a^{m+1}, M) \cap \bigcup_{j=1}^{n+1} (a^{m-j+1}, a^{m-j}, M) \right ).\nonumber \end{gather}
Using Lemmas \ref{T12Zcap(E_1)} and \ref{(_,_,X)0} we deduce that the considered intersection consists of the stability conditions for which
$a^{m}, a^{m+1}, M$ are semi-stable and some of the two systems of inequalities in \eqref{T12Zcap(E_1)ineq} or the system $\begin{array}{c} \phi(a^{m})<\phi(a^{m+1})<\phi(a^{m})+1 \\ \phi(a^{m})<\phi(M) \\ \phi(a^{m+1})<\phi(M)+1\end{array}$ holds. Since the first system in \eqref{T12Zcap(E_1)ineq} implies the last system we deduce that the intersection \eqref{T12Zcap union(E_1)1} is described by the inequalities:
\begin{gather} \begin{array}{c} \phi(a^{m})<\phi(a^{m+1})<\phi(a^{m})+1 \\ \phi(a^{m})<\phi(M) \\ \phi(a^{m+1})<\phi(M)+1\end{array} \ \mbox{or} \ \begin{array}{c} \phi\left ( a^{m}\right) < \phi\left (a^{m+1}\right) \\ \phi\left ( a^{m}\right) < \phi\left (M\right) \\ \phi(M)< \phi\left ( a^{m+1}\right) < \phi\left (M\right)+1\end{array}. \end{gather}
Now analogous arguments as in the last paragraph of the proof of Lemma \ref{T12Zcap(E_1)} show that the intersection \eqref{T12Zcap union(E_1)1} is contractible. The corollary follows.
\epr
We pass to the proof that $ \mk{T}_b^{st}$ is contractible. Let us denote
\begin{gather} \label{T43Z} W= (M,\_,\_) \cup \bigcup_{p\in \ZZ}(b^p,a^{p},a^{p+1}). \end{gather}
Corollary \ref{(X,_,_)} and Lemmas \ref{T12lemma1}, \ref{T12lemma3} imply (recall Remark \ref{VK}) that $W$ is contractible. From \eqref{T_43new} and \eqref{(_,_,M)} we see that: \begin{gather} \label{T43=Zcup} \mk{T}_{b}^{st} =W \cup (\_,\_,M')= W \cup \bigcup_{m\in \ZZ} (b^{m}, b^{m+1}, M').\ \end{gather}
The proof of the next Lemma \ref{T43Zcap(E_1)} is analogous to the proof of Lemma \ref{T12Zcap(E_1)}):
\begin{lemma} \label{T43Zcap(E_1)} The set $ (b^{m}, b^{m+1}, M') \cap W$ consists of the stability conditions $\sigma$ for which $ b^{m},b^{m+1}, M' $ are semistable and \begin{gather} \label{T43Zcap(E_1)ineq} \begin{array}{c} \phi(b^{m})-1<\phi(b^{m+1}[-1])< \phi(b^{m})\\ \phi(b^{m})-1<\phi(M'[-1])< \phi(b^{m}) \\ \arg_{(\phi(b^{m})-1,\phi(a^{m}))}(Z(b^{m})- Z(b^{m+1}))> \phi(M')-1 \end{array} \ \mbox{or} \ \begin{array}{c}\phi(M')<\phi(b^{m+1}) \\ \phi\left ( b^{m}\right) < \phi\left (b^{m+1}\right) \\ \phi\left ( b^{m}\right) < \phi\left (M'\right) \\ \phi\left ( b^{m+1}\right) < \phi\left (M'\right)+1\end{array}. \end{gather}
It follows that $ (b^{m}, b^{m+1}, M') \cap W$ and $ (b^{m}, b^{m+1}, M') \cup W$ are contractible.
\end{lemma}
\bpr In \eqref{T_43c} we have that $(b^{m}, b^{m+1}, M')\cap (b^{j}, a^{j} , b^{j+1})=\emptyset$ for $j\neq m$. Therefore (recall \eqref{T43Z})
\begin{gather}\label{T43Zcap(E_1)0} (b^{m}, b^{m+1}, M') \cap W = (b^{m}, b^{m+1}, M') \cap ( (b^{m},a^{m} , b^{m+1})\cup (M,\_,\_)).\end{gather}
We show first the inclusion $\subset$. Assume that $\sigma \in (b^{m}, b^{m+1}, M') \cap W$. Then $b^{m}, b^{m+1}, M'$ are semistable and from table \eqref{left right M} we see that
\begin{gather}\label{T43Zcap(E_1)1} \begin{array}{c} \phi\left ( b^{m}\right) < \phi\left (b^{m+1}\right) \\ \phi\left ( b^{m}\right) < \phi\left (M'\right) \\ \phi\left ( b^{m+1}\right) < \phi\left (M'\right)+1\end{array}. \end{gather} Taking into account \eqref{T43Zcap(E_1)0} we consider two cases.
\ul{If $\sigma \in (b^{m},a^{m} , b^{m+1})$,} then $a^{m} \in \sigma^{ss}$ and $\phi(a^{m})<\phi(b^{m+1})$ (see table \eqref{no M}). From $\hom(M',a^{m})\neq 0$ (see \eqref{nonvanishing1}) it follows $\phi(M')\leq \phi(a^{m})<\phi(b^{m+1})$ and we get the second system in \eqref{T43Zcap(E_1)ineq}.
\ul{If $\sigma \in (M, b^{j}, b^{j+1} )$,} then $M, b^{j}, b^{j+1}\in \sigma^{ss}$ and in table \eqref{left right M} we see that $\phi(M)+1<\phi(b^{j+1})$ and $\phi(b^{j})<\phi(b^{j+1})$. From $\hom^1(M',M)\neq 0$ it follows that $\phi(M')\leq \phi(M)+1<\phi(b^{j+1})$. Since we have $\phi(b^{m})<\phi(M')$, Remark \ref{inc dec seq} (a) implies that $m\leq j$.
If $m=j$, then $\phi(M')<\phi(b^{m+1})$ and we obtain the second system of inequalities.
If $m<j$, then we will show that the first system of inequalities in \eqref{T43Zcap(E_1)ineq} holds. Now $\phi(M')<\phi(b^{j+1})$ and $\hom^1(b^{j+1}, b^{m})\neq 0$, hence $\phi(b^{m+1})\leq \phi(b^{j})<\phi(b^{j+1})\leq \phi(b^{m})+1$, $\phi(M')<\phi(b^{j+1})\leq \phi(b^{m})+1$. We have also $M\in \sigma^{ss}$ and by $\hom^1(M',M)\neq 0$ and $\hom(M,b^{m})\neq 0$ it follows that $\phi(M')\leq \phi(M)+1$ and $\phi(M)\leq b^{m}$. These arguments together with \eqref{T43Zcap(E_1)1} imply
\begin{gather} \label{T43Zcap(E_1)3} \begin{array}{c} \phi(b^{m})-1<\phi(b^{m+1}[-1])< \phi(b^{m})\\ \phi(b^{m})-1<\phi(M'[-1])\leq \phi(M) \leq \phi(b^{m}) ; \ \ \ \ \ \phi(M'[-1]) < \phi(b^{m}) \end{array}. \end{gather}
Due to \eqref{Z(delta)}, to show the first system in \eqref{T43Zcap(E_1)ineq} it remains to derive the following inequality:
\begin{gather} \label{T43Zcap(E_1)4} \arg_{(\phi(b^{m})-1,\phi(b^{m}))}(Z(\delta))> \phi(M')-1. \end{gather}
From \eqref{T43Zcap(E_1)3} we see that $Z(\delta)$ and $Z(M'[-1])$ both lie in the half-plane\footnote{The notation $Z(b^{m})^c_-$ is explained in \eqref{complement of a line}.} $Z(b^{m})^c_-$. In \eqref{[delta]} we have aslo $Z(M)=Z(\delta) + Z(M'[-1])$, therefore the vector $Z(M)$ is in $Z(b^{m})^c_-$ as well. Now the equality $Z(M)=Z(\delta) + Z(M'[-1])$ implies that \eqref{T43Zcap(E_1)4} is equivalent to $\phi(M)>\phi(M'[-1]) $. Hence we have to show that $\phi(M) \neq \phi(M'[-1]) $. Indeed, on one hand $\phi(M'[-1]) = \phi(M)$ implies $ \arg_{(\phi(b^{m})-1,\phi(b^{m}))}(Z(\delta))= \phi(M)$. On the other hand,
$\sigma \in (M, b^{j}, b^{j+1} )$, $m<j$ and \eqref{T43Zcap(E_1)1} imply $\phi(M)+1< \phi(b^{j+1})\leq \phi(b^{m})+1 < \phi(M')+1\leq\phi(M)+2$. Thus, we see that $\phi(M'[-1]) = \phi(M)$ implies $Z(b^{j+1}) \in Z(\delta)^c_-$. However, from the first inequality in \eqref{T43Zcap(E_1)3} and Corollary \ref{from equal phases to} (d) it follows that $Z(b^{j+1}) \in Z(\delta)^c_+$, which is a contradiction, and \eqref{T43Zcap(E_1)4} follows.
So far we showed the inclusion $\subset$. We show now the inverse inclusion $\supset$.
We assume first that the second system of inequalities in \eqref{T43Zcap(E_1)ineq} holds. In particular $\sigma \in (b^{m}, b^{m+1}, M')$. By the inequality $\phi(M')<\phi(b^{m+1}) $ we can apply Proposition \ref{phi_1>phi_2} (b), hence the short exact sequence \eqref{short filtration 1} implies that $a^{m} \in \sigma^{ss}$ and $Z(M')+Z(b^{m+1})=Z(a^{m})$. We have also $\phi(b^{m+1})-1<\phi(M')<\phi(b^{m+1})$, and it follows that $\phi(M')<\phi(a^{m})<\phi(b^{m+1})$. If the inequality $\phi(b^{m+1})> \phi(b^{m})+1$ holds, then recalling that $\phi(b^{m})<\phi(M')$ we obtain that $\sigma \in (b^{m},a^{m},b^{m+1})\subset W$ (see table \eqref{no M}).
Therefore we reduce to the inequality $\phi(b^{m+1}[-1])\leq \phi(b^{m})$. Combining with $\phi(M'[-1])<\phi(b^{m+1}[-1])<\phi(M')$ and $\phi(b^{m})<\phi(M')$, we can write $\phi(M'[-1])< \phi(b^{m+1}[-1])\leq \phi(b^{m})<\phi(M')$ and then $(b^{m}, b^{m+1}[-1], M'[-1])$ is a $\sigma$-triple (see Definition \ref{sigma triple}).
Combining the triangles \eqref{short filtration 1} and \eqref{short filtration 2} we obtain the following sequence of triangles in $\mc T$: \be \label{T43Zcap(E_1)the filtation}
\begin{diagram}[size=1em]
0 & \rTo & & & M'[-1] & \rTo & & & a^{m}[-1] & \rTo & & & M \\
& \luDashto & & \ldTo & &\luDashto & & \ldTo & & \luDashto & & \ldTo & \\
& & M'[-1] & & & & b^{m+1}[-1] & & & & b^{m} &
\end{diagram}.
\ee
The conditions of Lemma \ref{three comp factors} (b) are satisfied with the triple $(b^{m}, b^{m+1}[-1], M'[-1])$ and the diagram above. Therefore $M\in \sigma^{ss}$ and $\phi(M)<\phi(b^{m})$.
If $\phi(b^{m+1}[-1])=\phi(b^{m})$, then we have also $\phi(M)+1<\phi(b^{m+1})$, and recalling that we have also $\phi(b^{m})<\phi(b^{m+1})$ we see that $\sigma \in (M, b^{m}, b^{m+1} )\subset W$ (see table \eqref{left right M}).
Therefore we can assume that $\phi(M'[-1])< \phi(b^{m+1}[-1]) < \phi(b^{m})<\phi(M')$. We will that $\sigma \in (M,b^{j}, b^{j+1})$ for some big $j$ in this case. Proposition \ref{all exc semistable} and Remark \ref{ext closure ab} ensure that $\{b^{j+1}\}_{j\in \ZZ}\subset \sigma^{ss}$. From $\phi(b^{m})<\phi(b^{m+1}) < \phi(b^{m})+1 $ and Corollary \ref{from equal phases to} (b) and (d) we see that $\phi(b^{j})<\phi(b^{j+1})$ and $Z(b^j)\in Z(\delta)_+^c$ for each $j$.
Now to show that $\sigma \in (M,b^{j}, b^{j+1})\subset W$ it is enough to derive $\phi(M)+1< \phi(b^j) $ for big enough $j$ (see table \eqref{left right M}).
Since we have $ \phi(b^{m}[-1])<\phi(M'[-1])< \phi(b^{m+1}[-1]) < \phi(b^{m})$ and $Z(\delta)=Z(b^{m})+Z(b^{m+1}[-1])$, we see that we can choose $t\in \RR$ so that $t< \phi(b^{m})<\phi(M')< \phi(b^{m+1})<t+1 $ and $Z(\delta)=\abs{Z(\delta)} \exp(\ri \pi t)$. Since $\hom^1(b^j,b^{m})\neq 0$, $\hom(b^{m},b^j)\neq 0$ for $j>m+1$ and by Corollary \ref{from equal phases to} (b), we have $\phi(b^{m})<\phi(b^j)< \phi(b^{m})+1 $ for $j>m+1$. These inequalities together with the incidence $Z(b^j)\in Z(\delta)_+^c$ imply that $\arg_{(t,t+1)}(Z(b^j))=\phi(b^j)$ for $j>m+1$ (see Remark \ref{arg remark} (c)). The formula \eqref{noncolinear ab2} in Corollary \ref{noncolinear ab} gives us the following:
\begin{gather}\label{T43Zcap(E_1)the limit} \lim_{j\rightarrow \infty} \phi(b^j)= t+1.\end{gather}
We showed that $\phi(M)<\phi(b^{m})$ (see below \eqref{T43Zcap(E_1)the filtation}) and we have also $\phi(b^{m})<\phi(M')$. From $\hom^1(M',M)\neq 0$ we derive $\phi(b^{m}[-1])<\phi(M)<\phi(b^{m})$. We showed also that $t< \phi(b^{m})<\phi(M')< \phi(b^{m+1})<t+1 $. Since $Z(M)+Z(M')=Z(\delta)=\abs{Z(\delta)} \exp(\ri \pi t)$, it follows that $Z(M)\in Z(\delta)^c_-$ and $\phi(M)<t$. Now \eqref{T43Zcap(E_1)the limit} ensures that $\phi(b^j)>\phi(M)+1$ for big enough $j$.
So far we showed that the second system of inequalities in \eqref{T43Zcap(E_1)ineq} implies $\sigma\in (b^{m}, b^{m+1}, M') \cap W$.
We pass to the first system of inequalities in \eqref{T43Zcap(E_1)ineq}.
So assume that $b^{m}, b^{m+1}, M' \in \sigma^{ss}$ and that these inequalities hold. They contain the inequalities defining $(b^{m}, b^{m+1}, M')$ (see table \eqref{left right M}), hence $\sigma \in (b^{m}, b^{m+1}, M')$. Furthermore, the first two inequalities show that $(b^{m}, b^{m+1}[-1], M'[-1])$ is a $\sigma$-triple and that the conditions of Lemma \ref{three comp factors} (a) are satisfied with this triple and the diagram \eqref{T43Zcap(E_1)the filtation}. Therefore $M\in \sigma^{ss}$ and $\phi(M)<\phi(b^{m})$. By $\hom^1(M',M)\neq 0$ we can write
$\phi(b^{m})-1< \phi(M'[-1])\leq \phi(M) < \phi(b^{m})$ (we use also \eqref{T43Zcap(E_1)ineq}), hence $Z(\delta), Z(M'[-1]), Z(M) \in Z(b^{m})^c_-$. Let us denote $t=\arg_{(\phi(b^{m})-1,\phi(b^{m}))}(Z(\delta))$. The third inequality in \eqref{T43Zcap(E_1)ineq} is the same as $t> \phi(M')-1 $. Combining these arguments with the equality $Z(M)=Z(\delta)+Z(M'[-1])$ we deduce that:
\begin{gather} \label{T43Zcap(E_1) help} \phi(b^{m}[-1])< \phi(M'[-1])< \phi(M) <t< \phi(b^{m}). \end{gather}
We will show that $\sigma \in (M, b^{j}, b^{j+1} )$ for some big enough $j$. We have $\phi(b^{m})<\phi(b^{m+1}) < \phi(b^{m})+1 $ (the first inequality in \eqref{T43Zcap(E_1)ineq}), which by Proposition \ref{all exc semistable} and Remark \ref{ext closure ab} implies that $\{b^{j+1}\}_{j\in \ZZ}\subset \sigma^{ss}$, and by Corollary \ref{from equal phases to} (b), (d) implies that $\phi(b^{j})<\phi(b^{j+1})$ and $Z(b^j)\in Z(\delta)_+^c$ for each $j$. Using Remark \ref{inc dec seq} one easily shows that $\phi(b^{m})<\phi(b^j)< \phi(b^{m})+1 $ for $j>m+1$. These inequalities together with the incidences $Z(b^j)\in Z(\delta)_+^c$, $\phi(b^m)\in(t,t+1)$ imply that $\arg_{(t,t+1)}(Z(b^j))=\phi(b^j)$ for $j>m+1$ (see Remark \ref{arg remark} (c)). The formula \eqref{noncolinear ab2} in Corollary \ref{noncolinear ab} leads to \eqref{T43Zcap(E_1)the limit} again. Now \eqref{T43Zcap(E_1) help} implies that $\phi(b^j)>\phi(M)+1$ for big $j$, hence $\sigma \in (M, b^{j}, b^{j+1} )\subset W$ (see table \eqref{left right M}).
The arguments showing that $ (b^{m}, b^{m+1}, M') \cap W$ and $ (b^{m}, b^{m+1}, M') \cup W$ are contractible are as in the last paragraph of the proof of Lemma \ref{T12Zcap(E_1)}.
The lemma is proved.
\epr
\begin{coro} \label{T43Zcap union(E_1)} The set $\mk{T}_{b}^{st}$ is contractible.
\end{coro}
\bpr Recall that $\mk{T}_{b}^{st}= W \cup \bigcup_{j\in \NN} (b^{j}, b^{j+1}, M') $ (see \eqref{T43=Zcup}). Using Lemmas \ref{T43Zcap(E_1)} and \ref{(_,_,X)0} one shows by induction that $W \cup \bigcup_{j=0}^n (b^{m-j+1}, b^{m-j}, M') $ is contractible for each $m\in \ZZ$ and each $n\in \NN$ (see the proof of Corollary \ref{T12Zcap union(E_1)} for details). Then the corollary follows from Remark \ref{VK}.
\epr
\section{Connecting \texorpdfstring{$\mk{T}_{a}^{st}$}{\space} and \texorpdfstring{$ \mk{T}_{b}^{st}$}{\space} by \texorpdfstring{$(\_,M,\_)$}{\space} and \texorpdfstring{$(\_,M',\_)$}{\space}} \label{connecting}
Due to the union \eqref{st with T}, to prove Theorem \ref{main theo} it remains to connect the contractible non-intersecting pieces $\mk{T}_{a}^{st}$, $\mk{T}_{b}^{st}$ by $(\_,M,\_)$ and $(\_,M',\_)$, and to show that in this procedure the contractibility is preserved. We describe first the building blocks of $(\_,M,\_)$ and $(\_,M',\_)$ by Proposition \ref{lemma for f_E(Theta_E)}:
From the list of triples $\mk{T}$ given in Corollary \ref{exceptional colleections} we see that (see also \eqref{(A,_,_)}):\begin{gather} \label{middle M ab} (\_,M\_) =\bigcup_{q\in \ZZ}(a^q,M,b^{q+1}) \qquad
(\_,M'\_)=\bigcup_{q\in \ZZ}(b^q,M',a^{q}) . \end{gather}
We apply Proposition \ref{lemma for f_E(Theta_E)} to the triples $(a^p,M,b^{p+1})$ and $(b^q,M',a^{q})$. Using Corollaries \ref{nonvanishings} and \ref{one nonvanishing degree} one shows that the coefficients $\alpha, \beta, \gamma$ defined in \eqref{alpha,beta,gamma} are $\alpha=0$, $\beta=\gamma=-1$ in both the cases. Thus we obtain the formulas in the first and the second column of table \eqref{middle M} for the contractible subsets $(a^p,M,b^{p+1}) \subset \st(D^b(\mc T))$ and $(b^q,M',a^{q})\subset \st(D^b(\mc T))$, respectively:
\begin{gather} \label{middle M} \begin{array}{| c | c |} \hline
(a^p,M,b^{p+1}) & (b^q,M',a^{q}) \\
\hline
\left \{a^p,M,b^{p+1} \in \sigma^{ss} : \begin{array}{c} \phi\left ( a^{p}\right) < \phi\left (M\right)+1 \\ \phi\left ( a^{p}\right) < \phi\left (b^{p+1}\right) \\ \phi\left ( M\right) < \phi\left (b^{p+1}\right)\end{array} \right \} & \left \{b^q,M',a^{q} \in \sigma^{ss} : \begin{array}{c} \phi\left ( b^{q}\right) < \phi\left (M'\right) +1\\ \phi\left ( b^{q}\right) < \phi\left (a^{q}\right) \\ \phi\left ( M'\right) < \phi\left (a^{q}\right)\end{array} \right \} \\ \hline
\end{array}.
\end{gather}
\begin{remark} \label{middle M rem} $(a^p,M,b^{p+1}[-1])$, $(b^q,M',a^{q}[-1])$ are Ext-exceptional triples (satisfy {\rm (a)} in Def. \ref{sigma triple}).
\end{remark}
In some steps of this section, when we need to show that certain exceptional objects are semi-stable, the tools in Section \ref{general remarks} are not efficient enough. For these cases we prove Lemmas \ref{nonsemistable a} and \ref{nonsemistable b} below. The relation $\bd R & \rDotsto & (S,E) \ed $ between a $\sigma$-regular object $R$ and an exceptional pair generated by it (introduced in \cite{DK1}) is utilized in the proof of these lemmas.
\begin{lemma} \label{nonsemistable a} Let $a^m \not \in \sigma^{ss}$ and $t=\phi_-(a^m)$, then one of the following holds:
{\rm \textbf{(a)}} $a^j\in \sigma^{ss}$ for some $j<m-1$ and $t=\phi(a^j)+1$; {\rm \textbf{(b)}} $a^j\in \sigma^{ss}$ for some $m<j$ and $t=\phi(a^j)$;
{\rm \textbf{(c)}} $b^j\in \sigma^{ss}$ for some $j<m$ and $t=\phi(b^j)+1$;
{\rm \textbf{(d)}} $b^j\in \sigma^{ss}$ for some $m<j$ and $t=\phi(b^j)$;
{\rm \textbf{(e)}} $M\in \sigma^{ss}$ and $t=\phi(M)+1$.
\end{lemma}
\bpr Recall that any $X\in \{E_i^j: j\in \NN, 1\leq i \leq 4 \}$ is a trivially coupling object (see \cite[after Lemma 10.28]{DK1}). Since $a^m[k]\in \{E_i^j: j\in \NN, 1\leq i \leq 4 \}$, where $k\in \{ 0,-1 \}$, from $a^m \not \in \sigma^{ss}$ and \cite[Lemma 6.3]{DK1} it follows that $a^m[k]$ is a $\sigma$-regular object, hence $a^m$ is a $\sigma$-regular object. Therefore we have $\begin{diagram} R & \rDotsto & (S,E) \end{diagram}$ for some exceptional pair $(S,E)$ (see \cite[Section 5]{DK1}). We will need the following two properties of the exceptional object $S$. The first is $S\in \sigma^{ss}$, $\phi(S)=\phi_-(a^m)$ (see \cite[formula (42) after Definition 5.2]{DK1}). The second property is $\hom(a^m,S)\neq 0$, which follows from \cite[(c) after formula (19)]{DK1} and the way $S$ was chosen (see \cite[Definition 5.2]{DK1}). Recall that there exists at most one nonzero element in the family $\{ \hom^k(a^m,X) \}_{k\in \ZZ}$ for any $X\in \mc T_{exc}$ (Corollary \ref{one nonvanishing degree}). By Remark \ref{exceptional objects} we have $S\in \{a^j[k], b^j[k]: j\in \ZZ, k\in \ZZ\}\cup\{M[k],M'[k] ; k \in \ZZ\}$.
Obviously $S\not = a^m[k]$ (since $a^m\not \in \sigma^{ss}$ and $S \in \sigma^{ss}$).
Now we will use the property $\hom(a^m,S)\neq 0$ and Corollary \ref{nonvanishings} to prove the lemma. By $\hom^*(a^m, M')= 0$ (see \eqref{nonvanishing1}) we exclude also the case $S=M'[k]$. It remains to consider the following cases (one of them must appear):
If $S=a^j[k]$ for some $j \neq m$ and $k\in \ZZ$, then by \eqref{nonvanishing5} we see that either $j<m-1$ and $k=1$, or $m<j$ and $k=0$.
If $S=b^j[k]$ for some $j \in \ZZ$ and $k\in \ZZ$, then by \eqref{nonvanishing4} we see that either $j<m$ and $k=1$, or $m<j$ and $k=0$.
If $S=M[k]$ for some $k\in \ZZ$, then by \eqref{nonvanishing2} we get $k=1$.
The lemma follows.
\epr
\begin{lemma} \label{nonsemistable b} Let $b^m \not \in \sigma^{ss}$ and $t=\phi_-(b^m)$, then one of the following holds:
{\rm \textbf{(a)}} $a^j\in \sigma^{ss}$ for some $j<m-1$ and $t=\phi(a^j)+1$; {\rm \textbf{(b)}} $a^j\in \sigma^{ss}$ for some $m\leq j$ and $t=\phi(a^j)$;
{\rm \textbf{(c)}} $b^j\in \sigma^{ss}$ for some $j<m-1$ and $t=\phi(b^j)+1$;
{\rm \textbf{(d)}} $b^j\in \sigma^{ss}$ for some $m<j$ and $t=\phi(b^j)$;
{\rm \textbf{(e)}} $M'\in \sigma^{ss}$ and $t=\phi(M')+1$.
\end{lemma}
\bpr By the same arguments as in the proof of Lemma \ref{nonsemistable a} one shows that $\hom(b^m,S)\neq 0$ and $\phi(S)=t$ for some $S \in \sigma^{ss}\cap \left ( \{a^j[k], M, M': j\in \ZZ, k\in \ZZ\}\cup\{b^j[k] ; k \in \ZZ, j\in \ZZ,j\neq m\}\right )$. Now we will use Corollaries \ref{nonvanishings} and \ref{one nonvanishing degree}.
By $\hom^*(b^m, M)=0$ (see \eqref{nonvanishing2}) we exclude the case $S=M[k]$. It remains to consider the following cases (one of them must appear):
If $S=a^j[k]$ for some $j \in \ZZ$ and $k\in \ZZ$, then by \eqref{nonvanishing3} we see that either $j<m-1$ and $k=1$, or $m\leq j$ and $k=0$.
If $S=b^j[k]$ for some $j \neq m$ and $k\in \ZZ$, then by \eqref{nonvanishing6} we see that either $j<m-1$ and $k=1$, or $m<j$ and $k=0$.
If $S=M'[k]$ for some $k\in \ZZ$, then by \eqref{nonvanishing2} we get $k=1$.
The lemma follows.
\epr
Lemmas \ref{semi-stability of a} and \ref{semi-stability of b} put together the arguments which ensure semi-stability, necessary later in the analysis of the intersections $(a^p,M,b^{p+1}) \cap \mk{T}_{a/b}^{st}$ and $(a^p,M,b^{p+1}) \cap (\mk{T}_{a}^{st} \cup (a^q,M,b^{q+1})\cup \mk{T}_{b}^{st})$.
\begin{lemma} \label{semi-stability of a} Let $\sigma \in (a^p,M,b^{p+1})$ and let the following inequality hold: \begin{gather} \label{semi-stability of a ineq} \begin{array}{c} \phi\left (b^{p+1}\right)-1< \phi\left ( M\right)<\phi\left (b^{p+1}\right)\end{array}. \end{gather}
Then we have the following:
{\rm \textbf{(a)}} $a^{p+1}\in \sigma^{ss}$ and $\phi(b^{p+1})-1<\phi(a^{p+1})-1<\phi(M)$.
{\rm \textbf{(b)}} If in addition to \eqref{semi-stability of a ineq} we have $\phi(a^p)<\phi(M)$, then $\sigma \in (a^p,a^{p+1},M)$.
{\rm \textbf{ (c)}} If in addition to \eqref{semi-stability of a ineq} we have \begin{gather} \label{semi-stability of a ineq1} \phi\left (b^{p+1}\right)-1< \phi\left ( a^{p}\right)<\phi\left (b^{p+1}\right), \end{gather} then $M'\in \sigma^{ss}$ and $\phi(b^{p+1})-1<\phi(M')=\arg_{(\phi\left (b^{p+1}\right)-1,\phi\left (b^{p+1}\right))}(Z(a^p)-Z(b^{p+1}))<\phi(a^{p})$.
{\rm \textbf{ (d)}} If \eqref{semi-stability of a ineq}, \eqref{semi-stability of a ineq1} hold and $\phi(M')<\phi(M)$, then $\sigma \in (a^j,a^{j+1},M)$ for some $j\in \ZZ$.
\end{lemma}
\bpr (a) We apply Proposition \ref{phi_1>phi_2} (b) to the triple $(a^p,M,b^{p+1})$ and since $a^{p+1}[-1]$ is in the extension closure of $M, b^{p+1}[-1]$ (by \eqref{short filtration 2}) it follows that $a^{p+1}\in \sigma^{ss}$. The inequality $\phi(b^{p+1})-1<\phi(a^{p+1})-1<\phi(M)$ follows from the given inequality \eqref{semi-stability of a ineq} and $Z(a^{p+1}[-1])=Z(M)+Z(b^{p+1}[-1])$.
(b) From the given inequalities it follows that $\phi(a^p)<\phi(b^{p+1})$. We have also $\phi(b^{p+1})<\phi(a^{p+1})<\phi(M)+1$ from (a). Therefore we obtain the inequalities $\phi(a^{p})<\phi(a^{p+1})$, $\phi(a^p)<\phi(M)$, $\phi(a^{p+1})<\phi(M)+1$, which means that $\sigma \in (a^p,a^{p+1},M)$ (see table \eqref{left right M}).
(c) Follows from Lemma \ref{two comp factors} applied to the Ext-triple $(a^p,M,b^{p+1}[-1])$ and the triangle \eqref{short filtration 1}.
(d) Now by the given inequalities and (c) we have $\phi(b^{p+1})-1<\phi(M')<\phi(M)<\phi(b^{p+1})$. Recalling that $Z(\delta)=Z(M')+Z(M)$, we see that we can choose $t\in \RR$ with $Z(\delta)=\abs{Z(\delta)}\exp(\ri \pi t)$ and $\phi(M')<t<\phi(M)<\phi(b^{p+1})<t+1$. If $\phi(a^p)<\phi(M)$, then (d) follows from (b).
So let $\phi(M)\leq \phi(a^p)$. Since we have also $\phi(a^p)<\phi(b^{p+1})$, we obtain $t<\phi(M)\leq \phi(a^p)<\phi(b^{p+1})<t+1$. Now Corollary \ref{noncolinear ab} shows that $\{Z(a^j), Z(b^j)\}_{j\in \ZZ}\subset Z(\delta)^c_+$ and that \eqref{noncolinear ab2}, \eqref{noncolinear ab1} hold for both the sequences $\{Z(a^j)\}_{j\in \ZZ}$ and $\{Z(b^j)\}_{j\in \ZZ}$. From (a) we see that $\phi(b^{p+1})<\phi(a^{p+1})<\phi(M)+1$, hence $t< \phi(a^{p+1})<t+2$, which combined with $Z(a^{p+1})\in Z(\delta)^c_+$ implies that $\phi(a^{p+1})<t+1$. Thus we obtain the inequalities $t<\phi(M)\leq \phi(a^p)<\phi(b^{p+1})<\phi(a^{p+1})<t+1$.
From \eqref{noncolinear ab2} we see that there exists $N\in \ZZ$, $N<p$ such that $t<\arg_{(t,t+1)}(Z(a^j))<\phi(M)$ for $j<N$. We will show below that $a^j\in \sigma^{ss}$ for $j<N$. Then (d) follows. Indeed, assume that $a^j\in \sigma^{ss}$ for each $j<N$. Then by \eqref{nonvanishing5} and Corollary \ref{from equal phases to} (a) it follows that $\phi(a^{p+1})-1<\phi(a^j) <\phi(a^{p+1})$ for $j<N$, therefore $t-1<\phi(a^j) <t+1$, which combined with $Z(a^j)\in Z(\delta)^c_+$ implies that $\arg_{(t,t+1)}(Z(a^j))=\phi(a^j)$. Putting the last equality in \eqref{noncolinear ab1} and in $\arg_{(t,t+1)}(Z(a^j))<\phi(M)$ we get $\phi(a^{j-1})<\phi(a^j)<\phi(M)$ which by table \eqref{left right M} implies that $\sigma \in (a^{j-1},a^j, M)$.
Suppose $a^j \not \in \sigma^{ss}$ for some $j<N$. From Remark \ref{ext closure ab} we know that $a^j$ is in the extension closure of $a^p$, $a^{p+1}[-1]$. It follows that $a^j\in\mc P[\phi(a^{p+1})-1,\phi(a^p)]$ and then $ \phi(a^{p+1})-1\leq \phi_-(a^j)$ (recall the paragraph after \eqref{sigma^{ss}}). We will use Lemma \ref{nonsemistable a} and show that each of the five cases given there leads to a contradiction. We fist derive \eqref{semi-stability of a ineq2}. The inequalities $ \phi(a^p)-1<\phi(a^{p+1})-1<\phi(M)
\leq \phi(a^p)$ can be used due to the previous steps. Therefore we have $a^j\in\mc P[\phi(a^{p+1})-1,\phi(a^p)] \subset \mc P(\phi(a^{p})-1,\phi(a^p)]$. Using $\phi(a^p)\in (t,t+1)$, $Z(a^j)\in Z(\delta)^c_+$, and Remark \ref{arg remark} (c) we get: $\arg_{(\phi(a^{p})-1,\phi(a^p)]}(Z(a^j))=\arg_{(t,t+1)}(Z(a^j))$. Now by Remark \ref{arg remark} (a) we get $\phi_-(a^j) < \arg_{(t,t+1)}(Z(a^j))$ and by our choice of $N$ we have $\arg_{(t,t+1)}(Z(a^j))<\phi(M)$. We combine these facts in the following inequalities:
\begin{gather} \label{semi-stability of a ineq2} \phi(a^{p+1})-1\leq \phi_-(a^j) < \arg_{(t,t+1)}(Z(a^j))<\phi(M)\leq \phi(a^p)<\phi(a^{p+1}). \end{gather}
One of the cases in Lemma \ref{nonsemistable a} must appear. In case (a) we have $\phi_-(a^j) =\phi(a^k)+1 $ for some $k<j-1$, hence by \eqref{semi-stability of a ineq2} it follows $\hom^1(a^p,a^k)=0$, which contradicts \eqref{nonvanishing5} and $j<N<p$.
In case (b): $\phi_-(a^j) =\phi(a^k) $ for some $k>j$. It follows that $\phi(a^k)=\arg_{(t,t+1)}(Z(a^k))$ (see Remark \ref{arg remark} (c)), hence by \eqref{semi-stability of a ineq2} and \eqref{phase formula} we get $\arg_{(t,t+1)}(Z(a^k))<\arg_{(t,t+1)}(Z(a^j))$, which contradicts \eqref{noncolinear ab1}.
In cases (c) and (d) we have $\phi_-(a^j) =\phi(b^k) $ of $\phi(b^k)+1$ for some $k\in \ZZ$, and then \eqref{semi-stability of a ineq2} implies $\hom(M,b^k)=0$, which contradicts \eqref{nonvanishing1}.
Case (e) in Lemma \ref{nonsemistable a} and \eqref{semi-stability of a ineq2} imply that $\phi(M)+1 <\phi(M)$ and we proved the lemma.
\epr
\begin{lemma} \label{semi-stability of b} Let $\sigma \in (a^p,M,b^{p+1})$ and let the following inequality hold: \begin{gather} \label{semi-stability of b ineq} \begin{array}{c} \phi\left (a^{p}\right)-1< \phi\left ( M\right)<\phi\left (a^{p}\right)\end{array}. \end{gather}
Then we have the following:
{\rm \textbf{(a)}} $b^{p}\in \sigma^{ss}$ and $\phi(M)<\phi(b^{p})<\phi(a^{p})$.
{\rm \textbf{(b)}} If in addition to \eqref{semi-stability of b ineq} we have $\phi(M)+1<\phi(b^{p+1})$, then $\sigma \in (M, b^p, b^{p+1})$.
{\rm \textbf{ (c)}} If in addition to \eqref{semi-stability of b ineq} we have \begin{gather} \label{semi-stability of b ineq1} \phi\left (a^{p}\right)-1< \phi\left ( b^{p+1}\right)-1<\phi\left (a^{p}\right), \end{gather} then $M'\in \sigma^{ss}$ and $\phi(b^{p+1})-1<\phi(M')=\arg_{(\phi\left (a^{p}\right)-1,\phi\left (a^{p}\right))}(Z(a^p)-Z(b^{p+1}))<\phi(a^{p})$.
{\rm \textbf{ (d)}} If \eqref{semi-stability of b ineq}, \eqref{semi-stability of b ineq1} hold and $\phi(M)<\phi(M')$, then $\sigma \in (b^j,b^{j+1},M')$ for some $j\in \ZZ$ or $\sigma \in (M, b^p, b^{p+1})$.
{\rm \textbf{ (e)}} If \eqref{semi-stability of b ineq}, \eqref{semi-stability of b ineq1} hold and $\phi(M)=\phi(M')$, then $\sigma \in (a^j,M,b^{j+1})$ for each $j<p$.
\end{lemma}
\bpr \textbf{(a)} We apply Proposition \ref{phi_1>phi_2} (a) to the triple $(a^p,M,b^{p+1})$ and since $b^{p}$ is in the extension closure of $M, a^p$ (by \eqref{short filtration 2}) it follows that $b^{p}\in \sigma^{ss}$ and $\phi(M)\leq \phi(b^{p})\leq \phi(a^{p})$. The inequality $\phi(M)<\phi(b^{p})<\phi(a^{p})$ follows from the given inequality \eqref{semi-stability of b ineq} and $Z(b^{p})=Z(M)+Z(a^{p})$.
\textbf{(b)} From the given inequalities we have $\phi(a^p)<\phi(M)+1<\phi(b^{p+1})$. In (a) we showed that $\phi(M)<\phi(b^{p})<\phi(a^{p})$. Therefore we obtain the inequalities $\phi(M)<\phi(b^{p})$, $\phi(M)+1<\phi(b^{p+1})$, $\phi(b^{p})<\phi(b^{p+1})$, which means that $\sigma \in (M,b^p,b^{p+1})$ (see table \eqref{left right M}).
\textbf{(c)} Follows from Lemma \ref{two comp factors} applied to the Ext-triple $(a^p,M,b^{p+1}[-1])$ and the triangle \eqref{short filtration 1}.
\textbf{(d)} Now by the given inequalities and (c) we have $\phi(a^{p})-1<\phi(M)<\phi(M')<\phi(a^{p})$. Recalling that $Z(\delta)=Z(M')+Z(M)$, we see that we can choose $t\in \RR$ with $Z(\delta)=\abs{Z(\delta)}\exp(\ri \pi t)$ and $\phi(M)<t<\phi(M')<\phi(a^{p})<\phi(M)+1$. If $\phi(M)+1<\phi(b^{p+1})$, then we apply (b).
So, let $\phi(b^{p+1})\leq\phi(M)+1$. Since we have also $\phi(a^p)<\phi(b^{p+1})$, we obtain $t<\phi(M')<\phi(a^p)<\phi(b^{p+1})\leq\phi(M)+1 <t+1$. Now Corollary \ref{noncolinear ab} ensures that $\{Z(a^j), Z(b^j)\}_{j\in \ZZ}\subset Z(\delta)^c_+$ and that \eqref{noncolinear ab2}, \eqref{noncolinear ab1} hold for both the sequences $\{Z(a^j)\}_{j\in \ZZ}$ and $\{Z(b^j)\}_{j\in \ZZ}$. From (a) we see that $\phi(M)<\phi(b^{p})<\phi(a^{p})$, hence $t-1< \phi(b^{p})<t+1$, which combined with $Z(b^{p})\in Z(\delta)^c_+$ implies that $t<\phi(b^{p})$. Hence we obtain the inequalities \begin{gather} \label{help ineq} t<\phi(b^p)< \phi(a^p)<\phi(b^{p+1})<t+1; \quad t<\phi(M')<\phi(a^p)<\phi(b^{p+1}) <t+1 . \end{gather}
From \eqref{noncolinear ab2} and $t<\phi(M')$ it follows that there exists $N\in \ZZ$, $N<p$ such that $t<\arg_{(t,t+1)}(Z(b^j))<\phi(M')$ for $j<N$. We will show below that $b^j\in \sigma^{ss}$ for $j<N$. Then (d) follows. Indeed, assume that $b^j\in \sigma^{ss}$ for each $j<N$. Then by \eqref{nonvanishing6} and Corollary \ref{from equal phases to} (a) it follows that $\phi(b^{p+1})-1<\phi(b^j) <\phi(b^{p+1})$ for $j<N$, and by \eqref{help ineq} we get $t-1<\phi(b^j) <t+1$, which combined with $Z(b^j)\in Z(\delta)^c_+$ implies that $\arg_{(t,t+1)}(Z(b^j))=\phi(b^j)$. Putting the last equality in \eqref{noncolinear ab1} and in $\arg_{(t,t+1)}(Z(b^j))<\phi(M')$ we obtain $\phi(b^{j-1})<\phi(b^j)<\phi(M')$, which implies $\sigma \in (b^{j-1},b^j, M')$.
Suppose $b^j \not \in \sigma^{ss}$ for some $j<N$. We apply Lemma \ref{nonsemistable b} and show that each of the five cases given there leads to a contradiction. We show first \eqref{semi-stability of b ineq2}. From Remark \ref{ext closure ab} we know that $b^j$ is in the extension closure of $b^p$, $b^{p+1}[-1]$ (recall that $N<p$) and we have $ \phi(b^p)-1<\phi(b^{p+1})-1 <t
< \phi(b^p)$ in \eqref{help ineq}. It follows that $b^j\in\mc P[\phi(b^{p+1})-1,\phi(b^p)] \subset \mc P(\phi(b^{p})-1,\phi(b^p)]$. Using $\phi(b^p)\in (t,t+1)$, $Z(b^j)\in Z(\delta)^c_+$, Remark \ref{arg remark} (c) and (a), we deduce that $\arg_{(\phi(b^{p})-1,\phi(b^p)]}(Z(b^j))=\arg_{(t,t+1)}(Z(b^j))>\phi_-(b^j)$. The incidence $b^j\in\mc P[\phi(b^{p+1})-1,\phi(b^p)]$ implies $\phi(b^{p+1})-1\leq \phi_-(b^j)$, and we get:
\begin{gather} \label{semi-stability of b ineq2} \phi(b^{p+1})-1\leq \phi_-(b^j) < \arg_{(t,t+1)}(Z(b^j))<\phi(M') <\phi(b^{p+1}). \end{gather}
One of the cases in Lemma \ref{nonsemistable b} must appear. In cases (a) and (b) we have $\phi_-(b^j) =\phi(a^k) $ of $\phi(a^k)+1$ for some $k\in \ZZ$, and then \eqref{semi-stability of b ineq2} implies $\hom(M',a^k)=0$, which contradicts \eqref{nonvanishing1}.
In case (c) we have $\phi_-(b^j) =\phi(b^k)+1 $ for some $k<j-1$, and \eqref{semi-stability of b ineq2} implies that $\hom^1(b^{p+1},b^k)=0$, which contradicts \eqref{nonvanishing6} and $k<j-1<p-1$.
In case (d) we have $\phi_-(b^j) =\phi(b^k) $ for some $k>j$. From $Z(b^k) \in Z(\delta)^c_+$, \eqref{semi-stability of b ineq2}, and $\phi(b^{p+1})\in (t,t+1)$ it follows that $\phi(b^k)=\arg_{(t,t+1)}(Z(b^k))$. Hence \eqref{semi-stability of b ineq2} and \eqref{phase formula} imply $\arg_{(t,t+1)}(Z(b^k))<\arg_{(t,t+1)}(Z(b^j))$, which contradicts $k>j$ and \eqref{noncolinear ab1}.
Case (e) in Lemma \ref{nonsemistable b} and \eqref{semi-stability of b ineq2} imply that $\phi(M')+1 <\phi(M')$. We proved completely part (d) of the lemma.
\textbf{(e)} Now by the given inequalities we have $\phi(a^{p})-1<\phi(M)=\phi(M')<\phi(a^{p})$. Recalling that $Z(\delta)=Z(M')+Z(M)$, we see that $t=\phi(M)=\phi(M')$ satisfies $Z(\delta)=\abs{Z(\delta)}\exp(\ri \pi t)$ and $t<\phi(a^{p})<t+1$. From (a) we get $t<\phi(b^{p})<\phi(a^{p})<t+1$. Now we can apply Corollary \ref{noncolinear ab full1}, which besides $\{Z(a^j), Z(b^j)\}_{j\in \ZZ}\subset Z(\delta)^c_+$ and formulas \eqref{noncolinear ab1}, \eqref{noncolinear ab2} gives us the inequalities \eqref{j<mleqi}.
We extend the inequality $t<\phi(b^{p})<\phi(a^{p})<t+1$ to \eqref{semi-stability of b ineq3} as follows. We already have that $a^p, b^p, b^{p+1}\in \sigma^{ss}$. In \eqref{semi-stability of b ineq1} is given that $\phi(a^{p})<\phi(b^{p+1})$.
From $\hom^1(b^{p+1},M')$ (see \eqref{nonvanishing2}) it follows $\phi(b^{p+1})\leq t+1$ and from $Z(b^{p+1})\in Z(\delta)^c_+$ we see that $\phi(b^{p+1})< t+1=\phi(M)+1$. We have also $\phi(M)<\phi(b^{p+1})$ (due to $\sigma \in (a^p,M,b^{p+1})$). Therefore $\phi(b^{p+1})-1<\phi(M)< \phi(b^{p+1})$ and from Lemma \ref{semi-stability of a} (a) we get $a^{p+1}\in \sigma^{ss}$ and $\phi(b^{p+1})-1<\phi(a^{p+1})-1<\phi(M)$. Thus, we derive:
\begin{gather} \label{semi-stability of b ineq3} \phi(a^{p})-1<\phi(b^{p+1})-1<\phi(a^{p+1})-1<t<\phi(b^{p})<\phi(a^{p})<\phi(b^{p+1})<\phi(a^{p+1})<t+1.\end{gather}
We will show below that $a^j$ and $b^j$ are semi-stable for each $j<p$. We claim that this implies $\sigma \in (a^{j}, M, b^{j+1})$ for $j<p$. Indeed, assume that $a^j, b^j \in \sigma^{ss}$ for each $j<p$. Then by \eqref{nonvanishing5}, \eqref{nonvanishing6} we get $\phi(a^{p+1})-1\leq \phi(a^j)\leq \phi(a^{p+1})$ and $\phi(b^{p+1})-1\leq \phi(b^j)\leq \phi(b^{p+1})$, which combined with $t<\phi(b^{p+1})<\phi(a^{p+1})<t+1$ and $Z(a^j), Z(b^j)\in Z(\delta)^c_+$ implies that $\phi(a^j),\phi(b^j) \in (t,t+1)$, in particular $\phi(a^j)=\arg_{(t,t+1)}(Z(a^j))$ and $\phi(b^j)=\arg_{(t,t+1)}(Z(b^j))$ for each $j<p$. The last two equalities hold also for $j=p$ by \eqref{semi-stability of b ineq3}. Putting these equalities in \eqref{j<mleqi} we get that $\phi(a^{j})<\phi(b^{j+1})$ for each $j<p$. Thus, we obtain $\phi(M)<\phi(a^{j})<\phi(b^{j+1})<\phi(M)+1$ for each $j<p$, which by table \eqref{middle M} gives $\sigma \in (a^{j}, M, b^{j+1})$.
Suppose that $b^j\not \in \sigma^{ss}$ for some $j<p$. Remark \ref{ext closure ab} asserts that $b^j$ is in the extension closure of $b^p$, $b^{p+1}[-1]$, therefore $b^j\in\mc P[\phi(b^{p+1})-1,\phi(b^p)]$, and hence $\phi(b^{p+1})-1\leq \phi_-(b^j)$, $\phi_+(b^j)\leq \phi(b^p)$. Due to \eqref{semi-stability of b ineq3} we can write $b^j\in\mc P[\phi(b^{p+1})-1,\phi(b^p)] \subset \mc P(\phi(a^{p})-1,\phi(a^p)]$. Using $\phi(a^p)\in (t,t+1)$, $Z(b^j)\in Z(\delta)^c_+$ and Remark \ref{arg remark} (c) we conclude that $\arg_{(\phi(a^{p})-1,\phi(a^p)]}(Z(b^j))=\arg_{(t,t+1)}(Z(b^j))$. Now using Remark \ref{arg remark} (a), we obtain:
\begin{gather} \label{semi-stability of b ineq4} \phi(b^{p+1})-1\leq \phi_-(b^j) < \arg_{(t,t+1)}(Z(b^j))<\phi_+(b^j)\leq \phi(b^p) <\phi(b^{p+1}). \end{gather}
We use Lemma \ref{nonsemistable b} and show that each of the five cases given there leads to a contradiction.
Case (a) ensures $\phi_-(b^j) = \phi(a^k)+1$ for some $k<j-1$ and \eqref{semi-stability of b ineq4} implies that $\hom^1(b^{p+1},a^k)=0$, which contradicts \eqref{nonvanishing3} (now $k<p$).
Case (b) ensures $\phi_-(b^j) =\phi(a^k) $ for some $k\geq j$, and then \eqref{semi-stability of b ineq4} and $Z(a^k) \in Z(\delta)^c_+$
imply $\arg_{(t,t+1)}(Z(a^k))=\phi(a^k)$, hence by \eqref{semi-stability of b ineq4} and \eqref{phase formula} we get $\arg_{(t,t+1)}(Z(a^k))<\arg_{(t,t+1)}(Z(b^j))$, which contradicts \eqref{j<mleqi} and $k\geq j$.
Case (c) ensures $\phi_-(b^j) =\phi(b^k)+1 $ for some $k<j-1<p-1$, and \eqref{semi-stability of b ineq4} implies that $\hom^1(b^{p+1},b^k)=0$, which contradicts \eqref{nonvanishing6}.
In case (d) we have $\phi_-(b^j) =\phi(b^k) $ for some $k>j$. It follows by $Z(b^k)\in Z(\delta)^c_+$ and \eqref{semi-stability of b ineq4} that $\phi(b^k)=\arg_{(t,t+1)}(Z(b^k))$, and then \eqref{semi-stability of b ineq4} gives $\arg_{(t,t+1)}(Z(b^k))<\arg_{(t,t+1)}(Z(b^j))$, which contradicts \eqref{noncolinear ab1}.
In case (e) using \eqref{semi-stability of b ineq4} we obtain $\phi(M')+1 < \phi(b^{p+1})$, which contradicts \eqref{nonvanishing2}.
Suppose that $a^j\not \in \sigma^{ss}$ for some $j<p$. Since $a^j$ is in the extension closure of $a^p$, $a^{p+1}[-1]$ (see Remark \ref{ext closure ab} ), therefore $a^j\in\mc P[\phi(a^{p+1})-1,\phi(a^p)]$, and hence $\phi_\pm(a^j) \in [\phi(a^{p+1})-1,\phi(a^p)]$. Due to \eqref{semi-stability of b ineq3} we have $a^j\in\mc P[\phi(a^{p+1})-1,\phi(a^p)] \subset \mc P(\phi(b^{p+1})-1,\phi(b^{p+1})]$ and Remark \ref{arg remark} (c) shows that that $\arg_{(\phi(b^{p+1})-1,\phi(b^{p+1})]}(Z(a^j))=\arg_{(t,t+1)}(Z(a^j))$. Now Remark \ref{arg remark} (a) completes the following:
\begin{gather} \label{semi-stability of b ineq5} \phi(a^{p+1})-1\leq \phi_-(a^j) < \arg_{(t,t+1)}(Z(a^j))<\phi_+(a^j)\leq \phi(a^p) <\phi(a^{p+1}). \end{gather}
We use Lemma \ref{nonsemistable a} to get a contradiction. One of the five cases given there must appear.
In case (a) of Lemma \ref{nonsemistable a} we have $\phi_-(a^j) =\phi(a^k)+1 $ for some $k<j-1<p-1$, and \eqref{semi-stability of b ineq5} implies $\hom^1(a^{p+1},a^k)=0$, which contradicts \eqref{nonvanishing5}.
Case (b) ensures $\phi_-(a^j) =\phi(a^k) $ for some $k>j$. It follows that $\phi(a^k)=\arg_{(t,t+1)}(Z(a^k))$ (see Remark \ref{arg remark} (c)), hence by \eqref{semi-stability of b ineq5} we get $\arg_{(t,t+1)}(Z(a^k))<\arg_{(t,t+1)}(Z(a^j))$, which contradicts \eqref{noncolinear ab1}.
In case (c) we have $\phi_-(a^j) = \phi(b^k)+1$ for some $k<j$ and \eqref{semi-stability of b ineq5} implies that $\hom^1(a^{p+1},b^k)=0$, which contradicts \eqref{nonvanishing4} (now $k<p$).
Case (d) ensures $\phi_-(a^j) =\phi(b^k) $ for some $ j<k$, and then $\arg_{(t,t+1)}(Z(b^k))=\phi(b^k)$ (see Remark \ref{arg remark} (c)),
hence by \eqref{semi-stability of b ineq5} we get $\arg_{(t,t+1)}(Z(b^k))<\arg_{(t,t+1)}(Z(a^j))$, which contradicts \eqref{j<mleqi}.
In case (e) we have $\phi_-(a^j) =\phi(M)+1$, and \eqref{semi-stability of b ineq5} implies $\hom^1(a^{p+1},M)=0$, which contradicts \eqref{nonvanishing2}.
The lemma is proved.
\epr
Next we glue $(a^p,M,b^{p+1})$ and $\mk{T}_{a}^{st}$.
\begin{lemma} \label{middle M cap left M'} For any $p\in \ZZ$ the set $(a^p,M,b^{p+1}) \cap \mk{T}_{a}^{st}$ consists of the stability conditions $\sigma$ for which $a^p,M,b^{p+1}$ are semistable and: \begin{gather} \label{middle M cap left M' sys}\begin{array}{c} \phi\left (b^{p+1}\right)-1< \phi\left ( M\right)<\phi\left (b^{p+1}\right)\\ \phi\left (b^{p+1}\right)-1< \phi\left ( a^{p}\right)<\phi\left (b^{p+1}\right) \\ \arg_{(\phi\left (b^{p+1}\right)-1,\phi\left (b^{p+1}\right))}(Z(a^p)-Z(b^{p+1}))<\phi(M) \end{array} \ \mbox{or} \ \begin{array}{c} \phi\left (a^p\right)< \phi\left ( M\right)\\ \phi\left (b^{p+1}\right)-1< \phi\left ( M\right)<\phi\left (b^{p+1}\right)\end{array}. \end{gather}
It follows that $(a^p,M,b^{p+1}) \cap \mk{T}_{a}^{st}$ and $ (a^p,M,b^{p+1}) \cup \mk{T}_{a}^{st}$ are contractible.
\end{lemma}
\bpr We start with the inclusion $\subset$. Assume that $\sigma \in (a^p,M,b^{p+1})$. Then $a^p,M,b^{p+1}$ are semi-stable and by table \eqref{middle M} we get \begin{gather} \label{middle M cap left M'1} \begin{array}{c} \phi\left ( a^{p}\right) < \phi\left (M\right)+1 \\ \phi\left ( a^{p}\right) < \phi\left (b^{p+1}\right) \\ \phi\left ( M\right) < \phi\left (b^{p+1}\right)\end{array} \end{gather}
Recalling \eqref{T_12new}, we see that we have to consider three cases.
\ul{If $\sigma \in (M',a^j, a^{j+1})$}, then $M',a^j, a^{j+1}$ are semi-stable and from table \eqref{left right M} we see that $\phi(M')+1<\phi(a^{j+1})$. Since we have also $\hom^1(b^{p+1},M')$, $\hom^1(a^{j+1},M)\neq 0$(see Corollary \ref{nonvanishings}), we obtain $ \phi\left (b^{p+1}\right) \leq \phi(M')+1<\phi(a^{j+1})\leq \phi(M)+1$, which combined with \eqref{middle M cap left M'1} implies
\begin{gather}\label{middle M cap left M'2} \phi\left (b^{p+1}\right)-1< \phi\left ( M\right)<\phi\left (b^{p+1}\right) \qquad \phi(M')< \phi(M). \end{gather} These non-vanishings and inequalities give also $\phi\left ( a^{p}\right) < \phi\left (b^{p+1}\right) \leq \phi(M')+1<\phi(a^{j+1})$. Using Remark \ref{inc dec seq} (a) we deduce that $p\leq j$.
We show now that $\phi(b^{p+1})<\phi(a^p)+1$. If $j=p$, then we immediately obtain this by $\hom^1(b^{p+1}, M')\neq 0$ and $ \phi(M')<\phi(a^p)$(see table \eqref{left right M}). If $j>p$, then $\hom(b^{p+1},a^j)\neq 0$ and $\hom^1(a^{j+1},a^p)\neq 0$ (see Corollary \ref{nonvanishings}) and we can write $\phi(b^{p+1})\leq\phi(a^{j})<\phi(a^{j+1})\leq \phi(a^{p})+1 $.
To obtain the first system of inequalities in \eqref{middle M cap left M' sys} it remains to show the third inequality.
From the triangle \eqref{short filtration 1} it follows that $\phi(b^{p+1})-1\leq \phi(M')\leq \phi(a^p)$ and $Z(M')=Z(a^p)-Z(b^{p+1})$, now $\phi(M')=\arg_{(\phi\left (b^{p+1}\right)-1,\phi\left (b^{p+1}\right))}(Z(a^p)-Z(b^{p+1}))<\phi(M)$ follows from the already proved $\phi\left (b^{p+1}\right)-1< \phi\left ( a^{p}\right)<\phi\left (b^{p+1}\right)$ and \eqref{middle M cap left M'2}.
\ul{If $\sigma\in (a^m,a^{m+1},M)$}, then $a^m,a^{m+1}$ are semistable as well and in table \eqref{left right M} we see that $\phi(a^m)<\phi(M)$, which together with the third inequality in \eqref{middle M cap left M'1} imply that $\phi(a^m)<\phi(b^{p+1})$ and hence $\hom(b^{p+1},a^m)=0$. By \eqref{nonvanishing3} we deduce that $p\geq m$.
If $p=m$, then we get immediately $\phi(a^p)<\phi(M)$. In table \eqref{left right M} we have $\phi(a^{p+1})<\phi(M)+1$ and in Corollary \ref{nonvanishings} we have $\hom(b^{p+1},a^{p+1})\neq 0$, hence $\phi(b^{p+1})<\phi(M)+1$ and we obtain the second system of inequalities in \eqref{middle M cap left M' sys}.
If $p>m$, then $\hom^1(b^{p+1},a^m)\neq 0$ and from the inequalities $\phi(a^m)<\phi(M)$, $\phi(a^m)<\phi(a^{m+1})$ (due to $\sigma \in (a^m,a^{m+1},M)$) it follows $\phi(b^{p+1})<\phi(M)+1$ and $\phi(b^{p+1})\leq \phi(a^m)+1<\phi(a^{m+1})+1\leq \phi(a^{p})+1$. Recalling \eqref{middle M cap left M'1} we see that we obtained the first two equalities in \eqref{middle M cap left M' sys}. Hence by Lemma \ref{semi-stability of a} (c) we get $M'\in \sigma^{ss}$ and $\phi(M')=\arg_{(\phi\left (b^{p+1}\right)-1,\phi\left (b^{p+1}\right))}(Z(a^p)-Z(b^{p+1}))$. From $\hom(M',a^m)\neq 0$ it follows $\phi(M')\leq \phi(a^m)<\phi(M)$ and we obtain the complete first system of inequalitites in \eqref{middle M cap left M' sys}.
\ul{If $\sigma\in (a^m, b^{m+1},a^{m+1})$}, then $a^m, b^{m+1},a^{m+1}\in \sigma^{ss}$ and in table \eqref{no M} we see that $\phi(a^m)+1<\phi(a^{m+1})$, hence Lemma \ref{from phases of big distance} and $a^p \in \sigma^{ss}$ imply that $p=m$ or $p=m+1$. If $p=m+1$, then by \eqref{middle M cap left M'1} we obtain $\phi(a^m)+1<\phi(a^{m+1})<\phi(b^{m+2})$, and hence $\hom^1(b^{m+2},a^m)=0$, which contradicts \eqref{nonvanishing3}. Thus, it remains to consider the case $m=p$. Now we have $\phi(a^p)+1<\phi(a^{p+1})$ and $\phi(b^{p+1})<\phi(a^{p+1})$(see table \eqref{no M}), which together with $\hom^1(a^{p+1},M)\neq 0$ imply $\phi(a^p)<\phi(M)$ and $\phi(b^{p+1})<\phi(M)+1$, hence we obtain the second system in \eqref{middle M cap left M' sys}. The inclusion $\subset$ is shown.
We show now the converse inclusion $\supset$. Assume that $a^p,M,b^{p+1}$ are semi-stable and that one of the two systems of inequalities in \eqref{middle M cap left M' sys} holds. In both the cases the given inequalities imply the inequalities \eqref{middle M cap left M'1}, therefore $\sigma \in (a^p,M,b^{p+1})$. If the second system in \eqref{middle M cap left M' sys} holds, then by Lemma \ref{semi-stability of a} (b) we get $\sigma \in (a^p,a^{p+1},M)\subset \mk{T}_{a}^{st}$. If the first system in \eqref{middle M cap left M' sys} holds, then by Lemma \ref{semi-stability of a} (c) and (d) we get $\sigma \in (a^j,a^{j+1},M)\subset \mk{T}_{a}^{st}$ for some $j\in \ZZ$, and the inclusion $\supset$ is proved as well.
As in the last paragraph of the proof of Lemma \ref{T12Zcap(E_1)} one shows that the two systems of inequalities in \eqref{middle M cap left M' sys} correspond to two contractible sets (the first is contractible by Corollary \ref{contractible 1a}), and it is easy to show that their intersection is homeomorphic to $\RR_{>0}^3 \times \{\phi_2-1<\phi_0<\phi_1 < \phi_2\}$, which is also contractible. Remark \ref{VK} shows that $ (a^p,M,b^{p+1}) \cap \mk{T}_{a}^{st}$ is contractible. Since $ (a^p,M,b^{p+1}) $ and $ \mk{T}_{a}^{st}$ are both contractible (Proposition \ref{lemma for f_E(Theta_E)} and Corollary \ref{T12Zcap union(E_1)}), Remark \ref{VK} shows that $ (a^p,M,b^{p+1}) \cup \mk{T}_{a}^{st}$ is contractible as well.
\epr
\begin{lemma} \label{middle M cap left M} For any $p\in \ZZ$ the set $(a^p,M,b^{p+1}) \cap \mk{T}_{b}^{st}$ consists of the stability conditions $\sigma$ for which $a^p,M,b^{p+1}$ are semistable and: \begin{gather} \label{middle M cap left M sys} \begin{array}{c}\phi\left (a^{p}\right)-1< \phi\left ( M\right)<\phi\left (a^{p}\right)\\
\phi\left (a^{p}\right)-1< \phi\left ( b^{p+1}\right)-1<\phi\left (a^{p}\right) \\ \arg_{(\phi\left (a^{p}\right)-1,\phi\left (a^{p}\right))}(Z(a^p)-Z(b^{p+1}))>\phi(M) \end{array} \ \mbox{or} \ \begin{array}{c} \phi\left (a^{p}\right)-1< \phi\left ( M\right)<\phi\left (a^{p}\right)\\ \phi(M)+1<\phi(b^{p+1})\end{array}. \end{gather}
It follows that $(a^p,M,b^{p+1}) \cap \mk{T}_{b}^{st}$ and $ (a^p,M,b^{p+1}) \cup \mk{T}_{b}^{st}$ are contractible.
\end{lemma}
\bpr We start with the inclusion $\subset$. Assume that $\sigma \in (a^p,M,b^{p+1})$. Then $a^p,M,b^{p+1}$ are semi-stable and by table \eqref{middle M} we get \begin{gather} \label{middle M cap left M1} \begin{array}{c} \phi\left ( a^{p}\right) < \phi\left (M\right)+1 \\ \phi\left ( a^{p}\right) < \phi\left (b^{p+1}\right) \\ \phi\left ( M\right) < \phi\left (b^{p+1}\right)\end{array} \end{gather}
Recalling \eqref{T_43new}, we see that we have to consider three cases.
\ul{If $\sigma \in (M,b^j, b^{j+1})$}, then $M,b^j, b^{j+1}$ are semi-stable and from table \eqref{left right M} we see that $\phi(M)<\phi(b^{j})$ and $\phi(M)+1<\phi(b^{j+1})$, hence $\phi(a^p)<\phi(b^{j+1})$ and $\hom(b^{j+1},a^p)=0$. From \eqref{nonvanishing3} it follows that $p\leq j$. If $j=p$, then $\phi(M)+1<\phi(b^{p+1})$ and by $\hom(b^{p},a^{p})\neq 0$ (see \eqref{nonvanishing3}) we get $\phi(M)<\phi(a^{p})$, which implies the second system in \eqref{middle M cap left M sys}. It remains to consider the case $p< j$.
In this case $\hom^1(b^{j+1}, a^p)\neq 0$ (see \eqref{nonvanishing3}) and we obtain $\phi(M)+1<\phi(b^{j+1})\leq \phi(a^{p})+1$, which combined with \eqref{middle M cap left M1} implies
$\phi\left (a^{p}\right)-1< \phi\left ( M\right)<\phi\left (a^{p}\right)$.
On the other hand, we have $\phi(b^{j})<\phi(b^{j+1})$ (see table \eqref{left right M}), and by $p<j$ we can write $\phi(b^{p+1})\leq \phi(b^{j})<\phi(b^{j+1})\leq \phi(a^{p})+1$, which combined with \eqref{middle M cap left M1} implies $ \phi\left (a^{p}\right)-1< \phi(b^{p+1})-1<\phi\left (a^{p}\right)$. Now we can use Lemma \ref{semi-stability of b} (c) to deduce that $M'\in \sigma^{ss}$ and $\phi(M')=\arg_{(\phi\left (a^{p}\right)-1,\phi\left (a^{p}\right))}(Z(a^p)-Z(b^{p+1}))$. From $\hom^1(b^{j+1},M')\neq 0$ and $\phi(M)+1<\phi(b^{j+1})$ it follows that $\phi(M)<\phi(M')$ and the first system in \eqref{middle M cap left M sys} follows.
\ul{If $\sigma\in (b^m,b^{m+1},M')$}, then $b^m,b^{m+1}, M'$ are semistable and in table \eqref{left right M} we see that $\phi(b^m)<\phi(M')$. By $\hom(M',a^p)\neq 0$ and $\hom(M,b^m)\neq 0$ (see \eqref{nonvanishing1}) we get:
\begin{gather} \label{middle M cap left M11} \phi(M)\leq \phi(b^m)<\phi(M')\leq \phi(a^p). \end{gather}
Whence $\phi(M)<\phi(a^p)$ and combining with \eqref{middle M cap left M1} we derive $ \phi\left (a^{p}\right)-1< \phi(M)<\phi\left (a^{p}\right)$. On the other hand, in \eqref{middle M cap left M11} we have also $\phi(b^m)<\phi(a^p)$, and hence $\hom(a^p,b^m)=0$, threfore by \eqref{nonvanishing4} we see that $p\geq m$. In \eqref{middle M cap left M11} we have also $\phi(M)<\phi(M')$. Taking into account Lemma \ref{semi-stability of b} (c), we see that if we show that $ \phi\left (a^{p}\right)-1< \phi\left (b^{p+1}\right)-1<\phi\left (a^{p}\right)$, then the first system in \eqref{middle M cap left M sys} follows. Since we have $\phi\left (a^{p}\right)< \phi\left (b^{p+1}\right)$ (see \eqref{middle M cap left M1}), it remains to show that $\phi\left (b^{p+1}\right)<\phi\left (a^{p}\right)+1$. If $p=m$, then from table \eqref{left right M} we obtain $\phi(b^{p+1})<\phi(M')+1$ and the inequality in question follows from $\phi(M')\leq \phi(a^p)$. If $m<p$, then $\hom^1(b^{p+1},b^m)\neq 0$ and we get $\phi(b^{p+1})\leq \phi(b^{m})+1<\phi\left (a^{p}\right)+1$ (see \eqref{middle M cap left M11}).
\ul{If $\sigma\in (b^m, a^{m},b^{m+1})$}, then $b^m, a^{m},b^{m+1} \in \sigma^{ss}$ and in table \eqref{no M} we see that $\phi(b^m)+1<\phi(b^{m+1})$, hence Lemma \ref{from phases of big distance} and $b^{p+1} \in \sigma^{ss}$ imply that $p=m$ or $p=m-1$. If $p=m-1$, then by \eqref{middle M cap left M1} we obtain $\phi(a^{m-1})+1<\phi(b^m)+1<\phi(b^{m+1})$, and hence $\hom^1(b^{m+1},a^{m-1})=0$, which contradicts \eqref{nonvanishing3}. Therefore we have $m=p$. Now we have $\phi(b^p)+1<\phi(b^{p+1})$ and $\phi(b^{p})<\phi(a^{p})$ (see table \eqref{no M}), which together with $\hom(M,b^p)\neq 0$ imply $\phi(M)+1<\phi(b^{p+1})$ and $\phi(M)<\phi(a^{p})$, hence the second system in \eqref{middle M cap left M sys} follows. Thus we showed the inclusion $\subset$.
We show now the inverse inclusion $\supset$. Assume that $a^p,M,b^{p+1}$ are semi-stable and that one of the two systems of inequalities in \eqref{middle M cap left M sys} holds. In both the cases the given inequalities imply the inequalities \eqref{middle M cap left M1}, therefore $\sigma \in (a^p,M,b^{p+1})$. If the second system in \eqref{middle M cap left M sys} holds, then by Lemma \eqref{semi-stability of b} (b) we get $\sigma \in (M, b^p,b^{p+1}) \subset \mk{T}_{b}^{st}$. If the first system in \eqref{middle M cap left M sys} holds, then the desired $\sigma \in \mk{T}_{b}^{st}$ follows from Lemma \eqref{semi-stability of b} (c) and (d). The inclusion $\supset$ is proved as well.
In Corollary \ref{T43Zcap union(E_1)} was shown that $\mk{T}_{b}^{st}$ is contractible. The proof that $(a^p,M,b^{p+1}) \cap \mk{T}_{b}^{st}$ and $ (a^p,M,b^{p+1}) \cup \mk{T}_{b}^{st}$ are contractible
is as in the last paragraph of Lemma \ref{middle M cap left M'}. The two systems in \eqref{middle M cap left M sys} correspond to contractible subsets of $(a^p,M,b^{p+1}) \cap \mk{T}_{b}^{st}$ (the first is contractible by Corollary \ref{contractible 1b}).
The intersection of these subsets is homeomorphic to $\RR_{>0}^3 \times \{\phi_0-1<\phi_1<\phi_2-1 < \phi_0\}$, which is also contractible. Now we apply Remark \ref{VK} twice and the lemma follows.
\epr
\begin{coro} \label{middle M cup left right M coro} For any $p\in \ZZ $ the set $\mk{T}_{a}^{st} \cup (a^p,M,b^{p+1})\cup \mk{T}_{b}^{st} $ is contractible. \end{coro}
\bpr In Lemma \ref{middle M cap left M'} we showed that $\mk{T}_{a}^{st} \cup (a^p,M,b^{p+1})$ is contractible. Since $\mk{T}_{a}^{st} \cap \mk{T}_{b}^{st} =\emptyset $ (see Subsection \ref{empty intersection}), it follows that $(\mk{T}_{a}^{st} \cup (a^p,M,b^{p+1}))\cap \mk{T}_{b}^{st} = (a^p,M,b^{p+1})\cap \mk{T}_{b}^{st}$, which is contractible by Lemma \ref{middle M cap left M}. Now we apply Remark \ref{VK}.
\epr
\begin{lemma} \label{middle M cap left right M} For any $q<p$ the set $(a^p,M,b^{p+1}) \cap (\mk{T}_{a}^{st} \cup (a^q,M,b^{q+1})\cup \mk{T}_{b}^{st} )$ consists of the stability conditions $\sigma$ for which $a^p,M,b^{p+1}$ are semistable and: \begin{gather} \begin{array}{c}\phi\left (a^{p}\right)-1< \phi\left ( M\right)<\phi\left (a^{p}\right)\\
\phi\left (a^{p}\right)-1< \phi\left ( b^{p+1}\right)-1<\phi\left (a^{p}\right) \end{array} \ \mbox{or} \ \begin{array}{c} \phi\left (a^{p}\right)-1< \phi\left ( M\right)<\phi\left (a^{p}\right)\\ \phi(M)+1<\phi(b^{p+1})\end{array} \nonumber \\[-2mm] \label{middle M cap left right M sys} \\[-2mm] \mbox{or} \begin{array}{c} \phi\left (a^p\right)< \phi\left ( M\right)\\ \phi\left (b^{p+1}\right)-1< \phi\left ( M\right)<\phi\left (b^{p+1}\right)\end{array} \ \mbox{or} \ \begin{array}{c} \phi\left (b^{p+1}\right)-1< \phi\left ( M\right)<\phi\left (b^{p+1}\right)\\ \phi\left (b^{p+1}\right)-1< \phi\left ( a^{p}\right)<\phi\left (b^{p+1}\right) \\ \arg_{(\phi\left (b^{p+1}\right)-1,\phi\left (b^{p+1}\right))}(Z(a^p)-Z(b^{p+1}))<\phi(M) \nonumber \end{array}.
\end{gather}
It follows that $(a^p,M,b^{p+1}) \cap (\mk{T}_{a}^{st} \cup (a^q,M,b^{q+1})\cup \mk{T}_{b}^{st} )$ and $(a^p,M,b^{p+1}) \cup (\mk{T}_{a}^{st} \cup (a^q,M,b^{q+1})\cup \mk{T}_{b}^{st} )$ are contractible.
\end{lemma}
\bpr We start with the inclusion $\subset$. Assume that $\sigma \in (a^p,M,b^{p+1})$. Then $a^p,M,b^{p+1}$ are semi-stable and by table \eqref{middle M} we get \begin{gather} \label{middle M cap left right M1} \begin{array}{c} \phi\left ( a^{p}\right) < \phi\left (M\right)+1 \\ \phi\left ( a^{p}\right) < \phi\left (b^{p+1}\right) \\ \phi\left ( M\right) < \phi\left (b^{p+1}\right)\end{array}. \end{gather}
\ul{If $\sigma \in (a^q,M,b^{q+1})$ and $q<p$,} then $a^q,b^{q+1} \in \sigma^{ss}$ and $\phi(M)<\phi(b^{q+1})$, $\phi\left ( a^{q}\right) < \phi\left (b^{q+1}\right) $ a well. By \eqref{nonvanishing3} we have $\hom(b^{q+1}, a^p)\neq 0$ and $\hom^1(b^{p+1}, a^q)\neq 0$, therefore $\phi(M)<\phi(b^{q+1})\leq \phi(a^{p})$ and $\phi(b^{p+1})\leq \phi(a^{q})+1<\phi(b^{q+1})+1\leq \phi\left (a^{p}\right)+1$. Combining with \eqref{middle M cap left right M1} we obtain the system in the first row and first column in \eqref{middle M cap left right M sys}.
\ul{If $\sigma \in \mk{T}_{a}^{st}$}, then by Lemma \ref{middle M cap left M'} some of the systems on the second row of \eqref{middle M cap left right M sys} follows.
\ul{If $\sigma \in \mk{T}_{b}^{st}$}, then by Lemma \ref{middle M cap left M} some of the systems on the first row of \eqref{middle M cap left right M sys} follows (\eqref{middle M cap left M sys} implies \eqref{middle M cap left right M sys}). So we showed the inclusion $\subset$.
We show now the inclusion $\supset$. So let $a^p,M,b^{p+1}$ be semi-stable. If some of the systems on the second row of \eqref{middle M cap left right M sys} holds, then by \ref{middle M cap left M'} it follows that $\sigma \in (a^p,M,b^{p+1})\cap \mk{T}_{a}^{st}$. If the system in the first row and second column of \eqref{middle M cap left right M sys} holds, then Lemma \ref{middle M cap left M} ensures that $\sigma \in (a^p,M,b^{p+1})\cap \mk{T}_{b}^{st}$.
Thus, it remains to consider the first system in \eqref{middle M cap left right M sys}. We assume till the end of the proof that
\begin{gather} \label{middle M cap left right M2}
\begin{array}{c}\phi\left (a^{p}\right)-1< \phi\left ( M\right)<\phi\left (a^{p}\right)\\
\phi\left (a^{p}\right)-1< \phi\left ( b^{p+1}\right)-1<\phi\left (a^{p}\right) \end{array}.
\end{gather}
Lemma \ref{semi-stability of b} (c) ensures that
\begin{gather} \label{middle M cap left right M3} M'\in \sigma^{ss}; \qquad \phi(b^{p+1})-1<\phi(M')=\arg_{(\phi\left (a^{p}\right)-1,\phi\left (a^{p}\right))}(Z(a^p)-Z(b^{p+1}))<\phi(a^{p}). \end{gather}
Now we consider three cases.
If $\phi(M')>\phi(M)$, then \eqref{middle M cap left right M2} and \eqref{middle M cap left right M3} yield the first system in \eqref{middle M cap left M sys} is satisfied and then Lemma \ref{middle M cap left M} says that $\sigma \in (a^p,M,b^{p+1})\cap \mk{T}_{b}^{st}$.
If $\phi(M')<\phi(M)$, then by $\hom^1(b^{p+1}, M')\neq 0$ it follows that $\phi(b^{p+1})-1<\phi(M)$. Combining this inequality with \eqref{middle M cap left right M2} one easily shows that:
\begin{gather} \label{middle M cap left right M6}\begin{array}{c}\phi\left (b^{p+1}\right)-1< \phi\left ( M\right)<\phi\left (b^{p+1}\right)\\
\phi\left (b^{p+1}\right)-1< \phi\left ( a^{p}\right)<\phi\left (b^{p+1}\right) \end{array}. \end{gather}
Having obtained \eqref{middle M cap left right M6} we can use Lemma \ref{semi-stability of a} (c) and due to $\phi(M')<\phi(M)$ we derive the first system in \eqref{middle M cap left M' sys}. Thus Lemma \ref{middle M cap left M'} ensures that $\sigma \in (a^p,M,b^{p+1})\cap \mk{T}_{a}^{st}$.
Finally, if $\phi(M)=\phi(M')$, then due to \eqref{middle M cap left right M2} we can apply Lemma \ref{semi-stability of b} \textbf{(e)}, which says that $\sigma \in (a^p,M,b^{p+1})\cap (a^q,M,b^{q+1})$ (recall that $q<p$). So far we showed the first part of the lemma.
We explain now, using the obtained representation through the systems of inequalities \eqref{middle M cap left right M sys}, that $(a^p,M,b^{p+1}) \cap (\mk{T}_{a}^{st} \cup (a^q,M,b^{q+1})\cup \mk{T}_{b}^{st} )$ is contractible. The four systems correspond to four open subsets of $(a^p,M,b^{p+1}) \cap (\mk{T}_{a}^{st} \cup (a^q,M,b^{q+1})\cup \mk{T}_{b}^{st} )$ (see the last paragraph of the proof of Lemma \ref{T12Zcap(E_1)}). We denote these subsets by $S_{11} , S_{12}, S_{21}, S_{22}$, where $S_{ij}$ corresponds to the system in the $i$-th row and $j$-th column of \eqref{middle M cap left right M sys}. The proved part of the lemma is the equality $(a^p,M,b^{p+1}) \cap (\mk{T}_{a}^{st} \cup (a^q,M,b^{q+1})\cup \mk{T}_{b}^{st} )=\bigcup_{1\leq i,j \leq 2} S_{ij}$. The subset $S_{22}$ is contractible by Corollary \ref{contractible 1a}. The subsets $S_{11} , S_{12}, S_{21}$ are contractible since they are homeomorphic to convex subsets of $\RR^6$. For example $S_{11}$ is homeomorphic to $$\RR_{>0}^3\times \left \{ (\phi_0,\phi_1,\phi_2) \in \RR^3: \begin{array}{c}\phi_0-1< \phi_1<\phi_0\\
\phi_0-1< \phi_2-1<\phi_0 \end{array}\right \}. $$
One easily shows that $S_{11}\cap S_{12}$ is homeomorphic to $\RR_{>0}^3 \times \{\phi_0-1<\phi_1<\phi_2-1 < \phi_0\}$, hence it is contractible, and by Remark \ref{VK} we deduce that $S_{11}\cup S_{12}$ is contractible. Note that in $S_{12}$ we have $\phi(M)+1<\phi(b^{p+1})$ and in $S_{22}$ we have $\phi(M)+1>\phi(b^{p+1})$ , therefore $S_{12}\cap S_{22}=\emptyset$. Hence $S_{22}\cap (S_{11}\cup S_{12})=S_{22}\cap S_{11}$. One easily shows that $S_{22}\cap S_{11}$ is homeomorphic to:
\begin{gather} \RR_{>0}^3 \times \left \{ (\phi_0,\phi_1,\phi_2) \in \RR^3: \begin{array}{c} r_i>0\\ \phi_2-1 < \phi_1 < \phi_0 < \phi_2 \\ \arg_{(\phi_2-1,\phi_2)}(r_0 \exp(\ri \pi \phi_0) - r_2 \exp(\ri \pi \phi_2))<\phi_1\end{array} \right \},\end{gather}
which by Corollary \ref{contractible 2a} is contractible as well. Thus, we see that $S_{22}\cap (S_{11}\cup S_{12})$ is contractible, therefore by Remark \ref{VK} we see that $S_{22}\cup S_{11}\cup S_{12}$ is contractible. In $S_{11}$ and $S_{12}$ we have $\phi(M)<\phi(a^p)$ and in $S_{21}$ we have $\phi(M)>\phi(a^p)$, therefore $S_{21}\cap(S_{22}\cup S_{11}\cup S_{12})=S_{21}\cap S_{22}$. On the other hand, one easily shows (by drawing a picture) that the intersection $S_{21}\cap S_{22}$ is homeomorphic to $\RR_{>0}^3 \times \{\phi_2-1<\phi_0<\phi_1 < \phi_2\}$, which is contractible as well, and hence $S_{21}\cap(S_{22}\cup S_{11}\cup S_{12})$ is contractible. Applying Remark \ref{VK} again ensures that $S_{21}\cup S_{22}\cup S_{11}\cup S_{12} = (a^p,M,b^{p+1}) \cap (\mk{T}_{a}^{st} \cup (a^q,M,b^{q+1})\cup \mk{T}_{b}^{st} )$ is contractible. In Corollary \ref{middle M cup left right M coro} is shown that $\mk{T}_{a}^{st} \cup (a^q,M,b^{q+1})\cup \mk{T}_{b}^{st}$ is contractible and with one more reference to Remark \ref{VK} we prove the lemma.
\epr
\begin{coro} \label{middle M cup left right M} The set $(\_,M,\_)\cup \mk{T}_{a}^{st}\cup \mk{T}_{b}^{st}$ is contractible.
\end{coro}
\bpr Recall that $ (\_,M\_) =\bigcup_{q\in \ZZ}(a^q,M,b^{q+1})$ (see \eqref{middle M ab}). We will prove that for each $p\in \ZZ$ and for each $k\geq 1$ the set \eqref{middle M cup left right M1} below is contractible, and the corollary follows from Remark \ref{VK}:
\begin{gather} \label{middle M cup left right M1} \bigcup_{i=0}^k (a^{p-i},M,b^{p+1-i})\cup (\mk{T}_{a}^{st} \cup \mk{T}_{b}^{st} ).\end{gather}
In the previous lemma was shown that for $k=1$ and any $p\in \ZZ$ the set \eqref{middle M cup left right M1} is contractible. Assume that for some $k\geq 1$ this set is contractible for each $p\in \ZZ$. Take now any $p\in \ZZ$. We have
\begin{gather}\label{middle M cup left right M2}\bigcup_{i=0}^{k+1} (a^{p-i},M,b^{p+1-i})\cup (\mk{T}_{a}^{st} \cup \mk{T}_{b}^{st} )= (a^{p},M,b^{p+1})\cup \left (\bigcup_{i=1}^{k+1} (a^{p-i},M,b^{p+1-i})\cup (\mk{T}_{a}^{st} \cup \mk{T}_{b}^{st} ) \right ). \end{gather}
Proposition \ref{lemma for f_E(Theta_E)} and the induction assumption say that the two components on RHS of \eqref{middle M cup left right M2} are contractible. Since the intersection analyzed in Lemma \ref{middle M cap left right M} does not depend on $q$, we can write:
$$(a^{p},M,b^{p+1})\cap\left ( \bigcup_{i=1}^{k+1} (a^{p-i},M,b^{p+1-i})\cup (\mk{T}_{a}^{st} \cup \mk{T}_{b}^{st} ) \right )=(a^{p},M,b^{p+1})\cap\left ( (a^{p-1},M,b^{p})\cup (\mk{T}_{a}^{st} \cup \mk{T}_{b}^{st} ) \right ), $$
which by Lemma \ref{middle M cap left right M} is contractible. Now Remark \ref{VK} ensures that that \eqref{middle M cup left right M2} is contractible.
\epr
The next step is to glue $(\_,M,\_)\cup \mk{T}_{a}^{st}\cup \mk{T}_{b}^{st}$ and $(b^p,M',a^p)$. This is done in several substeps: Lemmas \ref{middle M' cap left M}, \ref{middle M' cap left M'}, \ref{semi-stability of a'}, \ref{semi-stability of b'}, which lead to Corollary \ref{middle M' cap left right middle M}. In the next two lemmas we prove inclusions in only one direction not equality of sets.
\begin{lemma} \label{middle M' cap left M} Let $p\in \ZZ$. If $\sigma \in (b^p,M',a^{p}) \cap \mk{T}_{b}^{st}$, then $b^p,M',a^{p}$ are semistable and: \begin{gather} \label{middle M' cap left M sys} \begin{array}{c}\phi\left (a^{p}\right)-1< \phi\left ( M'\right)<\phi\left (a^{p}\right)\\
\phi\left (a^{p}\right)-1< \phi\left ( b^{p}\right)<\phi\left (a^{p}\right) \end{array} \ \mbox{or} \ \begin{array}{c} \phi\left (a^{p}\right)-1< \phi\left ( M'\right)<\phi\left (a^{p}\right)\\ \phi(b^{p})<\phi(M')\end{array}. \end{gather}
\end{lemma}
\bpr In table \eqref{middle M} we see that $b^p,M',a^{p}$ are semi-stable and:
\begin{gather} \label{middle M' cap left M1} \begin{array}{c} \phi\left ( b^{p}\right) < \phi\left (M'\right) +1\\ \phi\left ( b^{p}\right) < \phi\left (a^{p}\right) \\ \phi\left ( M'\right) < \phi\left (a^{p}\right)
\end{array} \end{gather}
Recalling \eqref{T_43new}, we see that we have to consider three cases.
\ul{If $\sigma \in (M,b^j, b^{j+1})$}, then $M,b^j, b^{j+1}$ are semi-stable and from table \eqref{left right M} we see that $\phi(M)<\phi(b^{j})$ and $\phi(M)+1<\phi(b^{j+1})$. By $\hom^1(a^p,M)\neq 0$ and $\hom^1(b^{j+1},M')\neq 0$ (see \eqref{nonvanishing2}) we can write $\phi(a^p)\leq \phi(M)+1 <\phi(b^{j+1})\leq \phi(M')+1$, therefore (see also \eqref{middle M' cap left M1}) we get
\begin{gather} \label{middle M' cap left M2} \phi\left (a^{p}\right)-1< \phi\left ( M'\right)<\phi\left (a^{p}\right).\end{gather}
Since $\phi(b^p)<\phi(a^p)\leq \phi(M)+1<\phi(b^{j+1})$, due to \eqref{nonvanishing6} the inequality $p\leq j$ must hold.
If $j=p$, then the inequality $\phi(M)<\phi(b^{p})$ (coming from $\sigma \in (M,b^j, b^{j+1})$) implies $\phi(a^p)-1\leq \phi(M) <\phi(b^{p})$ and combining with \eqref{middle M' cap left M1} and \eqref{middle M' cap left M2} we obtain the first system in \eqref{middle M' cap left M sys}.
If $p<j$, then we have $\hom(a^p, b^j)\neq 0$ (see \eqref{nonvanishing4}) and $\hom^1(b^{j+1}, b^p)\neq 0$, hence $\phi(a^p)\leq \phi(b^j)<\phi(b^{j+1})\leq \phi(b^p)+1$ and again the first system in \eqref{middle M' cap left M sys} follows.
\ul{If $\sigma\in (b^m,b^{m+1},M')$}, then $b^m,b^{m+1}, M'$ are semistable and in table \eqref{left right M} we see that $\phi(b^m)<\phi(M')$, therefore $\phi(b^m)<\phi(M')<\phi(a^p)$ and $\hom(a^p,b^m)=0$. From \eqref{nonvanishing4} we deduce that $m\leq p$.
If $m=p$, then the incidence $\sigma\in (b^m,b^{m+1},M')$ gives $\phi(b^p)<\phi(M')$ and $\phi(b^{p+1})-1<\phi(M')$ (see table \eqref{left right M}), and from $\hom(a^p,b^{p+1})\neq 0$ we obtain $\phi(a^{p})-1<\phi(M')$, therefore the second system in \eqref{middle M' cap left M sys} holds.
Let $m<p$. Then we have $\phi(b^m)<\phi(b^{m+1})$ and $\phi(b^m)<\phi(M')$ (see table \eqref{left right M}). Using $\hom^1(a^p,b^m)\neq 0$ (see \eqref{nonvanishing4}) we deduce $\phi(a^p)\leq \phi(b^m)+1<\phi(b^{m+1})+1\leq \phi(b^{p})+1$ and $\phi(a^p)\leq \phi(b^m)+1<\phi(M')+1$, which combined with \eqref{middle M' cap left M1} produces the first system in \eqref{middle M' cap left M sys}.
\ul{If $\sigma\in (b^m, a^{m},b^{m+1})$}, then $b^m, a^{m},b^{m+1} \in \sigma^{ss}$ and in table \eqref{no M} we see that $\phi(b^m)+1<\phi(b^{m+1})$, hence Lemma \ref{from phases of big distance} and $b^{p} \in \sigma^{ss}$ imply $p=m$ or $p=m+1$. If $p=m+1$, then by \eqref{middle M' cap left M1} we obtain $\phi(b^m)+1<\phi(b^{m+1})<\phi(a^{m+1})$, and hence $\hom^1(a^{m+1},b^{m})=0$, which contradicts \eqref{nonvanishing4}. Therefore we have $m=p$. In table \eqref{no M} we see that $\phi(b^p)+1<\phi(b^{p+1})$ and $\phi(a^p)<\phi(b^{p+1})$. From $\hom^1(b^{p+1}, M')\neq 0$ it follows that $ \phi(b^p)+1<\phi(b^{p+1})\leq \phi(M')+1$ and $\phi(a^p)<\phi(b^{p+1})\leq \phi(M')+1$. These inequalities together with \eqref{middle M' cap left M1} produce the second system in \eqref{middle M' cap left M sys}.
\epr
\begin{lemma} \label{middle M' cap left M'} Let $p\in \ZZ$. If $\sigma \in (b^p,M',a^{p}) \cap \mk{T}_{a}^{st}$, then $b^p,M',a^{p}$ are semistable and: \begin{gather} \label{middle M' cap left M' sys}\begin{array}{c} \phi\left (b^{p}\right)-1< \phi\left ( M'\right)<\phi\left (b^{p}\right)\\ \phi\left (b^{p}\right)-1< \phi\left ( a^{p}\right)-1<\phi\left (b^{p}\right) \end{array} \ \mbox{or} \ \begin{array}{c} \phi\left (b^{p}\right)-1< \phi\left ( M'\right)<\phi\left (b^{p}\right)\\ \phi\left ( M'\right)+1<\phi\left (a^p\right)\end{array}. \end{gather}
\end{lemma}
\bpr In table \eqref{middle M} we see that $b^p,M',a^{p}$ are semi-stable and: \begin{gather} \label{middle M' cap left M'1} \begin{array}{c} \phi\left ( b^{p}\right) < \phi\left (M'\right) +1\\ \phi\left ( b^{p}\right) < \phi\left (a^{p}\right) \\ \phi\left ( M'\right) < \phi\left (a^{p}\right)
\end{array} \end{gather}
Recalling \eqref{T_12new}, we see that we have to consider the following three cases.
\ul{If $\sigma \in (M',a^j, a^{j+1})$}, then $M',a^j, a^{j+1}$ are semi-stable and $\phi(M')<\phi(a^{j})$, $\phi(M')+1<\phi(a^{j+1})$, $\phi(a^{j})<\phi(a^{j+1})$ (see table \eqref{left right M}). On the other hand $\phi(b^p)<\phi(M')+1$, hence $\hom(a^{j+1},b^p)=0$ and \eqref{nonvanishing4} implies that $p-1\leq j$. If $p-1=j$, then we have $\phi(M')+1<\phi(a^{p})$ and $\phi(M')<\phi(a^{p-1}) \leq \phi(b^p)$ (see also \eqref{nonvanishing4}) and combining with \eqref{middle M' cap left M'1} we derive the second system in \eqref{middle M' cap left M' sys}.
Let $p\leq j$. Then by \eqref{nonvanishing4} we have $\hom^1(a^{j+1},b^p)\neq 0$ and we can write $\phi(M')+1<\phi(a^{j+1})\leq\phi(b^p)+1$ and $\phi(a^p)\leq \phi(a^j)<\phi(a^{j+1})\leq \phi(b^p)+1$, therefore $\phi(M')<\phi(b^p)$ and $\phi(a^p)<\phi(b^p)+1$, which combined with \eqref{middle M' cap left M'1} amounts to the first system in \eqref{middle M' cap left M' sys}.
\ul{If $\sigma\in (a^m,a^{m+1},M)$}, then $a^m,a^{m+1}, M$ are semistable as well and in table \eqref{left right M} we see that $\phi(a^m)<\phi(M)$, $\phi(a^{m+1})<\phi(M)+1$, $\phi(a^m)<\phi(a^{m+1})$. Since $\hom(M',a^m)\neq0$ and $\hom(M, b^p)\neq 0$, it follows
that $\phi(M')\leq\phi(a^m)<\phi(M)\leq\phi( b^p)$ and hence (see also \eqref{middle M' cap left M'1}):
\begin{gather} \label{middle M' cap left M'11} \phi\left (b^{p}\right)-1< \phi\left ( M'\right)<\phi\left (b^{p}\right) \end{gather}
On the other hand, $\phi(a^m)<\phi(M)$ and $\hom(M, b^p)\neq 0$ imply that $\phi(a^m)<\phi( b^p)$ and $\hom(b^p,a^m)=0$. Now from \eqref{nonvanishing3} we deduce that $m<p$. If $m=p-1$, then we have $\phi(a^{p})<\phi(M)+1\leq\phi( b^p)+1$, which together with \eqref{middle M' cap left M'11} and \eqref{middle M' cap left M'1} amounts to the first system in \eqref{middle M' cap left M' sys}.
If $m<p-1$, then $\hom^1(a^p, a^m)\neq 0$ and $\hom( a^{m+1},b^p)\neq 0$ (see \eqref{nonvanishing4}). Therefore we have $\phi(a^p)\leq \phi(a^m)+1<\phi(a^{m+1})+1\leq \phi(b^p)+1$ and the first system in \eqref{middle M' cap left M' sys} follows again.
\ul{If $\sigma\in (a^m, b^{m+1},a^{m+1})$}, then $a^m, b^{m+1},a^{m+1}\in \sigma^{ss}$ and in table \eqref{no M} we see that $\phi(a^m)+1<\phi(a^{m+1})$, hence Lemma \ref{from phases of big distance} and $a^p \in \sigma^{ss}$ imply $p=m$ or $p=m+1$. If $p=m$, then by \eqref{middle M' cap left M'1} we obtain $\phi(b^m)+1<\phi(a^m)+1<\phi(a^{m+1})$, and hence $\hom^1(a^{m+1},b^m)=0$, which contradicts \eqref{nonvanishing4}. Thus, it remains to consider the case $m=p-1$. Now we have $\phi(a^{p-1})+1<\phi(a^{p})$ and $\phi(a^{p-1})<\phi(b^{p})$(see table \eqref{no M}), which together with $\hom(M',a^{p-1})\neq 0$ imply $\phi(M')+1<\phi(a^{p})$ and $\phi(M')<\phi(b^{p})$, hence we obtain the second system of inequalities in \eqref{middle M' cap left M' sys}.
\epr
\begin{lemma} \label{semi-stability of a'} Let $\sigma \in (b^p,M',a^{p})$ and let the following inequality hold: \begin{gather} \label{semi-stability of a' ineq} \phi\left (b^{p}\right)-1< \phi\left ( M'\right)<\phi\left (b^{p}\right). \end{gather}
Then we have the following:
{\rm \textbf{(a)}} $a^{p-1}\in \sigma^{ss}$ and $\phi(M')<\phi(a^{p-1})<\phi(b^{p})<\phi(a^{p})$.
{\rm \textbf{(b)}} If in addition to \eqref{semi-stability of a' ineq} we have $\phi\left ( M'\right)+1<\phi\left (a^p\right)$, then $\sigma \in (M',a^{p-1},a^{p})$.
{\rm \textbf{ (c)}} If in addition to \eqref{semi-stability of a' ineq} we have $\phi\left (b^{p}\right)-1< \phi\left ( a^{p}\right)-1<\phi\left (b^{p}\right),$ then $\sigma \in (a^{p-1},M,b^p)$.
\end{lemma}
\bpr (a) We apply Proposition \ref{phi_1>phi_2} (a) to the triple $ (b^p,M',a^{p})$ and since $a^{p-1}$ is in the extension closure of $M', b^{p}$ (by \eqref{short filtration 1}) it follows that $a^{p-1}\in \sigma^{ss}$, $\phi(M')\leq \phi(a^{p-1})\leq \phi(b^{p})$ . The inequality $\phi(M')<\phi(a^{p-1})<\phi(b^{p})$ follows from the given inequality \eqref{semi-stability of a' ineq}, formula \eqref{phase formula} and $Z(a^{p-1})=Z(M')+Z(b^{p})$. The inequality $\phi(b^{p})<\phi(a^{p})$ follows from $\sigma \in (b^p,M',a^{p})$ (see table \eqref{middle M}).
(b) From the given inequalities and (a) we have $\phi(M')<\phi(a^{p-1})$, $\phi\left ( M'\right)+1<\phi\left (a^p\right)$, and $\phi(a^{p-1})<\phi(a^{p})$, then table \eqref{left right M} shows that $\sigma \in (M',a^{p-1},a^{p})$.
(c) From Lemma \ref{two comp factors} applied to the Ext-triple $(b^p,M',a^{p}[-1]) $ and the triangle \eqref{short filtration 2} we obtain $M\in \sigma^{ss}$ and $\phi(a^{p})-1<\phi(M)<\phi(b^{p})$. In (a) we got $a^{p-1}\in \sigma^{ss}$ and $\phi(a^{p-1})<\phi(a^{p})$, therefore $\phi(a^{p-1})<\phi(M)+1$. In (a) we have also $\phi(a^{p-1})<\phi(b^p)$.
Looking at table \eqref{middle M} we see that $\sigma \in (a^{p-1},M,b^p)$.
\epr
\begin{lemma} \label{semi-stability of b'} Let $\sigma \in (b^p,M',a^{p})$ and let the following inequality hold: \begin{gather} \label{semi-stability of b' ineq} \phi\left (a^{p}\right)-1< \phi\left ( M'\right)<\phi\left (a^{p}\right). \end{gather}
Then we have the following:
{\rm \textbf{(a)}} $b^{p+1}\in \sigma^{ss}$ and $\phi\left (b^{p}\right)-1< \phi\left (a^{p}\right)-1<\phi\left (b^{p+1}\right)-1 <\phi\left ( M'\right)$.
{\rm \textbf{(b)}} If in addition to \eqref{semi-stability of b' ineq} we have $ \phi(b^{p})<\phi(M')$, then $\sigma \in (b^p,b^{p+1},M')$.
{\rm \textbf{ (c)}} If in addition to \eqref{semi-stability of b' ineq} we have $\phi\left (a^{p}\right)-1< \phi\left ( b^{p}\right)<\phi\left (a^{p}\right) $, then $\sigma \in (a^{p},M,b^{p+1})$
\end{lemma}
\bpr (a) We apply Proposition \ref{phi_1>phi_2} (b) to the triple $ (b^p,M',a^{p})$ and since $b^{p+1}[-1]$ is in the extension closure of $M', a^{p}[-1]$ (by \eqref{short filtration 1}) it follows that $b^{p+1}\in \sigma^{ss}$, $ \phi\left (a^{p}\right)-1\leq \phi\left (b^{p+1}\right)-1 \leq \phi\left ( M'\right)$. The inequality $ \phi\left (a^{p}\right)-1<\phi\left (b^{p+1}\right)-1 <\phi\left ( M'\right)$ follows from the given inequality \eqref{semi-stability of b' ineq}, formula \eqref{phase formula}, and $Z(b^{p+1}[-1])=Z(M')+Z(a^{p}[-1])$. The inequality $\phi(b^{p})<\phi(a^{p})$ follows from $\sigma \in (b^p,M',a^{p})$.
(b) From the given inequalities and (a) we have $\phi(b^{p})<\phi(b^{p+1})$, $\phi(b^{p})<\phi(M')$ and $\phi(b^{p+1})<\phi(M')+1$. Now in table \eqref{left right M} we see that $\sigma \in (b^p,b^{p+1},M')$.
(c) From Lemma \ref{two comp factors} applied to the Ext-triple $(b^p,M',a^{p}[-1]) $ and the triangle \eqref{short filtration 2} we obtain $M\in \sigma^{ss}$ and $\phi(a^{p})-1<\phi(M)<\phi(b^{p})$. In (a) we showed that $b^{p+1}\in \sigma^{ss}$ and $\phi(a^{p})<\phi(b^{p+1})$, $ \phi\left (b^{p}\right)<\phi\left (b^{p+1}\right)$. Now all the conditions determining $(a^{p},M,b^{p+1})$ (given in table \eqref{middle M}) are satisfied, hence $\sigma \in (a^{p},M,b^{p+1})$.
\epr
\begin{coro} \label{middle M' cap left right middle M} For any $p\in \ZZ$ the set $(b^p,M',a^{p}) \cap (\mk{T}_{a}^{st} \cup (\_,M,\_)\cup \mk{T}_{b}^{st} )$ consists of the stability conditions $\sigma$ for which $ b^p,M',a^{p} $ are semistable and: \begin{gather} \begin{array}{c}\phi\left (a^{p}\right)-1< \phi\left ( M'\right)<\phi\left (a^{p}\right)\\
\phi\left (a^{p}\right)-1< \phi\left ( b^{p}\right)<\phi\left (a^{p}\right) \end{array} \ \mbox{or} \ \begin{array}{c} \phi\left (a^{p}\right)-1< \phi\left ( M'\right)<\phi\left (a^{p}\right)\\ \phi(b^{p})<\phi(M')\end{array} \nonumber \\[-2mm] \label{middle M' cap left right middle M sys} \\[-2mm] \mbox{or} \begin{array}{c} \phi\left (b^{p}\right)-1< \phi\left ( M'\right)<\phi\left (b^{p}\right)\\ \phi\left (b^{p}\right)-1< \phi\left ( a^{p}\right)-1<\phi\left (b^{p}\right) \end{array} \ \mbox{or} \ \begin{array}{c} \phi\left (b^{p}\right)-1< \phi\left ( M'\right)<\phi\left (b^{p}\right)\\ \phi\left ( M'\right)+1<\phi\left (a^p\right)\end{array}.
\nonumber\end{gather}
It follows that $(b^p,M',a^{p}) \cap (\mk{T}_{a}^{st} \cup (\_,M,\_)\cup \mk{T}_{b}^{st} )$ and $(b^p,M',a^{p}) \cup (\mk{T}_{a}^{st} \cup (\_,M,\_)\cup \mk{T}_{b}^{st} )$ are contractible.
\end{coro}
\bpr Due to Lemmas \ref{middle M' cap left M} and \ref{middle M' cap left M'}, to show the inclusion $\subset$ it remains only to show that the incidence $\sigma \in (b^p,M',a^{p}) \cap (\_,M,\_)$ implies some of the systems in \eqref{middle M' cap left right middle M sys}. Assume that $\sigma \in (b^p,M',a^{p}) \cap (a^q,M,b^{q+1})$ for some $q\in \ZZ$. From table \eqref{middle M} we see that $b^p,M',a^{p},a^q,M,b^{q+1}$ are semi-stable and:
\begin{gather}\label{middle M' cap left right middle M1} \begin{array}{c} \phi\left ( b^{p}\right) < \phi\left (M'\right) +1\\ \phi\left ( b^{p}\right) < \phi\left (a^{p}\right) \\ \phi\left ( M'\right) < \phi\left (a^{p}\right)
\end{array} \ \mbox{and} \ \begin{array}{c} \phi\left ( a^{q}\right) < \phi\left (M\right)+1 \\ \phi\left ( a^{q}\right) < \phi\left (b^{q+1}\right) \\ \phi\left ( M\right) < \phi\left (b^{q+1}\right)\end{array}. \end{gather}
If $p\leq q$, then the non-vanishings $\hom(a^p,a^q)\neq 0$, $\hom^1(b^{q+1},M')\neq 0$, and $\hom(M,b^p)\neq 0$ (see Corollary \ref{nonvanishings}) together with \eqref{middle M' cap left right middle M1} imply the following inequalities $\phi(a^p)\leq \phi(a^q)<\phi(M)+1\leq \phi(b^p)+1$ and $\phi(a^p)\leq \phi(a^q)<\phi(b^{q+1})\leq \phi(M')+1$, which combined with \eqref{middle M' cap left right middle M1} amount to the system in the first row and the first column of \eqref{middle M' cap left right middle M sys}.
If $q<p$, then the non-vanishings $\hom(M',a^q)\neq 0$, $\hom(b^{q+1},b^p)\neq 0$, and $\hom^1(a^p,M)\neq 0$ together with \eqref{middle M' cap left right middle M1} imply the inequalities $\phi(M')\leq \phi(a^q)<\phi(b^{q+1})\leq \phi(b^{p})$ and $\phi(a^p)\leq \phi(M)+1<\phi(b^{q+1})+1\leq \phi(b^{p})+1$. The system in the second row and the first column in \eqref{middle M' cap left right middle M sys} follows.
So far we showed the incusion $\subset$.
Assume that $b^p,M',a^{p}\subset \sigma^{ss}$ and that \eqref{middle M' cap left right middle M sys} holds. Each of the systems in \eqref{middle M' cap left right middle M sys} contains in it the inequalities of $(b^p,M',a^{p})$ from table \eqref{middle M}, hence $\sigma \in (b^p,M',a^{p})$. Lemmas \ref{semi-stability of a'} and \ref{semi-stability of b'} ensure that $\sigma \in (\mk{T}_{a}^{st} \cup (\_,M,\_)\cup \mk{T}_{b}^{st} )$ as well and the first part of the corollary follows.
Now the arguments are analogous to those given in the end of the proof of Lemma \ref{middle M cap left right M}.
The four systems in \eqref{middle M' cap left right middle M sys} correspond to four open subsets of $(b^p,M',a^{p}) \cap (\mk{T}_{a}^{st} \cup (\_,M,\_)\cup \mk{T}_{b}^{st} )$. We denote these subsets by $S_{11} , S_{12}, S_{21}, S_{22}$, where $S_{ij}$ corresponds to the system in the $i$-th row and $j$-th.
The first part of the corollary and Remark \ref{VK} reduce the proof of the last statement to proving that
$\bigcup_{1\leq i,j \leq 2} S_{ij}$ is contractible .
All of $S_{11} , S_{12}, S_{21}, S_{22}$ are contractible since they are homeomorphic to convex subsets of $\RR^6$.
One easily shows that:
\begin{itemize}
\item $S_{11}\cap S_{12}$ is homeomorphic to $\RR_{>0}^3 \times \{\phi_2-1<\phi_0<\phi_1 < \phi_2\}$
\item $S_{21}\cap (S_{11}\cup S_{12})=S_{21}\cap S_{11}$ is homeomorphic to $\RR_{>0}^3 \times \{\phi_2-1<\phi_1<\phi_0 < \phi_2\}$
\item $S_{22}\cap (S_{11}\cup S_{12}\cup S_{21})=S_{22}\cap S_{21}$ is homeomorphic to $\RR_{>0}^3 \times \{\phi_0-1<\phi_1<\phi_2-1 < \phi_0\}$.
\end{itemize}
Since the obtained subsets of $\RR^6$ are convex, in particular contractible, it follows by Remark \ref{VK} that $\bigcup_{1\leq i,j \leq 2} S_{ij}$ is contractible. The corollary follows.
\epr
We can prove now Theorem \ref{main theo}:
\begin{theorem} $\st(D^b(Q ))$ is contractible.
\end{theorem}
\bpr Recall that $\st(\mc T)=\mk{T}_{a}^{st} \cup(\_,M',\_)\cup(\_,M,\_)\cup \mk{T}_{b}^{st} $ (see \eqref{st with T}). Recalling \eqref{middle M ab} we get:
\begin{gather} \label{one union} \st(D^b(Q ))=\mk{T}_{a}^{st} \cup(\_,M,\_)\cup \mk{T}_{b}^{st} \cup \bigcup_{k\in \ZZ}(b^{k},M',a^{k}). \end{gather}
Corollary \ref{middle M cup left right M} says that $ \mk{T}_{a}^{st}\cup(\_,M,\_)\cup \mk{T}_{b}^{st}$ is contractible and it remains to show that after adding $ \bigcup_{k\in \ZZ}(b^{k},M',a^{k})$ the result is still contractible.
We first show that for any two integers $q>p$ we have: \begin{gather} \label{one inclusion} (b^p,M',a^{p}) \cap (b^q,M',a^{q}) \subset (b^p,M',a^{p}) \cap (\mk{T}_{a}^{st} \cup (\_,M,\_)\cup \mk{T}_{b}^{st} ).
\end{gather}
Assume that $\sigma \in (b^p,M',a^{p}) \cap (b^q,M',a^{q})$. Then in table \eqref{middle M} we see that
\begin{gather}\label{two middle M'} \begin{array}{c} \phi\left ( b^{p}\right) < \phi\left (M'\right) +1\\ \phi\left ( b^{p}\right) < \phi\left (a^{p}\right) \\ \phi\left ( M'\right) < \phi\left (a^{p}\right)
\end{array} \ \mbox{and} \ \begin{array}{c} \phi\left ( b^{q}\right) < \phi\left (M'\right) +1\\ \phi\left ( b^{q}\right) < \phi\left (a^{q}\right) \\ \phi\left ( M'\right) < \phi\left (a^{q}\right)
\end{array}. \end{gather}
Since $p<q$, we have the non-vanishings $\hom(a^p,b^q)\neq 0$ and $\hom^1(a^{q},b^{p})\neq 0$ (see \eqref{nonvanishing4}). We combine with \eqref{two middle M'} as follows $\phi(a^p)\leq \phi(b^q)<\phi(a^q)\leq \phi(b^p)+1$ and $\phi(a^p)\leq \phi(b^q)< \phi(M')+1$, hence $\phi(a^p)-1< \phi(b^p)$ and $\phi(a^p)-1<\phi(M')$. In \eqref{two middle M'} we have also $\phi(b^p)<\phi(a^p)$ and $\phi(M')<\phi(a^p)$ and the system in the first row and the first column of \eqref{middle M' cap left right middle M sys} follows. Therefore by Corollary \ref{middle M' cap left right middle M} we get $\sigma \in (b^p,M',a^{p}) \cap (\mk{T}_{a}^{st} \cup (\_,M,\_)\cup \mk{T}_{b}^{st} )$ and we showed the inclusion \eqref{one inclusion}. This implies that for any $p\in \ZZ$ and any $n\geq 1$ holds the following equality:
\begin{gather} (b^p,M',a^{p}) \cap\left ( \mk{T}_{a}^{st} \cup (\_,M,\_)\cup \mk{T}_{b}^{st} \bigcup_{k=1}^n(b^{p+k},M',a^{p+k}) \right ) = (b^p,M',a^{p}) \cap\left ( \mk{T}_{a}^{st} \cup (\_,M,\_)\cup \mk{T}_{b}^{st}\right ) .
\nonumber
\end{gather}
In Corollary \ref{middle M' cap left right middle M} we showed that $(b^p,M',a^{p}) \cap (\mk{T}_{a}^{st} \cup (\_,M,\_)\cup \mk{T}_{b}^{st} )$ and $(b^p,M',a^{p}) \cup (\mk{T}_{a}^{st} \cup (\_,M,\_)\cup \mk{T}_{b}^{st} )$ are contractible (for any $p\in \ZZ$). Now using the equality above and Remark \ref{VK} one easily shows by induction that
$ \mk{T}_{a}^{st} \cup (\_,M,\_)\cup \mk{T}_{b}^{st}\cup \bigcup_{k=0}^n(b^{p+k},M',a^{p+k}) $ is contractible for any $p\in \ZZ$ and any $n\geq 1$. Applying Remark \ref{VK} again we deduce that the right-hand side of \eqref{one union} is contractible as well. Therefore $\st(D^b(Q ))$ is contractible.
\epr
|
\section{Introduction}
\label{intro}
Anharmonic oscillators or resonators involving Kerr-type nonlinearities have been an
interesting research topic due to their broad applications
in technology and fundamental physics \cite{kov}, \cite{dyk}.
Recently, this systems with strong parameters of the nonlinearity are also
attracting considerable attention for investigation of various quantum effects. The quantum dynamics of the Kerr nonlinear resonator (KNR) is naturally described by Fock states, which have a defnite number of
energy quanta. For systems with strong Kerr nonlinearity, leading to the photon-photon interaction in QED systems, the nonlinearity makes frequencies of transitions between
adjacent oscillatory energy levels different, i.e. nonlinearity effects in an anharmonic oscillator break the equidistance
of oscillatory energy levels. Thus, in the case of strong nonlinearity the oscillatory energy levels are well resolved
and spectroscopic selective excitation of transitions between Fock states at the level of a few quanta becomes to be possible. This regime has been demonstrated in the photon blockade (PB) \cite{Liu,A,Mir,gor,nori}, generation of Fock states and superposition states in KNRs \cite{nori,gev2,mod,sup} and on changing PB into electromagnetically-induced
transparency \cite{nori1}.
Note, that nonstationary PB in a context of Fock states generation was predicted in \cite{schmid} for short-time evolution of a kicked KNR and then in \cite{imam6}. The nonstationary Kerr PB is often referred to as a nonlinear optical-state truncation or nonlinear quantum scissors \cite{leon25}.
The important parameter responsible for KNR to reach quantum regimes is the ratio between the parameter of the Kerr-type nonlinearity and damping of the oscillatory mode. Thus, efficiency of quantum
nonlinear effects requires a high nonlinearity with respect to dissipation.
In this respect, strong nonlinearities on a few-photon level can be produced by interaction between photons and an atom in a cavity \cite{Birn2,Bish,Fink,Far}, in systems with interacting photons or polaritons in
arrays of cavities coupled to atoms or qubits \cite{Hart,Utus,Schm,Tom}, in optomechanical systems and Kerr type nonlinear cavities \cite{Lia,Ferr}. An important implementation of Kerr-type microwave resonator has been recently
achieved in the context of superconducting devices based
on the nonlinearity of the Josephson junction revealing nonlinear behavior even
at the single-photon level \cite{lang,hof,kir}.
Note, that most theoretical proposals on investigation of quantum effects in nonlinear Kerr resonators
are focused on using idealized cases of zero temperature
reservoirs since they can lead to the study of pure quantum effects. However, consideration of the reservoir at finite temperatures leads to applications in simulating of more realistic systems as well as to study of unusual quantum
phenomena connecting quantum engineering and temperature.
In this paper we investigate finite temperature reservoir effects in the dissipative KNR considering selective excitation in transitions between Fock states. For this goal
we investigate selective excitation in the pulsed regime, considering KNR driven by a sequence
of classical Gaussian pulses separated by time intervals. This system has been used in the field of PB and generation of Fock states \cite{gor,gev2,mod} as well as for production of superpositions states in a mesoscopic range \cite{gev1} and for demonstration of chaos at a low level of quanta \cite{chaos1,chaos2}.
It has been demonstrated in the cited papers, that in the specific pulsed regime the obtained results considerably differ from those derived for the case of continuous-wave driving. On the whole, the production of the Fock states
as well as the superposition of the Fock states can be controlled by shape of pulses and is realized for time-intervals exceeding the characteristic
time of decoherence. Thus, the other goal of the present paper is investigation of various regimes of selective excitation in dependence from the parameters of pulses. In this way, we demonstrate improving of the degree of quantum effects in KNR by applying the sequence of tailored pulses. This approach was recently exploited for formation of high degree continuous-variable entanglement in the nondegenerate optical parametric
oscillator \cite{adam,adam2}
Our consideration is based on the Hamiltonian of KNR coupled with finite temperature reservoir and the master
equation for the reduced density matrix. We focus on analysis of the mean photon number, the probability distributions of photons and the second-order correlation functions of photons for zero-delay time that describes quantum statistics of oscillatory mode in thermal Kerr resonator. It should be mentioned that thermal reservoir effects have been widely investigated in nonlinear optical processes, particularly, for the resonance fluorescence in monochromatic field \cite{res,kry,env} and for the resonance fluorescence in bichromatic field \cite{kry1,kry2,kry3}.
The paper is arranged as follows. In Sec. II we describe periodically pulsed KNR coupled with a finite temperature reservoir. In Sec. III we consider one-photon and two-photon selective excitations on the base of populations of photon-number states and the second-order correlation functions. In subsection 3.3 we also discuss dynamics of Fock states populations in dependence from both the duration of pulses and amplitude of pump field. We summarize our results in Sec. IV.
\section{ Kerr nonlinear resonator coupled with thermal reservoir}
In this section, we give the theoretical description of the
system. The Kerr nonlinear resonator driven by the field at the central frequency $\omega$ and interacting with a reservoir
is described by the following Hamiltonian:
\begin{eqnarray}
H=\hbar \omega_{0}a^{+}a + \hbar \chi a^{+{2}}a^{2}+ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\nonumber\\
+ \hbar f(t)({\Omega} e^{-i\omega t}a^{+} + {\Omega^{*}}e^{i\omega t}a)+H_{loss}.~~~~~
\label{H}
\end{eqnarray}
Here, time-dependent coupling constant $\Omega f(t)$ proportional to the driving field amplitude consists of the Gaussian pulses with the duration $T$ separated by time intervals $\tau$
\begin{equation}
f(t)=\sum{e^{-(t - t_{0} - n\tau)^{2}/T^{2}}}, \label{driving}
\end{equation}
while $a^{+}$, $a$ are the creation and annihilation operators, $ \omega_{0} $ is an oscillatory frequency, $\chi$ is the
nonlinearity parameter.
$H_{loss}=a \Gamma^{+} + a^{+}\Gamma $
is responsible for the linear losses of oscillatory states, due to
couplings with heat reservoir operators giving rise to the
damping rate $\gamma$.
We trace out the reservoir degrees of freedom in the Born-Markov limit assuming that system and environment are
uncorrelated at initial time t = 0. This procedure leads to the master
equation for the reduced density matrix. The master equation within the
framework of the rotating-wave approximation, in the interaction
picture corresponding to the transformation $\rho \rightarrow e^{-i\omega a^{+} a t}\rho e^{i\omega a^{+} a t} $ reads as
\begin{equation}
\frac{d \rho}{dt} =-\frac{i}{\hbar}[H_{eff}, \rho] +
\sum_{i=1,2}\left( L_{i}\rho
L_{i}^{\dagger}-\frac{1}{2}L_{i}^{\dagger}L_{i}\rho-\frac{1}{2}\rho L_{i}^{\dagger}
L_{i}\right)\label{master},
\end{equation}
where $L_{1}=\sqrt{(n_{th}+1)\gamma}a$ and $L_{2}=\sqrt{n_{th}\gamma}a^+$ are the
Lindblad operators, $\gamma$ is a dissipation rate, and $n_{th}$ denotes the mean
number of quanta of a heat bath, $n_{th}=\frac{1}{e^{\hbar \omega/ kT_{B} }-1}$. The effective Hamiltonian reads as
\begin{equation}
H_{eff}=\hbar \Delta a^{+}a + \hbar \chi a^{+2}a^{2} +
\hbar f(t)(\Omega a^{+} + \Omega^{*}a),\label{Hamiltonian}
\end{equation}
where $\Delta=\omega_{0} -\omega$ is the detuning
between the mean frequency of the driving field and the frequency of the
oscillator.
Note, that analytical results for a dissipative driven nonlinear oscillator in continius-wave, steady-state regimes have
been obtained in terms of the solution of the Fokker-Planck equation for the quasiprobability distribution function $P(\alpha, \alpha^{*})$ in complex
P-representation \cite{drum}. This approach based on the
method of potential equations leads to the analytic solution for the quasiprobability distribution function $P(\alpha, \alpha^{*})$ within the framework
of an exact nonlinear treatment of quantum
fluctuations. In this way, the probability distribution of
photon-number states and Wigner functions have been also obtained \cite{a33,a34,kh}.
This model seems experimentally feasible and can be realized in
several physical systems. Particularly, the effective Hamiltonian (\ref{Hamiltonian}) (with $f(t)=1$ ) describes a
qubit off-resonantly coupled to a driven cavity. In fact, it is well known that the Hamiltonian of a two-level atom interacting with the cavity mode in the dispersive approximation, if the two-level system remains in its ground state, can be reduced to the effective Hamiltonian (\ref{Hamiltonian}). This model also describes a nanomechanical oscillator with $a^{\dagger}$ and $a$ raising and lowering operators related to the position and momentum operators of a mode of quantum motion. The other implementation is a transmission-line resonator involved the Josephson
junction with quadratic part of the Josephson potential. In this case the $a^{+}$ and $a$ raising and lowering operators describe the normal mode of the resonator junction circuit \cite{bias}.
Kerr-like systems seems to be important also for production of quantum correlation and maximally entangled stated. They were discussed, for instance, in \cite {leonski1,said,leonski2,said2}.
\section{Selective excitation of Fock states}\label{SecPD}
In the absence of any driving, states of the Hamiltonian (\ref{Hamiltonian})
are the Fock photon states $|n\rangle$ which are spaced in the energies $E_{n} = E_{0} +
\hbar\omega_{0} n + \hbar\chi n(n-1)$ with $n = 0, 1, ...$. The
levels form an anharmonic ladder with
anharmonicity that is given by $E_{21}-E_{10}=2\hbar\chi$. For strong
nonlinearities
$\chi/\gamma >1$ the nonlinear shifts of oscillatory energetic levels exceed the radiative widths of these levels. In this case the selective resonance excitations of Fock states in the transitions $|m\rangle\rightarrow|n\rangle$ are possible due to interaction with driving field. Considering the various selective excitations from vacuum state, in the multiphoton transitions $|0\rangle\rightarrow|n\rangle$, we obtain the resonance n-photon transitions between oscillatory initial and final states on the frequencies $n\omega_n=E_{n0}=n\omega_0+\chi n(n-1)$. Thus, the detuning for the resonant frequencies $\Delta_{n}=\omega_{0} -\omega_{n}$ is calculated as $\Delta_{n}=-\chi(n-1)$. For one-photon transition, $n=1$, and $E_{10}=\hbar \omega_0$, the resonance frequency is $\omega_1=\omega_0$,
for two-photon transition, $n=2$, and $E_{20}=2 \hbar \omega_0 + 2 \chi $, the two-photon resonance frequency is $\omega_2=\omega_0+ \chi $, while for $n=3$ , $E_{30}=3 \hbar \omega_0 + 6 \chi $, the resonance frequency is $\omega_3=\omega_0+ 2\chi $.
Below, we concentrate on quantum regimes
for the parameters leading to resolved oscillatory energy levels calculating the mean photon number, the photon-number distributions and the second-order correlation functions for photonic mode.
Considering the pulsed regimes of Kerr nonlinear reservoir we assume that the spectral widths of pulses should be smaller
than the nonlinear shifts of the oscillatory energy levels. It means that the
duration of pulses should be larger than $1/\chi $. Thus, for strong
nonlinearities
$\chi/\gamma >1$, we arrive to the following inequalities for the duration of Gaussian pulses $1/\gamma > T > 1/\chi$.
Note, that the temporal pulse separation is
larger than the cavity photon lifetime.
We solve the master equation Eq. (\ref{master}) numerically based on quantum
state diffusion method. The applications of this method for studies of driven anharmonic oscillator
can be found in \cite{gev2,mod,gor,chaos1}.
In the calculations, a finite basis of number states $|n\rangle$ is kept large
enough (where $n_{max}$ is typically 50) so that the highest energy states are
never populated appreciably. In the following the main photon number and the distributions of photon numbers
$P(n)=\langle n|\rho|n\rangle$ will be analyzed for various level of thermal photons.
We also turn to calculation of the normalized second-order correlation function for zero delay time $ g^{(2)}$ defined as:
\begin{equation}
g^{(2)}(t)=\frac{\langle a^{\dagger}(t)a^{\dagger}(t)a(t)a(t)\rangle}{(\langle a^{\dagger}(t)a(t)\rangle)^2}
\end{equation}
for observation of photon statistics and photon correlation effects. Moreover, the correlation function is expressed through the mean photon number fluctuations as $\langle (\Delta n)^2\rangle = \langle n\rangle + \langle n\rangle ^2(g^{(2)}-1)$.
Thus, the condition $g^{(2)}<1$ corresponds to the sub-Poissonian statistics, $\langle (\Delta n)^2\rangle < \langle n\rangle$ .
In the pulsed regime these quantities are nonstationary and exhibit a periodic time-dependent behavior, i.e. repeat the periodicity of the pump laser in an over transient regime.
Besides this, the results depend on the parameters of Gaussian pulses such as the amplitude, the duration of pulses and the time-interval between them, dissipation rates and Kerr-interaction coupling. Thus, we investigate production of photonic states for an arbitrary interaction time-intervals including also time-intervals
exceeding the characteristic time of dissipative processes, $t\gg\gamma^{-1}$.
In the cited paper \cite{gor}, only the case of KNR with the mean number of reservoir photons $ N=0.001$ has been considered. In addition to this investigation, bellow we present systematic analyse of KNR for the wide range of thermal photon numbers including also vacuum case, $N=0$.
\begin{figure}
\resizebox{0.5\textwidth}{7.5cm}
\includegraphics{fig1.eps}
}
\caption{The selective excitations in KNR at zero temperature for both cases of the detuning $\Delta_{1}=0$ (a), (b) and $\Delta_{2}=\omega_{0} -\omega_{2}=-\chi$ (c), (d). The populations of photon-number states versus time-intervals (a), (c). The averaged photon numbers and the second-order correlation function versus time-intervals (b), (d). The parameters are: (a), (b)
$\chi /\gamma = 15$, the maximum amplitude of pump field $\Omega/\gamma =6$; (c), (d)
$\chi /\gamma = 30$, the maximum amplitude of pump field $\Omega/\gamma =12$ . The mean number of reservoir photons $n_{th}=0$ and
$\tau = 5.5{\gamma}^{-1}$, $T=0.4{\gamma}^{-1}.$}
\label{0}
\end{figure}
\subsection{The case of reservoir at zero temperature}
In this subsection, we consider the case of reservoir at zero temperature when only pure quantum effects occur. The typical results at the detuning $\Delta=\Delta_{1}=0$, i. e. at the frequency $\omega = \omega_{0}$, corresponding to one-photon transition $|0\rangle\rightarrow|1\rangle$ are depicted in Figs.1 for
$T = 0 K$.
The populations
$P_0, P_1$ of the photon-number states $|n\rangle$ versus time-intervals is shown in Fig.\ref{0} (a), while the mean photon number and the correlation function are displayed in Fig.\ref{0} (b) for zero detuning $\Delta=0$. These results show the selective excitations of the Fock state $|1\rangle$. In this case the excitation power is also small therefore the high excitation numbers have not been occupied. In the result only the $|1\rangle$ is effectively excitated. As we see, the population $ P_1$ reachs up to 0.8 at the maxima of the average number of photons. The nonstationary
correlation function versus dimensionless time intervals is shown in Fig.\ref{0} (b)
with the curve of mean photon number. As we see, the average number of photons at frequency $\omega_{0}$ follows almost the train of Gaussian pulses. The second small peak corresponds to Rabi oscillations and the broadening of the shapes in comparisson with Gaussian envelopes are due to the radiative decays of the mode in a resonator. The mean photon number equals to $\langle n\rangle$=0.9 and is approximatelly presented as $\langle
n\rangle=P_1+2P_2$=0.9 in this regime. Thus, time-dependence of ${\langle n(t)\rangle}$ repeats the behaviour of the population $ P_1$.
During the pulses, if $P_1$ reachs to the maximum, the probability of generation of
a second photon in the mode at the frequency $\omega = \omega_{0}$ is suppressed that is demonstrated in the behavior of $g^{(2)}(t)$. In this case oscillatory mode acquires
sub-poissonian statistics with the second-order correlation function $ g^{(2)}<
1$. As calculations show,
$g^{(2)}=0.12$ at the maximum values ${\langle n(t)\rangle}=0.9$ and is zero at the second small peak. This results are in accordance with the result that for the pure $|1\rangle$ Fock
state the normalized second-order correlation function is zero.
\begin{figure}
\resizebox{0.5\textwidth}{7.5cm}
\includegraphics{fig2.eps}
}
\caption{Reservoir effects in the selective excitations in KNR for two cases of the detuning: $\Delta_{1}=0$ (a), (b) and $\Delta_{2}=\omega_{0} -\omega_{2}=-\chi$ (c), (d). The populations of photon-number states (a), (c). The averaged photon numbers and the second-order correlation function (b), (d). The parameters are: (a), (b)
$\chi /\gamma = 15$, the maximum amplitude of pump field $\Omega/\gamma =6$; (c), (d)
$\chi /\gamma = 30$, the maximum amplitude of pump field $\Omega/\gamma =12$ . The mean number of reservoir photons $n_{th}=0.1$ and the parameters of pulses are:
$\tau = 5.5{\gamma}^{-1}$, $T=0.4{\gamma}^{-1}.$}
\label{01}
\end{figure}
\begin{figure}
\resizebox{0.5\textwidth}{7.5cm}
\includegraphics{fig6.eps}
}
\caption{The maximal values of population $P_1(t)$ versus $n_{th}$. The parameters are: $\Delta=0$, $\chi /\gamma = 15$, the maximum amplitude of pump field $\Omega/\gamma =6$ ,
$\tau = 5.5{\gamma}^{-1}$, $T=0.4{\gamma}^{-1}.$}
\label{Stat}
\end{figure}
The results for the other operational regime for zero-temperature reservoir are depicted in Fig.1 (c),(d). Here, we assume
the regime of two-photon excitation of resonator mode, choosing $\Delta=\Delta_{2}=\omega_{0} -\omega_{2}$, at the frequency of resonance transition $\omega_2=\omega_0+ \chi $. Thus, in this case the detuning is $\Delta/\gamma=-\chi/\gamma$.
These results indicate an effective two-photon selective excitation as illustrated in Fig.1 (c). As we see,
in this regime of two-photon excitation the maximal population $P_2=0.64$ exceeds the maximal population of one-photon state $P_1=0.3$ . The time evolution of the mean photon number versus
dimensionless time is depicted in Fig. \ref{0} (d) with the plot of the second-order correlation function. In this case, the maximal value of mean photon numbers reachs to ${\langle n(t)\rangle}=1.9$ while the level of photon-number correlation at this time interval is $g^{(2)}=0.6$ and the variance of photon number fluctuations equals to $\langle (\Delta n)^2\rangle =0.24 \langle n\rangle $, i.e. the oscillatory mode has sub-poissonian statistics.
The result on the second-order correlation function is in accordance with analytical result for pure state $|2\rangle$. Indeed, it is easy to calculate that $\langle 2|(a^{+})^2 a^2|2\rangle = 2$ and thus $g^{(2)}=0.5$ for $|2\rangle$ Fock state.
\subsection{Reservoir effects in selective excitations of Fock states}
Considering the reservoir at finite temperatures we realize that temperature effects can lead to decreasing of the populations in comparison to the case of zero-temperature reservoir on one side and change photon statistics of mode on the other side. The results for thermal photons with $n_{th}=0.1$ are shown in Fig. \ref{01} (a),(c). In order to illustrate the difference between the case of zero-temperature resonator we assume here the parameters
as in the previous case depicted in Fig.\ref{0} (a),(c). As we see, the populations only are slightly decreased in the presence of temperature noise with $n_{th}=0.1$ for two considered regimes (a) and (c). We show in Fig. \ref{01} (b), (d) how the mean photon numbers and the second-order correlation function explicitly depend on the time-interval. As it can be seen, the maximal values of mean photon numbers are approximately the same as in the case of pure resonator, while quantum statistics of oscillatory mode is considerably changed due to the thermal noise. Indeed, in this case initial time-evolution of the system until a time-interval corresponding to the first pulse coming to cavity is described by the master equation without the interaction part, $\Omega f(t)=0$. Thus, in this range the quantity $g^{(2)}$ describes the statistics of mode of an anharmonic oscillator in thermal reservoir. It displays time-dependent fluctuations around the level $g^{(2)}=2$ that correspond to the statitics of thermal light mode. From the results in Fig. \ref{01} (b), (d) it can also be infered that the correlation function is sharply increasing in the fronts of pulses. However, the peak values strongly depend on the parameters of nonlinearity and amplitude of pump field. So, for Fig.\ref{01} (b) the maximal value is observed as $g^{(2)}=0.45$, while for Fig.\ref{01}(d) the correation function equals to $g^{(2)}=7$, (for ${\langle n(t)\rangle}=0.5$), and corresponds to super-poissonan statistic of mode, $\langle (\Delta n)^2\rangle =4 \langle n\rangle $ at this time-interval.
\begin{figure}
\resizebox{0.5\textwidth}{7.5cm}
\includegraphics{fig3.eps}
}
\caption{Time-evolution of the averaged photon number and the second-order correlation function for two regimes of KNR. The parameters are: (a),(c) $\Delta=0$,
$\chi /\gamma = 15$, $\Omega/\gamma =6$; (b),(d) $\Delta_{2}=\omega_{0} -\omega_{2}=-\chi$,
$\chi /\gamma = 30$, $\Omega/\gamma =12$. The mean number of reservoir photons: (a),(b) $n_{th}=0.58$; (c),(d) $n_{th}=1.9$. The other parameters are:
$\tau = 5.5{\gamma}^{-1}$, $T=0.4{\gamma}^{-1}.$}
\label{05819}
\end{figure}
It is interesting to analyse excitations of resonator mode for high levels of thermal photon numbers. In this way, in Fig. \ref{Stat} we plot the maximal values of the population $P_1$ in dependence on $n_{th}$ for the operational regime of KNR depicted in Fig.\ref{0} and Fig.\ref{01}. As we see the population $P_1$ is strongly decreasing with increasing of the thermal photon level.
The correlation functions and the mean photon numbers for high levels of thermal photons are considered in details for two cases, with $n_{th}=0.58$ and $n_{th}=1.9$, in Fig. \ref{05819}. We analyse finite temperature reservoir effects for one-photon and two-photon excitation regimes, corresponding to two values of the detuning: $\Delta_{1}=0$ and $\Delta_{2}=\omega_{0} -\omega_{2}=-\chi/\gamma$.
As we see, in both these cases $g^{(2)}$ describes mainly the statistics of dissipative oscillatory mode with the dips that correspond to the peaks of the averaged photon numbers. Particularly, for the regime depicted in Fig. \ref{05819} (a) the peak of mean photon number $n=0.8$ while the corresponding dip on $g^{(2)}=0.3$ shows subpoissonian staistics of mode, $\langle (\Delta n)^2\rangle =0.56 \langle n\rangle $. The location of these dips and peaks in Fig. \ref{05819} are determined by time-intervals of Gaussian pulses. We also found that these dips and peaks deacrese, if the level of thermal noise increases.
\subsection{Pulse-parameter effects in dynamics of the populations}
It should be mentioned that the parameters of the Gaussian pulses in above consideration are free parameters and they might be chosen in order to control pulsed selective excitation of resonator mode. In the end of this section we shortly illustrate behavior of the populations by increasing the duration of pulses on one side and by increasing the amplitude of driving field on the other side for zero-temperature reservoir. In this way, we describe two regimes of excitation: one-photon and two-photon excitations of oscillatory mode.
\begin{figure}
\resizebox{0.5\textwidth}{4cm}
\includegraphics{fig4.eps}
}
\caption{Time-evolution of the populations $P_1$ and $P_2$ . The parameters are: (a) $\chi /\gamma = 15$, the maximum amplitude of pump field $\Omega/\gamma =8$ , the detuning $\Delta=0$; (b) $\chi /\gamma = 30$, the maximum amplitude of pump field $\Omega/\gamma =12$ , the detuning $\Delta_{2}=\omega_{0} -\omega_{2}=-\chi$.
The other parameters are: $\tau = 4{\gamma}^{-1}$, $T=0.8{\gamma}^{-1}.$}
\label{08T}
\end{figure}
\begin{figure}
\resizebox{0.5\textwidth}{4cm}
\includegraphics{fig5.eps}
}
\caption{Time-evolution of the populations $P_1$ and $P_2$ . The parameters are: (a) $\chi /\gamma = 15$, the maximum amplitude of pump field $\Omega/\gamma =14$ , the detuning $\Delta=0$; (b) $\chi /\gamma = 30$, the maximum amplitude of pump field $\Omega/\gamma =25$ , the detuning $\Delta_{2}=\omega_{0} -\omega_{2}=-\chi$.
The other parameters are: $\tau = 5.5{\gamma}^{-1}$, $T=0.4{\gamma}^{-1}.$}
\label{pulse_effect}
\end{figure}
For the case when the detuning $\Delta=\Delta_{1}=0$, i.e. the frequency of driving field $\omega = \omega_{0}$ and only one-photon transition $|0\rangle\rightarrow|1\rangle$ is effectively realized, the results for the population of $|1\rangle$ Fock
state are shown in Fig. \ref{08T} (a) for the duration of pulses $T=0.8{\gamma}^{-1}$.
In this case, only two energetic levels of mode are effectively
involved in the Rabi-like oscillations of the populations $P_0$ and
$P_1$ of vacuum and first excitation number states that is
demonstrated in Fig.\ref{08T} (a). Comparing this result with analogous one depicted in Fig. \ref{0}(a) we conclude that the number of Rabi oscillations increases with increasing of the duration of pulses. The population $P_2$ is shown in Fig. \ref{08T} (b) for the case of two-photon excitation assuming $\Delta=-\chi$. As we see, time-dependence of the population displays Rabi oscillation in difference from the population shown in Fig.\ref{0} (c) for the case of more short pulses. The analogous situation takes place for photon populations in increasing of the amplitude of pump field in comparisson with the case of Fig. \ref{0}(c). The results for the case of more strong pump field, $\Omega/\gamma =14$ and $\Omega/\gamma =25$ are presented in Figs. \ref{pulse_effect}. As we see, the Rabi-like oscillations of the populations increase if the pump amplitude increases.
\section{Conclusion}\label{Conclusion}
In conclusion, we have investigated selective excitation of photon-states in a lossy Kerr-nonlinear resonator at finite temperatures driven by a sequence of Gaussian pulses. In quantum regime realized for strong Kerr nonlinearities with respect to the rate of damping of
the oscillatory mode one-photon excitation $|0\rangle\rightarrow|1\rangle$ at the frequency $E_{10}=\hbar \omega_0$, as well as two-photon excitation $|0\rangle\rightarrow|2\rangle$
at the frequency $E_{20}=2 \hbar \omega_0 + 2 \chi $ have been analyzed in details.
We have
demonstrated that the larger photon-number populations of the
resonator can be reached if shaped pulses are implemented.
The dynamics of mean number of photons and populations of Fock states follow almost to the train of Gaussian pulses and display Rabi oscillations and the broadening of the shapes in comparison with Gaussian envelopes. These oscillations are due to the radiative decays of the mode in a resonator for time-intervals between pulses. We have illustrated the increasing of the number of Rabi oscillations with increasing of the duration of pulses as well as with increasing of the amplitude of driving field.
We have demonstrated that temperature effects leads to decreasing of the populations in comparison with the zero-temperature case on one side and also change cardinally photon statitics of mode on the other side. Particularly, this behaviour on the maximal values of the population $P_1$ in dependence on $n_{th}$ has been illustrated in Fig. 3 for wide range of the thermal photon level. The quantum statistics of mode is described on the framework of non-stationary second-order correlation function $g^{(2)}$ and the mean photon number fluctuations $\langle (\Delta n)^2\rangle$. In this way, we have studied engineering of photon statistics via a thermal reservoir that is illustrated in Fig. 2(b), (d) and Fig. \ref{05819}. In case of high level of reservoir photons, $g^{(2)}$ describes mainly the statistics of dissipative mode of the anharmonic oscillator in thermal reservoir with the dips that correspond to the peaks of the averaged photon numbers. The location of these dips showing photon antibunching are determined by time-intervals of Gaussian pulses.
We acknowledge support from the Armenian State Committee of Science, the
Project No.13-1C031. G. Yu. K. acknowledges discussions with Lock Yue Chew, I. A. Shelykh and S. R. Shahinyan.
|
\section{Introduction}\label{sec: 1}
It has been debated that Alfv$\acute{\rm e}$n field-line resonances are directly related to various phenomena occurring at the auroral region, such as auroral activation and deformation [Samson et al., 1992; Lysak and Song, 2008; Rae et al., 2009], ionospheric density disturbances [Rankin et al., 2004], auroral particle acceleration and turbulences [G$\acute{\rm e}$not et al., 2000; Chaston et al., 2003], and high-$\beta$ plasma instability in the magnetosphere [Erickson et al., 2000]. A steep cavity of the Alfv$\acute{\rm e}$n velocity has been known to exist below a height of $\approx 6000$ km, where so-called ionospheric Alfv$\acute{\rm e}$n resonant (IAR) waves are excited [e.g., Carlson et al., 1998]. The IAR waves are distributed in a high frequency range of 1--10 Hz [Hebden et al., 2005]. On the other hand, there are global field-line resonant modes in a low frequency range of 10--100 mHz [e.g., Rae et al. 2009] that lies in a slower cavity at the magnetospheric side. These two work on a low-$\beta$ plasma medium with magnetosphere-ionosphere (MI) coupling.
Generation mechanisms of the inertial Alfv$\acute{\rm e}$n waves with a field-line wavelength of $<1000$ km and a perpendicular one of $<1$ km at the ionospheric cavity region have been vigorously studied by a particle-in-cell simulation [G$\acute{\rm e}$not et al., 2004] and satellite observations [Chaston et al., 2006]. By assuming the Alfv$\acute{\rm e}$n velocity as $v_{\rm A}\approx 1000$ km/s, they roughly correspond to the above IAR modes with frequency of $>1$ Hz; similarly, waves with $<100$ mHz correspond to a long wavelength range of $>10^4$ km. Accompanied by auroral activation at the substorm onset, magnetic pulsations (20--80 s) called Pi 1 and 2 [Holter et al., 1995; Park et al., 2012] and auroral kilometric electromagnetic radiations (AKR) [Morioka et al., 2010] have been known to occur. These phenomena are good implications for the existence of the above field line resonant modes. The IAR wave trapping can be followed by an enhancement of field-aligned currents, which leads to excitation of the inertial Alfv$\acute{\rm e}$n waves and electron acceleration.
How much knowledge we get about generation of the IAR waves and inertial Alfv$\acute{\rm e}$n waves from the viewpoint of natural conditioning? Is it the unique solution that an impulse propagates from the magnetosphere and some waves are trapped in the ionospheric cavity, although the related auroral arc slowly develops? We have made a series of studies on the characteristics of the Alfv$\acute{\rm e}$n resonant modes destabilized in a full field line system with MI coupling. A three-dimensional magnetohydrodynamic (MHD) simulation demonstrated that the feedback instability of field line resonant modes induces vortex structures at the magnetic equator [Watanabe, 2010]; the system is a slab magnetic field and $v_{\rm A} = {\rm const}$. Considering a dipole magnetic field and steep gradients of $v_{\rm A}$, a linear analysis revealed that low-frequency modes [HW, 2011], IAR high-frequency modes, and hybrid Alfv$\acute{\rm e}$n resonant (HAR) modes [HW, 2012] become feedback unstable; the third mode is generated by a coupling of the former two. If one includes the other, or, if the IAR and HAR modes naturally grow in the process of nonlinear evolution of the low-frequency modes, a problem could be unified: the relationship between the enhancement of global convections and auroral particle acceleration along with turbulent structures.
The purpose of this study is to comprehend the nonlinear evolution of Alfv$\acute{\rm e}$n waves underlying phenomena related to auroral intensification such as Pi 1 and 2, AKR, and electron acceleration. We clarify how the position and timing of wave growth change in response to changes in the $v_{\rm A}$ cavities. The main points are summarized below. i) On the time lag of wave amplification at the ionosphere and the magnetic equator. In case of $v_{\rm A}={\rm const}$, although the primary (feedback) instability is triggered by the MI coupling region, a secondary increase in flow and current rapidly starts at the magnetic equator side. We clarified this mechanism in a quantitative analysis. Considering the ionospheric cavity, we found that the IAR short-wavelength wave is trapped through the growth of a long-wavelength wave. ii) On the formation of an energy transport pass from the ionosphere to the magnetic equator. Considering both the ionospheric and magnetospheric cavities, we found that the IAR wave is trapped, and then part of the wave propagates to the magnetic equator side, and a flow shear instability occurs due to excitation of the HAR waves.
\section{Model Description}\label{sec: 2}
In order to elucidate the physics involved in auroral structures, nonlinear evolution of shear Alfv$\acute{\rm e}$n waves propagating along the dipole magnetic field $\mbox{\boldmath $B$}_0$ can be modeled by using two-field reduced MHD equations [Chmyrev et al., 1988]. The waves slightly slip ($\Omega / k_\perp \ll v_{\rm A}$) through the feedback coupling to density waves at the ionosphere, where $\Omega$: frequency, $k_\perp$: perpendicular wavenumber, and $v_{\rm A}$: Alfv$\acute{\rm e}$n velocity. As the model formulation in this paper is about the same as Hiraki [2014], we simply explain the coordinates, equations, and numerical techniques, while the different parts in detail. The system of interest is a field line with a length of $l\approx 7\times 10^4$ km and at a latitude of 70$^\circ$, where auroral arcs develop. The field line position $s$ is defined as $s=0$ at the ionosphere and $s=l$ at the magnetic equator (a radial distance of $\approx8.5$ Earth radius). We set a local flux tube: a square of ($l_\perp \times l_\perp$) with $l_\perp = 10^{-3} l \approx 70$ km at $s=0$, a rectangle of ($h_\nu l_\perp \times h_\varphi l_\perp$) at $s$, and ($\approx 3300$ km $\times$ $\approx 1700$ km) at $s=l$ using dipole metrics $h_\nu(s)$ and $h_\varphi(s)$ with $B_0(s) = 1/h_\nu h_\varphi$ [HW, 2011].
The electric field $\mbox{\boldmath $E$}$ is partitioned into a background convective part $\mbox{\boldmath $E$}_0$ ($\perp \mbox{\boldmath $B$}_0$) and the Alfv$\acute{\rm e}$nic perturbation $\mbox{\boldmath $E$}_1 = B_0 \mbox{\boldmath $\nabla $}_\perp \phi$. The magnetic perturbation is expressed as $\mbox{\boldmath $B$}_1 = \mbox{\boldmath $\nabla $}_\perp \psi \times \mbox{\boldmath $B$}_0$. The equations at $0<s \le l$ are written as
\begin{eqnarray}
\displaystyle && \partial_t \omega + \mbox{\boldmath $v$}_\perp \cdot \mbox{\boldmath $\nabla$}_\perp \omega = v_{\rm A}^2 \nabla_\parallel j_\parallel \\
&& \partial_t \psi + \mbox{\boldmath $v$}_0 \cdot \mbox{\boldmath $\nabla$}_\perp \psi + \frac{1}{B_0} \nabla_\parallel B_0 \phi = -\eta j_\parallel.
\end{eqnarray}
The convective drift velocity $\mbox{\boldmath $v$}_0=\mbox{\boldmath $E$}_0 \times \mbox{\boldmath $B$}_0/B_0^2$ is set so that $E_0$ satisfies the equi-potential condition, while $\mbox{\boldmath $v$}_\perp = \mbox{\boldmath $v$}_0 + \mbox{\boldmath $v$}_1(\mbox{\boldmath $E$}_1)$, vorticity $\omega = \nabla_\perp^2 \phi$, field-aligned current $j_\parallel = - \nabla_\perp^2 \psi$, and $\nabla_\parallel = \partial_s + \mbox{\boldmath $b$}_0 \cdot \mbox{\boldmath $\nabla $}_\perp \times \mbox{\boldmath $\nabla $}_\perp \psi$.
Ionospheric plasma motion including density waves is described by the two fluid equations. Considering the current dynamo layer (height of 100--150 km), we can assume that ions and electrons respectively yield the Pedersen drift $\mbox{\boldmath $v$}_{\rm i} = \mu_{\rm P} \mbox{\boldmath $E$}-D \mbox{\boldmath $\nabla $}_\perp \ln n_{\rm i}$ and the Hall drift $\mbox{\boldmath $v$}_{\rm e} = \mu_{\rm H} \mbox{\boldmath $E$} \times \mbox{\boldmath $B$}_0/B_0 - \mbox{\boldmath $j$}_\parallel/e n_{\rm e}$, with $\mu_{\rm P,H}$: mobilities and $D$: diffusion coefficient. By integrating the continuity equations over the dynamo layer, the equations at $s=0$ become
\begin{eqnarray}
\displaystyle && \partial_t n_{\rm e} + \mbox{\boldmath $v$}_\perp \cdot \mbox{\boldmath $\nabla$}_\perp n_{\rm e} = j_\parallel - R n_{\rm e} \\
&& \mbox{\boldmath $\nabla$}_\perp \cdot (n_{\rm e} \mu_{\rm P} \mbox{\boldmath $E$}) - \mbox{\boldmath $v$}_\perp \cdot \mbox{\boldmath $\nabla$}_\perp n_{\rm e} = D \nabla_\perp^2 n_{\rm e} - j_\parallel,
\end{eqnarray}
where $R n_{\rm e}$ is a linearized recombination term, and the Hall mobility is normalized to be unity. We assume that $j_\parallel$ is carried by thermal electrons.
\begin{figure}[t!]
\includegraphics[width=1.0\columnwidth, bb=0 0 360 252, clip]{fig1_cavi.pdf}
\caption{Alfv$\acute{\rm e}$n velocity profiles 1--8 along the field line $s$, used in our simulations; the values are normalized by the average velocity $v_{\rm A}'$ (see text).}
\end{figure}
\begin{figure}[t!]
\includegraphics[width=1.0\columnwidth, bb=0 0 360 252, clip]{fig2_cavi.pdf}
\caption{Average electron density at the ionosphere as function of time; $\tau_{\rm A}$ is the Alfv$\acute{\rm e}$n transit time. Shown are for the eight cases of $v_{\rm A}$ profiles in Fig.\ 1.}
\end{figure}
\begin{figure*}[t]
\includegraphics[scale=0.7, bb=220 80 1000 620, clip]{fig3_cavi.png}
\caption{Vorticity $\omega(x,y)$ at the magnetic equator $s=l$ at $t/\tau_{\rm A} = 0$ (a), 3 (b), and 3.5 (c), respectively, in the case 1 of $v_{\rm A}$; the values are normalized. (d) Temporal variation in the average vorticity along the field line. (e) Vorticity at the ionosphere $s=0$ at $t/\tau_{\rm A} = 3.5$.}
\end{figure*}
We used the 4th-order finite difference methods in space and time to solve Eqs.\ (1)--(4), including sharp gradients of $v_{\rm A}$, with grid numbers of (256, 256, 512) for the $\nu$, $\varphi$, and $s$ directions, respectively. The time resolution was changed in accord with the Courant condition: $\max(\mbox{\boldmath $v$}_1 / \Delta x(s)) \Delta t < 0.25$. A higher-order low-pass filter [Lele, 1992] was used along with the numerical viscosity and resistivity $\nu_{\rm v} = \eta = 1\times 10^{-7} /B_0(s)$. Regarding the calculation domain $\mbox{\boldmath $x$}_\perp(s=0) \equiv [x,y]$, $x$ and $y$ pointed southward and eastward, respectively, in the southern hemisphere. We set a periodic boundary in the $\mbox{\boldmath $x$}_\perp$ direction, e.g., at $x$, $y = 0$ and $l_\perp = 70$ km (thus $\Delta x \approx 0.27$ km) at the ionosphere $s=0$. An asymmetric boundary for the magnetic field $\psi = 0$ (or $j_\parallel = 0$) was set at the magnetic equator $s=l$. At the ionospheric boundary of $\phi$, Eq.\ (4) was solved using the multigrid-BiCGStab method. Here, values of the control parameters were referred to Hiraki [2014]; $B_0=5.7\times10^{-5}$ T, $\mu_{\rm P} / \mu_{\rm H}=0.5$, $D=4\times 10^5$ m$^2$/s, $R=2\times10^{-3}$ /s, and the ambient density $n_{\rm e0} = 3.8 \times10^4$ cm$^{-3}$ at $s=0$. In this paper, the convection electric field was set to point southward with an amplitude of $E_0 = 60$ mV/m at $s=0$, i.e., an westward drift of $v_0 \approx 1.1$ km/s.
Different from $v_{\rm A}={\rm const}$ in our previous study [Hiraki, 2014], a field line variation in the Alfv$\acute{\rm e}$n velocity $v_{\rm A}(s) \equiv B_0(s) / \sqrt{\mu_0 m_{\rm i} n_{\rm i}(s)}$ was taken into account. Here, $\mu_0$ is the permeability, $m_{\rm i}$ the proton mass, and $n_{\rm i}$ the ion density, respectively. The modeled ion density [HW, 2011]
\begin{equation}
\displaystyle n_{\rm i}(s) = n_{\rm a} {\rm e}^{-(\hat{r}(s) - 1)/h_{\rm a}} + n_{\rm b} \hat{r}(s)^{-q} + n_{\rm c}
\end{equation}
was used. Here, $\hat{r}$ is the normalized distance from the Earth center to the position $s$; $\hat{r}(0)=1$ and $\hat{r}(l)\approx 8.5$. We fixed $n_{\rm i}(0) = 7\times 10^5$ cm$^{-3}$, $h_{\rm a} = 1/6$, and $n_{\rm c} = 0$ cm$^{-3}$. By changing $n_{\rm b}=8.5\times10^1$--$7\times10^5$ cm$^{-3}$ and $q=2.4$--6.4, we obtained eight $v_{\rm A}$ profiles characterized by ionospheric and magnetospheric cavities shown in Fig.\ 1. These profiles were renormalized so that the Alfv$\acute{\rm e}$n transit time is fixed constant as $\tau_{\rm A} \equiv \int_0^l {\rm d}s/v_{\rm A}(s) = l/v_{\rm A}'\approx 47$ s with the average velocity $v_{\rm A}' = 1.5 \times 10^3$ km/s.
We solved a linearized set of Eqs.\ (1)--(4) to determine the eigenfunctions $(\tilde{\phi}(s), \tilde{\psi}(s), \tilde{n}_{\rm e}(0))$ and frequency $\Omega$ of Alfv$\acute{\rm e}$n waves as functions of the perpendicular wavenumber $\mbox{\boldmath $k$}_\perp$ and the field-line harmonic number. Although field variables of an initially placed auroral arc were treated in our previous study [Hiraki, 2014], we provide only the perturbed fields $(\tilde{\phi}, \tilde{\psi}, \tilde{n}_{\rm e})$ to shed light on the pure nonlinear coupling of Alfv$\acute{\rm e}$n eigenmodes in this study. The fundamental mode with ${\rm Re}(\Omega) \approx 16$ mHz and $\mbox{\boldmath $k$}_\perp / 2 \pi = (k_x, k_y) = (1, 2)$ was found to have the maximum growth rate $\gamma \equiv {\rm Im}(\Omega)$ for the case 1 of $v_{\rm A}$ in Fig.\ 1. We performed 8 runs using each $v_{\rm A}$ in Fig.\ 1 and yielding the perturbed fields of the (1, 2) mode along with a stable $(2, -1)$ mode with $\gamma<0$; i.e., $\phi = \epsilon \tilde{\phi}_{1,2} (1 + \delta {\rm e}^{{\rm i}k_{(2,-1)} x_\perp})$ with $\epsilon = 10^{-5}$ and $\delta = 10^{-2}$ at $t = 0$, $\psi$ and $n_{\rm e}$ as well but $\delta=0$.
\section{Results}\label{sec: 3}
\subsection{Overall trend}\label{sec: 3.1}
We performed nonlinear simulations of Alfv$\acute{\rm e}$n waves using Eqs.\ (1)--(4) for 8 cases of the velocity profiles as shown in Fig.\ 1. Figure 2 shows the temporal evolution of the root mean square density $\langle n_{\rm e} \rangle$ at the ionosphere; we illustrated the density that has a slower change than the other variables, vorticity $\omega$ and current $j_\parallel$. Note that slight differences in the initial values are originated from a dependence of eigenfunction $\tilde{n}_{\rm e1,2}$ on $v_{\rm A}(s)$. As expected from the $v_{\rm A}$ profile, we can divide the results into three groups, that is, cases 1--3, cases 4--6, and cases 7, 8.
In the cases 1 and 2 including a weak gradient of $v_{\rm A}$, the density grows with an extremely high rate ($t/\tau_{\rm A} \approx 1$--2) to a high value of $\langle n_{\rm e} \rangle = 8$--10; note that the ambient density is set as $n_{\rm e0}=10$ in this unit. It causes the linearly stable mode $(2, -1)$ to gain energies, and $\langle n_{\rm e} \rangle$ shows a saturation through a nonlinear coupling to the (1, 2) mode. The density decreases into a low level as $\langle n_{\rm e} \rangle = 2$--3 after the waves are stabilized along the field line. The cases 4--6 are characterized by a developed ionospheric cavity. Since the fundamental eigenmode has a small growth rate in these cases [HW, 2012], the density slowly increases until $t/\tau_{\rm A} \approx 8$. Though we mention later in Sec.\ \ref{sec: 3.3}, a short-wavelength wave is trapped in the cavity region, a gradual oscillation with a period of $t/\tau_{\rm A}\approx 1.6$ appears, and after that the density reaches a high level ($\langle n_{\rm e} \rangle \approx 8$) during a rapid growth of IAR modes. The cases 7 and 8, respectively, have high growth rates compared to those in the cases 2 and 1. This arises from the situation that the fundamental eigenmode, escaping from the magnetospheric cavity, can penetrate into the ionosphere without any loss of amplitude; see HW [2011] for details. The remarkable feature is a fast oscillation with a period of $t/\tau_{\rm A} \approx 0.6$ during the growth phase; waves are trapped in the ionospheric cavity in this period. After $t/\tau_{\rm A} \approx 8$, $\langle n_{\rm e} \rangle$ reaches a steady state but is a higher level ($\langle n_{\rm e} \rangle \approx 4.5$) than the cases 1--3.
\begin{figure}[b]
\includegraphics[width=1.0\columnwidth, bb=0 0 360 252, clip]{fig4_cavi.pdf}
\caption{Temporal variation in the average vorticity at the magnetic equator $s=l$, in the cases 0--2 of $v_{\rm A}$ (see text).}
\end{figure}
\subsection{Cases 1--3 of $v_{\rm A}$}\label{sec: 3.2}
Figure 3 (a)--(c) shows the temporal variation in vorticity $\omega(x,y)$ at the magnetic equator $s=l$ in the case 1; panels are for $t/\tau_{\rm A} = 0$, 3, and 3.5, respectively. Also, the average vorticity $\langle \omega \rangle(s)$ at each field line position $s$ is shown in panel (d), while $\omega(x,y)$ at the ionosphere at $t/\tau_{\rm A}=3.5$ is in panel (e). We found from Fig.\ 3 (b) and (c) that the initial eigenmode at $s=l$ strongly deforms into a thinner structure of $\omega>0$, and the vortex street forms just after that. We also found from Fig.\ 3 (d) that this secondary instability at the nonlinear stage starts around a localized area of $s=10$--11 R$_{\rm E}$; erosion, or pileup of vortices, into the lower altitude occurs shortly. On the other hand, the initial eigenmode structure keeps until $t/\tau_{\rm A}=3.5$ in the ionospheric side (e), though areas of $\omega>0$ becomes thinner. We see formation of thin hair-like structures on the both ends.
In order to elucidate the cause of the secondary instability depicted above, we have performed one more run with $v_{\rm A}={\rm const}$ (hereafter, case 0) to be compared with the cases 1 and 2. Figure 4 shows the results of $\langle \omega \rangle(s=l)$. It is clear that there is a dramatic change of growth of the vorticity (flow) perturbations at $t/\tau_{\rm A}\approx 3.3$ for the cases 0 and 1, and at $t/\tau_{\rm A}\approx 4.1$ for the case 2. The values of $\langle \omega \rangle(s=l)\approx0.024$ and $\langle v_{1y} \rangle \approx v_0 \approx 26$ km/s at these times were estimated. It means that, when the wave perturbation offsets the background convection field, a vortex street forms by the velocity shear on both sides of the maxima of $|\omega|$. In addition, we note that there is no feature of the above-like secondary instability in the case 3, which is situated between these cases 0--2 and case 4.
\begin{figure}[h]
\includegraphics[scale=0.7, bb=80 80 700 580, clip]{fig5_cavi.png}
\caption{Temporal variation in the average field-aligned current along the field line (upper panel), in the case 4 of $v_{\rm A}$. Average vorticity and current at $s=0$ and vorticity at $s=l$ as function of time (bottom panel).}
\end{figure}
\begin{figure}[b]
\includegraphics[scale=0.75, bb=100 80 700 730, clip]{fig6_cavi.png}
\caption{Field-aligned current $j_\parallel(x,y)$ at the ionosphere $s=0$ at $t/\tau_{\rm A}=2.3$ (a), 5.6 (b), and 9 (c), respectively, in the case 4 of $v_{\rm A}$; the values are normalized by $j' = 660$ $\mu$A/m$^2$.}
\end{figure}
\subsection{Cases 4--6 of $v_{\rm A}$}\label{sec: 3.3}
Next, let us see the behaviors of $\omega$ and $j_\parallel$ in the case 4 where only the ionospheric cavity is deepened. Figure 5 shows the temporal variation in the average field-aligned current $\langle j_\parallel \rangle(s)$ (upper panel), besides, in $\langle \omega \rangle$ and $\langle j_\parallel \rangle$ at $s=0$, and $\langle \omega \rangle$ at $s=l$ (lower panel). The upper panel reveals that the wave magnetic perturbation, or $\langle j_\parallel \rangle$, begins to be trapped in the ionospheric cavity region $s=0$--2 R$_{\rm E}$ at $t/\tau_{\rm A}=2.3$. We also find from the lower panel that $\langle j_\parallel \rangle(s=0)$ increases and has a weak oscillation that means incidence and reflection of waves at the ionosphere. An oscillatory behavior of $\langle \omega \rangle(s=0)$ in phase with $\langle j_\parallel \rangle(s=0)$ occurs at $t/\tau_{\rm A}=4$, and these show a faster growth than in the previous period. The vorticity $\langle \omega \rangle(s=l)$ also oscillates from $t/\tau_{\rm A}=6$ and grows to be a high level until $t/\tau_{\rm A}\approx10$ since a moderate-amplitude wave propagates from the ionosphere.
Figure 6 shows the temporal variation in field-aligned current $j_\parallel(x,y)$ at $s=0$ at $t/\tau_{\rm A}=2.3$, 5.6, and 9. We find that the initially given mode becomes wavy at $t/\tau_{\rm A}=2.3$, which indicates wave trapping in the ionospheric cavity as mentioned above. High wavenumber modes in $j_\parallel$ appear, involving a growth of $\langle \omega \rangle(s=0)$ (see Fig.\ 5) at $t/\tau_{\rm A}=5.6$. It proves that the first harmonic eigenmode is generated from the fundamental perturbation due to a slight phase difference in $\omega$, $j_\parallel$, and $n_{\rm e}$. The initial structure of $j_\parallel>0$ strongly deforms, and small patches are newly produced at each left-upper side at $t/\tau_{\rm A}=9$. The field-aligned current reaches a high value of $\langle j_\parallel \rangle \approx 20$ $\mu$A/m$^2$ at this time.
\begin{figure}[t]
\includegraphics[width=1.0\columnwidth, bb=0 0 360 252, clip]{fig7_cavi.pdf}
\caption{Temporal variation in the average field-aligned current along the field line, in the case 7 of $v_{\rm A}$.}
\end{figure}
\begin{figure*}[t]
\includegraphics[scale=0.67, bb=200 80 1000 600, clip]{fig8_cavi.png}
\caption{Vorticity $\omega(x,y)$ at the magnetic equator $s=l$ (a--c) and the ionosphere $s=0$ (d--f) at $t/\tau_{\rm A} = 2$, 3.5, and 4.5, respectively, in the case 7 of $v_{\rm A}$.}
\end{figure*}
\subsection{Cases 7 and 8 of $v_{\rm A}$}\label{sec: 3.4}
Let us finally see the behaviors of variables in the case 7 where there is a unique oscillation in the linear stage (see Fig.\ 2). Figure 7 shows the temporal variation in $\langle j_\parallel \rangle(s)$. We find that an initially placed eigenfunction quickly deforms at $t/\tau_{\rm A} = 2$, and the ionospheric Alfv$\acute{\rm e}$n resonance (IAR) occurs. The half-wavelength IAR wave oscillates, while a wavy structure forms at $s=7$--11 R$_{\rm E}$, at $t/\tau_{\rm A} = 3$--3.5. The latter means that part of IAR waves, through their ionospheric feedback coupling, is shot to the magnetic equator. The one-wavelength IAR wave is produced at $s=0$ at $t/\tau_{\rm A} = 4.5$. A rapid growth of $\langle j_\parallel \rangle(s)$ up to 70 $\mu$A/m$^2$ in the cavity region is found. The current amplifies until $t/\tau_{\rm A} = 5.3$ in the magnetospheric side, extending down to $s \approx 5$ R$_{\rm E}$; pileup of vortices occurs in the background. It is clear that the hybrid Alfv$\acute{\rm e}$n resonant (HAR) waves [HW, 2012] grow.
Figure 8 shows the temporal variation in vorticity $\omega(x,y)$ at the magnetic equator $s=l$ (a--c) and the ionosphere $s=0$ (d--f) at $t/\tau_{\rm A} = 2$, 3.5, and 4.5. The first striking change appears in the $s=0$ side, that is, splitting of initial arcs along with thin hair-like structures on the poleward side. This feature is related to the trapping of IAR waves depicted in Fig.\ 7. The effect of this splitting is transmitted (as part of IAR waves) to the $s=l$ side until $t/\tau_{\rm A} = 3.5$, and a vortex structure forms at $t/\tau_{\rm A} = 4.5$. At the same time, a distinct structure, or an oblique slice forms in the initial arcs at $s=0$. We clarified the timing that vortices appear at $s=l$. The same as the cases 0--2, vorticity and the $y$-component of perturbed velocity at $s=l$ reach $\langle \omega \rangle(s=l)\approx0.024$ and $\langle v_{1y} \rangle \approx v_0 \approx 26$ km/s, respectively, at $t / \tau_{\rm A} \approx 4.3$. Vortices in this case are slightly distorted compared to the case 1 shown in Fig.\ 3, because their formation is strongly affected by IAR waves propagating from the ionosphere. However, we emphasize that the secondary nonlinear instability at $s=l$ is triggered by a flow shear formed after the wave perturbation offsets the background convection field. It is also marvelous that the total perturbed velocity peaks at $\langle v_{1} \rangle_{\max} \approx 65$ km/s and $|v_{1}|_{\max} \approx 220$ km/s at the latter period $t/\tau_{\rm A} = 6.2$.
\section{Discussion}\label{sec: 4}
Let us summarize the effects of the Alfv$\acute{\rm e}$n velocity cavities on the wave growth obtained above. For the case 0 with $v_{\rm A}(s)={\rm const}$ and a dipole $B_0(s)$, we find that the convection drift is offset through a growth of perturbed velocity fields and a secondary instability occurs at the magnetic equator. The similar behavior was found in the simpler slab geometry of $B_0 = {\rm const}$ [Watanabe, 2010]. Much simpler system that describes shear Alfv$\acute{\rm e}$n wave dynamics is set up by removing the ionosphere; however, it provides an almost self-evident solution, or a uniform propagation of given waves. We call the case 0 the minimal model of shear Alfv$\acute{\rm e}$n waves developed in the MI coupling system. Only the growth of internal waves on the flux tube of our interest, without any external force, produces a strong flow perturbation at the magnetic equator. This is originated from the fact that the fundamental eigenmode has a wave form of the same order of amplitude throughout the field line. The maximum of the flow perturbations at $s=l$ in this case is estimated to be $|v_{1}|_{\max} \approx 230$ km/s at $t/\tau_{\rm A} = 5$ and the $x$-component is one third of that, which is similar to the case 7 shown in Sec.\ \ref{sec: 3.4}. So-called "bursty flow" that occurs in substorm onsets has been known to have a magnitude of $v_x = 100$--300 km/s [Lee et al., 2012; Sergeev et al., 2012]. We suppose that the flow perturbation obtained in this study could not be the bursty flow itself but be the first indication of it.
Next, we consider what knowledge is given by the expansion from cases 0--2 to cases 4--5. That is that trapping of IAR waves occurs after the fundamental mode structure deforms. This is originated from the fact that the first harmonic eigenmode has a wave form of which the half wavelength is trapped in the cavity region (see Fig.\ 5). It is marvelous that, when the IAR modes grow, the secondary instability (vortex formation) at the magnetic equator side is suppressed. When $\omega(s=l)$ becomes large after $t/\tau_{\rm A}\approx 10$ as shown in Fig.\ 5, structures are surely changed to be turbulent. By the next expansion to cases 7 and 8, we clarified that a high-amplitude wave is trapped in the cavity side, grows, and the field energy is transported to the magnetic equator side. It causes the flow shear instability. It means that these cases combine the essences of the above two groups (0--2 and 4--5). We emphasize that the first harmonic wave is excited in the cavity region, and then the second- and the third harmonic waves are excited to form the hybrid modes, resulting in the strong flow perturbation at the magnetic equator. It brings a directivity "from one side (cavity) to the other side (magnetic equator)" into our system. The existence of the directivity is supported by observations of AKR [Morioka et al., 2010]. They claimed that an enhancement of the field-aligned current occurs at the acceleration region (nearly equals to the height of $v_{\rm A}$ peaks) at substorm onsets, and then it causes the enhancement of flow fields at the magnetic equator.
When the background cavity becomes deep, a nonlinear coupling between the stable and unstable fundamental modes is promoted due to a perpendicular phase shift of $n_{\rm e}$ and $\phi$. The conversion from the initially placed fundamental mode (cases 1--3) to the IAR and HAR waves ($\parallel \mbox{\boldmath $B$}_0$) occurs simultaneously. This property is seen in the values of $\langle n_{\rm e} \rangle$, especially those in the quasi-steady states after saturation. The values in the cases 4--6 (IAR) are largest and those in the cases 7 and 8 (HAR) are the next one at $t/\tau_{\rm A} \ge 8$. On the other hand, the fundamental mode in the cases 1--3 is considered not to bring any strong energy in the ionosphere once it is stabilized. If we relate our result "the vortices at the magnetic equator associated with IAR wave propagation" to the evolution of auroral structures, we need to investigate the relationship between the vortices and the field-aligned electric field $E_\parallel$. By doing a test particle simulation under the parallel field as done by Chaston et al.\ [2006], our next problem is to clarify which physical quantity directly concerns the auroral luminosity. Although there has been many theoretical studies [Chaston et al., 2008; Watt et al., 2009; Lysak and Song, 2008] on the electron acceleration by Alfv$\acute{\rm e}$n waves, a 3D nonlinear calculation that treats the interaction of shear Alfv$\acute{\rm e}$n and inertial Alfv$\acute{\rm e}$n waves along a full field line has not been implemented. For the realistic setup, we should adopt a deeper cavity by one order than that in the case 8 of this study: $v_{\rm A, \min}\approx 10^6$ m/s to $v_{\rm A,\max} \approx 10^8$ m/s. The IAR and HAR modes could be generated to play a major role in auroral structuring and energy transport in the MI coupling system.
\section{Conclusion}
We performed MHD simulations of shear Alfv$\acute{\rm e}$n waves in a full field line system with MI coupling and Alfv$\acute{\rm e}$n velocity cavities.
Feedback instability of the Alfv$\acute{\rm e}$n resonant modes showed various nonlinear features: i) a secondary flow shear instability occurs at the magnetic equator, ii) trapping of the ionospheric Alfv$\acute{\rm e}$n resonant modes facilitates deformation of auroral fine structures, and iii) waves emitted from the ionospheric cavity cause vortices and magnetic oscillations around the magnetic equator side.
Essential features in the initial brightening of auroral arc at substorm onsets could be explained by growth of the Alfv$\acute{\rm e}$n resonant modes, which are the nature of the field line system responding to a rapid change in the background conditions.
The IAR and HAR modes could play a major role in auroral structuring and energy transport in the MI coupling system.
\newpage
|
\section{Introduction} \label{s:infinitam_introduction}
Volumetric models have become a popular representation for 3D scenes in recent
years. While being used as early as~\cite{Curless96:VMB}, one of the
breakthroughs leading to their popularity was
KinectFusion~\cite{newcombe_ismar_2011,izadi_uist_2011}. The focus of
KinectFusion and a number of related works~\cite{ReconstructME_2012,kinfu_2011,Reitmayr13:KFusion}
is on 3D reconstruction using RGB-D sensors. However, monocular SLAM has since also
been tackled with very similar approaches~\cite{Newcombe11:DTA,Pradeep13:MFR}.
In the monocular case, various dense stereo techniques are used to create depth
images and then the same integration methods and volumetric representations are
used to fuse the information into a consistent 3D world model. Storing the
underlying truncated signed distance function (TSDF) volumetrically makes for
most of the simplicity and efficiency that can be achieved with GPU
implementations of these systems. However, this representation is also
memory-intensive and limits the applicability to small scale reconstructions.
Several avenues have been explored for overcoming this limitation. A first line
of works uses a moving volume~\cite{Roth12:MVK,Whelan13:RRT,kinfuls_2012} to
follow the camera while a sparse point cloud or mesh representation is computed
for the parts of the scene outside the active volume. In a second line of
research, the volumetric representation is split into a set of blocks aligned
with dominant planes~\cite{Henry13:PVS} or even reduced to a set of bump maps
along such planes~\cite{Thomas13:FSR}. A third category of approaches employs
an octree representation of the 3D volume~\cite{Zeng13:OFR,Chen13:SRV,
Steinbruecker14:V3D}. Finally, a hash lookup for a set of sparsely allocated
subblocks of the volume is used in~\cite{neissner_tog_2013}. Some of these
works, e.g.~\cite{kinfuls_2012,Chen13:SRV,neissner_tog_2013,Henry13:PVS}, also
provide methods for swapping data from the limited GPU memory to some larger
host memory to further expand the scale of the reconstructions.
With the aim of providing for a fast and flexible 3D reconstruction pipeline,
we propose a new, unifying framework called InfiniTAM. The core idea is that
individual steps like camera tracking, scene representation and integration of
new data can easily be replaced and adapted to the needs of the user.
Along with the framework we also provide a set of components for scalable
reconstruction: two implementations of camera trackers, based on RGB data and
on depth data, two representations of the 3D volumetric data, a dense volume
and one based on hashes of subblocks~\cite{neissner_tog_2013}, and an optional
module for swapping subblocks in and out of the typically limited GPU memory.
Given these components, a wide range of systems can be developed for specific
purposes, ranging from very efficient reconstruction of small scale volumes
with limited hardware resources up to full scale reconstruction of large-scale
scenes. While such systems can be tailored for the development of higher level
applications, the framework also allows users to focus on the development of
individual new components, reusing only parts of the framework.
Although most of the ideas used in our implementation have already been
presented in related works, there are a number of differences
and novelties that went into the engineering of our framework. For example in
Section~\ref{s:swapping} we define an engine for swapping data to and from the
GPU in a way that can deal with slow read and write accesses on the host,
e.g.\ on a disk. It also has a fixed maximum number of data transfers between
GPU and host to ensure an interactive online framework.
One general aim of the whole framework was to keep the implementation portable,
adaptable and simple. As we show in Section~\ref{s:results}, InfiniTAM has
minimal dependencies and natively builds on Linux, Mac OS and Windows platforms.
The remainder of this report and in particular Sections~\ref{s:itm_overview}
through~\ref{s:swapping} describe technical implementation details of the
InfiniTAM framework. These sections are aimed at closing the gap between the
theoretical description of volumetric depth map fusion and the actual software
implementation in our InfiniTAM package. Some advice on
compilation and practical usage of InfiniTAM is given in Section~\ref{s:results}
and some concluding remarks follow in Section~\ref{s:conclusions}.
\section{Architecture Overview}\label{s:itm_overview}
We first present an overview of the overall processing pipeline of our
framework and discuss the cross device implementation architecture. These
serve as the backbone of the framework throughout this report.
\subsection{Processing Pipeline}\label{s:pipeline}
\begin{figure}[tb]
\centering
\includegraphics[width=\linewidth]{images/itm_pipeline.pdf}
\caption{InfiniTAM processing pipeline}
\label{fig:itm_pipeline}
\end{figure}
The main processing steps are illustrated in Figure~\ref{fig:itm_pipeline}.
As with the typical and well known KinectFusion
pipeline~\cite{newcombe_ismar_2011}, there is a \textit{tracking} step for
localizing the camera, an \textit{integration} step for updating the 3D world
model and a \textit{raycasting} step for rendering the 3D data in preparation
of the next tracking step. To accommodate the additional processing requirements
for octrees~\cite{Steinbruecker14:V3D}, voxel block
hashing~\cite{neissner_tog_2013} or other data structures, an
\textit{allocation} step is included, where the respective data structures are
updated in preparation for integrating data from a new frame. Also an optional
\textit{swapping} step is included for transferring data between GPU and CPU.
A detailed discussion of each individual stage follows in
Section~\ref{s:itm_methodstages}. For now notice that our implementation
follows the chain-of-responsibility design pattern, which means that the
relevant data structures (e.g.\ \reftocode{ITMImage})
are passed between several processing engines
(e.g.\ \reftocode{ITMSceneReconstructionEngine}). The engines
are stateless and each engine is responsible for one specific aspect of
the overall processing pipeline. The state is passed on in objects containing
the processed information. Finally one central class (\reftocode{ITMMainEngine})
holds instances of all objects and engines and controls the flow of information.
\subsection{Cross Device Implementation Architecture} \label{s:itm_codearch}
\begin{figure}[tb]
\centering
\includegraphics[width=0.75\linewidth]{images/itm_enginestruct.pdf}
\caption{InfiniTAM Cross Device Engine Architecture}
\label{fig:itm_enginestruct}
\end{figure}
Each engine is further split into three layers as shown in
Figure~\ref{fig:itm_enginestruct}. The topmost, \textit{Abstract
Layer} is accessed by the library's main engine and is in general just an
blank interface, although some common code can be shared at this point. The
interface is implemented in the \textit{Device Specific Layer},
which will be very different depending on whether it runs on a CPU, a GPU,
on OpenMP or other hardware acceleration architectures. In the \textit{Device
Agnostic Layer} there may be some inline code that is called from the higher
layers and recompiled for the different architectures.
Considering the example of a tracking engine,
the \textit{Abstract Layer} contains
code for the generic optimisation of an error function, the \textit{Device
Specific Layer} contains a loop or GPU kernel call to evaluate the error
function for all pixels in an image, and the \textit{Device Agnostic Layer}
contains a simple inline C-function to evaluate the error in a single pixel.
\section{Volumetric Representation}\label{s:itm_data_structures}
The volumetric representation is the central data structure within the
InfiniTAM framework. It is used to store a truncated signed distance function
that implicitly represents the 3D geometry by indicating for each point in
space how far in front or behind the scene surface it is.
Being truncated, only a small band of values around the current surface
estimate is actually being stored, while beyond the band the values are
set to a maximum and minimum value, respectively.
The data structure used for this representation crucially defines the memory
efficiency of the 3D scene model and the computational efficiency of
interacting with this model. We therefore keep it modular and interchangeable
in our framework and provide two implementations, namely for a fixed size dense
volume (\reftocode{ITMPlainVoxelArray}) and voxel block hashing
(\reftocode{ITMVoxelBlockHash}). Both are described in the following, but first
we briefly discuss how the representation is kept interchangeable throughout
the framework.
\subsection{Framework Flexibility}
Core data classes dealing with the volumetric signed distance function are
templated on (i) the type of voxel information stored, and (ii) the data
structure used for indexing and accessing the voxels. The relevant engine
classes are equally templated to provide the specific implementations
depending on the choosen data structures.
An example for a class representing the actual information at each voxel
is given in the following listing:
\begin{lstlisting}
struct ITMVoxel_s_rgb
{
_CPU_AND_GPU_CODE_ static short SDF_initialValue() { return 32767; }
_CPU_AND_GPU_CODE_ static float SDF_valueToFloat(float x)
{ return (float)(x) / 32767.0f; }
_CPU_AND_GPU_CODE_ static short SDF_floatToValue(float x)
{ return (short)((x) * 32767.0f); }
static const bool hasColorInformation = true;
/** Value of the truncated signed distance transformation. */
short sdf;
/** Number of fused observations that make up @p sdf. */
uchar w_depth;
/** RGB colour information stored for this voxel. */
Vector3u clr;
/** Number of observations that made up @p clr. */
uchar w_color;
_CPU_AND_GPU_CODE_ ITMVoxel_s_rgb()
{
sdf = SDF_initialValue();
w_depth = 0;
clr = (uchar)0;
w_color = 0;
}
};
\end{lstlisting}
where the macro \reftocode{\_CPU\_AND\_GPU\_CODE\_} identifies methods and
functions that can be run both as host and as device code and is defined as:
\begin{lstlisting}
#if defined(__CUDACC__) && defined(__CUDA_ARCH__)
#define _CPU_AND_GPU_CODE_ __device__ // for CUDA device code
#else
#define _CPU_AND_GPU_CODE_
#endif
\end{lstlisting}
Alternative voxel types provided along with the current implementation are
based on floating point values instead of short, or they do not contain colour
information. Note that the member \reftocode{hasColorInformation} is used in
the provided integration methods to decide whether or not to gather colour
information, which has an impact on the processing speed accordingly.
Two examples for the index data structures that allow accessing the voxels are
given below. Again, note that the processing engines choose different
implementations according to the selected indexing class. For example the
allocation step for voxel block hashing has to modify a hash table, whereas
for a dense voxel array it can return immediately and does not have to do
anything.
\subsection{Dense Volumes}
We first discuss the naive way of using a dense volume of limited size, as
presented in the earlier works of~\cite{newcombe_ismar_2011,izadi_uist_2011,
Newcombe11:DTA,Pradeep13:MFR}. This is very well suited for understanding the
basic workings of the algorithms, it is trivial to parallelise on the GPU, and
it is sufficient for small scale reconstruction tasks.
In the InfiniTAM framework the class \reftocode{ITMPlainVoxelArray} is used to
represent dense volumes and a simplified listing of this class is given below:
\begin{lstlisting}
class ITMPlainVoxelArray
{
public:
struct ITMVoxelArrayInfo {
/// Size in voxels
Vector3i size;
/// offset of the lower left front corner of the volume in voxels
Vector3i offset;
ITMVoxelArrayInfo(void)
{
size.x = size.y = size.z = 512;
offset.x = -256;
offset.y = -256;
offset.z = 0;
}
};
typedef ITMVoxelArrayInfo IndexData;
private:
IndexData indexData_host;
public:
ITMPlainVoxelArray(bool allocateGPU)
...
~ITMPlainVoxelArray(void)
...
/** Maximum number of total entries. */
int getNumVoxelBlocks(void) { return 1; }
int getVoxelBlockSize(void) { return indexData_host.size.x * indexData_host.size.y * indexData_host.size.z; }
const Vector3i getVolumeSize(void) { return indexData_host.size; }
const IndexData* getIndexData(void) const
...
};
\end{lstlisting}
Note that the subtype \reftocode{IndexData} as well as the methods
\reftocode{getIndexData()}, \reftocode{getNumVoxelBlocks()} and
\reftocode{getVoxelBlockSize()} are used extensively within the processing
engines and on top of that the methods \reftocode{getNumVoxelBlocks()} and
\reftocode{getVoxelBlockSize()} are used to determine the overall number of
voxels that have to be allocated.
Depending on the choice of voxel type, each voxel requires at least 3 bytes
and, depending on the available memory a total volume of about
$768\times{}768\times{}768$
voxels is typically the upper limit for this dense representation. To make the
most of this limited size, the initial camera is typically placed towards the
centre of the rear plane of this voxel cube, as identified by the
\reftocode{offset} in our implementation, so that there is maximum overlap
between the view frustum and the volume.
\subsection{Voxel Block Hashing}
The key idea for scaling the SDF based representation to larger 3D environments is
to drop the empty voxels from outside the truncation band and to represent only
the relevant parts of the volume. In~\cite{neissner_tog_2013} this is achieved
using a hash lookup of subblocks of the volume. Our framework provides an
implementation of this method that we explain in the following.
Voxels are grouped in blocks of predefined size (currently $8\times8\times8$
voxels). All the voxel blocks are stored in a contiguous array, referred
henceforth as the \textit{voxel block array} or VBA. In the current
implementation this has a defined size of $2^{18}$ elements.
To handle the voxel blocks we further need:
\begin{itemize}
\item Hash Table and Hashing Function: Enable fast access to voxel blocks in the voxel block array -- details in Subsection \ref{ss:itm_hashtable}.
\item Hash Table Operations: Insertion, retrieval and deletion of voxel blocks -- details in Subsection \ref{ss:itm_hashopers}.
\end{itemize}
\subsubsection{Hash Table and Hashing Function} \label{ss:itm_hashtable}
To quickly and efficiently find the position of a certain voxel block in the
voxel block array, we use a hash table. This hash table is a contiguous array
of \reftocode{ITMHashEntry} objects of the following form:
\begin{lstlisting}
struct ITMHashEntry
{
/** Position of the corner of the 8x8x8 volume, that identifies the entry. */
Vector3s pos;
/** Offset in the excess list. */
int offset;
/** Pointer to the voxel block array.
- >= 0 identifies an actual allocated entry in the voxel block array
- -1 identifies an entry that has been removed (swapped out)
- <-1 identifies an unallocated block
*/
int ptr;
};
\end{lstlisting}
The hash function \reftocode{hashIndex} for locating entries in the hash table
takes the corner coordinates \reftocode{vVoxPos} of a 3D voxel block and
computes an index as follows~\cite{neissner_tog_2013}:
\begin{lstlisting}
template<typename T> _CPU_AND_GPU_CODE_ inline int hashIndex(const InfiniTAM::Vector3<T> vVoxPos, const int hashMask){
return ((uint)(((uint)vVoxPos.x * 73856093) ^ ((uint)vVoxPos.y * 19349669) ^ ((uint)vVoxPos.z * 83492791)) & (uint)hashMask);
}
\end{lstlisting}
To deal with \textit{hash collisions}, each hash index points to a
\textit{bucket} of fixed size (typically 2), which we consider the
\textit{ordered} part of the hash table. There is an additional
\textit{unordered} excess list that is used once an ordered bucket fills up.
In either case, each \reftocode{ITMHashEntry} in the hash table stores an offset
in the voxel block array and can hence be used to localise the voxel data for
this specific voxel block. This overall structure is illustrated in
Figure~\ref{fig:itm_datastruct}.
\begin{figure}[tb]
\centering
\includegraphics[width=0.85\linewidth]{images/itm_datastruct.pdf}
\caption{Hash table logical structure}
\label{fig:itm_datastruct}
\end{figure}
\subsubsection{Hash Table Operations} \label{ss:itm_hashopers}
The three main operations used when working with a hash table are the
\textbf{insertion}, \textbf{retrieval} and \textbf{removal} of entries. In the
current version of InfiniTAM we support the former two, with \textbf{removal}
not currently required or implemented. The code used by the \textbf{retrieval}
operation is shown below:
\begin{lstlisting}
template<class TVoxel>
_CPU_AND_GPU_CODE_ inline TVoxel readVoxel(const TVoxel *voxelData, const ITMVoxelBlockHash::IndexData *voxelIndex, const Vector3i & point, bool &isFound)
{
const ITMHashEntry *hashTable = voxelIndex->entries_all;
const TVoxel *localVBA = voxelData;
TVoxel result; Vector3i blockPos; int offsetExcess;
int linearIdx = vVoxPosParse(point, blockPos);
int hashIdx = hashIndex(blockPos, SDF_HASH_MASK) * SDF_ENTRY_NUM_PER_BUCKET;
isFound = false;
//check ordered list
for (int inBucketIdx = 0; inBucketIdx < SDF_ENTRY_NUM_PER_BUCKET; inBucketIdx++)
{
const ITMHashEntry &hashEntry = hashTable[hashIdx + inBucketIdx];
offsetExcess = hashEntry.offset - 1;
if (hashEntry.pos == blockPos && hashEntry.ptr >= 0)
{
result = localVBA[(hashEntry.ptr * SDF_BLOCK_SIZE3) + linearIdx];
isFound = true;
return result;
}
}
//check excess list
while (offsetExcess >= 0)
{
const ITMHashEntry &hashEntry = hashTable[SDF_BUCKET_NUM * SDF_ENTRY_NUM_PER_BUCKET + offsetExcess];
if (hashEntry.pos == blockPos && hashEntry.ptr >= 0)
{
result = localVBA[(hashEntry.ptr * SDF_BLOCK_SIZE3) + linearIdx];
isFound = true;
return result;
}
offsetExcess = hashEntry.offset - 1;
}
return result;
}
\end{lstlisting}
Both \textbf{insertion} and \textbf{retrieval} work by iterating through the
elements of the list stored within the hash table. Given a target 3D voxel
location in world coordinates, we first compute its corresponding voxel block
location, by dividing the voxel location by the size of the voxel blocks. Next,
we call the hashing function \reftocode{hashIndex} to compute the index of the
bucket from the ordered part of the hash table. All elements in the bucket are
then checked, with \textbf{retrieval} looking for the target block location and
\textbf{insertion} for an unallocated hash entry. If this is found,
\textbf{retrieval} returns the voxel stored at the target location within the
block addressed by the hash entry. \textbf{Insertion} (i) reserves a block
inside the voxel block array and (ii) populates the hash table with a new entry
containing the reserved voxel block array address and target block 3D world
coordinate location.
If all locations in the bucket are exhausted, the enumeration of the list moves
to the linked list in the unordered part of the hash table, using the
\reftocode{offset} field to provide the location of the next hash entry. The
enumeration finishes when \reftocode{offset} is found to be smaller or equal to
$-1$. At this point, if the target location still has not been found,
\textbf{retrieval} returns an empty voxel. \textbf{Insertion} (i) reserves an
unallocated entry in the unordered part of the hash table and a block inside
the voxel block array, (ii) populates the hash table with a new entry
containing the reserved voxel block array address and target block 3D world
coordinate location and (iii) changes the \reftocode{offset} field in the
previous entry in the linked list to point to the newly populated one.
The reserve operations used for the unordered part of the hash table and for
the voxel block array use prepopulated allocation lists and, in the GPU code,
atomic operations.
All hash table operations are done through these functions and there is no
direct memory access encouraged or indeed permitted by the current version of
the code.
\section{Individual Method Stages}\label{s:itm_methodstages}
\begin{figure}[tb]
\centering
\includegraphics[width=.9\linewidth]{images/itm_classdiagram.pdf}
\caption{Simplified diagram of InfiniTAM namespaces, classes and collaborations.}
\label{fig:itm_classdiagram}
\end{figure}
A general outline of the InfiniTAM processing pipeline has already been given
in Section~\ref{s:pipeline} and Figure~\ref{fig:itm_pipeline}. Details of each
processing stage will be discussed in the following. The distinct stages are
implemented using individual engines and a simplified diagram of the
corresponding classes and their collaborations is given in
Figure~\ref{fig:itm_classdiagram}. The stages are:
\begin{itemize}
\item \textbf{Tracking}: The camera pose for the new frame is obtained by fitting the current depth (or optionally colour) image to the projection of the world model from the previous frame. This is implemented using the \reftocode{ITMTracker}, \reftocode{ITMColorTracker} and \reftocode{ITMDepthTracker} engines.
\item \textbf{Allocation}: Based on the depth image, new voxel blocks are allocated as required and a list of all visible voxel blocks is built. This is implemented inside the \reftocode{ITMSceneReconstructionEngine}.
\item \textbf{Integration}: The current depth and colour frames are integrated within the map. This is implemented inside the \reftocode{ITMSceneReconstructionEngine}.
\item \textbf{Swapping In and Out}: If required, map data is swapped in from host memory to device memory and merged with the present data. Parts of the map that are not required are swapped out from device memory to host memory. This is implemented using the \reftocode{ITMSwappingEngine} and \reftocode{ITMGlobalCache}.
\item \textbf{Raycasting}: The world model is rendered from the current pose (i) for visualisation purposes and (ii) to be used by the tracking stage at the next frame. This uses the \reftocode{ITMVisualisationEngine}.
\end{itemize}
As illustrated in Figure~\ref{fig:itm_classdiagram}, the main processing
engines are contained within the \reftocode{ITMLib} namespace.
Apart from these, the UI and image acquisition engines (\reftocode{UIEngine},
\reftocode{ImageSourceEngine} and \reftocode{OpenNIEngine}) are contained in
the \reftocode{InfiniTAM} namespace.
The \reftocode{ITMLib} namespace also contains the additional engine class
\reftocode{ITMLowLevelEngine} that we do not discuss in detail. It is used
for low level image processing such as computation of image gradients and a
resolution hierarchy.
In this section we discus the tracking, allocation, integration and raycasting
stages in greater detail. We delay a discussion of the swapping until
Section~\ref{s:swapping}.
\subsection{Tracking}
In the tracking stage the camera pose consisting of the rotation matrix
$\mathbf{R}$ and translation vector $\mathbf{t}$ has to be
determined given the RGB-D image and the current 3D world model. Along with
InfiniTAM we provide the
three engines \reftocode{ITMDepthTracker} and \reftocode{ITMRenTracker} that performs tracking based on the
new depth image, and \reftocode{ITMColorTracker} that is using the colour
image. All three of these implement the abstract \reftocode{ITMTracker} class and
have implementations running on the CPU and on CUDA.
In the \reftocode{ITMDepthTracker} we follow the original alignment process as
described in~\cite{newcombe_ismar_2011,izadi_uist_2011}:
\begin{itemize}
\item Render a map $\mathcal{V}$ of surface points and a map $\mathcal{N}$ of
surface normals from the viewpoint of an initial guess for $\mathbf{R}$
and $\mathbf{t}$ -- details in Section~\ref{s:raycasting}
\item Project all points $\mathbf{p}$ from the depth image onto points
$\bar{\mathbf{p}}$ in $\mathcal{V}$ and $\mathcal{N}$ and compute their
distances from the planar approximation of the surface,
i.e.\ $d = \left(\mathbf{R} \mathbf{p} + \mathbf{t} - \mathcal{V}(\bar{\mathbf{p}})\right)^T \mathcal{N}(\bar{\mathbf{p}})$
\item Find $\mathbf{R}$ and $\mathbf{t}$ minimising the linearised sum of the
squared distances by solving a linear equation system
\item Iterate the previous two steps until convergence
\end{itemize}
A resolution hierarchy of the depth image is used in our implementation to
improve the convergence behaviour.
\reftocode{ITMRenTracker} is used as the local refinement after \reftocode{ITMDepthTracker}, and we implemented a variation of the tracking algorithm described in \cite{Ren12:ECCV}:
\begin{itemize}
\item Given an initial guess for $\mathbf{R}$ and $\mathbf{t}$, the method back projects all points $\mathbf{p}$ from the depth image into points $\bar{\mathbf{p}}$ in the map $\mathcal{V}$ and computes the robust cost $\mathcal{E} = -4\exp\{\sigma\mathcal{V}(\bar{\mathbf{p}})\}/(\exp\{\sigma\mathcal{V}(\bar{\mathbf{p}})\} + 1)^2$, where $\sigma$ is a fixed parameter that controls the basin of attraction
\item Find $\mathbf{R}$ and $\mathbf{t}$ minimising the sum of the cost using Gauss Newton optimization algorithm
\end{itemize}
When \reftocode{ITMRenTracker} is chosen, the system first runs \reftocode{ITMDepthTracker} on coarse hierarchies of the depth image, then \reftocode{ITMRenTracker} runs on the high-resolution depth image for the final pose refinement.
Alternatively the colour image can be used within an \reftocode{ITMColorTracker}.
In this case the alignment process is as follows:
\begin{itemize}
\item Create a list $\mathcal{V}$ of surface points and a corresponding list
$\mathcal{C}$ of colours from the viewpoint of an initial guess --
details in Section~\ref{s:raycasting}
\item Project all points from $\mathcal{V}$ into the current colour image $I$
and compute the Euclidean norm of the difference in colours,
i.e.\ $d = \left\|I(\pi(\mathbf{R} \mathcal{V}(i) + \mathbf{t})) - \mathcal{C}(i)\right\|_2$
\item Find $\mathbf{R}$ and $\mathbf{t}$ minimising the sum of the squared
differences using the Levenberg-Marquardt optimisation algorithm
\end{itemize}
Again a resolution hierarchy in the colour image is used and the list of surface
points is subsampled by a factor of 4. A flag in \reftocode{ITMLibSettings}
allows to select which tracker is used and the default is the
\reftocode{ITMDepthTracker}.
The three main classes \reftocode{ITMDepthTracker}, \reftocode{ITMRenTracker} and \reftocode{ITMColorTracker}
actually only implement a shared optimisation framework, including e.g.\ the
Levenberg-Marquardt algorithm, Gauss Newton algorithm and solving the linear equation systems. These
are always running on the CPU. Meanwhile the evaluation of the error function
value, gradient and Hessian is implemented in derived, CPU and CUDA specific
classes and makes use of parallelisation.
\subsection{Allocation}
In the allocation stage the underlying data structure for the representation
of the truncated signed distance function is prepared and updated, so that a new
depth image can be integrated. In the simple case of a dense voxel grid, the
allocation stage does nothing. In contrast, for voxel block hashing the goal
is to find all the voxel blocks affected by the current depth image and to
make sure that they are allocated in the hash table~\cite{neissner_tog_2013}.
In our implementation of voxel block hashing, the aim was to minimise the use
of blocking operations (e.g.\ atomics) and to completely avoid the use of
critical sections. This has led us to doing the processing three separate
stages, as we explain in the following.
In the first stage we project each pixel from the depth image to 3D space
and create a line segment along the ray from depth $d-\mu$ to $d+\mu$, where
$d$ is the measured depth at the pixel and $\mu$ is the width of the truncation
band of the signed distance function. This line segment intersects a number of
voxel blocks, and we search for these voxel blocks in the hash table. If one of
the blocks is not allocated yet, we find a free hash bucket space for it. As a
result for the next stage we create two arrays, each of the same size as the
number of elements in the hash table. The first array contains a bool
indicating the visibility of the voxel block referenced by the hash table entry,
the second contains information about new allocations that have to be performed.
Note that this requires only
simple, non-atomic writes and if more than one new block has to be allocated
with the same hash index, only the most recently written allocation will
actually be performed. We tolerate such artefacts from intra-frame hash
collisions, as they will be corrected in the next frame automatically for
small intra-frame camera motions.
In the second stage we allocate voxel blocks for each non-zero entry in the
allocation array that we built previously. This is done using a single atomic
subtraction on a stack of free voxel block indices i.e.\ we decrease the number
of remaining blocks by one and add the previous head of the stack to the hash
entry.
In the third stage we build a list of live voxel blocks, i.e.\ a list of the
blocks that project inside the visible view frustum. This is later going to be
used by the integration and swapping stages.
\subsection{Integration}
In the integration stage, the information from the most recent RGB-D image is
incorporated into the 3D world model. In case of a dense voxel grid this is
identical to the integration in the original KinectFusion
algorithm~\cite{newcombe_ismar_2011,izadi_uist_2011} and for voxel block
hashing the changes are minimal after the list of visible voxel blocks has been
created in the allocation step.
For each voxel in any of the visible voxel blocks, or for each voxel in the
whole volume for dense grids, the function
\reftocode{computeUpdatedVoxelDepthInfo} is called. If a voxel is
\textit{behind} the surface observed in the new depth image by more than the
truncation band of the signed distance function, the image does not contain any
new information about this voxel, and the function returns without writing
any changes. If the voxel is close to or in front of the observed surface, a
corresponding observation is added to the accumulated sum. This is illustrated
in the listing of the function \reftocode{computeUpdatedVoxelDepthInfo} below,
and there is a similar function in the code that additionally updates the colour
information of the voxels.
\begin{lstlisting}
template<class TVoxel>
_CPU_AND_GPU_CODE_ inline float computeUpdatedVoxelDepthInfo(TVoxel &voxel, Vector4f pt_model, Matrix4f M_d, Vector4f projParams_d,
float mu, int maxW, float *depth, Vector2i imgSize)
{
Vector4f pt_camera; Vector2f pt_image;
float depth_measure, eta, oldF, newF;
int oldW, newW;
// project point into image
pt_camera = M_d * pt_model;
if (pt_camera.z <= 0) return -1;
pt_image.x = projParams_d.x * pt_camera.x / pt_camera.z + projParams_d.z;
pt_image.y = projParams_d.y * pt_camera.y / pt_camera.z + projParams_d.w;
if ((pt_image.x < 1) || (pt_image.x > imgSize.x - 2) || (pt_image.y < 1) || (pt_image.y > imgSize.y - 2)) return - 1;
// get measured depth from image
depth_measure = depth[(int)(pt_image.x + 0.5f) + (int)(pt_image.y + 0.5f) * imgSize.x];
if (depth_measure <= 0.0) return -1;
// check whether voxel needs updating
eta = depth_measure - pt_camera.z;
if (eta < -mu) return eta;
// compute updated SDF value and reliability
oldF = TVoxel::SDF_valueToFloat(voxel.sdf); oldW = voxel.w_depth;
newF = MIN(1.0f, eta / mu);
newW = 1;
newF = oldW * oldF + newW * newF;
newW = oldW + newW;
newF /= newW;
newW = MIN(newW, maxW);
// write back
voxel.sdf = TVoxel::SDF_floatToValue(newF);
voxel.w_depth = newW;
return eta;
}
\end{lstlisting}
The main difference between the dense voxel grid and the voxel block hashing
representations is that the aforementioned update function is called for
a different number of voxels and from within different loop constructs.
\subsection{Raycast}\label{s:raycasting}
As the last step in the pipeline, an image is computed from the updated 3D
world model to provide input for the tracking at the next frame. This image can
also be used for visualisation. The main process underlying this rendering is
raycasting, i.e.\ for each pixel in the image a ray is being cast from the
camera up until an intersection with the surface is found. This essentially
means checking the value of the truncated signed distance function at each
voxel along the ray until a zero-crossing is found, and the same raycasting
engine can be used for a range of different representations, as long as an
appropriate \reftocode{readVoxel()} function is called for reading values from
the SDF.
As noted in the original KinectFusion paper~\cite{newcombe_ismar_2011}, the
performance of the raycasting can be improved significantly by taking larger
steps along the ray. The value of the truncated signed distance function can
serve as a conservative estimate for the distance to the nearest surface, hence
this value can be used as step size. To additionally handle empty space in the
volumetric representation, where no corresponding voxel block has been
allocated, we introduce a state machine with the following states:
\begin{lstlisting}
enum {
SEARCH_BLOCK_COARSE,
SEARCH_BLOCK_FINE,
SEARCH_SURFACE,
BEHIND_SURFACE,
WRONG_SIDE
} state;
\end{lstlisting}
Starting from \reftocode{SEARCH\_BLOCK\_COARSE}, we take steps of the size of
each block, i.e.\ $8$ voxels, until an actually allocated block is encountered.
Once the ray enters an allocated block, we take a step back and enter state
\reftocode{SEARCH\_BLOCK\_FINE}, indicating that the step length is now limited
by the truncation band $\mu$ of the signed distance function. Once we enter a
valid block and the values in that block indicate we are still in front of the
surface, the state is changed to \reftocode{SEARCH\_SURFACE} until a negative
value is read from the signed distance function, which indicates we are now
in state \reftocode{BEHIND\_SURFACE}. This terminates the raycasting iteration
and the exact location of the surface is now found using two trilinear
interpolation steps. The state \reftocode{WRONG\_SIDE} is entered if we are
searching for a valid block in state \reftocode{SEARCH\_BLOCK\_FINE} and
encounter negative SDF values, indicating we are behind the surface as soon as
we enter a block. In this case the ray hits the surface from behind for
whichever reason, and we do not want to count the boundary between the
unallocated, empty space and the block with the negative values as an object
surface.
Another measure for improving the performance of the raycasting is to select a
plausible search range. If a sparse voxel block representation is used, then
we are given a list of visible blocks from the allocation step, and we can
render these blocks by forward projection to give us an idea of the maximum and
minimum depth values to expect at each pixel. Within InfiniTAM this can be done
using the method \reftocode{CreateExpectedDepths()} of an
\reftocode{ITMVisualisationEngine}. A naive implementation on the
CPU computes the 2D bounding box of the projection of each voxel block into the
image and fills this area with the maximum and minimum depth values of the
corresponding 3D bounding box of the voxel block, correctly handling
overlapping bounding boxes, of course.
To parallelise this process on the GPU we split it into two steps. First we
project each block down into the image, compute the bounding box, and create a
list of $16\times{}16$ pixel fragments, that are to be filled with specific
minimum and maximum depth values. Apart from a prefix sum to count the number
of fragments, this is trivially parallelisable. Second we go through the list
of fragments and actually render them. Updating the minimum and maximum depth
for a pixel requires atomic operations, but by splitting the process into
fragments we reduce the number of collisions to typically a few hundreds or
thousands of pixels in a $640\times{}480$ image and achieve an efficiently
parallelised overall process.
\section{Swapping}\label{s:swapping}
Voxel hashing already enables much larger maps to be created, compared to the much simpler dense 3D volumes. Video card memory capacity however is often quite limited. Practically an off-the-shelf video card can roughly hold the map of a single room at 4mm voxel resolution in active memory, even with voxel hashing. This problem can be mitigated using a traditional method from the graphics community, that is also employed e.g.\ in~\cite{kinfuls_2012,Chen13:SRV,neissner_tog_2013,Henry13:PVS}. We only hold the \textit{active} part of the map in video card memory, i.e.\ only parts that are inside or close to the current view frustum. The remainder of the map is swapped out to host memory and swapped back in as needed.
We have designed our swapping framework aiming for the following three objectives: (O1) the transfers between host and device should be minimised and have guaranteed maximum bounds, (O2) host processing time should be kept to a minimum and (O3) no assumptions should be made about the type and speed of the host memory, i.e.\ it could be a hard drive. These objectives lead to the following design considerations:
\begin{itemize}
\item \textbf{O1}: All memory transfers use a host/device buffer of \textit{fixed} user-defined size.
\item \textbf{O2}: The host map memory is configured as a voxel block array of size equal to the number of entries in the hash table. Therefore, to check if a hash entry has a corresponding voxel block in the host memory, only the hash table index needs to be transferred and checked. The host does not need to perform any further computations, e.g.\ as it would have to do if a separate host hash table were used. Furthermore, whenever a voxel block is deallocated from device memory, its corresponding hash entry is not deleted but rather marked as unavailable in device memory, and, implicitly, available in host memory. This (i) helps maintain consistency between device hash table and host voxel block storage and (ii) enables a fast visibility check for the parts of the map stored only in host memory.
\item \textbf{O3}: Global memory speed is a function of the type of storage device used, e.g.\ faster for RAM and slower for flash or hard drive storage. This means that, for certain configurations, host memory operations can be considerably slower than the device reconstruction. To account for this behaviour and to enable stable tracking, the device is constantly integrating new live depth data even for parts of the scene that are known to have host data that is not yet in device memory. This might mean that, by the time all visible parts of the scene have been swapped into the device memory, some voxel blocks might hold large amounts of new data integrated by the device. We could replace the newly fused data with the old one from the host stored map, but this would mean disregarding perfectly fine map data. Instead, after the swapping operation, we run a secondary integration that fuses the host voxel block data with the newly fused device map.
\end{itemize}
\begin{figure}[tb]
\centering
\includegraphics[width=0.7\linewidth]{images/itm_swap_overall.pdf}
\caption{Swapping pipeline}
\label{fig:itm_swap_overall}
\end{figure}
The design considerations have led us to the swapping in/out pipeline shown in Figure~\ref{fig:itm_swap_overall}. We use the allocation stage to establish which parts of the map need to be swapped in, and the integration stage to mark which parts need to swapped out. A voxel needs to be swapped (i) from host once it projects within a small (tunable) distance from the boundaries of live visible frame and (ii) to disk after data has been integrated from the depth camera.
\begin{figure}[tb]
\centering
\includegraphics[width=0.9\linewidth]{images/itm_swap_in.pdf}
\caption{Swapping in: First, the hash table entry at address \textbf{1} is copied into the device transfer buffer at address \textbf{2}. This is then copied at address \textbf{3} in the host transfer buffer and used as an address inside the host voxel block array, indicating the block at address \textbf{4}. This block is finally copied back to location \textbf{7} inside the device voxel block array, passing through the host transfer buffer (location \textbf{5}) and the device transfer buffer (location \textbf{6}).}
\label{fig:itm_swap_in}
\end{figure}
The swapping in stage is exemplified for a single block in Figure \ref{fig:itm_swap_in}. The indices of the hash entries that need to be swapped in are copied into the device transfer buffer, up to its capacity. Next, this is transferred to the host transfer buffer. There the indices are used as addresses inside the host voxel block array and the target blocks are copied to the host transfer buffer. Finally, the host transfer buffer is copied to the device where a single kernel integrates directly from the transfer buffer into the device voxel block memory.
\begin{figure}[tb]
\centering
\includegraphics[width=0.9\linewidth]{images/itm_swap_out.pdf}
\caption{Swapping Out. The hash table entry at location \textbf{1} and the voxel block at location \textbf{3} are copied into the device transfer buffer at locations \textbf{2} and \textbf{4}, respectively. The entire used transfer buffer is then copied to the host, at locations \textbf{5}, and the hash entry index is used to copy the voxel block into location \textbf{7} inside the host voxel block array.}
\label{fig:itm_swap_out}
\end{figure}
An example for the swapping out stage is shown in Figure \ref{fig:itm_swap_out} for a single block. Both indices and voxel blocks that need to be swapped out are copied to the device transfer buffer. This is then copied to the host transfer buffer memory and again to host voxel memory.
All swapping related variables and memory is kept inside the \reftocode{ITMGlobalCache} object and all swapping related operations are done by the \reftocode{ITMSwappingEngine}.
\section{Compilation, UI, Usage and Examples}\label{s:results}
As indicated in the class diagram in Figure~\ref{fig:itm_classdiagram}, the
InfiniTAM software package is split into two major parts. The bulk of the
implementation is grouped in the namespace \reftocode{ITMLib}, which contains
a stand alone 3D reconstruction library for use in other applications.
The \reftocode{InfiniTAM} namespace contains further supporting classes for
image acquisition and a GUI -- these parts would almost certainly be replaced
in user applications. Finally there is a single sample application included
allowing the user to quickly test the framework.
The project comes with a Microsoft\textregistered{} Visual
Studio\textregistered{} solution file as well as with a cmake build file and
has been tested on Microsoft\textregistered{} Windows\textregistered{} 8,
openSUSE Linux 12.3 and Apple\textregistered{} Mac\textregistered{} OS X\textregistered{} 10.9 platforms. Apart from a basic C++
software development environment, InfiniTAM depends on the following external
third party libraries:
\begin{itemize}
\item \textbf{OpenGL / GLUT} (e.g.\ freeglut 2.8.0): This is required for
the visualisation in the \reftocode{InfiniTAM} namespace and the
sample application, but the \reftocode{ITMLib} library should run
without. Freeglut is available from
\url{http://freeglut.sourceforge.net/}
\item \textbf{NVIDIA\textcopyright{} CUDA\texttrademark{} SDK}
(e.g.\ version 6.0): This is required for all GPU accelerated code.
The use of GPUs is optional however, and it is still possible to
compile the CPU part of the framework without CUDA. The CUDA SDK is
available from \url{https://developer.nvidia.com/cuda-downloads}
\item \textbf{OpenNI} (e.g.\ version 2.2.0.33): This is optional and the
framework compiles without OpenNI, but it is required to get live
images from suitable hardware sensors. Again, it is only referred to
in the \reftocode{InfiniTAM} namespace, and without OpenNI the
system will still run of previously recorded images stored on disk.
OpenNI is available from \url{http://structure.io/openni}
\end{itemize}
Finally the framework comes with a Doxygen reference documentation, that can
be built separately. More details on the build process can be found in the
\texttt{README} file provided alongside the framework.
\begin{figure}[tb]
\centering
\includegraphics[width=0.9\linewidth]{images/itm_ui.pdf}
\caption{InfiniTAM UI Example}
\label{fig:itm_ui}
\end{figure}
The UI for our InfiniTAM sample application is shown in Figure~\ref{fig:itm_ui}.
We display the raycasted reconstruction rendering, live depth image and live
colour image from the camera. Furthermore the processing time per frame and the
keyboard shortcuts available for the UI are displayed near the bottom. The
keyboard shortcuts allow the user to process the next frame, to process
continuously and to exit. Other functionalities such as exporting a 3D model
from the current reconstruction and rendering the reconstruction from arbitrary
viewpoints are not currently implemented. The UI window and interactivity is
implemented in \reftocode{UIEngine} and depends on the OpenGL and GLUT
libraries.
Running InfiniTAM requires the intrinsic calibration data for the depth (and
optionally colour) camera. The calibration is specified through an external
file, an example of which is shown below.
\begin{lstlisting}
640 480
504.261 503.905
352.457 272.202
640 480
573.71 574.394
346.471 249.031
0.999749 0.00518867 0.0217975 0.0243073
-0.0051649 0.999986 -0.0011465 -0.000166518
-0.0218031 0.00103363 0.999762 0.0151706
1135.09 0.0819141
\end{lstlisting}
This includes (i) for each camera (RGB and depth) the image size, focal length
and principal point in pixels, as per the Matlab Calibration
Toolbox~\cite{Bouguet04:CCT} (ii) the Euclidean transformation matrix mapping
points in the RGB camera coordinate system to the depth camera
and (iii) the calibration converting Kinect-like disparity values to depths.
If the depth tracker is used, the
calibration for the RGB camera and the Euclidean transformation between the two
are ignored. If live data from OpenNI is used, the disparity calibration is
ignored.
We also provide two example command lines using different data sources, OpenNI
and image files. These data sources are selected based on the number of
arguments passed to InfiniTAM and for example in the bash shell with a
Linux environment these are:
\begin{small}
\begin{verbatim}
$ ./InfiniTAM Teddy/calib.txt
$ ./InfiniTAM Teddy/calib.txt Teddy/Frames
\end{verbatim}
\end{small}
The first line starts InfiniTAM with the specified calibration file and live
input from OpenNI, while the second uses the given calibration file and the
RGB and depth images specified in the other two arguments.
We tested the OpenNI input with a Microsoft Kinect for XBOX 360, with a
PrimeSense Carmine 1.08
and with the Occipital Structure Sensor.
For offline use with images from files, these have to be in PPM/PGM format
with the RGB images being standard PPM files and the depth images being 16bit
big-endian raw Kinect disparities.
All internal library settings are defined inside the \reftocode{ITMLibSettings}
class, and they are:
\begin{lstlisting}
/// Use GPU or run the code on the CPU instead.
bool useGPU;
/// Enables swapping between host and device.
bool useSwapping;
/// Tracker types
typedef enum {
//! Identifies a tracker based on colour image
TRACKER_COLOR,
//! Identifies a tracker based on depth image
TRACKER_ICP
} TrackerType;
/// Select the type of tracker to use
TrackerType trackerType;
/// Number of resolution levels for the tracker.
int noHierarchyLevels;
/// Number of resolution levels to track only rotation instead of full SE3.
int noRotationOnlyLevels;
/// For ITMColorTracker: skip every other point in energy function evaluation.
bool skipPoints;
/// For ITMDepthTracker: ICP distance threshold
float depthTrackerICPThreshold;
/// Further, scene specific parameters such as voxel size
ITMLib::Objects::ITMSceneParams sceneParams;
\end{lstlisting}
The \reftocode{ITMSceneParams} further contain settings for the voxel size in
millimetres and the truncation band of the signed distance function.
Furthermore the file \reftocode{ITMLibDefines.h} contains definitions that
select the type of voxels and the voxel index used in the compilation of
\reftocode{ITMMainEngine}:
\begin{lstlisting}
/** This chooses the information stored at each voxel. At the moment, valid
options are ITMVoxel_s, ITMVoxel_f, ITMVoxel_s_rgb and ITMVoxel_f_rgb
*/
typedef ITMVoxel_s ITMVoxel;
/** This chooses the way the voxels are addressed and indexed. At the moment,
valid options are ITMVoxelBlockHash and ITMPlainVoxelArray.
*/
typedef ITMLib::Objects::ITMVoxelBlockHash ITMVoxelIndex;
\end{lstlisting}
For using InfiniTAM as a 3D reconstruction library in other applications, the
class \reftocode{ITMMainEngine} is recommended as the main entry point to the
\reftocode{ITMLib} library. It performs the whole 3D reconstruction algorithm.
Internally it stores the latest image as well as the 3D world model and it
keeps track of the camera pose. The intended use is as follows:
\begin{enumerate}
\item Create an \reftocode{ITMMainEngine} specifying the internal settings,
camera parameters and image sizes.
\item Get the pointer to the internally stored images with the method
\reftocode{GetView()} and write new image information to that memory.
\item Call the method \reftocode{ProcessFrame()} to track the camera and
integrate the new information into the world model.
\item Optionally access the rendered reconstruction or another image for
visualisation using \reftocode{GetImage()}.
\item Iterate the above three steps for each image in the sequence.
\end{enumerate}
The internally stored information can be accessed through member variables
\reftocode{trackingState} and \reftocode{scene}.
\section{Conclusions}\label{s:conclusions}
We tried to keep the InfiniTAM system as simple, portable and usable as
possible. We hope that the external dependencies on CUDA, OpenNI, OpenGL and
GLUT are easy to meet and that the cmake and MSVC project files allow users to
build the framework on a wide range of platforms. While the user interface and
surroundings are fairly minimalistic, the underlying library should be reliable,
robust and well tested.
We also tried to keep an eye on expandability. For example it
is fairly easy to integrate a different form of tracking into the
pipeline by reimplementing the \reftocode{ITMTracker} interface. Also different
media for swapping can be made accessible by replacing the
\reftocode{ITMGlobalCache} class. Finally it is still a manageable
process to port InfiniTAM to different kinds of hardware platforms by
reimplementing the \textit{Device Specific Layer} of the engines.
While we hope that InfiniTAM provides a reliable basis for further research,
no system is perfect. We will try to fix any problems we hear of and provide
future updates to InfiniTAM at the project website.
\small
\bibliographystyle{plain}
|
\section{Introduction}
\label{section:Introduction}
Since the introduction of {\em directed treewidth} in \cite{Reed1999,JohnsonRobertsonSeymourThomas2001}
much effort has been devoted into trying to identify algorithmically useful digraph width measures.
Such a width measure should ideally satisfy two properties. First, it should be small on several interesting instances of digraphs. Second,
many combinatorial problems should become polynomial time tractable on digraphs of constant width. While the first property is satisfied by most of the digraph width measures introduced so far
\cite{Barat2006,BerwangerDawarHunterKreutzerObdrzalek2012,BerwangerGradel2004,BerwangerGradelKaiserRabinovich2012,GruberHolzer2008,HunterKreutzer2008,Reed1999,Safari2005},
the goal of identifying large classes of problems that can be solved in polynomial time when these
measures are bounded by a constant has proven to be extremely hard to achieve. On the positive side, Johnson, Robertson, Seymour and Thomas
showed already in their seminal paper \cite{JohnsonRobertsonSeymourThomas2001} that certain linkage problems, such as Hamiltonicity and $k$-disjoint
paths (for constant $k$), can be solved in polynomial time on digraphs of constant directed treewidth.
Subsequently, It was shown in \cite{DankelmannGutinKimJung2009} that for each constant $k\in \N$, one can decide in polynomial time
the existence of a spanning tree with at most $k$ leaves on digraphs of constant directed treewidth.
More recently, it was shown in \cite{BerwangerDawarHunterKreutzerObdrzalek2012} that determining the winner for some classes of parity games
can be solved in polynomial time on digraphs of constant DAG-width \cite{BerwangerDawarHunterKreutzerObdrzalek2012}.
In this work we enrich the list of problems that can be solved in polynomial time on digraphs of constant
directed treewidth. More precisely, we devise the first algorithmic metatheorem connecting {\em directed} treewidth
to the monadic second order logic of graphs with edge set quantifications ({$\mbox{MSO}_2$\;} logic).
We show that most of the positive algorithmic results obtained so far on digraphs of constant {\em directed} treewidth can be
reformulated in terms of our metatheorem. Additionally we show how to use our metatheorem to provide polynomial time
algorithms for a parameterized version of the minimum spanning strong subgraph problem, and for a parameterized version
of the problem of counting subgraphs satisfying a given minor closed property.
We note that celebrated results due to Courcelle \cite{Courcelle1990MSO} and
Arnborg, Lagergren and Seese \cite{ArnborgLagergrenSeese1991} state that any problem expressible in
{$\mbox{MSO}_2$\;} logic can be solved in linear time on graphs of constant {\em undirected} treewidth. Additionally,
an equally famous result due to Courcelle, Makowsky and Rotics states that any problem expressible in
{$\mbox{MSO}$\;} logic (without edge set quantifications) can be solved in linear time on graphs of constant clique-width \cite{CourcelleMakowskyRotics2000}.
However, we observe that there are families of digraphs of constant {\em directed} treewidth, but simultaneously
unbounded {\em undirected} treewidth and clique-width \cite{CourcelleMakowskyRotics2000}.
For instance, the $n\times n$ grid, in which all horizontal edges are oriented to the right and all vertical edges are oriented upwards,
has {\em directed} $\mbox{treewidth $0$}$, but {\em undirected} treewidth $\Theta(n)$ and clique-width $\Theta(n)$.
Thus our algorithmic metatheorem is not implied by the results in \cite{Courcelle1990MSO,ArnborgLagergrenSeese1991,CourcelleMakowskyRotics2000}. On the
other hand, the fact that $3$-colorability is MSO expressible
implies that a complete analog of Courcelle's is theorem for digraphs of
constant directed treewidth cannot be achieved unless P=NP, since $3$-colorability is already NP-complete on DAGs.
Before stating our main theorem we will introduce some notation.
An edge-weighting function for a digraph $G=(V,E)$ is a function $\graphweightingfunction:E\rightarrow \Omega$ where $\Omega$
is a finite commutative semigroup of size polynomial in $|V|$. We will always assume that $\Omega$
has an identity element. We define the size of $G$ as $|G|= |V|+|E|$.
The weight of a subgraph $H=(V',E')$ of $G$ is defined as $\graphweightingfunction(H)=\sum_{e\in E'} \graphweightingfunction(e)$.
We say that $H$ is the union of $k$ directed paths if there exist directed simple paths
$\mathfrak{p}_1,\mathfrak{p}_2,...,\mathfrak{p}_k$ with $\mathfrak{p}_i=(V_i,E_i)$ for $i\in \{1,...,k\}$ such that
$H=\mathfrak{p}_1\cup \mathfrak{p}_2\cup ...\cup \mathfrak{p}_k = (\cup_{i=1}^k V_i,\cup_{i=1}^k E_i)$.
We note that the unions we consider are not necessarily vertex-disjoint nor edge-disjoint.
\begin{theorem}[Main Theorem]
\label{theorem:MainTheoremDirectedTreewidth}
Let $\varphi$ be an {$\mbox{MSO}_2$\;} sentence and let $k,w\in \N$. There is a computable function
$f(\varphi,w,k)$ such that, given a weighted digraph $G=(V,E,\graphweightingfunction\!:\!E\rightarrow \Omega)$ of directed treewidth $w$,
a positive integer $l<|V|$, and an element $\alpha \in \Omega$, one can count in time $f(\varphi,w,k)\cdot |G|^{O(k\cdot(w+1))}$ the number of subgraphs $H$ of $G$
simultaneously satisfying the following four properties:
\begin{enumerate}[(i)]
\item $H\models \varphi$,
\item $H$ is the union of $k$ directed paths,
\item $H$ has $l$ vertices,
\item\label{item:Weight} $H$ has weight $\graphweightingfunction(H)=\alpha$. \newline
\end{enumerate}
\end{theorem}
We note that in \cite{deOliveiraOliveira2013IPEC} we proved an analog theorem for digraphs of constant {\em directed} pathwidth.
Nevertheless it can be shown that there exist families of digraphs of constant directed treewidth but unbounded {\em directed}
pathwidth \cite{BerwangerDawarHunterKreutzerObdrzalek2012}. Therefore, Theorem \ref{theorem:MainTheoremDirectedTreewidth} is a strict generalization of the results in \cite{deOliveiraOliveira2013IPEC}.
To prove Theorem \ref{theorem:MainTheoremDirectedTreewidth} we will introduce two new technical tools which may be of independent
interest. The first, the tree-zig-zag number of a digraph, is a new directed width measure that is at most a constant times directed
treewidth. The second, the notion of $z$-saturated tree slice languages, is a new framework for the manipulation of infinite families of digraphs.
\subsection{Applications}
\label{section:Applications}
The parameters $l$ and $\alpha$ in Theorem \ref{theorem:MainTheoremDirectedTreewidth} are upper bounded by
$|V|^{O(1)}$. By varying these parameters we can consider different flavours of optimization problems.
For instance, we can choose to count the number of subgraphs of $G$ that are the union of $k$ directed paths, satisfy
$\varphi$ and have maximal/minimal number of vertices, or maximal/minimal weight. In this section we provide a list of
natural combinatorial problems that can be solved in polynomial time on digraphs of constant directed treewidth
using Theorem \ref{theorem:MainTheoremDirectedTreewidth}. In Subsection \ref{subsection:ApplicationsKnown}, we show
how to use Theorem \ref{theorem:MainTheoremDirectedTreewidth} to rederive three known positive algorithmic results
for digraphs of constant directed treewidth. In Subsection \ref{subsection:ApplicationsNotKnown}, we show
how Theorem \ref{theorem:MainTheoremDirectedTreewidth} can be used to solve in polynomial time two interesting classes
of combinatorial problems which have not yet been studied in the context of digraph width measures. Concerning the
first class of problems, we
show how to count the number of {\em minimum spanning strong subgraphs} that are the union of $k$ directed paths. Concerning the second class,
we show how to count the number of maximal subgraphs that are the union of $k$ directed paths and satisfy some given minor closed property.
\subsubsection{First Examples}
\label{subsection:ApplicationsKnown}
In order to use Theorem \ref{theorem:MainTheoremDirectedTreewidth} to solve a counting problem
in polynomial time, we need to exhibit an {$\mbox{MSO}_2$\;} sentence $\varphi$ specifying a suitable class
of digraphs to be counted, and to specify values for the parameters $l$ and $\alpha$ which respectively determine
the number of vertices and the weight of the subgraphs being counted. We observe that the
class of digraphs specified by $\varphi$ is fixed and does not vary with the input digraph.
The parameters $l$ and $\alpha$ on the other hand, may vary with the input.
\paragraph{\textbf{Counting Hamiltonian Cycles}}
We set $\varphi$ to be an {$\mbox{MSO}_2$\;} sentence defining cycles, i.e., connected digraphs in which each vertex has precisely one
incoming edge and one outgoing edge. We set $l=|V|$ since we are only interested in counting sub-cycles of $G$ that span all
of its vertices. Finally, since any cycle is the union of $2$ directed paths, we set $k=2$.
We observe that counting Hamiltonian cycles on digraphs of constant directed treewidth can also be done
via an adaptation of the techniques in \cite{JohnsonRobertsonSeymourThomas2001}.
\paragraph{\textbf{Counting $\sigma$-Linkages}}
Given a sequence $\sigma= (s_1,t_1,s_2,t_2,...,s_k,t_k)$ of $2k$ not necessarily distinct vertices, a $\sigma$-linkage
is a set of internally disjoint directed paths $\mathfrak{p}_1,\mathfrak{p}_2,...,\mathfrak{p}_k$
where for each $i\in \{1,...,k\}$, the path $\mathfrak{p}_i$ connects $s_i$ to $t_i$. To count the number of $\sigma$-linkages
on a digraph $G$ we first assign a distinct color to each vertex in the set
$\{s_1,...,s_k,t_1,...,t_k\}$ and assume that all other vertices of $G$ are uncolored.
Then we define an {$\mbox{MSO}_2$\;} sentence $\varphi_{\sigma}$ that is true in a digraph $H$ whenever
it consists of the union of $k$ internally disjoint paths $\mathfrak{p}_1,...,\mathfrak{p}_k$
where for each $i$, the path $\mathfrak{p}_i$ connects a vertex of color $c(s_i)$ to a vertex of color $c(t_i)$ in such
a way that all internal vertices of $\mathfrak{p}_i$ are uncolored.
For each $l\in \{1,...,|V|\}$ we can use Theorem \ref{theorem:MainTheoremDirectedTreewidth} to
count the number of $\sigma$-linkages of size $l$. We observe that counting $\sigma$-linkages
can also be done by via an adaptation of the results in \cite{JohnsonRobertsonSeymourThomas2001}.
\paragraph{\textbf{Counting Spanning-Out Trees with at most $k$-leaves}}
A spanning-out tree is a spanning tree in which all edges are directed towards the leaves.
To count the number of spanning-out trees with at most $k$-leaves we set
$\varphi$ to be an {$\mbox{MSO}_2$\;} sentence defining trees with at most $k$-leaves. In other words, $\varphi$ defines
connected digraphs without cycles in which at most $k$ vertices have no out-going edge. Since the tree has to span
all vertices of $G$, we set $l=|V|$. Finally, we note that any spanning-out tree with at most $k$ leaves is the union of $k$ directed paths.
We observe that counting spanning-out trees can also be done via an adaptation of the results in
in \cite{DankelmannGutinKimJung2009}.
\subsubsection{New Applications}
\label{subsection:ApplicationsNotKnown}
In this section we exhibit two natural classes of counting problems that can be
solved in polynomial time on digraphs of constant directed treewidth using Theorem
\ref{theorem:MainTheoremDirectedTreewidth}. To the best of our knowledge these
problems cannot be addressed in polynomial time using previously existing techniques.
\paragraph{\bf Minimum Spanning Strong Subgraph}
The classic {\em Minimum Spanning Strong Subgraph (MSSS) problem} is defined as follows.
Given a strongly connected digraph $G$, find a spanning strongly connected subgraph of $G$ with
the minimum number of edges. This problem is in general NP complete since it generalizes
the Hamiltonian cycle problem. Even though the MSSS problem has received a considerable amount of
attention \cite{Aho1972transitive,BangHuangYeo2003Strongly,BessyThomasse2003,BangYeo2001Minimum,Gabow2004special,Vetta2001Approximating},
the connections between this problem and directed width measures are, to the limit of our knowledge, unexplored. Here we show
that a parameterized version of the MSSS problem can be solved in polynomial time on digraphs
of constant directed treewidth. A $k$-MSSS is a minimum spanning strong subgraph that is the union of $k$ directed paths.
We note that determining the existence of a $\mbox{$k$-MSSS}$ on general digraphs is still NP-complete for each constant $k\geq 2$,
since any Hamiltonian cycle is a $2$-MSSS. Using Theorem \ref{theorem:MainTheoremDirectedTreewidth}
we can not only determine the existence of a $k$-MSSS on digraphs of constant directed treewidth,
but also count in polynomial time the number of occurrences of such subgraphs. All we need to do is to set $l=|V|$, since the subgraphs
we are counting are spanning, and to set $\varphi_{\mathit{str}}$ as the monadic second order sentence
that is true in a digraph if and only if it is strongly connected.
We observe that the techniques in \cite{JohnsonRobertsonSeymourThomas2001} cannot be directly applied
to solve the $k$-MSSS problem in polynomial time due to the fact that the $k$ paths covering a
$k$-MSSS need not to be internally disjoint. For instance, in Figure \ref{figure:MSSS} we show a family $H_1,H_2,...$
of digraphs where for each $n\in \N$, $H_n$ is the union of $2$ paths. Note however that one needs
$2n$ internally disjoint paths to cover all vertices and edges of $H_n$.
\begin{figure}[!hb]
\centering
\includegraphics[scale=0.30]{MSSS}
\vspace{-5pt}
\caption{For each $n\in \N$, the digraph $H_n$ is the union of $2$ paths. On the other hand,
$2n$ internally disjoint paths are necessary to cover all vertices and edges of $H_n$.}
\label{figure:MSSS}
\end{figure}
\vspace{-5pt}
\paragraph{\bf Subgraphs Excluding a Minor}
An undirected graph $H$ is a minor of an undirected graph $G$ if $H$
can be obtained from a subgraph of $G$ by a sequence of edge contractions.
A family of undirected graphs $\mathcal{F}$ is said to be minor closed if whenever
a graph $G$ belongs to $\mathcal{F}$, any minor of $G$ is also in $\mathcal{F}$.
Many interesting graph families are minor closed, such as, planar graphs, outerplanar graphs,
graphs of bounded genus, forests, series-parallel graphs, graphs of bounded undirected treewidth, etc.
Given a minor closed family $\mathcal{F}$ and a graph $G$ it is often NP-complete to find
a maximal subgraph of $G$ that belongs to the family $\mathcal{F}$. For instance, the following
problems are NP-complete: finding a maximal outerplanar subgraph
\cite{Cimikowski1996Sizes,Poranen2005Heuristics}, finding a maximal planar subgraph
\cite{JungerMutzel1996,Cualinescu1998} and finding a maximal subgraph of a given genus $g$ \cite{Cualinescu1996Finding}.
By the celebrated graph minor theorem of Robertson and Seymour \cite{RobertsonSeymour2004}, for any minor closed
family of undirected graphs $\mathcal{F}$ there exists a finite set of undirected graphs $\hat{\mathcal{F}}$,
such that for each graph $H$, $H\in \mathcal{F}$ if and only if none of the graphs in $\hat{\mathcal{F}}$
is a minor of $H$. Thus, using the finite set $\hat{\mathcal{F}}$ one can define an {$\mbox{MSO}_2$\;} sentence
$\varphi_{\mathcal{F}}$ such that $\varphi_{\mathcal{F}}$ is true in a graph $H$ if and only if $H\in \mathcal{F}$ (see for instance \cite{CourcelleEngelfriet2012}).
This fact implies that Theorem \ref{theorem:MainTheoremDirectedTreewidth} can be used to count in polynomial time,
on digraphs of constant directed treewidth, the number of subgraphs that are the union of $k$ directed
paths and whose underlying undirected graph satisfy a minor closed property. More precisely,
if $H$ is a directed graph, let $\stackrel{\leftrightarrow}{H}$ denote the undirected graph obtained from $H$ by
forgetting the directions of the edges in $H$. We have the following corollary of Theorem \ref{theorem:MainTheoremDirectedTreewidth}.
\begin{corollary}
\label{corollary:StructuralCounting}
Let $\mathcal{F}$ be a minor closed family of undirected graphs, $G$ be a digraph of directed treewidth $w$, and let $k\in \N$.
Then one can count in time $f(\varphi_{\mathcal{F}},k,w)\cdot |G|^{O(k\cdot (w+1))}$ the number of (maximal) subgraphs $H$
of $G$ subject to the following restrictions:
\begin{enumerate}
\item $H$ is the union of $k$ directed paths.
\item $\stackrel{\leftrightarrow}{H}$ belongs to $\mathcal{F}$.
\end{enumerate}
\end{corollary}
For instance, Corollary \ref{corollary:StructuralCounting} implies that we can count in polynomial time, on digraphs of constant directed
treewidth, maximal planar subgraphs that are the union of $k$ directed paths, or maximal subgraphs that are the
union of $k$ directed paths and can be embedded on a torus. In our opinion it is rather surprising that the problems addressed in
Corollary \ref{corollary:StructuralCounting} can be solved in polynomial time, in view of the complexity of the subgraphs
that are being counted, and in view of the fact that digraphs of constant directed treewidth may have simultaneously unbounded
{\em undirected} treewidth and clique-width.
\subsection{Hardness Results}
We argue briefly that under the assumption that the \textbf{W} hierarchy does not collapse to \textbf{FPT}, a widely believed assumption in parameterized
complexity theory \cite{DowneyFellows1992}, the dependence on $w$ and
on $k$ on the exponent of the running time $f(\varphi,w,k)\cdot |G|^{O(k\cdot (w+1))}$ of Theorem \ref{theorem:MainTheoremDirectedTreewidth} cannot be removed.
Concerning the dependence on $k$, we note that the problem of determining whether there exists
$k$ disjoint paths on DAGs from prescribed pairs of nodes is \textbf{W[1]} hard with respect to $k$ \cite{Slivkins2003}. Since any
DAG has {\em directed} treewidth $0$, we have that the existence of $k$-disjoint paths is already \textbf{W[1]}-hard even for digraphs of directed treewidth $0$.
Concerning the dependence on $w$, we note that it can be shown \cite{LampisKaouriMitsou2011} that finding Hamiltonian paths on digraphs is \textbf{W[2]}
hard with respect to the cycle-rank of the digraph in question. Since constant directed treewidth is more expressive than constant
cycle rank, the hardness results in \cite{LampisKaouriMitsou2011} extends to {\em directed} treewidth. Thus the dependence on $w$ in the
exponent of the running time $f(\varphi,w,k)\cdot |G|^{O(k\cdot (w+1))}$ cannot be removed even if $k=1$.
\subsection{Proof Techniques and Organization of the Paper}
\label{subsection:ProofTechniques}
We will prove Theorem \ref{theorem:MainTheoremDirectedTreewidth} using slice theoretic techniques.
The notion of slice language was introduced in \cite{deOliveiraOliveira2010}
and used to solve several problems in the partial order theory of concurrency.
Subsequently, slice languages were lifted to the context of digraphs and used to provide the first algorithmic
metatheorem for digraphs of constant {\em directed} pathwidth \cite{deOliveiraOliveira2013IPEC}.
In this work we extend the results in \cite{deOliveiraOliveira2013IPEC} by introducing the
notions of {\em tree slice language} and {\em slice tree automata}. We use {\em tree} slice-languages to provide the
first algorithmic metatheorem for digraphs of constant {\em directed} treewidth (Theorem \ref{theorem:MainTheoremDirectedTreewidth}).
More precisely, we will show that the problem of counting the number of subgraphs satisfying
the conditions ($i$)-($iv$) of Theorem \ref{theorem:MainTheoremDirectedTreewidth} can be reduced to
the problem of counting the number of terms accepted by a suitable deterministic slice tree-automaton.
We note that the results in this work strictly generalize the results in \cite{deOliveiraOliveira2013IPEC},
since there are families of digraphs of constant {\em directed} treewidth but unbounded {\em directed} pathwidth.
Below we give a brief description of the main technical tools used in this paper and how they fit
together to yield a proof of Theorem \ref{theorem:MainTheoremDirectedTreewidth}. All notions introduced
in the following paragraphs will be re-defined more carefully along the paper.
A unit slice of arity $r$ is a digraph ${\mathbf{S}}$ whose vertex set is partitioned into a center $C$,
an out-frontier $F_0$ and $r$ in-frontiers $F_1,...,F_r$ in such a way that the center $C$ has at
most one vertex and each frontier vertex is incident with precisely one edge of ${\mathbf{S}}$ (Figure \ref{figure:Slices}).
Intuitively a slice ${\mathbf{S}}$
can be glued to a slice ${\mathbf{S}}'$ at frontier $j$ if the out-frontier of ${\mathbf{S}}$ can be matched with the $j$-th in-frontier of
${\mathbf{S}}'$. A finite set $\mathbold{\Sigma}$ of slices with possibly distinct arities is called a slice alphabet. In this paper we will only
be interested in slices of arity $0$, $1$ and $2$. A term over $\mathbold{\Sigma}$
is a tree-like expression $\mathbf{T}$ in which each node $p$ is labeled with
a slice $\mathbf{T}[p]$ whose arity is equal to the number of children of $p$. We say that $\mathbf{T}$ is a unit decomposition if
for each position $pj$ the slice $\mathbf{T}[pj]$ can be glued to the slice $\mathbf{T}[p]$ at its $j$-th frontier (Figure \ref{figure:UnitDecomposition}).
Each unit decomposition $\mathbf{T}$ gives rise to a digraph $\composedT$ which is intuitively obtained by glueing
each two adjacent slices in $\mathbf{T}$ along their matching frontiers. Conversely, for each digraph $G$ there is a suitable
slice alphabet $\mathbold{\Sigma}$ and unit decomposition $\mathbf{T}$ over $\mathbold{\Sigma}$ such that $\composedT$ is
isomorphic to $G$. We can represent infinite families of digraphs via tree-automata over slice alphabets. We say that
such an automaton ${\mathcal{A}}$ is a slice tree-automaton if all terms generated by ${\mathcal{A}}$ are
unit decompositions. With a slice tree-automaton ${\mathcal{A}}$ one can associate two types of languages. The first,
the slice language $\lang({\mathcal{A}})$, is simply
the set of all unit decompositions accepted by ${\mathcal{A}}$. The second, the graph language $\lang_{{\mathcal{G}}}({\mathcal{A}})$
is the set of all digraphs represented by unit decompositions in $\lang({\mathcal{A}})$.
We say that a unit decomposition $\mathbf{T}$ has tree-zig-zag number $z$ if each {\em simple} path in the
digraph $\composedT$ represented by $\mathbf{T}$ crosses each frontier of each slice in $\mathbf{T}$ at most $z$ times $\mbox{(Figure \ref{figure:UnitDecomposition})}.$
A slice tree-automaton ${\mathcal{A}}$ has tree-zig-zag number $z$ if each unit decomposition $\mathbf{T}\in \lang({\mathcal{A}})$
has tree-zig-zag number $z$. Finally, we say that a slice tree-automaton ${\mathcal{A}}$ is $z$-saturated over a slice alphabet $\mathbold{\Sigma}$
if the presence of a digraph $H$ in the graph language $\lang_{{\mathcal{G}}}({\mathcal{A}})$ implies that each
unit decomposition $\mathbf{T}$ of tree-zig-zag number $z$ representing $H$ belongs to $\lang({\mathcal{A}})$.
The importance of the notion of saturation stems from the following fact. Given a slice tree-automaton
${\mathcal{A}}$ of tree-zig-zag number $z$ and a slice tree-automaton ${\mathcal{A}}'$ that
is $z$-saturated, it is possible to show that $\lang_{{\mathcal{G}}}({\mathcal{A}}\cap {\mathcal{A}}') = \lang_{{\mathcal{G}}}({\mathcal{A}})\cap \lang_{{\mathcal{G}}}({\mathcal{A}}')$.
In other words, the set of digraphs represented by the intersection ${\mathcal{A}}\cap {\mathcal{A}}'$ is equal
to the intersection of the sets of digraphs represented by ${\mathcal{A}}$ and ${\mathcal{A}}'$ separately.
We note that this crucial property is not satisfied by general slice tree-automata.
Within this framework, the proof of Theorem \ref{theorem:MainTheoremDirectedTreewidth} can be divided into the following
steps.
\begin{enumerate}
\item In the first step, we will show that given a digraph $G$ of directed treewidth $w$
one can construct a unit decomposition $\mathbf{T}$ of $G$ of tree-zig-zag number $z\leq 9w+18$. This construction
will follow from a combination of Theorem \ref{theorem:ComparisonWithOtherMeasures} with
Proposition \ref{proposition:OliveTreeDecompositionUnitDecomposition}. Subsequently, we will show that using
$\mathbf{T}$ one can construct a slice tree-automaton ${\mathcal{A}}(\mathbf{T},k\cdot z)$ of tree-zig-zag number $z$
whose graph language contains all subgraphs of $G$ that are the union of $k$ directed paths. The construction of
${\mathcal{A}}(\mathbf{T},k\cdot z)$ will be given in the proof of Lemma \ref{lemma:SubgraphsC}.
\item In the second step we will show that given an {$\mbox{MSO}_2$\;} sentence $\varphi$ and an integer $k$, one can automatically
construct a $z$-saturated slice tree-automaton ${\mathcal{A}}(\varphi,k,z)$ whose graph language $\lang_{{\mathcal{G}}}({\mathcal{A}}(\varphi,k,z))$
consists precisely of the digraphs which at the same time satisfy $\varphi$ and are the union of $k$ directed paths
(Theorem \ref{theorem:MonadicSliceTreeAutomataZSaturated}).
Additionally, given a positive integer $k\in \N$ and a weight $\alpha \in \Omega$, we can use ${\mathcal{A}}(\varphi,k,z)$
to construct another $z$-saturated tree-automaton ${\mathcal{A}}(\varphi,k,z,l,\alpha)$ whose graph language contains
only those digraphs generated by ${\mathcal{A}}(\varphi,k,z)$ which have $l$ vertices and weight $\alpha$ (Lemma \ref{lemma:AutomatonSizeWeight}).
\item Finally, in the third step we will show that the slice language of the tree-automaton
${\mathcal{A}}(\mathbf{T},k\cdot z)\cap {\mathcal{A}}(\varphi,k,z,l,\alpha)$ has precisely one unit decomposition
$\mathbf{T}$ for each subgraph of $G$ that is the union of $k$ directed paths,
satisfy $\varphi$ and have prescribed length $l$ and weight $\alpha$. This claim will follow from Lemma \ref{lemma:IntersectionSubdecompositions}
using the fact that ${\mathcal{A}}(\mathbf{T},k\cdot z)$ has tree-zig-zag number $z$ and that ${\mathcal{A}}(\varphi,k,z,l,\alpha)$ is $z$-saturated.
At this point, the problem of counting subgraphs of $G$ satisfying these four properties boils down to the problem of
counting the number of unit decompositions accepted by ${\mathcal{A}}(\mathbf{T},k\cdot z)\cap {\mathcal{A}}(\varphi,k,z,l,\alpha)$.
Since the latter automaton is deterministic, this counting process can be carried in polynomial time.
This step will be carried in Theorem \ref{theorem:CountingSizeWeight} via an application of Theorem \ref{theorem:CountingSubgraphs}.
\end{enumerate}
The remainder of this paper is organized as follows. Next, in Section \ref{section:DirectedTreewidth} we recall
the definition of directed treewidth \cite{JohnsonRobertsonSeymourThomas2001}. Subsequently, in
Section \ref{section:TreeZigZagNumber}, we introduce the {\em tree-zig-zag} number of a digraph, a new directed width
measure.
In Section \ref{section:ComparisonDirectedTreewidth}, we show that the tree-zig-zag number of a
digraph is at most a constant times its directed treewidth. In Section \ref{section:TreeAutomata} we recall
some of the main definitions of tree-automata theory.
In Section \ref{section:TreeSliceLanguage} we introduce tree slice languages and
slice tree-automata. In Section \ref{section:zSaturationAndCounting} we introduce the notion of $z$-saturation
and state a slice theoretic metatheorem (Theorem \ref{theorem:CountingSubgraphs}).
In Section \ref{section:MSOandTreeSliceLanguages}
we will show that for any {$\mbox{MSO}_2$\;} sentence $\varphi$ and any $k,z\in \N$ one can construct a $z$-saturated slice automaton
${\mathcal{A}}(\varphi,k,z)$ whose graph language consists of all digraphs that are the union of $k$ directed paths and satisfy $\varphi$.
In $\mbox{Section \ref{section:ProofOfMainTheorem}}$ we will show how to restrict ${\mathcal{A}}(\varphi,k,z)$ into an automaton
${\mathcal{A}}(\varphi,k,z,l,\alpha)$ whose graph language consists precisely of the digraphs that, at the same time, are the union of
$k$ paths, satisfy $\varphi$, have $l$ vertices and weight $\alpha$. In the same section we prove our main theorem,
Theorem \ref{theorem:MainTheoremDirectedTreewidth}. Finally, in Section \ref{section:Conclusion} we make
some final remarks and discuss some future directions.
\section{Directed Treewidth}
\label{section:DirectedTreewidth}
In this section we recall the definitions of arboreal decomposition and {\em directed} treewidth. For a matter of uniformity with other notions of
tree decompositions encountered in this paper, our notation slightly differs from the notation used in \cite{JohnsonRobertsonSeymourThomas2001}.
Let $\{1,...,r\}^*$ denote the set of all strings over $\{1,...,r\}$ and let ${\lambda}$ denote
the empty string. A subset
$N\subseteq \{1,...,r\}^*$ is prefix closed if for every $p\in \{1,...,r\}^*$ and every $j\in \{1,...,r\}$, $pj\in N$ implies that $p\in N$.
We note that the empty string ${\lambda}$ is an element of any prefix closed subset of $\{1,...,r\}^*$.
We say that $N\subseteq \{1,...,r\}^*$ is well numbered if for every $p\in \{1,...,r\}^*$ and every $j\in \{1,...,r\}$,
the presence of $pj$ in $N$ implies that $p1,...,p(j-1)$ also belong to $N$.
An $r$-ary tree is a pair $T=(N,F)$ whose set of nodes $N$ is a finite prefix
closed, well numbered subset of $\{1,...,r\}^*$, and whose set of arcs $F$ is defined as $F = \{(p,pj)\;|\; p,pj\in N, j \in \{1,...,r\}\}$.
Observe that by our definition, the root of an $r$-ary tree is the empty string ${\lambda}$.
A binary tree is an $r$-ary tree in which $r=2$.
If $pj\in N$, then we say that
$pj$ is a child of $p$, or interchangeably, that $p$ is the parent of $pj$. A {\em leaf} is a node $p\in N$ without children.
If $pu\in N$ for $u\in \{1,...,r\}^*$, then we say that
$pu$ is a descendant of $p$. For a node $p\in N$ we let $N(p) =\{pu\in N\; |\; u\in \{1,...,r\}^*\}$
denote the set of all descendants of $p$. Note that $p$ is a descendant of itself and therefore $p\in N(p)$.
Let $G=(V,E)$ be a digraph and let $Z$ and $K$ be two disjoint subsets of vertices of $G$.
We say that $K$ is $Z$-normal if there is no directed walk in $V\backslash Z$ with first and
last vertex in $K$ that uses a vertex of $V\backslash (Z\cup K)$. In other words, $K$ is $Z$-normal if every walk which
starts and ends in $K$ is either wholly contained in $K$ or uses a vertex of $Z$.
An arboreal decomposition of a digraph $G=(V,E)$ is a four-tuple $\mathcal{D} = (N,F,W,Z)$
where $(N,F)$ is an $r$-ary tree for some $r\in \N$,\; $W:N\rightarrow 2^{V}$ is a function that associates
with each node $p\in N$ a {\em non-empty} set of vertices $W(p)\subseteq V$, and $Z:F\rightarrow 2^{V}$ is a function that
associates with each arc $(p,pj)\in F$, a set of vertices $Z(p,pj)$. In the sequel, we may refer to the
sets $W(p)$ as the {\em bags} of $\mathcal{D}$. For a node $p\in N$ we let
$V(p,\mathcal{D}) = \bigcup_{u\in N(p)} W(u)$ denote the set of all vertices of $G$ that belong
to some bag associated with a descendant of $p$.
The functions $W$ and $Z$ satisfy the following two properties.
\begin{enumerate}[1)]
\itemsep0.2em
\item \label{item:ArborealOne} $\{W(p)\;|\; p\in N\}$ is a partition of $V$ into non-empty sets.
\item \label{item:ArborealTwo} For each $(p,pj)\in F$, the set $V(pj,\mathcal{D})$ is $Z(p,pj)$-normal.
\end{enumerate}
Intuitively, Condition \ref{item:ArborealTwo} says that for each $(p,pj)\in F$, the set of all vertices of
$G$ that belong to bags associated with descendants of $pj$ is $Z(p,pj)$ normal.
If $e$ is an arc in $F$ and $p$ is a node in $N$ then we write $e\sim p$ to indicate that $p$
is incident with $e$. In other words, $e\sim p$ means that either $e=(p,p')$ or $e=(p',p)$ for some $p'\in N$.
The width $w(\mathcal{D})$ of the arboreal decomposition $\mathcal{D}$ is the least
integer $w$ such that for every node $p\in N$, $|W(p) \cup \bigcup_{e\sim p} Z(e)| \leq w+1$.
The {\em directed treewidth} of $G$ is the
least integer $w$ such that there is an arboreal decomposition of $G$ of width $w$. An arboreal
decomposition $\mathcal{D}=(N,F,W,Z)$ of a digraph $G$ is {\em good} if additionally the following condition is satisfied.
\begin{enumerate}[3)]
\item For each position $p\in N$, if $pi\in N$ and $pj\in N$ with $i<j$, then there is no edge in $G$ with source in $V(pj,\mathcal{D})$
and target in $V(pi,\mathcal{D})$.
\end{enumerate}
A haven of order $w$ in a digraph $G=(V,E)$ is a function $\beta$ that assigns to each set $Z\subseteq V$ with $|Z|<w$,
the vertex-set of a strongly connected component of the digraph $G\backslash Z$, in such a way
that for each two sets of vertices $Z,Z' \subseteq V$, if $Z'\subseteq Z$ with $|Z|< w$, then $\beta(Z)\subseteq \beta(Z')$.
It can be shown that if $G$ has a haven of order $w$ then its directed treewidth
is at least $w-1$. Theorem $3.3$ of reference \cite{JohnsonRobertsonSeymourThomas2001} states that a digraph $G$ either has
directed treewidth at most $3w-2$, or it has a haven of order $w$. The proof of this theorem is algorithmic. The algorithm
either constructs a good arboreal decomposition of $G$ of width $3w-2$ or declares that $G$ has a haven of order $w$. Since
a haven of order $w$ is a certificate that the directed treewidth of $G$ is at least $w-1$, one can be sure that
if the directed treewidth of $G$ is at most $w-2$, a good arboreal decomposition of $G$ of width at most $3w-2$ will be found.
Equivalently, if $G$ has directed treewidth at most $w$ then one can always find an arboreal decomposition for $G$ of width
at most $3w+4$.
\begin{theorem}[\cite{JohnsonRobertsonSeymourThomas2001}]
\label{theorem:ConstructionGoodArborealDecomposition}
Let $G$ be a digraph of directed treewidth at most $w$. One can construct in time $|G|^{O(w)}$ a good arboreal
decomposition of $G$ of width at most $3w+4$.
\end{theorem}
\section{Olive-Tree Decompositions and the Tree-Zig-Zag Number of a Digraph}
\label{section:TreeZigZagNumber}
In this section we will introduce the tree-zig-zag number of a digraph, a new directed width measure.
Next, in Section \ref{section:TreeZigZagNumber} we will show that the tree-zig-zag number of a digraph
is at most a constant times its directed treewidth.
\begin{definition}
\label{definition:OliveTreeDecomposition}
An {\em olive-tree decomposition} of a digraph $G=(V,E)$ is a triple $\mathcal{T} = (N,F,\mathfrak{m})$ where
$(N,F)$ is a binary tree and $\mathfrak{m}:V\rightarrow N$ is an injective map from vertices of $G$ to nodes of $T$.
\end{definition}
The notion of olive-tree decomposition is similar to the notion of carving decomposition introduced
by Seymour and Thomas in \cite{SeymourThomas1994}. The only difference is that in a carving decomposition, as defined in \cite{SeymourThomas1994}, the vertices
of the digraph $G$ are bijectively mapped to the leaves of the tree, while in our definition these vertices can also be mapped to
the internal nodes of the tree, and the mapping is required to be injective, but not necessarily bijective.
If $\mathcal{T} = (N,F,\mathfrak{m})$ is an olive-tree decomposition of a digraph $G=(V,E)$
then we let $V(p,\mathcal{T})= \mathfrak{m}^{-1}(N(p))$ denote the set of all vertices of $G$ that are mapped to some
descendant of $p$.
If $V_1,V_2\subseteq V$ are two subsets of vertices of $G$ with $V_1\cap V_2 =\emptyset$, then
we let $E(V_1,V_2)$ denote the set of all edges of $G$ with one endpoint in $V_1$ and another endpoint in $V_2$.
The width $w(p)$ of a node $p\in N$ is defined as $w(p)=|E(V(p,\mathcal{T}),V\backslash V(p,\mathcal{T}))|$. The width $w(\mathcal{T})$
of $\mathcal{T}$ is defined as the maximum width of a node in $N$. More precisely,
$w(\mathcal{T}) = \max \{w(p)\;|\; p\in N\}$.
We observe that an olive-tree decomposition of a digraph $G=(V,E)$ has width at most $|E|$.
In this work we will not be interested in olive-tree decompositions of minimum width.
Rather, we will be concerned with decompositions having small {\em tree-zig-zag number}, a digraph width measure
that will be defined below.
Let $\mathcal{T}= (N,F,\mathfrak{m})$ be
an olive-tree decomposition of a digraph $G=(V,E)$, $H=(V',E')$ be a subgraph of $G$,
and $\mathfrak{m}|_{V'}:V'\rightarrow N$ be the restriction of $\mathfrak{m}$ to $V'$.
We say that the
triple $\mathcal{T}' = (N,F,\mathfrak{m}|_{V'})$ is the olive-tree decomposition of $H$ induced by $\mathcal{T}$.
A simple path in a digraph $G$ is an alternated sequence $\mathfrak{p}=v_1e_1v_2e_2....v_{n-1}e_{n-1}v_n$ of vertices
and edges of $G$ such that for each $i\in \{1,...,n-1\}$, the edge $e_i$ has $v_{i}$ as source and $v_{i+1}$ as target, and such that $v_i\neq v_j$ for
each $i,j$ with $i\neq j$.
We view $\mathfrak{p}$ as a subgraph of $G$ by setting $\mathfrak{p}=(V_{\mathfrak{p}},E_{\mathfrak{p}})$ where
$V_{\mathfrak{p}}=\{v_1,v_2,...,v_n\}$ and $E_{\mathfrak{p}}=\{e_1,e_2,...,e_{n-1}\}$. We let
$$w(\mathcal{T},\mathfrak{p}) = \max_{u\in N} |\; E_{\mathfrak{p}} \cap E(V(u,\mathcal{T}), V\backslash V(u,\mathcal{T}))\;|$$
be the width of the path $\mathfrak{p}$ along the olive-tree decomposition $\mathcal{T}$. Intuitively,
$w(\mathcal{T},\mathfrak{p})$ quantifies the amount of times the path $\mathfrak{p}$ enters or leaves the set $V(u,\mathcal{T})$ for
each $u\in N$.
\begin{definition}[Tree-Zig-Zag Number]
Let $G=(V,E)$ be a digraph and $\mathcal{T}=(N,F,\mathfrak{m})$ be an olive-tree decomposition of $G$. The
tree-zig-zag number of $\mathcal{T}$ is defined as
$$
{tzn}(\mathcal{T}) = \max\{w(\mathcal{T}, \mathfrak{p}) \; | \; \mathfrak{p} \mbox{ is a simple path in $G$} \}.
$$
The tree-zig-zag number of $G$ is defined as the minimum tree-zig-zag number of an olive-tree decomposition of $G$:
$${tzn}(G) = \min\{ {tzn}(\mathcal{T}) \; |\; \mathcal{T} \mbox{ is an olive-tree decomposition of $G$}\}.$$
\end{definition}
In Equation \ref{equation:ComparisonWidthMeasures} below we compare the tree-zig-zag number of a digraph with several
other directed width measures. In \cite{deOliveiraOliveira2013IPEC} we defined the zig-zag number ${zn}(G)$
of a digraph $G$ as a measure that quantifies the amount of times a directed path is allowed enter
or leave any initial segment of a total ordering of the vertices of $G$. The tree-zig-zag number
${tzn}(G)$ may be regarded as an analog of ${zn}(G)$ which quantifies the amount
of times a directed path is allowed to enter or leave any sub-tree of an olive-tree decomposition of $G$.
If $G$ is a digraph, we write $\mathit{dtw}(G)$ for its directed treewidth \cite{JohnsonRobertsonSeymourThomas2001},
$\mathit{Dw}(G)$ for
its {\em $D$-width} \cite{Gruber2012}, $\mathit{dagw}(G)$ for its DAG-width \cite{BerwangerDawarHunterKreutzerObdrzalek2012},
$\mathit{dpw}(G)$ for its directed path-width \cite{Barat2006}, $\mathit{Kelw}(G)$ for its Kelly-width
\cite{GanianHlinenyKneisLangerObdrzRossmanith2014},
$\mathit{ddp}(G)$ for its DAG-depth \cite{GanianHlinenyKneisLangerObdrzRossmanith2014},
$\mathit{Kw}(G)$ for its K-width \cite{GanianHlinenyKneisLangerObdrzRossmanith2014},
$s(G)$ for its weak separator number \cite{Gruber2012} and $\mathit{cr}(G)$ for its cycle rank \cite{Gruber2012}.
A dashed arrow $A\dashrightarrow B$ from measure $A$ to measure $B$ indicates that $A$ is
at least as expressive as $B$. More precisely, there exist constants $\alpha_1,\alpha_2\in \N$ such that
for every digraph $G$, $A(G)\leq \alpha_1\cdot B(G) + \alpha_2$. A full arrow $A\rightarrow B$ indicates that
the measure $A$ is strictly more expressive than measure $B$. More precisely, $A$ is at least as expressive as
$B$, and there exists an infinite class of digraphs in which $A$ is bounded by a constant, but
$B$ is unbounded.
\begin{equation}
\label{equation:ComparisonWidthMeasures}
\begin{diagram}
\mathbold{n}[3]{\mathit{zn}(G)}
\arrow{sse,t,1}{\mbox{\tiny{\cite{deOliveiraOliveira2013IPEC}}}}
\\
\mathbold{n}[3]{\mathit{Kelw}(G)}
\arrow{se,t,1}{\mbox{\tiny{\cite{GanianHlinenyKneisLangerObdrzRossmanith2014}}}}
\mathbold{n}[2]{\mathit{Kw}(G)}
\\
\mathbold{n}{\mathbold{{tzn}(G)}}
\arrow{e,t,..}{\mathit{Th.}\;\ref{theorem:ComparisonWithOtherMeasures}}
\arrow[2]{ne,..}
\mathbold{n}{\mathit{dtw}(G)}
\arrow{ne,t}{\mbox{\tiny{\cite{HunterKreutzer2008}}}}
\arrow{e,t}{\mbox{\tiny{\cite{BerwangerDawarHunterKreutzerObdrzalek2012}}}}
\arrow{se,t}{\mbox{\tiny{\cite{AmiriKaiserKreutzerRabinovichSiebertz2015}}}}
\mathbold{n}{\mathit{dagw}(G)}
\arrow{e,t}{\mbox{\tiny{\cite{BerwangerDawarHunterKreutzerObdrzalek2012}}}}
\mathbold{n}{\mathit{dpw}(G)}
\arrow{e,t}{\mbox{\tiny{\cite{Gruber2012}}}}
\mathbold{n}{\mathit{cr}(G)}
\arrow{n,r}{\mbox{\tiny{\cite{GanianHlinenyKneisMeisterObdrzalekRossmanithSikdar2010}}}}
\arrow{s,r}{\mbox{\tiny{\cite{GanianHlinenyKneisMeisterObdrzalekRossmanithSikdar2010}}}}
\\
\mathbold{n}{\frac{\mathit{cr}(G)}{\log |G|}}
\arrow{e,t,..}{\mbox{\tiny{\cite{Gruber2012}}}}
\mathbold{n}{s(G)}
\arrow{e,t,..}{\mbox{\tiny{\cite{Gruber2012}}}}
\mathbold{n}{\mathit{Dw}(G)} \arrow{ne,t}{\mbox{\tiny{\cite{Safari2005,deOliveiraOliveira2013IPEC}}}}
\mathbold{n}[2]{\mathit{ddp}(G)}
\end{diagram}
\end{equation}
The numbers above each arrow $A\rightarrow B$ ($A\dashrightarrow B$) in Equation \ref{equation:ComparisonWidthMeasures}
refer to the works in which the corresponding relation between the measures $A$ and $B$ was established.
All relations listed above can be inferred from the literature, except for the relation
$\mathbold{{tzn}}(G)\dashrightarrow \mathit{zn}(G)$, which is immediate, and the relation
$\mathbold{{tzn}(G)} \dashrightarrow dtw(G)$, which will be formally stated in
Theorem \ref{theorem:ComparisonWithOtherMeasures} below and proved in Section \ref{section:ComparisonDirectedTreewidth}.
The fact that DAG-width, Kelly-width and D-width are strictly more expressive than directed pathwidth follows
from the fact that the width of the complete undirected\footnote{In this setting each undirected edge is represented
by two directed edges in opposite directions.} binary tree on $n$ leaves is bounded with respect to these three
measures, but unbounded ($\Omega(\log n)$) with respect to directed pathwidth \cite{deOliveiraOliveira2013IPEC}.
It is worth noting that the precise statement of our main theorem (Theorem \ref{theorem:MainTheoremDirectedTreewidth})
holds if the parameter $w$ corresponds to the width of the digraph $G$ with respect to any measure reachable from
$\mathit{dtw}(G)$ in Equation \ref{equation:ComparisonWidthMeasures}. More Precisely,
Theorem \ref{theorem:MainTheoremDirectedTreewidth} also holds when $w$ is the Kelly width, DAG-width, D-width, directed pathwidth,
cycle rank, K-width or DAG-depth of $G$.
{Theorem \ref{theorem:MainTheoremDirectedTreewidth}} can also be applied if the parameter $w$ is the
tree-zig-zag number of $G$. However, in this particular case, an explicit olive-tree decomposition of $G$ of
tree-zig-zag number $O(w)$ must be given in the input. For directed tree-width and less expressive measures,
such an olive-tree decomposition of width $O(w)$ can be automatically constructed in time $|G|^{O(w)}$.
This construction will be carried in Section \ref{section:ComparisonDirectedTreewidth} together with the proof of
Theorem \ref{theorem:ComparisonWithOtherMeasures}.
\begin{theorem}
\label{theorem:ComparisonWithOtherMeasures}
Let $G$ be a digraph, ${tzn}(G)$ be its tree-zig-zag number and $dtw(G)$ be its directed treewidth.
Then ${tzn}(G)\leq 9\cdot dtw(G) + 18$.
\end{theorem}
\section{Tree-Zig-Zag Number vs Directed Treewidth}
\label{section:ComparisonDirectedTreewidth}
In this section we will prove Theorem \ref{theorem:ComparisonWithOtherMeasures}.
First we will state a couple of propositions concerning $Z$-normal sets.
\begin{proposition}
\label{proposition:NormalComplement}
Let $G=(V,E)$ be a digraph and $K,Z\subseteq V$ be such that $K$ is $Z$-normal. Then for each subset $X\subseteq K$,
$K\backslash X$ is $Z\cup X$-normal.
\end{proposition}
\begin{proof}
The proof is by contradiction.
Assume that there is an $X\subseteq K$ such that $K\backslash X$ is not $Z\cup X$-normal. Then there is a walk in $G\backslash (Z\cup X)$ that
starts and ends in $K\backslash X$, but that uses a vertex from $V\backslash ((Z\cup X)\cup (K\backslash X))= V\backslash (Z\cup K)$. This contradicts the
assumption that $K$ is $Z$-normal.
\end{proof}
If $G=(V,E)$ is a digraph, $Z$ is a subset of $V$ and $\mathfrak{p} = v_1e_1v_2....v_{n-1}e_{n-1}v_n = (V_{\mathfrak{p}},E_{\mathfrak{p}})$ is a path on $G$ then we say that
$\mathfrak{p}$ is internally disjoint from $Z$ if $Z\cap V_{\mathfrak{p}} \subseteq \{v_1,v_n\}$. In other words, $\mathfrak{p}$ is internally disjoint
from $Z$ if none of its internal vertices belongs to $Z$. The next proposition says that if $K$ is a $Z$-normal subset of $V$ and
$\mathfrak{p}$ is a path that is internally disjoint from $Z$, then $\mathfrak{p}$ can enter or leave $K$ at most $2$ times.
\begin{proposition}
\label{proposition:InternallyDisjointPath}
Let $G=(V,E)$ be a digraph and $K,Z \subseteq V$ be subsets of $V$ such that $K$ is $Z$-normal. Let
$\mathfrak{p} = (V_{\mathfrak{p}},E_{\mathfrak{p}})$ be a path in $G$ that is internally disjoint from $Z$. Then $$|E_{\mathfrak{p}} \cap E(K,V\backslash K)| \leq 2.$$
\end{proposition}
\begin{proof}
The proof is by contradiction. Assume that $|E_{\mathfrak{p}} \cap E(K,V\backslash K)| \geq 3$. Let $e_1$, $e_2$ and $e_3$ be the first three
edges of $\mathfrak{p}$ that have one endpoint in $K$ and other endpoint in $V\backslash K$. Then
$\mathfrak{p} = \mathfrak{p}_0 e_1 \mathfrak{p}_1 e_2 \mathfrak{p}_2 e_3 \mathfrak{p}_3$ where for each $i\in \{1,2,3\}$, the source of $e_i$ is
the last vertex of $\mathfrak{p}_{i-1}$ and the target of $e_i$ is the first vertex of $\mathfrak{p}_i$. Since the path $\mathfrak{p}$ is internally disjoint
from $Z$, we have that either $\mathfrak{p}_1$ is entirely contained in $K$ and $\mathfrak{p}_2$ is entirely contained in $V\backslash (K\cup Z)$, or
$\mathfrak{p}_1$ is entirely contained in $V\backslash (K\cup Z)$ and $\mathfrak{p}_2$ is entirely contained in $K$. Therefore
either $e_1\mathfrak{p}_1e_2$ or $e_2\mathfrak{p}_2e_3$ is a path that starts and finishes at $K$ and uses a vertex of $V\backslash (K\cup Z)$.
This contradicts the assumption that $K$ is $Z$-normal.
\end{proof}
The next proposition says that if $\mathfrak{p}$ is a path of $G$, then the number of edges of $\mathfrak{p}$ crossing
a $Z$-normal set is upper bounded by $2\cdot |Z|+2$.
\begin{proposition}
\label{proposition:NumberOfEdges}
Let $G=(V,E)$ be a digraph and $K,Z \subseteq V$ be subsets of $V$ such that $K$ is $Z$-normal. Then for each
path $\mathfrak{p} = (V_{\mathfrak{p}},E_{\mathfrak{p}})$ in $G$, $$|E_{\mathfrak{p}} \cap E(K,V\backslash K)| \leq 2\cdot |Z|+2.$$
\end{proposition}
\begin{proof}
Let $\mathfrak{p}=(V_{\mathfrak{p}},E_{\mathfrak{p}})$ be a path in $G$ and assume that $V_{\mathfrak{p}}\cap Z = \{v_1,...,v_k\}$. We may assume without
loss of generality that for each $i\in \{1,...,k-1\}$, $v_i$ occurs before $v_{i+1}$ in $\mathfrak{p}$. In other words we may assume that
$\mathfrak{p} = \mathfrak{p}_0 \cup \mathfrak{p}_1 \cup ... \cup \mathfrak{p}_k$ where $\mathfrak{p}_0,\mathfrak{p}_1,...,\mathfrak{p}_k$ are internally disjoint paths in
which for each $i\in \{1,...,k\}$, $v_i$ is the last vertex of $\mathfrak{p}_{i-1}$ and the first vertex of $\mathfrak{p}_i$.
We note that for each $i\in \{1,...,k\}$ the path $\mathfrak{p}_i=(V_{\mathfrak{p}_i},E_{\mathfrak{p}_i})$ is internally disjoint from $Z$. Therefore, from Proposition
\ref{proposition:InternallyDisjointPath} we have that $|E_{\mathfrak{p}_i} \cap E(K,V\backslash K)| \leq 2$.
This implies that $|E_{\mathfrak{p}} \cap E(K,V\backslash K)| \leq \sum_{i=0}^k |E_{\mathfrak{p}_i}\cap E(K,V\backslash K)| \leq 2k+2 \leq 2|Z|+2$.
\end{proof}
The next proposition says that if $G$ is a digraph and $K_1$ and $K_2$ are disjoint subsets of vertices of $G$ such that
no edge has source in $K_2$ and target in $K_1$, then any path crossing $K_1\cup K_2$ at most $2$ times, crosses
$K_2$ at most $3$ times.
\begin{proposition}
\label{proposition:UnionSetsPath}
Let $G=(V,E)$ be a digraph and $K_1,K_2$ be subsets of vertices of $G$ such that
$K_1\cap K_2 = \emptyset$ and such that there is no edge with source in $K_2$ and
target in $K_1$. Let $\mathfrak{p}=(V_{\mathfrak{p}},E_{\mathfrak{p}})$ be a path in $G$ such that
$|E_{\mathfrak{p}}\cap E(K_1\cup K_2, V\backslash (K_1\cup K_2))| \leq 2$. Then $|E_{\mathfrak{p}}\cap E(K_2,V\backslash K_2)| \leq 3$.
\end{proposition}
\begin{proof}
Let $\mathfrak{p}=(V_{\mathfrak{p}},E_{\mathfrak{p}})$ be a path in $G$.
If $|E_{\mathfrak{p}}\cap E(K_1\cup K_2, V\backslash (K_1\cup K_2))|=0$ then $\mathfrak{p}$ is either entirely contained in $K_1\cup K_2$
or entirely contained in $V\backslash (K_1\cup K_2)$ and the proposition holds trivially. Now let $|E_{\mathfrak{p}}\cap E(K_1\cup K_2, V\backslash (K_1\cup K_2))|=1$
and let $e_1$ be the unique edge with one endpoint in $K_1\cup K_2$ and another endpoint in $V\backslash (K_1\cup K_2)$.
Then $\mathfrak{p} = \mathfrak{p}_0e_1\mathfrak{p}_1$ where the source of $e_1$ is the last vertex of $\mathfrak{p}_0$ and the target of $e_1$ is the first vertex
of $\mathfrak{p}_1$. Note that for each $i\in \{0,1\}$, either $\mathfrak{p}_i$ is entirely contained in $K_1\cup K_2$ or entirely contained
in $V\backslash (K_1 \cup K_2)$. Since there is no edge with source in $K_2$ and target in $K_1$, we have that $\mathfrak{p}_0$ and
$\mathfrak{p}_1$ can each cross $K_2$ at most one time. In other words, $|E_{\mathfrak{p}_i}\cap E(K_2,V\backslash K_2)| \leq 1$.
Therefore $\mathfrak{p}_0$, $e_1$ and $\mathfrak{p}_1$ together cross $K_2$ at most three times and the proposition holds in this case.
Finally, let $|E_{\mathfrak{p}}\cap E(K_1\cup K_2, V\backslash (K_1\cup K_2))|=2$.
Let $e_1$ and $e_2$ be the only edges of $\mathfrak{p}$ with one endpoint in $K_1\cup K_2$ and the other endpoint in $V\backslash (K_1\cup K_2)$,
and assume that $e_1$ is visited before $e_2$. Then there are paths $\mathfrak{p}_1,\mathfrak{p}_2,\mathfrak{p}_3$ such that $\mathfrak{p} = \mathfrak{p}_0e_1\mathfrak{p}_1e_2\mathfrak{p}_2$
and for $i\in \{1,2\}$, the source of $e_i$ is the last vertex of $\mathfrak{p}_{i-1}$ and the target of $e_i$ is the source of $\mathfrak{p}_i$.
Note that for $i\in \{0,1,2\}$, $\mathfrak{p}_i$ is either entirely contained in $K_1\cup K_2$ or entirely contained in $V\backslash (K_1\cup K_2)$. Note also
that each $\mathfrak{p}_i$ crosses $K_2$ at most one time, since there is no edge with source in $K_2$ and target in $K_1$.
This already implies that $\mathfrak{p}$ can cross $K_2$ at most $5$ times. We claim that with further analysis it can be shown that the number of
crossings is at most $3$, which is optimal. The analysis is as follows. If the source of $e_1$ belongs to $V\backslash (K_1\cup K_2)$, then the
target of $e_2$ is also in $V\backslash (K_1\cup K_2)$. In this case
both $\mathfrak{p}_0$ and $\mathfrak{p}_2$ are entirely contained in $V\backslash (K_1\cup K_2)$, and therefore only $e_1,e_2$ and $\mathfrak{p}_1$ have
the possibility of crossing $K_2$. If the source of $e_1$ belongs
to $K_1\cup K_2$, then the target of $e_2$ also belongs to $K_1\cup K_2$. This implies that $\mathfrak{p}_1$ is entirely contained in
$V\backslash (K_1\cup K_2)$. In this situation there are two sub-cases to be analysed.
If the target of $e_2$ belongs to $K_2$ then $\mathfrak{p}_2$ is entirely contained in $K_2$ and only $e_1,e_2$ and $\mathfrak{p}_0$ have the
possibility of crossing $K_2$. On the other hand, if the target of $e_2$ is in $K_1$ then $e_2$ does not cross $K_2$ and thus only
$e_1$, $\mathfrak{p}_0$ and $\mathfrak{p}_2$ have the possibility to cross $K_2$.
\end{proof}
Using Proposition \ref{proposition:UnionSetsPath} we can show that if $K_1$ and $K_2$ are disjoint subsets of vertices
of a digraph $G$ such that $K_1\cup K_2$ is a $Z$-normal and such that there is no edge with source in $K_2$ and target
in $K_1$, then each path in $G$ crosses $K_2$ at most $3|Z|+3$ times.
\begin{proposition}
\label{proposition:UnionSets}
Let $G=(V,E)$ be a digraph and $K_1,K_2,Z \subseteq V$ be subsets of vertices of $G$ such that
$K_1\cup K_2$ is $Z$-normal, $K_1\cap K_2 =\emptyset$, and such that there is no edge with source in $K_2$ and target in $K_1$.
Then for each path $\mathfrak{p} = (V_{\mathfrak{p}},E_{\mathfrak{p}})$ in $G$ we have that $$|E_{\mathfrak{p}}\cap E(K_2,V\backslash K_2)| \leq 3\cdot |Z| + 3.$$
\end{proposition}
\begin{proof}
Analogously to the proof of Proposition \ref{proposition:NumberOfEdges} we let $V_{\mathfrak{p}}\cap Z = \{v_1,...,v_k\}$, and assume that
$\mathfrak{p} = \mathfrak{p}_0\cup \mathfrak{p}_1 \cup ... \cup \mathfrak{p}_k$ where $\mathfrak{p}_1,\mathfrak{p}_2,...,\mathfrak{p}_k$ are internally disjoint paths such that
for each $i\in \{1,...,k\}$, $v_i$ is the last vertex of $\mathfrak{p}_{i-1}$ and the first vertex of $\mathfrak{p}_i$. We note that
for each $i\in \{1,...,k\}$ the path $\mathfrak{p}_i$ is internally disjoint from $Z$. Therefore, since $K_1\cup K_2$ is $Z$-normal,
by Proposition \ref{proposition:InternallyDisjointPath} we have that
$|E_{\mathfrak{p}_i}\cap E(K_1\cup K_2, V\backslash(K_1\cup K_2))| \leq 2$.
Now, since there is no edge with source in $K_2$ and target in $K_1$, we can apply Proposition \ref{proposition:UnionSetsPath}
to infer that $|E_{\mathfrak{p}_i}\cap E(K_2,V\backslash K_2)|\leq 3$.
This implies that $|E_{\mathfrak{p}} \cap E(K_2,V\backslash K_2)| \leq \sum_{i=0}^k |E_{\mathfrak{p}_i}\cap E(K_2,V\backslash K_2)| \leq 3k+3 \leq 3|Z|+3$.
\end{proof}
The main technical lemma of this section states that each good arboreal decomposition $\mathcal{D}$ of width
$w$ can be transformed into an olive-tree decomposition $\mathcal{T}$ of tree-zig-zag number at most $3w+6$.
\begin{lemma}
\label{lemma:FromArborealPredecompositionToOliveTreeDecomposition}
Let $G=(V,E)$ be a digraph and $\mathcal{D} = (N,F,W,Z)$ be a good arboreal decomposition of $G$ of width
$w$. One can construct in time $O(w\cdot |N|)$ an olive-tree decomposition $\mathcal{T} = (N',F',\mathfrak{m})$ of
$G$ of tree-zig-zag number at most $3w+6$.
\end{lemma}
\begin{proof}
We start by defining the sets of nodes and arcs of the olive-tree decomposition $\mathcal{T}$. Intuitively, $\mathcal{T}$
is obtained by replacing each node $p$ of $\mathcal{D}$, labeled with a bag $W(p)$ and having $r$ children, with a line
$L_p \equiv a_p^0a_p^1...a_p^{|W(p)|}b_p^1b_p^2...b_p^{r}$ as depicted in Figure \ref{figure:conversionDirectedTreewidthCarving}.
\begin{figure}[!hf]
\centering
\includegraphics[scale=0.25]{arborealDecompositionToOliveTreeDecomposition}
\caption{From a good arboreal decomposition of width $w$ to an olive-tree decomposition of tree-zig-zag number at most $3w+6$. }
\label{figure:conversionDirectedTreewidthCarving}
\end{figure}
Each vertex in $W(p)$ is mapped by $\mathfrak{m}$ to a node in $\{a_p^1,...,a_p^{|W(p)|}\}$ in such a way that no two distinct vertices
in $W(p)$ are mapped to the same node. No vertex of $G$ is mapped to the node
$a_p^0$ nor to the nodes $b_p^1,b_p^2,...,b_p^r$. These nodes are used to connect the line $L_p$ corresponding to the node $p$ of $\mathcal{D}$
to lines corresponding to other positions. In particular for each $j\in \{1,...,r\}$, $b_p^j$ is connected to $a^0_{pj}$. In other words, $b_p^j$ is connected
to the first vertex of the line $L_{pj}$. We will show below that the olive-tree decomposition defined in this way has width
at most $3w+6$. First however we formally define the sets $N'$ and $F'$ and the mapping function $\mathfrak{m}$.
\begin{equation}
\label{equation:setOfNodes}
N' = \{a_p^{i}\;|\;p\in N, 0\leq i \leq |W(p)| \} \cup \{b_p^j \;|\; (p,pj)\in F \}
\end{equation}
\begin{equation}
\label{equation:setofArcs}
\begin{array}{l}
F' = \{(a_{p}^i,a_p^{i+1})\;|\;p\in N,\;0\leq i \leq |W(p)|-1 \}\; \cup\; \{ (a_p^{|W(p)|}, b_{p}^1)\;|\; p\in N \} \; \cup \\
\\
\hspace{1cm} \{(b_p^i,b_p^{i+1})\;|\; p\in N, 1\leq i\leq r-1\} \;\cup\; \{(b_p^{j}, a_{pj}^0)\;|\; (p,pj) \in F\}
\end{array}
\end{equation}
Finally, the labeling function $\mathfrak{m}:V\rightarrow N'$ is chosen arbitrarily as long as it satisfies the following
condition for each node $p\in N$.
\begin{equation}
\label{equation:MapBags}
\mathfrak{m}(W(p))= \{a_p^1,...,a_p^{|W(p)|}\}.
\end{equation}
In other words we choose $\mathfrak{m}$ in such a way that for each position $p\in N$, the vertices in $W(p)$ are
bijectively mapped to the nodes in $\{a_p^1, ...,a_p^{|W(p)|}\}$.
We argue that $\mathcal{T}$ is indeed an olive-tree decomposition of
tree-zig-zag number at most $3w+6$.
Recall that if $G=(V,E)$ is a digraph and $\mathcal{D}$ is an arboreal decomposition of $G$, then
for each node $p$ of $\mathcal{D}$, $V(p,\mathcal{D})$ denotes the set of
vertices of $G$ that belong to some bag associated with a descendant of $p$ (including $p$ itself).
Analogously, if $\mathcal{T}$ is an olive-tree decomposition of $G$ then for each node $u$ of $\mathcal{T}$,
$V(u,\mathcal{T})$ denotes the set of vertices of $G$ that are mapped to some descendant of $u$.
To show that $\mathcal{T}$ has tree-zig-zag number at most $3w+6$ we
need to show that for each node $u\in N'$, and each path $\mathfrak{p}$ in $G$,
$$| E_{\mathfrak{p}}\cap E(V(u,\mathcal{T}),V\backslash V(u,\mathcal{T}))|\leq 3w+6.$$
There are two cases to be considered, depending on whether $u=a_p^i$ or whether $u=b_p^j$. We analyse each of these cases below.
\begin{enumerate}
\item ($u=a_p^i$) We start by noting that for each node $p\in N$, $V(p,\mathcal{D}) = V(a_p^0,\mathcal{T})$.
If $p$ is the root of $\mathcal{D}$ then $V(p,\mathcal{D}) = V$ and thus both $V(p,\mathcal{D})$
and $V(a_{p}^0,\mathcal{T})$ are $\emptyset$-normal.
If $p$ is not the root of $\mathcal{D}$ then $p$ has a parent $p'$. In this case by definition of arboreal decomposition
we have that $V(p,\mathcal{D})$ is $Z(p',p)$-normal. Thus $V(a_p^0,\mathcal{T})$ is also $Z(p',p)$-normal.
We let $X_p$ be equal to $\emptyset$ if $p$ is the root of $\mathcal{D}$, and equal to $Z(p',p)$ if $p'$ is the parent of $p$.
Thus we can simply say that $V(p,\mathcal{D})$ is $X_p$-normal.
Now let $j\in \{1,...,|W(p)|\}$. Then
$V(a_p^j,\mathcal{T}) = V(a_p^0,\mathcal{T}) \backslash \mathfrak{m}^{-1}(\{a_p^1,...,a_p^{j-1}\})$. In other words,
$V(a_p^j,\mathcal{T})$ is equal to $V(a_p^0,\mathcal{T})$ minus the vertices of $G$ that are mapped by
$\mathfrak{m}$ to some node in $\{a_p^1,...,a_p^{j-1}\}$. This implies, by
Proposition \ref{proposition:NormalComplement} that the set $V(a_p^j,\mathcal{T})$ is
$X_p \cup \mathfrak{m}^{-1}(\{a_p^1,...,a_p^{j-1}\})$-normal.
Since by the construction of $\mathcal{T}$, $\mathfrak{m}^{-1}(\{a_p^1,...,a_p^{j-1}\}) \subseteq W(p)$
(Equation \ref{equation:MapBags}), and since $\mathcal{D}$ has width $w$, we have that
$|X_p \cup \mathfrak{m}^{-1}(\{a_p^1,...,a_p^{j-1}\})| \leq w+1$.
Therefore, by Proposition \ref{proposition:NumberOfEdges}, for each $j\in \{0,...,|W(p)|\}$,
$$|E_{\mathfrak{p}}\cap E(V(a_p^j,\mathcal{T}),V\backslash V(a_p^j,\mathcal{T}))| \leq 2(w+1)+2 \leq 3w +6.$$
\item ($u=b_p^j$) Let $u=b_p^j$ for some $p\in N$ and $j\in \{1,...,r\}$. Since by the construction of $\mathcal{T}$,
no vertex of $G$ is mapped to $b_p^j$, we
have that $V(b_{p}^j,\mathcal{T}) = \bigcup_{k=j}^r V(a_{pk}^0,\mathcal{T})$. Additionally, since
$V(a_p^0,\mathcal{T}) = V(p,\mathcal{D})$ for each $p\in N$, we have that
$V(b_{p}^j,\mathcal{T}) = \bigcup_{k=j}^r V(pk,\mathcal{D})$.
Note that $V(b_p^j,\mathcal{T})\subseteq V(b_p^1,\mathcal{T})$ for each $j\in \{1,...,r\}$.
Since $\mathcal{D}$ is a good arboreal decomposition, we have that for $i,j\in \{1,...,r\}$ with $i<j$
there is no edge with source in $V(pj,\mathcal{D})$ and target in $V(pi,\mathcal{D})$.
This implies that for each $j\in \{1,...,r\}$, there is no edge with source in $V(b_{p}^j,\mathcal{T})$ and
target in $V(b_p^1,\mathcal{T})\backslash V(b_{p}^j,\mathcal{T})$. Now let $X_p$ be equal
to $\emptyset$ if $p$ is the root of $\mathcal{D}$, and equal to $Z(p',p)$ if $p'$ is the parent of $p$.
We note that $V(a_p^0,\mathcal{T})=V(p,\mathcal{D})$ is $X_p$-normal, and that
$V(b_p^1,\mathcal{T})= V(a_p^0,\mathcal{T}) \backslash W(p)$. Therefore, by Proposition \ref{proposition:NormalComplement},
$V(b_p^1,\mathcal{T})$ is $X_p\cup W(p)$-normal.
Now, we can apply Proposition \ref{proposition:UnionSets} with
$K_1 = V(b_p^1,\mathcal{T})\backslash V(b_p^j,\mathcal{T})$, $K_2= V(b_p^j,\mathcal{T})$,
and $Z=X_p \cup W(p)$ to infer that
$$|E_{\mathfrak{p}}\cap E(V(b_p^j,\mathcal{T}), V\backslash V(b_p^j,\mathcal{T}))|\leq 3|Z|+3 \leq 3(w+1) + 3 = 3w+6.$$
The inequality $3|Z|+3\leq 3(w+1)+3$ follows from the fact that $|Z|=|X_p\cup W(p)|\leq w+1$, since $\mathcal{D}$ has width $w$.
\end{enumerate}
\end{proof}
Finally we are in a position to prove Theorem \ref{theorem:ComparisonWithOtherMeasures}.
\paragraph{\bf Proof of Theorem \ref{theorem:ComparisonWithOtherMeasures}}
By Lemma \ref{theorem:ConstructionGoodArborealDecomposition}, given a digraph
$G$ of directed treewidth $w$ one can construct in time $|G|^{O(w)}$ a good arboreal
decomposition $\mathcal{D}$ of $G$ of width at most $3w+4$. By Lemma \ref{lemma:FromArborealPredecompositionToOliveTreeDecomposition} one can transform
$\mathcal{D}$ into an olive-tree decomposition $\mathcal{T}$ of $G$ of tree-zig-zag number at most $3(3w+4)+6 = 9w+18$.
$\square$
\section{Tree Automata}
\label{section:TreeAutomata}
In this section we recall some of the main concepts of tree-automata theory. For an extensive
treatment of the subject we refer the reader to the standard reference \cite{Tata2007}. As two
non-standard applications, we consider the problem of counting the number of terms of depth
$d$ accepted by a deterministic tree-automaton, and the problem of generating terms having a prescribed weight.
A ranked alphabet is a finite set $\mathbold{\Sigma} = \mathbold{\Sigma}_0 \cup \mathbold{\Sigma}_1 \cup ... \cup \mathbold{\Sigma}_r$ of
function symbols where the elements of $\mathbold{\Sigma}_i$ are function symbols of arity $i$.
Intuitively, the arity of a function symbol specifies its number of inputs.
Constants are regarded as function symbols of arity $0$. If $f$ is a function symbol in $\mathbold{\Sigma}$
then we let $\mathfrak{a}(f)$ denote the arity of $f$. In other words $\mathfrak{a}(f)=i$ if and only if $f\in \mathbold{\Sigma}_i$.
The set $\mathit{Ter}(\mathbold{\Sigma})$ of all terms\footnote{In this work we will {\em not} be interested in terms containing variables.
In other words, all terms considered here are ground terms.} over $\mathbold{\Sigma}$ is inductively defined as follows:
\begin{itemize}
\item if $f\in \mathbold{\Sigma}_0$ then $f$ is a term in $\mathit{Ter}(\mathbold{\Sigma})$,
\item if $f\in \mathbold{\Sigma}_{\mathfrak{a}(f)}$ and $t_1,...,t_{\mathfrak{a}(f)}$ are terms in $\mathit{Ter}(\mathbold{\Sigma})$
then $f(t_1,t_2,...,t_{\mathfrak{a}(f)})$ is a term in $\mathit{Ter}(\mathbold{\Sigma})$.
\end{itemize}
Let $t=f(t_1,...,t_{\mathfrak{a}(f)})$ be a term over the ranked alphabet $\mathbold{\Sigma}$. Then we define $\mathbold{ls}(t)=f$ as the leading
symbol of $t$. We denote by $\mathit{Pos}(t)$ the set of positions of $t$, which is a prefix closed subset of $\{1,...,r\}^*$ used to index the subterms of $t$.
More precisely, if $t=f(t_1,...,t_{\mathfrak{a}(f)})$ then
$$\mathit{Pos}(t)=\{\lambda\} \cup \bigcup_{ j \in \{1,...,\mathfrak{a}(f)\}} \{ j p\;|\; p\in \mathit{Pos}(t_{ j })\}.$$
We note that if $t$ is a constant, i.e., a function symbol of arity $0$, then $\mathit{Pos}(t)=\{\lambda\}$.
If $t\in \mathit{Ter}(\mathbold{\Sigma})$ then we let $|t|=|\mathit{Pos}(t)|$.
The subterm $t|_p$ of $t$ at position $p$ is
inductively defined as follows: $t|_{{\lambda}} = t$; if $t=f(t_1,t_2,...,t_{\mathfrak{a}(f)})$, then for each $ j \in [\mathfrak{a}(f)]$ and each position
$j p\in \mathit{Pos}(t)$, $t|_{ j p} = t_{ j }|_{p}$.
If $t$ is a term and $p\in \mathit{Pos}(t)$ then $t[p] = \mathbold{ls}(t|_p)$ denotes the leading symbol of the subterm of $t$ at position $p$.
A tree-language over a ranked alphabet $\mathbold{\Sigma}$ is any subset $\lang\subseteq \mathit{Ter}(\mathbold{\Sigma})$. In particular
the empty set $\emptyset$ is a tree-language. A {\em bottom-up tree-automaton} over $\mathbold{\Sigma}$ is a tuple ${\mathcal{A}}=(Q,\mathbold{\Sigma},Q_{F},\Delta)$
where $Q$ is a set of states, $Q_{F}\subseteq Q$ a set of final states and $\Delta=\Delta_0 \cup \Delta_1 \cup ....\cup \Delta_r$ is a transition relation
where $\Delta_0\subseteq \Sigma_0\times Q$ and $\Delta_i\subseteq Q^i\times \mathbold{\Sigma}_i \times Q$ for each $i\in \{1,...,r\}$.
The size of ${\mathcal{A}}$, which is defined as $|{\mathcal{A}}|=|Q|+|\Delta|$, measures the number of states in $Q$ plus the number of transitions in $\Delta$.
The set $\lang({\mathcal{A}},{\mathfrak{q}},i)$ of all terms reaching a state ${\mathfrak{q}}\in Q$ in depth at most $i$ is inductively defined as follows.
\vspace{-2pt}
\begin{equation}
\label{equation:InductiveDefinition}
\lang({\mathcal{A}},{\mathfrak{q}},1) = \{a\in \mathbold{\Sigma}_0\; |\; (a,{\mathfrak{q}})\in \Delta_0\}
\end{equation}
\begin{equation*}
\begin{array}{l}
\lang({\mathcal{A}},{\mathfrak{q}},i) = \lang({\mathcal{A}},{\mathfrak{q}},i-1)\; \cup\; \{f(t_1,...,t_{\mathfrak{a}(f)})\;|\; ({\mathfrak{q}}_1,...,{\mathfrak{q}}_{\mathfrak{a}(f)},f,{\mathfrak{q}})\in \Delta_{\mathfrak{a}(f)}, \\
\hspace{8.5cm} t_j\in \lang({\mathcal{A}},{\mathfrak{q}}_j,i-1)\}
\end{array}
\end{equation*}
We denote by $\lang({\mathcal{A}})$ the set of all terms reaching a final state in $Q_{F}$ at any finite depth.
\begin{equation}
\lang({\mathcal{A}}) =\bigcup_{{{\mathfrak{q}}\in Q_{F},i\in \N}} \lang({\mathcal{A}},{\mathfrak{q}},i)
\end{equation}
We say that the set $\lang({\mathcal{A}})$ is the language generated by ${\mathcal{A}}$.
Let ${\mathcal{A}}=(Q,\mathbold{\Sigma},Q_{F},\Delta)$ be a tree-automaton. We say that ${\mathcal{A}}$ is
{\em deterministic} if for every function symbol $f\in \mathbold{\Sigma}$ and every tuple $({\mathfrak{q}}_1,...,{\mathfrak{q}}_{\mathfrak{a}(f)})\in Q^{\mathfrak{a}(f)}$
there exists at most one ${\mathfrak{q}}\in Q$ such that $({\mathfrak{q}}_1,...,{\mathfrak{q}}_{\mathfrak{a}(f)},f,{\mathfrak{q}})\in \Delta_{\mathfrak{a}(f)}$. We say that
${\mathcal{A}}$ is complete if for every function symbol $f$ and every tuple $({\mathfrak{q}}_1,...,{\mathfrak{q}}_{\mathfrak{a}(f)}) \in Q^{\mathfrak{a}(f)}$
there exists at least one ${\mathfrak{q}}\in Q$ for which $({\mathfrak{q}}_1,...,{\mathfrak{q}}_{\mathfrak{a}(f)},f,{\mathfrak{q}})\in \Delta_{\mathfrak{a}(f)}$. Observe that
from any tree-automaton ${\mathcal{A}}$ one can derive a complete tree-automaton ${\mathcal{A}}'$ generating the same
language by adding a dead state ${\mathfrak{q}}_{\mathit{dead}}$, and creating a transition $({\mathfrak{q}}_1,...,{\mathfrak{q}}_{\mathfrak{a}(f)},f,{\mathfrak{q}}_{\mathit{dead}})$
whenever there is no transition in ${\mathcal{A}}$ whose left side is $({\mathfrak{q}}_1,...,{\mathfrak{q}}_{\mathfrak{a}(f)},f)$.
If $t$ is a term in $\mathit{Ter}(\mathbold{\Sigma})$, then the depth of $t$ is defined as $\max\{|p|: p\in \mathit{Pos}(t)\}$. In other
words, the depth of a term $t$ is the size of the longest path from the root of $t$ to one of its leaves. We denote
by $\mathit{depth}(t)$ the depth of $t$.
If ${\mathcal{A}}$ is a tree-automaton and $t\in \lang({\mathcal{A}})$ is a term of depth $d$, then we say that
${\mathcal{A}}$ accepts $t$ in depth $d$.
The next lemma says that for any deterministic tree-automaton
${\mathcal{A}}$ and any $d\in \N$, one can count in polynomial time the number of terms accepted by ${\mathcal{A}}$
in depth at most $d$.
\begin{lemma}
\label{lemma:CountingTrees}
Let ${\mathcal{A}}$ be a deterministic tree-automaton and let $d\in \N$. One can
count in time $d^{O(1)}\cdot|{\mathcal{A}}|^{O(1)}$ the number of terms accepted by ${\mathcal{A}}$ in
depth at most $d$.
\end{lemma}
\begin{proof}
The proof follows by a standard dynamic programming argument.
First we write a recursive formula that counts the number of terms that reach a given state ${\mathfrak{q}}$ in depth $i$:
\begin{equation}
\label{equation:CountingAcceptedTermsOne}
|\lang({\mathcal{A}},{\mathfrak{q}},1)|= |\{(f,{\mathfrak{q}})\;|\; f\in \mathbold{\Sigma}_0,\;{\mathfrak{q}} \in Q\}|
\end{equation}
\begin{equation}
\label{equation:CountingAcceptedTermsTwo}
|\lang({\mathcal{A}},{\mathfrak{q}},i)|= \sum_{({\mathfrak{q}}_1,...,{\mathfrak{q}}_{\mathfrak{a}(f)},f,{\mathfrak{q}}) \in \Delta} \;\; \prod_{j=1}^{\mathfrak{a}(f)} |\lang({\mathcal{A}},{\mathfrak{q}}_j,i-1)|
\end{equation}
Now the number of terms accepted by ${\mathcal{A}}$ in depth $d$ is the number of terms that reach a final state at depth
$d$.
\begin{equation}
\label{equation:FinalCounting}
|\lang({\mathcal{A}})| = \sum_{{\mathfrak{q}}\in Q_{F}, i\leq d} |\lang({\mathcal{A}},{\mathfrak{q}}, i)|.
\end{equation}
Thus to determine $|\lang({\mathcal{A}})|$, one can use Equation \ref{equation:CountingAcceptedTermsOne} to
compute and store in memory the value $|\lang({\mathcal{A}},{\mathfrak{q}},1)|$ for each ${\mathfrak{q}}\in Q$. Subsequently, using the values
stored in memory, one can use Equation \ref{equation:CountingAcceptedTermsTwo} to compute and store in memory the values $|\lang({\mathcal{A}},{\mathfrak{q}},2)|$ and so on. We repeat this process
until we have computed all values $|\lang({\mathcal{A}},{\mathfrak{q}},d)|$. At this point, we apply Equation \ref{equation:FinalCounting}
to determine the number of terms that reach a final term in depth at most $d$. Since in this work, $\mathfrak{a}(f)$ is
bounded by a constant (indeed in our applications $\mathfrak{a}(f)\leq 2$), this counting process can clearly be done in time
$d^{O(1)}\cdot |{\mathcal{A}}|^{O(1)}$.
\end{proof}
\subsection{Properties of Tree-Automata}
\label{subsection:PropertiesOfTreeAutomata}
If $\lang$ is a tree language over
a ranked alphabet $\mathbold{\Sigma}$ then the complement of $\lang$ is defined as $\overline{\lang} = \mathit{Ter}(\mathbold{\Sigma})\backslash \lang$.
A projection between ranked alphabets $\mathbold{\Sigma}$ and $\mathbold{\Sigma}'$ is any arity preserving total mapping
$\mathbold{\pi}:\mathbold{\Sigma}\rightarrow \mathbold{\Sigma}'$. By arity preserving we mean that if $f$ is a function symbol of arity $r$ in
$\mathbold{\Sigma}$, then $\mathbold{\pi}(f)$ is a function symbol of arity $r$ in $\mathbold{\Sigma}'$.
Recall that if $t$ is a term and $p\in \mathit{Pos}(t)$ then $t[p] = \mathbold{ls}(t|_p)$ denotes the leading symbol of the subterm of $t$ rooted at position $p$.
A projection $\mathbold{\pi}:\mathbold{\Sigma}\rightarrow \mathbold{\Sigma}'$ can be homomorphically extended to a mapping $\mathbold{\pi}:\mathit{Ter}(\mathbold{\Sigma})\rightarrow \mathit{Ter}(\mathbold{\Sigma}')$ between terms
by setting $\mathbold{\pi}(t)[p]=\mathbold{\pi}(t[p])$ for each position $p\in \mathit{Pos}(t)$.
Additionally, such mapping $\mathbold{\pi}$ can
be further extended to tree languages $\lang\subseteq \mathit{Ter}(\mathbold{\Sigma})$ by setting $\mathbold{\pi}(\lang) = \{\mathbold{\pi}(t)\;|\;t\in \mathit{Ter}(\mathbold{\Sigma})\}$.
Finally, given a projection $\mathbold{\pi}:\mathbold{\Sigma}\rightarrow \mathbold{\Sigma}'$ and a tree language $\lang$ over $\mathbold{\Sigma}'$,
the inverse homomorphic image of $\lang$ under $\mathbold{\pi}$ is defined as $\mathbold{\pi}^{-1}(\lang)=\{t\in \mathit{Ter}(\mathbold{\Sigma})\;|\;\mathbold{\pi}(t)\in \lang \}$,
i.e., the set of all terms over $\mathit{Ter}(\mathbold{\Sigma})$ which are mapped to some term in $\lang$.
In Lemma \ref{lemma:PropertiesOfTreeAutomata} below we list several well known closure properties
of tree languages recognizable by tree-automata (see for instance \cite{Tata2007}).
\begin{lemma}[Properties of Tree Automata]
\label{lemma:PropertiesOfTreeAutomata}
Let ${\mathcal{A}}$ be an arbitrary tree-automaton over a ranked alphabet $\mathbold{\Sigma}$ and let ${\mathcal{A}}_1$ and ${\mathcal{A}}_2$ be
deterministic complete tree-automata over $\mathbold{\Sigma}$.
\begin{enumerate}[(i)]
\item \label{lemma:PropertiesOfTreeAutomata:Deterministic}
There exists a unique minimal deterministic complete tree-automaton $\mathit{det}({\mathcal{A}})$
such that $\lang(\mathit{det}({\mathcal{A}})) = \lang({\mathcal{A}})$. Additionally,
$\mathit{det}({\mathcal{A}})$ can be constructed in time $2^{O(|{\mathcal{A}}|)}$.
\item \label{lemma:PropertiesOfTreeAutomata:Complement}
One can construct in time $O(|{\mathcal{A}}_1|)$ a
deterministic complete tree-automaton $\overline{{\mathcal{A}}_1}$ such that
$\lang({\mathcal{A}}_1)= \lang(\overline{{\mathcal{A}}_1})$.
\item \label{lemma:PropertiesOfTreeAutomata:UnionIntersection}
One can construct in time $O(|{\mathcal{A}}_1|\cdot |{\mathcal{A}}_2|)$ deterministic complete
tree-automata ${\mathcal{A}}_1\cup {\mathcal{A}}_2$ and ${\mathcal{A}}_1\cap {\mathcal{A}}_2$ such that
$\lang({\mathcal{A}}_1\cup {\mathcal{A}}_2) = \lang({\mathcal{A}}_1) \cup \lang({\mathcal{A}}_2)$ and
$\lang({\mathcal{A}}_1\cap {\mathcal{A}}_2) = \lang({\mathcal{A}}_1) \cap \lang({\mathcal{A}}_2)$.
\item \label{lemma:PropertiesOfTreeAutomata:TermRecognition}
Let $t\in \mathit{Ter}(\mathbold{\Sigma})$ be a term over $\mathbold{\Sigma}$.
Then one may determine in time $O(|{\mathcal{A}}|\cdot |t|)$ whether $t\in \lang({\mathcal{A}})$.
\item \label{lemma:PropertiesOfTreeAutomata:Projection}
Let $\mathbold{\pi}:\mathbold{\Sigma}\rightarrow \mathbold{\Sigma}'$ be a projection. Then one can construct in
time $O(|{\mathcal{A}}|)$ a tree-automaton $\mathbold{\pi}({\mathcal{A}})$
over $\mathbold{\Sigma}'$ such that $\lang(\mathbold{\pi}({\mathcal{A}})) = \mathbold{\pi}(\lang({\mathcal{A}}))$.
\item \label{lemma:PropertiesOfTreeAutomata:InverseHomomorphism}
Let $\mathbold{\pi}:\mathbold{\Sigma}'\rightarrow \mathbold{\Sigma}$
be a projection. One can construct in time $O(|\mathbold{\Sigma}'|\cdot |{\mathcal{A}}|)$ a tree-automaton $\mathbold{\pi}^{-1}({\mathcal{A}})$
over $\mathbold{\Sigma}'$ such that $\lang(\mathbold{\pi}^{-1}({\mathcal{A}})) = \mathbold{\pi}^{-1}(\lang({\mathcal{A}}))$. Additionally,
if ${\mathcal{A}}$ is deterministic, then $\mathbold{\pi}^{-1}({\mathcal{A}})$ is also deterministic.
\end{enumerate}
\end{lemma}
\subsection{Weighted Terms}
\label{subsection:WeightedTerms}
Let $\mathbold{\Sigma}$ be a ranked alphabet, $\Xi$ be a finite semigroup, and
let $\automataweightingfunction:\mathbold{\Sigma} \rightarrow \Xi$ be a function that associates with
each symbol $f\in\mathbold{\Sigma}$, a weight $\automataweightingfunction(f)\in \Xi$. The weight of a
term $t\in \mathit{Ter}(\mathbold{\Sigma})$ is inductively defined as follows.
\begin{equation}
\automataweightingfunction(t)=\left\{\begin{array}{lcl}
\automataweightingfunction(f) & & \mbox{if $t=f$ for $f\in \mathbold{\Sigma}_{0}$} \\
\automataweightingfunction(f) + \sum_{i=1}^{\mathfrak{a}(f)} \automataweightingfunction(t_i) & & \mbox{if $t=f(t_1,...,t_{\mathfrak{a}(f)})$ and $\mathfrak{a}(f)\geq1$} \\
\end{array}\right.
\end{equation}
The following lemma says that given an alphabet $\mathbold{\Sigma}$, a weighting function $\automataweightingfunction:\mathbold{\Sigma} \rightarrow \Xi$, and
a weight $a\in \Xi$, one can construct a tree-automaton ${\mathcal{A}}(\mathbold{\Sigma},\automataweightingfunction,a)$ generating precisely
the terms in $\mathit{Ter}(\mathbold{\Sigma})$ with weight $a$.
\begin{lemma}
\label{lemma:WeightedTerms}
Let $\mathbold{\Sigma} = \mathbold{\Sigma}_0 \cup ...\cup \mathbold{\Sigma}_r$ be a ranked alphabet and $\automataweightingfunction:\mathbold{\Sigma}\rightarrow \Xi$
be a weighting function on $\mathbold{\Sigma}$. Then for each weight $a\in \Xi$, one can construct in time $|\mathbold{\Sigma}|\cdot |\Xi|^{O(r)}$
a tree-automaton ${\mathcal{A}}(\mathbold{\Sigma},\automataweightingfunction,a)$ such that
$\lang({\mathcal{A}}(\mathbold{\Sigma},\automataweightingfunction,a))= \{t\in \mathit{Ter}(\mathbold{\Sigma})\;|\; \automataweightingfunction(t) = a\}$.
\end{lemma}
\begin{proof}
Let ${\mathcal{A}} = (Q,\mathbold{\Sigma},Q_{F},\Delta)$ where
$$Q=\{{\mathfrak{q}}_b\;|\; b\in \Xi\} \mbox{\hspace{1cm}} Q_F=\{{\mathfrak{q}}_a\}$$
$$\Delta = \{(f,{\mathfrak{q}}_{\automataweightingfunction(f)})\;|\; f\in \mathbold{\Sigma}_0\} \cup
\{({\mathfrak{q}}_{b_1},...,{\mathfrak{q}}_{b_{\mathfrak{a}(f)}},f,{\mathfrak{q}}_b) \; |\; f\in \mathbold{\Sigma},\; \mathfrak{a}(f)\geq 1, \; b=\automataweightingfunction(f) + \sum_{i=1}^{\mathfrak{a}(f)} b_i\}$$
We will show that ${\mathcal{A}}$ generates precisely the terms in $\mathit{Ter}(\mathbold{\Sigma})$ of weight $a$.
First, we claim that for each $b\in \Xi$ and each $i\in \N$,
\begin{equation}
\label{equation:LevelsWeight}
\lang({\mathcal{A}},{\mathfrak{q}}_b,i) = \{t\in \mathit{Ter}(\mathbold{\Sigma}) \; |\; \automataweightingfunction(t) = b,\; t \mbox{ has depth }i \}.
\end{equation}
The proof of this claim follows by induction on $i$. Equation \ref{equation:LevelsWeight} is true for $i=1$,
since in this case $\lang({\mathcal{A}},{\mathfrak{q}}_b,1)=\{f\in \mathbold{\Sigma}_0\; |\; (f, {\mathfrak{q}}_b)\in \Delta,\; \automataweightingfunction(f)=b\}$.
Assume that Equation \ref{equation:LevelsWeight} holds for $i\in \N$. We show that it also holds for $i+1$. By Equation \ref{equation:InductiveDefinition},
we have that
\begin{equation*}
\begin{array}{lcl}
\lang({\mathcal{A}},{\mathfrak{q}}_b,i+1) = \lang({\mathcal{A}},{\mathfrak{q}}_b,i) & \cup &
\{f(t_1,...,t_{\mathfrak{a}(f)})\;|\; ({\mathfrak{q}}_{b_1},...,{\mathfrak{q}}_{b_{\mathfrak{a}(f)}},f,{\mathfrak{q}}_b)\in \Delta_{\mathfrak{a}(f)}',\; \\
& & \hspace{2.7cm} t_j\in \lang({\mathcal{A}},{\mathfrak{q}}_{b_j},i)\;\} \\
\end{array}
\end{equation*}
By the induction hypothesis, $\automataweightingfunction(t_j)=b_j$ for each $t_j\in \lang({\mathcal{A}}, {\mathfrak{q}}_{b_j},i)$. Therefore
the weight of $f(t_1,...,t_{\mathfrak{a}(f)})$ is
$\automataweightingfunction(f) + \sum_{j=1}^{\mathfrak{a}(f)} \automataweightingfunction(t_i) = \automataweightingfunction(f) + \sum_{j=1}^{\mathfrak{a}(f)} b_j$.
Since $({\mathfrak{q}}_{b_1},...,{\mathfrak{q}}_{b_{\mathfrak{a}(f)}},f,{\mathfrak{q}}_b)\in \Delta_{\mathfrak{a}(f)}$ if and only if $b=\automataweightingfunction(f) + \sum_{j=1}^{\mathfrak{a}(f)} b_j$,
our claim is proved. Note that ${\mathfrak{q}}_a$ is the only final state in $Q_F$. Therefore the language accepted by ${\mathcal{A}}$ is
\begin{equation}
\label{equation:Union}
\lang({\mathcal{A}}) = \bigcup_{i\in \N} \lang({\mathcal{A}},{\mathfrak{q}}_a, i).
\end{equation}
Since for each $i\in \N$ the language $\lang({\mathcal{A}},{\mathfrak{q}}_a,i)$ consists of all terms of weight $a$ accepted in
depth $i$, we have that $\lang({\mathcal{A}})$ is the set of all terms of weight $a$ accepted in any finite depth, proving
in this way the lemma.
\end{proof}
\section{Tree Slice Languages}
\label{section:TreeSliceLanguage}
As mentioned in the introduction, the proof of Theorem \ref{theorem:MainTheoremDirectedTreewidth} is based on the framework
of tree slice languages. We dedicate this section to the introduction of this framework. We start by defining, in Subsection
\ref{subsection:Slices}, the notion of {\em slice} of arity $r$. Intuitively, a slice of arity $r$ is a digraph whose vertex set is
partitioned into a center $C$, an out-frontier $F_0$ and $r$ in-frontiers $F_1,...,F_r$. Each such a slice should be
regarded as a function symbol of arity $r$. Within this point of view, a finite set $\mathbold{\Sigma}$ of
slices with possibly distinct arities can be regarded as a ranked alphabet (Subsection \ref{subsection:SliceAlphabets}).
In Subsection \ref{subsection:GluabilityOfSlices} we introduce a notion of gluability for slices.
A slice ${\mathbf{S}}$ can be glued to a slice ${\mathbf{S}}'$ at frontier $j$ if the out-frontier of ${\mathbf{S}}$ can be matched with the $j$-th in-frontier of ${\mathbf{S}}'$.
In Subsection \ref{subsection:TermsAndUnitDecompositions} we define {\em unit decompositions}, and {\em tree slice languages}. A unit
decomposition is a term $\mathbf{T}$ over a slice alphabet $\mathbold{\Sigma}$ satisfying the property that
each two slices associated with consecutive positions of $\mathbf{T}$ can be glued along their matching frontiers.
A tree slice language $\lang$ is a tree language over a slice alphabet
$\mathbold{\Sigma}$ such that each term $\mathbf{T}\in \lang$ is a unit decomposition. A {\em slice tree automaton} is a
tree automaton ${\mathcal{A}}$ generating a tree slice language $\lang({\mathcal{A}})$. In Subsection \ref{subsection:DigraphsUnitDecompositions}
we show how to associate with each unit decomposition $\mathbf{T}$, a digraph $\composedT$ which is intuitively obtained by
gluing each two consecutive slices of $\mathbf{T}$. We can extend this association to slice languages. Namely, the graph language $\lang_{{\mathcal{G}}}$
derived from a slice language $\lang$ is the set of all digraphs associated to unit decompositions in $\lang$.
In Subsection \ref{subsection:Subslices} we will introduce the notion of sub-decompositions
of unit decompositions. Sub-decompositions should be regarded as a slice theoretic analog of the notion of subgraph.
A key idea of this paper is to reduce the problem of counting subgraphs of a digraph to the problem of counting
sub-decompositions of a unit decomposition.
In Subsection \ref{subsection:InitialSliceTreeAutomata} we show that given any slice alphabet $\mathbold{\Sigma}$, one
can construct a slice automaton ${\mathcal{A}}(\mathbold{\Sigma})$ whose slice language consists of all unit decompositions over $\mathbold{\Sigma}$.
Finally, in Subsection \ref{subsection:SliceProjection} we introduce the notion of slice projection, which will be used in many places
along this paper.
Our main application for slice languages will be given in Section \ref{section:zSaturationAndCounting} where we will introduce the notion
of {\em $z$-saturated tree slice language}. We will use this notion to count subgraphs satisfying interesting properties on digraphs of constant tree-zig-zag number.
Since the tree-zig-zag number of a digraph is at most a constant times its directed treewidth, we will also be able to count subgraphs
satisfying interesting properties on digraphs of constant directed treewidth.
\subsection{\bf Slices}
\label{subsection:Slices}
A {\em slice} of arity $r\geq 0$ is a digraph ${\mathbf{S}} = (V,E,s,t,\rho,\xi,[C,F_0,F_1,...,F_r])$ with vertex set
$V = C\cup F_0 \cup ... \cup F_r$ and edge set $E$.
The function $s:E\rightarrow V$ associates with each edge $e\in E$ a source
vertex $e^s$, while the function $t:E\rightarrow V$ associates with each edge $e\in E$ a target vertex $e^t$. We say that $e^s$ and $e^t$
are the endpoints of $e$. The function $\rho:C \rightarrow \Gamma_1$ labels each vertex in $C$ with an element
from a finite set of labels $\Gamma_1$, and $\xi: E \rightarrow \Gamma_2$ labels each edge in $E$ with an element from a
finite set of labels $\Gamma_2$. We say that $C$ is the center of ${\mathbf{S}}$, $F_0$ is the out-frontier of ${\mathbf{S}}$,
and for each $j\in \{1,...,r\}$, $F_j$ is the $j$-th in-frontier of ${\mathbf{S}}$. A slice is subject to the following restrictions.
\begin{enumerate}[s1)]
\item The sets $C,F_0,...,F_r$ are pairwise disjoint. For concreteness, we assume that $C$ is either empty or
$C=\{1,...,n\}$ for some $n\in \N$, and that for each $j\in \{0,...,r\}$, the frontier $F_j$ is either
empty or $F_j=\{[j,i_{j,1}],...,[j,i_{j,c_j}]\}$ for some $c_j\in \N$, and $i_{j,1}< ... < i_{j,c_j}\in \N$.
\item No edge in $E$ has both endpoints in the same frontier.
\item Each frontier vertex $v\in F_0\cup F_1 \cup ...\cup F_r$ is the endpoint of a unique edge $e$.
\end{enumerate}
We say that ${\mathbf{S}}$ is a {\em unit slice} if the center $C$ has at most one vertex. In other words in a unit
slice the center is either empty or the singleton $\{1\}$. In this work we will only be interested in
unit slices. We say that a frontier $F_j$ is normalized if $i_{j,k}=k$ for each $k\in \{1,...,c_j\}$. A slice
${\mathbf{S}}$ is {\em normalized} if all of its frontiers are normalized.
Non-normalized slices will play an important role in Subsection \ref{subsection:Subslices} when considering
the notion of sub-slice.
A slice of arity $0$
is a slice with no in-frontier. In this case ${\mathbf{S}} = (V,E,s,t,\rho,\xi,[C,F_0])$ with $V=C \cup F_0$.
A slice of arity $0$ should not be confused with a slice in which all in-frontiers are empty. Rather, in such a slice the in-frontiers simply do not
exist.
In Figure \ref{figure:Slices} we depict three examples of unit slices.
\begin{figure}[!hf]
\centering
\includegraphics[scale=0.35]{slices}
\caption{${\mathbf{S}}$ is a slice of arity $0$, ${\mathbf{S}}'$ is a slice of arity $1$ and ${\mathbf{S}}''$ is a slice of arity $2$. The out-frontier
$F_0$ is always drawn on the top. The in-frontiers $F_1,...,F_r$ are drawn at the bottom and in increasing order from left to right.
For each frontier vertex $[j,i]$ we draw a black dot at frontier $j$ and write the number $i$ near from it. Within each frontier, the black
dots are drawn in increasing order from left to right. The center vertex, if any, is drawn in the center of each box. The edges are drawn in red.
The slices ${\mathbf{S}}$ and ${\mathbf{S}}'$ are normalized. The slice ${\mathbf{S}}''$ is not normalized because $F_1 = \{[1,2],[1,4],[1,5]\}$, instead of
$\{[1,1],[1,2],[1,3]\}$.}
\label{figure:Slices}
\end{figure}
\subsection{\bf Slice Alphabets}
\label{subsection:SliceAlphabets}
A slice alphabet is simply a finite set $\mathbold{\Sigma}$ of slices, possibly with different arities.
Slice alphabets will be used to define terms over slices and to provide sliced representations of digraphs.
Let ${\mathbf{S}}$ be a slice with frontiers $F_j = \{[j,i_{j,1}],...,[j,i_{j,c_j}]\}$ for $j\in \{0,...,r\}$.
The width $w({\mathbf{S}})$ of ${\mathbf{S}}$ is the size of its largest frontier, i.e., $w({\mathbf{S}}) = \max_{j} \{c_j\}$.
The {\em extra-width} $ew({\mathbf{S}})$ of ${\mathbf{S}}$ is the greatest number occurring in a frontier of ${\mathbf{S}}$. More precisely,
$ew({\mathbf{S}}) = \max_{j} \{i_{j,c_j}\}$. For instance, in Figure \ref{figure:Slices} the extra-width of the slice ${\mathbf{S}}''$ is $5$.
For any $c,q,r\in \N$ with $q\geq c$, and any finite sets of labels $\Gamma_1$ and $\Gamma_2$, we let $\mathbold{\Sigma}_r(c,q,\Gamma_1,\Gamma_2)$
denote the set of all unit slices of arity $r$, width at most $c$, extra-width at most $q$, whose center vertex (if any) is labelled with an element of $\Gamma_1$,
and whose edges are labelled with elements of $\Gamma_2$.
Now consider the set
$$\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2) = \mathbold{\Sigma}_0(c,q,\Gamma_1,\Gamma_2) \cup
\mathbold{\Sigma}_1(c,q,\Gamma_1,\Gamma_2) \cup ...\cup \mathbold{\Sigma}_r(c,q,\Gamma_1,\Gamma_2).$$
We can view $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)$ as a ranked alphabet by regarding each
slice in $\mathbold{\Sigma}_j(c,q,\Gamma_1,\Gamma_2)$ as a function symbol of arity $j$.
We let $\mathbold{\Sigma}_j(c,\Gamma_1,\Gamma_2)$ denote the subset of $\mathbold{\Sigma}_j(c,c,\Gamma_1,\Gamma_2)$ consisting only of normalized slices and
set $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2) = \mathbold{\Sigma}_0(c,\Gamma_1,\Gamma_2) \cup \mathbold{\Sigma}_1(c,\Gamma_1,\Gamma_2) \cup ...\cup \mathbold{\Sigma}_r(c,\Gamma_1,\Gamma_2)$. In this work we are only interested in slices of arity at most $2$. Therefore, when considering
the slice alphabets $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)$ and $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$ defined above, we assume that $r=2$.
\subsection{\bf Gluability of Slices}
\label{subsection:GluabilityOfSlices}
\begin{figure}[!hf]
\centering
\includegraphics[scale=0.3]{treeSliceDecomposition}
\caption{$G$ is a digraph, $\mathcal{T}=(N,F,\mathfrak{m})$ is an olive-tree decomposition of $G$ and
$\mathbf{T}$ is a unit-decomposition of $G$. Note that $\mathfrak{m}(a_1) = 111$, $\mathfrak{m}(a_2) = 11$, $\mathfrak{m}(a_3) = \lambda$,
$\mathfrak{m}(a_4) = 12$ and $\mathfrak{m}(a_1) = 112$. The unit decomposition $\mathbf{T}$ is compatible with $\mathcal{T}$ since
the map defined by $a_1\rightarrow v_{111}$, $a_2\rightarrow v_{11}$, $a_3\rightarrow v_{\lambda}$,
$a_4\rightarrow v_{12}$ and $a_5\rightarrow v_{112}$ is an isomorphism
from $G$ to $\stackrel{\circ}{\mathbf{T}}$. The unit decomposition $\mathbf{T}'$ is a sub-decomposition of $\mathbf{T}$.}
\label{figure:UnitDecomposition}
\end{figure}
If ${\mathbf{S}}$ is a slice and $[j,i]$ is a vertex in the $j$-th frontier of ${\mathbf{S}}$, then we denote by
$e({\mathbf{S}},j,i)$ the unique edge of ${\mathbf{S}}$ that has $[j,i]$ as endpoint.
Let ${\mathbf{S}} = (V,E,\rho,\xi,[C,F_0,F_1,...,F_r])$ and ${\mathbf{S}}' = (V',E',\rho',\xi',[C',F_0',F_1', ...,F_r'])$
be two slices in $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)$. We say that ${\mathbf{S}}$ can be glued to ${\mathbf{S}}'$ at frontier $j$, for $1\leq j\leq r$, if the out-frontier
of ${\mathbf{S}}$ can be coherently matched with the $j$-th in-frontier of ${\mathbf{S}}'$. Formally, ${\mathbf{S}}$ can be
glued to ${\mathbf{S}}'$ at frontier $j$ if the following conditions are satisfied.
\begin{enumerate}
\renewcommand{\theenumi}{g\arabic{enumi}}
\item\label{gluingone} For each $i\in \{1,...,q\}$, $[0,i]\in F_0$ if and only if $[j,i]\in F_j'$.
\item\label{gluingtwo} $\xi(e({\mathbf{S}},0,i)) = \xi(e({\mathbf{S}}',j,i))$.
\item\label{gluingthree} Either $[0,i]$ is the target of $e({\mathbf{S}},0,i)$ and $[j,i]$ is the source of $e({\mathbf{S}}',j,i)$
or $[0,i]$ is the source of $e({\mathbf{S}},0,i)$ and $[j,i]$ is the target of $e({\mathbf{S}}',j,i)$.
\end{enumerate}
Intuitively, Condition \ref{gluingone} says that the vertex $[0,i]$ in the out-frontier of ${\mathbf{S}}$ is matched with the
vertex $[j,i]$ in the $j$-th in-frontier of ${\mathbf{S}}'$. Condition \ref{gluingtwo} says that the unique edge of ${\mathbf{S}}$ having $[0,i]$
as endpoint has the same label as the unique edge of ${\mathbf{S}}'$ having $[j,i]$ as endpoint. Finally, Condition \ref{gluingthree}
says that these edges must also agree in direction. For instance, in Figure \ref{figure:UnitDecomposition}, the slice
$\mathbf{T}[11]$ can be glued to the slice $\mathbf{T}[1]$ at frontier $1$. While $\mathbf{T}[12]$ can be
glued to $\mathbf{T}[1]$ at frontier $2$.
\subsection{\bf Terms over Slices, Unit Decompositions and Tree Slice Languages}
\label{subsection:TermsAndUnitDecompositions}
As observed in Subsection \ref{subsection:SliceAlphabets}, a slice alphabet $\mathbold{\Sigma}$ can be
regarded as a ranked alphabet where each slice ${\mathbf{S}}\in \mathbold{\Sigma}$ of arity $r$ is
a function symbol of arity $r$. In this paper $\mathbold{\Sigma}$ will be typically the
slice alphabet $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)$ or the normalized slice alphabet $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$, both
defined in Subsection \ref{subsection:SliceAlphabets}.
We let $\mathit{Ter}(\mathbold{\Sigma})$ denote the set of all terms
formed with slices from $\mathbold{\Sigma}$.
In this work however we will be only interested on terms over $\mathbold{\Sigma}$ that can give rise
to digraphs. These terms are called {\em unit decompositions}.
\begin{definition}[Unit Decomposition]
\label{definition:UnitDecomposition}
Let $\mathbold{\Sigma}$ be an alphabet of unit slices. A term $\mathbf{T}\in \mathit{Ter}(\mathbold{\Sigma})$ is a {\em unit pre-decomposition} if for each
two consecutive positions $p, pj\in \mathit{Pos}(\mathbf{T})$, the slice $\mathbf{T}[pj]$ can be glued to the slice $\mathbf{T}[p]$ at frontier $j$. A term
$\mathbf{T}$ is a {\em unit decomposition} if it is a unit pre-decomposition in which the slice $\mathbf{T}[\lambda]$ at the root of $\mathbf{T}$ has empty out-frontier.
\end{definition}
The width $w(\mathbf{T})$ of a unit decomposition $\mathbf{T}$ is the maximum width of a slice occurring in it.
A unit decomposition is normalized if for each position $p\in \mathit{Pos}(\mathbf{T})$
the slice $\mathbf{T}[p]$ is normalized. For instance, the unit decomposition $\mathbf{T}$ in Figure \ref{figure:UnitDecomposition}
is normalized while the unit decomposition $\mathbf{T}'$ in the same figure is not.
We let $\lang(\mathbold{\Sigma})$ be the set of all unit decompositions in $\mathit{Ter}(\mathbold{\Sigma})$.
A tree slice language over $\mathbold{\Sigma}$ is any subset $\lang$ of $\lang(\mathbold{\Sigma})$.
We say that a tree slice language $\lang\subseteq \lang(\mathbold{\Sigma})$ is normalized if all unit
decompositions in $\lang$ are normalized. We will see in the next subsection
that with each unit decomposition $\mathbf{T}$ one can associate a digraph $\composedT$ which is intuitively obtained
by gluing each two consecutive slices in $\mathbf{T}$. Thus with any slice language $\lang$ one can associate a graph
language $\lang_{{\mathcal{G}}}$ consisting of all digraphs that correspond to unit decompositions in $\lang$.
Of particular importance to us are the slice languages that can be effectively represented via tree-automata
over slice alphabets. We call these automata {\em slice tree-automata}.
\begin{definition}[Slice Tree-Automaton]
\label{definition:SliceTreeAutomaton}
Let $\mathbold{\Sigma}$ be a slice alphabet. We say that a tree-automaton ${\mathcal{A}}=(Q,\mathbold{\Sigma},Q_{F},\Delta)$ over $\mathbold{\Sigma}$
is a {\em slice tree-automaton} if for each term $\mathbf{T}\in \lang({\mathcal{A}})$, $\mathbf{T}$ is a unit decomposition over $\mathbold{\Sigma}$.
\end{definition}
In other words, ${\mathcal{A}}$ is a slice tree-automaton if $\lang({\mathcal{A}})\subseteq \lang(\mathbold{\Sigma})$. In this case
we say that $\lang({\mathcal{A}})$ is the slice language generated by ${\mathcal{A}}$. We say that a slice tree automaton ${\mathcal{A}}$
is normalized if the slice language $\lang({\mathcal{A}})$ is normalized.
\subsection{\bf Digraphs associated with Unit Decompositions}
\label{subsection:DigraphsUnitDecompositions}
Each unit decomposition $\mathbf{T}\in \lang(\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2))$ can be
associated with a digraph $\composedT$ which is intuitively obtained by gluing together each two
consecutive slices in $\mathbf{T}$. For instance, gluing the slices of the unit decomposition
$\mathbf{T}$ of Figure \ref{figure:UnitDecomposition} we get the digraph $G$. To make this notion
of gluing more precise, it will be convenient to define the notion of {\em sliced edge sequence}.
Intuitively, each edge $e$ of the digraph $\composedT$ will be defined with basis on a sliced
edge sequence that contains all "sliced parts" of $e$.
Below, $\mathit{support}(\mathbf{T})$ denotes the set of all positions in $\mathit{Pos}(\mathbf{T})$ for
which the slice $\mathbf{T}[p]$ has non-empty center.
\begin{definition}[Sliced Edge Sequence]
\label{definition:SlicedEdgeSequence}
Let $\mathbf{T}$ be a unit decomposition over a slice alphabet $\mathbold{\Sigma}$.
Let $p,p'$ be two positions in $\mathit{support}(\mathbf{T})$.
A sliced edge sequence from $p$ to $p'$ is a sequence
\begin{equation}
\label{equation:slicedEdgeSequence}
K\equiv(p_1,a_{1},e_{1},b_{1})(p_2,a_{2},e_{2},b_{2})...(p_n,a_{n},e_{n},b_{n})
\end{equation}
\noindent where $p_1=p$, $p_n=p'$, and the following conditions are satisfied.
\begin{enumerate}
\itemsep0.35em
\item \label{item:edgesequence1} For each $i\in \{1,...,n\}$, $e_{i}$ is an edge in $\mathbf{T}[p_i]$ with source $a_{i}$ and target $b_{i}$.
\item \label{item:edgesequence2} $a_{1}$ is the center vertex of $\mathbf{T}[p_1]$ and $b_{n}$ is the center vertex of $\mathbf{T}[p_{n}]$.
\item \label{item:edgesequence3} For each $i\in \{1,...,n-1\}$, there is a $j$ such that either $p_i = p_{i+1} j$ or $p_{i+1} = p_i j$.
\item \label{item:edgesequence4} If $p_{i} = p_{i+1} j$ then for some $k\in \{1,...,q\}$, $b_i = [0,k]$ and $a_{i+1} = [j,k]$.
\item \label{item:edgesequence5} If $p_{i+1} = p_i j$ then for some $k\in \{1,...,q\}$, $b_i= [j,k]$ and $a_{i+1} = [0,k]$.
\end{enumerate}
\end{definition}
We note that Conditions \ref{item:edgesequence1}-\ref{item:edgesequence5} of Definition \ref{definition:SlicedEdgeSequence} together with the fact that $\mathbf{T}$ is a unit
decomposition ensures that the $p_i\neq p_j$ for $i\neq j$.
To illustrate Definition \ref{definition:SlicedEdgeSequence} we note that in the unit decomposition $\mathbf{T}$ of Figure
\ref{figure:UnitDecomposition} there is a sliced edge sequence from position $\lambda$ to position $12$, a sliced edge sequence from $12$ to $112$ and so
on. Intuitively, each sliced edge sequence $K$ gives rise to an edge $e_K$ in the digraph $\composedT$ that is obtained by gluing
all of its sliced parts $e_{1},...,e_{n}$. Condition \ref{item:edgesequence1} says that $e_{i}$ is the sliced part of $e_K$
lying at the slice $\mathbf{T}[p_i]$. Condition
\ref{item:edgesequence2} says that the source of the first sliced part of $e_K$ is the center vertex of $\mathbf{T}[p_1]$ and the target of the last
sliced part of $e_K$ is the center vertex of $\mathbf{T}[p_n]$. Condition \ref{item:edgesequence3} says that for each $i\in \{1,...,n-1\}$, $e_{i}$ and $e_{{i+1}}$
lie in neighboring slices of $\mathbf{T}$. If $p_{i+1}=p_i j$ then the edge $e_i$ is intuitively directed towards the $j$-th in-frontier of
$\mathbf{T}[p_i]$. In this case, Condition \ref{item:edgesequence4}, says that the target of $e_{i}$ lies in the $j$-th in-frontier of $\mathbf{T}[p_i]$
while the source of $e_{{i+1}}$ lies in the out-frontier of $\mathbf{T}[p_{i+1}]$.
On the other hand, if $p_i=p_{i+1}j$ then the edge $e_{i}$ is intuitively directed towards the out-frontier of $\mathbf{T}[p_{i}]$.
In this case, Condition \ref{item:edgesequence5} says that the target of $e_{i}$ lies in
the out-frontier of $\mathbf{T}[p_i]$ and the source of $e_{{i+1}}$ lies in the $j$-th in frontier of $\mathbf{T}[p_{i+1}]$.
Let $\mathbf{T}$ be a unit decomposition and for each $p\in \mathit{Pos}(\mathbf{T})$ let $\mathbf{T}[p] = (V_p,E_p,\rho_p,\xi_p)$ be the slice of
$\mathbf{T}$ at position $p$. The digraph $\composedT = (V,E,\rho,\xi)$ associated with $\mathbf{T}$ is defined as follows.
First, for each position $p\in \mathit{support}(\mathbf{T})$, we add a vertex $v_p$ to the vertex set $V$. Subsequently, for each two positions
$p,p'\in \mathit{Pos}(\mathbf{T})$ and each sliced edge sequence $K$ from $p$ to $p'$ we add an edge $e_K$ to $E$ and set its source as $e_K^s = v_p$ and its target
as $e_K^t = v_{p'}$. Observe that multiple edges are allowed in $\composedT$ since for some pair of positions $p,p'$ there may exist
more than one sliced edge sequence from $p$ to $p'$. For each $p\in \mathit{Pos}(\mathbf{T})$, the
vertex $v_p$ receives the same label as the center vertex of $\mathbf{T}[p]$. In other words, $\rho(v_p) = \rho_p(1)$
\footnote{We recall that if the center of a unit slice is not empty then the center is the singleton $\{1\}$.}. We note that if
$K$ is a sliced edge sequence as defined in Equation \ref{equation:slicedEdgeSequence} then Conditions \ref{item:edgesequence4} and \ref{item:edgesequence5}
of Definition \ref{definition:SlicedEdgeSequence} together with Condition \ref{gluingtwo} (of Subsection \ref{subsection:GluabilityOfSlices})
guarantee that all edges $e_{1},e_{2},...,e_{n}$ have the same label.
Thus the label of the edge $e_K$ is set as $\xi(e_K) = \xi_{p_1}(e_{1})=...=\xi_{p_n}(e_{n})$.
If $G=(V,E,\rho,\xi)$ is a digraph where $\rho:V \rightarrow \Gamma_1$ and $\xi:E\rightarrow \Gamma_2$
are vertex and edge labeling functions respectively, then for each two vertices $v,v'\in V$ and each label $b\in \Gamma_2$,
we let $\overrightarrow{E}(v,v',b) = \{e\;|\;e^s = v,\;e^t = v', \xi(e) = b\}$ denote the set of all edges in $E$ which have
$v$ as source vertex, $v'$ as target vertex and $b$ as label.
An isomorphism from a digraph $G_1=(V_1,E_1,\rho_1,\xi_1)$ to a digraph $G_2 = (V_2,E_2,\rho_2,\xi_2)$
is a bijection $\phi:V_1\rightarrow V_2$ from $V_1$ to $V_2$ such that for each $v\in V_1$,
$\rho_1(v) = \rho_2(\phi(v))$, and such that for each two vertices $v,v'\in V_1$ and each label $b\in \Gamma_2$,
$|\overrightarrow{E}_1(v,v',b)| = |\overrightarrow{E}_2(\phi(v),\phi(v'),b)|$.
A canonization function for finite digraphs is a function $[\;\cdot\;]$ satisfying two properties. First, for every digraph $G$,
$[G]$ is a digraph isomorphic to $G$. Second, for every two digraphs $G_1$ and $G_2$, $G_1$ is isomorphic to $G_2$ if and
only if $[G_1]=[G_2]$. We say that $[G]$ is the canonical form of $G$. In this paper we let $[\;\cdot\;]$ be an arbitrary but fixed canonization
function for finite digraphs.
We say that a term $\mathbf{T}$ is a unit decomposition of a digraph $G$ if the digraph $\composedT$ is isomorphic to $G$.
Since with any unit decomposition $\mathbf{T}$ one can associate a digraph $\composedT$, with any tree slice language $\lang$ one
can associate a possibly infinite family of digraphs.
If $\lang$ is a tree slice language over an alphabet $\mathbold{\Sigma}$ of unit slices, then the graph language derived from $\lang$ is the set $\lang_{{\mathcal{G}}}$
of canonical forms of digraphs obtained by composing the slices of each unit decomposition in $\lang$. Formally,
\begin{equation}
\label{equation:GraphLanguage}
\lang_{{\mathcal{G}}} = \{[\composedT]\;|\; \mathbf{T}\in \lang\}.
\end{equation}
For convenience, in some places we may simply say that a digraph
$H$ belongs to $\lang_{{\mathcal{G}}}$ instead of saying that $[H]$ belongs
to $\lang_{{\mathcal{G}}}$. If ${\mathcal{A}}$ is a slice tree automaton then
we denote by $\lang_{{\mathcal{G}}}({\mathcal{A}})$ the graph language derived
from $\lang({\mathcal{A}})$.
\subsection{Sub-slices and Sub-Decompositions}
\label{subsection:Subslices}
In this subsection we introduce the notions of {\em sub-slice} and of {\em sub-decomposition}. Intuitively,
the notion of sub-decomposition is a sliced version of the notion of subgraph.
Let ${\mathbf{S}}=(V,E,\rho,\xi,[C,F_0,F_1,...,F_r])$ be a slice of arity $r$. We say that a slice
${\mathbf{S}}'$ is a {\em sub-slice} of ${\mathbf{S}}$ if ${\mathbf{S}}'=(V',E',\rho',\xi',[C',F_0',F_1',...,F_r'])$ where $V'\subseteq V$,
$E'\subseteq E$, $\rho'=\rho|_{V'}$, $\xi'= \xi|_{E'}$, $C'\subseteq C$ and
$F_j'\subseteq F_j$ for each $j\in {0,1,...,r}$. In other words, a sub-slice of ${\mathbf{S}}$ is a subgraph of ${\mathbf{S}}$ that
is also a slice. Labels of vertices and edges in a sub-slice are inherited from the original slice.
We note that even if ${\mathbf{S}}$ is a normalized slice, a sub-slice ${\mathbf{S}}'$ of ${\mathbf{S}}$ may not be
normalized. For instance, in Figure \ref{figure:UnitDecomposition}, the slice $\mathbf{T}'[1]$ is a sub-slice of $\mathbf{T}[1]$.
Note that $\mathbf{T}'[1]$ is not normalized even though $\mathbf{T}[1]$ is. We also call attention to the fact that a sub-slice
has always the same arity as the original slice, and that the empty slice ${\bm{\varepsilon}}_r$ of arity $r$ is a sub-slice of any slice of
arity $r$.
\begin{definition}[Sub-decomposition]
\label{definition:Subdecomposition}
Let $\mathbold{\Sigma}$ be a slice alphabet and let $\mathbf{T}$ and $\mathbf{T}'$ be unit decompositions in $\lang(\mathbold{\Sigma})$.
We say that $\mathbf{T}'$ is a sub-decomposition of $\mathbf{T}$ if the following conditions are satisfied.
\begin{enumerate}[i)]
\item\label{subdecomposition1} $\mathit{Pos}(\mathbf{T})=\mathit{Pos}(\mathbf{T}')$,
\item\label{subdecomposition2} for each $p\in \mathit{Pos}(\mathbf{T})$ the unit slice $\mathbf{T}'[p]$ is a sub-slice of $\mathbf{T}[p]$,
\item\label{subdecomposition3} for each two consecutive positions $p,pj \in \mathit{Pos}(\mathbf{T})$ the slice
$\mathbf{T}'[pj]$ can be glued to the slice $\mathbf{T}'[p]$ at frontier $j$.
\end{enumerate}
\end{definition}
Conditions \ref{subdecomposition1}-\ref{subdecomposition3} of Definition \ref{definition:Subdecomposition} guarantee that if $\mathbf{T}'$ is a sub-decomposition of $\mathbf{T}$ then
the digraph $\composedTprime$ is a subgraph of $\composedT$.
We emphasize that $\composedTprime$ is an actual subgraph of $\composedT$ and not merely isomorphic to a subgraph of $\composedT$.
Conversely, for each subgraph $H$ of $\composedT$ there is a sub-decomposition $\mathbf{T}'$ of $\mathbf{T}$ for which $\composedTprime = H$.
Again at this point we are speaking about strict equality, and not merely isomorphism.
Thus each sub-decomposition of $\mathbf{T}$ unequivocally corresponds to a subgraph of $\composedT$.
A crucial step towards the proof of Theorem \ref{theorem:MainTheoremDirectedTreewidth} will consist in reducing the
problem of counting subgraphs of a digraph to the problem of counting sub-decompositions of a unit
decomposition.
\subsection{Initial Slice Tree-Automata}
\label{subsection:InitialSliceTreeAutomata}
In this section we will show that for each slice alphabet $\mathbold{\Sigma}$ one can construct a deterministic slice tree-automaton
${\mathcal{A}}(\mathbold{\Sigma})$ whose slice language consists of all unit decompositions that can be formed with elements from $\mathbold{\Sigma}$.
We say that ${\mathcal{A}}(\mathbold{\Sigma})$ is the initial tree-automaton for $\mathbold{\Sigma}$.
Before proceeding, we define the notion of identity slice, which will be used below in the proof of Proposition \ref{proposition:InitialAutomaton},
and later, in the proof of Lemma \ref{lemma:SubgraphsC}.
An {\em identity slice} in $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)$
is a slice $\mathbf{I}=(V,E,\rho,\xi,[C,F_0,F_1])$ of arity $1$ with empty center
($C=\emptyset$) in which all edges are "parallel". In other words for each $e\in E$, there exists a $k\in \{1,...,q\}$ such that either $e^s = [0,k]$ and
$e^t=[1,k]$ or $e^s=[1,k]$ and $e^t=[0,k]$.
\begin{figure}[!hf]
\centering
\includegraphics[scale=0.25]{IdentitySlice}
\caption{An identity slice $\mathbf{I}$ and two other slices ${\mathbf{S}}$ and ${\mathbf{S}}'$. A slice ${\mathbf{S}}$ can be glued to ${\mathbf{S}}'$ at
frontier $j$ if and only if there is a unique identity slice $\mathbf{I}$ such that ${\mathbf{S}}$ can be glued to $\mathbf{I}$, and
such that $\mathbf{I}$ can be glued to ${\mathbf{S}}'$ at $\mbox{frontier $j$}$. In this case $\mathbf{I} = \mathbf{I}({\mathbf{S}})$.
}
\label{figure:IdentitySlice}
\end{figure}
Our only interest in identity slices stems from the
fact that for each slice ${\mathbf{S}}$ of arity $r$ there is a unique identity slice $\mathbf{I}$ for which
${\mathbf{S}}$ can be glued to $\mathbf{I}$. We denote this unique identity slice by $\mathbf{I}({\mathbf{S}})$.
Additionally, for each $j\in \{1,...,r\}$ there exists a unique identity slice
$\mathbf{I}_j$ such that $\mathbf{I}_j$ can be glued to ${\mathbf{S}}$ at frontier $j$. This implies that a
slice ${\mathbf{S}}$ can be glued to a slice ${\mathbf{S}}'$ at $\mbox{frontier $j$}$ if and only if $\mathbf{I}({\mathbf{S}})$ can be
glued to ${\mathbf{S}}'$ at frontier $j$ (See Figure \ref{figure:IdentitySlice}). We observe that we consider ${\bm{\varepsilon}}_1$, the empty slice
of arity one, as an identity slice.
\begin{proposition}[Initial Slice Tree-Automaton]
\label{proposition:InitialAutomaton}
Let $\mathbold{\Sigma}$ be a slice alphabet and let $r$ be the maximum arity of a slice in $\mathbold{\Sigma}$. Then
one can construct in time $O(|\mathbold{\Sigma}|)$ a slice tree-automaton ${\mathcal{A}}(\mathbold{\Sigma})$ whose
slice language $\lang({\mathcal{A}}(\mathbold{\Sigma}))$ is the set of all unit decompositions over $\mathbold{\Sigma}$.
\end{proposition}
\begin{proof}
We construct the automaton ${\mathcal{A}}(\mathbold{\Sigma}) = (Q,\mathbold{\Sigma},Q_F,\Delta)$ explicitly.
First, we define the set $\mathbf{I}(\mathbold{\Sigma}) = \{\mathbf{I}({\mathbf{S}})\;|\; {\mathbf{S}}\in \mathbold{\Sigma}\}$ which
consist of all identity slices $\mathbf{I}$ for which some slice in $\mathbold{\Sigma}$ can be glued to $\mathbf{I}$. The
set of states $Q$ has one state ${\mathfrak{q}}_{\mathbf{I}}$ for each identity slice $\mathbf{I}$ in $\mathbf{I}(\mathbold{\Sigma})$.
The set of final states is the singleton $Q_F=\{{\mathfrak{q}}_{{\bm{\varepsilon}}_1}\}$. The transition relation
$\Delta$ has one transition $({\mathfrak{q}}_{\mathbf{I}_1},...,{\mathfrak{q}}_{\mathbf{I}_r},{\mathbf{S}},{\mathfrak{q}}_{\mathbf{I}({\mathbf{S}})})$ for
each slice ${\mathbf{S}}$ of arity $r$ in $\mathbold{\Sigma}$, where for each $j\in \{1,...,r\}$, $\mathbf{I}_i$ is
the unique identity slice that can be glued to ${\mathbf{S}}$ at frontier $j$. Observe that since the states ${\mathfrak{q}}_{\mathbf{I}_1},...,{\mathfrak{q}}_{\mathbf{I}_r}$ and
${\mathfrak{q}}_{\mathbf{I}({\mathbf{S}})}$ are completely determined by ${\mathbf{S}}$, the relation $\Delta$ has $|\mathbold{\Sigma}|$ transitions.
By the construction of the transition relation $\Delta$ we have that for each term $\mathbf{T}$ accepted by ${\mathcal{A}}(\mathbold{\Sigma})$
and each two consecutive positions $p,pj$ in $\mathit{Pos}(\mathbf{T})$, the slice $\mathbf{T}[pj]$ can be glued
to the slice $\mathbf{T}[p]$ at frontier $j$. Since the unique accepting state is ${\mathfrak{q}}_{{\bm{\varepsilon}}_1}$ we also have that $\mathbf{T}[\lambda]$
has empty out-frontier. This implies that each such term $\mathbf{T}$ is a unit decomposition. For the converse, let
$\mathbf{T}$ be a unit pre-decomposition over $\mathbold{\Sigma}$. We will show by induction on the height of $\mathbf{T}$ that
$\mathbf{T}$ reaches the state ${\mathfrak{q}}_{\mathbf{I}(\mathbf{T}[\lambda])}$. This implies in particular that if $\mathbf{T}$ is a unit
decomposition, then $\mathbf{T}$ reaches the unique accepting state ${\mathfrak{q}}_{{\bm{\varepsilon}}_1}$, since in this case $\mathbf{T}[\lambda]$ can
be glued to ${\bm{\varepsilon}}_1$. In the base case, let $\mathbf{T}$ be a unit pre-decomposition
of height $0$. Then $\mathbf{T}$ consists of a single slice ${\mathbf{S}}$ of $\mbox{arity $0$}$. By definition of $\Delta$, we have that there is a
transition $({\mathbf{S}},{\mathfrak{q}}_{\mathbf{I}({\mathbf{S}})})\in \Delta$ and therefore ${\mathbf{S}}$ reaches the state ${\mathfrak{q}}_{\mathbf{I}({\mathbf{S}})}$.
Now suppose that the claim is valid for every unit pre-decomposition of height at most $h$ and let $\mathbf{T}$ be a unit pre-decomposition
of height $h+1$. Let the slice $\mathbf{T}[\lambda]$ at the root of $\mathbf{T}$ have arity $r$. By the induction hypothesis, for each
$i\in \{1,...,r\}$, the subterm $\mathbf{T}|_{i}$ rooted at position $i$, reaches the state ${\mathfrak{q}}_{\mathbf{I}_i}$ where
$\mathbf{I}_i = \mathbf{I}(\mathbf{T}|_{i}[\lambda])$. Since by the construction of $\Delta$ the transition
$({\mathfrak{q}}_{\mathbf{I}_1},...,{\mathfrak{q}}_{\mathbf{I}_r},\mathbf{T}[\lambda],{\mathfrak{q}}_{\mathbf{I}(\mathbf{T}[\lambda])})$ belongs to $\Delta$,
we have that $\mathbf{T}$ reaches the state ${\mathfrak{q}}_{\mathbf{I}(\mathbf{T}[\lambda])}$. This proves the inductive step.
\end{proof}
\subsection{Normalizing Projection and Unweighting Projection}
\label{subsection:SliceProjection}
We say that a mapping $\mathbold{\pi}:\mathbold{\Sigma} \rightarrow \mathbold{\Sigma}'$
between slice alphabets is a slice projection if $\mathbold{\pi}$ is arity preserving, gluing preserving, and empty-frontier preserving.
By arity preserving we mean that $\mathfrak{a}({\mathbf{S}}) = \mathfrak{a}(\mathbold{\pi}({\mathbf{S}}))$. By gluing preserving we mean that if ${\mathbf{S}}$ can be
glued to ${\mathbf{S}}'$ at frontier $j$ then $\mathbold{\pi}({\mathbf{S}})$ can be glued to $\mathbold{\pi}({\mathbf{S}}')$ at frontier $j$. And by empty-frontier
preserving we mean that if a frontier $F_i$ is empty in ${\mathbf{S}}$ then the corresponding frontier in $\mathbold{\pi}({\mathbf{S}})$ is also
empty. Two classes of slice projections will be of particular importance to us. The normalizing projections, and the unweighting projections which
are defined below.
\begin{figure}[!hf]
\centering
\includegraphics[scale=0.40]{normalizingProjection}
\caption{The normalizing projection ${\mathbold{\eta}}_{c,q}$ normalizes each frontier of a slice in such a way that
the order of the vertices in each frontier is preserved. In this example $c=3$ and $q=6$. The unnormalized slice ${\mathbf{S}}_1$ has frontiers
$F_0 = \{[0,3],[0,6]\}$ and $F_1 = \{[1,2],[1,4],[1,5]\}$. After the normalization the frontiers become $F_0' = \{[0,1],[0,2]\}$
and $F_1'= \{[1,1],[1,2],[1,3]\}$. The unweighting projection ${\mathbold{\zeta}}$ simply erases the weights associated to each
edge of the slice. In this example, the weights $\alpha_1,\alpha_2,\alpha_3$ attached to the edges of the slice ${\mathbf{S}}_2$ are erased by ${\mathbold{\zeta}}$.
}
\label{figure:NormalizingAndUnweighting}
\end{figure}
The normalizing projection ${\mathbold{\eta}}_{c,q}:\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)\rightarrow \mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$
acts on each slice ${\mathbf{S}}$ in $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)$ by renumbering the frontier vertices of ${\mathbf{S}}$ in
such a way that the new resulting slice ${\mathbold{\eta}}_{c,q}({\mathbf{S}})$ is normalized and in such a way that the ordering
of the vertices inside each frontier is preserved. More precisely, for each $j\in \{0,...,r\}$, let $F_j =\{[j,i_{j,1}],[j,i_{j,2}], ..., [j,i_{j,c_j}]\}$
where $c_j\leq c$ and $i_{j,1} < i_{j,2} < ... < i_{j,c_j} \leq q$.
Then the slice ${\mathbold{\eta}}_{c,q}({\mathbf{S}})$ is obtained from ${\mathbf{S}}$ by replacing each frontier vertex $[j,i_{j,k}]$ with the vertex $[j,k]$.
After the application of the normalizing projection ${\mathbold{\eta}}_{c,q}$ the $j$-th frontier of ${\mathbf{S}}$ becomes
$F_j' =\{[j,1],...,[j,c_j]\}$ (Figure \ref{figure:NormalizingAndUnweighting}).
We note that if $\mathbf{T}$ is a unit decomposition over $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)$,
then ${\mathbold{\eta}}_{c,q}(\mathbf{T})$ is a normalized unit decomposition over $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$ representing the same
digraph. In other words, if $\mathbf{T}'={\mathbold{\eta}}_{c,q}(\mathbf{T})$ then $\composedTprime = \composedT$.
If $\Gamma_2$ is a set of edge labels and $\Omega$ is a set of edge weights, then the set $\Gamma_2\times \Omega$ can be
regarded as a new set of edge labels.
The unweighting projection ${\mathbold{\zeta}}_{\Omega}:\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2\times \Omega)\rightarrow \mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)$
is a function that takes a slice ${\mathbf{S}}\in \mathbold{\Sigma}(c,q,\Gamma_1, \Gamma_2\times \Omega)$ and erases the weight coordinate
from the label of each edge. More precisely, if ${\mathbf{S}} = (V,E,\rho,\xi\times \graphweightingfunction,[C,F_0,...,F_j])$
where $\xi\times \graphweightingfunction: E\rightarrow \Gamma_2 \times \Omega$, then
${\mathbold{\zeta}}_{\Omega}({\mathbf{S}}) = (V,E,\rho,\xi,[C,F_0,...,F_j])$ where $\xi:E\rightarrow \Gamma_2$ is
the projection of $\xi\times \graphweightingfunction$ to its first coordinate. Unweighting projections and normalizing projections
will be used in Section \ref{section:ProofOfMainTheorem} to construct the slice tree-automaton ${\mathcal{A}}(\varphi,k,z,l,a)$ mentioned in the introduction
(Section \ref{subsection:ProofTechniques}).
\section{$z$-Saturated Tree Slice Languages}
\label{section:zSaturationAndCounting}
In this section we will define the notion of tree-zig-zag number of a unit decomposition and the notion of {\em $z$-saturated} tree
slice language. We will show that given a $z$-saturated tree slice language $\lang$ generated by a slice tree-automaton
${\mathcal{A}}$ and a unit decomposition $\mathbf{T}$ of tree-zig-zag $\mbox{number $z$,}$ we can count in polynomial time the number of subgraphs
of $\composedT$ that are isomorphic to some digraph in $\lang_{{\mathcal{G}}}$.
This seemingly abstract result is a crucial step towards the proof of our main theorem (Theorem \ref{theorem:MainTheoremDirectedTreewidth}).
The next crucial step, which will be carried in Section \ref{section:MSOandTreeSliceLanguages}, consists in
showing that for any {$\mbox{MSO}_2$\;} logic sentence $\varphi$ and any $k,z\in \N$, one can define a $z$-saturated slice tree-automaton
generating precisely the set of digraphs that at the same time are the union of $k$ directed paths and satisfy $\varphi$.
\subsection{$z$-Saturation}
\label{subsection:zSaturation}
Let $\mathbf{T}$ be a unit decomposition over $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)$ and let $\composedT=(V,E,\rho,\xi)$
be the digraph represented by $\mathbf{T}$.
We say that $\mathbf{T}$ is compatible with an olive-tree decomposition $\mathcal{T}=(N,F,\mathfrak{m})$
of a digraph $G=(V',E',\rho',\xi')$ if both $\mathbf{T}$ and $\mathcal{T}$ have the
same tree-structure (i.e. $N=\mathit{Pos}(\mathbf{T})$), and the map $\beta:V'\rightarrow V$ given by $\beta(u)=v_{\mathfrak{m}(u)}$
is an isomorphism from $G$ to $\composedT$. For instance, in Figure \ref{figure:UnitDecomposition}, the unit decomposition $\mathbf{T}$ is
compatible with the olive-tree decomposition $\mathcal{T}$. Note that for each unit decomposition $\mathbf{T}$
there is a unique olive-tree decomposition $\mathcal{T} = (N,F,\mathfrak{m})$ of the digraph $\composedT$ such that
$\mathbf{T}$ is compatible with $\mathcal{T}$. In this olive-tree decomposition, $N=\mathit{Pos}(\mathbf{T})$ and $\mathfrak{m}$ is defined by setting
$\mathfrak{m}(v_p) = p$ for each position $p\in \mathit{Pos}(\mathbf{T})$.
We say that a unit decomposition $\mathbf{T}$ has {\em tree-zig-zag number} ${tzn}(\mathbf{T}) = z$ if $\mathbf{T}$ is compatible with an olive-tree decomposition
of tree-zig-zag number $z$. Intuitively, $\mathbf{T}$ has tree-zig-zag $\mbox{number $z$}$ if each simple path of $\composedT$ crosses each frontier of each slice in $\mathbf{T}$
at most $z$ times. For instance, in Figure \ref{figure:UnitDecomposition}, the unit decomposition $\mathbf{T}$ has tree-zig-zag number $2$. Note that
the olive-tree decomposition $\mathcal{T}$ in Figure \ref{figure:UnitDecomposition} that is compatible with $\mathbf{T}$ has also tree-zig-zag number
$2$. We say that a slice language $\lang$ has tree-zig-zag number $z$ if each unit decomposition in $\lang$
has tree-zig-zag number $z$.
Let $H$ be a digraph, $\mathbold{\Sigma}$ be a slice alphabet and
$$\mathbold{ud}(\mathbold{\Sigma},H,z)=\{\mathbf{T}\in \lang(\mathbold{\Sigma}) |\composedT \simeq H, {tzn}(\mathbf{T})\leq z\}.$$
We say that a tree slice language $\lang$ over $\mathbold{\Sigma}$ is {\em $z$-saturated} with respect to $\mathbold{\Sigma}$,
if for every digraph $H$, the fact that $[H]\in \lang_{{\mathcal{G}}}$ implies that $\mathbold{ud}(\mathbold{\Sigma},H,z) \subseteq \lang$. In other words $\lang$ is $z$-saturated if whenever a canonical form $[H]$ belongs to the graph language $\lang_{{\mathcal{G}}}$, all unit decompositions
of tree-zig-zag number $z$ of $H$ belong to the slice language $\lang$. If the alphabet $\mathbold{\Sigma}$
is clear from the context we may say simply that $\lang$ is $z$-saturated, instead of saying that
$\lang$ is $z$-saturated with respect to $\mathbold{\Sigma}$.
A slice tree-automaton ${\mathcal{A}}$ is $z$-saturated if $\lang({\mathcal{A}})$ is $z$-saturated.
Proposition \ref{proposition:Intersection} below justifies our interest in the concept of $z$-saturation.
\begin{proposition}
\label{proposition:Intersection}
Let $\lang$ and $\lang'$ be tree slice languages over $\mathbold{\Sigma}$ such that
$\lang$ has tree-zig-zag $\mbox{number $z$}$ and such that $\lang'$ is $z$-saturated with respect to $\mathbold{\Sigma}$.
Then $(\lang\cap \lang')_{{\mathcal{G}}} = \lang_{{\mathcal{G}}}\cap \lang'_{{\mathcal{G}}}$.
\end{proposition}
\begin{proof}
The inclusion $(\lang\cap \lang')_{{\mathcal{G}}} \subseteq \lang_{{\mathcal{G}}}\cap \lang_{{\mathcal{G}}}'$ holds for any two slice
languages $\lang$ and $\lang'$ irrespectively of whether they are saturated or not. To see this, let $H$ be a digraph and
let $[H] \in (\lang \cap \lang')_{{\mathcal{G}}}$. Then $H$ has a unit decomposition $\mathbf{T}$
in $\lang\cap \lang'$. Since $\mathbf{T}\in \lang$, $[H]\in \lang_{{\mathcal{G}}}$ and, since $\mathbf{T}\in \lang'$, $[H]\in \lang_{{\mathcal{G}}}'$.
Thus $(\lang \cap \lang')_{{\mathcal{G}}} \subseteq \lang_{{\mathcal{G}}}\cap \lang'_{{\mathcal{G}}}$. Now we prove that if $\lang$ has
tree-zig-zag number $z$ and $\lang'$
is $z$-saturated, the converse inclusion also holds. Let $H$ be a digraph and let $[H]\in \lang_{{\mathcal{G}}}\cap \lang'_{{\mathcal{G}}}$. Since
$\lang$ has tree-zig-zag number $z$, $H$ has a unit decomposition $\mathbf{T}$ of tree-zig-zag number $z$ in $\lang$. Since $\lang'$ is $z$-saturated
with respect to $\mathbold{\Sigma}$, each unit-decomposition of $H$
over $\mathbold{\Sigma}$ of tree-zig-zag number $z$ is in $\lang'$, and in particular $\mathbf{T}\in \lang'$. Therefore
$\mathbf{T}\in \lang\cap \lang'$ and $[H]\in (\lang\cap \lang')_{{\mathcal{G}}}$.
\end{proof}
In other words, whenever $\lang$ has tree-zig-zag number $z$ and $\lang'$ is $z$-saturated, the intersection $\lang_{{\mathcal{G}}}\cap \lang_{{\mathcal{G}}}'$
of their graph languages is precisely the graph language of the intersection $\lang\cap \lang'$.
It is worth noting that Proposition \ref{proposition:Intersection} would not be true if none of the slice languages $\lang$ and $\lang'$ were saturated.
For instance if $\lang=\{\mathbf{T}\}$ and $\lang'=\{\mathbf{T}'\}$ for two distinct unit decomposition $\mathbf{T}$ and $\mathbf{T}'$ of a digraph $H$
then $\lang_{{\mathcal{G}}}=\lang_{{\mathcal{G}}}' = \{ [H] \}$ but $\lang\cap \lang' = \emptyset$!
Proposition \ref{proposition:OliveTreeDecompositionUnitDecomposition} below says that any olive-tree decomposition
$\mathcal{T}$ of a digraph $G$ can be efficiently converted into a unit decomposition $\mathbf{T}$ of $G$
that is compatible with $\mathcal{T}$. Note that there may be several unit decompositions of $G$ compatible
with $\mathcal{T}$. In the proof of Proposition \ref{proposition:OliveTreeDecompositionUnitDecomposition}
we provide an algorithm for computing one of these unit decompositions.
\begin{proposition}
\label{proposition:OliveTreeDecompositionUnitDecomposition}
Let $\mathcal{T}$ be an olive-tree decomposition of a digraph $G=(V,E,\rho,\xi)$ of width $q=w(\mathcal{T})$.
Then one can construct in time $O(|\mathcal{T}|\cdot |E|)$ a normalized unit decomposition $\mathbf{T}$ over
$\mathbold{\Sigma}(q,\Gamma_1,\Gamma_2)$ compatible with $\mathcal{T}$.
\end{proposition}
\begin{proof}
Let $\mathcal{T} = (N,F,\mathfrak{m})$ be an olive-tree decomposition of $G=(V,E,\rho,\xi)$. First we tag
each edge $e\in G$ with a number $\tau(e)\in \{1,...,|E|\}$ in such a way that no two edges
are tagged with the same number. We will construct a non-normalized unit decomposition $\mathbf{T}'$ over $\mathbold{\Sigma}(q,|E|,\Gamma_1,\Gamma_2)$
such that $\composedTprime\simeq G$. A normalized unit decomposition $\mathbf{T}$ over $\mathbold{\Sigma}(q,\Gamma_1,\Gamma_2)$ such that
$\composedT\simeq G$ can be obtained from $\mathbf{T}'$ by an application of the normalizing projection
${\mathbold{\eta}}_{q,|E|}:\mathbold{\Sigma}(q,|E|,\Gamma_1,\Gamma_2)\rightarrow \mathbold{\Sigma}(q,\Gamma_1,\Gamma_2)$.
To construct $\mathbf{T}'$ it is enough to specify the slice $\mathbf{T}'[p]$ for each position $p\in \mathit{Pos}(\mathbf{T}') = N$.
Instead of specifying each such slice $\mathbf{T}'[p]$ separately we will proceed in a more intuitive way. Namely, we will first
define which unit slices of $\mathbf{T}'$ have a center vertex, and subsequently, for each edge $e$ in $G$ and each
$p\in \mathit{Pos}(\mathbf{T}')$ we will specify which sliced part of $e$ (if any) belongs to $\mathbf{T}'[p]$. The first part
is easy. A slice $\mathbf{T}'[p]$ has a center vertex if and only if some vertex of $G$ is mapped by $\mathfrak{m}$ to
the position $p\in N$ in the olive-tree decomposition $\mathcal{T}$. For simplicity, let $v_p$ be the
vertex of $G$ for which $\mathfrak{m}(v_p) = p$. We label the center vertex of $\mathbf{T}'[p]$ with the same label
as the vertex $v_p$ in $G$. For each edge $e\in E$ with source $e^s = v_p$ and target $e^t=v_{p'}$
we create a sliced edge sequence $K\equiv(p_1,a_{1},e_1,b_{1})(p_2,a_{2},e_2,b_{2})...(p_n,a_{n},e_{n},b_{n})$ where $p_1=p$, $p_2=p'$,
$p_1p_2...p_n$ is the unique minimum path from $p$ to $p'$ in the tree $(N,F)$. For each $i\in \{1,...,n\}$,
the vertices $a_{i}$ and $b_{i}$ and the edge $e_{i}$ belong to the slice $\mathbf{T}'[p_i]$. The vertex $a_{1}$ is the center vertex
of $\mathbf{T}'[p_1]$ and $b_{n}$ is the center vertex of $\mathbf{T}'[p_n]$. For each $i\in \{1,...,n-1\}$, if $p_{i+1} = p_ij$ then
$b_{i}$ is the vertex $[j,\tau(e)]$ at the $j$-th in-frontier of $\mathbf{T}'[p_i]$ and $a_{{i+1}}$ is the vertex $[0,\tau(e)]$
at the out-frontier of $\mathbf{T}'[p_i]$. On the other hand, if $p_{i}=p_{i+1}j$ then $b_{i} = [0,\tau(e)]$ and $a_{{i+1}}=[j,\tau(e)]$.
Finally we label each edge $e_i$ of the sliced edge sequence $K$ with the same label as the edge $e$ in $G$.
One can readily check that the sequence $K$ defined in this way is indeed a sliced edge sequence, and therefore that
$\composedTprime = G$. As a final step, we obtain the unit decomposition $\mathbf{T}$ by an application of
the normalizing projection ${\mathbold{\eta}}_{q,|E|}$ to $\mathbf{T}'$. In other words, $\mathbf{T} = {\mathbold{\eta}}_{q,|E|}(\mathbf{T}')$.
\end{proof}
\subsection{Counting Subgraphs via $z$-Saturated Slice Languages}
\label{subsection:CountingSubgraphsViaZSaturatedSliceLanguages}
In this Subsection we will introduce the main technical tool of this paper. We will
show that given a $z$-saturated tree-automaton ${\mathcal{A}}$ representing digraphs that are the union of
$k$ paths, and a unit decomposition $\mathbf{T}$ of tree-zig-zag number $z$, one can count in polynomial time the number of subgraphs of $\composedT$
that are isomorphic to some digraph in $\lang_{{\mathcal{G}}}({\mathcal{A}})$. The proof will proceed in two steps.
First, we will show that from a normalized unit decomposition $\mathbf{T}$ one can construct a (non-normalized) deterministic slice tree-automaton ${\mathcal{A}}(\mathbf{T},k\cdot z)$
whose slice language $\lang({\mathcal{A}}(\mathbf{T},k\cdot z))$ consists of all sub-decompositions of $\mathbf{T}$ of width at most $k\cdot z$.
Each such sub-decomposition of $\mathbf{T}$ unequivocally identifies a subgraph of $\composedT$.
As a partial converse, each subgraph of $\composedT$ that is the union of $k$ directed paths has a representative unit decomposition in
$\lang({\mathcal{A}}(\mathbf{T},k\cdot z))$. Note that $\lang({\mathcal{A}}(\mathbf{T},k\cdot z))$ still
may contain unit decompositions of digraphs that are not the union of $k$ directed paths. However these undesired unit
decompositions are irrelevant, since they will be eliminated in the next step. In our second step, we will
show that the intersection ${\mathcal{A}}\cap {\mathcal{A}}(\mathbf{T},k\cdot z)$
is a deterministic slice tree-automaton whose graph language consists precisely of the subgraphs of
$\composedT$ that are isomorphic to some digraph in $\lang_{{\mathcal{G}}}({\mathcal{A}})$. Note that
${\mathcal{A}}\cap {\mathcal{A}}(\mathbf{T},k\cdot z)$ accepts a finite number of terms, and that the depth
of each such accepted term is equal to the depth of $\mathbf{T}$.
At this point, the problem of counting subgraphs of $\composedT$ that are isomorphic to digraphs in $\lang_{{\mathcal{G}}}({\mathcal{A}})$ boils down
to counting the number of terms accepted by ${\mathcal{A}} \cap {\mathcal{A}}(\mathbf{T},k\cdot z)$ in depth $\mathit{depth}(\mathbf{T})$. We can count
these terms in polynomial time using Lemma \ref{lemma:CountingTrees}.
Lemma \ref{lemma:SubgraphsC} below says that given any unit decomposition $\mathbf{T}$ of width $q$, and any $c\leq q$, one can
construct a slice tree-automaton whose slice language consists of all sub-decompositions of $\mathbf{T}$ of width at most
$c$.
\begin{lemma}
\label{lemma:SubgraphsC}
Let $\mathbf{T}$ be a normalized unit decomposition in $\lang(\mathbold{\Sigma}(q,\Gamma_1,\Gamma_2))$ and let
$c\leq q$. Then one may construct in time $|\mathbf{T}|\cdot q^{O(c)}$ a slice tree-automaton ${\mathcal{A}}(\mathbf{T},c)$ over
$\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)$ with $|\mathbf{T}|\cdot q^{O(c)}$ states satisfying
the following properties.
\begin{enumerate}
\item \label{SubgraphsC-item1} ${\mathcal{A}}(\mathbf{T},c)$ is deterministic.
\item \label{SubgraphsC-item2} $\lang({\mathcal{A}}(\mathbf{T},c)) = \{\mathbf{T}'\in
\lang(\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)) \;|\;\mbox{$\mathbf{T}'$ is a sub-decomposition of $\mathbf{T}$}\}$
\end{enumerate}
\end{lemma}
\begin{proof}
Let $\mathbf{T}$ be a unit decomposition in $\lang(\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2))$. We will construct a slice
tree-automaton ${\mathcal{A}} = {\mathcal{A}}(\mathbf{T},c) = (Q,\mathbold{\Sigma},Q_F,\Delta)$ over $\mathbold{\Sigma} = \mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)$ whose slice language consists of all sub-decompositions of $\mathbf{T}$ of width at most $c$. The set of states $Q$ has one state ${\mathfrak{q}}_{p,\mathbf{I}}$ for each position
$p\in \mathit{Pos}(\mathbf{T})$ and each identity slice $\mathbf{I}$ in $\mathbold{\Sigma}_1(c,q,\Gamma_1,\Gamma_2)$. We note that
since the empty slice of arity one, ${\bm{\varepsilon}}_1$, is also an identity slice, the state ${\mathfrak{q}}_{p,{\bm{\varepsilon}}_1}$
belongs to $Q$ for each $p\in \mathit{Pos}(\mathbf{T})$.
The set of final states is the singleton $Q_F = \{{\mathfrak{q}}_{\lambda,{\bm{\varepsilon}}_1}\}$.
Now we will construct the transition relation
$\Delta = \Delta_0 \cup \Delta_1 \cup \Delta_2$. First we recall that for each position $p\in \mathit{Pos}(\mathbf{T})$, if
$\mathbf{T}[p]$ is a slice of arity $r$ then any sub-slice ${\mathbf{S}}$ of $\mathbf{T}[p]$ has also arity $r$.
Recall that if ${\mathbf{S}}$ is a slice, then $\mathbf{I}({\mathbf{S}})$ denotes the unique identity slice such that ${\mathbf{S}}$ can be
glued to $\mathbf{I}({\mathbf{S}})$.
For each $r\in \{0,1,2\}$, and each position $p\in \mathit{Pos}(\mathbf{T})$ such that $\mathbf{T}[p]$ has arity $r$, the relation $\Delta_r$ has
one transition $({\mathfrak{q}}_{p1,\mathbf{I}_1},..., {\mathfrak{q}}_{pr,\mathbf{I}_r}, {\mathbf{S}}, {\mathfrak{q}}_{p,\mathbf{I}({\mathbf{S}})})$ for
each sub-slice ${\mathbf{S}}$ of $\mathbf{T}[p]$ satisfying the following conditions:
\begin{enumerate}[(i)]
\item \label{condition:one} ${\mathbf{S}}$ has width at most $c$, i.e., ${\mathbf{S}}\in \mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)$,
\item \label{condition:two} for each $j\in \{1,...,r\}$, $\mathbf{I}_j$ is the unique identity slice that can be glued to ${\mathbf{S}}$ at frontier $j$.
\end{enumerate}
To see that ${\mathcal{A}}$ is deterministic, note that
for each slice ${\mathbf{S}}$ there is a unique identity slice $\mathbf{I}$ such that ${\mathbf{S}}$ can be glued to $\mathbf{I}$.
Therefore, for each tuple $({\mathfrak{q}}_{p1,\mathbf{I}_1},{\mathfrak{q}}_{p2,\mathbf{I}_2}, ..., {\mathfrak{q}}_{pj,\mathbf{I}_r}, {\mathbf{S}})$ there is a unique state
${\mathfrak{q}}_{p,\mathbf{I}}$ such that $({\mathfrak{q}}_{p1,\mathbf{I}_1},{\mathfrak{q}}_{p2,\mathbf{I}_2}, ..., {\mathfrak{q}}_{pr,\mathbf{I}_r}, {\mathbf{S}},{\mathfrak{q}}_{p,\mathbf{I}})$ belongs
to $\Delta_r$. This also implies that each term $\mathbf{T}'$ accepted by ${\mathcal{A}}$
is a unit pre-decomposition, i.e., each two consecutive positions of $\mathbf{T}'$ can be glued. Additionally, the fact that ${\mathfrak{q}}_{\lambda,{\bm{\varepsilon}}_1}$ is the unique accepting state of ${\mathcal{A}}$
implies that the slice at the root of $\mathbf{T}'$ has empty-frontier, since this slice must be glueable to ${\bm{\varepsilon}}_1$.
Therefore each such term $\mathbf{T}'$ is a unit decomposition.
It remains to show that a unit decomposition $\mathbf{T}'$ is accepted by ${\mathcal{A}}$ if and only if $\mathbf{T}'$ is a sub-decomposition of $\mathbf{T}$ of
width at most $c$.
\begin{enumerate}
\item (if direction) Let $\mathbf{T}'$ be a sub-decomposition of $\mathbf{T}$ of width at most $c$. We claim that for each position
$p\in \mathit{Pos}(\mathbf{T}') = \mathit{Pos}(\mathbf{T})$
the subterm $\mathbf{T}'|_p$ of $\mathbf{T}'$ rooted at $p$ reaches the state ${\mathfrak{q}}_{p,\mathbf{I}(\mathbf{T}'[p])}$. This claim implies in particular
that the whole term $\mathbf{T}' = \mathbf{T}'|_{\lambda}$ reaches the unique accepting state ${\mathfrak{q}}_{\lambda,{\bm{\varepsilon}}_1}$.
The proof is by induction on the height of the position $p$.
In the base case, $p$ is a leaf of the set $\mathit{Pos}(\mathbf{T})$.
In this case, the slice $\mathbf{T}[p]$ has arity zero, and thus the sub-slice $\mathbf{T}'[p]$ has also arity zero. By the construction
of the transition relation $\Delta_0$ given above, the transition $(\mathbf{T}'[p], {\mathfrak{q}}_{p,\mathbf{I}(\mathbf{T}'[p])})$ belongs to
$\Delta_0$ and thus $\mathbf{T}'|_p$ reaches the state ${\mathfrak{q}}_{p,\mathbf{I}(\mathbf{T}'[p])}$.
Now assume by induction that the claim is valid for every position $p'$ of height $h$.
Let $p$ be a position in $\mathit{Pos}(\mathbf{T})$ of height $h+1$ with children are $p1,...,pr$
for some $r\in \{1,2\}$. By the induction hypothesis, for each $i\in \{1,...,r\}$ the term $\mathbf{T}'|_{pi}$ reaches the
state ${\mathfrak{q}}_{pi,\mathbf{I}(\mathbf{T}'[pi])}$. By the definition of the transition relation $\Delta_r$, we have that the transition
$({\mathfrak{q}}_{p1,\mathbf{I}(\mathbf{T}'[p1])}, ...,{\mathfrak{q}}_{pr,\mathbf{I}(\mathbf{T}'[pr])},\mathbf{T}'[p], {\mathfrak{q}}_{p,\mathbf{I}(\mathbf{T}'[p])})$ belongs to $\Delta_r$,
and thus $\mathbf{T}'|_p$ reaches the state ${\mathfrak{q}}_{p,\mathbf{I}(\mathbf{T}'[p])}$. This proves our claim.
\item (only if direction) For the converse, let $\mathbf{T}'$ be a unit decomposition accepted by ${\mathcal{A}}$.
We will prove that $\mathbf{T}'$ is a sub-decomposition of $\mathbf{T}$ by showing that
$\mathit{Pos}(\mathbf{T}) = \mathit{Pos}(\mathbf{T}')$ and that for each position $p\in \mathit{Pos}(\mathbf{T}')$, $\mathbf{T}'[p]$ is a sub-slice of
$\mathbf{T}[p]$.
We claim that for each $p\in \mathit{Pos}(\mathbf{T}')$, the subterm $\mathbf{T}'|_p$ of $\mathbf{T}'$ rooted at $p$ reaches the state
${\mathfrak{q}}_{p,\mathbf{I}(\mathbf{T}'[p])}$. By the construction of the transition relation $\Delta$ this claim implies both
that $\mathbf{T}'[p]$ is a sub-slice of $\mathbf{T}[p]$ for each $p\in \mathit{Pos}(\mathbf{T}')$ and that $\mathit{Pos}(\mathbf{T}')=\mathit{Pos}(\mathbf{T})$, as
desired.
The proof of this claim is by induction on the depth of $p$. In the base case, $p=\lambda$. In this
case, $\mathbf{T}'|_{\lambda} = \mathbf{T}'$ reaches the unique accepting state ${\mathfrak{q}}_{\lambda,{\bm{\varepsilon}}_1} = {\mathfrak{q}}_{\lambda,\mathbf{I}(\mathbf{T}'[\lambda])}$.
Now assume that for every position $p$ of depth at most $d$, the term $\mathbf{T}'|_{p}$ reaches the state ${\mathfrak{q}}_{p,\mathbf{I}(\mathbf{T}'[p])}$.
We will show that the claim holds for every position $p$ of depth $d+1$.
Let $p\in \mathit{Pos}(\mathbf{T}')$ be a position of depth $d$. By the induction hypothesis, $\mathbf{T}'|_{p}$ reaches the state ${\mathfrak{q}}_{p,\mathbf{I}(\mathbf{T}'[p])}$.
Let $\mathbf{T}'[p]$ have arity $r$ for some $r\in \{1,2\}$. Since $\mathbf{T}'|_{p}$ reaches ${\mathfrak{q}}_{p,\mathbf{I}(\mathbf{T}'[p])}$, there exist states $\mathfrak{q}_{1},...,\mathfrak{q}_r$ such that
the transition $(\mathfrak{q}_1,...,\mathfrak{q}_r,\mathbf{T}'[p],{\mathfrak{q}}_{p,\mathbf{I}(\mathbf{T}'[p])})$ belongs to $\Delta$ and $\mathbf{T}'|_{pj}$ reaches
$\mathfrak{q}_j$ for each $j\in \{1,...,r\}$.
By the definition of $\Delta$, for each $j\in \{1,...,r\}$, $\mathfrak{q}_{j} = {\mathfrak{q}}_{pj,\mathbf{I}_j}$ where $\mathbf{I}_{j}$ is the
unique identity slice that can be glued to $\mathbf{T}'[p]$ at frontier $j$.
Since $\mathbf{T}'$ is a unit decomposition, $\mathbf{T}'[pj]$ can be glued to $\mathbf{T}'[p]$ at frontier $j$. Therefore $\mathbf{I}_j = \mathbf{I}(\mathbf{T}'[pj])$.
Thus $\mathfrak{q}_j = {\mathfrak{q}}_{pj,\mathbf{I}(\mathbf{T}'[pj])}$ and $\mathbf{T}'|_{pj}$ reaches ${\mathfrak{q}}_{pj,\mathbf{I}(\mathbf{T}'[pj])}$.
This proves our inductive step.
\end{enumerate}
\end{proof}
Proposition \ref{proposition:UnionZigZag} below establishes a relation between the minimum number of paths necessary to
cover all edges and vertices of a digraph $H$, and the width of a unit decomposition of $H$ of tree-zig-zag number
$z$. Intuitively, if $\mathbf{T}$ is a unit decomposition of tree-zig-zag number $z$ of a digraph $H$, then each directed simple path $\mathfrak{p}$ in $H$
crosses each frontier of a slice in $\mathbf{T}$ at most $z$ times. Therefore, if $H$ is the union of $k$ directed simple paths $\mathfrak{p}_1,...,\mathfrak{p}_k$,
then all such paths together cross each frontier of each slice of $\mathbf{T}$ at most $k\cdot z$ times.
\begin{proposition}
\label{proposition:UnionZigZag}
Let $H$ be a digraph that is the union of $k$-paths and $\mathbold{\Sigma}$ be a slice alphabet.
Then any unit decomposition $\mathbf{T}\in \lang(\mathbold{\Sigma})$ of $H$ of tree-zig-zag number $z$ has width at most $k\cdot z$.
\end{proposition}
\begin{proof}
Let $H = (V,E)$ be a digraph that is the union of $k$ directed paths $\mathfrak{p}_1, ..., \mathfrak{p}_k$ where for each $i\in \{1,...,k\}$, $\mathfrak{p}_i = (V_{\mathfrak{p}_i},E_{\mathfrak{p}_i})$.
Let $\mathbf{T}\in \lang(\mathbold{\Sigma})$ be a unit decomposition of tree-zig-zag number $z$ of a digraph $H$. Then $\mathbf{T}$ is compatible
with an olive-tree decomposition $\mathcal{T} = (N,F,\mathfrak{m})$ of tree-zig-zag number $z$. Additionally, the width
of $\mathbf{T}$ is equal to the width of $\mathcal{T}$. Since $\mathcal{T}$ has
tree-zig-zag number $z$, for each position $p\in N$ and each $i\in \{1,...,k\}$
we have that $$|E(V(p,\mathcal{T}), V\backslash V(p,\mathcal{T}))\;\cap \; E_{\mathfrak{p}_i}| \leq z.$$
This implies that $$|E(V(p,\mathcal{T}), V\backslash V(p,\mathcal{T}))\; \cap\; \bigcup_{i=1}^k E_{\mathfrak{p}_i}\;| \leq k\cdot z.$$
But since $E = \cup_{i=1}^k E_{\mathfrak{p}_i}$, we have that $|E(V(p,\mathcal{T}),V\backslash V(p,\mathcal{T}))| \leq k\cdot z$.
Thus $\mathcal{T}$ has width at most $k\cdot z$, implying in this way that $\mathbf{T}$ has also width
at most $k\cdot z$.
\end{proof}
Next we state the main lemma of this section. Intuitively Lemma \ref{lemma:IntersectionSubdecompositions} below says that if $\mathbf{T}$ is a unit decomposition
of tree-zig-zag number $z$ of a digraph $G$, and if ${\mathcal{A}}$ is a $z$-saturated tree-automaton representing only digraphs that are
the union of $k$ directed paths, then the slice language $\lang({\mathcal{A}}(\mathbf{T},k\cdot z)\cap {\mathcal{A}})$ has precisely one unit decomposition
for each subgraph of $\composedT$ that is isomorphic to a digraph in $\lang_{{\mathcal{G}}}({\mathcal{A}})$. In this sense, the problem of counting
the number of subgraphs of $\composedT$ that are isomorphic to a digraph in $\lang_{{\mathcal{G}}}({\mathcal{A}})$ boils down to counting the number
of unit-decompositions in $\lang({\mathcal{A}}(\mathbf{T},k\cdot z) \cap {\mathcal{A}})$. This counting step will be detailed in Theorem
\ref{theorem:CountingSubgraphs}.
\begin{lemma}
\label{lemma:IntersectionSubdecompositions}
Let $\mathbf{T}$ be a unit decomposition of tree-zig-zag number $z$ over $\mathbold{\Sigma}(q,\Gamma_1,\Gamma_2)$. Let ${\mathcal{A}}$ be a deterministic
$z$-saturated slice automaton over $\mathbold{\Sigma}(k\cdot z,q,\Gamma_1,\Gamma_2)$ such that each digraph in $\lang_{{\mathcal{G}}}({\mathcal{A}})$
is the union of $k$ directed paths.
\begin{enumerate}
\item \label{lemma:IntersectionSubdecompositions-One} The tree-automaton ${\mathcal{A}}(\mathbf{T},k\cdot z) \cap {\mathcal{A}}$ is deterministic and all its
accepted unit decompositions have depth at most $\mathit{depth}(\mathbf{T})$.
\item \label{lemma:IntersectionSubdecompositions-Two} $H$ is a subgraph of $\composedT$ such that $[H] \in \lang_{{\mathcal{G}}}({\mathcal{A}})$ if and only if
there exists a unit decomposition $\mathbf{T}' \in \lang({\mathcal{A}}(\mathbf{T},k\cdot z) \cap {\mathcal{A}})$ such that $\composedTprime=H$.
\end{enumerate}
\end{lemma}
\begin{proof}
Item \ref{lemma:IntersectionSubdecompositions-One} is straightforward. The automaton ${\mathcal{A}}(\mathbf{T},k\cdot z) \cap {\mathcal{A}}$ is
deterministic because both ${\mathcal{A}}(\mathbf{T},k\cdot z)$ and ${\mathcal{A}}$ are deterministic
(Lemma \ref{lemma:PropertiesOfTreeAutomata}.\ref{lemma:PropertiesOfTreeAutomata:UnionIntersection}). Since by construction the construction
of ${\mathcal{A}}(\mathbf{T},k\cdot z)$, all unit decompositions
accepted by ${\mathcal{A}}(\mathbf{T},k\cdot z)$ have depth $\mathit{depth}(\mathbf{T})$, we have that all unit decompositions accepted by
${\mathcal{A}}(\mathbf{T},k\cdot z)\cap {\mathcal{A}}$ also have depth $\mathit{depth}(\mathbf{T})$.
Now we proceed to prove item \ref{lemma:IntersectionSubdecompositions-Two}.
\begin{enumerate}[(a)]
\item (if direction) Let $\mathbf{T}'$ be a unit decomposition in $\lang({\mathcal{A}}(\mathbf{T},k\cdot z)\cap {\mathcal{A}})$ such
that $\composedTprime = H$. Since $\mathbf{T}'\in \lang({\mathcal{A}}(\mathbf{T},k\cdot z))$, by Lemma \ref{lemma:SubgraphsC},
$\mathbf{T}'$ is a sub-decomposition of $\mathbf{T}$. Therefore $H$ is a subgraph of $\composedT$. Additionally,
since $\mathbf{T}'\in \lang({\mathcal{A}})$, $[H]\in \lang_{{\mathcal{G}}}({\mathcal{A}})$.
\item (only if direction) Let $H$ be a subgraph of $\composedT$ such that $[H]\in \lang_{{\mathcal{G}}}({\mathcal{A}})$.
Since $H$ is a subgraph of $\composedT$, there is a sub-decomposition $\mathbf{T}'$ of $\mathbf{T}$ such that $\composedTprime = H$.
We will show that $\mathbf{T}'$ belongs to $\lang({\mathcal{A}}(\mathbf{T},k\cdot z)\cap {\mathcal{A}})$.
Since $\mathbf{T}$ has tree-zig-zag number $z$, $\mathbf{T}'$ has tree-zig-zag number at most $z$. Now, since
$H\in \lang_{{\mathcal{G}}}({\mathcal{A}})$, we have that $H$ is the union of $k$ directed paths. By Proposition
\ref{proposition:UnionZigZag}, each unit decomposition of $H$ of tree-zig-zag number at most $z$ has width at most
$k\cdot z$. Thus $\mathbf{T}'$ has width at most $k\cdot z$. Finally, since ${\mathcal{A}}$ is $z$-saturated with
respect to $\mathbold{\Sigma}(k\cdot z,q,\Gamma_1,\Gamma_2)$ we have that $\mathbf{T}'$ belongs
to $\lang({\mathcal{A}})$. Thus $\mathbf{T}'\in \lang({\mathcal{A}}(\mathbf{T},k\cdot z) \cap {\mathcal{A}})$.
\end{enumerate}
\vspace{-3pt}
\end{proof}
The next Theorem is the main application for Lemma \ref{lemma:IntersectionSubdecompositions}. Intuitively, given a unit decomposition $\mathbf{T}$
of tree-zig-zag number $z$, and a $z$-saturated tree-automaton ${\mathcal{A}}$ representing only digraphs that are the union of $k$ directed paths, where
$z$ and $k$ are constants, one can count in polynomial time the number
of subgraphs of $\composedT$ that are isomorphic to some digraph in $\lang({\mathcal{A}})$. The idea is that Lemma \ref{lemma:IntersectionSubdecompositions}
allow us to reduce this counting problem to the problem of counting the number of accepted unit decompositions in the
tree-automaton ${\mathcal{A}}(\mathbf{T},k\cdot z) \cap {\mathcal{A}}$.
\begin{theorem}[Slice Theoretic Metatheorem]
\label{theorem:CountingSubgraphs}
Let $\mathbf{T}$ be a unit decomposition over $\mathbold{\Sigma}(q,\Gamma_1,\Gamma_2)$ and let ${\mathcal{A}}$ be a
deterministic $z$-saturated slice tree-automaton over $\mathbold{\Sigma}(k\cdot z,q,\Gamma_1,\Gamma_2)$ satisfying the
property that each digraph in $\lang_{{\mathcal{G}}}({\mathcal{A}})$ is the union of $k$ directed paths. Then one can count in
time $|\mathbf{T}|^{O(1)}\cdot q^{O(k\cdot z)}\cdot |{\mathcal{A}}|^{O(1)}$ the number of subgraphs of $\composedT$ that are isomorphic
to a digraph in $\lang_{{\mathcal{G}}}({\mathcal{A}})$.
\end{theorem}
\begin{proof}
First, we construct in time $|\mathbf{T}|\cdot q^{O(k\cdot z)}$ the tree-automaton ${\mathcal{A}}(\mathbf{T},k\cdot z)$ of
Lemma \ref{lemma:SubgraphsC} whose slice language consists of the set of all sub-decompositions of $\mathbf{T}$ of
width at most $k\cdot z$. Subsequently, using Lemma \ref{lemma:PropertiesOfTreeAutomata}.\ref{lemma:PropertiesOfTreeAutomata:UnionIntersection},
we construct the tree-automaton ${\mathcal{A}}(\mathbf{T},k\cdot z) \cap {\mathcal{A}}$ in time $|\mathbf{T}| \cdot q^{O(k\cdot z)} \cdot |{\mathcal{A}}|$.
By Lemma \ref{lemma:IntersectionSubdecompositions} a subgraph of $\composedT$ is isomorphic to some digraph in
$\lang_{{\mathcal{G}}}({\mathcal{A}})$ if and only if there exists a sub-decomposition $\mathbf{T}'$ of $\mathbf{T}$ in $\lang({\mathcal{A}}(\mathbf{T},k\cdot z)\cap {\mathcal{A}})$
for which $\composedTprime = H$. Therefore, counting the subgraphs of $\composedT$ that are isomorphic to some
digraph in $\lang_{{\mathcal{G}}}({\mathcal{A}})$ amounts to counting the number of unit decompositions of depth $\mathit{depth}(\mathbf{T})$ in
$\lang({\mathcal{A}}(\mathbf{T},k\cdot z)\cap {\mathcal{A}})$.
In other words, this problem is equivalent to the problem of counting the number of terms accepted by ${\mathcal{A}}\cap {\mathcal{A}}(\mathbf{T},k\cdot z)$
in depth $\mathit{depth}(\mathbf{T})$. Since $\mathit{depth}(\mathbf{T})\leq |\mathbf{T}|$, by Lemma \ref{lemma:CountingTrees}, this counting process can be done in time
$|\mathbf{T}|^{O(1)} \cdot q^{O(k\cdot z)} \cdot |{\mathcal{A}}|^{O(1)}$.
\end{proof}
Next, in Section \ref{section:MSOandTreeSliceLanguages} we will show that for each {$\mbox{MSO}_2$\;} sentence
$\varphi$ and each $k,z\in \N$, one can construct a $z$-saturated tree-automaton ${\mathcal{A}}(\varphi,k,z)$
representing the set of all digraphs that at the same time are the union of $k$ directed paths and satisfy $\varphi$. Subsequently,
in Section \ref{section:ProofOfMainTheorem} we will show how to restrict ${\mathcal{A}}(\varphi,k,z)$ into a
$z$-saturated slice tree-automaton ${\mathcal{A}}(\varphi,k,z,l,\alpha)$ representing only the digraphs in $\lang_{{\mathcal{G}}}({\mathcal{A}}(\varphi,k,z))$
which have a prescribed number $l$ of vertices and a prescribed weight $a\in \Omega$. The proof of Theorem \ref{theorem:MainTheoremDirectedTreewidth}
will follow by plugging the tree-automaton ${\mathcal{A}}(\varphi,k,z,l,\alpha)$ into Theorem \ref{theorem:CountingSubgraphs}.
\section{{$\mbox{MSO}_2$\;} Logic and Tree Slice Languages}
\label{section:MSOandTreeSliceLanguages}
Defining interesting families of digraphs via $z$-saturated tree-automata is a difficult task.
The difficulty relies on the fact that to construct a $z$-saturated tree-automaton we have to make sure
that for each digraph $H$ in the graph language $\lang_{{\mathcal{G}}}({\mathcal{A}})$,
all unit decompositions of $H$ with tree-zig-zag number at most $z$ are in the
slice language $\lang({\mathcal{A}})$. In this section we will introduce a suitable way of circumventing
this difficulty by using the monadic second order logic of graphs with edge set quantifications, or {$\mbox{MSO}_2$\;} logic
for short. This logic, which extends first order logic by incorporating quantification over sets of vertices and over sets of edges,
is able to express a large variety of natural graph properties \cite{CourcelleEngelfriet2012}.
We will show that for any {$\mbox{MSO}_2$\;} sentence $\varphi$ and any $k,z\in \N$ we can automatically construct a $z$-saturated slice tree-automaton ${\mathcal{A}}(\varphi,k,z)$ whose graph language consists of
all digraphs that at the same time satisfy $\varphi$ and are the union of $k$ directed paths
(Theorem \ref{theorem:MonadicSliceTreeAutomataZSaturated}).
Let $\Gamma_1$ be a set of vertex labels and $\Gamma_2$ be a set of edge labels.
A $(\Gamma_1,\Gamma_2)$-labeled digraph is a relational structure $G=(V,E,s,t,\rho,\xi)$
comprising a set of vertices $V$, a set of edges $E$, source and target relations $s,t \subseteq E\times V$,
a vertex-labeling relation $\rho \subseteq V\times \Gamma_1$ and an edge-labeling relation $\xi \subseteq E\times \Gamma_2$.
The language of {$\mbox{MSO}_2$\;} logic for $(\Gamma_1,\Gamma_2)$-labeled digraphs includes the connectives $\vee,\wedge,\neg$, variables for vertices,
edges, sets of vertices and sets of edges, the quantifier $\exists$ that can be applied to these variables, and the following predicates:
\begin{enumerate}
\item $x\in X$ where $x$ is a vertex variable and $X$ a vertex set variable,
\item $y\in Y$ where $y$ is an edge variable and $Y$ an edge set variable,
\item Equality, $=$, of variables representing vertices, edges, sets of vertices and sets of edges.
\item $s(y,x)$ where $y$ is an edge variable, $x$ a vertex variable, and the interpretation is that $x$ is the source of $y$.
\item $t(y,x)$ where $y$ is an edge variable, $x$ a vertex variable, and the interpretation is that $x$ is the target $y$.
\item For each $a\in \Gamma_1$, a predicate $\rho(x,a)$ where $x$ is a vertex variable, and the interpretation is
that $x$ is a vertex labeled with $a$.
\item For each $b\in \Gamma_2$, a predicate $\xi(x,b)$ where $x$ is an edge variable, and the interpretation is
that $x$ is an edge labeled with $b$.
\end{enumerate}
Let $\mathcal{X}$ be a set of free first order variables and second order variables. An interpretation of $\mathcal{X}$
in $G$ is a function $M:\mathcal{X}\rightarrow (V\cup E) \cup (2^{V}\cup 2^{E})$ that associates with
each vertex variable $x\in \mathcal{X}$, a vertex in $V$, with each edge variable $y\in \mathcal{X}$, an edge in $E$,
with each vertex set variable $X\in \mathcal{X}$ a set of vertices and with each edge set variable $Y\in \mathcal{X}$,
a set of edges. The semantics of a formula $\varphi$ with free variables
$\mathcal{X}$ being true under interpretation $M$ is the standard one. An {$\mbox{MSO}_2$\;} {\em sentence} is an {$\mbox{MSO}_2$\;} formula without free
variables.
If $G=(V,E,s,t,\rho,\xi)$ is a $\mbox{$(\Gamma_1,\Gamma_2)$-labeled}$ digraph and $\varphi$
is an {$\mbox{MSO}_2$\;} sentence then we write $G\models \varphi$ to indicate that $G$ satisfies $\varphi$.
Let $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)= \mathbold{\Sigma}_0(c,\Gamma_1,\Gamma_2) \cup \mathbold{\Sigma}_1(c,\Gamma_1,\Gamma_2)
\cup \mathbold{\Sigma}_2(c,\Gamma_1,\Gamma_2)$ be the ranked slice alphabet defined in Section \ref{subsection:SliceAlphabets} (with $r=2$).
Lemma \ref{lemma:MonadicSliceTreeAutomata} below, which will be proved in $\mbox{Section \ref{subsection:ProofLemmaMonadicSliceTreeAutomata},}$
establishes a connection between {$\mbox{MSO}_2$\;} logic and slice tree-automata.
\begin{lemma}
\label{lemma:MonadicSliceTreeAutomata} For every {$\mbox{MSO}_2$\;} sentence $\varphi$ over $(\Gamma_1,\Gamma_2)$-labeled digraphs
and every $c\in \N$, one can construct a normalized deterministic slice tree-automaton ${\mathcal{A}}(\varphi,c)$ over
$\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$ generating the following tree slice language:
\begin{equation}
\lang({\mathcal{A}}(\varphi,c)) = \{\mathbf{T} \in \lang(\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2))\;|\; \composedT \models \varphi\}.
\end{equation}
\end{lemma}
In other words, Lemma \ref{lemma:MonadicSliceTreeAutomata} says that given an {$\mbox{MSO}_2$\;} sentence $\varphi$ and a number $c\in N$ we
can construct a slice tree-automaton whose tree slice language consists of all unit decompositions of width at most
$c$ representing a digraph that satisfies $\varphi$. Another way of interpreting $\mbox{Lemma \ref{lemma:MonadicSliceTreeAutomata}}$
is as a "sliced" version of Courcelle's celebrated model checking theorem \cite{Courcelle1990MSO}. Recall that Courcelle's theorem
states that graphs of constant {\em undirected} treewidth can be model checked in linear time against {$\mbox{MSO}_2$\;} sentences.
In analogy with Courcelle's theorem, Lemma \ref{lemma:MonadicSliceTreeAutomata} says that digraphs admitting unit decompositions of constant width can be model checked in linear time against {$\mbox{MSO}_2$\;} sentences.
In order to verify whether a digraph $G$ admitting a unit decomposition of width at most $c$ satisfies an {$\mbox{MSO}_2$\;} sentence $\varphi$,
all one needs to do is to find a normalized unit decomposition $\mathbf{T}$ of $G$ of width at most $c$,
and then check in linear time if $\mathbf{T}$ is accepted by ${\mathcal{A}}(\varphi,c)$.
In this work however we will not be interested in model checking properties on digraphs admitting unit decompositions of
constant width. Instead we will use Lemma \ref{lemma:MonadicSliceTreeAutomata} to construct $z$-saturated slice tree-automata
representing families of digraphs that are the union of $k$ directed paths and satisfy a given prescribed {$\mbox{MSO}_2$\;} property. These automata,
which will be constructed in the proof of Theorem \ref{theorem:MonadicSliceTreeAutomataZSaturated} below,
can be coupled to Theorem \ref{theorem:CountingSubgraphs} to provide a way of counting subgraphs satisfying interesting
properties on digraphs of constant tree-zig-zag number, and hence, on digraphs of constant {\em directed} treewidth.
At this point our approach differs substantially from Courcelle's theorem \cite{Courcelle1990MSO} as well as from
the approaches in \cite{ArnborgLagergrenSeese1991,CourcelleMakowskyRotics2000} in the sense
that, as mentioned in the introduction, digraphs of constant {\em directed} treewidth may have simultaneously unbounded {\em undirected}
treewidth and unbounded clique-width.
\begin{theorem}
\label{theorem:MonadicSliceTreeAutomataZSaturated} For every {$\mbox{MSO}_2$\;} sentence $\varphi$ and every $k,z\in \N$,
one can effectively construct a normalized deterministic $z$-saturated slice tree-automaton ${\mathcal{A}}(\varphi,k,z)$ over
the slice alphabet $\mathbold{\Sigma}(k\cdot z,\Gamma_1,\Gamma_2)$ representing the following graph language.
\begin{equation}
\lang_{{\mathcal{G}}}({\mathcal{A}}(\varphi,k,z)) = \{[H]\;|\; H \models \varphi,\; H\mbox{ is the union of $k$ directed paths}\}.
\end{equation}
\end{theorem}
\begin{proof}
Let $\gamma(k)$ be the {$\mbox{MSO}_2$\;} sentence that is true in a digraph $H$ if and only if $H$ is the union of $k$ directed paths.
Using Lemma \ref{lemma:MonadicSliceTreeAutomata} we construct a normalized deterministic tree-automaton
${\mathcal{A}}(k\cdot z, \varphi\wedge \gamma(k))$ generating the set of all unit decompositions $\mathbf{T}$ over
$\mathbold{\Sigma}(k\cdot z,\Gamma_1,\Gamma_2)$ for which the digraph $\composedTprime$ satisfies $\varphi\wedge \gamma(k)$.
In other words if $[H]\in \lang_{{\mathcal{G}}}({\mathcal{A}}(k\cdot z,\varphi\wedge \gamma(k)))$
then $H$ satisfies $\varphi$ and is the union of $k$ directed paths.
For the converse, suppose that $H$ is a digraph that is the union of $k$ directed paths and satisfies $\varphi$.
Then by Proposition \ref{proposition:UnionZigZag}, each unit decomposition $\mathbf{T}$ of $H$ of tree-zig-zag number at most $z$ has width
at most $k\cdot z$. Therefore, $\mathbf{T} \in \lang({\mathcal{A}}(\varphi,k,z))$. This implies not only that $[H]\in \lang_{{\mathcal{G}}}({\mathcal{A}}(\varphi,k,z))$
but also that $\lang({\mathcal{A}}(\varphi,k,z))$ is $z$-saturated.
\end{proof}
\subsection{Proof of Lemma \ref{lemma:MonadicSliceTreeAutomata}}
\label{subsection:ProofLemmaMonadicSliceTreeAutomata}
To prove Lemma \ref{lemma:MonadicSliceTreeAutomata} we need to translate each {$\mbox{MSO}_2$\;} sentence $\varphi$
expressing a property of $(\Gamma_1,\Gamma_2)$-labeled digraphs into an {$\mbox{MSO}_2$\;} sentence $\psi$ expressing a property
of unit decompositions over $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$ in such a way that
for each unit decomposition $\mathbf{T}$ over $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$, $\mathbf{T}$ satisfies $\psi$ if and
only if the digraph $\composedT$ represented by $\mathbf{T}$ satisfy $\varphi$.
With this goal in mind we need to define a new {$\mbox{MSO}_2$\;} vocabulary which is suitable for
expressing properties of unit decompositions.
The language of {$\mbox{MSO}_2$\;} logic for unit decompositions
over $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$ has the connectives $\vee,\wedge,\neg$, vertex variables, edge variables,
vertex set variables and edge set variables, the quantifier $\exists$ that can be applied to these variables,
and the following predicates:
\begin{enumerate}
\item $x\in X$ where $x$ is a vertex variable and $X$ a vertex set variable,
\item $y\in Y$ where $y$ is an edge variable and $Y$ an edge set variable,
\item Equality, $=$, of variables representing vertices, edges, sets of vertices and sets of edges.
\item $\hat{s}(y,x)$ where $y$ is an edge variable, $x$ a vertex variable, and the interpretation is that for some position $p\in \mathit{Pos}(\mathbf{T})$,
$x$ is a vertex of $\mathbf{T}[p]$, $y$ is an edge of $\mathbf{T}[p]$ and $x$ is the source of $y$.
\item $\hat{t}(y,x)$ where $y$ is an edge variable, $x$ a vertex variable, and the interpretation is that for some position $p\in \mathit{Pos}(\mathbf{T})$,
$x$ is a vertex of $\mathbf{T}[p]$, $y$ is an edge of $\mathbf{T}[p]$ and $x$ is the target of $y$.
\item For each $a\in \Gamma_1$, a predicate $\hat{\rho}(x,a)$ where $x$ is a vertex variable, and the interpretation is
that for some $p\in \mathit{Pos}(\mathbf{T})$, $x$ is a vertex of $\mathbf{T}[p]$ labeled with $a$.
\item For each $b\in \Gamma_2$, a predicate $\hat{\xi}(y,b)$ where $y$ is an edge variable, and the interpretation is
that for some $p\in \mathit{Pos}(\mathbf{T})$, $y$ is an edge of $\mathbf{T}[p]$ labeled with $b$.
\item For each $j\in \{0,1,2\}$ and each $i\in \{1,...,c\}$, the predicate $F_{j,i}(x)$ where $x$ is a vertex variable and
the interpretation is that for some position $p\in \mathit{Pos}(\mathbf{T})$, $x$ is the frontier vertex $[j,i]$ of the slice $\mathbf{T}[p]$.
\item The predicate $C(x)$ where $x$ is a vertex variable and the interpretation is that for some position $p\in \mathit{Pos}(\mathbf{T})$, $x$ is the unique
center vertex of the slice $\mathbf{T}[p]$.
\item The predicate $\mathit{Neighbors}(x_1,x_2)$ where $x_1,x_2$ are vertex variables and the interpretation is that for some position $p\in \mathit{Pos}(\mathbf{T})$
and some $j\in \{1,2\}$, $x_1$ is a vertex of $\mathbf{T}[p]$ and $x_2$ a vertex of $\mathbf{T}[pj]$.
\end{enumerate}
Recall from Section \ref{subsection:DigraphsUnitDecompositions} that if $\mathbf{T}$ is a unit decomposition then the digraph
$\composedT$ has an edge $e_K$ with source $e^{s}_K=v_{p}$ and target $e^t_K=v_{p'}$ if and only if there exists a sliced edge
sequence
$$K \equiv (p_1,a_{1},e_{1},b_{1})(p_2,a_{2},e_{2},b_{2})...(p_n,a_{n},e_{n},b_{n})$$
from $p_1=p$ to $p_n=p'$.
We note that each edge of each slice occurring in $\mathbf{T}$ belongs
to a unique sliced edge sequence. In particular, each sliced edge sequence $K$ is unequivocally determined by its first edge $e_{1}$.
Using conditions \ref{item:edgesequence1}-\ref{item:edgesequence5} of Definition \ref{definition:SlicedEdgeSequence},
it is straightforward to write an {$\mbox{MSO}_2$\;} formula $\theta(u,y,v)$ in the vocabulary of unit decompositions with free vertex variables $u,v$ and
free edge variable $y$, which is true in a unit decomposition $\mathbf{T}$ if and only if there exist positions $p$ and $p'$ in $\mathbf{T}$
such that $u$ is the center vertex of $\mathbf{T}[p]$, $y$ is an edge in $\mathbf{T}[p]$ with source $u$, $v$ is the center vertex of $\mathbf{T}[p']$, and there exists
a sliced edge sequence from $p$ to $p'$ whose first edge is $y$. Using the formula $\theta(u,y,v)$ we can map formulas in the
vocabulary of $(\Gamma_1,\Gamma_2)$-labeled digraphs to formulas in the vocabulary of unit decompositions over $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$,
as done below in Proposition \ref{proposition:TranslationFormula}.
\begin{proposition}
\label{proposition:TranslationFormula}
Let $\varphi$ be an {$\mbox{MSO}_2$\;} formula in the vocabulary of $(\Gamma_1,\Gamma_2)$-labelled graphs. There
is a formula $\psi$ in the vocabulary of unit decompositions over $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$ such
that for each unit decomposition $\mathbf{T}$ over $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$,
$\mathbf{T} \models \psi$ if and only if $\composedT \models \varphi$.
\end{proposition}
\begin{proof}
As mentioned above, using the predicates $C(x)$, $F_{j,i}(x)$, $\mathit{Neighbors}(x_1,x_2)$,
$\hat{s}(y,x)$ and $\hat{t}(y,x)$ we can define a formula $\theta(u,y,v)$ that is true in a unit decomposition
$\mathbf{T}$ if and only if there is a sliced edge sequence with first vertex $u$, first edge $y$ and
last vertex $v$.
The translation from $\varphi$ to $\psi$ proceeds as follows. We replace each occurrence of the predicate $\rho(x,a)$ in $\varphi$
with the predicate $\hat{\rho}(x,a)$, each occurrence of $\xi(x,a)$
with $\hat{\xi}(x,a)$, each occurrence of $s(y,x)$ with $(\exists v)\theta(x,y,v)$ where $v$ is a new variable
not occurring in $\varphi$, and each occurrence of $t(y,x')$ with $(\exists u) \theta(u,y,x')$, where $u$ is a new variable
not occurring in $\varphi$. Now it is straightforward to prove by induction of the structure of $\varphi$ that
for each given unit decomposition $\mathbf{T}\in \mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$, $\mathbf{T}\models \psi$ if
and only if $\composedT\models \varphi$.
\end{proof}
In the last step of the proof of Lemma \ref{lemma:MonadicSliceTreeAutomata} we will show that for each {$\mbox{MSO}_2$\;} sentence
$\psi$ in the vocabulary of unit decompositions over $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$, it is possible to
construct a slice tree-automaton ${\mathcal{A}} = {\mathcal{A}}(\psi,\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2))$ such that
$\mathbf{T} \in \lang({\mathcal{A}})$ if and only if $\mathbf{T}$ satisfies $\psi$.
Let $\mathcal{X}$ be a set of variables, and ${\mathbf{S}} \in \mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$ be a unit slice with $r$ vertices and $r'$ edges
(including the frontier vertices). We represent an interpretation of
$\mathcal{X}$ in ${\mathbf{S}}$ as a $|\mathcal{X}|\times (r+r')$ boolean matrix $I$ whose rows are
indexed by the variables in $\mathcal{X}$ and the columns are indexed by the
vertices and edges of ${\mathbf{S}}$. Intuitively, if $x$ is a vertex (edge) variable
and $u$ is a vertex (edge) in ${\mathbf{S}}$ then we set $I_{x,u}=1$ if and only if $u$ is
assigned to $x$. On the other hand, if $X$ is a vertex (edge) set variable
then $I_{X,u}=1$ if and only if $u$ belongs to the set of vertices (edges) assigned to $X$.
If $\mathbf{T}$ is a unit decomposition in $\lang(\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2))$
then an interpretation of $\mathcal{X}$ in $\mathbf{T}$ is a function $\mathcal{I}$ that
associates with each position $p\in \mathit{Pos}(\mathbf{T})$, an interpretation $\mathcal{I}(p)$ of $\mathcal{X}$ in the slice $\mathbf{T}[p]$.
We define the $\mathcal{X}$-interpreted extension of $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$ as the following set.
\begin{equation*}
\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X})=\bigcup_{{\mathbf{S}}\in \mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)} {\mathbf{S}}^{\mathcal{X}}
\end{equation*}
where for each slice ${\mathbf{S}}\in \mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$,
\begin{equation}
{\mathbf{S}}^{\mathcal{X}} = \{ ({\mathbf{S}},I)\;|\; I \mbox{ is an interpretation of } \mathcal{X} \mbox{ in } {\mathbf{S}}\}.
\end{equation}
If $\mathbf{T}$ is a unit decomposition in $\lang(\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2))$ and $\mathcal{I}$ is an
interpretation of $\mathcal{X}$ in $\mathbf{T}$ then we write $\mathbf{T}^{\mathcal{I}}$ to denote the term in
$\lang(\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X}))$ in which $\mathbf{T}^{\mathcal{I}}[p] = (\mathbf{T}[p],\mathcal{I}(p))$ for each position $p\in \mathit{Pos}(\mathbf{T})$.
We say that $\mathbf{T}^{\mathcal{I}}$ is an interpreted term.
Now we are in a position to prove Lemma \ref{lemma:MonadicSliceTreeAutomata}. For each {$\mbox{MSO}_2$\;} formula $\psi$
in the vocabulary of unit decompositions over $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$
with free variables $\mathcal{X}$ we will construct a tree-automaton ${\mathcal{A}}(\psi,\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X}))$
over $\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X})$ whose slice language
$\lang({\mathcal{A}}(\psi,\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X})))$ consists of all interpreted terms $\mathbf{T}^{\mathcal{I}}\in \lang(\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X}))$
for which $\mathbf{T} \models \psi(\mathcal{X})$ with interpretation $\mathcal{I}$. The tree-automaton
${\mathcal{A}}(\psi,\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2,\mathcal{X}))$ is constructed inductively with
respect to the structure of the formula $\psi$.
\paragraph{Base Case} In the base case the formula $\psi$ is one of the predicates
$x\in X$, $x_1=x_2$,
$C(x)$, $F_{j,i}(x)$ for $j\in \{0,1,2\}$, $\mathit{Neighbors}(x_1,x_2)$, $\hat{s}(y,x)$, $\hat{t}(y,x)$, $\hat{\rho}(x,a)$ or
$\hat{\xi}(y,b)$. Below, we describe the behavior of the tree-automaton ${\mathcal{A}} = {\mathcal{A}}(\psi,\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X}))$
when $\psi$ is each of these predicates. If $x$ is a vertex (edge) variable in $\psi$, then we say that an interpreted term
$\mathbf{T}^{\mathcal{I}}$ passes the singleton test with respect to $x$ if there exists a unique position $p\in \mathit{Pos}(\mathbf{T})$
and a unique vertex (edge) $u$ in $\mathbf{T}[p]$ such that $\mathcal{I}(p)_{x,u}=1$. We note that this condition can be easily
checked by a tree-automaton over $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2,\mathcal{X})$. Intuitively, $\mathbf{T}^{\mathcal{I}}$ passes the singleton
test with respect to $x$ if precisely one vertex (edge) of some slice of $\mathbf{T}$ is assigned to $x$.
\begin{enumerate}
\item If $\psi \equiv (x_1 = x_2)$ where $x_1$ and $x_2$ are vertex (edge) variables,
then ${\mathcal{A}}$ accepts $\mathbf{T}^{\mathcal{I}}$ if and only if $\mathbf{T}^{\mathcal{I}}$ passes the singleton test with
respect to both $x_1$ and $x_2$, and for each position $p\in \mathit{Pos}(\mathbf{T})$, and each vertex (edge)
$u\in \mathbf{T}[p]$, $\mathcal{I}(p)_{x_1,u} = \mathcal{I}(p)_{x_2,u}$.
\item If $\psi \equiv (X_1 = X_2)$ where $X_1$ and $X_2$ are vertex (edge) set variables,
then ${\mathcal{A}}$ accepts $\mathbf{T}^{\mathcal{I}}$ if and only if $\mathcal{I}(p)_{x_1,u} = \mathcal{I}(p)_{x_2,u}$
for each position $p\in \mathit{Pos}(\mathbf{T})$, and each vertex (edge) $u\in \mathbf{T}[p]$.
\item If $\psi \equiv x\in X$ where $x$ is a vertex (edge) variable and $X$ is a vertex (edge) set
variable then ${\mathcal{A}}$ accepts $\mathbf{T}^{\mathcal{I}}$ if and only if $\mathbf{T}^{\mathcal{I}}$ passes the singleton test with
respect to $x$ and for each position $p\in \mathit{Pos}(\mathbf{T})$, and each vertex (edge) $u\in \mathbf{T}[p]$,
$\mathcal{I}(p)_{x,u}=1$ implies that $\mathcal{I}(p)_{X,u}=1$.
\item If $\psi \equiv \hat{s}(y,x)$ (resp. $\psi = \hat{t}(y,x)$) then ${\mathcal{A}}$ accepts $\mathbf{T}^{\mathcal{I}}$ if and only if $\mathbf{T}^{\mathcal{I}}$ passes the
singleton test with respect to both $y$ and $x$, and there exists a position $p\in \mathit{Pos}(\mathbf{T})$, a vertex $v\in \mathbf{T}[p]$ and
an edge $e$ in $\mathbf{T}[p]$ such that $v$ is the source (target) of $e$ and $\mathcal{I}(p)_{x,v}=1$ and $\mathcal{I}(p)_{y,e}=1$.
\item If $\psi \equiv \hat{\rho}(x,a)$ then ${\mathcal{A}}$ accepts $\mathbf{T}^{\mathcal{I}}$ if and only if $\mathbf{T}^{\mathcal{I}}$ passes
the singleton test with respect to $x$ and there is a position $p\in \mathit{Pos}(\mathbf{T})$ and a vertex $v\in \mathbf{T}[p]$ such that
$\mathcal{I}(p)_{x,v}=1$ and $v$ is labeled with $a$.
\item If $\psi \equiv \hat{\xi}(y,b)$ then ${\mathcal{A}}$ accepts $\mathbf{T}^{\mathcal{I}}$ if and only if $\mathbf{T}^{\mathcal{I}}$ passes
the singleton test with respect to $y$ and there is a position $p\in \mathit{Pos}(\mathbf{T})$ and an edge $e\in \mathbf{T}[p]$ such that
$\mathcal{I}(p)_{y,e}=1$ and $e$ is labeled with $b$.
\item If $\psi \equiv C(x)$ (resp. $\psi\equiv F_{i,j}(x)$) then ${\mathcal{A}}$ accepts $\mathbf{T}^{\mathcal{I}}$ if and only if $\mathbf{T}^{\mathcal{I}}$
passes the singleton test with respect to $x$ and there exists a position $p\in \mathit{Pos}(\mathbf{T})$ and a vertex $v\in \mathbf{T}[p]$ such that
$\mathcal{I}(p)_{x,v}=1$ and $v$ is the center vertex of $\mathbf{T}[p]$ (resp. $v$ is the vertex $[j,i]$ at the $j$-th frontier of $\mathbf{T}[p]$).
\item If $\psi \equiv \mathit{Neighbors}(x_1,x_2)$ then ${\mathcal{A}}$ accepts $\mathbf{T}^{\mathcal{I}}$ if and only if $\mathbf{T}^{\mathcal{I}}$ passes
the singleton test with respect to both $x_1$ and $x_2$, and there exists a position $p\in \mathit{Pos}(\mathbf{T})$, a number $j\in \{1,2\}$,
a vertex $v\in \mathbf{T}[p]$ and a vertex $v'\in \mathbf{T}[pj]$ such that $\mathcal{I}(p)_{x_1,v}=1$ and $\mathcal{I}(pj)_{x_2,v'} = 1$.
\end{enumerate}
\paragraph{Disjunction, conjunction and negation}
The three boolean operations $\vee,\wedge,\neg$ are handled using the fact that tree-automata are effectively
closed under union, intersection and complement (Lemma \ref{lemma:PropertiesOfTreeAutomata}).
Below we let ${\mathcal{A}}(\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X}))$ be the slice tree-automaton generating the tree slice language $\lang(\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X}))$,
i.e., the set of all unit decompositions over $\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X})$.
\begin{equation}
\begin{array}{c}
{\mathcal{A}}(\psi \vee \psi',\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X})) = {\mathcal{A}}(\psi ,\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X})) \cup {\mathcal{A}}(\psi',\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X})) \\
\\
{\mathcal{A}}(\psi \wedge \psi',\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X})) = {\mathcal{A}}(\psi ,\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X})) \cap {\mathcal{A}}(\psi',\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X})) \\
\\
{\mathcal{A}}(\neg \psi,\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X})) = \overline{{\mathcal{A}}(\psi ,\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X}))} \cap
{\mathcal{A}}(\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X}))\\
\end{array}
\end{equation}
Observe that in the definition of ${\mathcal{A}}(\neg \psi,\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X}))$, the
intersection with the tree-automaton ${\mathcal{A}}(\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X}))$ guarantees
that the language generated by ${\mathcal{A}}(\neg \psi, \newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X}))$ has no
term that is not an unit decomposition.
\paragraph{Existential Quantification}
To eliminate existential quantifiers we proceed as follows: For each variable $X$,
define the slice projection $\mathit{Proj}_{X}:\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X})\rightarrow \newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X}\backslash\{X\})$ that
sends each interpreted slice $({\mathbf{S}},I)\in \newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X})$ to the interpreted slice $({\mathbf{S}},I\backslash X)$ in
the slice alphabet $\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X}\backslash\{X\})$
where $I\backslash X$ denotes the matrix $I$ with the row corresponding to the variable $X$ deleted. Subsequently,
we extend $\mathit{Proj}_{X}$ homomorphically to terms by setting $\mathit{Proj}_X(\mathbf{T})[p] = \mathit{Proj}_X(\mathbf{T}[p])$
to each position $p$ in $\mathit{Pos}(\mathbf{T})$. Finally, we extend $\mathit{Proj}_X$ to tree slice languages by applying
the projection to each term of the language.
Then we set $${\mathcal{A}}(\exists X \psi(\mathcal{X}), \newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X}\backslash\{X\})) = \mathit{Proj}_{X}({\mathcal{A}}(\psi(\mathcal{X}),\newslicealphabet(c,\vertexlabel,\edgelabel,\mathcal{X}))).$$
We note that if $\psi$ is a sentence, i.e., a formula without free variables, then by the end of this
inductive process all variables occurring in $\psi$ will have been projected. In this way, the slice language
$\lang({\mathcal{A}}(\psi,\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)))$
will consist precisely of the unit decompositions $\mathbf{T}$ over $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$
for which $\mathbf{T} \models \psi$.
To finalize the proof of Lemma \ref{lemma:MonadicSliceTreeAutomata}, let $\varphi$ be a sentence in the vocabulary of $(\Gamma_1,\Gamma_2)$-labeled digraphs.
We apply Proposition \ref{proposition:TranslationFormula} to translate $\varphi$ into a sentence $\psi$ in the vocabulary of unit decompositions
over $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$ such that $\mathbf{T}\models \psi$
if and only if $\composedT \models \varphi$. By setting ${\mathcal{A}}(\varphi,c) = {\mathcal{A}}(\psi,\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2))$
we have that ${\mathcal{A}}(\varphi,c)$ accepts $\mathbf{T}$ if and only if $\mathbf{T}\models \psi$ if and only if $\composedT\models \varphi$. This concludes
the proof of Lemma \ref{lemma:MonadicSliceTreeAutomata}.
$\square$
\section{Proof of Theorem \ref{theorem:MainTheoremDirectedTreewidth}}
\label{section:ProofOfMainTheorem}
In this section we will prove Theorem \ref{theorem:MainTheoremDirectedTreewidth}, which states that given an
{$\mbox{MSO}_2$\;} sentence $\varphi$, a digraph $G$ of directed treewidth $w$ and a number $k\in \N$ one can count in polynomial
time the number of subgraphs of $G$ that are the union of $k$ directed paths, satisfy $\varphi$, and have prescribed
size $l$ and weight $\alpha$. First, in Subsection \ref{subsection:PrescribedSize} we will show how to construct slice tree-automata
representing digraphs of a prescribed size. Subsequently, in Subsection \ref{subsection:PrescribedWeight} we will
show how to construct slice tree-automata representing digraphs of a prescribed weight. In Subsection \ref{subsection:InverseHomomorphicImage}
we will define a suitable notion of inverse homomorphic image for slice languages. Using the results in these three subsections
in conjunction with the tree-automaton ${\mathcal{A}}(\varphi,k,z)$ of $\mbox{Theorem \ref{theorem:MonadicSliceTreeAutomataZSaturated},}$
we will show, in Subsection \ref{subsection:Restriction}, how to construct a $z$-saturated slice tree-automaton ${\mathcal{A}}(\varphi,k,z,l,\alpha)$ representing all digraphs
that are the union of $k$ directed paths, satisfy $\varphi$, and have prescribed size $l$ and $\mbox{weight $\alpha$}$.
The proof of Theorem \ref{theorem:MainTheoremDirectedTreewidth}, which will be detailed in Subsection \ref{subsection:TheProof},
will follow by plugging ${\mathcal{A}}(\varphi,k,z,l,\alpha)$ into Theorem \ref{theorem:CountingSubgraphs}.
\subsection{Generating Digraphs of a Prescribed Size}
\label{subsection:PrescribedSize}
In this subsection we will show that given an arbitrary slice alphabet $\mathbold{\Sigma}$ of unit slices, and a number $l$, one can construct
a deterministic tree-automaton generating precisely the unit decompositions in $\lang(\mathbold{\Sigma})$
that give rise to digraphs with $l$ vertices. Let $\Z_m$ denote the integers with addition modulo $m$. Consider
the following weighting function $\automataweightingfunction_{\Z_m}:\mathbold{\Sigma} \rightarrow \Z_{m}$:
\begin{equation}
\label{equation:WeightingCounting}
\automataweightingfunction_{\Z_m}({\mathbf{S}})=
\left\{\begin{array}{l}
0 \mbox{ if ${\mathbf{S}}$ has empty center,}\\
1 \mbox{ if ${\mathbf{S}}$ has a center vertex.}\\
\end{array}
\right.
\end{equation}
Recall from Section \ref{subsection:WeightedTerms} that given a term $\mathbf{T}\in \mathit{Ter}(\mathbold{\Sigma})$, the
weight of $\mathbf{T}$ is defined as
\begin{equation}
\label{equation:WeightTerm}
\automataweightingfunction_{\Z_m}(\mathbf{T}) = \sum_{p\in \mathit{Pos}(\mathbf{T})} \automataweightingfunction_{\Z_m}(\mathbf{T}[p]).
\end{equation}
In other words, the weight of $\mathbf{T}$ is simply the sum of the weights of all slices occurring in $\mathbf{T}$.
One can readily check that $\mathbf{T}$ has weight
$\automataweightingfunction_{\Z_m}(\mathbf{T}) = l$ if and only there are $l\;(\mathrm{mod}\; m)$ slices in $\mathbf{T}$ with non-empty center.
In particular, if $\mathbf{T}$ is a unit decomposition in $\lang(\mathbold{\Sigma})$,
then $\automataweightingfunction_{\Z_m}(\mathbf{T})$ is the number of vertices in the digraph $\composedT$ represented by $\mathbf{T}$, modulo $m$.
\begin{observation}
\label{observation:CountingVertices}
Let $\mathbf{T}$ be a unit decomposition in $\lang(\mathbold{\Sigma})$.
Then $$\automataweightingfunction_{\Z_m}(\mathbf{T})= |\!\composedT\!|\; (\mathrm{mod}\;m).$$
\end{observation}
Recall from Lemma \ref{lemma:WeightedTerms} that if $\mathbold{\Sigma}$ is a slice alphabet,
$\automataweightingfunction:\mathbold{\Sigma} \rightarrow \Xi$ is a weighting function on $\mathbold{\Sigma}$
and $a\in \Xi$, then the automaton ${\mathcal{A}}(\mathbold{\Sigma},\automataweightingfunction,a)$
generates precisely the set of terms $\mathbf{T}\in \mathit{Ter}(\mathbold{\Sigma})$ whose weight is $\automataweightingfunction(\mathbf{T})=a$.
By setting $\Xi = \Z_m$, $\automataweightingfunction = \automataweightingfunction_{\Z_{m}}$ and $a=l$, the tree-automaton
${\mathcal{A}}(\mathbold{\Sigma},\automataweightingfunction_{\Z_m},l)$ generates the set of all terms over
$\mathbold{\Sigma}$ which have $l\;(\mathrm{mod}\;m)$ slices with non-empty center. Let ${\mathcal{A}}(\mathbold{\Sigma})$ be the
slice tree-automaton generating the set of all unit decomposition over $\mathbold{\Sigma}$.
Then for each $l\in \{0,...,m-1\}$,
\begin{equation*}
\label{equation:CountingVerticesB}
\lang({\mathcal{A}}(\mathbold{\Sigma},\automataweightingfunction_{\Z_m},l) \cap {\mathcal{A}}(\mathbold{\Sigma}))
=
\{\mathbf{T}\in \lang(\mathbold{\Sigma})\;|\; |\!\composedT\!| = l\; (\mathrm{mod}\; m)\}.
\end{equation*}
In other words ${\mathcal{A}}(\mathbold{\Sigma},\automataweightingfunction_{\Z_m},l) \cap {\mathcal{A}}(\mathbold{\Sigma})$
generates all unit decompositions over $\mathbold{\Sigma}$ whose corresponding digraph has $l\;(\mathrm{mod}\; m)$ vertices.
\subsection{Generating Digraphs of a Prescribed Weight}
\label{subsection:PrescribedWeight}
Let $G = (V,E,\rho,\xi)$ be a $(\Gamma_1,\Gamma_2)$-labeled digraph and $\graphweightingfunction:E\rightarrow \Omega$
be a function that weights the edges in $E$ with elements from a finite semigroup $\Omega$. We say that the pair
$(G,\mu)$ is a weighted digraph. Alternatively, we can view $(G,\mu)$ as the digraph $(V,E,\rho,\xi\times \graphweightingfunction)$
where $\xi\times \graphweightingfunction:E\rightarrow \Gamma_2\times \Omega$ is a function that labels each edge
$e\in E$, with the element $[\xi\times \graphweightingfunction](e) = (\xi(e),\graphweightingfunction(e))$.
In this way we consider that unit decompositions of weighted digraphs are formed with elements of the slice alphabet $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2\times \Omega)$.
We insist in having two label sets $\Gamma_2$ and $\Omega$ because while we consider that the set $\Gamma_2$ is fixed,
the set $\Omega$ may vary with the input digraph.
For a unit decomposition $\mathbf{T}$ over $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2\times \Omega)$ we
let $\graphweightingfunction(\composedT)$ be the sum of the weights of all edges in the digraph $\composedT$.
Let ${\mathbf{S}}$ be a slice in $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2\times \Omega)$ and let $E$ be the edge set
of ${\mathbf{S}}$. We denote by $E_{\mathit{out}}$ the set of all edges that have an endpoint
in the out-frontier of ${\mathbf{S}}$ and the other endpoint in the center of ${\mathbf{S}}$. We denote by $E_{\mathit{in}}$
the set of edges whose endpoints lie in distinct in-frontiers of ${\mathbf{S}}$.
Let $\automataweightingfunction_{\Omega}:\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2\times \Omega)\rightarrow \Omega$ be
a weighting function on $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2\times \Omega)$ that associates
with each slice ${\mathbf{S}}\in \mathbold{\Sigma}(c,q,\Gamma_1, \Gamma_2\times \Omega)$ the value
$$\automataweightingfunction_{\Omega}({\mathbf{S}}) = \sum_{e\in E_{\mathit{out}}}\graphweightingfunction(e) - \sum_{e\in E_{\mathit{in}}}\graphweightingfunction(e).$$
Note that edges that have only one endpoint at an in-frontier of ${\mathbf{S}}$ do not have their weights counted neither positively,
nor negatively. The weight of a unit decomposition $\mathbf{T}$ over the slice alphabet $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2\times \Omega)$ is defined as
\begin{equation}
\label{equation:Weight}
\automataweightingfunction_{\Omega}(\mathbf{T}) = \sum_{p\in \mathit{Pos}(\mathbf{T})} \automataweightingfunction_{\Omega}(\mathbf{T}[p]).
\end{equation}
The next proposition says that the weight $\automataweightingfunction_{\Omega}(\mathbf{T})$ of
$\mathbf{T}$ is equal to the weight $\graphweightingfunction(\composedT)$ of the digraph $\composedT$ represented by $\mathbf{T}$.
\begin{proposition}
\label{proposition:CountingWeight}
Let $\mathbf{T}$ be a unit decomposition in $\lang(\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2 \times \Omega))$.
Then
\begin{equation}
\label{equation:EqualityWeight}
\automataweightingfunction_{\Omega}(\mathbf{T}) = \graphweightingfunction(\composedT).
\end{equation}
\end{proposition}
\begin{proof}
Recall that each edge $e_K$ in the digraph $\composedT$ represented by $\mathbf{T}$ is specified by a sliced edge sequence
$K\equiv (p_1,a_{1},e_{1},b_{1})(p_2,a_{2},e_{2},b_{2})...(p_n,a_{n},e_{n},b_{n})$, where $e_{i}$ is
the sliced part of $e_K$ lying at slice $\mathbf{T}[p_i]$. Recall that by definition, for each $i\in \{1,...,n\}$,
the weight of $e_{i}$ in $\mathbf{T}[p_i]$ is equal to the weight of $e_K$ in $\composedT$. We claim that the overall contribution of the weights of
the edges in $K$ to the sum in Equation \ref{equation:Weight} is equal to the weight of $e_K$. This claim implies
Equation \ref{equation:EqualityWeight}. There are three cases to be considered. If $p_1$ is a descendant of $p_n$, then $e_{1}$ is the only
sliced part of $e_K$ whose weight contributes positively to the sum in Equation \ref{equation:Weight}. The
weights of all other sliced parts $e_{2},...,e_{n}$ are not counted at all. This happens because,
in this case, $e_1$ is the only edge of $K$ that has a center vertex and an out-frontier vertex as endpoints.
All other edges of $K$ have one endpoint in some in-frontier and another endpoint in the center or out-frontier, and
for this reason their weights are not counted.
Analogously, if $p_n$ is a descendant of $p_1$, then $e_{n}$ is the only
sliced part of $e_K$ that has its weight contributed positively to the sum in $\mbox{Equation \ref{equation:Weight}.}$
The weights of all other sliced parts $e_{1},...e_{{n-1}}$ are not counted at all. Finally, if neither $p_1$ is a
descendant of $p_n$, nor $p_n$ is a descendant of $p_1$, then both sliced parts $e_{1}$ and $e_{n}$ have
their weights contributed positively to the sum in Equation \ref{equation:Weight}. Nevertheless, in this case there exists some $k$, with $1<k<n$ such that
$e_{k}$ has both of its endpoints in distinct in-frontiers of $\mathbf{T}[p_k]$.
Indeed, $p_k$ is the position farthest away from the root with the property that both $p_1$ and $p_n$ are descendants of $p_k$.
Therefore we have that the weight of $e_k$ is counted negatively.
The weights of all other edges $e_2...e_{k-1}$ and $e_{k+1}...e_{n-1}$ are not counted at all, since each of these edges
have one endpoint in some in-frontier and another endpoint at an out-frontier. This proves our claim.
\end{proof}
In view of Proposition \ref{proposition:CountingWeight}, if ${\mathcal{A}}(\mathbold{\Sigma},\automataweightingfunction_{\Omega},\alpha)$ is the tree-automaton of Lemma
\ref{lemma:WeightedTerms} in which
$\automataweightingfunction = \automataweightingfunction_{\Omega}$ and $a=\alpha$, then the slice language
of ${\mathcal{A}}(\mathbold{\Sigma},\automataweightingfunction_{\Omega},\alpha) \cap {\mathcal{A}}(\mathbold{\Sigma})$
is the set of all unit decompositions over $\mathbold{\Sigma}$ which represent a digraph of weight $\alpha$.
More precisely,
\begin{equation}
\label{equation:CountingSizeWeightB}
\lang({\mathcal{A}}(\mathbold{\Sigma},\automataweightingfunction_{\Omega},\alpha) \cap {\mathcal{A}}(\mathbold{\Sigma}))
= \{\mathbf{T}\in \lang(\mathbold{\Sigma})\;|\; \graphweightingfunction(\composedT) = \alpha \}.
\end{equation}
\subsection{Inverse Homomorphic Image of Slice Languages}
\label{subsection:InverseHomomorphicImage}
Let $\mathbold{\pi}:\mathbold{\Sigma} \rightarrow \mathbold{\Sigma}'$ be a slice projection such as defined in
Section \ref{subsection:SliceProjection}. If $\lang$ is a slice language over $\mathbold{\Sigma}'$ then the inverse homomorphic
image $\mathbold{\pi}^{-1}(\lang)$, as defined in Section \ref{subsection:PropertiesOfTreeAutomata}, is not necessarily
a slice language since for a unit decomposition $\mathbf{T}\in \lang$, the inverse set $\mathbold{\pi}^{-1}(\mathbf{T})$ consisting of all terms whose image is $\mathbf{T}$
may have some terms over $\mathbold{\Sigma}$ that are not unit decompositions. To fix this we intersect $\mathbold{\pi}^{-1}(\mathbf{T})$
with the slice language $\lang(\mathbold{\Sigma})$ of all unit decompositions over $\mathbold{\Sigma}$. More precisely, if
$\mathbf{T}$ is a unit decomposition over $\mathbold{\Sigma}'$ then we define $\mathbf{inv}(\mathbold{\pi},\mathbf{T}) = \mathbold{\pi}^{-1}(\mathbf{T})\cap \lang(\mathbold{\Sigma})$.
Going further, if $\lang$ is a slice language over $\mathbold{\Sigma}'$ then
\begin{equation}
\label{equation:InverseImageUnweighting}
\mathbf{inv}(\mathbold{\pi},\lang) = \bigcup_{\mathbf{T}\in \lang} \mathbf{inv}(\mathbold{\pi},\mathbf{T}) = \mathbold{\pi}^{-1}(\lang) \cap \lang(\mathbold{\Sigma}).
\end{equation}
For instance, if ${\mathbold{\eta}}_{c,q}:\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)\rightarrow \mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$
is a normalizing projection and $\mathbf{T}$ is a unit decomposition over $\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$, then
$\mathbf{inv}({\mathbold{\eta}}_{c,q}, \mathbf{T})$ consists of all unit decompositions that are obtained from $\mathbf{T}$ by renumbering the
vertices on each frontier of each slice of $\mathbf{T}$ with numbers from $\{1,...,q\}$ in such a way that the order
in each frontier is preserved. Therefore $\mathbf{inv}({\mathbold{\eta}}_{c,q},\lang)$ is the maximal unnormalized slice language
whose image under ${\mathbold{\eta}}_{c,q}$ is $\lang$. Note that if $\lang$ is a $z$-saturated slice language over
$\mathbold{\Sigma}(c,\Gamma_1,\Gamma_2)$ then $\mathbf{inv}({\mathbold{\eta}},\lang)$ is a $z$-saturated slice language
over $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)$.
Analogously, if ${\mathbold{\zeta}}_{\Omega}:\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2\times \Omega)\rightarrow \mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)$
is an unweighting projection, and $\mathbf{T}$ is a unit decomposition over $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)$, then
$\mathbf{inv}({\mathbold{\zeta}}_{\Omega},\mathbf{T})$ consists of all unit decompositions over $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2\times \Omega)$ that
are obtained from $\mathbf{T}$ by weighting the edges of each slice in $\mathbf{T}$ with elements from $\Omega$ in such a way that gluability
of slices is preserved. Thus $\mathbf{inv}({\mathbold{\zeta}}_{\Omega},\lang)$ is a slice language consisting of all weighted versions of
unit decompositions in $\lang$. We note that if $\lang$ is a $z$-saturated slice language over $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2)$ then
$\mathbf{inv}({\mathbold{\zeta}}_{\Omega},\lang)$ is a $z$-saturated slice language over $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2\times \Omega)$.
\subsection{Restricting ${\mathcal{A}}(\varphi,k,z)$}
\label{subsection:Restriction}
In Theorem \ref{theorem:MonadicSliceTreeAutomataZSaturated} we showed that given any {$\mbox{MSO}_2$\;} sentence $\varphi$ in the
vocabulary of $\mbox{$(\Gamma_1,\Gamma_2)$-labeled}$ digraphs, and any $z,k\in \N$ one can construct a normalized
$z$-saturated slice tree-automaton ${\mathcal{A}}(\varphi,k,z)$ over the slice alphabet
$\mathbold{\Sigma}(k\cdot z,\Gamma_1,\Gamma_2)$ whose graph language $\lang_{{\mathcal{G}}}({\mathcal{A}}(\varphi,k,z))$
consists precisely of the digraphs that are the union of $k$ directed paths and satisfy $\varphi$.
In this section we show how to construct a $z$-saturated tree-automaton ${\mathcal{A}}(\varphi,k,z,l,\alpha)$
over the slice alphabet $\mathbold{\Sigma}(c,q,\Gamma_1,\Gamma_2\times \Omega)$ whose graph $\lang_{{\mathcal{G}}}({\mathcal{A}}(\varphi,k,z,l,\alpha))$
contains only the digraphs in $\lang_{{\mathcal{G}}}({\mathcal{A}}(\varphi,k,z))$ that have a prescribed size $l$ and prescribed weight
$\alpha\in \Omega$. If $\varphi$ is an {$\mbox{MSO}_2$\;} sentence in the vocabulary of $(\Gamma_1,\Gamma_2)$-labeled digraphs, $G=(V,E)$ is a
$(\Gamma_1,\Gamma_2)$-labeled digraph and $\mu:E\rightarrow \Omega$ is a weighting function, then we say that
the weighted digraph $(G,\mu)$ satisfies $\varphi$ if $G$ satisfies $\varphi$. In other words, a weighted digraph satisfies an {$\mbox{MSO}_2$\;}
sentence if its unweighted version does.
\begin{lemma}
\label{lemma:AutomatonSizeWeight}
Let $\varphi$ be an {$\mbox{MSO}_2$\;} sentence over $(\Gamma_1,\Gamma_2)$-labeled digraphs, $q,k,z,l,m\in \N$ be positive integers with $l<m$, $q\geq k\cdot z$,
and let $\alpha\in \Omega$. For some computable function $g$, one can construct in time
$g(\varphi,k,z,|\Gamma_1|,|\Gamma_2|)\cdot q^{O(k\cdot z)} \cdot |\Omega|^{O(k\cdot z)}\cdot m^{O(1)}$
a $z$-saturated slice tree-automaton ${\mathcal{A}}(\varphi,k,z,l,\alpha)$ over $\mathbold{\Sigma}(k\cdot z,q,\Gamma_1,\Gamma_2\times \Omega)$
such that
\begin{equation*}
\label{equation:AutomatonPrescribedLenghtWeight}
\begin{array}{lcr}
\lang({\mathcal{A}}(\varphi,k,z,l,\alpha)) = \{\mathbf{T}\in \mathbold{\Sigma}(k\cdot z,q,\Gamma_1,\Gamma_2\times \Omega) \; |\;
\mbox{ $\composedT\models \varphi$, $\composedT$ is the union of $k$ directed paths,} \\
\hspace{8.3cm} \mbox{ $\composedT$ has $l\;(\mathrm{mod}\;m)$ vertices, $\composedT$ has weight $\alpha$}\}.
\end{array}
\end{equation*}
\end{lemma}
\begin{proof}
Let ${\mathbold{\eta}}_{k\cdot z,q}:\mathbold{\Sigma}(k\cdot z,q,\Gamma_1,\Gamma_2)\rightarrow \mathbold{\Sigma}(k\cdot z,\Gamma_1,\Gamma_2)$
be a normalizing projection and let
${\mathbold{\zeta}}_{\Omega}:\mathbold{\Sigma}(k\cdot z,q,\Gamma_1,\Gamma_2\times \Omega) \rightarrow \mathbold{\Sigma}(k\cdot z,q,\Gamma_1,\Gamma_2)$
be an unweighting projection. By the discussion in Section \ref{subsection:InverseHomomorphicImage}, the slice tree-automaton
\begin{equation}
\label{equation:Inverses}
\mathbf{inv}({\mathbold{\zeta}}_{\Omega},\mathbf{inv}({\mathbold{\eta}}_{k\cdot z,q}, {\mathcal{A}}(\varphi,k,z)))
\end{equation}
is a deterministic $z$-saturated tree-automaton over $\mathbold{\Sigma}(k\cdot z, q,\Gamma_1,\Gamma_2\times \Omega)$ whose graph
language consists of all weighted versions of digraphs in ${\mathcal{A}}(\varphi,k,z)$.
We will restrict the tree-automaton in Equation \ref{equation:Inverses} so that it represents only digraphs with weight $\alpha$ and
$l\;\mathrm{mod}\;m$ vertices. For simplicity of notation, let
$\mathbold{\Sigma} = \mathbold{\Sigma}(k\cdot z ,q,\Gamma_1,\Gamma_2\times \Omega)$. Recall that the deterministic tree-automaton
${\mathcal{A}}(\mathbold{\Sigma},\automataweightingfunction_{\Z_m},l) \cap {\mathcal{A}}(\mathbold{\Sigma})$ constructed in
Section \ref{subsection:PrescribedSize} generates precisely the set of unit decompositions over $\mathbold{\Sigma}$ that give
rise to a digraph with $l\;\mathrm{mod}\;m$ vertices. Recall also that the deterministic tree-automaton
${\mathcal{A}}(\mathbold{\Sigma},\automataweightingfunction_{\Omega},\alpha) \cap {\mathcal{A}}(\mathbold{\Sigma})$
constructed in Section \ref{subsection:PrescribedWeight} generates precisely the unit decompositions over $\mathbold{\Sigma}$
which give rise to digraphs of weight $\alpha$. Therefore, the slice tree-automaton
\begin{equation*}
\label{equation:treeAutomaton}
{\mathcal{A}}(\varphi,k,z,l,\alpha) = \mathbf{inv}({\mathbold{\zeta}}_{\Omega},\mathbf{inv}({\mathbold{\eta}},{\mathcal{A}}(\varphi,k,z))) \cap
{\mathcal{A}}(\mathbold{\Sigma},\automataweightingfunction_{\Z_m},l) \cap {\mathcal{A}}(\mathbold{\Sigma},\automataweightingfunction_{\Omega},\alpha) \cap
{\mathcal{A}}(\mathbold{\Sigma})
\end{equation*}
is a $z$-saturated, deterministic slice tree-automaton over the alphabet $\mathbold{\Sigma}$ whose graph language
$\lang_{{\mathcal{G}}}({\mathcal{A}}(\varphi,k,z,l,\alpha))$ consists of all digraphs that are the union of $k$ directed paths,
satisfy $\varphi$, have $l\;(\mathrm{mod}\;m)$ vertices and weight $\alpha$.
To finalize the proof, we need to estimate the size of ${\mathcal{A}}(\varphi,k,z,l,\alpha)$.
By Lemma \ref{lemma:PropertiesOfTreeAutomata}.\ref{lemma:PropertiesOfTreeAutomata:InverseHomomorphism}
the automaton in Equation \ref{equation:Inverses} can be constructed in time $|{\mathcal{A}}(\varphi,k,z)|\cdot \mathbold{\Sigma}^{O(1)}$.
By Proposition \ref{proposition:InitialAutomaton}, the automaton ${\mathcal{A}}(\mathbold{\Sigma})$ can be constructed in time
$O(|\mathbold{\Sigma}|)$. By Lemma \ref{lemma:WeightedTerms} the automaton
${\mathcal{A}}(\mathbold{\Sigma},\automataweightingfunction_{\Z_m},l)$ can be constructed in time $|\mathbold{\Sigma}|\cdot |\Z_m|^{O(1)}$
and the automaton ${\mathcal{A}}(\mathbold{\Sigma},\automataweightingfunction_{\Omega},\alpha)$
can be constructed in time $|\mathbold{\Sigma}|\cdot |\Omega|^{O(1)}$.
Therefore, given ${\mathcal{A}}(\varphi,k,z)$, the tree automaton
${\mathcal{A}}(\varphi,k,z,l,\alpha)$ can be constructed in time $|{\mathcal{A}}(\varphi,k,z)|\cdot |\mathbold{\Sigma}|^{O(1)} \cdot |\Omega|^{O(k\cdot z)} \cdot m^{O(1)}$.
Note that the size of the alphabet $\mathbold{\Sigma}(k\cdot z,q,\Gamma_1,\Gamma_1\times \Omega)$ is bounded by
$2^{O(k\cdot z \log k\cdot z)}\cdot |\Gamma_1| \cdot |\Gamma_2|^{O(k\cdot z)}\cdot |\Omega|^{O(k\cdot z)} \cdot q^{O(k\cdot z)}$.
Thus, the tree-automaton ${\mathcal{A}}(\varphi,k,z,l,\alpha)$ can be constructed in time
$$g(\varphi,k,z,|\Gamma_1|,|\Gamma_2|) \cdot q^{O(k\cdot z)} \cdot |\Omega|^{O(k\cdot z)} \cdot m^{O(1)},$$
where $g(\varphi,k,z,|\Gamma_1|,\Gamma_2)$ is the time necessary to construct ${\mathcal{A}}(\varphi,k,z)$ times $2^{O(k\cdot z \log k\cdot z)}$.
\end{proof}
\subsection{Proof of Theorem \ref{theorem:MainTheoremDirectedTreewidth}}
\label{subsection:TheProof}
The proof of Theorem \ref{theorem:MainTheoremDirectedTreewidth} will follow as
a corollary of the following theorem, whose proof is obtained by plugging
the automaton ${\mathcal{A}}(\varphi,k,z,l,\alpha)$, constructed in Lemma \ref{lemma:AutomatonSizeWeight},
into Theorem \ref{theorem:CountingSubgraphs}.
\begin{theorem}
\label{theorem:CountingSizeWeight}
Let $\mathbf{T}\in \lang(\mathbold{\Sigma}(q,\Gamma_1,\Gamma_2\times \Omega))$ be a normalized unit decomposition of width $q$ and tree-zig-zag number $z$.
Let $\varphi$ be an {$\mbox{MSO}_2$\;} sentence in the vocabulary of $(\Gamma_1,\Gamma_2)$-labeled digraphs.
Then for each $k,l\in \N$ and each $\alpha\in \Omega$ one can count in time $f(\varphi,k,z)\cdot \mathbf{T}^{O(1)}\cdot q^{O(k\cdot z)}\cdot |\Omega|^{O(k\cdot z)}$
the number of subgraphs $H$ of $\composedT$ simultaneously satisfying the following four properties:
\begin{enumerate}
\item \label{CountingSizeWeight-One} $H\models \varphi$,
\item \label{CountingSizeWeight-Two} $H$ is the union of $k$ directed paths,
\item \label{CountingSizeWeight-Three} $H$ has $l$ vertices,
\item \label{CountingSizeWeight-Four} $H$ has weight $\graphweightingfunction(H) = \alpha$.
\end{enumerate}
\end{theorem}
\begin{proof}
The proof follows by a combination of Theorem \ref{theorem:CountingSubgraphs} with
Lemma \ref{lemma:AutomatonSizeWeight}. First, Theorem \ref{theorem:CountingSubgraphs} says that given a $z$-saturated slice tree automaton ${\mathcal{A}}$
over $\mathbold{\Sigma}(k\cdot z,q,\Gamma_1,\Gamma_2\times \Omega)$, we can count in time $|\mathbf{T}|^{O(k\cdot z)}\cdot |{\mathcal{A}}|^{O(1)}$
the number of subgraphs of $\composedT$ which are isomorphic to some digraph in $\lang_{{\mathcal{G}}}({\mathcal{A}})$.
Second by Lemma \ref{lemma:AutomatonSizeWeight}, we can construct a $z$-saturated slice tree-automaton ${\mathcal{A}}(\varphi,k,z,l,\alpha)$
such that a digraph $H$ belongs to the graph language $\lang_{{\mathcal{G}}}({\mathcal{A}}(\varphi,k,z,l,\alpha))$ if and only if
$H$ satisfies $\varphi$, is the union of $k$ directed paths, has $l\;(\mathrm{mod}\;m)$ vertices, and weight $\alpha$.
Since any subgraph of $\composedT$ has at most $|\mathbf{T}|$ vertices, if we set $m=|\mathbf{T}|+1$ and
${\mathcal{A}} = {\mathcal{A}}(\varphi,k,z,l,\alpha)$, then Theorem \ref{theorem:CountingSubgraphs}
provides us with an algorithm for counting all the subgraphs of $\composedT$ that
satisfy Conditions \ref{CountingSizeWeight-One}-\ref{CountingSizeWeight-Four} of the present theorem.
Since $|{\mathcal{A}}(\varphi,k,z,l,\alpha)| \leq g(\varphi,k,z,|\Gamma_1|,|\Gamma_2|) \cdot q^{O(k\cdot z)}\cdot |\Omega|^{O(k\cdot z)}\cdot m^{O(1)}$,
and since the label sets $\Gamma_1$ and $\Gamma_2$ are fixed, the algorithm runs in time
$$f(\varphi,k,z)\cdot \mathbf{T}^{O(1)}\cdot q^{O(k\cdot z)}\cdot |\Omega|^{O(k\cdot z)},$$
\noindent where $f(\varphi,k,z) = g(\varphi,k,z,|\Gamma_1|,|\Gamma_2|)^{O(1)}$ and $|\Gamma_1|$ and $|\Gamma_2|$ are treated as constants.
\end{proof}
Finally, the proof of our main theorem (Theorem \ref{theorem:MainTheoremDirectedTreewidth}) follows as an
application of Theorem \ref{theorem:CountingSizeWeight}.
\paragraph{\textbf{Proof of Theorem \ref{theorem:MainTheoremDirectedTreewidth}}}
Let $G=(V,E,\rho,\xi\times \graphweightingfunction)$ be a $(\Gamma_1,\Gamma_2\times \Omega)$-labeled digraph of {\em directed}
treewidth $w$. By Theorem \ref{theorem:ConstructionGoodArborealDecomposition},
we can construct in time $|G|^{O(w)}$ a good arboreal decomposition $\mathcal{D}$ of $G$ of width $O(w)$.
By Theorem \ref{theorem:ComparisonWithOtherMeasures}, from $\mathcal{D}$ we can construct an olive-tree decomposition
$\mathcal{T}$ of tree-zig-zag number $z$ for some $z\leq 9w+18$.
Using Proposition \ref{proposition:OliveTreeDecompositionUnitDecomposition} we can use $\mathcal{T}$
to construct a normalized unit decomposition $\mathbf{T}$ over $\mathbold{\Sigma}(q,\Gamma_1,\Gamma_2\times \Omega)$ such that $\mathbf{T}$
has tree-zig-zag number $z$ and $\composedT=G$. Therefore given an {$\mbox{MSO}_2$\;} sentence $\varphi$, and positive integers $k,z\in \N$, we can
apply Theorem \ref{theorem:CountingSizeWeight} to count in time $f(\varphi,k,z)\cdot |\mathbf{T}|^{O(1)}\cdot q^{O(k\cdot z)}\cdot |\Omega|^{O(k\cdot z)}$,
the number of subgraphs of $\composedT$ that are the union of $k$ directed paths, satisfy $\varphi$, have $l$ vertices and weight
$\alpha$. Since $|\mathbf{T}|\leq |G|^{O(1)}$, $q\leq |E|$, and by assumption $|\Omega|\leq |G|^{O(1)}$, we have that the total running time of
the algorithm is $f(\varphi,k,z)\cdot |G|^{O(k\cdot z)}$. Since $z\leq 9w+18$, the running time of the algorithm stated in terms
of directed treewidth is $f(\varphi,k,z)\cdot |G|^{O(k(w+1))}$. Here we write $w+1$ in the exponent, to emphasize that the treewidth of
$G$ can be $0$.
$\square$
\section{Conclusion}
\label{section:Conclusion}
In this work we devised the first algorithmic metatheorem for digraphs of constant directed treewidth.
We showed that most of the previously known positive algorithmic results for this class of digraphs
can be re-stated in terms of our metatheorem. Additionally, we showed
how to use our metatheorem to provide polynomial time algorithms for
two classes of counting problems whose polynomial-time solvability is not
implied by previously existing techniques. Namely, for each fixed $k$, we showed how to count in polynomial time on digraphs of
constant directed treewidth, the number of minimum spanning strong subgraphs that are the union of $k$ directed paths,
and the number of subgraphs that are the union of $k$ directed paths and satisfy a given minor closed property.
To prove our main theorem we introduced two new theoretical tools which in our opinion are of
independent interest. The first, the tree-zig-zag number of a digraph, is a new directed width
measure that is at most a constant times its directed treewidth. Concerning this measure,
we leave open the problem of determining whether there exist families of digraphs of constant
tree-zig-zag number but unbounded directed treewidth, or whether the directed treewidth of a digraph
is always bounded by a function of its tree-zig-zag number. The second theoretical tool we have introduced
is the notion of $z$-saturated tree-automata.
By Theorem \ref{theorem:CountingSubgraphs}, given a digraph $G$ of constant directed treewidth, and
a $z$-saturated tree-automaton ${\mathcal{A}}$ generating only digraphs that are the union of $k$ directed paths,
one can count the number of subgraphs of $G$ that are isomorphic to some digraph in $\lang_{{\mathcal{G}}}({\mathcal{A}})$. It would be
interesting to study ways of constructing $z$-saturated tree-automata without the help of {$\mbox{MSO}_2$\;} logic.
Such a construction would open the possibility of using Theorem \ref{theorem:CountingSubgraphs} to solve counting problems,
on digraphs of constant directed treewidth, that may not be approachable via Theorem \ref{theorem:MainTheoremDirectedTreewidth}.
\section{Acknowledgements.}
The author is currently supported by the European Research Council, ERC grant agreement 339691, within the context
of the project Feasibility, Logic and Randomness (FEALORA).
\bibliographystyle{abbrv}
|
\section{LHCb experiment}
\section{The vertex locator}
\label{Section:velo}
The LHCb experiment is a forward single-arm spectrometer collecting data at the Large Hadron Collider (LHC) at CERN.
The Vertex Locator (VELO) detector allows LHCb to measure vertices very accurately, such that primary vertices (from proton-proton collisions) and secondary vertices (from decays of short-lived particles) can be separated.
The VELO is a silicon strip detector made of 42 modules, where each module has strips in the $R$ and $\phi$-directions on alternating sides.
The modules are placed along the beam line over roughly a meter, such that each particle which originates from the interaction region at an angle in the range from \SIrange{15}{390}{\milli\radian} can be reconstructed.
The VELO modules are cooled using evaporative $\text{CO}_2$ cooling.
The modules are kept in a vacuum, which is separated from the LHC vacuum by a \SI{300}{\micro\meter} thick RF-foil~\cite{reoptimisedtdr}; the foil stops the beams from inducing currents in the modules.
\Cref{Figure:rffoil} shows the complex shape of the RF-foil.
Nominally the VELO is \SI{27}{\milli\meter} from the LHC beams to ensure they do not damage the VELO.
During data taking, once the beams are stable, the modules move in so that the sensor edges are \SI{8.2}{\milli\meter} from the beams.
More details about the VELO and LHCb can be found in~\cite{reoptimisedtdr,velotdr,lhcbdetector}.
The performance of the VELO and the effects of radiation damage are discussed in~\cite{performance,radiation}.
\begin{figure}[tbp]
\centering
\includegraphics[height=4.5cm,keepaspectratio]{velo_foil.png}
\hspace{1em}
\includegraphics[height=4.5cm,keepaspectratio]{VP_foilandsensors.png}
\caption{
Comparison of the current RF-foil (left) and the upgrade RF-foil (right).
The left image shows the two halves of the current RF-foil, while the inset image zooms into the region of the RF-foil near the beams.
The right image shows one half of the simulated upgrade RF-foil with some modules in place.
}
\label{Figure:rffoil}
\label{Figure:upgraderffoil}
\end{figure}
\section{The LHCb upgrade}
\label{Section:lhcbupgrade}
The LHC does not deliver the maximum instantaneous luminosity that it could to LHCb; the beams are separated to reduce the number of collisions.
The general purpose detectors (ATLAS and CMS) have a maximum instantaneous luminosity of \SI{7e33}{\per\square\centi\meter\per\second}, whereas LHCb receives \SI{4e32}{\per\square\centi\meter\per\second}, almost a factor of 20 less.
The hardware trigger in LHCb uses transverse energy ($E_t$) in the calorimeters to trigger events with interesting hadronic decays.
When the luminosity is increased a harder cut must be made on $E_t$ to keep data rates manageable, but this reduces the trigger efficiency.
This means the trigger yield saturates as luminosity is increased~\cite{lhcbupgradeloi}.
To overcome this limitation LHCb is planning an upgrade during LHC long shut-down 2 (LS2).
This upgrade will allow the full detector to be read out into a software trigger, bypassing the hardware trigger; this will enable LHCb to run at a luminosity of \SI{2e33}{\per\square\centi\meter\per\second}.
This rise in luminosity increases the occupancy of the VELO to unacceptable levels.
Further, running in these conditions pushes the radiation damage beyond the maximum allowable and so the VELO will need to be upgraded.
\section{The vertex locator upgrade}
\label{Section:upgrade}
The upgrade VELO will use pixel sensors~\cite{veloupgradetdr} instead of strip sensors; this is the largest change between the current and the upgrade VELO.
Aside from this, many aspects will remain the same: the modules will move in close to the beams when they become stable, evaporative $\text{CO}_2$ cooling will be used and most of the support infrastructure outside of the vacuum will remain the same.
In the next section the upgraded detector will be described and the major differences highlighted.
\subsection{Detector overview}
\label{Section:upgrade:improvements}
Each VELO upgrade module will contain 4 planar silicon sensors, as can be seen in \cref{Figure:modules}.
Each sensor will be read out by 3 VeloPix chips, which are based on Timepix3~\cite{timepix}.
The VeloPix measures $14.07\times{}14.07\,$\si{\square\milli\meter} and contains $256\times{}256$ pixels, which leads to a pitch of \SI{55}{\micro\meter}.
The VeloPix will use a binary data-driven readout of $4\times{}2$ super-pixels.
\begin{figure}[tbp]
\centering
\includegraphics[width=0.8\textwidth,keepaspectratio]{modules_new.png}
\caption{
The diagram shows the upgrade VELO modules and one possible design of the supports which will hold the modules in place.
The silicon substrate (turquoise) with microchannel cooling inside is shown, however it is mostly hidden beneath the hybrid (brown).
The sensors (red) will be evenly distributed on opposite sides of the module.}
\label{Figure:modules}
\end{figure}
The sensors will need a bias voltage of \SI{1000}{\volt} after the radiation damage received while collecting an integrated luminosity of \SI{50}{\per\femto\ensuremath{\rm \,b}\xspace}~\cite{veloupgradetdr}.
It is difficult to have this high voltage without sparking between the sensor and the VeloPix and so a guard ring of \SI{450}{\micro\meter} has been specifically designed to reduce the risk of sparking.
Additionally the sensor edge may be coated to further reduce the risk of sparking.
The VeloPix and sensors will be bump-bonded and together glued onto the silicon substrate, which will be cooled using microchannels.
These are made by etching \SI{200}{\micro\meter} wide and \SI{120}{\micro\meter} deep channels into \SI{260}{\micro\meter} thick silicon.
Then another \SI{140}{\micro\meter} thick piece of silicon is overlaid to seal the channels.
Liquid $\text{CO}_2$ is pumped through these channels at a pressure of \SIrange{20}{30}{\bar}~\cite{veloupgradetdr} where it evaporates and cools the module.
Unlike other methods of cooling, there will be no mismatch in coefficients of thermal expansion and so the module will not deform when cooled and heated.
Further, this is a particularly low-material cooling method since it requires no additional conductive material around the sensors.
Thermal simulations of the microchannel cooling have determined that the tip of the sensor could extend \SI{5}{\milli\meter} from the edge of the substrate, which further reduces the material that particles must pass through.
The microchannels must not leak into the VELO vacuum; to ensure this, the pressure was successfully cycled a thousand times from \SIrange{0}{160}{\bar} using a prototype with no leakage.
The active silicon sensor will be moved almost $40\%$ closer to the beams, from \SI{8.2}{\milli\meter} to \SI{5.1}{\milli\meter}; this reduces the extrapolation distance to the vertices and so improves the vertex and IP resolution.
The RF-foil is responsible for $50\%$ of material that particles pass through in the VELO, so its design is critical.
The RF-foil is shaped to fit tightly around the VELO upgrade modules and is corrugated to minimise its contribution to the material; this leads to a very complicated shape as pictured in \cref{Figure:upgraderffoil}.
It will be milled out of a solid aluminium block to a thickness of \SI{250}{\micro\meter} and then possibly chemically thinned around the area near the beams.
\subsection{Simulated performance}
\label{Section:upgrade:performance}
The impact parameter (IP) of a track is the perpendicular distance between the track direction and its associated primary vertex; this is an important property for LHCb, because it effects physics quantities such as the lifetime resolution and the IP can also be used to trigger events which contain long-lived particles like \ensuremath{\PB}\xspace{} mesons.
In \cref{Figure:ipres} the IP resolution is shown as a function of inverse transverse momentum ($1/\mbox{$p_{\rm T}$}\xspace{}$), where it can be seen that the upgrade VELO will significantly improve the IP resolution.
The improvements come from the low-material modules, the thinner RF-foil and from moving the active silicon closer to the beams.
The expected tracking efficiency of the current and upgrade VELO detectors during Run~3 is shown in \cref{Figure:trackingefficiency}.
The increase in occupancy in Run~3 creates many fake tracks in the current strip-based VELO and reduces its tracking efficiency, but the upgrade pixel-based VELO does not suffer from this issue.
The excellent performance of the upgrade VELO is only useful if it can withstand the effects of being irradiated.
\Cref{Figure:radiationdamage} shows the expected effect of radiation on the performance of the upgrade VELO with sensors biased to \SI{500}{\volt}.
There is only a relatively small degradation in performance after collecting an integrated luminosity of \SI{50}{\per\femto\ensuremath{\rm \,b}\xspace}, which can almost completely be removed by further increasing the sensor bias.
This demonstrates that the detector can maintain the excellent performance even after heavy irradiation.
\begin{figure}[tbp]
\centering
\includegraphics[width=0.32\textwidth,keepaspectratio]{TDR_long_IPx.pdf}
\includegraphics[width=0.32\textwidth,keepaspectratio]{VELO_eff_Phi.pdf}
\includegraphics[width=0.32\textwidth,keepaspectratio]{ip_resolution_500V.pdf}
\caption{
The left plot shows the IP resolution as a function of $1/\mbox{$p_{\rm T}$}\xspace{}$ for the current VELO (black circles) and the upgrade VELO (red squares), with a typical $1/\mbox{$p_{\rm T}$}\xspace{}$ spectrum in grey.
The middle plot shows tracking efficiency as a function of azimuthal angle ($\phi$) for the current and upgrade VELO detectors.
The right plot shows the effect of radiation damage on the IP resolution; the points show the performance after \SI{0}{\per\femto\ensuremath{\rm \,b}\xspace} (black circles), \SI{10}{\per\femto\ensuremath{\rm \,b}\xspace} (red squares), \SI{30}{\per\femto\ensuremath{\rm \,b}\xspace} (yellow upward triangles) and \SI{50}{\per\femto\ensuremath{\rm \,b}\xspace} (blue downward triangles).
}
\label{Figure:ipres}
\label{Figure:trackingefficiency}
\label{Figure:radiationdamage}
\end{figure}
\section{Conclusion}
\label{Section:conclusion}
The LHCb experiment is set for a significant upgrade, which will be ready for Run~3 of the LHC in 2020.
This upgrade will allow LHCb to run at a significantly higher instantaneous luminosity and collect an integrated luminosity of \SI{50}{\per\femto\ensuremath{\rm \,b}\xspace} by the end of Run~4.
In this process the VELO will be upgraded to a pixel-based detector.
The upgraded VELO will improve upon the current detector by being closer to the beams and having lower material modules with microchannel cooling and a thinner RF-foil.
Simulations have shown that it will maintain its excellent performance, even after the radiation damage caused by collecting an integrated luminosity of \SI{50}{\per\femto\ensuremath{\rm \,b}\xspace}.
|
\section{Introduction}
Macroscopic models of moving crowd identify the swarm through some density that is transported by a velocity vector field, see for instance a review paper by {\sc B. Maury} \cite{Ma}. For example, to describe the traffic jams, one may use the one-dimensional fluid model describing two-phase flow
\eq{
\left.
\begin{array}{r}
\partial_{t}\alpha+\partial_{x}(\alpha u)=0\\
\partial_{t}(\alpha u)+\partial_{x}(\alpha u^2+\pi)=0
\end{array}
\right\}
\quad\text{in } (0,T)\times \mathbb{R}\label{s0}}
with the following restrictions
\eq{0\leq\alpha\leq 1,\qquad (1-\alpha)\pi=0\label{cs0}.}
Here $\alpha$ denotes the liquid volume fraction that plays the role of the crowd density, $u$ denotes the velocity and $\pi\geq0$ denotes some singular pressure term appearing only when $\alpha=1$.\\
System (\ref{s0}-\ref{cs0}) is known in the literature as the {\it pressureless gas system with unilateral
constraint} and has been studied for example by {\sc F. Berthelin} in \cite{Be}. It can be interpreted as a coupling of two systems in the respective domains where $\alpha<1$ (liquid-gas mixture) and where $\alpha=1$ (pure liquid).
{This system can be formally derived from bi-fluid system (see f.i. {\sc F. Bouchut} {\it et al.} in \cite{BBCR} where a hierarchy for gas-liquid two-phase flows is also presented} ).
\bigskip
\noindent A generalization of (\ref{s0}-\ref{cs0}) to the multi-dimensional viscous case is described {by} the compressible/incompressible Navier-Stokes type of system
\begin{equation}\label{CM0}
\left. \begin{array}{r}
\partial_t\rho + {\rm div} (\rho \vc{u}) = 0\\
0\leq\rho\leq \rho^*\\
\partial_t (\rho \vc{u}) + {\rm div} (\rho \vc{u} \otimes \vc{u}) + \nabla p +\rho^*\nabla\pi-{\rm div}\,\bf{S} = \vc{0}\\
{\rm div}\lr{\rho^*\vc{u}}=0 \quad \text{a. e. in}\quad \{\rho=\rho^*\}\\
\pi \geq 0 \quad \text{a. e. in}\quad \{\rho=\rho^*\}\\
\pi=0 \quad \text{a. e. in}\quad \{\rho<\rho^*\}\\
\end{array}\right\} \quad \text{in} \ (0,T)\times\Omega,
\end{equation}
in which the homogeneous congestion constraint $\alpha{\leq} 1$ has been replaced by the inhomogeneous one $\rho\leq \rho^*(x)$.\\
The unknowns here are the density $\rho$, the velocity vector field $\vc{u}$ and the pressure $\pi$, which is the Lagrange multiplier associated with the incompressibility constraint ${\rm div}\lr{\rho^*\vc{u}}=0$ a.e. in $\{\rho=\rho^*\}$. Note that as in the previous example, $\pi$ is apparent only in the congested regions $\{\rho=\rho^*\}$. In fact, conditions \eqref{CM0}$_{5}$, \eqref{CM0}$_{6}$ can be rewritten as one constraint
\eq{\rho\pi=\rho^*\pi\geq 0.\label{CON}}
Furthermore, the stress tensor $\vc{S}$ and the internal pressure $p$ are known functions of $\rho$ and $\vc{u}$ which are typical for barotropic flow of Newtonian fluid, i.e.
\[ \vc{S}(\vc{u}) = 2 \mu D(\vc{u}) + \lambda {\rm div} \vc{u} \, \vc{I}, \quad \mu>0,\quad 2\mu+\lambda>0,\]
\eqh{p(\rho)=\rho^\gamma,\quad \gamma>1.}
System \eqref{CM0} mixes the free flow and congested regions denoted by $\Omega_f(t)$ and $\Omega_c(t)$ respectively, where
$$\Omega_f(t) = \{x: \rho(t,x) <\rho^*(x)\}, \qquad
\Omega_c (t) = \{x: \rho(t,x)=\rho^*(x)\}.$$
It can be thus seen as a free boundary problem, in which the the free interface separating $\Omega_c$ from $\Omega_f$ can become very irregular, as it is illustrated on the figure below.
\bigskip
\begin{center}
\includegraphics{petite_patate_1}
\end{center}
\bigskip
The objective of this paper is to mathematically justify that the solution to problem \eqref{CM0} can be obtained as a limit of $(\rho_n,\vc{u}_n)$-- the solutions to the isentropic compressible Navier-Stokes equations
\begin{equation}\label{CM0prim}
\left. \begin{array}{r}
\partial_t\rho_n + {\rm div} (\rho_n \vc{u}_n) = 0\\
\partial_t (\rho_n \vc{u}_n) + {\rm div} (\rho_n \vc{u}_n \otimes \vc{u}_n) + \nabla p +\rho^*\nabla\pi_{{n}}-{\rm div}\,\bf{S} = \vc{0}
\end{array}\right\}
\end{equation}
where $\pi_{n}$ is some approximation of the limit pressure $\pi$.
We want to find such an approximation that would guarantee uniform boundedness of the sequence approximating the density $\rho_{n}\leq\rho^*$. This feature is very important for numerical purposes, see for example \cite{Ma}.
\bigskip
\noindent Below we present two possible ways of approximating the pressure $\pi$ appearing in the limit system \eqref{CM0}.
\medskip
\noindent{\bf The barotropic pressure.} The first kind of approximation uses the classical barotropic pressure
\begin{equation}\label{power_pres}\displaystyle \pi_{n}=\pi_{\gamma_n}=a\left(\frac{\rho_n}{\rho^*}\right)^{\gamma_n}
\end{equation}
with $a >0$ fixed and the adiabatic exponent $\gamma_n>1$ being the approximation parameter.\\
Mathematical analysis of system \eqref{CM0prim} with the above pressure for fixed $\gamma_n$ is based on the existence theory for barotropic Navier-Stokes equations developed by {\sc P.--L. Lions} in \cite{PLL} and {\sc E. Feireisl} in \cite{EF2001}.\\
In this framework, only the homogeneous case $\rho^*=1$ has been studied. It has been justified by {\sc P.--L. Lions} and {\sc N.~Masmoudi} in \cite{LiMa} that the limit system \eqref{CM0} may be recovered from (\ref{CM0prim}-\ref{power_pres}) letting $\gamma_n\to\infty$. Later on, {\sc S.~Labb\'e} and {\sc E.~Maitre} performed the same limit passage for more complex system. They considered the viscosity coefficients $\mu,\ \lambda$ depending on the density and an additional surface tension term in the momentum equation \eqref{CM0}$_3$.\\
A similar asymptotic limit has been studied also for a model of tumour growth by {\sc B.~Perthame}, {\sc F.~Guir\'os} and {\sc J.L.~V\'azquez} \cite{PeQuVa}. They used the cell population density model with the pressure of the form $\frac{m}{m-1}\lr{\rho/\rho^*}^{m-1}$ with parameter $m>1$ and the maximum packing density for the cells denoted by $\rho^*$ which is constant. In the limit $m\to\infty$ they obtained a free boundary model of Hele-Shaw type.
\bigskip
\noindent{\bf The singular pressure.} The second approximation uses a pressure that becomes singular close to some threshold value of the density $\rho^*$, for example
\begin{equation}\label{sing_pres}
\displaystyle \pi_{\varepsilon_n}= \varepsilon_n\dfrac{\left(\dfrac{\rho_n}{\rho^*}\right)^\alpha}{\left(1-\dfrac{\rho_n}{\rho^*}\right)^\beta}
\end{equation}
with $\beta,\alpha>0$ fixed and $\varepsilon_n>0$ being the approximation parameter. \\
Such type of degeneration of the pressure can be used to model various phenomena. It appears in kinetic theory of dense gases where the interaction between the molecules is strongly repulsive at very short distance. The mutual reluctance of neighbouring molecules to share a certain amount of space (covolume) leads to the Van der Waals equation of state (see \cite{ChCo})
\[\pi = \frac{NRT}{V-Nb} - a\frac{N^2}{V^2}, \]
where $N$ is the number of molecules, $V$ the volume. The two terms on the right hand side (r.h.s.) represent respectively the repulsive and attractive forces. We see that the pressure becomes singular when $Nb$ approaches $V$, which corresponds to the state where motion is no longer possible. Other equations of state, modifying the representation of repulsive forces, were proposed for instance by {\sc N.F.~Carnahan} and {\sc K.E.~Starling} in \cite{CaSt}. \\
Similar form of the pressure has been also recently considered by {\sc P.~Degond}, {\sc J.~Hua} and {\sc L.~Navoret} to model collective motion in \cite{DeHuNa}, and by {\sc F.~Berthelin} and {\sc D.~Broizat} to model traffic flow in \cite{BeBr}. Their approximation is of the form $\nabla \pi_{\varepsilon_n}$ with
\[
\displaystyle \pi_{\varepsilon_n}= \dfrac{\varepsilon_n}{\left(\dfrac{1}{\rho_n}-\dfrac{1}{\rho^*}\right)^\beta} ,
\]
while in \eqref{CM0prim} we take $\rho^*\nabla\pi_{\varepsilon_n}$. However, as we will see later on, the factor $\rho^*$ is necessary in order to obtain the energy equality when $\rho^*$ is non-constant.\\
Finally, systems involving such kind of degeneration are used in the theory of granular flows (see \cite{AnFoPo} and \cite{PiLe}). For instance, {\sc F.M.~Auzerais}, {\sc R.~Jackson} and {\sc W.B.~Russel} proposed in \cite{AuJaRu} a model for sedimentation using an empirical pressure of the form
\[\pi=\dfrac{C_0\phi^s}{\phi^*-\phi}\]
with $2\leq s\leq5$. In this framework the density of the fluid is replaced by $\phi$--the volume fraction of the solid phase ($0\leq\phi\leq1$) with some constant threshold value $\phi^*=0.64$.
\medskip
\noindent From the mathematical point of view, the first result for system \eqref{CM0prim} with singular pressure is due to {\sc E. Feireisl, H. Petzeltov\'a, E.~Rocca} and {\sc G.~Schimperna}, \cite{Fe}. They studied a model of two-phase compressible fluid flow with a Cahn-Hilliard type equation for a phase variable. As a corollary of their result we get the existence of global in time weak solution for \eqref{CM0prim} and \eqref{sing_pres} with $\rho^*=const.$, $\beta>3$ and $\varepsilon_n>0$ being fixed.
\medskip
\noindent{To our knowledge, the justification of the limit passage $\varepsilon_n\to0$, which formally gives system \eqref{CM0}, is still an open problem in the general case.} The only result so far concerns the one-dimensional homogeneous ($\rho^*=1$) case, studied by {\sc D. Bresch}, {\sc C. Perrin} and {\sc E. Zatorska} in \cite{BrPeZa}. First of all they proved that for $\beta,\gamma>1$ and $\varepsilon>0$ fixed
there exists a {regular} solution $(\rho_\varepsilon,u_\varepsilon)$ to
\eq{
& \partial_t\rho_\varepsilon + \partial_x(\rho_\varepsilon u_\varepsilon) = 0, \\
& \partial_t (\rho_\varepsilon u_\varepsilon) + \partial_x (\rho_\varepsilon u_\varepsilon \otimes u_\varepsilon) + \partial_x p(\rho_\varepsilon) + \varepsilon\partial_x \frac{\rho_\varepsilon^\gamma}{(1-\rho_\varepsilon)^\beta} -(\lambda+\mu) \partial^2_{xx} u_\varepsilon = 0,
\label{1D}}
such that
\begin{equation}\label{b_rho}
0< c \le \rho_\varepsilon \le C(\varepsilon) < 1,
\end{equation}
for some constants $c$ and $C(\varepsilon)$.
Secondly, they justified that system \eqref{CM0} possesses a weak solution being a limit of regular solutions to \eqref{1D}.
\bigskip
\noindent The main difficulty in justifying the limit passage in both approaches is that the energy estimate does not give a uniform bound on the pressure term. More precisely, when $\rho^*=1$, the energy estimate for the power law pressure reads
\[\frac{ d}{dt}\intOB{\frac{1}{2}\rho_n |\vc{u}_n|^2 + \frac{a}{\gamma_n-1}\rho_n^{\gamma_n}} + \intOB{\mu|\nabla \vc{u}_n|^2+(\lambda+2\mu)({\rm div} \vc{u}_n)^2} = 0 \]
but it does not imply that $\rho^{\gamma_n}$ is bounded uniformly with respect to $\gamma_n\to\infty$.\\
Similarly, for the degenerate pressure we have
\[\frac{ d}{dt}\intOB{\frac{1}{2}\rho_n |\vc{u}_n|^2 + \rho_n\Gamma_{\varepsilon_n}(\rho_n)} + \intOB{\mu|\nabla \vc{u}_n|^2+(\lambda+2\mu)({\rm div} \vc{u}_n)^2} = 0 \]
with
\[\Gamma_{\varepsilon_n}(\rho_n)=\int_0^{\rho_n}{\dfrac{\pi_{\varepsilon_n}(s)}{s^2}\mathrm{d}s},\]
but this does not yield the uniform pressure estimate either. Indeed, take for instance $\pi_{\varepsilon_n}=\varepsilon_n\rho_n^2(1-\rho_n)^{-4}$ for which $\Gamma_{\varepsilon_n}=\varepsilon_n/(3 (1-\rho_n)^{3})$, then the growth of $\rho_n\Gamma_{\varepsilon_n}$ around singularity is one order less than of $\pi_{\varepsilon_n}$. Additional information on the pressure is thus necessary and requires more sophisticated tools such as application of the Bogovskii operator.
\bigskip
\noindent The most important difference in these two ways of approximation lies in the uniform estimate of the density. Indeed, approximation $\pi_{\gamma_n}=\rho_n^{\gamma_n}$ does not guarantee the validity of the congestion constraint $0\le \rho_{n}\le 1$ for fixed $\gamma_n$. The main advantage of approximation based on \eqref{sing_pres} is validity of this restriction uniformly with respect to $\varepsilon_n$. In fact, the degenerate pressure plays the role of natural barrier (see \cite{Ma} by {\sc B. Maury}) and
makes the second approach more suitable for numerical schemes. \\
\noindent Our goal is therefore to extend the result from the previous work \cite{BrPeZa} to the global weak solutions framework in the multi-dimensional space case.
The main difference between the one-dimensional case and the multi-dimensional case, is that in the latter, {\it the sufficiently regular solutions are not known to exist}. Therefore, the strong convergence of the density and validity of the congestion constraint \eqref{CON} cannot be deduced directly from the {\it a priori} estimates. Eventually, the validity of the r.h.s. inequality in \eqref{b_rho} follows from an additional level of approximation using truncations of singular part of the pressure.
However, recovering system \eqref{CM0} after this step requires {\it equi-integrability of the pressure term}, for which some restriction of the strength of singularity need to be imposed. The equi-integability of the pressure similar to \eqref{sing_pres} for $\rho^*=const.$ and $\beta>3 $ was proved by {\sc E.Feireisl} {\it et al.} in \cite{Fe} and we heavily relay on their approach to this issue.
The purpose of the paper is also to generalize \cite{BrPeZa} to the heterogeneous case, i.e. when the constant upper bound on the density is replaced by a prescribed function $\rho^*=\rho^*(x)>0$. This generalization has many applications. For instance, in \cite{BeBr}, {\sc F.~Berthelin} and {\sc D.~Broizat} use a non-constant maximal constraint to study the dynamics of traffic jams and the influence of the number of lanes on the road.
It is also important in the study of models of flow through the closed pipes of non-uniform height $h^*(x)$ as said by {\sc F.~Berthelin} in \cite{Be}. In these models, the surface of the flowing fluid described by $h$ may be either free when $h < h^*$ or additionally pressurized when it touches the pipe wall $h=h^*$, see the picture below.
\begin{center}
\includegraphics{conduite_1}
\end{center}
The study of the non-viscous systems can be found in {\sc C.~Bourdarias, M.~Ersoy, S.~Gerbi} \cite{BoEr}, see also references therein.
\section{Formulation of the main problem}
System \eqref{CM0} is supplemented with the homogeneous Dirichlet boundary conditions
\eq{\vc{u}\big|_{\partial\Omega}=\vc{0}.
\label{bc}}
We assume that the threshold density $\rho^*=\rho^*(x)>0$ is a $\mathcal{C}^1(\overline{\Omega})$ function
and that the initial data
\eq{\rho\big|_{t=0}=\rho_0,\quad (\rho\vc{u})\big|_{t=0}=\vc{m}_0\label{ic}}
satisfy
\begin{equation}\label{ini_data}
\begin{gathered}
0\leq \rho_0(x) < \rho^*(x) \quad\text{a.e. in}\quad \Omega,\quad p(\rho_0) \in L^1(\Omega), \\
\vc{m}_0\in (L^2(\Omega))^3, \quad \vc{m}_0\vc{1}_{\{\rho_0=0\}}=0 \quad\text{a.e. in}\quad \Omega,\\
\frac{|\vc{m}_0|^2}{\rho_0}\vc{1}_{\{\rho_0>0\}} \in L^1(\Omega),
\end{gathered}
\end{equation}
\begin{equation}\label{add_hyp_ini}
M_0=\frac{1}{|\Omega|}\intO{\rho_0} <\inf_{x\in \Omega} \rho^*(x).
\end{equation}
\begin{rmk} Condition \eqref{add_hyp_ini} plays an essential role in the proof of uniform $L^1((0,T)\times\Omega)$ bound for a sequence approximating $\pi$. Note that in the case $\rho^*$ constant, condition \eqref{add_hyp_ini} is directly satisfied since $\rho_0<\rho^*$.
\end{rmk}
Below we introduce the notion of a weak solution to system \eqref{CM0}.
\begin{df}[Weak solution of the limit system]\label{Def1}
A triple $(\rho,\vc{u},\pi)$ is called a weak solution to \eqref{CM0} with \eqref{bc} and \eqref{ic} if equations
\begin{equation*}
\begin{array}{c}
\partial_t\rho + {\rm div} (\rho \vc{u}) = 0,\\
\partial_t (\rho \vc{u}) + {\rm div} (\rho \vc{u} \otimes \vc{u}) + \nabla p +\rho^*\nabla\pi-{\rm div}\bf{S} = \vc{0}
\end{array}
\end{equation*}
are satisfied in the sense of distributions, the divergence free condition ${\rm div}\lr{\rho^*\vc{u}}=0$ is satisfied a.e. in $\{\rho=\rho^*\}$, the constraint $0\leq \rho\leq\rho^*$ is satisfied a.e. in $(0,T)\times \Omega$,
and the following regularity properties hold
\begin{align*}
&\rho\in \mathcal{C}([0,T];L^p(\Omega)), \quad 1\leq p < \infty, & \\
& \vc{u} \in L^2(0,T;(W^{1,2}_0(\Omega))^3),\quad \dfrac{|\vc{m}|^2}{\rho} \in L^{\infty}(0,T; L^1(\Omega)),&\\
&\pi\in {\cal M}^+ ((0,T)\times \Omega).&
\end{align*}
Moreover, $\pi$ is sufficiently regular so that the condition
\eq{(\rho-\rho^*)\pi=0.\label{Cons}}
is satisfied in the sense of distributions.
\end{df}
\begin{rmk} Similarly to the homogeneous case studied bt {\sc P.-L. Lions} and {\sc N. Masmoudi} in {\rm \cite{LiMa}}, we can prove that the constraint $\eqref{CM0}_2$ and the divergence free condition $\eqref{CM0}_4$ are "compatible" (see {\rm Lemma \ref{LMP}}). More precisely, if $\vc{u}$ belongs to $L^2(0,T;(W_0^{1,2}(\Omega))^3)$, $\rho$ belongs to $L^2((0,T)\times\Omega)$ and the couple $(\rho,\vc{u})$ satisfies the continuity equation $\eqref{CM0}_1$, then $\eqref{CM0}_2$ and $\eqref{CM0}_4$ with $0\leq \rho_0\leq \rho^*$ are equivalent.
As it will be explained later, this fact is a natural consequence of the renormalized theory applied to the equations satisfied by the quantities $\rho/\rho^*$ and $\rho-\rho^*$.
\end{rmk}
\bigskip
\noindent Now we define the notion of weak solution to approximate system \eqref{CM0prim} with the approximate pressure of the form
\begin{equation*}
\displaystyle \pi_\varepsilon(r)= \left\{\begin{array}{ll}
\varepsilon\dfrac{r^{\alpha}}{(1-r)^{\beta}} &\text{if } r < 1 \\
\infty &\text{if } r \geq 1 \end{array}\right.\quad \alpha,\beta >3.
\end{equation*}
\begin{df}[Weak solution of the approximate system]\label{weak_app_sol} A pair $(\rho_\varepsilon,\vc{u}_\varepsilon)$ is called a weak solution to (\ref{CM0prim}) if it satisfies
\begin{itemize}
\item the approximate continuity equation
\eq{ \intTO{\rho_\varepsilon\partial_t\phi} +\intO{\rho_0\phi(0)} + \intTO{\rho_\varepsilon \vc{u}_\varepsilon \cdot \nabla \phi} \\
= \intO{(\rho_\varepsilon \phi)(T) }\label{AC}}
for all $\phi\in \mathcal{D}([0,T]\times \Omega)$;
\item the approximate momentum equation
\eq{
& \intTO{\rho_\varepsilon \vc{u}_\varepsilon \cdot \partial_t \varphi} + \intO{\vc{m}_0\cdot \varphi(0)} \\
+&\intTO{\rho_\varepsilon(\vc{u}_\varepsilon\otimes\vc{u}_\varepsilon): \nabla \varphi}- \intTO{\vc{S} : \nabla \varphi}\\
+&\intTO{p{\rm div}\varphi}
+\intTO{\pi_\varepsilon \left(\frac{\rho_\varepsilon}{\rho^*}\right){\rm div} (\rho^*\varphi)} \\
&\qquad\qquad\qquad\qquad\qquad\qquad\qquad= \intO{ (\vc{m}_\varepsilon \cdot\varphi)(T)},
\label{AM}}
for all $\varphi\in (\mathcal{D}([0,T]\times \Omega))^3$.
\end{itemize}
\end{df}
\bigskip
\noindent The objective of this paper is to justify that the weak solution from \mbox{Definition \ref{Def1}} can be obtained as a limit of weak solutions from Definition \ref{weak_app_sol}. \\
\medskip
\noindent The main theorem of this paper reads.
\begin{Theorem} \label{main}
\begin{enumerate}
\item Let $\varepsilon>0$ be fixed, then there exists a global weak solution $(\rho_\varepsilon,\vc{u}_\varepsilon)$ to \eqref{CM0prim} in the sense of Definition \ref{weak_app_sol}, moreover,
\eq{0\leq\rho_\vep\leq\rho^*.\label{un_main}}
\item For $\varepsilon\to0$, there exists a subsequence $(\rho_\vep,\vc{u}_\vep,\pi_{\varepsilon})$ converging to $(\rho,\vc{u},\pi)$ a solution of system \eqref{CM0} in the sense of Definition \ref{Def1}.
More precisely
\eqh{
&\rho_\vep \to \rho \quad\text{weakly\ in }\quad L^p((0,T)\times\Omega) \quad \forall 1\leq p <+\infty,\\
&\vc{u}_\vep \to \vc{u} \quad\text{weakly\ in }\quad L^2(0,T;(W^{1,2}_0(\Omega))^3),\\
&\pi_\varepsilon \to \pi \quad\text{weakly\ in }\quad {\cal M}^+((0,T)\times\Omega).
}
\end{enumerate}
\end{Theorem}
\medskip
The paper is organized as follows. In Section \ref{S_rho1}
we present details of approximation and prove the first part of Theorem \ref{main}. Then, in Section \ref{S_ep}, we recover the original system by letting $\varepsilon\to 0$. Section \ref{Appendix} is the Appendix in which we recall some basic facts about the Bogovskii operator, the Riesz transform and the renormalized continuity equation that are used in several places in the course of the proof.
\section{Existence of approximate solutions}\label{S_rho1}
In order to prove the first part of Theorem \ref{main}, we consider further approximation with additional truncation parameter $\delta$ and artificial pressure $\kappa\rho^{K}$ where $K$ is sufficiently large positive number that will be determined later on. Then we will show that (\ref{AC}-\ref{AM}) can be recovered by letting $\delta\to0$ and $\kappa\to0$ respectively. In this section $\varepsilon$ is fixed and we drop it when no confusion can arise.
\subsection{Basic level of approximation}
For fixed $\kappa,\delta>0$, we consider the basic level of approximation
\eq{\label{appr_model_delta}
& \partial_t\rho_\delta + {\rm div} (\rho_\delta \vc{u}_\delta) = 0 \\
& \partial_t (\rho_\delta \vc{u}_\delta) + {\rm div} (\rho_\delta \vc{u}_\delta \otimes \vc{u}_\delta) + \nabla p(\rho_\delta) +\rho^*\nabla\pi_{\kappa,\delta}\lr{\frac{\rho_\delta}{\rho^*}} -{\rm div}\vc{S} = 0
}
with the approximate pressure
given by
\begin{equation}\label{pressure_ap_p2}
\displaystyle \pi_{\kappa,\delta}(s) =\kappa s^K+\left\{
\begin{array}{ll}
\vspace{0.2cm}
\displaystyle\varepsilon \frac{s^{\alpha}}{(1-s)^{\beta}} & \text{ if } s < 1-\delta, \\
\displaystyle\varepsilon \frac{s^{\alpha}}{\delta^{\beta}} & \text{ if } s\geq 1-\delta .
\end{array}\right.
\end{equation}
\bigskip
For all parameters fixed, the approximate pressure $\pi_{\kappa,\delta}$ is a monotone increasing function of $\rho$. For such pressure the issue of existence of global in time weak solutions is a straightforward adaptation of the proof from the case of barotropic system see for example \cite{PLL}, \cite{FN} or \cite{NS}. Below we will only make some general comments concerning the {\it a priori} estimates that are necessary to state the analogous existence result.
First of all, let us formally write the basic energy equality. Multiplying the second equation of \eqref{appr_model_delta} by $\vc{u}$, integrating by parts and using the mass equation we easily get
\[\frac{ d}{dt} \intO{\frac{1}{2} \rho_\delta |\vc{u}_\delta|^2} + \intO{\big(\nabla p+ \rho^*\nabla \pi_{\kappa,\delta}\big)\cdot \vc{u}_\delta}+ \intO{\vc{S}:\nabla \vc{u}_\delta}=0.\]
Let us explain how to deal with the term coming from the pressure for reader's convenience but also to check that everything works with the heterogeneous maximal density:
\begin{align*}
\intO{\big(\nabla p+ \rho^*\nabla \pi_{\kappa,\delta}\big)\cdot \vc{u}_\delta} & = \intO{\gamma\rho_\delta^{\gamma-2}\nabla\rho_\delta\cdot (\rho_\delta\vc{u}_\delta)}\\
& \quad + \intO{\dfrac{\rho^*}{\rho_\delta}\pi_{\kappa,\delta}'\left(\dfrac{\rho_\delta}{\rho^*}\right)\nabla\left(\dfrac{\rho_\delta}{\rho^*}\right) \cdot(\rho_\delta\vc{u}_\delta)} \\
& = \intO{\nabla\lr{\dfrac{\gamma}{\gamma-1}\rho_\delta^{\gamma-1}+Q_{\kappa,\delta}\lr{\dfrac{\rho_\delta}{\rho^*}}} \cdot \big(\rho_\delta\vc{u}_\delta\big)} \\
& = -\intO{\lr{\dfrac{\gamma}{\gamma-1}\rho_\delta^{\gamma-1}+Q_{\kappa,\delta}\lr{\dfrac{\rho_\delta}{\rho^*}}} {\rm div}\,(\rho_\delta\vc{u}_\delta)} \\
& = \intO{\lr{\dfrac{\gamma}{\gamma-1}\rho_\delta^{\gamma-1}+Q_{\kappa,\delta}\lr{\dfrac{\rho_\delta}{\rho^*}}} \partial_t\rho_\delta}
\end{align*}
where we denoted $\displaystyle Q_{\kappa,\delta}'(r)=\frac{\pi_{\kappa,\delta}'(r)}{r}$. Therefore, the energy equality reads
\eq{ \label{energ_est}
\frac{ d}{dt} &\intO{\lr{\frac{1}{2} \rho_\delta |\vc{u}_\delta|^2 + \frac{1}{\gamma-1}\rho_\delta^\gamma+ \rho_\delta\Gamma_{\kappa,\delta}\lr{\frac{\rho_\delta}{\rho^*}}}} \\
&+ \intO{\vc{S}:\nabla\vc{u}_\delta}=0
}
with $\Gamma_{\kappa,\delta}$ such that $\Gamma_{\kappa,\delta}(r)+r\Gamma_{\kappa,\delta}'(r)=Q_{\kappa,\delta}(r)$. To find the expression of $\Gamma_{\kappa,\delta}$ in terms of $\pi_{\kappa,\delta}$, we use the definition of $Q_{\kappa,\delta}$, integrate by part and use that
$[\pi_{\kappa,\delta}(r)/r]\vert_{r=0} =0$. We get
\eq{\label{Gamma}
\Gamma_{\kappa,\delta}(s) = \int_0^{s} {\frac{\pi_{\kappa,\delta}(r)}{r^2}\, {\rm d}r}.}
Integrating \eqref{energ_est} with respect to time gives rise to the following estimates
\begin{equation} \label{un_delta}
\begin{gathered}
\sup_{t\in[0,T]}\lr{\|\sqrt{\rho_\delta} \vc{u}_\delta (t)\|_{L^2(\Omega)}
+\|\rho_\delta (t)\|_{L^{\gamma}(\Omega)}
+\left\|\rho_\delta \Gamma_{\kappa,\delta}\lr{\frac{\rho_\delta}{\rho^*}}(t) \right\|_{L^1(\Omega)} }\leq C,\\
\intT{{\|\vc{u}_\delta\|_{W^{1,2}(\Omega)}^2}} \leq C.
\end{gathered}
\end{equation}
In particular, since
\begin{align*}
\displaystyle \rho_\delta \Gamma_{\kappa,\delta}\lr{\frac{\rho_\delta}{\rho^*}}= \rho_\delta\int_0^{\rho_\delta}{\dfrac{\pi_{\kappa,\delta}(s)}{s^2}\,\mathrm{d}s}
\geq
\dfrac{\kappa}{K-1}\rho_\delta^K
\end{align*}
we obtain that $\rho_\delta$ is bounded in $L^\infty(0,T; L^K(\Omega))$.
\begin{rmk} This computation shows that the approximate of the form
\[\displaystyle\varepsilon\rho^* \nabla \dfrac{\left(\dfrac{\rho}{\rho^*}\right)^\alpha}{\left(1-\dfrac{\rho}{\rho^*}\right)^\beta}\]
gives a good contribution to the energy. The form of the so-called potential energy from \eqref{energ_est} is a natural extension of the one obtained in {\rm \cite{BrPeZa}} and {\rm \cite{Fe}} to the heterogeneus case.
This would not be the case for the approximation
\[\displaystyle\varepsilon\nabla \dfrac{1}{\left(\dfrac{1}{\rho}-\dfrac{1}{\rho^*}\right)^\beta}\]
proposed in {\rm \cite{BeBr}} or {\rm \cite{DeHuNa}}.
\end{rmk}
The above estimate may be used to improve integrability of $\rho_\delta$. Indeed, taking $K$ sufficiently large, say $K>4$, we can test \eqref{AM} by
\[\displaystyle \varphi=\dfrac{\psi(t)}{\rho^*}{\cal B}\lr{\rho_{\delta}-\frac{1}{|\Omega|}\int_\Omega{\rho_{\delta}(y){\rm d}y}},\]
where ${\cal B}$ is the Bogovskii operator (for definition and properties see Lemma \ref{lem_bog} and Proposition \ref{prop_bog} in Appendix).
The details of this testing will be given in Section \ref{uniform_est}. In particular, it gives an additional control (but not uniform with respect to $\kappa$) for the density
\eq{\|\rho_{\delta}\|_{L^{K+1}((0,T)\times \Omega)}\leq C.\label{un_bog}}
Combination of these estimates can be used to deduce that
\eq{\|\rho_\delta|\vc{u}_\delta|^2\|_{L^2(0,T;L^{\frac{6K}{4K+3}}(\Omega))}
+\|\rho_\delta\vc{u}_\delta\|_{L^\infty(0,T; L^{\frac{2K}{K + 1}}(\Omega))\cap L^2(0,T; L^{\frac{6K}{K + 6}}(\Omega))}\leq C.\label{interp}}
These observations allow us to apply methods developed for the barotropic system from \cite{FN, PLL} to justify the following statement.
\begin{prop}[Existence of weak solutions] Let $\varepsilon$, $\kappa$, $\delta$ be fixed and positive. Then, there exists a couple $(\rho_{\delta}, \vc{u}_{\delta})$ solving (\ref{appr_model_delta}-\ref{pressure_ap_p2}) in the sense of distributions on $(0,T)\times\Omega$ with the following regularity properties
\begin{equation}\label{0_reg}
\begin{gathered}
\rho_{\delta} \in L^\infty(0,T;L^{K}(\Omega)) \cap L^{K+1}((0,T)\times \Omega),\\
\rho_{\delta} \in \mathcal{C}([0,T];L^{K}_{\rm weak}(\Omega)) \cap \mathcal{C}([0,T];L^p(\Omega)) , \quad 1\leq p <K, \\
\rho_{\delta} \geq 0 \quad\text{a.e in } (0,T)\times\Omega,
\\
\vc{u}_{\delta}\in L^2(0,T;(W_0^{1,2}(\Omega))^3),
\\
\rho_{\delta} |\vc{u}_{\delta}|^2 \in L^\infty(0,T;L^1(\Omega))\cap L^2(0,T;L^{\frac{6K}{4K+3}}(\Omega)),\\
\rho_{\delta} \vc{u}_{\delta} \in \mathcal{C}([0,T]; (L^{\frac{2K}{K + 1}}_{\rm weak}(\Omega))^ 3)\cap L^2(0,T;(L^{\frac{6K}{K+6}}(\Omega))^3)
\end{gathered}
\end{equation}
and such that the energy inequality
\eq{ \label{energ_in}
\frac{ d}{dt} \intOB{\frac{1}{2} \rho_\delta |\vc{u}_\delta|^2 + \frac{1}{\gamma-1}\rho_\delta^\gamma
+ \rho_\delta\Gamma_{\kappa,\delta}\lr{\frac{\rho_\delta}{\rho^*}}}
+ \intOB{\vc{S}:\nabla\vc{u}_\delta} \leq0
}
is satisfied in the sense of distributions with respect to time.
Moreover $(\rho_{\delta}, \vc{u}_{\delta})$ extended by $0$ outside $\Omega$ is the renormalized solution to the continuity equation in the sense of Definition \ref{df2}.
\end{prop}
\subsection{Uniform estimates}\label{uniform_est}
The goal of this subsection is to provide estimates which are uniform with respect to $\delta$. One of them is estimate \eqref{un_delta}. Note, however, that it does not assure boundedness of the singular part of the pressure $\pi_{\kappa,\delta}$. In fact, such a bound follows from Bogovskii estimate announced in the previous section and giving rise to \eqref{un_bog}. We now check that this estimate assures also the uniform $L^1((0,T)\times\Omega)$ for the pressure.
\bigskip
\ni{\bf $L^1$ bound of the pressure.}
From estimates \eqref{un_delta} it follows that $\rho_\delta\in L^{K}((0,T)\times\Omega)$ with $K \geq 4$, therefore we can test the momentum equation by
\[\varphi(t,x) = \frac{\psi(t)}{\rho^*}\mathcal{B}\left(\rho_\delta - \Ov{\rho_\delta}\right),\qquad \psi(t)\in \mathcal{C}^\infty_0((0,T)),\quad \psi\geq0,\]
where we denoted $\Ov{\rho_\delta}= \frac{1}{|\Omega|}\int_\Omega{\rho_\delta}{\rm d}y$.
We then obtain
\eq{\label{B}
&\intTO{ \psi{\pi_{\kappa,\delta}\lr{\frac{\rho_\delta}{\rho^*}}\left(\rho_\delta - \overline{\rho_\delta}\right)}}
+\intTO{\psi{ p(\rho_\delta)\frac{\rho_\delta}{\rho^*}}}\\
& \qquad\qquad= \intTO{\partial_t(\rho_\delta\vc{u}_\delta)\cdot \varphi} -\intTO{\rho_\delta(\vc{u}_\delta\otimes \vc{u}_\delta) : \nabla \varphi} \\
& \qquad\qquad\quad+ \intTO{\vc{S} : \nabla \varphi}
+\intTO{\psi{ p(\rho_\delta)\frac{\Ov{\rho_\delta}}{\rho^*}}}\\
& \qquad\qquad\quad-\intTO{\psi{ p(\rho_\delta)\nabla\lr{\frac{1}{\rho^*}}\cdot{\cal B}(\rho_\delta-\Ov{\rho_\delta})}}\\
& \qquad\qquad=\sum_{i=1}^5 I_{i}}
A priori estimates obtained in \eqref{un_delta} allow to control the r.h.s. Indeed for the first term we may write
\begin{align*}
I_{1}&=-\intTO{\psi'{\frac{\rho_\delta}{\rho^*}\vc{u}_\delta \cdot \mathcal{B} \left(\rho_\delta - \overline{\rho_\delta}\right)}} -\intTO{\psi {\frac{\rho_\delta}{\rho^*}\vc{u}_\delta \cdot \partial_t\mathcal{B} \left(\rho_\delta - \overline{\rho_\delta}\right)}}.
\end{align*}
Therefore, if only $K>3$ and since $\psi, \psi'$ and $\lr{\rho^*}^{-1}$ are bounded in $L^\infty((0,T)\times\Omega)$ and thanks to the mass equation we get
\eqh{
|I_{1}|& \leq C\intT{\|{\rho_\delta}\|_{L^{3}(\Omega)}\|\vc{u}_\delta\|_{L^6(\Omega)}\|\mathcal{B} \left(\rho_\delta - \overline{\rho_\delta}\right)\|_{L^{6/5}(\Omega)}}\\
& \quad+ C\intT{\|{\rho_\delta}\|_{L^3(\Omega)}\|\vc{u}_\delta\|_{L^6(\Omega)}\|\mathcal{B} \left({\rm div} (\rho_\delta\vc{u}_\delta )\right)\|_{L^{6/5}(\Omega)}} \\
& \leq C\intT{\|{\rho_\delta}\|^2_{L^{3}(\Omega)}\|\vc{u}_\delta\|_{L^6(\Omega)}}\\
& \quad+ C\intT{\|{\rho_\delta}\|_{L^3(\Omega)}^2\|\vc{u}_\delta\|_{L^6(\Omega)}^2} \\
& \leq C.
}
Similarly, for the second term, we have
\eqh{
I_2=&
\intTO{\psi\frac{\rho_\delta}{\rho^*}\lr{\vc{u}_\delta\otimes\vc{u}_\delta}:\nabla \mathcal{B}\left(\rho_\delta - \Ov{\rho_\delta}\right)}\\
&+\intTO{\psi{\rho_\delta(\vc{u}_\delta\otimes \vc{u}_\delta)\nabla\lr{\frac{1}{\rho^*}}\cdot{\cal B}(\rho_\vep-\Ov{\rho_\vep})}},}
thus
\begin{align*}
|I_{2}|&\leq C\intTO{{\rho_\delta|\vc{u}_\delta|^2\lr{|\nabla \mathcal{B}\big(\rho_\delta-\overline{\rho_\delta}\big) |+| \mathcal{B}\big(\rho_\delta-\overline{\rho_\delta}\big) |}}}\\
&\leq C\intT{\|\rho_\delta\|^2_{L^3(\Omega)}\|\vc{u}_\delta\|^2_{L^6(\Omega)}} \leq C.
\end{align*}
For the stress tensor $I_3$ we may write
\eqh{
|I_{3}|& \leq C\intT{\|\nabla\vc{u}_\delta\|_{L^2(\Omega)}\|\rho_\delta\|_{L^2(\Omega)}}\leq C.
}
The remaining pressure terms $I_{4}$ and $I_5$ can be easily controlled using the uniform $L^\infty(0,T;L^K(\Omega))$ bound for $\rho_\delta$ obtained in \eqref{un_delta}, if only $K$ is sufficiently large
\eqh{
|I_4|\leq C\intT{\|\rho_\delta\|_{L^1(\Omega)}\|\rho_\delta\|_{L^\gamma(\Omega)}^\gamma}\leq C
}
for $K>\gamma$ and
\eqh{
|I_5|\leq C\intT{\|\rho_\delta\|_{L^{\frac{2\gamma K}{K-1}}(\Omega)}^\gamma\|\rho_\delta\vc{u}_\delta\|_{L^{\frac{2K}{K+1}}(\Omega)}}
\leq C}
for $K>2\gamma+1$. Note that in above estimate we essentially use the $L^\infty((0,T)\times\Omega)$ bound on $\nabla\lr{\dfrac{1}{\rho^*}}$, which follows from assumptions on $\rho^*$ in Theorem \ref{main}.
Collecting these estimates, we verify that the l.h.s. of \eqref{B} is bounded uniformly with respect to $\delta$.
Now, let us split the first term into two parts as follows
\begin{align} \label{splitting}
&\intTO{\psi{\pi_{\kappa,\delta}\lr{\dfrac{\rd}{\rs}}\left(\rho_\delta -\Ov{\rho_\delta}\right)}} \nonumber \\
\displaystyle &\quad = \intTO{\psi{\pi_{\kappa,\delta}\lr{\dfrac{\rd}{\rs}}\left(\rho_\delta -\Ov{\rho_\delta}\right)\vc{1}_{\displaystyle\left\lbrace\rho_\delta < \dfrac{\inf\rho^* + M_0}{2}\right\rbrace}}} \\
\displaystyle & \qquad + \intTO{\psi{\pi_{\kappa,\delta}\lr{\dfrac{\rd}{\rs}}\left(\rho_\delta -\Ov{\rho_\delta}\right)\vc{1}_{\displaystyle\left\lbrace\rho_\delta \geq \dfrac{\inf\rho^* + M_0}{2}\right\rbrace}}}\nonumber
\end{align}
In the first integral, the term $\pi_{\kappa,\delta}\lr{\dfrac{\rd}{\rs}}$ is far from the singularity $\rho^*$, so the integral is finite.
For the second integral we may write
\begin{align*} \displaystyle&\intTO{\psi{\pi_{\kappa,\delta}\lr{\dfrac{\rd}{\rs}}\left(\rho_\delta -\Ov{\rho_\delta}\right)\vc{1}_{\displaystyle\left\lbrace\rho_\delta \geq \dfrac{\inf\rho^* + M_0}{2}\right\rbrace}}}\\
\displaystyle& \quad \geq \frac{\inf \rho^* -M_0}{2} \intTO{\psi{\pi_{\kappa,\delta}\lr{\dfrac{\rd}{\rs}}}\vc{1}_{\displaystyle\left\lbrace\rho_\delta \geq \dfrac{\inf\rho^* + M_0}{2}\right\rbrace}}.
\end{align*}
Thanks to \eqref{add_hyp_ini}, we deduce from the second integral of \eqref{splitting} that
\eqh{\intTO{\pi_{\kappa,\delta}\lr{\dfrac{\rd}{\rs}}} \leq C
\label{pressure_bound_delta}}
and then the control of the first integral of \eqref{splitting} yields
\eqh{{\intTO{\rho_\delta \pi_{\kappa,\delta}\lr{\dfrac{\rd}{\rs}}}} \leq C.}
In particular, as mentioned in \eqref{un_bog}, $\rho_\delta$ is bounded in $L^{K+1}((0,T)\times\Omega)$
uniformly with respect to $\delta$, but not uniformly with respect to $\kappa$.
\bigskip
\ni{\bf Equi-integrability of the pressure.} In order to perform the limit passage $\delta\rightarrow 0$ and to recover system (\ref{AC}-\ref{AM}) with parameter $\kappa$, the weak limit of $\pi_{\kappa,\delta}\lr{\dfrac{\rd}{\rs}}$ has to be more regular than merely a positive measure. To show that it converges weakly in $L^1((0,T)\times \Omega)$ to $\pi_{\kappa}\lr{\dfrac{\rho_\kappa}{\rho^*}}$ we want to apply the {\sc De~La~Vall\'ee-Poussin} criterion and we test the momentum equation by
\[\varphi(t,x) = \frac{\psi(t)}{\rho^*}\mathcal{B}\left(\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right) - \Ov{\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right)}\right), \quad\psi(t)\in {\mathcal C}^\infty_0((0,T)),\quad \psi\geq0,\]
where
\begin{equation}\label{eta_function}
\eta_\delta(s) = \begin{cases} \,\,\log (1-s) & \text{ if } s \leq 1-\delta, \\
\,\,\log (\delta) & \text{ if } s > 1-\delta. \end{cases}
\end{equation}
This testing results in
\begin{align}\label{e-i}
&\intTO{ \psi\lr{\pi_{\delta}\lr{\frac{\rho_\delta}{\rho^*}}+\frac{p(\rho_\delta)}{\rho^*}}\left(\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right) - \overline{\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right)}\right)} \nonumber \\
& = \intTO{\partial_t(\rho_\delta\vc{u}_\delta)\cdot \varphi} -\intTO{\rho_\vep(\vc{u}_\vep\otimes \vc{u}_\vep) : \nabla \varphi} \nonumber\\
& \quad+ \intTO{\vc{S} : \nabla \varphi}-\intTO{\psi{ p(\rho_\delta)\nabla\lr{\frac{1}{\rho^*}}\cdot{\cal B}\lr{\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right)-\Ov{\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right)}}}}\nonumber\\
& =\sum_{i=1}^4 J_{i}.\end{align}
Similarly as in the previous paragraph, the most demanding term $J_1$ equals to
\eqh{J_1 = &-\intTO{\psi' {\frac{\rho_\delta}{\rho^*}\vc{u}_\delta \cdot \mathcal{B} \left(\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right) - \overline{\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right)}\right)}} \\
&-\intTO{\psi {\frac{\rho_\delta}{\rho^*}\vc{u}_\delta \cdot \partial_t\mathcal{B} \left(\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right) - \overline{\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right)}\right)}}}
We can generalize the notion of renormalized solutions of the continuity equation (see the Appendix and \cite{NS} section 6.2) to functions
\[b_k(s) = \begin{cases} & b(s) \text{ if } s\in[0,k) \\
& b(k) \text{ if } s\in[k,\infty) \end{cases} \quad \text{with } b\in {\cal C}^1[0,\infty)\]
to deduce that $\eta_\delta$ satisfies the equation
\begin{align} \label{renorm_eta}
&\partial_t \eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right) + {\rm div}\left(\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right) \vc{u}_\delta\right) + \left((\eta_\delta)'_+\left(\dfrac{\rho_\delta}{\rho^*}\right)\dfrac{\rho_\delta}{\rho^*} -\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right)\right) {\rm div} \vc{u}_\delta \nonumber\\
&\quad + (\eta_\delta)'_+\left(\dfrac{\rho_\delta}{\rho^*}\right)\dfrac{\rho_\delta}{\rho^*}\vc{u}_\delta\cdot\nabla\log\rho^*=0.
\end{align}
In the above formula $(\eta_\delta)'_+$ denotes the right derivative of $\eta_\delta$
\[(\eta_\delta)'_+ (s)=\begin{cases} \, \dfrac{1}{1-s} &\text{ if } (x,t)\in\{s(x,t) < 1-\delta\} \\
\, 0 & \text { if }(x,t)\in\{s(x,t) \geq 1-\delta\}\end{cases}.\]
We thus obtain
\eq{\label{J1}
J_1=& -\intTO{\psi'{\frac{\rho_\delta}{\rho^*}\vc{u}_\delta \cdot \mathcal{B} \left(\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right) - \overline{\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right)}\right)}} \\
&- \intTO{\psi {\frac{\rho_\delta}{\rho^*}\vc{u}_\delta \cdot \mathcal{B} \left({\rm div}\left(\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right)\vc{u}_\delta\right)\right)}} \\
& -\intTO{ \psi\frac{\rho_\delta}{\rho^*}\vc{u}_\delta\cdot\mathcal{B}\left[F_\delta\right]}
}
where we denoted
\begin{align*}
F_\delta=&\left(\eta_\delta'\left(\dfrac{\rho_\delta}{\rho^*}\right)\dfrac{\rho_\delta}{\rho^*} -\dfrac{\rho_\delta}{\rho^*}\right) {\rm div} \vc{u}_\delta + \eta_\delta'\left(\dfrac{\rho_\delta}{\rho^*}\right)\dfrac{\rho_\delta}{\rho^*}\vc{u}_\delta\cdot\nabla\log\rho^* \\ &-\frac{1}{|\Omega|}\int_\Omega{\left[\left(\eta_\delta'\left(\dfrac{\rho_\delta}{\rho^*}\right)\dfrac{\rho_\delta}{\rho^*} -\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right)\right) {\rm div} \vc{u}_\delta + \eta_\delta'\left(\dfrac{\rho_\delta}{\rho^*}\right)\dfrac{\rho_\delta}{\rho^*}\vc{u}_\delta\cdot\nabla\log\rho^* \right]{\rm d}y}.
\end{align*}
Therefore, to control the last integral of $J_1$ we need bounds on $\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right)$ and $\eta_\delta'\left(\dfrac{\rho_\delta}{\rho^*}\right)\dfrac{\rho_\delta}{\rho^*}$. To simplify, we assume that $\alpha, \beta$ are integers. For $\dfrac{\rho_\delta}{\rho^*} < 1-\delta$ we have
\begin{align*}
\displaystyle\rho_\delta \Gamma_{\kappa,\delta}\lr{\dfrac{\rho_\delta}{\rho^*}} &
\geq \rho_\delta \int_0^{\frac{\rho_\delta}{\rho^*}}{\varepsilon\dfrac{s^{\alpha-2}}{(1-s)^\beta} {\rm d}s} \\
\displaystyle & = \varepsilon \rho_\delta \int_{1-\frac{\rho_\delta}{\rho^*}}^1{\dfrac{(1-s)^{\alpha-2}}{s^\beta}ds} \\
& = \varepsilon \rho_\delta \sum_{k=0}^{\alpha-2}\begin{pmatrix} \alpha-2 \\ k \end{pmatrix} (-1)^k\int_{1-\frac{\rho_\delta}{\rho^*}}^1{s^{k-\beta}ds}\\
& = \dfrac{\varepsilon \rho_\delta}{\beta-1}\dfrac{1}{\lr{1-\dfrac{\rho_\delta}{\rho^*}}^{\beta-1}}-\dfrac{\varepsilon \rho_\delta}{\beta-1}\\
&\quad+\varepsilon \rho_\delta \sum_{k=1}^{\alpha-2}\begin{pmatrix} \alpha-2 \\ k \end{pmatrix} (-1)^k\int_{1-\frac{\rho_\delta}{\rho^*}}^1{s^{k-\beta}ds} \\
& \geq \dfrac{\varepsilon \rho_\delta}{\lr{1-\dfrac{\rho_\delta}{\rho^*}}^{\beta-1}}\left(\dfrac{1}{\beta-1}-\delta C(\alpha,\beta,\delta)\right) -C_2
\end{align*}
where, in the last inequality, $C_2$ controls all the non-degenerate terms. The constant $C(\alpha,\beta,\delta)$ is such that $\lim_{\delta\rightarrow 0} |C| < +\infty$ and it controls all the degenerate terms of lower order. Taking $\delta$ sufficiently small, we ensure that
\[\delta|C(\alpha,\beta,\delta)|<\dfrac{1}{2(\beta-1)}.\]
Thus it follows that
\begin{equation}\label{ineg_equi}
\displaystyle\rho_\delta \Gamma_{\kappa,\delta}\lr{\frac{\rho_\delta}{\rho^*}} \geq C_1\frac{\varepsilon\rho_\delta}{\lr{1-\frac{\rho_\delta}{\rho^*}}^{\beta-1}} - C_2, \qquad C_1 >0 .
\end{equation}
Now, since for $\dfrac{\rho_\delta}{\rho^*} < 1-\delta$
\[\eta'_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right) = \dfrac{1}{\left(1-\dfrac{\rho_\delta}{\rho^*}\right)}\]
we obtain using the Young inequality ($\beta>3$)
\[C_1|\rho_\delta\eta'_\delta(\rho_\delta)|^2 \leq \rho_\delta \Gamma_{\kappa,\delta}\lr{\frac{\rho_\delta}{\rho^*}} +C_2.\]
This inequality is still satisfied for $\dfrac{\rho_\delta}{\rho^*} \geq 1-\delta$ because $\eta'\equiv 0$ on this set.\\
Concerning the control of $\eta_\delta$, if $\dfrac{\rho_\delta}{\rho^*} < 1-\delta$
we observe that
\[(1-R)^{\beta-1}|\log(1-R)|^q \underset{R\rightarrow 1^-}{\longrightarrow} 0 \quad \forall 1\leq q< \infty\]
which proves that there exists $C_1$ and $C_2$ uniform with respect to $\delta$ such that
\begin{equation} \label{bound_eta}
C_1(q,\varepsilon,\beta)\left|\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right)\right|^q \leq \rho_\delta \Gamma_{\kappa,\delta}\left(\dfrac{\rho_\delta}{\rho^*}\right) +C_2(q,\varepsilon,\beta)\quad \forall q\in [1,\infty). \end{equation}
Since $\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right)$ is constant for $\dfrac{\rho_\delta}{\rho^*} \geq 1-\delta$, we deduce that this inequality is satisfied for all $\dfrac{\rho_\delta}{\rho^*}$.
These estimates allow to control $\mathcal{B}\left[F_\delta\right] $ from \eqref{J1}
uniformly in $L^2(0,T;L^q(\Omega))$ with $q<3/2$. So, the full integrant in the third term of $J_1$ is controlled provided $\rho_\delta\vc{u}_\delta\in L^2(0,T;(L^{p}(\Omega))^3)$ with $p> 3$, which due to \eqref{interp} asks for $K>6$.
The other $J_k$ from \eqref{e-i} are bounded thanks to \eqref{bound_eta} similarly as in the case of $L^1$ bound on the pressure obtained in previous paragraph.
Finally, from \eqref{e-i} we deduce that
\begin{equation}\label{equiintegr}
\intTO{\pi_{\delta}\lr{\frac{\rho_\delta}{\rho^*}}\eta_\delta\left(\dfrac{\rho_\delta}{\rho^*}\right)} \leq C,
\end{equation}
which implies, thanks to the {\sc De La Vall\'ee-Poussin} criterion, the equi-integrability of the pressure $\pi_{\delta}$.
\subsection{Passage to the limit $\delta \to 0$, $\kappa\to0$}
First, let us perform the limit passage $\delta\to0$. In fact the second limit passage $\kappa\to0$ is an immediate consequence of the first one and of new uniform estimates. Indeed, as it will turn out at the end of this subsection, after letting $\delta$ to $0$ one recovers estimate \eqref{un_main}. This estimate can be used to substitute all $\kappa$-dependent estimates used to pass to the limit with $\delta$.
\noindent We start with an observation that according to \eqref{un_delta}, there exist subsequences $\rho_\delta,\vc{u}_\delta$ such that
\eq{\label{conv}
&\rho_\delta\longrightarrow \rho \qquad \text{weakly-* in } L^{\infty}(0,T;L^{K}(\Omega)), \\
& \vc{u}_\delta \longrightarrow \vc{u} \qquad \text{weakly in } L^2(0,T;(W^{1,2}(\Omega))^3).
}
Next, directly from the continuity equation it follows that $\rho_\delta$ is equi-continuous with values in $W^{-1,\frac{2K}{K+1}}(\Omega)$. Moreover $\rho_\delta$ belongs to $\mathcal{C}([0,T];L^{K}_{\rm weak}(\Omega))$ and is uniformly bounded in $L^\infty(0,T; L^{K}(\Omega))$, thus by the Arzel\`a-Ascoli theorem we verify that
\begin{equation}\label{cvg_dens}
\rho_\delta\longrightarrow \rho \qquad \text{in }\ \mathcal{C}([0,T];L^{K}_{\rm weak}(\Omega)).
\end{equation}
Similarly, we prove that
\begin{equation}\label{cvg_momentum}
\rho_\delta\vc{u}_\delta \longrightarrow \rho \vc{u} \quad \text{in }\ \mathcal{C}([0,T]; (L_{\rm weak}^{\frac{2K}{K+1}}(\Omega))^3).
\end{equation}
Indeed, from the momentum equation we conclude that $\rho_\delta\vc{u}_\delta$ is equi-continuous with values in $W^{-1,s}(\Omega)$ with $s=\frac{K+1}{K}$. Moreover, on account of \eqref{interp} it is uniformly bounded in $L^\infty(0,T;(L^{\frac{2K}{K+1}}(\Omega))^3)$ and belongs to $\mathcal{C}([0,T]; (L_{\rm weak}^{\frac{2K}{K+1}}(\Omega))^3)$, thus the Arzel\`a-Ascoli theorem yields \eqref{cvg_momentum}.
\bigskip
\noindent{\bf Passage to the limit in the continuity and the momentum equations.}
Convergences \eqref{cvg_dens} and \eqref{cvg_momentum} allow us to pass to the limit in the approximate continuity equation, which is now satisfied in the sense of distributions on $(0,T)\times\Omega$.
To pass to the limit in the momentum equation we need to justify the convergence in the nonlinear terms. For the convective term we have
\begin{equation}\label{cvg_conv}
\rho_\delta\vc{u}_\delta\otimes \vc{u}_\delta \longrightarrow \rho \vc{u} \otimes \vc{u} \quad \text{weakly in } L^q((0,T)\times \Omega)
\end{equation}
for some $q>1$. It follows from \eqref{interp} which gave us uniform bound on $\rho_\delta |\vc{u}_\delta|^2$ in $L^2(0,T;L^p(\Omega))$
with $\displaystyle p= \frac{6K}{4K + 3}>1$, the convergence established in \eqref{cvg_momentum} and in
\eqref{conv} together with compact imbeddings.
Finally, thanks to \eqref{equiintegr} we deduce existence of a subsequence such that
\eq{\pi_{\kappa,\delta}\lr{\dfrac{\rd}{\rs}}\to \overline{\pi_{\varepsilon}\lr{\frac{\rho}{\rho^*}}} \text{ weakly in } L^1((0,T)\times \Omega),}
however, identification of the limit term cannot be done yet, since $\rho_\delta$ is only known to converge weakly to $\rho$. Nevertheless, with these convergences at hand, the passage to the limit in the continuity equation is automatic. Letting $\delta\rightarrow 0$ in the weak formulation of the momentum equation, we get
\eq{\label{weak_mom_delta}
&\intTO{\rho \vc{u}\cdot\partial_t\varphi }+ \intO{ \vc{m}_0\cdot\varphi(0)} \\
&+\intTO{\rho(\vc{u}\otimes \vc{u}):\nabla\varphi} -\intTO{\vc{S} : \nabla\varphi } \\
&+\intTO{\overline{p(\rho)} {\rm div}\varphi} +\intTO{ \overline{\pi_{\varepsilon}\lr{\frac{\rho}{\rho^*}}}{\rm div}(\rho^*\varphi)}= 0
}
satisfied for any $\varphi\in C^\infty_c\lr{[0,T)\times\Ov{\Omega}}^3$ such that $\varphi\big|_{\partial\Omega}=0$.
The final step is to identify
\eq{\label{identy}
\overline{p(\rho)}=p(\rho), \quad \mbox{and}\quad \overline{\pi_{\kappa}\lr{\frac{\rho}{\rho^*}}}= \pi_{\kappa}\lr{\frac{\rho}{\rho^*}}.}
To do that we need to show the strong convergence of the density.
\bigskip
\noindent {\bf Strong convergence of the density.} Proving the strong convergence of the density is a standard difficulty in the study of compressible Navier-Stokes equations (see \cite{PLL}, \cite{NS}). The main ingredients are the theory of renormalized solutions to the continuity equation (whose definition is recalled in Appendix with the main existence result) and the compactness of a quantity called {\it the effective flux}. We follow in this part the standard theory of compressible Navier-Stokes equations.
The renormalized version of the continuity equation \eqref{renormal} with $b(\rho_\delta)=\rho_\delta\log\rho_\delta$ (we extend $\rho_\delta$ and $\vc{u}_\delta$ by $0$ outside $\Omega$) reads
\begin{equation}\label{renorm}
\partial_t(\rho_\delta\log\rho_\delta) + {\rm div}(\rho_\delta\log\rho_\delta\vc{u}_\delta) = -\rho_\delta{\rm div} \vc{u}_\delta \quad \text{in}\quad \mathcal{D}'((0,T)\times \mathbb{R}^3).
\end{equation}
We now want to pass to the limit $\delta\to 0$ in the above equality. Due to \eqref{hyp_renorm_2} with $\lambda_1=1$ and the weak-* convergence of $\rho_\delta$ in $L^\infty(0,T;L^K(\Omega))$ established in previous paragraph, we have
\begin{eqnarray*}
& b(\rho_\delta)\longrightarrow \overline{b(\rho)} \quad \text{weakly-* in } L^\infty(0,T;L^{\frac{K}{2}}(\Omega)) &\\
& (\rho_\delta b'(\rho_\delta)-b(\rho_\delta)){\rm div}\vc{u}_\delta \rightarrow \overline{(\rho b'(\rho)-b(\rho)){\rm div} \vc{u}} \quad \text{weakly in } L^2(0,T;L^{\frac{2K}{K+4}}(\Omega))&
\end{eqnarray*}
To prove the convergence in the term $\rho_\delta\log\rho_\delta\,\vc{u}_\delta$ we need the following compensated compactness lemma (see for instance \cite{PLL} Lemma 5.1).
\begin{Lemma}\label{lem_compac} Let $g_n$, $h_n$ converge weakly to $g$, $h$ respectively in $L^{p_1}(0,T;L^{p_2}(\Omega))$, $L^{q_1}(0,T;L^{q_2}(\Omega))$ where $1\leq p_1,p_2 \leq \infty$, $\frac{1}{p_1}+ \frac{1}{q_1}=\frac{1}{p_2}+ \frac{1}{q_2}= 1$. We assume in addition that
\begin{itemize}
\item $\partial_t g_n$ is bounded in $L^1(0,T;W^{-m,1}(\Omega))$ for some $m\geq 0$ independent of $n$.
\item $\|h_n-h_n(t,\cdot+\xi)\|_{L^{q_1}(L^{q_2})} \rightarrow 0$ as $|\xi|\rightarrow 0$, uniformly in $n$.
\end{itemize}
Then $g_nh_n$ converges to $gh$ in $\mathcal{D}'$.
\end{Lemma}
This result applied to $g_\delta=\rho_\delta\log\rho_\delta$ and $h_\delta=\vc{u}_\delta$ gives the convergence of $\rho_\delta\log\rho_\delta\vc{u}_\delta$ to $\overline{\rho \log\rho}\vc{u}$, and, passing to the limit $\delta \rightarrow 0$ in \eqref{renorm}, we check that
\eq{\partial_t (\overline{\rho \log\rho} )+ {\rm div}(\overline{\rho\log \rho}\vc{u}) =-\overline{\rho {\rm div} \vc{u}}\label{lim_ren}}
in the sense of distributions.
Comparing \eqref{lim_ren} and the renormalized continuity equation satisfied by the limit functions $\rho$, $\vc{u}$ with $b(\rho)=\rho\log\rho$ we obtain
\begin{equation}\label{ineg_log}
\partial_t(\overline{\rho\log\rho}-\rho\log\rho) + {\rm div}[(\overline{\rho\log\rho}-\rho\log\rho)\vc{u}]=-\overline{\rho{\rm div} \vc{u}} +\rho{\rm div} \vc{u}.
\end{equation}
We will exploit this equality to show compactness of the, so called, effective viscous flux.
To derive a key equality for this reasoning, we introduce the the inverse divergence operator
$\mathcal{A}=\nabla\Delta^{-1}$
and the double Riesz transform
$\mathcal{R}=\nabla\otimes\nabla\Delta^{-1}$
specified by \eqref{defA} and \eqref{defR} in Appendix, where we also recall some of their basic properties.
\bigskip
Applying the operator $\nabla \Delta^{-1}$ to the approximate continuity equation, we get
\[\partial_t\nabla\Delta^{-1}(\rho_\delta\mathbf{1}_\Omega) + \nabla \Delta^{-1}{\rm div}(\rho_\delta\vc{u}_\delta) = 0.\]
Due to Lemma \ref{LA1}, it follows that $\nabla\Delta^{-1}(\rho_\delta\mathbf{1}_\Omega) \in L^{\alpha+1}(0,T;W^{1,\alpha+1}(\Omega))\cap L^\infty(0,T;W^{1,\alpha}(\Omega))$ and $\partial_t\nabla\Delta^{-1}(\rho_\delta\mathbf{1}_\Omega) \in L^2((0,T)\times\Omega)$.
Thus
$$\varphi_\delta = \psi\phi\nabla\Delta^{-1}[\mathbf{1}_\Omega\rho_\delta]$$
with $\phi\in \mathcal{D}(\Omega)$ and $\psi\in\mathcal{D}((0,T))$ is an admissible test function for the momentum equation. After straightforward calculation we obtain
\eq{\label{eff_flux_delta}
&\intT{\psi\intO{\phi\rho_\delta\left[p(\rho_\delta)+\rho^*\pi_{\kappa,\delta}\lr{\dfrac{\rd}{\rs}} - (2\mu+\lambda){\rm div}\vc{u}_\delta\right]}} \\
& = -\intT{\psi\intO{\left[p(\rho_\delta)+\rho^*\pi_{\kappa,\delta}\lr{\dfrac{\rd}{\rs}}\right] \partial_i\phi\mathcal{A}_i(\rho_\delta\mathbf{1}_\Omega)}} \\
& \quad-\intT{\psi\intO{\phi\pi_{\kappa,\delta}\lr{\dfrac{\rd}{\rs}} \partial_i\rho^*\mathcal{A}_i(\rho_\delta\mathbf{1}_\Omega)}} \\
& \quad + (\mu+\lambda)\intT{\psi\intO{\partial_i\phi{\rm div}\vc{u}_\delta\mathcal{A}_i(\rho_\delta\mathbf{1}_\Omega)}}\\
&\quad+ \mu\intT{\psi\intO{\partial_j\phi\partial_ju_\delta^i\mathcal{A}_i(\rho_\delta\mathbf{1}_\Omega)}}\\
& \quad -\intT{\psi\intO{\partial_j\phi\rho_\delta u_\delta^i u_\delta^j\mathcal{A}_i(\rho_\delta\mathbf{1}_\Omega)}} \\
&\quad -\intT{\partial_t\psi\intO{\phi\rho_\delta u_\delta^i\mathcal{A}_i(\rho_\delta\mathbf{1}_\Omega)}} \\
& \quad +\intT{\psi\intO{u_\delta^j\left[\rho_\delta\mathbf{1}_\Omega\mathcal{R}_{ij}\big(\rho_\delta u_\delta^i\phi\big) -\rho_\delta u_\delta^i\phi\mathcal{R}_{ij}(\rho_\delta\mathbf{1}_\Omega)\right]}}
}
Using the convergences established above and the properties of operator $\mathcal{A}_i$ (Lemma \ref{LA1}), we may pass to the limit $\delta\to 0$ in all the terms of the above formula except the last one.
Since $\int_{\mathbb{R}^3}{\mathcal{R}_{ij}(f)g} = \int_{\mathbb{R}^3}{\mathcal{R}_{ij}(g)f}$, for all $ f\in L^r(\mathbb{R}^3),\, g\in L^{r'}(\mathbb{R}^3)$
we deduce that the last term can be rewritten as
\eq{\label{com}
&\intT{\psi\intO{\rho_\delta\left[u_\delta^j\mathcal{R}_{ij}\big(\rho_\delta u_\delta^i\phi\big) -\mathcal{R}_{ij}\big(\rho_\delta u_\delta^i\phi u_\delta^j\big)\right]}} \\
& = \intT{\psi\intO{\rho_\delta[u_\delta^j,\mathcal{R}_{ij}](\phi\rho_\delta u_\delta^i)}}.
}
Due to \eqref{un_delta} and \eqref{interp}, Lemma \ref{LA3} may be applied to control\begin{equation*}
[\vc{u}_\delta,\mathcal{R}](\phi\rho_\delta\vc{u}_\delta) \text{ in } L^1(0,T;W^{1,q}(\Omega)) \text{ with } \frac{1}{q} = \frac{1}{2}+\frac{\alpha+ 6}{6\alpha}=\frac{2\alpha+3}{3\alpha},
\end{equation*}
so, since $\alpha>3$, $q\in (1,\frac{3}{2})$.\\
Next, we identify the limit of $[\vc{u}_\delta,\mathcal{R}](\phi\rho_\delta\vc{u}_\delta)$. Due to \eqref{cvg_conv} we already know that $\mathcal{R}_{ij}(\rho_\delta u_\delta^i u_\delta^j)$ converges weakly to $\mathcal{R}_{ij}(\rho u^i u^j)$ thus we only have to justify the weak convergence of $u_\delta^j\mathcal{R}_{ij}(\rho_\delta u_\delta^i)$. This easily follows by application of Lemma \ref{lem_compac} to $g_\delta=\mathcal{R}_{ij}(\phi \rho_\delta\vc{u}_\delta^i)$ and $h_\delta=u_\delta^j$. Thus, Lemma \ref{LA2} yields
\begin{equation}
[u_\delta^j,\mathcal{R}_{ij}](\phi\rho_\delta u_\delta^i) \quad\text{converges weakly to }[u^j,\mathcal{R}_{ij}](\phi\rho u^i)\quad \text{in } L^1(0,T,L^p(\Omega)),
\end{equation}
for $p$ such that $W^{1,q}(\Omega) \hookrightarrow L^p(\Omega)$, namely
\[p < \frac{3q}{3-q} = \frac{3\alpha}{\alpha{+}3}.\]
Observe that since $\alpha>3$ thus $p\leq3/2$ and therefore the integral on the r.h.s. of \eqref{com} is bounded provided $\rho_\delta\in L^\infty(0,T;L^{p'}(\Omega))$ with $p'\geq 3$, which is satisfied for $K\geq3$. Finally, applying once more Lemma \ref{lem_compac} this time with $g_\delta=\rho_\delta$ and $h_\delta=[u_\delta^j,\mathcal{R}_{ij}](\phi\rho_\delta u_\delta^i) $ we identify the limit of \eqref{com}.
Therefore, passing to the limit $\delta\to 0$ in \eqref{eff_flux_delta}, we obtain
\eq{\label{eff_flux_lim}
&\lim_{\delta\to 0}\intT{\psi\intO{\phi\rho_\delta\left[p(\rho_\delta)+\rho^*\pi_{\kappa,\delta}\lr{\dfrac{\rd}{\rs}} - (2\mu+\lambda){\rm div}\vc{u}_\delta\right]}} \\
& = -\intT{\psi\intO{\left[\Ov{p(\rho)}+\rho^* \overline{\pi_{\varepsilon}\lr{\frac{\rho}{\rho^*}}}\right] \partial_i\phi\mathcal{A}_i(\rho_\delta\mathbf{1}_\Omega)}} \\
& \quad-\intT{\psi\phi\intO{ \overline{\pi_{\varepsilon}\lr{\frac{\rho}{\rho^*}}}\partial_i\rho^*\mathcal{A}_i(\rho\mathbf{1}_\Omega)}} \\
& \quad + (\mu+\lambda)\intT{\psi\intO{\partial_i\phi{\rm div} \vc{u}\mathcal{A}_i(\rho\mathbf{1}_\Omega)}}\\
&\quad+ \mu\intT{\psi\intO{\partial_j\phi\partial_j u^i\mathcal{A}_i(\rho\mathbf{1}_\Omega)}}\\
& \quad -\intT{\psi\intO{\partial_j\phi\rho u^i u^j\mathcal{A}_i(\rho\mathbf{1}_\Omega)}} -\intT{\partial_t\psi\intO{\phi\rho u^i\mathcal{A}_i(\rho\mathbf{1}_\Omega)}} \\
& \quad +\intT{\psi\intO{\rho\left[u^j,\mathcal{R}_{ij}\right](\phi\rho u^i)}}.
}
This is to be compared with analogous expression obtained for a limit momentum equation \eqref{weak_mom_delta} with
$\varphi = \psi\phi\nabla\Delta^{-1}[\mathbf{1}_\Omega\rho]$, where $\phi\in \mathcal{D}(\Omega)$ and $\psi\in\mathcal{D}((0,T))$, we have
\eq{\label{eff_flux2}
&\intT{\psi\intO{\phi\rho\left[\overline{p(\rho)}+ \rho^*\overline{\pi_{\varepsilon}\lr{\frac{\rho}{\rho^*}}} - (2\mu+\lambda){\rm div} \vc{u}\right]}} \\
& = -\intT{\psi\intO{\left[\overline{p(\rho)}+ \rho^*\overline{\pi_{\varepsilon}\lr{\frac{\rho}{\rho^*}}}\right] \partial_i\phi\mathcal{A}_i(\rho\mathbf{1}_\Omega)}} \\
& \quad-\intT{\psi\phi\intO{ \overline{\pi_{\varepsilon}\lr{\frac{\rho}{\rho^*}}}\partial_i\rho^*\mathcal{A}_i(\rho\mathbf{1}_\Omega)}} \\
& \quad + (\mu+\lambda)\intT{\psi\intO{\partial_i\phi{\rm div} \vc{u}\mathcal{A}_i(\rho\mathbf{1}_\Omega)}}\\
&\quad+ \mu\intT{\psi\intO{\partial_j\phi\partial_j u^i\mathcal{A}_i(\rho\mathbf{1}_\Omega)}}\\
& \quad -\intT{\psi\intO{\partial_j\phi\rho u^i u^j\mathcal{A}_i(\rho\mathbf{1}_\Omega)}} -\intT{\partial_t\psi\intO{\phi\rho u^i\mathcal{A}_i(\rho\mathbf{1}_\Omega)}} \\
& \quad +\intT{\psi\intO{\rho\left[u^j,\mathcal{R}_{ij}\right](\phi\rho u^i)}}.
}
Comparing \eqref{eff_flux_lim} with \eqref{eff_flux2} we get
\eq{\label{ef}
&\lim_{\delta\to 0}\intT{\psi\intO{\phi\rho_\delta\left[p(\rho_\delta)+\rho^* \pi_{\kappa,\delta}\lr{\dfrac{\rd}{\rs}}- (2\mu+\lambda){\rm div}\vc{u}_\delta\right]}} \\
& \qquad = \intT{\psi\intO{\phi\rho\left[\overline{p(\rho)}
+\rho^*\overline{\pi_{\kappa}\lr{\frac{\rho}{\rho^*}}}- (2\mu+\lambda){\rm div} \vc{u}\right]}}
.}
On one hand, due to monotonicity of $p(\rho)$ we have
\[\intTO{\phi\big(p(\rho_\delta)\rho_\delta-\Ov{p(\rho)}\rho\big)} \geq 0 .\]
On the other hand, since $\cdot\mapsto\pi_{\kappa,\delta}(\cdot)$ is non-decreasing for fixed $\delta$ and $\pi_{\kappa,\delta^1}(\cdot) \geq \pi_{\kappa,\delta^2}(\cdot)$ provided $\delta^1\leq \delta^2$, therefore
\eq{ \liminf_{\delta\rightarrow 0} \intTO{\phi\rho^*\lr{\pi_{\kappa,\delta}\lr{\dfrac{\rd}{\rs}}\rho_\delta-\overline{\pi_{\kappa}\lr{\frac{\rho}{\rho^*}}}\rho}}\\
\geq
\liminf_{\delta\rightarrow 0}\intTO{\phi\rho^* \lr{\pi_{\kappa,\delta^0}\lr{\frac{\rho_\delta}{\rho^*}}\rho_\delta
-\overline{\pi_{\kappa}\lr{\frac{\rho}{\rho^*}}}\rho}}\\
\geq
\liminf_{\delta\rightarrow 0}\intTO{\phi \rho^*\lr{\Ov{\pi_{\kappa,\delta^0}\lr{\frac{\rho}{\rho^*}}} -\overline{\pi_{\kappa}\lr{\frac{\rho}{\rho^*}}}}\rho},
}
where the last inequality is again a consequence of the monotonicity of $\pi_{\delta^0}$ for any fixed $\delta^0$. Now, one can see that the r.h.s. vanishes due to the strong convergence of $\Ov{\pi_{\delta^0}\lr{\frac{\rho}{\rho^*}}} $ to $\overline{\pi_{\varepsilon}\lr{\frac{\rho}{\rho^*}}}$ for $\delta^0\to 0$. This follows from the equi-integrability of the singular pressure \eqref{equiintegr}.
These inequalities imply that \eqref{ef} can be reduced to
\[\intTO{\overline{\rho{\rm div} \vc{u}}} \geq \intTO{\rho{\rm div} \vc{u}}.\]
Coming back to \eqref{ineg_log} and using the convexity of the function $s\mapsto s\log s$ we get
\[\overline{\rho\log \rho} = \rho\log \rho\]
which yields the strong convergence of $\rho_\delta$ in $L^p((0,T)\times\Omega)$, $p<\alpha+1$.\\
Having disposed of the problem of strong convergence of the density, we can identify the limits \eqref{identy} in the momentum equation \eqref{weak_mom_delta}.
In addition, using Mosco convergence, one can let $\delta\to 0$ also in the energy inequality \eqref{energ_in}, we have
\eq{\label{energ_delta}
\frac{ d}{dt} \intOB{\frac{1}{2} \rho |\vc{u}|^2 + \frac{1}{\gamma-1}\rho^\gamma
+ \rho\Gamma_{\kappa}\lr{\frac{\rho}{\rho^*}}}
+ \intOB{\vc{S}:\nabla\vc{u}} \leq0
}
satisfied in the sense of distributions on $(0,T)$.\\
\bigskip
\noindent{\bf Upper bound on the limit density.} To conclude this section, let us prove the uniform bound for limit density $\rho_\vep$, which is the main advantage of our approximation scheme in comparison with \cite{LiMa}.
Recall that the basic energy estimate \eqref{un_delta} implies in particular that
\[\displaystyle \sup_{t\in[0,T]} \intO{\rho_{\delta} \Gamma_{\kappa,\delta}\lr{\frac{\rho_\delta}{\rho^*}}} \leq C.\]
Thus, the upper bound for $\rho_\kappa$ follows from
\begin{align*}
\int_\Omega{\rho_\delta\Gamma_{\kappa,\delta}\left(\frac{\rho_\delta}{\rho^*}\right){\rm d} {x}} &\geq \int_\Omega{\rho_\delta\Gamma_{\kappa,\delta}\left(\frac{\rho_\delta}{\rho^*}\right)\vc{1}_{\{\rho_\delta/\rho^*\geq 1-\delta\}}{\rm d} {x}} \\
& \geq \int_\Omega{\rho_\delta\left(\int_0^{1-\delta}{\frac{s^{\alpha-2}}{(1-s)^\beta}\mathrm{d}s}\right)\vc{1}_{\{\rho_\delta/\rho^*\geq 1-\delta\}}{\rm d} {x}} \\
&=\varepsilon\int_\Omega{\rho_\delta\left(\int_\delta^{1}{\frac{(1-s)^{\alpha-2}}{s^\beta}\mathrm{d}s}\right)\vc{1}_{\{\rho_\delta/\rho^*\geq 1-\delta\}}{\rm d} {x}} \\
&= \varepsilon\int_\Omega{\rho_\delta\left(\sum_{k=0}^{\alpha-2}\begin{pmatrix}\alpha -2 \\ k \end{pmatrix}(-1)^k\int_\delta^1{s^{k-\beta}\mathrm{d}s}\right)\vc{1}_{\{\rho_\delta/\rho^*\geq 1-\delta\}}{\rm d} {x}} \\
& =\frac{\varepsilon}{\beta-1}\intO{\rho_\delta\lr{\frac{1}{\delta^{\beta-1}}-1}\vc{1}_{\{\rho_\delta/\rho^*\geq 1-\delta\}}}
\\
&\quad+ \varepsilon\int_\Omega{\rho_\delta\left(\sum_{k=1}^{\alpha-2}\begin{pmatrix}\alpha -2 \\ k \end{pmatrix}(-1)^k\int_\delta^1{s^{k-\beta}\mathrm{d}s}\right)\vc{1}_{\{\rho_\delta/\rho^*\geq 1-\delta\}}{\rm d} {x}}
\end{align*}
Similarly to \eqref{ineg_equi} we get
\[\geq \varepsilon\frac{C_1(\varepsilon,\alpha,\beta)}{\delta^{\beta-1}}\left|\left\{\frac{\rho_\delta}{\rho^*}\geq 1-\delta\right\}\right|-C_2(\varepsilon,\alpha,\beta)\]
which implies that $\left|\left\lbrace\dfrac{\rho_\delta}{\rho^*}\geq 1-\delta\right\rbrace\right| \leq C(\varepsilon)\delta^{\beta-1}$. Recall that $\beta>3$, thus finally, after letting $\delta\to 0$ we obtain $\left|\left\lbrace\dfrac{\rho_\kappa}{\rho^*}\geq 1\right\rbrace\right|=0$, which implies \eqref{un_main}.
Having obtained this uniform bound, passage to the limit $\kappa\to 0$ is just a repetition of the steps from above and thus,
\begin{equation}\label{rho_bound}
0 \leq \rho_\varepsilon \leq 1,
\end{equation}
and the first part of Theorem \ref{main} is proved .$\Box$
\section{Recovering of the two-phase system}\label{S_ep}
The purpose of this section is to perform the last limit passage, i.e. $\varepsilon\to 0$ and so to prove the second part of Theorem \ref{main}.
\bigskip
\noindent {\bf Uniform bounds.} We first summarize what kind of uniform estimates are available at this level of approximation.
Directly from \eqref{energ_delta} it follows that
\eq{
& (\rho_\vep |\vc{u}_\vep|^2)_{\{\varepsilon>0\}}\quad \text{is bounded in } L^{\infty}(0,T;L^1(\Omega)) \\
& (\rho_\vep )_{\{\varepsilon>0\}} \quad \text{is bounded in } L^{\infty}((0,T)\times \Omega) \\
& \lr{\rho_\vep \Gamma_\varepsilon\lr{\frac{\rho_\vep}{\rho^*}}}_{\{\varepsilon>0\}}\quad \text{is bounded in } L^{\infty}(0,T;L^1(\Omega)) \\
& (\vc{u}_\vep)_{\{\varepsilon>0\}}\quad \text{ is bounded in } L^2(0,T;(W^{1,2}(\Omega))^3).
}
Next, since the density sequence is uniformly bounded in $L^\infty((0,T)\times \Omega)$, we can test the momentum equation by
\[\phi(t,x) = \frac{\psi(t)}{\rho^*}\mathcal{B}\left(\rho_\vep - \Ov{\rho_\vep}\right),\qquad \psi(t)\in {\mathcal C}^\infty_0((0,T)),\quad \psi\geq0,\]
to get the uniform bound for the pressure
\eq{\label{pressure_bound_eps}
\intTO{\pi_{\vep}\lr{\frac{\re}{\rs}}} \leq C,\quad\text{and}\quad {\intTO{\rho_\vep \pi_{\vep}\lr{\frac{\re}{\rs}}}} \leq C.
}
\bigskip
\noindent{\bf Passage to the limit in the continuity and the momentum equations.} Using the approximate continuity and momentum equations we may repeat the steps leading to \eqref{cvg_dens} to get
\begin{equation}
\rho_\vep \longrightarrow \rho \quad\text{ in } \ \mathcal{C}([0,T];L_{\rm weak}^p(\Omega)) \quad \forall 1\leq p <+\infty.
\end{equation}
Moreover, the compensated compactness lemma \ref{lem_compac} allows to justify the convergence of $\rho_\vep\vc{u}_\vep$ and $\rho_\vep\vc{u}_\vep\otimes\vc{u}_\vep$ in the sense of distributions to $\rho \vc{u}$ and $\rho \vc{u}\otimes \vc{u}$ respectively, as it was done in \eqref{cvg_momentum} and \eqref{cvg_conv}. These observations are sufficient in order to pass to the limit in the approximate continuity equation in order to obtain \eqref{AC}.
Further, thanks to \eqref{pressure_bound_eps} we can extract the subsequences such that
\begin{align*}
& \pi_{\vep}\lr{\frac{\re}{\rs}} \longrightarrow {\pi} \quad \text{weakly in } \mathcal{M}_+((0,T)\times \Omega), \\
& \rho_\vep\pi_{\vep}\lr{\frac{\re}{\rs}}\longrightarrow {\pi}_1\quad \text{weakly in } \mathcal{M}_+((0,T)\times \Omega).
\end{align*}
This allows us to pass to the limit in the momentum equation \eqref{weak_mom_delta}, we obtain
\eq{\label{weak_mom}
&\intTO{\rho \vc{u}\cdot\partial_t\varphi }+ \intO{ \vc{m}_0\cdot\varphi(0)} \\
&+\intTO{\rho(\vc{u}\otimes \vc{u}):\nabla\varphi} -\intTO{\vc{S} : \nabla\varphi } \\
&+\intTO{\overline{p(\rho)} {\rm div}\varphi} +\intTO{ {\pi}{\rm div}(\rho^*\varphi)}= 0,
}
where the last product of the l.h.s. has to be understood in the sense of distribution since $\pi$ is merely a measure. In particular, this justify the regularity imposed on $\rho^*$ (that it belongs too $ \mathcal{C}^1$). \\
As in the previous section, we still need to identify the limit $\Ov{p(\rho)}$ but also to prove that the limit measure $\pi$ satisfies the constraint from \eqref{Cons}.
\bigskip
\noindent{\bf Strong convergence of the density.} The strong convergence of the sequence approximating the density will allow us to identify
\eq{\label{identy2}
\overline{p(\rho)}=p(\rho).}
In order to prove that, we need to derive a variant of effective viscous flux equality \eqref{ef}. The basic idea is the same as previously, we want to test the approximate momentum equation \eqref{pressure_bound_eps} by $\nabla \Delta^{-1}[\mathbf{1}_\Omega\rho_\vep]$, which is now bounded in $L^{\infty}(0,T;W^{1,\infty}(\Omega))$, pass to the limit and compare it with analogous expression obtained for the limit equation \eqref{weak_mom}.
Note, however, that we do not have enough regularity on the limit pressure $\pi$ to test the limit momentum equation by $\nabla \Delta^{-1}[\mathbf{1}_\Omega\rho]$.
To justify this step we regularize in time and space the weak limits $\rho$ and $\pi$ by means of standard multipliers. Indeed, taking in \eqref{weak_mom} the supremum over $\varphi=\frac{\phi}{\rho^*}$ for all $\phi \in {\cal D}((0,T)\times\Omega)$ and using the uniform estimates, we can verify that the limit pressure $\pi$ is more regular.
Indeed, considering the limit momentum equation
\[\partial_t(\rho \vc{u}) + {\rm div}\,(\rho\vc{u}\otimes \vc{u}) -{\rm div} \,({\bf S}) + \nabla p +\rho^*\nabla \pi = 0 \]
satisfied in the sense of distributions, we can divide this equation by $\rho^*$ and apply the operator $\nabla\Delta^{-1}$ to obtain
\begin{equation}\label{regu_pi}
\pi \in W^{-1,\infty}(0,T;W^{1,2}(\Omega)) + L^p(0,T;L^q(\Omega))\quad p,q>1.\end{equation}
On the other hand, from the continuity equation, we easily get
\[\rho \in \mathcal{C}([0,T];L^p)\cap \mathcal{C}^1([0,T];W^{-1,2}(\Omega))\quad 1\leq p< +\infty.\]
Therefore for $\rho_n = \rho\ast\omega_n$, $\pi_n=\pi\ast \omega_n$, where $\omega_n$ is a mollifying sequence, we have the following convergences
\eq{\label{convol}
&\rho_n\longrightarrow \rho \quad \text{in } \mathcal{C}([0,T]; L^p_{\rm weak}(\Omega)) \cap \mathcal{C}^1([0,T],W^{-1,2}(\Omega)) \\
& \pi_n\longrightarrow \pi \quad \text{in } W^{-1,\infty}(0,T;W^{1,2}(\Omega)) + L^p(0,T; L^q(\Omega)).
}
Then the product $\rho\pi$ can be expressed as
\eq{\label{rpi}
\rho\pi=\rho_n\pi_n+(\rho-\rho_n)\pi_n+\rho(\pi-\pi_n)
}
and we can apply Lemma \ref{lem_compac} to pass to the limit in the r.h.s.
With this justification, the same computations as in the previous section, give rise to the effective flux equality
\eqh{
&\intTO{\psi\phi\big(\rho^*\pi_1 +\overline{\rho p(\rho)} - (\lambda+2\mu)\overline{\rho {\rm div} \vc{u}}\big)} \\
&- \intTO{\phi\big(\rho^*\rho\pi + \rho\overline{p(\rho)}- (\lambda+2\mu)\rho {\rm div} \vc{u}\big)}=0,
}
where $\phi \in {\cal D}(\Omega)$, $\psi\in {\cal D}$((0,T)). Therefore
\eq{ \label{eff_flux}
&(2\mu+\lambda)\intTO{\psi\phi(\overline{\rho{\rm div} \vc{u}}-\rho {\rm div} \vc{u})}\\
& = \intTO{\psi\phi\big(\overline{p(\rho)\rho}-\overline{p(\rho)}\rho + \rho^*\pi_1-\rho^*\rho\pi\big)} \\
& \geq \intTO{\psi\phi\rho^*\big(\pi_1-\rho\pi\big) },
}
where the last inequality is a consequence of monotonicity of $p(\rho)$.
In order to treat the r.h.s. of \eqref{eff_flux} we show that
\eq{\pi_1 = \rho^*\pi\label{c0}}
a.e. in $(0,T)\times\Omega$. To prove this fact we first observe that
\eq{\label{pi_rho_pi}
\rho_\vep \pi_\varepsilon\lr{\dfrac{\rho_\vep}{\rho^*}} = & \varepsilon \rho_\vep \dfrac{\left(\dfrac{\rho_\vep}{\rho^*}\right)^\alpha}{\left(1-\dfrac{\rho_\vep}{\rho^*}\right)^\beta} \\
= & \varepsilon \rho^* \dfrac{\left(\dfrac{\rho_\vep}{\rho^*}\right)^\alpha}{\left(1-\dfrac{\rho_\vep}{\rho^*}\right)^\beta} + \varepsilon (\rho_\vep-\rho^*) \dfrac{\left(\dfrac{\rho_\vep}{\rho^*}\right)^\alpha}{\left(1-\dfrac{\rho_\vep}{\rho^*}\right)^\beta} \\
= & \rho^*\pi_\varepsilon\lr{\dfrac{\rho_\vep}{\rho^*}} -\varepsilon\frac{\rho^*\lr{\dfrac{\rho_\vep}{\rho^*}}^\alpha}{\left(1-{\dfrac{\rho_\vep}{\rho^*}}\right)^{\beta-1}}.
}
Then, letting $\varepsilon\rightarrow 0$ and using the $L^1((0,T)\times\Omega)$ bound on $\pi_\varepsilon$, we show that
\[\varepsilon\frac{\rho^*\lr{\dfrac{\rho_\vep}{\rho^*}}^\alpha}{\left(1-{\dfrac{\rho_\vep}{\rho^*}}\right)^{\beta-1}} \longrightarrow 0 \quad \text{strongly in } L^{\beta/(\beta-1)}((0,T)\times \Omega).\]
Therefore, passing to the limit $\varepsilon\to0$ in \eqref{pi_rho_pi}, we arrive at \eqref{c0}.
Inserting this to \eqref{eff_flux} we may deduce that
\eqh{
(2\mu+\lambda)\intTO{\psi\phi(\overline{\rho{\rm div} \vc{u}}-\rho {\rm div} \vc{u})} \geq \intTO{\psi\phi\rho^*\big(\rho^*-\rho\big)\pi} \geq 0,
}
where the last inequality follows from the fact that $\rho\leq \rho^*$ a.e. in $(0,T)\times\Omega$.
Note that both pairs $(\rho,\vc{u})$ and $(\rho_\vep,\vc{u}_\vep)$ satisfy the renormalized continuity equation \eqref{renormal}, thus the above equality implies that taking $b(\rho_\vep)=\rho_\vep\log\rho_\vep$ we pass to the weak limit and compare with the equation on $\rho\log\rho$. The same arguments as in the previous section show that $\rho_\vep$ converges to $\rho$ strongly in $L^p((0,T)\times \Omega)$, $1\leq p<+\infty$.
\bigskip
\noindent {\bf Recovery of the congestion constraint.} The strong convergence of the density enables also to identify the limit in the singular pressure, we have
\eq{\pi_1=\rho \pi,\label{c1}}
where the meaning to the product on the r.h.s. is given as in \eqref{rpi}.
Finally, comparing \eqref{c0} with \eqref{c1} we get that
$$\rho^*\pi=\pi_1=\rho\pi,$$
which gives the congestion constraint \eqref{Cons}.
\bigskip
\noindent {\bf The divergence free condition.} To justify that the limit triple $(\rho,\vc{u},\pi)$ is a solution to system \eqref{CM0} in the sense of Definition \eqref{Def1} one has to check that the divergence free condition ${\rm div}\,(\rho^*\vc{u})=0$ is satisfied a.e. in $\{\rho=\rho^*\}$.
We will show that it follows from a certain compatibility between \eqref{CM0}$_{1}$ and conditions \eqref{CM0}$_{2,4}$ for the limit system. We have the following generalization of Lemma 2.1 from \cite{LiMa} to the heterogeneous case.
\begin{Lemma}\label{LMP}
Let $\vc{u}\in L^2(0,T;(W^{1,2}_0(\Omega))^3)$ and $\rho\in L^2((0,T)\times\Omega)$ such that
$$\partial_{t}\rho+{\rm div}(\rho \vc{u})=0\quad in \ (0,T)\times\Omega,\quad \rho(0)=\rho_0$$
then the following two assertions are equivalent
\begin{itemize}
\item[(i)] ${\rm div}(\rho^* \vc{u}) = 0$ a.e on $\{\rho \geq \rho^*\}$ and $0 \leq \rho_0 \leq \rho^*$,
\item[(ii)] $0\leq\rho(x,t)\leq\rho^*(x)$.
\end{itemize}
\end{Lemma}
{\noindent\it Proof.~} We first prove implication $(ii)\to (i)$. Denote $\displaystyle R = \frac{\rho}{\rho^*}$, then $R$ satisfies the equation
\begin{equation}\label{eq_R}
\partial_t R + {\rm div}(R \vc{u}) + R\vc{u}\cdot\nabla\log \rho^* = 0.
\end{equation}
Let $\beta \in \mathcal{C}^1([0,\infty))$, multiplying the previous equation by $\beta'$ we get
\begin{equation} \label{renorm_R}
\partial_t \beta(R) + {\rm div}(\beta(R) \vc{u}) + (\beta'(R)R -\beta(R)) {\rm div} \vc{u} + \beta'(R)R\vc{u}\cdot\nabla\log\rho^*=0.
\end{equation}
Due to assumptions $0 \leq R \leq 1$, using in \eqref{renorm_R} $\beta(R) = R^k$ for any integer $k$, we obtain
\[\partial_t R^k + {\rm div}(R^k\vc{u}) + (k-1)R^k{\rm div} \vc{u} + kR^k\vc{u}\cdot\nabla\log\rho^* = 0.\]
Since $R^k \in L^\infty((0,T)\times\Omega)$, thus $\partial_tR^k\in W^{-1,\infty}((0,T)\times \Omega)$. On the other hand, $|R^k\vc{u}|\leq |R\vc{u}|\in L^\infty(0,T;L^2(\Omega))$, so ${\rm div}(R^k\vc{u})$ is bounded in $L^\infty(0,T;W^{-1,2}(\Omega))$ and $R^k{\rm div} \vc{u}$ is bounded in $L^2((0,T)\times \Omega)$. As a conclusion, we see that\\
$kR^k({\rm div} \vc{u} + \vc{u}\cdot\nabla\log\rho^*)$ is a bounded distribution. Hence, letting $k$ go to infinity we justify that
\[R^k({\rm div} \vc{u} + \vc{u}\cdot\nabla\log\rho^*) \rightharpoonup 0 \quad \text{in } \mathcal{D}'((0,T)\times\Omega).\]
However, we also know that
\[R^k({\rm div} \vc{u} + \vc{u}\cdot\nabla\log\rho^*) \to \mathbf{1}_{\{R=1\}}({\rm div} \vc{u} + \vc{u}\cdot\nabla\log\rho^*) \quad \text{a.e. in} \ (0,T)\times\Omega \]
and since $|R^k({\rm div} \vc{u} + \vc{u}\cdot\nabla\log\rho^*)|\leq |{\rm div} \vc{u} + \vc{u}\cdot\nabla\log\rho^*|$, where ${\rm div} \vc{u},\ \vc{u}\cdot \nabla\log\rho^* \in L^2((0,T)\times \Omega)$, we get
\[ {\rm div} \vc{u} + \vc{u}\cdot\nabla\log\rho^* = 0 \quad \text{a.e. on} \quad \{R= 1\}.\]
\medskip
To prove the implication $(i)\to (ii)$ we set $d(\rho)=\rho-\rho^*$ which satisfies the equation
\eq{\partial_td + {\rm div}(d \vc{u}) + {\rm div}(\rho^* \vc{u}) = 0.
\label{eq_d}}
Let us next consider $b(d)$ the positive part of $d(\rho)$, regularized around $0$ in the following way
\[ b_\eta(s) = \begin{cases} b(s) \quad & \text{if } |s|\geq \eta, \\
\frac{1}{4\eta}(s+\eta)^2 \quad & \text{if } |s|< \eta. \end{cases}\]
Multiplying \eqref{eq_d} by $b_\eta'(d)$, we obtain the equation
\eq{\partial_tb_\eta(d) + {\rm div}(b_\eta(d) \vc{u}) + (b_\eta'(d)d-b_\eta(d)){\rm div} \vc{u} + b_\eta'(d){\rm div}(\rho^* \vc{u}) = 0.\label{eq_ren_d}}
Note that
\[ b_\eta(d)'d-b_\eta(d) = \begin{cases} 0 \quad & \text{if } |\rho-\rho^*|\geq \eta, \\
\frac{1}{4\eta}(d^2-\eta^2) \quad & \text{if } |\rho-\rho^*|< \eta, \end{cases}\]
and that for $\eta\to0$, $b_\eta'(d)d-b_\eta(d)$, $b'_\eta(d)$ converge strongly in $L^2((0,T)\times\Omega)$ to 0 and $b'_+(d)$ respectively, where $b_+'$ is defined by
\[ b_+'(d) = \begin{cases} & 0 \quad \text{if}\quad \rho< \rho^* \\
& 1 \quad \text{if}\quad \rho\geq \rho^* \end{cases}.\]
Thus letting $\eta \rightarrow 0$ in \eqref{eq_ren_d} we get
\begin{equation}\label{eq_renorm_b}
\partial_tb(d) + {\rm div}(b(d) \vc{u}) + b_+'(d){\rm div}(\rho^* \vc{u}) = 0.
\end{equation}
By assumption, ${\rm div}(\rho^* \vc{u})=0$ on $\{\rho \geq \rho^*\}$ and $b(d)(t=0)= 0$. Therefore, integrating \eqref{eq_renorm_b} over $(0,t)\times\Omega$, we obtain $\intO{b(d)(t)} = \intO{b(d)(0)}= 0$ for $t\geq 0$. Since $b(d)$ is a nonnegative function, we conclude that $d\leq 0$, which means that $\rho \leq \rho^*$.
$\Box$
\bigskip
The above result together with estimate \eqref{rho_bound} guarantee that the second part of Theorem \ref{main} is proven. $\Box$
\section{Appendix}\label{Appendix}
{\bf The Bogovskii operator.} The Bogovskii operator is defined in the following lemma.
\begin{Lemma}[\cite{NS}, lemma 3.17]\label{lem_bog}
Let $\Omega$ be a bounded Lipschitz domain in $\mathbb{R}^3$. Then there exits a linear operator $\mathcal{B}_\Omega=\big(\mathcal{B}_\Omega^1,\mathcal{B}_\Omega^2,\mathcal{B}_\Omega^3)$ with the following properties :
\[\mathcal{B}_\Omega\,: \overline{L^p}(\Omega) \rightarrow (W_0^{1,p}(\Omega))^3, \quad 1<p<\infty;\]
\[{\rm div}\,(\mathcal{B}_\Omega(f))=f \quad \text{a.e. in } \Omega,\,f\in \overline{L^p}(\Omega);\]
\[\|\nabla \mathcal{B}_\Omega(f)\|_{L^p(\Omega)} \leq c(p,\Omega)\|f\|_{L^p(\Omega)}, \quad 1<p<\infty\]
\[\text{If } f={\rm div}\,(g),\text{ with } g\in L^p(\Omega),\, {\rm div}\,(g)\in L^q(\Omega), \, 1<q<\infty, \text{ then }\]
\[\|\mathcal{B}_\Omega(f)\|_{L^q(\Omega)}\leq c(q,\Omega)\|g\|_{L^q(\Omega)}\]
where $\overline{L^p(\Omega)}=\{f\in L^p(\Omega):\int_\Omega{f}(y){\rm d}y=0\}$.
\end{Lemma}
\label{prop_bog}\begin{prop}[\cite{FN}, Theorem 10.11]
Let $\Omega$ be a bounded Lipschitz domain in $\mathbb{R}^3$, $\mathcal{B}_\Omega$ can be uniquely extended as a bounded linear operator
\[\mathcal{B}_\Omega \,: [\overline{W^{1,p'}}(\Omega)]^* = \{f\in [W^{1,p'}(\Omega)]^*; \langle f, 1 \rangle = 0 \} \rightarrow (L^p(\Omega))^3\]
in such way that
\[-\int_\Omega{\mathcal{B}_\Omega(f)\cdot \nabla v} = \langle f,v\rangle_{\{[\overline{W^{1,p'}}(\Omega)]^*,W^{1,p'}\}}, \quad \forall v \in W^{1,p'}(\Omega);\]
\[\|\mathcal{B}_\Omega(f)\|_{L^p(\Omega)} \leq c\|f\|_{[\overline{W^{1,p'}}(\Omega)]^*}.\]
\end{prop}
\bigskip
\noindent{\bf The Riesz transform.} The inverse divergence operator
$\mathcal{A}=\nabla\Delta^{-1}$
and the double Riesz transform
$\mathcal{R}=\nabla\otimes\nabla\Delta^{-1}$
are defined as
\begin{equation}\label{defA}
\mathcal{A}_j[v]=\left(\nabla\Delta^{-1}\right)_j v=-\mathcal{F}^{-1}\left(\frac{i\xi_j}{|\xi|^{2}}\mathcal{F}(v)\right),
\end{equation}
\begin{equation}\label{defR}
\mathcal{R}_{i,j}[v]=\partial_i\mathcal{A}_j[v]
=\left(\nabla\otimes\nabla\Delta^{-1}\right)_{i,j} v=\mathcal{F}^{-1}\left(\frac{\xi_i\xi_j}{|\xi|^{2}}\mathcal{F}(v)\right).
\end{equation}
Here, the inverse Laplacian is identified through the Fourier transform $\mathcal{F}$ and the inverse Fourier transform $\mathcal{F}^{-1}$ as
\begin{equation*}
(-\Delta)^{-1}(v)=\mathcal{F}^{-1}\left(\frac{1}{|\xi|^{2}}\mathcal{F}(v)\right).
\end{equation*}
In what follows we recall some of basic properties of these operators.
\begin{Lemma}\label{LA1}
The operator ${\cal R}$ is a continuous linear operator from
$L^p(\mathbb{R}^3)$ into $L^p(\mathbb{R}^3)$
for any $1<p<\infty$. In particular, the following estimate holds true:
$$\|{\cal R}[v]\|_{L^p(\mathbb{R}^3)}\leq c(p)\|v\|_{L^p(\mathbb{R}^3)}\quad for\ all\ v\in L^p(\mathbb{R}^3).$$
The operator ${\cal A}$ is a continuous linear operator from
$L^1(\mathbb{R}^3)\cap L^2(\mathbb{R}^3)$ into $L^2(\mathbb{R}^3)+L^{\infty}(\mathbb{R}^3)$,
and from $L^p(\mathbb{R}^3)$ into $L^{\frac{3p}{3-p}}(\mathbb{R}^3)$ for any $1<p<3$.
Moreover,
$$\|\nabla {\cal A}[v]\|_{L^p(\mathbb{R}^3)}\leq C(p)\|v\|_{L^p(\mathbb{R}^3)},\quad 1<p<\infty.$$
\end{Lemma}
The proof of this lemma can be found e.g. in \cite{FN}, Section 10.16. In what follows we present two important properties of commutators involving Riesz operator.
The first result is a straightforward consequence of the {\it Div-Curl} lemma, its proof can be found in \cite{EF2001}, Lemma 5.1.
\begin{Lemma}\label{LA2}
Let
\begin{eqnarray*}
\vc{V}_{\varepsilon}\rightharpoonup \vc{V}\quad {\rm weakly\ in}\ L^{p}(\mathbb{R}^{3}),\quad
r_{\varepsilon}\rightharpoonup r\quad {\rm weakly\ in}\ L^{q}(\mathbb{R}^{3}),
\end{eqnarray*}
where
\begin{equation*}
\frac{1}{p}+\frac{1}{q}=\frac{1}{s}<1.
\end{equation*}
Then
\begin{equation*}
\vc{V}_{\varepsilon}\mathcal{R}(r_{\varepsilon})-r_{\varepsilon}\mathcal{R}(\vc{V}_{\varepsilon})
\rightharpoonup
\vc{V}\mathcal{R}(r)-r\mathcal{R}(\vc{V})\quad weakly\ in\ L^{s}(\mathbb{R}^{3}).
\end{equation*}
\end{Lemma}
The next lemma can be deduced from the general results of {Baj{\v{s}}anski} and {Coifman} \cite{BC}, and {Coifman} and {Meyer} \cite{CM}.
\begin{Lemma}\label{LA3}
Let $w\in W^{1,r}(\mathbb{R}^3)$ and $\vc{V}\in L^p(\mathbb{R}^3)$ be given, where
$1<r<3$, $1<p<\infty$, $\frac{1}{r}+\frac{1}{p}-\frac{1}{3}<\frac{1}{s}<1.$
Then for all such $s$ we have
$$\|{\cal R}[w\vc{V}]-w{\cal R}[\vc{V}]\|_{W^{\alpha,s}(\mathbb{R}^3)}
\leq c(s,p,r)\|w\|_{W^{1,r}(\mathbb{R}^3)}\|\vc{V}\|_{L^p(\mathbb{R}^3)},$$
where $\alpha$ is given by
$\frac{\alpha}{3}=\frac{1}{s}+\frac{1}{3}-\frac{1}{p}-\frac{1}{r}.$
\end{Lemma}
Here, $W^{\alpha,s}(\mathbb{R}^3)$ for $\alpha\in(0,\infty) \setminus \mathbb{N}$ denotes the Sobolev-Slobodeckii space (see e.g. \cite{T78}).
The proof can be found in \cite{FN}, Section 10.17.
\bigskip
\noindent{\bf The renormalized continuity equation.} Below we recall the definition of the renormalized solution to the continuity equation.
\begin{df}\label{df2}
The pair $(\rho,\vc{u})$ is called a renormalized solution to the continuity equation, if equation \eqref{CM0}$_1$ holds in $\mathcal{D}'((0,T)\times \mathbb{R}^3)$ provided that $\rho,\vc{u}$ is prolonged by $0$ outside $\Omega$ and equation
\begin{equation}\label{renormal}
\partial_{t} b(\rho)+{\rm div} \lr{b(\rho)\vc{u}}+\lr{\rho b'(\rho)-b(\rho)}{\rm div}\vc{u}=0,
\end{equation}
holds in $\mathcal{D}'((0,T)\times \mathbb{R}^3)$ provided that $\rho,\vc{u}$ is prolonged by $0$ outside $\Omega$, for any function $b\in {\cal C}[0,\infty)\cap {\cal C}^{1}(0,\infty)$, such that $sb'(s)\in {\cal C}[0,\infty)$.
\end{df}
The following result is a consequence of technique introduced and developed by {DiPerna} and {Lions} \cite{DPL}.
\begin{Lemma}\label{LX}
Let $\rho \in L^{p}((0,T)\times\Omega)$, $p\geq 2$, $\rho\geq 0$, a.e. in $\Omega$ and $\vc{u}\in L^2(0,T; W^{1,2}(\Omega))$ satisfy continuity equation
$$\partial_{t}\rho+{\rm div}(\rho\vc{u})=0$$
in $\mathcal{D}'((0,T)\times \mathbb{R}^3)$ (prolonged by $0$ outside $\Omega$). Then the renormalized continuity equation \eqref{renormal} holds in $\mathcal{D}'((0,T)\times \mathbb{R}^3)$ for any function $b$ satisfying
\begin{equation}\label{hyp_renorm}
b\in \mathcal{C}[0,+\infty)\cap \mathcal{C}^1(0,+\infty), \quad |b'(t)|\leq ct^{-\lambda_0} \quad t\in (0,+\infty), \quad \lambda_0<1
\end{equation}
and growth conditions at infinity
\begin{equation}\label{hyp_renorm_2}
|b'(t)|\leq ct^{\lambda_1}, \quad t\geq 1 \quad \text{where } c>0, \quad -1<\lambda_1\leq \frac{p}{2}-1.
\end{equation}
\end{Lemma}
A general reference here is \cite{FN}, Section 10.18, see also \cite{NS}.\\
\subsection*{Acknowledgment}
The authors wish to thank Didier Bresch for suggesting the problem, stimulating
conversations and help during the preparation of the paper.
The first author acknowledges support from the ANR-13-BS01-0003-01 project DYFICOLTI and
support from the "Projet Exploratoire 2014 de la cellule Energie du CNRS" Dynafilm.
The second author was supported by MN grant IdPlus2011/000661.
|
\section{Introduction}
\label{sec:intro}
A series of experiments with ultra cold atoms carried out in the last decade
\cite{greiner02_fast_tunnability,kinoshita06_non_thermalization,hofferberth07_nonequilibrium_dynamics_1D_bosons,trotzky11_relaxation_BH_10101010,chenau12_light_cone_experimental_bosons,gring12_pre-thermalization_isolated_bose_gas,langen13_local_equilibrium} exhibited absence of dissipation in the many-particle system and therefore essentially unitary time evolution on long time scales. This motivated a great deal of activity involving the study of the dynamics of interacting quantum systems that are driven out of equilibrium by preparing them in an initial state that is not in the eigenbasis of the Hamiltonian. Several interesting problems arise in these systems such as the thermalization mechanisms in integrable and non-integrable models~(see Refs.~\onlinecite{cazalilla10_quenches,dziarmaga10_quenches,polkovnikov11_nonequilibrium_dynamics} and references therein) and more generally the emergence of thermodynamics in isolated systems.
Much of the theoretical effort has been devised to investigate exactly solvable models and integrable systems, which are special since the large number of integrals of motion that constrain the nonequilibrium dynamics are believed to preclude the relaxation to thermal equilibrium. Instead, in many cases the long-times steady state is captured by a statistical description based on a generalized Gibbs ensemble (GGE)~\cite{rigol07_generalized_gibbs_hcbosons} which results from the maximization of the entropy subjected to the constraints imposed by the conserved quantities. In such a description a different temperature is associated with each conserved quantity.
Interestingly, it was shown in Refs.~\onlinecite{rigol06_hcbosons_dynamics,rigol11_quench_initial_state_dependence}
that certain kinds of initial states can lead to asymptotic values of the observables whose GGE description is essentially indistinguishable from the one computed with a standard thermal Gibbs ensemble. This effect turns out to be generic for integrable models that can be mapped onto quadratic, bosonic or fermionic models and initial states for which two sets of modes are strongly entangled~\cite{chung12_thermalization_entanglement}. However, the GGE cannot reproduce the behavior of all observables~\cite{iucci09_quench_LL}, and in particular it fails to capture energy fluctuations. Therefore, the effective temperature that emerges from the standard Gibbs distribution description characterizes the asymptotic thermal correlations and constitute a measure of the entanglement between the eigenmodes in the initial state, but does not have the usual thermodynamic meaning.
One important relation in equilibrium statistical mechanics both quantum and classical is the Fluctuation-Dissipation theorem (FDT), that relates linear response and correlation functions in a model and observable-independent fashion. Even though the FDT is strictly valid for systems in thermodynamic equilibrium, in many out-of-equilibrium situations, the FDT turns out to be more relevant for the analysis of thermalization issues than the functional decay of observables~\cite{cugliandolo11_effective_temperature_slow_dynamics}. It was shown to hold out of equilibrium after relaxation, in both nonintegrable~\cite{essler12_dynamical_gge,khatami13_FDT_noneq_nonintegrable} and integrable~\cite{essler12_dynamical_gge} systems. However, in the latter case only a basic form of it holds, implying that the way in which deviations from equilibrium states originated in external perturbations and random fluctuations dissipate in time are related, but a detailed balancing relation between the probabilities of energy absorption and release involving only the temperature of the system breaks down. Still, it is possible to define an effective temperature from the FDT~\cite{cugliandolo11_effective_temperature} in the context of quantum quenches as was done for example in integrable models such as the Luttinger model~\cite{mitra11_quench_mode_coupling,mitra12_quench_thermalization_dissipation} and the transverse field Ising chain~\cite{foini11_quench_FDT,foini12_quench_FDT_long}. The effective temperatures defined in this way depend on the momentum and frequency being considered and more important, change according to the observable under study.
In this work we analyze how these ideas apply in the context of a quantum quench for which two sets of modes are strongly entangled in the initial state and that as a consequence exhibits signs of thermalization in the decay of their correlations. We compute dynamic correlation functions of local and non-local operators in a model that is describable in terms of free fermions, from which we extract effective temperatures by forcing the FDT. We show that all the effective temperatures obtained for local operators have a well defined limit (at least in a certain range of frequencies) when the initial entanglement is strong, that is given by the effective temperature of the system after relaxation. On the other hand, effective temperatures extracted from correlatons of non-local operators exhibit a similar behavior, but its frequency dependence at large values of the initial entanglement show small deviations from that limit.
The rest of this article is organized as follows: In section \ref{sec:model} we present the model (a 1D hard-core boson in presence of a superlattice potential) and the known results in the generalized Gibbs ensemble. In section \ref{sec:corr} we study the dynamic two-time correlation functions of Fermi, density and non-local operators. In section \ref{sec:fdr} we introduce the concept of fluctuation-dissipation relations (FDRs) and compute effective temperatures for the operators analyzed in the previous section. In section \ref{sec:conclusions} we present our conclusions and discuss some implications of our work.
\section{The model}
\label{sec:model}
Let us consider a model that describes a system of hard-core bosons in one dimension that initially move in the presence of a superlattice potential. After performing a Jordan-Wigner transformation, this model maps onto the following Hamiltonian
\begin{equation}
H_0=-\sum_j^Lf^\dagger_j f_{j+1}+\mathrm{h.c.}+\Delta\sum_j^L(-1)^j f^\dagger_j f_j, \label{eq:H0}
\end{equation}
written in terms of noninteracting spinless fermions creation $f_j$ and destruction $f_j^\dagger$ operators at site $j$ ($j=1,\ldots,L$, for a lattice of $L$ sites). Periodic boundary conditions (b.c.) in the bosonic model translate into either periodic or antiperiodic b.c. in the corresponding fermionic model depending on whether the number of bosons (fermions) in the system $N$ is odd or even, while open b.c map into open boundary conditions. The system is driven out of equilibrium by preparing it in an initial state in contact with a thermal reservoir at a temperature $T$ , i.e., it is described by a density matrix $\rho_0 = Z^{-1}e^{H_0/T}$ (such that $\Tr \rho_0 = 1$). For $t > 0$, the superlattice potential is switched off and the system evolves unitarily with a Hamiltonian $H$ obtained from $H_0$ by setting $\Delta=0$.
Let us first recall the results of Ref.~\onlinecite{chung12_thermalization_entanglement} and show that correlation functions acquire a thermal form for long times. After Fourier transforming, $H_0$ and $H$ become
\begin{equation}
H_0=H+\Delta\sum_k \left(f^\dagger_{k+\pi}f_k+f^\dagger_k f_{k+\pi}\right)
\end{equation}
and
\begin{equation}
H=\sum_k \omega_k \left(f^\dagger_k f_k-f^\dagger_{k+\pi}f_{k+\pi}\right),
\end{equation}
where $\omega_k=-2\cos k$ and $-\pi/2\leqslant k\leqslant\pi/2$. The existence of the coupling $\Delta$ in $H_0$ implies that in the initial state there are correlations (i.e. bi-partite entanglement) between the eigenmodes at $k$ and $k+\pi$, i.e. $\langle f^\dagger_{k+\pi}f_k\rangle\neq 0$. A Bogoliubov rotation finally renders $H_0$ diagonal with dispersion $E_k=\sqrt{\omega_k^2+\Delta^2}$.
Dephasing makes static correlations at long times to be described by a GGE density matrix that is obtained using the maximum entropy principle taking into account that the system dynamics is constrained by the existence of the set of integrals of motion given by $I_k = f^\dagger_kf_k$ (and $I_{k+\pi} = f^\dagger_{k+\pi}f_{k+\pi}$ (with $k$ restricted to the first Brioulin zone). The GGE density matrix thus obtained reads:
\begin{equation}
\rho_\mathrm{GGE}=\frac{1}{Z_\mathrm{GGE}}\exp\left\{-\sum_k\lambda_k \left( f^\dagger_k f_k-f^\dagger_{k+\pi}f_{k+\pi}\right)\right\},\label{eq:rho_GGE}
\end{equation}
where, at $T=0$ for simplicity,
\begin{equation}
\lambda_k=\log\frac{E_k+\omega_k}{E_k-\omega_k}.
\end{equation}
For $\Delta\gg\omega_k$, $E_k$ can be approximated by $\Delta$ and therefore $\lambda_k=2\omega_k/\Delta$. Thus, the GGE density matrix, equation (\ref{eq:rho_GGE}), reduces to a standard Gibbs ensemble with temperature $T^\mathrm{G}_\mathrm{eff}=\Delta/2$ and the system exhibits thermal correlations.
\section{Dynamic correlations}
\label{sec:corr}
In this section we present our results for the dynamic correlations of several quantities relevant for our model. We study (anti)symmetrized two-time correlations of two operators $A$ and $B$ in the Heisenberg representation, $A_H(t) = e^{iHt} A e^{-iHt}$,
\begin{equation}
C^{AB}_\pm(t,t_0 )= \langle [A (t+t_0 ),B (t_0 )]_\pm \rangle, \label{eq:corrs}
\end{equation}
where $[X,Y ]_\pm = (XY \pm Y X)/2$ and $\langle \cdots \rangle$ represents the trace over the initial state $\rho_0$. Without loss of generality we consider operators with zero mean value, i.e. $O(t) = O(t)-\langle O(t) \rangle$. We focus on the (anti)symmetric correlator $C_+(C_-)$ and the retarded (or linear response) function, which can be constructed by using $C_\pm$
\begin{equation}
R^{AB}_\pm(t,t_0) = 2i \theta (t) C_\pm^{AB}(t,t_0). \label{eq:resp}
\end{equation}
$R_{\pm}^{AB}$ vanishes for $t<0$ respecting causality. In thermal equilibrium it is related to the correlation function $C_{\pm}^{AB}$ by means of the fluctuation-dissipation theorem (FDT) explained in section \ref{sec:fdr}. While the usual (bosonic) FDT involves $R_-$ and $C_+$, a fermionic version can be constructed by using $R_+$ and $C_-$. We examine these functions in time domain in section \ref{sec:corrtime}, and in the frequency domain in section \ref{sec:corrfrec}. The latter is in turn used to compute the effective temperature for each pair of operators.
\subsection{Time dependence}
\label{sec:corrtime}
Before starting with the specific two-time correlators calculation, we remark some aspects of the procedure followed and state general results. We are concerned with the computation of the two-time correlation functions
\begin{equation}
C^A_{\pm,(n,m)}(t,t_0 )= \langle [A_n (t+t_0 ),A_m (t_0 )]_\pm \rangle, \label{eq:corrsnm}
\end{equation}
where the subindices $n,m$ represent the position in the lattice and $A_n$ are generic operators. The mean value $\left\langle \cdots\right\rangle$ is taken over the ground state of the system before the quantum quench, i.e. the ground state of $H_0$: $\rho_0= \left|\psi_0\right\rangle \left\langle\psi_0\right|$. We work in the thermodynamic limit $L \rightarrow \infty$ which we impose by taking the analytic limit or considering a system of $L=1000$ lattice sites in the case of numerical results. In the limit $t_0\rightarrow\infty$, correlation functions reach a stationary regime, in which, as in equilibrium, they only depend on the time difference $t$: $C_{\pm}^A(t,t_0\rightarrow\infty) =C_{\pm}^A(t)$. This regime is relevant for extracting effective temperatures and is imposed analytically, by using the Riemann-Lebesgue lemma, or numerically, by taking $ t_0=100 $. Within the thermodynamic limit and the stationary regime the linear response function $R_\pm$ and correlator $C_\pm$ of all the operators studied in this paper, show an independence on specific site $n$ and $m$ for periodic boundary conditions; they only depend on the site difference $l = n-m $, $C_{\pm,(n,m)}^A(t,t_0\rightarrow\infty) =C_{\pm,l}^A(t)$. In the case of open b.c, this rule does not apply, but is nearly fulfilled by taking $n$ and $m$ near the center of the lattice.
We shall study the time dependence of $C_\pm$ and $R_\pm$ for several operators in the limits mentioned above, analyzing their dependence with site difference $l$, the initial superlattice potential strength $\Delta$ and initial temperature $T$.
\subsubsection{Local Operators}
Let us start by studying the quasiparticle Fermi operator $f_n$ correlation functions. Following the definition \eqref{eq:corrs}, we shall consider
\begin{equation}
C_{\pm,(n,m)}^f(t,t_0)= \langle \left[ f_n(t+t_0),f^\dagger_m(t_0) \right]_\pm \rangle\label{eq:C+c},
\end{equation}
where we shall employ $C_+$ to build the linear response (retarded) function $R_+$. As we mentioned before, in the thermodynamic ($L\rightarrow \infty$) and stationary ($t_0\rightarrow \infty$) limits, these functions have only dependence on $t_0$ and the lattice site difference $l=n-m$. In these regimes, the linear response function $R_+$ results
\begin{equation}
R^f_{+,l}(t) = \begin{cases}
-t\theta(t)\phantom{}_{1}\tilde{F}_{2}\left(1;\frac{3-l}{2},\frac{3+l}{2};- t^{2}\right) & \text{$l$ odd}\\
\phantom{-}i \theta(t)\phantom{}_{1}\tilde{F}_{2}\left(1;\frac{2-l}{2},\frac{2+l}{2};- t^{2}\right) & \text{$l$ even}\\
\end{cases}
\label{eq:Cc}
\end{equation}
where $ \phantom{}_{1}\tilde{F}_{2}$ represents a generalized (regularized) hypergeometric function. Interestingly, $R^f_{l}$ is independent of $\Delta$ which may lead to the conclusion that in the stationary regime the initial state correlations have been lost. Nevertheless, some information remains as $R^f$ is different for even and odd site difference, which is a consequence of the different translational symmetries of $H$ and $H_0$. On the other hand, the antisymmetric correlator $C^f_-$ in the stationary regime,
\begin{equation}
C_{-,l}^f(t) = \intop_{-\pi/2}^{\pi/2} \frac{dk}{4 \pi} \cos (k l) \frac{\omega_k \left(e^{-i \omega_k t}-e^{i\pi l}e^{i \omega_k t}\right) }{\sqrt{\omega_k^2+ \Delta^2}}
\label{eq:Rc}
\end{equation}
does depend on the supperlatice potential. In Fig. \ref{fig:Rc} we plot the real and imaginary parts of the response $R_+$ and antisymmetric correlator $C_-$ for different values of site difference $l$.
\begin{figure}[ht]
\centering
\subfloat{\label{fig:ReCf}} \subfloat{\label{fig:ReRf}} \subfloat{\label{fig:ImCf}} \subfloat{\label{fig:ImRf}}
\includegraphics{fermi.pdf}
\caption{Correlation functions for the Fermi operators varying the site difference $l$ with $\Delta =1$. In the panels \protect\subref{fig:ReCf} and \protect\subref{fig:ImCf} we show the real and imaginary part of $C^f$, respectively, while in the right panels \protect\subref{fig:ReRf} and \protect\subref{fig:ImRf} the same information is displayed for the response function $R^f$. }
\label{fig:Rc}
\end{figure}
We observe that both functions are real or pure imaginary for odd $l$ or even $l$, respectively. Also, we notice the presence of the so-called light-cone effect~\cite{calabrese06_quench_CFT}, in which the functions are expected to be constant up to a time $t_e = l/v_e$ ($l/2$ in this case) where $v_e$ is the quasiparticle (excitation) velocity. On the other hand, the change in $\Delta$ reduces the amplitude of $C_-^f$. Moreover, for large $\Delta$,
\begin{equation}
C_{-,l}^f(t) \approx \begin{cases}
-\frac{\sqrt{\pi}\phantom{}_{2}\tilde{F}_{3}\left(1,\frac{3}{2};\frac{1}{2},\frac{3-l}{2},\frac{3+l}{2};- t^{2}\right)}{2\Delta} & \text{$l$ odd}\\
-\frac{i t\phantom{}_{1}\tilde{F}_{2}\left(1;\frac{2-l}{2},\frac{2+l}{2};- t^{2}\right)}{\Delta} & \text{$l$ even},
\end{cases}
\label{eq:CcD}
\end{equation}
while the long time behavior is well represented by
\begin{equation}
C_{-,l}^f(t) \approx \frac{\alpha}{\sqrt{t}} I_l \cos \left( 2 t + \phi_l \right) \label{eq:Rct},
\end{equation}
where $\alpha = \alpha (\Delta)$, $\phi_l$ a phase that depends on the site difference and $I_l=i$ for even $l$ and $I_l=1$ in other case. The decay rate is universal ($t^{-1/2}$), clearly independent from $l$ or $\Delta$. Both $R$ and $C$ in the stationary regime, show the same decay rate as the density and one time $ \langle f^\dagger_{-l}(t) f_l(t) \rangle$ correlation functions. As we shall see, the rather simple structure of the Fermi operator correlation functions will allow us to extract a simple expression for the effective temperature, which coincides with the one expected in the GGE.
At this point one wonders whether the properties observed above are unique of the quasiparticle correlations or manifest in other type of correlation functions. For instance, we shall consider the case of the density-density correlator,
\begin{equation}
C_{+,l}^n(t,t_0)= \langle \left[ n_n(t+t_0), n_m(t_0) \right]_\pm \rangle.
\end{equation}
As $n_i(t)$ is a bosonic operator, we study the usual correlation functions $R_-$ and $C_+$. Fig. \ref{fig:dens-dens} shows the $\Delta$ and lattice site difference $l$ dependence of these functions in the stationary regime. Both functions show a $t^{-1}$ universal decay,
\begin{figure}[ht]
\centering
\subfloat{\label{fig:Cnl}}
\subfloat{\label{fig:Rnl}}
\subfloat{\label{fig:CnD}}
\subfloat{\label{fig:RnD}}
\includegraphics{dens.pdf}
\caption{Density-density two-time correlation functions. Figs. \protect\subref{fig:Cnl} and \protect\subref{fig:Rnl} show the site difference dependence for $\Delta =1 $ ($C_+^n$ in \protect\subref{fig:Cnl}, $R_-^n$ in \protect\subref{fig:Rnl}). The change in $\Delta$ for $C_+^n$ \protect\subref{fig:CnD} and $R_-^n$ \protect\subref{fig:RnD} with $l=0$. In the double logarithmic scale plot \protect\subref{fig:RnD} the dashed lines represent a $t^{-1}$ decay, compatible with both correlators. }
\label{fig:dens-dens}
\end{figure}
\begin{align}
C_{+,l}^n(t) \approx & \frac{\alpha_c}{t} (\beta + \sin 4t)\\
R_{-,l}^n(t) \approx & \frac{\alpha_r}{t} \cos 4t
\end{align}
for the $t\gg 1$ regime, which is also shown by the out of equilibrium one time density correlation $\left \langle n_{l}(t) n_j(t) \right \rangle$. The light-cone effect is also present. As in the previous correlators an increase in the initial superlattice potential intensity decreases the correlation functions amplitude. In the large $\Delta$ limit, both functions can be written as the product of hypergeometric functions:
\begin{align}
C_{+,l}^n(t) \approx & (-1)^l (P_l^2(t)+Z_l^2(t))\label{eq:Cnt} \\
R_{-,l}^n(t) \approx & -4 i (-1)^l \theta(t)P_l(t)Z_l(t) \label{eq:Rnt}
\end{align}
where $P_l$ represents $\phantom{}_{1}\tilde{F}_{2}\left(1;1-\frac{l}{2},1+\frac{l}{2};- t^{2}\right)/2$ for even $l$ and $-\phantom{}_{1}\tilde{F}_{2}\left(1;\frac{3-l}{2},\frac{3+l}{2};- t^{2}\right)i t/2$ for odd $l$ whereas $Z_l$ is $-\phantom{}_{1}\tilde{F}_{2}\left(2;2-\frac{l}{2},2+\frac{l}{2};- t^{2}\right)it/\Delta$ when $l$ is even and $-\phantom{}_{2}\tilde{F}_{3}\left(1,\frac{3}{2};\frac{1}{2},\frac{3-l}{2},\frac{3+l}{2};- t^{2}\right)\sqrt{\pi}/2\Delta$ in the other case.
\subsubsection{Non-local operators}
The last set of operators we shall consider are the hard-core bosons creation and annihilation non-local operators $b_n$ and $b_m^\dagger$ written in terms of the local operators as
\begin{equation}
b_{n}=\prod_{j=1}^{m-1}(1-2f_{j}^{\dagger}f_{j}) f_{m},\quad b_{m}^{\dagger}=f_{m}^{\dagger}\prod_{j=1}^{m-1}(1-2f_{j}^{\dagger}f_{j}). \label{eq:jw}
\end{equation}
Non-local two-time correlations have been already studied in Refs. \onlinecite{rossini10_quantum_ising_quench,foini11_quench_FDT,foini12_quench_FDT_long} for the quantum Ising model in a transverse magnetic field. In these papers the computation $\langle \sigma^x_n(t+t_0) \sigma^x_m(t_0)\rangle$ involves calculating the four-spin correlation function done by means of a $2L\times 2L$ T\"{o}plitz determinant. The two-spin correlator is then recovered by taking the thermodynamic limit and making use of the cluster property. For our model, the fermionic Hamiltonian (equation \eqref{eq:H0}) does not contain anomalous terms and therefore we can make use of a simpler straightforward approach. We start by defining the hermitian combination $B_i=b_i+b_i^\dagger$ and considering the two-point correlation functions $C_\pm^B$,
\begin{equation}
C_\pm^B(t,t_0)= \langle \left[ B_n(t+t_0), B_m(t_0) \right]_\pm \rangle \label{eq:noloc2t}
\end{equation}
from which we can calculate the response function. We observe that only one of the two terms in equation \eqref{eq:noloc2t} is needed, as $C_+^B=\Real \langle B_n B_m \rangle$ and $C_-^B=i\Imag \langle B_n B_m \rangle$. Using the definition, we obtain
\begin{equation}
\langle B_n(t) B_m(t') \rangle = \langle b_n(t) b_m^\dagger(t') \rangle+ \langle b_n^\dagger(t) b_m(t') \rangle
\label{eq:nlB}
\end{equation}
since the remaining terms vanish. The first term in equation \eqref{eq:nlB} can be computed by extending the approach presented in Ref. \onlinecite{rigol05_groundstate_hcbosons} for different times. We can write
\begin{equation}
\left\langle \Psi\right|b_n(t)b_m^\dagger(t')\left| \Psi \right\rangle= \left\langle \Psi(t)\right|b_n e^{-i Ht}e^{i Ht'} b_m^\dagger\left| \Psi(t') \right\rangle \label{eq:noloc1}
\end{equation}
where $b_m^\dagger (b_n)$ can be mapped to fermions by the equation \eqref{eq:jw} and $\Psi(t)$ is the time evolved ground state:
\begin{align}
\left| \Psi(t') \right\rangle &= \prod_{\nu=1}^N e^{-iHt'} c_\nu^\dagger |0\rangle= \prod_{\nu=1}^N e^{-iHt'} c_\nu^\dagger e^{iHt'} e^{-iHt'} |0\rangle \nonumber \\
&= \prod_{\nu=1}^N \sum_{j=1}^L f_j^\dagger \varphi_{\nu}(j,t')\left| 0\right\rangle
\end{align}
where $c^\dagger_\nu$ are the operators that render $H_0$ diagonal and $\varphi_{\nu}(j,t')$ the time dependent eigenfunctions of $H_0$. Then
\begin{equation*}
b^\dagger_m \left| \Psi(t') \right\rangle = f_{m}^{\dagger}\prod_{j=1}^{m-1}(1-2f_{j}^{\dagger}f_{j}) \prod_{\nu=1}^N \sum_{j=1}^L f_j^\dagger \varphi_{\nu}(j,t')\left| 0\right\rangle.
\end{equation*}
Then we define a $L\times N$ matrix $P(t')$ with elements $\varphi_{\nu}(j,t')$. Then the action of $b_m^\dagger$ on $\left| \Psi(t') \right\rangle$ amounts to change the signs of elements $P_{j\nu}$ with $j\leq m-1$ and the further creation of a particle at site $m$ implies the addition of a column to $P$ with elements $P_{i,N+1}=\delta_{im}$. Thus, we can write
\begin{align}
e^{i Ht'} b_m^\dagger\left| \Psi(t') \right\rangle&= \prod_{\nu=1}^{N+1} \sum_{j=1}^L f_j^\dagger(t') P'_{j\nu}(t')\left| 0\right\rangle\\
&= \prod_{\nu=1}^{N+1} \sum_{i=1}^L f_i^\dagger Q_{i\nu}(t') \left| 0\right\rangle
\end{align}
where $P'$ is obtained by changing the required signs and adding the new column, and $Q(t')= e^{iht'}P'(t)$ is again a $L\times N$ matrix, where $h$ is the matrix representation of the Hamiltonian $H$. Hence, we can rewrite equation \eqref{eq:noloc1} as
\begin{align}
\left\langle \Psi\right|b_n(t)b_m^\dagger(t')\left| \Psi \right\rangle&= \det Q^\dagger(t) Q(t')\\
&= \det P'^\dagger_n (t)e^{-ih(t-t')} P'_m(t') \label{eq:nl1}.
\end{align}
The second term in the correlator \eqref{eq:nlB} is more involved since we can no longer create a new column in $P$ as the fermionic creation and destruction operators are permuted with respect to the ground state operators,
\begin{align}
\langle b_n^\dagger(t) b_m(t') \rangle =& \prod_{\mu,\nu=1}^N\sum_{j,l=1}^L \varphi^\ast_{\mu}(l,t) \varphi_{\nu}(j,t') \times \nonumber \\
&\left\langle 0 \right| f_l \cdots f_{n}^{\dagger} e^{-i H(t-t')} f_{m}\cdots f_j^\dagger \left| 0\right\rangle
\end{align}
We circumvent this issue by employing the following property: Calling $\tau = t-t'$:
\begin{align}
f_n^\dagger e^{-iH\tau}f_m=& e^{-iH\tau} f_n^\dagger(\tau) f_m = e^{-iH\tau}\sum_{j=1}^N f_j^\dagger (e^{ih \tau})_{jn} f_m \nonumber \\
=- e^{-iH\tau}& f_m e^{iH\tau}f_n^\dagger e^{-iH\tau} +e^{-iH\tau}(e^{ih \tau})_{mn}\label{eq:nl2}
\end{align}
Then $\langle b_n^\dagger(t) b_m(t') \rangle$ can be written as
\begin{align}
\langle b_n^\dagger(t) b_m(t') \rangle =& \det P^{\dagger}_n (t) e^{-ih (t-t')} P_m(t')\nonumber \\& \quad - \det O_n^{m\dagger}(t,t') O_m^{n}(t',t)
\end{align}
where $P_m$ is $P'_m$ with no additional column and $O_m^n$ is a $L \times N+1$ matrix defined by
\begin{equation}
O_m^{n}(t',t)=
\begin{cases}
(e^{iht'}P^m(t'))_{j \nu} & \mbox{for } \nu = 1, \ldots ,N \\
(e^{iht})_{j n} & \mbox{for } \nu = N+1
\end{cases}
\end{equation}
for $j = 1, \ldots , L $. We can recover $\langle B_n (t+t_0)B_m(t_0)\rangle $ by adding expressions \eqref{eq:nl1} and \eqref{eq:nl2} and taking $t'\rightarrow t_0$ and $t\rightarrow t+t_0$. Thus, for our model, this approach reduces the computation of non-local correlations to the evaluation of $(N+1)\times (N+1)$ matrix determinants, instead of determinants of $2L\times 2L$ T\"{o}plitz matrices.
We compute the non-local correlation function using a system with 1000 lattice sites with open-boundary conditions, half-filled ($N=L/2$) and taking $t_0 = 100$ as the stationary limit. In Fig. \ref{fig:noloc} we show the results obtained for $C_+$ (\ref{fig:nolocC}) and linear response function (\ref{fig:nolocR}).
\begin{figure}[ht]
\vspace*{-12pt}
\subfloat{\label{fig:nolocC}}
\subfloat{\label{fig:nolocR}}
\includegraphics{noloc.pdf}
\caption{Non-local two-time correlation functions. Figs. \protect\subref{fig:nolocC} and \protect\subref{fig:nolocR} show the site difference dependence for $C^B_+$ and the $\Delta$ dependence in $R^B_-$, respectively. In the insets the same information in semi logarithmic axis along with exponential decays. The insets also present the exponential decay constants $\gamma_c(\Delta)$ and $\gamma_r(\Delta)$ respectively. }
\label{fig:noloc}
\end{figure}
These functions present an exponential decay whose rate depends on the initial superlattice potential $\Delta$, and is independent of the lattice difference (shown in the insets of Figs. \ref{fig:nolocC} and \ref{fig:nolocR}). The long time behavior is well fitted by
\begin{align}
C_{+,l}^{B}(t)=& \alpha_c e^{- \gamma_c t} \left( \beta_c + \frac{\sin (2t+\phi)}{\sqrt{t}} \right)\\
R_{-,l}^{B}(t)=& \alpha_r e^{- \gamma_r t} \left( \beta_r + \frac{\sin (2t+\phi)}{\sqrt{t}} \right)
\end{align}
i.e, damped oscillations modulated by an exponential decay dictated by $\gamma_i = \gamma_i (\Delta)$.
\subsubsection{Initial state at finite temperature}
\label{sec:timetemp}
We extend our analysis to the case in which the initial state is a thermal state with temperature $T$, described by $\rho_0 = \exp(-H_0/T)/Z$, which involves working in the grand canonical ensemble (GCE). This raises a new problem as the border terms $f_L^\dagger f_{L+1}$ are treated by imposing (anti-)periodic boundary conditions which depend on the number of particles $N$ in the system, and $N$ is not fixed in the GCE. One possible workaround could be to calculate the correlations using open boundary conditions, but this approach complicates the analytical results. We address this issue by keeping the simplicity of analytically calculated periodic boundary conditions correlators and checking the relevance of the border terms comparing these results with the ones obtained by solving the problem numerically with open-boundary conditions (shown as dots in Fig. \ref{fig:timetemps}). We checked the independence of the boundary conditions for the correlators in the zero temperature case far from the lattice borders.
Following the zero temperature analysis done before, we start by studying the Fermi operator correlators. We compute $C_{\pm,l}^f(t,t_0)= \langle \left[ f_n(t+t_0),f^\dagger_m(t_0) \right]_\pm \rangle$ where $\left\langle \cdots\right\rangle $ now represents $\Tr [\rho_0 \cdots]$. In the thermodynamic limit and stationary regime the $R_+$ correlator is the same as in the zero temperature case (equation \eqref{eq:Cc}), i.e. it has neither $\Delta$ nor initial temperature dependence. The differences between this result and the one obtained numerically with open boundary conditions are negligible. The temperature $T$ and superlattice potential $\Delta$ dependences are only contained in the linear response function
\begin{equation}
C_{-,l}^f(t,T)= \intop_{-\pi/2}^{\pi/2} dk\; \mathcal{I}\, \tanh(\sqrt{\omega_k^2+ \Delta^2}/2T),
\end{equation}
where $\mathcal{I}$ represents the integrand in equation \eqref{eq:Rc}. In Figs. \ref{fig:ImCcparT} and \ref{fig:ReCcimparT} we show this function (solid lines) and the numerical calculations (dots) varying the reservoir temperature $T$. The agreement of both calculations, periodic and open boundary, shows that the border terms are not significant.
\begin{figure}[ht]
\vspace*{-12pt}
\subfloat{\label{fig:ImCcparT}} \subfloat{\label{fig:ReCcimparT}} \subfloat{\label{fig:CnT}}\subfloat{\label{fig:RnT}}
\includegraphics{timetemp.pdf}
\caption{Initial temperature dependence of the two-time correlators. Figs. \protect\subref{fig:ImCcparT} and \protect\subref{fig:ReCcimparT} show the $C_-$ correlator for Fermi operators with different site difference, while \protect\subref{fig:CnT} and \protect\subref{fig:RnT} are $C_+$ and $R_-$ for density operators. In both cases $\Delta=1$ and $l=0$. The solid lines represent periodic boundary conditions while the dots are open boundary conditions correlators.}
\label{fig:timetemps}
\end{figure}
We notice that the limit $T \rightarrow 0$ is well defined as we recover the zero temperature result. Varying the initial temperature has a similar behavior in $C_-$ as changing the supperlattice potential strength $\Delta$. Moreover, the large $T$ limit as in the Fermi case is identical to $\Delta \gg 1$ regime (equation \eqref{eq:CcD}) taking $\Delta = 2T$, while the strong insulator limit is the same as in the $T=0$ case. Furthermore, for large time difference ($t\gg 1$) it has the same behavior as in zero temperature, shown in equation \eqref{eq:Rct}, with $\alpha = \alpha(\Delta,T)$.
The analysis of density-density correlators with an initial thermal state, shown in Figs. \ref{fig:CnT} and \ref{fig:RnT}, shows similar features than the Fermi correlators. The effect of rising $T$ is similar to the one produced by increasing $\Delta$ and the high temperatures limits is well described by equations \eqref{eq:Cnt} and \eqref{eq:Rnt} taking $\Delta=2T$. As in the Fermi case, open (dots) and periodic boundary (lines) conditions correlators coincide, showing that the border terms do not play an important role in the studied correlations.
\subsection{Frequency dependence}
\label{sec:corrfrec}
In this section we analyze the frequency dependence of the correlation calculated in section \ref{sec:corrtime}. More specifically, we study the Fourier transform of the linear response function imaginary part and the (anti-)symmetric correlator in the stationary and thermodynamic limits, both of the functions related by the fluctuation-dissipation theorem. Following the order established in section \ref{sec:corrtime}, we start by analyzing the simpler Fermi correlations, whose linear response function imaginary part in the frequency space is
\begin{equation}
\mathrm{Im}R^f_{+,l}(\omega) = \theta \left(1- \omega^2/4\right) \frac{e^{i \pi l}T_l(\omega/2)}{\sqrt{1-\omega^2/4}}, \label{eq:Cfom}
\end{equation}
where $T_n(x)$ are the Chebyshev polynomials of the first kind and degree $n$. The higher contribution to $R_+$ comes from frequencies from the bands' edge ($\omega \approx\pm 2$), while the $T_n$ polynomials mostly modify the center of the band as the site difference $l$ increases. Furthermore, the antisymmetric correlator is
\begin{equation}
C_{-,l}^{f} (\omega) = \mathrm{Im} R_{+,l}^f (\omega) \frac{\omega}{\sqrt{\omega^2+ \Delta^2}},
\end{equation}
which shows the same bandwidth and functional dependence in $l$. The main effect of increasing the supperlatice potential strength is to reduce the contribution of the frequencies in the center of the band to $C_-$. Since $\mathrm{Im}R_+$ can be factorized from $C_-$, the effective temperature can be easily extracted (see equation \eqref{eq:Teffc}). On the other hand, when the system is in contact with a thermal reservoir the temperature dependence appears in $C_-^f$ through a multiplicative factor,
\begin{equation}
C_{-,l}^{f} (\omega,T) = C_{-,l}^{f} (\omega) \tanh \frac{\sqrt{\omega^2+\Delta^2}}{2T}.
\end{equation}
Even though it clearly modifies the response function, the main consequence of rising the temperature of the initial reservoir is similar to the one produced by increasing $\Delta$: decreasing the contribution of the low frequency modes in the correlation function
as $T \rightarrow \infty$. As expected from the results shown in section \ref{sec:timetemp}, the linear response function is independent of $T$, coinciding with equation \eqref{eq:Cfom}.
\begin{figure}[ht]
\vspace*{-12pt}
\subfloat{\label{fig:Comega}} \subfloat{\label{fig:ImromegaT}} \subfloat{\label{fig:Cnolocomega}}\subfloat{\label{fig:Imrnolocomega}}
\includegraphics{frecs.pdf}
\caption{Density and non-local correlators in frequency space. Panel \protect\subref{fig:Comega} and \protect\subref{fig:ImromegaT} show the density symmetric correlator and linear response imaginary part for different $\Delta$ values. Panels \protect\subref{fig:Cnolocomega} and \protect\subref{fig:Imrnolocomega} present the same functions for non-local operators. In all the cases $l=0$.}
\label{fig:frecs}
\end{figure}
Next, we study the frequency dependence of the density and non-local correlations in the thermodynamic limit and stationary regime, by performing a discrete Fourier transform over the time-dependent correlators in $t\in[0,100]$ with a time interval $\tau=0.25$. In Fig. \ref{fig:frecs} we plot these functions, only showing the positive frequency sector as both functions have definite parity ($C_+$ is even and $\textrm{Im}R_-$ is odd). Both density correlators (Figs. \ref{fig:Comega} and \ref{fig:ImromegaT}) present a contribution from frequencies between $-4 \leq\omega \leq 4$. For small values of $\Delta$ the contribution of higher frequencies to $C_+^n$ is important, but as the initial potential increases the lower frequency modes become more relevant. In the case of $\textrm{Im}R_-^n$, the amplitude seems to be inversely proportional to $\Delta$, decreasing the contribution of all frequency modes for higher potential values. Finally, the non-local correlators present a different panorama, as both functions amplitude decrease as the frequency increases. Analyzing the variation with $\Delta$, we notice that the symmetric correlator remains almost unchanged, only becomes smoother with this change. The linear response imaginary part presents a peak around $\omega\sim 1.3$, which reduces its amplitude and shifts to higher frequencies as the initial superlattice potential rises. The frequency-dependent correlators obtained in this section shall be employed in the calculation of effective temperature, depicted in section \ref{sec:efftemp}.
\section{Effective temperatures from FDRs}
\label{sec:fdr}
In this section we compute the effective temperatures from the correlators studied in section \ref{sec:corr}, analyzing both zero and finite temperature initial states. Let us start by stating some generalities of the fluctuation-dissipation theorem (FDT).
For typical observables having bosonic properties, the correlation function $C^-$ is used to construct the retarded function $R_-^{AB}(t,t_0)=2i\theta(t)C^{AB}_-(t,t_0)$, while, in the case of Fermi operators which do not commute, the retarded function is defined employing the commutator, $R_+^{AB}(t,t_0)=2i\theta(t)C^{AB}_+(t,t_0)$. The FDT relates the functions $R_\pm^{AB}$ and $C_\mp^{AB}$ in equilibrium at inverse temperature $\beta$. In the frequency domain, where
\begin{equation}
R_\pm^{AB}(\omega)=\int_{-\infty}^{\infty} dt e^{i\omega t} R_\pm^{AB}(t),
\end{equation}
it takes the form
\begin{equation}
\Im R_\pm^{AB}(\omega)=\left[\tanh\frac{\omega\beta}{2}\right]^{\mp 1}C_\mp^{AB}(\omega).
\end{equation}
Before obtaining specific results for effective temperatures from FDT for this model, let us state a general result valid for quasi-free systems whose static correlations relax to the GGE. In this case dynamic correlations of local operators are also asymptotically described by the GGE~\cite{essler12_dynamical_gge}. By using a spectral decomposition in terms of eigenstates of the Hamiltonian one can show that a basic form of the FDT holds out of equilibrium for long times~\cite{essler12_dynamical_gge}:
\begin{equation}
-\frac{1}{\pi}\Im\chi_{\mathcal{A}\mathcal{B}}\left(\omega\right)=S_{\mathcal{A}\mathcal{B}}(\omega)-S_{\mathcal{B}\mathcal{A}}(-\omega).
\end{equation}
However, differently from the usual FDT for systems in thermodynamic equilibrium, the negative $S_{\mathcal{A}\mathcal{B}}(\omega)$ and positive $S_{\mathcal{B}\mathcal{A}}(-\omega)$ parts of the spectral function in general are not simply related by $S_{\mathcal{B}\mathcal{A}}(-\omega)=e^{-\beta\omega}S_{\mathcal{A}\mathcal{B}}(\omega)$, where $\beta$ is the inverse temperature. We will show that after relaxation from a quantum quench it is possible to establish an analogous relation for correlations of quasiparticle creation and destruction operators.
Consider a general bilinear Hamiltonian $H_b=\sum_{i,j} f^\dagger_i h_{ij} f_j$ where $f^\dagger_i$ and $f^\dagger_j$ are destruction and creation fermionic operators and $h$ a symmetric matrix. $H_b$ is diagonalized by a canonical transformation $f_j=\sum_\nu\mathcal{U}_{j,\nu}f_\nu$, $H_b=\sum_\nu\varepsilon_\nu f^\dagger_\nu f_\nu$ where $\varepsilon_\nu$ is the dispersion relation. Consider the correlation function for the Fermi field
\begin{align}
C^{ij}(t,t_0)=&\langle f_i(t+t_0)f_j^\dagger(t_0)\rangle\\ =&\sum_{\mu\nu}\mathcal{U}^\ast_{i\mu}\mathcal{U}_{j\nu} e^{-i\varepsilon_\nu(t+t_0)}e^{i\varepsilon_\mu t_0}\langle f_\mu f^\dagger_\nu\rangle\label{eq:correlator}
\end{align}
where $\langle\ldots\rangle$ is the initial state. Even though the correlator $\langle f_\mu f^\dagger_\nu\rangle$ is not diagonal for initial states that are not translation invariant, for rather standard conditions the non diagonal contributions decay rapidly and vanish in the thermodynamic limit\cite{cazalilla12_thermalization_correlations,barthel08_quench} which constitutes the way by which dephasing takes place. In the specific model we are analyzing, the eigenmode correlator is not diagonal in momentum space, but the only contribution outside the diagonal is the correlation between modes at $k$ and $k+\pi$, $\langle f_k f^\dagger_{k+\pi}\rangle =\Delta/E_k$. In the thermodynamic limit these terms yield a smooth function of $k$ and therefore by application of the Riemann-Lebesgue theorem do not contribute to Eq. (\ref{eq:correlator}):
\begin{equation}
\lim_{t_0\to +\infty}C^{ij}(t,t_0)=\sum_{\mu}\mathcal{U}^\ast_{i\mu}\mathcal{U}_{j\mu} e^{-i\varepsilon_\mu t}\left[1-N_0(\varepsilon_\mu)\right]
\end{equation}
where $N_0(\varepsilon_\nu)=1/[e^{\lambda(\varepsilon_\nu)}+1]$ are the mode occupations in the initial state. From this correlator we can construct the response and the correlation function, which in frequency space read
\begin{align}
\Im R_+(\omega)=&\pi\sum_{\mu}\mathcal{U}^\ast_{i\mu}\mathcal{U}_{j\mu} \delta(\omega-\varepsilon_\mu)\\
C_-(\omega)=&\pi\sum_{\mu}\mathcal{U}^\ast_{i\mu}\mathcal{U}_{j\mu} \delta(\omega-\varepsilon_\mu)[1-2N_0(\varepsilon_\mu)]
\end{align}
Therefore, both functions are related as
\begin{equation}
\Im R_{+}(\omega)= \left[\tanh\frac{\lambda(\omega)}{2}\right]^{-1} C_-(\omega),
\end{equation}
and therefore we have a frequency-dependent effective temperature $1/ T_\mathrm{eff}(\omega)=\lambda(\omega)/\omega$. We notice that this result is generic for initial states and quenches to quasi-free models for which the long-times regime is captured by the GGE.
\subsection{Effective temperatures for local and non-local operators}
\label{sec:efftemp}
After obtaining this general result, we wish to explore the effective temperatures extracted from the correlators calculated for our model shown in section \ref{sec:corr}. In general, these can be written as
\begin{equation}
T^{AB}_{\mathrm{eff}}(\omega)=\frac{\omega}{2}\mathrm{arctanh}^{-1}\left[\left(\frac{\Im R_\pm^{AB}(\omega)}{C_\mp^{AB}(\omega)}\right)^{\mp 1}\right].
\end{equation}
For out of equilibrium systems these temperatures usually depend on frequency and the operators studied. However, if the system achieves a thermal state after long times, all of the $T_\mathrm{eff}$ should be equal and frequency independent, at least for a value of $t_0$ large enough.
\begin{figure}[ht]
\centering
\subfloat{\label{fig:Teffn}}\subfloat{\label{fig:TeffB}}\subfloat{\label{fig:TeffnT}}
\includegraphics{Teff.pdf}
\caption{Effective temperatures for the different autocorrelation functions: density operators ($T_{\mathrm{eff}}^n(\omega)$) in Fig. \protect\subref{fig:Teffn}, non-local operators ($T_{\mathrm{eff}}^B(\omega)$) in Fig. \protect\subref{fig:TeffB} and density operators ($T_{\mathrm{eff}}^n({\omega,T)}$) for an initial thermal state in Fig. \protect\subref{fig:TeffnT}.}
\label{fig:Teff}
\end{figure}
Let us start with the Fermi operators correlations, whose effective temperature $T^f_{\mathrm{eff}}$ can be calculated analytically, being
\begin{equation}
T^f_{\mathrm{eff}}(\omega) = \frac{\omega}{2}\mathrm{arctanh}^{-1}\left[\frac{\omega}{\sqrt{\omega^2 + \Delta^2}}\right] \label{eq:Teffc}.
\end{equation}
Thus, we obtain a frequency dependent effective temperature that is is independent of the site difference, even though the correlation functions depend on this difference. Nevertheless, one can check the fidelity of $T_\mathrm{eff}^f$: by reducing the size of the quench by taking $\Delta \rightarrow 0$, $T_\mathrm{eff}^f\rightarrow 0$ and equilibrium is recovered. As we expected from the general result above, $T_\mathrm{eff}^f$ coincides with the temperature calculated in the GGE ($T_\mathrm{eff}^\mathrm{GGE}$) and therefore is $T^f_\mathrm{eff} \approx \Delta/2$ in the $\Delta \gg 1$ regime.
At this point the relevant question is whether these characteristics are shared by the effective temperatures that correspond to other observables. In Figs. \ref{fig:Teffn} and \ref{fig:TeffB} we show the temperatures obtained for the autocorrelation functions ($l=0$) of density and non-local operators, respectively. As one could expect, they do not share the same frequency dependence and are different from $T_\mathrm{eff}^f$. However, as $\Delta$ increases, the effective temperature from density correlations smooths out and reduce its amplitude approaching the value $\Delta/2$ predicted by the GGE temperature, as is shown in the inset of Fig. \ref{fig:Teffn}. Although in this regime the system seems to approach a standard Gibbs ensemble with temperature $T=\Delta/2$, the remaining frequency dependence, as in the case of $T_\mathrm{eff}^f$, discards thermalization. In the non local case (Fig. \protect\subref{fig:TeffB}) the effective temperature seems to approach $\Delta/2$ for large values of $\Delta$. However its deviations from this value at intermediate frequencies are larger than in the local case, and do not vanish in the limit $\Delta\to\infty$.
When the system is connected with a thermal reservoir before the quench, the properties of the effective temperatures are quite similar to the ones above. For the Fermi operators, the additional temperature dependence in $T_\mathrm{eff}^f(\omega,T)$ is given by an extra factor in the argument of the hyperbolic arctangent,
\begin{equation}
T^f_{\mathrm{eff}}(\omega,T) = \frac{\omega}{2}\mathrm{arctanh}^{-1}\left[\frac{\omega \tanh \left( \frac{\sqrt{\omega^2 + \Delta^2}}{2T}\right)}{\sqrt{\omega^2 + \Delta^2}} \right].
\end{equation}
As $T_\mathrm{eff}^f(\omega)$, it shows an independence on the site difference $l$. It also presents a well defined ``equilibrium'' limit approaching $T$ as $\Delta \rightarrow 0$, while in the $\Delta \gg 1$ regime follows the GGE temperature. As expected by the results in section \ref{sec:corr}, the high temperature regime is $T_\mathrm{eff}^f(\omega,T) \approx T$, but as a residual frequency dependence remains, a thermal state is not reached in this regime. The density-density autocorrelation function, shown in Fig. \ref{fig:TeffnT}, presents a similar panorama. Its frequency dependence is different from the correlators above, although as $\Delta$ or $T$ rises its value approaches $\Delta/2$ or $T$, respectively. Comparing Figs. \ref{fig:Teffn} and \ref{fig:TeffnT}, it seems that one can reach a state similar to a standard Gibbs state faster by increasing the reservoir temperature than by rising $\Delta$, as the inset in Fig. \ref{fig:TeffnT} shows a smaller dispersion than the inset in Fig. \ref{fig:Teffn}. This can be explained by the initial thermal reservoir, which favors an incoherent evolution of the system. Nevertheless, the persistent frequency dependence hints a non thermal state. We stress that the system does not reach a Gibbsian unique temperature state even after long times, as if it did, all the calculated effective temperatures should be equal and constant.
\section{Summary}
\label{sec:conclusions}
To conclude, we analyzed various dynamic correlation functions, for local and non-local operators after a quantum quench in an exactly solvable model in which the statistical description in terms of the GGE essentially leads to the emergence of thermal correlations. This is due to the existence of bi-partite eigenmode entanglement and a gap in the spectrum of the Hamiltonian that describes the initial state. For these correlations, the imposition of the FDT in the non-equilibrium context leads to the appearance of an effective temperature depending on frequency (and eventually momentum or position) that is different for each operator considered. Nevertheless, in the limit of strong initial entanglement, in agreement with the emergence of thermal behavior from the GGE, the local operators effective temperatures approach a well defined value (in a certain frequency region). However, the remaining frequency dependence of these temperatures and the fact that the non-local temperature does not follow this limit, discards thermalization to a standard Gibbs state in a strict sense. Finally, it is of particular interest the case of the frequency-dependent effective temperature obtained from the application of the FDT to the quasiparticle correlation function, evaluated at the dispersion relation of the Hamiltonian that performs the evolution. This effective temperature is directly related to the GGE Lagrange multipliers.
\acknowledgments
This work was partially supported by CONICET (PIP 0662), ANPCyT (PICT 2010-1907) and UNLP (PID X497), Argentina.
|
\section{Introduction}\label{se:intro}
It is typical in Computer Science to classify problems according to the amount of resources that are needed to solve them. Hence, problems are usually classified according to the amount of time or to the amount of memory that a specific model of computation requires for their solution.
This epistemological need of classifying problems finds, in the Graph Drawing field, a very original interpretation. A Graph Drawing problem can be broadly described as follows: Given a graph of a certain family and a drawing convention (e.g.\ all edges should be straight-line segments), draw the graph optimizing some specific features. Among those features a fundamental one is the amount of geometric space that the drawing spans and a natural question is: Which is the amount of space that is required for drawing a planar graph, or a tree, or a bipartite graph? Hence, besides classifying problems according to the above classical coordinates, Graph Drawing classifies problems according to the amount of geometric space that a drawing that solves that problem requires.
Of course, such a space requirement can be influenced by the class of graphs (one can expect that the area required to draw an $n$-vertex tree is less than the one required to draw an $n$-vertex general planar graph) and by the drawing convention (straight-line drawings look more constrained than drawings where edges can be polygonal lines).
The attempt of classifying graph drawing problems with respect to the space required spurred, over the last fifty years, a large body of research. On one hand, techniques have been devised to compute geometric lower bounds that are completely original and do not find counterparts in the techniques adopted in Computer Science to find time or memory lower bounds. On the other hand, the uninterrupted upper bound hunting has produced several elegant algorithmic techniques.
In this paper we survey the state of the art on such algorithmic and lower bound techniques for several families of planar graphs. Indeed, drawing planar graphs without crossings is probably the most classical Graph Drawing topic and many researches gave fundamental contributions on planar drawings of trees, outerplanar graphs, series-parallel graphs, etc.
We survey the state of the art focusing on the impact of the most popular drawing conventions on the geometric space requirements. In Section~\ref{se:straight-line} we discuss straight-line drawings. In Section~\ref{se:poly-line} we analyze drawings where edges can be polygonal lines. In Section~\ref{se:upward} we describe upward drawings, i.e.\ drawings of directed acyclic graphs where edges follow a common vertical direction. In Section~\ref{se:convex} we describe convex drawings, where the faces of a planar drawing are constrained to be convex polygons. Proximity drawings, where vertices and edges should enforce some proximity constraints, are discussed in Section~\ref{se:proximity}. Section~\ref{se:clustered} is devoted to drawings of clustered graphs.
We devote special attention to put in evidence those that we consider the main open problems of the~field.
\section{Preliminaries}\label{se:preliminaries}
In this section we present preliminaries and definitions. For more about graph drawing, see~\cite{dett-gd-99,kw-dgmm-01}.
\subsection*{Planar Drawings, Planar Embeddings, and Planar Graphs} \label{se:graphs-planarembeddings}
All the graphs that we consider are \emph{simple}, i.e., they contain no multiple edges and loops. A \emph{drawing} of a graph $G(V,E)$ is a mapping of each vertex of $V$ to a point in the plane and of each edge of $E$ to a simple curve connecting its endpoints. A drawing is \emph{planar} if no two edges intersect except, possibly, at common endpoints. A \emph{planar graph} is a graph admitting a planar drawing.
A planar drawing of a graph determines a circular ordering of the edges incident to each vertex. Two drawings of the same graph are \emph{equivalent} if they determine the same circular ordering around each vertex and a \emph{planar embedding} (sometimes also called {\em combinatorial embedding}) is an equivalence class of planar drawings. A graph is \emph{embedded} when an embedding of it has been decided. A planar drawing partitions the plane into topologically connected regions, called \emph{faces}. The unbounded face is the \emph{outer face}, while the bounded faces are the \emph{internal faces}. The outer face of a graph $G$ is denoted by $f(G)$. A graph together with a planar embedding and a choice for its outer face is a \emph{plane graph}. In a plane graph, \emph{external} and \emph{internal} vertices are defined as the vertices incident and not incident to the outer face, respectively. Sometimes, the distinction is made between \emph{planar embedding} and \emph{plane embedding}, where the former is an equivalence class of planar drawings and the latter is a planar embedding together with a choice for the outer face. The \emph{dual graph} of an embedded planar graph $G$ has a vertex for each face of $G$ and has an edge $(f,g)$ for each two faces $f$ and $g$ of $G$ sharing an edge.
\subsection*{Maximality and Connectivity} \label{se:graphs-connectivity}
A plane graph is \emph{maximal} (or equivalently is a \emph{triangulation}) when all its faces are delimited by \emph{$3$-cycles}, that is, by cycles of three vertices. A planar graph is \emph{maximal} when it can be embedded as a triangulation. Algorithms for drawing planar graphs usually assume to deal with maximal planar graphs. In fact, any planar graph can be augmented to a maximal planar graph by adding some ``dummy'' edges to the graph. Then the algorithm can draw the maximal planar graph and finally the inserted dummy edges can be removed obtaining a drawing of the input graph.
A graph is \emph{connected} if every pair of vertices is connected by a path. A graph with at least $k+1$ vertices is \emph{$k$-connected} if removing any (at most) $k-1$ vertices leaves the graph connected; $3$-connected, $2$-connected, and $1$-connected graphs are also called \emph{triconnected}, \emph{biconnected}, and \emph{connected} graphs, respectively. A \emph{separating cycle} is a cycle whose removal disconnects the graph.
\subsection*{Classes of Planar Graphs} \label{se:graphs-classes}
A \emph{tree} is a connected acyclic graph. A \emph{leaf} in a tree is a node of degree one. A \emph{caterpillar} $C$ is a tree such that the removal from $C$ of all the leaves and of their incident edges turns $C$ into a path, called the \emph{backbone} of the caterpillar.
A \emph{rooted tree} is a tree with one distinguished node called \emph{root}. In a rooted tree each node $v$ at distance (i.e., length of the shortest path) $d$ from the root is the \emph{child} of the only node at distance $d-1$ from the root $v$ is connected to. A \emph{binary tree} (a \emph{ternary tree}) is a rooted tree such that each node has at most two children (resp. three children). Binary and ternary trees can be supposed to be rooted at any node of degree at most two and three, respectively. The \emph{height} of a rooted tree is the maximum number of nodes in any path from the root to a leaf. Removing a non-leaf node $u$ from a tree disconnects the tree into connected components. Those containing children of $u$ are the \emph{subtrees} of $u$.
A \emph{complete tree} is a rooted tree such that each non-leaf node has the same number of children and such that each leaf has the same distance from the root. Complete trees of degree three and four are also called \emph{complete binary trees} and \emph{complete ternary trees}, respectively.
A rooted tree is \emph{ordered} if a clockwise order of the neighbors of each node (i.e., a planar embedding) is specified. In an ordered binary tree and in an ordered ternary tree, fixing a linear ordering of the children of the root yields to define the \emph{left} and \emph{right child} of a node, and the \emph{left}, \emph{middle}, and \emph{right child} of a node, respectively. If the tree is ordered and binary (ternary), the subtrees rooted at the left and right child (at the left, middle, and right child) of a node $u$ are the \emph{left} and the \emph{right subtree} of $u$ (the \emph{left}, the \emph{middle}, and the \emph{right subtree} of $u$), respectively. Removing a path $P$ from a tree disconnects the tree into connected components. The ones containing children of nodes in $P$ are the \emph{subtrees} of $P$. If the tree is ordered and binary (ternary), then each component is a \emph{left} or \emph{right subtree} (a \emph{left}, \emph{middle}, or \emph{right subtree}) of $\mathcal P$, depending on whether the root of such subtree is a left or right child (is a left, middle, or right child) of a node in $\mathcal P$, respectively.
An \emph{outerplane graph} is a plane graph such that all the vertices are incident to the outer face. An \emph{outerplanar embedding} is a planar embedding such that all the vertices are incident to the same face. An \emph{outerplanar graph} is a graph that admits an outerplanar embedding. A \emph{maximal outerplane graph} is an outerplane graph such that all its internal faces are delimited by cycles of three vertices. A \emph{maximal outerplanar embedding} is an outerplanar embedding such that all its faces, except for the one to which all the vertices are incident, are delimited by cycles of three vertices. A \emph{maximal outerplanar graph} is a graph that admits a maximal outerplanar embedding. Every outerplanar graph can be augmented to maximal by adding dummy edges to it.
If we do not consider the vertex corresponding to the outer face of $G$ and its incident edges then the dual graph of an outerplane graph $G$ is a tree. Hence, when dealing with outerplanar graphs, we talk about the \emph{dual tree} of an outerplanar graph (meaning the dual graph of an outerplane embedding of the outerplanar graph). The nodes of the dual tree of a maximal outerplane graph $G$ have degree at most three. Hence the dual tree of $G$ can be rooted to be a binary tree.
\emph{Series-parallel graphs} are the graphs that can be inductively constructed as follows. An edge $(u,v)$ is a series-parallel graph with \emph{poles} $u$ and $v$. Denote by $u_i$ and $v_i$ the poles of a series-parallel graph $G_i$. Then, a \emph{series composition} of a sequence $G_1,G_2,\dots,G_k$ of series-parallel graphs, with $k\geq 2$, constructs a series-parallel graph that has poles $u=u_1$ and $v=v_k$, that contains graphs $G_i$ as subgraphs, and such that vertices $v_i$ and $u_{i+1}$ have been identified to be the same vertex, for each $i=1,2,\dots,k-1$. A \emph{parallel composition} of a set $G_1,G_2,\dots,G_k$ of series-parallel graphs, with $k\geq 2$, constructs a series-parallel graph that has poles $u=u_1=u_2=\cdots=u_k$ and $v=v_1=v_2=\cdots=v_k$, that contains graphs $G_i$ as subgraphs, and such that vertices $u_1,u_2,\cdots,u_k$ (vertices $v_1,v_2,\cdots,v_k$) have been identified to be the same vertex. A \emph{maximal series-parallel graph} is such that all its series compositions construct a graph out of exactly two smaller series-parallel graphs $G_1$ and $G_2$, and such that all its parallel compositions have a component which is the edge between the two poles. Every series-parallel graph can be augmented to maximal by adding dummy edges to it. The \emph{fan-out} of a series-parallel graph is the maximum number of components in a parallel composition.
A graph $G$ is \emph{bipartite} if its vertex set $V$ can be partitioned into two subsets $V_1$ and $V_2$ so that every edge of $G$ is incident to a vertex of $V_1$ and to a vertex of $V_2$. A \emph{bipartite planar graph} is both bipartite and planar. A \emph{maximal bipartite planar graph} admits a planar embedding in which all its faces have exactly four incident vertices. Every bipartite planar graph with at least four vertices can be augmented to maximal by adding dummy edges to it.
\subsection*{Drawing Standards}
A {\em straight-line drawing} is a drawing such that each edge is represented by a straight-line segment. A {\em poly-line drawing} is a drawing such that each edge is represented by a sequence of consecutive segments. The points in which two consecutive segments of the same edge touch are called \emph{bends}. A {\em grid drawing} is a drawing such that vertices and bends have integer coordinates. An {\em orthogonal drawing} is a poly-line drawing such that each edge is represented by a sequence of horizontal and vertical segments. A {\em convex drawing} (resp. {\em strictly-convex drawing}) is a planar drawing such that each face is delimited by a convex polygon (resp. strictly-convex polygon), that is, every interior angle of the drawing is at most $180^{\circ}$ (resp. less than $180^{\circ}$) and every exterior angle is at least $180^{\circ}$ (resp. more than $180^{\circ}$). An \emph{order-preserving drawing} is a drawing such that the order of the edges incident to each vertex respects an order fixed in advance. An \emph{upward drawing} (resp. \emph{strictly-upward drawing}) of a rooted tree is a drawing such that each edge is represented by a non-decreasing curve (resp. increasing curve). A \emph{visibility representation} is a drawing such that each vertex is represented by a horizontal segment $\sigma(u)$, each edge $(u,v)$ is represented by a vertical segment connecting a point of $\sigma(u)$ with a point of $\sigma(v)$, and no two segments cross, except if they represent a vertex and one of its incident edges.
\subsection*{Area of a Drawing}
The \emph{bounding box} of a drawing is the smallest rectangle with sides parallel to the axes that contains the drawing completely. The \emph{height} and \emph{width} of a drawing are the height and width of its bounding box. The \emph{area} of a drawing is the area of its bounding box. The \emph{aspect ratio} of a drawing is the ratio between the maximum and the minimum of the height and width of the drawing. Observe that the concept of area of a drawing only makes sense once a \emph{resolution rule} is fixed, i.e., a rule that does not allow vertices to be arbitrarily close (\emph{vertex resolution rule}), or edges to be arbitrarily short (\emph{edge resolution rule}). Without any of such rules, one could just construct drawings with arbitrarily small area. It is usually assumed in the literature that graph drawings in small area have to be constructed on a grid. In fact all the algorithms we will present in Sects.~\ref{se:straight-line},~\ref{se:poly-line},~\ref{se:upward},~\ref{se:convex}, and~\ref{se:clustered} assign integer coordinates to vertices. The assumption of constructing drawings on the grid is usually relaxed in the context of proximity drawings (hence in Sect.~\ref{se:proximity}), where in fact it is assumed that no two vertices have distance less than one unit.
\subsection*{Directed Graphs and Planar Upward Drawings}
A \emph{directed acyclic graph} (\emph{DAG} for short) is a graph whose edges are oriented and containing no cycle $(v_1,\dots,v_n)$ such that edge $(v_i,v_{i+1})$ is directed from $v_i$ to $v_{i+1}$, for $i=1,\dots,n-1$, and edge $(v_n,v_1)$ is directed from $v_n$ to $v_1$. The \emph{underlying graph} of a DAG $G$ is the undirected graph obtained from $G$ by removing the directions on its edges. An \emph{upward drawing} of a DAG is such that each edge is represented by an increasing curve. An \emph{upward planar drawing} is a drawing which is both upward and planar. An \emph{upward planar DAG} is a DAG that admits an upward planar drawing. In a directed graph, the \emph{outdegree} of a vertex is the number of edges leaving the vertex and the \emph{indegree} of a vertex is the number of edges entering the vertex. A \emph{source} (resp. \emph{sink}) is a vertex with indegree zero (resp. with outdegree zero). An \emph{st-planar DAG} is a DAG with exactly one source $s$ and one sink $t$ that admits an upward planar embedding in which $s$ and $t$ are on the outer face. \emph{Bipartite DAGs} and \emph{directed trees} are DAGs whose underlying graphs are bipartite graphs and trees, respectively. A \emph{series-parallel DAG} is a DAG that can be inductively constructed as follows. An edge $(u,v)$ directed from $u$ to $v$ is a series-parallel DAG with \emph{starting pole} $u$ and \emph{ending pole} $v$. Denote by $u_i$ and $v_i$ the starting and ending poles of a series-parallel DAG $G_i$, respectively. Then, a \emph{series composition} of a sequence $G_1,G_2,\dots,G_k$ of series-parallel DAGs, with $k\geq 2$, constructs a series-parallel DAG that has starting pole $u=u_1$, that has ending pole $v=v_k$, that contains DAGs $G_i$ as subgraphs, and such that vertices $v_i$ and $u_{i+1}$ have been identified to be the same vertex, for each $i=1,2,\dots,k-1$. A \emph{parallel composition} of a set $G_1,G_2,\dots,G_k$ of series-parallel DAGs, with $k\geq 2$, constructs a series-parallel DAG that has starting pole $u=u_1=u_2=\cdots=u_k$, that has ending pole $v=v_1=v_2=\cdots=v_k$, that contains DAGs $G_i$ as subgraphs, and such that vertices $u_1,u_2,\cdots,u_k$ (vertices $v_1,v_2,\cdots,v_k$) have been identified to be the same vertex. We remark that series-parallel DAGs are a subclass of the upward planar DAGs whose underlying graph is a series-parallel graph.
\subsection*{Proximity Drawings}
A \emph{Delaunay drawing} of a graph $G$ is a straight-line drawing such that no three vertices are on the same line, no four vertices are on the same circle, and three vertices $u$, $v$, and $z$ form a $3$-cycle $(u,v,z)$ in $G$ if and only if the circle passing through $u$, $v$, and $z$ in the drawing contains no vertex other than $u$, $v$, and $z$. A \emph{Delaunay triangulation} is a graph that admits a Delaunay drawing.
The \emph{Gabriel region} of two vertices $x$ and $y$ is the disk having segment $\overline{xy}$ as diameter. A \emph{Gabriel drawing} of a graph $G$ is a straight-line drawing of $G$ having the property that two vertices $x$ and $y$ of the drawing are connected by an edge if and only if the Gabriel region of $x$ and $y$ does not contain any other vertex. A \emph{Gabriel graph} is a graph admitting a Gabriel drawing.
A \emph{relative neighborhood drawing} of a graph $G$ is a straight-line drawing such that two vertices $x$ and $y$ are adjacent if and only if there is no vertex whose distance to both $x$ and $y$ is less than the distance between $x$ and $y$. A \emph{relative neighborhood graph} is a graph admitting a relative neighborhood drawing.
A \emph{nearest neighbor drawing} of a graph $G$ is a straight-line drawing of $G$ such that each vertex has a unique closest vertex and such that two vertices $x$ and $y$ of the drawing are connected by an edge if and only if $x$ is the vertex of $G$ closest to $y$ or viceversa. A \emph{nearest neighbor graph} is a graph admitting a nearest neighbor drawing.
A \emph{$\beta$-drawing} is a straight-line drawing of $G$ having the property that two vertices $x$ and $y$ of the drawing are connected by an edge if and only if the $\beta$-region of $x$ and $y$ does not contain any other vertex. The \emph{$\beta$-region} of $x$ and $y$ is the line segment $\overline{xy}$ if $\beta=0$, it is the intersection of the two closed disks of radius $d(x,y)/(2\beta)$ passing through both $x$ and $y$ if $0<\beta <1$, it is the intersection of the two closed disks of radius $d(x,y)/(2\beta)$ that are centered on the line through $x$ and $y$ and that respectively pass through $x$ and through $y$ if $1\leq \beta <\infty$, and it is the closed infinite strip perpendicular to the line segment $\overline{xy}$ if $\beta =\infty$.
\emph{Weak proximity drawings} are such that there is no geometric requirement on the pairs of vertices not connected by an edge. For example, a \emph{weak Gabriel drawing} of a graph $G$ is a straight-line drawing of $G$ having the property that if two vertices $x$ and $y$ of the drawing are connected by an edge then the Gabriel region of $x$ and $y$ does not contain any other vertex, while there might exist two vertices whose Gabriel region is empty and that are not connected by an edge.
A \emph{Euclidean minimum spanning tree} $T$ of a set $P$ of points is a tree spanning the points in $P$ (that is, the nodes of $T$ coincide with the points of $P$ and no ``Steiner points'' are allowed) and having minimum total edge length.
A \emph{greedy drawing} of a graph $G$ is a straight-line drawing of $G$ such that, for every pair of nodes $u$ and $v$, there exists a \emph{distance-decreasing path}, where a path $(v_0,v_1,\ldots,v_m)$ is distance-decreasing if $d(v_{i},v_{m})<d(v_{i-1},v_{m})$, for $i=1,\ldots,m$, where $d(p,q)$ denotes the Euclidean distance between two points $p$ and $q$.
For more about proximity drawings, see Chapter 7 in~\cite{gd-handbook}.
\subsection*{Clustered Graphs and $c$-Planar Drawings}
A \emph{clustered graph} is a pair $C(G,T)$, where $G$ is a graph, called \emph{underlying graph}, and $T$ is a rooted tree, called \emph{inclusion tree}, such that the leaves of $T$ are the vertices of $G$. Each internal node $\nu$ of $T$ corresponds to the subset of vertices of $G$, called \emph{cluster}, that are the leaves of the subtree of $T$ rooted at $\nu$. A clustered graph $C(G,T)$ is \emph{$c$-connected} if each cluster induces a connected subgraph of $G$, it is \emph{non-$c$-connected} otherwise.
A \emph{drawing} $\Gamma$ of a clustered graph $C(G,T)$ consists of a drawing of $G$ (each vertex is a point in the plane and each edge is as Jordan curve between its endvertices) and of a representation of each node $\mu$ of $T$ as a simple closed region containing all and only the vertices that belong to $\mu$. A drawing is \emph{$c$-planar} if it has no edge crossings (i.e., the drawing of the underlying graph is planar), no edge-region crossings (i.e., an edge intersects the boundary of a cluster at most once), and no region-region crossings (i.e., no two cluster boundaries cross).
A \emph{$c$-planar embedding} is an equivalence class of $c$-planar drawings of $C$, where two $c$-planar drawings are equivalent if they have the same order of the edges incident to each vertex and the same order of the edges incident to each cluster.
\section{Straight-line Drawings}\label{se:straight-line}
In this section, we discuss algorithms and bounds for constructing small-area planar straight-line drawings
of planar graphs and their subclasses.
In Sect.~\ref{se:straight-planar} we deal with general planar graphs,
in Sect.~\ref{se:straight-bipartite} we deal with $4$-connected and bipartite graphs,
in Sect.~\ref{se:straight-series-parallel} we deal with series-parallel graphs,
in Sect.~\ref{se:straight-outerplanar} we deal with outerplanar graphs,
and in Sect.~\ref{se:straight-trees} we deal with trees.
Table~\ref{ta:straight-line} summarizes the best known area bounds for straight-line planar drawings of planar graphs and their subclasses. Observe that the lower bounds of the table that refer to general planar graphs, $4$-connected planar graphs, and bipartite planar graphs hold true for {\em plane} graphs.
\begin{table}[!htb]\footnotesize
\centering
\linespread{1.2}
\selectfont
\begin{tabular}{|c|c|c|c|c|}
\cline{2-5}
\multicolumn{1}{c|}{} & \emph{Upper Bound} & \emph{Refs.} & \emph{Lower Bound} & \emph{Refs.} \\
\hline
{\em General Planar Graphs} & $\frac{8n^2}{9}+O(n)$ & \cite{fpp-hdpgg-90,s-epgg-90,b-dpg89a-08} & $\frac{4n^2}{9}-O(n)$ & \cite{Val81,fpp-hdpgg-90,FratiP07,mnra-madp3t-10}\\
\hline
{\em $4$-Connected Planar Graphs} & $\lfloor\frac{n}{2}\rfloor \times (\lceil\frac{n}{2}\rceil-1)$ & \cite{mnn-gd4pg-01} & $\lfloor\frac{n}{2}\rfloor \times (\lceil\frac{n}{2}\rceil-1)$ & \cite{mnn-gd4pg-01}\\
\hline
{\em Bipartite Planar Graphs} & $\lfloor\frac{n}{2}\rfloor \times (\lceil\frac{n}{2}\rceil-1)$ & \cite{bb-dpbgsa-05} & $\lfloor\frac{n}{2}\rfloor \times (\lceil\frac{n}{2}\rceil-1)$ & \cite{bb-dpbgsa-05}\\
\hline
{\em Series-Parallel Graphs} & $O(n^2)$ & \cite{fpp-hdpgg-90,s-epgg-90,zhn-sgdpgbb-10} & $\Omega(n 2^{\sqrt{\log n}})$ & \cite{f-lbarspg-j10}\\
\hline
{\em Outerplanar Graphs} & $O(n^{1.48})$ & \cite{df-sadog-j09} & $\Omega(n)$ & \emph{trivial}\\
\hline
{\em Trees} & $O(n \log n)$ & \cite{cdp-noad-92} & $\Omega(n)$ & \emph{trivial}\\
\hline
\end{tabular}
\vspace{2mm}
\caption{\small A table summarizing the area requirements for straight-line planar drawings of several classes of planar graphs. Notice that $4$-connected planar graphs have been studied only with the additional constraint of having at least four vertices on the outer face.}
\label{ta:straight-line}
\end{table}
\subsection{General Planar Graphs} \label{se:straight-planar}
In this section, we discuss algorithms and bounds for constructing small-area planar straight-line drawings of general planar graphs. Observe that, in order to derive bounds on the area requirements of general planar graphs, it suffices to restrict the attention to maximal planar graphs, as every planar graph can be augmented to maximal by the insertion of ``dummy'' edges. Moreover, such an augmentation can be performed in linear time~\cite{r-nmdg-87}.
We start by proving that every plane graph admits a planar straight-line drawing~\cite{w-bzv-36,s-cm-51}. The simplest and most elegant proof of such a statement is, in our opinion, the one presented by F\'ary in 1948~\cite{f-srpg-48}.
F\'ary's algorithm works by induction on the number $n$ of vertices of the plane graph $G$; namely, the algorithm inductively assumes that a straight-line planar drawing of $G$ can be constructed with the further constraint that the outer face $f(G)$ is drawn as an arbitrary triangle $\Delta$. The inductive hypothesis is trivially satisfied when $n=3$. If $n>3$, then two cases are possible. In the first case $G$ contains a separating $3$-cycle $c$. Then let $G_1$ (resp. $G_2$) be the graph obtained from $G$ by removing all the vertices internal to $c$ (resp. external to $c$). Both $G_1$ and $G_2$ have less than $n$ vertices, hence the inductive hypothesis applies first to construct a straight-line planar drawing $\Gamma_1$ of $G_1$ in which $f(G_1)$ is drawn as an arbitrary triangle $\Delta$, and second to construct a straight-line planar drawing $\Gamma_2$ of $G_2$ in which $f(G_2)$ is drawn as $\Delta(c)$, where $\Delta(c)$ is the triangle representing $c$ in $\Gamma_1$ (see Fig.~\ref{fig:fary1}(a)). Thus, a straight-line drawing $\Gamma$ of $G$ in which $f(G)$ is represented by $\Delta$ is obtained.
\begin{figure}[htb]
\centering
\begin{tabular}{c c c}
\mbox{\epsfig{figure=Fary2.eps,scale=0.2,clip=}} \hspace{5mm} &
\mbox{\epsfig{figure=Fary3.eps,scale=0.27,clip=}} \hspace{5mm} &
\mbox{\epsfig{figure=Fary4.eps,scale=0.27,clip=}} \\
(a) \hspace{5mm} & (b) \hspace{5mm} & (c)\\
\end{tabular}
\caption{(a) Induction in F\'ary's algorithm if $G$ contains a separating $3$-cycle. (b)--(c) Induction in F\'ary's algorithm if $G$ contains no separating $3$-cycle.}
\label{fig:fary1}
\end{figure}
In the second case, $G$ does not contain any separating $3$-cycle, i.e. $G$ is $4$-connected. Then, consider any internal vertex $u$ of $G$ and consider any neighbor $v$ of $u$. Construct an $(n-1)-$vertex plane graph $G'$ by removing $u$ and all its incident edges from $G$, and by inserting ``dummy'' edges between $v$ and all the neighbors of $u$ in $G$, except for the two vertices $v_1$ and $v_2$ forming faces with $u$ and $v$. The graph $G'$ is simple, as $G$ contains no separating $3$-cycle. Hence, the inductive hypothesis applies to construct a straight-line planar drawing $\Gamma'$ of $G'$ in which $f(G')$ is drawn as $\Delta$. Further, dummy edges can be removed and vertex $u$ can be introduced in $\Gamma'$ together with its incident edges, without altering the planarity of $\Gamma'$. In fact, $u$ can be placed at a suitable point in the interior of a small disk centered at $v$, thus obtaining a straight-line drawing $\Gamma$ of $G$ in which $f(G)$ is represented by $\Delta$ (see Figs.~\ref{fig:fary1}(b)--(c)).
The first algorithms for constructing planar straight-line grid drawings of planar graphs in polynomial area were presented (fifty years later than F\'ary's algorithm!) by de Fraysseix, Pach, and Pollack~\cite{fpp-sssfepg-88,fpp-hdpgg-90} and, simultaneously and independently, by Schnyder~\cite{s-epgg-90}. The approaches of the two algorithms, that we sketch below, are still today the base of every known algorithm to construct planar straight-line grid drawings of triangulations.
The algorithm by de Fraysseix \emph{et al.}~\cite{fpp-sssfepg-88,fpp-hdpgg-90} relies on two main ideas.
First, any $n$-vertex maximal plane graph $G$ admits a total ordering $\sigma$ of its vertices, called \emph{canonical ordering}, such that (see Fig.~\ref{fig:canonical1}(a)): (i) the subgraph $G_k$ of $G$ induced by the first $k$ vertices in $\sigma$ is biconnected, for each $k=3,\dots,n$; and (ii) the $k$-th vertex in $\sigma$ lies in the outer face of $G_{k-1}$, for each $k=4,\dots,n$.
Second, a straight-line drawing of an $n$-vertex maximal plane graph $G$ can be constructed starting from a drawing of the $3$-cycle induced by the first three vertices in a canonical ordering $\sigma$ of $G$ and incrementally adding vertices to the partially constructed drawing in the order defined by $\sigma$.
To construct the drawing of $G$ one vertex at a time, the algorithm maintains the invariant that the outer face of $G_k$ is delimited by a polygon composed of a sequence of segments having slopes equal to either $45^{\degree}$ or $-45^{\degree}$. When the next vertex $v_{k+1}$ in $\sigma$ is added to the drawing of $G_k$ to construct a drawing of $G_{k+1}$, a subset of the vertices of $G_k$ undergoes a horizontal shift that allows for $v_{k+1}$ to be introduced in the drawing still maintaining the invariant that the outer face of $G_{k+1}$ is delimited by a polygon composed of a sequence of segments having slopes equal to either $45^{\degree}$ or $-45^{\degree}$ (see Fig.~\ref{fig:canonical1}(b)--(c)).
The area of the constructed drawings is $(2n-4) \times (n-2)$. The described algorithm has been proposed by de Fraysseix~\emph{et~al.} together with an $O(n \log n)$-time implementation. The authors conjectured that its complexity could be improved to $O(n)$. This bound was in fact achieved a few years later by Chrobak and Payne in~\cite{chrobak95lineartime}.
\begin{figure}[htb]
\centering
\begin{tabular}{c c c}
\mbox{\epsfig{figure=deFraysseix1.eps,scale=0.28,clip=}} \hspace{3mm} &
\mbox{\epsfig{figure=deFraysseix2.eps,scale=0.35,clip=}} \hspace{3mm} &
\mbox{\epsfig{figure=deFraysseix3.eps,scale=0.35,clip=}} \\
(a) \hspace{3mm} & (b) \hspace{3mm} & (c)\\
\end{tabular}
\caption{(a) A canonical ordering of a maximal plane graph $G$. (b) The drawing of $G_k$ constructed by the algorithm of de Fraysseix~\emph{et~al.} (c) The drawing of $G_{k+1}$ constructed by the algorithm of de Fraysseix~\emph{et~al.}}
\label{fig:canonical1}
\end{figure}
The ideas behind the algorithm by Schnyder~\cite{s-epgg-90} are totally different from the ones of de Fraysseix~\emph{et~al.} In fact, Schnyder's algorithm constructs the drawing by determining the coordinates of all the vertices in one shot. The algorithm relies on results concerning planar graph embeddings that are indeed less intuitive than the canonical ordering of a plane graph used by de Fraysseix~\emph{et~al.}
First, Schnyder introduces the concept of \emph{barycentric representation} of a graph $G$ as an injective function $v\in V(G)\rightarrow (x(v),y(v),z(v))$ such that $x(v)+y(v)+z(v)=1$, for all vertices $v\in V(G)$, and such that, for each edge $(u,v)\in E(G)$ and each vertex $w\notin \{u,v\}$, $x(u)<x(w)$ and $x(v)<x(w)$ hold, or $y(u)<y(w)$ and $y(v)<y(w)$ hold, or $z(u)<z(w)$ and $z(v)<z(w)$ hold. Schnyder proves that, given any graph $G$, given any barycentric representation $v\rightarrow (x(v),y(v),z(v))$ of $G$, and given any three non-collinear points $\alpha$, $\beta$, and $\gamma$ in the three-dimensional space, the mapping $f:v\in V(G)\rightarrow v_1 \alpha +v_2 \beta + v_3 \gamma$ is a straight-line planar embedding of $G$ in the plane spanned by $\alpha$, $\beta$, and $\gamma$.
Second, Schnyder introduces the concept of a \emph{realizer} of $G$ as an orientation and a partition of the interior edges of a plane graph $G$ into three sets $T_1$, $T_2$, and $T_3$, such that: (i) the set of edges in $T_i$, for each $i=1,2,3$, is a tree spanning all the internal vertices of $G$ and exactly one external vertex; (ii) all the edges of $T_i$ are directed towards this external vertex, which is the root of $T_i$; (iii) the external vertices belonging to $T_1$, to $T_2$, and to $T_3$ are distinct and appear in counter-clockwise order on the border of the outer face of $G$; and (iv) the counter-clockwise order of the edges incident to $v$ is: Leaving $T_1$, entering $T_3$, leaving $T_2$, entering $T_1$, leaving $T_3$, and entering~$T_2$. Fig.~\ref{fig:schnyder}(a) illustrates a realizer for a plane graph~$G$. Trees $T_1$, $T_2$, and $T_3$ are sometimes called \emph{Schnyder woods}.
Third, Schnyder describes how to get a barycentric representation of a plane graph $G$ starting from a realizer of $G$; this is essentially done by looking, for each vertex $v\in V(G)$ at the paths $P_i(v)$, that are the only paths composed entirely of edges of $T_i$ connecting $v$ to the root of $T_i$ (see Fig.~\ref{fig:schnyder}(b)), and counting the number of the faces or the number of the vertices in the regions $R_1(v)$, $R_2(v)$, and $R_3(v)$ that are defined by $P_1(v)$, $P_2(v)$, and $P_3(v)$. The area of the constructed drawings is $(n-2)\times (n-2)$.
\begin{figure}[htb]
\centering
\begin{tabular}{c c}
\mbox{\epsfig{figure=Schnyder1.eps,scale=0.25,clip=}} \hspace{5mm} &
\mbox{\epsfig{figure=Schnyder2.eps,scale=0.25,clip=}} \\
(a) \hspace{5mm} & (b) \\
\end{tabular}
\caption{(a) A realizer for a plane graph $G$. (b) Paths $P_1(v)$, $P_2(v)$, and $P_3(v)$ (represented by green, red, and blue edges, respectively) and regions $R_1(v)$, $R_2(v)$, and $R_3(v)$ (delimited by $P_1(v)$, $P_2(v)$, and $P_3(v)$, and by the edges incident to the outer face of $G$).}
\label{fig:schnyder}
\end{figure}
Schnyder's upper bound has been unbeaten for almost twenty years. Only recently Brandenburg~\cite{b-dpg89a-08} proposed an algorithm for constructing planar straight-line drawings of triangulations in $\frac{8n^2}{9} + O(n)$ area. Such an algorithm is based on a geometric refinement of the de Fraysseix \emph{et al.}~\cite{fpp-sssfepg-88,fpp-hdpgg-90} algorithm combined with some topological properties of planar triangulations due to Bonichon et al.~\cite{bsm-wtr-02}, that will be discussed in Sect.~\ref{se:poly-line}.
A quadratic area upper bound for straight-line planar drawings of plane graphs is asymptotically optimal. In fact, almost ten years before the publication of such algorithms, Valiant observed in~\cite{Val81} that there exist $n$-vertex plane graphs (see Fig.~\ref{fig:lowerboundnested}(a)) requiring $\Omega(n^2)$ area in any straight-line planar drawing (in fact, in every poly-line planar drawing). It was then proved by de Fraysseix \emph{et al.}~in~\cite{fpp-hdpgg-90} that \emph{nested triangles graphs} (see Fig.~\ref{fig:lowerboundnested}(b)) require $\left(\frac{2n}{3}-1\right) \times\left(\frac{2n}{3}-1\right)$ area in any straight-line planar drawing (in fact, in every poly-line planar drawing). Such a lower bound was only recently improved to $\frac{4n^2}{9}-\frac{2n}{3}$ by Frati and Patrignani~\cite{FratiP07}, for all $n$ multiple of $3$ (see Fig.~\ref{fig:lowerboundnested}(c)), and then by Mondal \emph{et al.}~\cite{mnra-madp3t-10} to $\left \lfloor \frac{2n}{3}-1 \right\rfloor \times \left \lfloor \frac{2n}{3} \right\rfloor $, for all $n\geq 6$.
\begin{figure}[htb]
\centering
\begin{tabular}{c c c}
\mbox{\epsfig{figure=LowerBoundValiant.eps,scale=0.3,clip=}} \hspace{5mm} &
\mbox{\epsfig{figure=NestedTrianglesGraph.eps,scale=0.3,clip=}} \hspace{5mm} &
\mbox{\epsfig{figure=LowerBoundG6big.eps,scale=0.3,clip=}}\\
(a) \hspace{5mm} & (b) \hspace{5mm} & (c) \\
\end{tabular}
\caption{(a) A graph~\cite{Val81} requiring quadratic area in any straight-line and poly-line drawing. (b) A graph~\cite{fpp-hdpgg-90} requiring $\left(\frac{2n}{3}-1\right) \times\left(\frac{2n}{3}-1\right)$ area in any straight-line and poly-line drawing. (c) A graph~\cite{FratiP07} requiring $\frac{4n^2}{9}-\frac{2n}{3}$ area in any straight-line drawing.}
\label{fig:lowerboundnested}
\end{figure}
However, the following remains open:
\begin{problem}
Close the gap between the $\frac{8n^2}{9} + O(n)$ upper bound and the $\frac{4n^2}{9} - O(n)$ lower bound for the area requirements of straight-line drawings of plane graphs.
\end{problem}
\subsection{$4$-Connected and Bipartite Planar Graphs} \label{se:straight-bipartite}
In this section, we discuss algorithms and bounds for constructing planar straight-line drawings of $4$-connected and bipartite planar graphs. Such different families of graphs are discussed in the same section since the best known upper bound for the area requirements of bipartite planar graphs uses a preliminary augmentation to $4$-connected planar graphs.
Concerning $4$-connected plane graphs, tight bounds are known for the area requirements of planar straight-line drawings if the graph has at least four vertices incident to the outer face. Namely, Miura \emph{et al.}~proved in~\cite{mnn-gd4pg-01} that every such a graph has a planar straight-line drawing in $(\lceil{\frac{n}{2}}\rceil-1) \times (\lfloor{\frac{n}{2}}\rfloor)$ area, improving upon previous results of He~\cite{h-gefcpg-97}. The authors show that this bound is tight, by exhibiting a class of $4$-connected plane graphs with four vertices incident to the outer face requiring $(\lceil{\frac{n}{2}}\rceil-1) \times (\lfloor{\frac{n}{2}}\rfloor)$ area (see Fig.~\ref{fig:straight-line-fourconnected}(a)).
\begin{figure}[htb]
\centering
\begin{tabular}{c c}
\mbox{\epsfig{figure=Four-Connected-LB.eps,scale=0.4,clip=}} \hspace{5mm} &
\mbox{\epsfig{figure=Bipartite-LB.eps,scale=0.4,clip=}} \\
(a) \hspace{5mm} & (b) \\
\end{tabular}
\caption{(a) A $4$-connected plane graph requiring $(\lceil{\frac{n}{2}}\rceil-1) \times (\lfloor{\frac{n}{2}}\rfloor)$ area in any straight-line planar drawing. (b) A bipartite plane graph requiring $(\lceil{\frac{n}{2}}\rceil-1) \times (\lfloor{\frac{n}{2}}\rfloor)$ area in any straight-line planar drawing.}
\label{fig:straight-line-fourconnected}
\end{figure}
The algorithm of Miura \emph{et al.}~divides the input $4$-connected plane graph $G$ into two graphs $G'$ and $G''$ with the same number of vertices. This is done by performing a \emph{4-canonical ordering} of $G$ (see~\cite{kh-rel4cpgiagdp-97}). The graph $G'$ ($G''$, respectively) is then drawn inside an isosceles right triangle $\Delta'$ (resp.\ $\Delta''$) whose width is $\frac{n}{2}-1$ and whose height is half of its width. To construct such drawings of $G'$ and $G''$, Miura \emph{et al.}~design an algorithm that is similar to the algorithm by de Fraysseix \emph{et al.}~\cite{fpp-hdpgg-90}. In the drawings produced by their algorithm the slopes of the edges incident to the outer faces of $G'$ and $G''$ have absolute value which is at most $45\degree$. The drawing of $G''$ is then rotated by $180\degree$ and placed on top of the drawing of $G'$. This allows for drawing the edges connecting $G'$ with $G''$ without creating crossings. Fig.~\ref{fig:Miura} depicts the construction of the Miura \emph{et al.}'s algorithm.
As far as we know, no bound better than the one for general plane graphs is known for $4$-connected plane graphs (possibly having three vertices incident to the outer face), hence the following is open:
\begin{problem}
Close the gap between the $\frac{8n^2}{9} + O(n)$ upper bound and the $\frac{n^2}{4} - O(n)$ lower bound for the area requirements of straight-line drawings of $4$-connected plane graphs.
\end{problem}
\begin{figure}[htb]
\centering{
\mbox{\epsfig{figure=Miura-algorithm.eps,scale=0.4,clip=}}}
\caption{The algorithm by Miura \emph{et al.}~to construct straight-line drawings of $4$-connected plane graphs~\cite{mnn-gd4pg-01}.}
\label{fig:Miura}
\end{figure}
Biedl and Brandenburg~\cite{bb-dpbgsa-05} show how to construct planar straight-line drawings of bipartite planar graphs in $(\lceil{\frac{n}{2}}\rceil-1) \times (\lfloor{\frac{n}{2}}\rfloor)$ area. To achieve such a bound, they exploit a result of Biedl \emph{et al.}~\cite{bkk-tpgfcc-98} stating that all planar graphs without separating triangles, except those ``containing a star'' (see~\cite{bb-dpbgsa-05} and observe that in this case a star is not just a vertex plus some incident edges), can be augmented to $4$-connected by the insertion of dummy edges; once such an augmentation is done, Biedl and Brandenburg use the algorithm of Miura \emph{et al.}~\cite{mnn-gd4pg-01} to draw the resulting $4$-connected plane graph. In order to be able to use Miura \emph{et al.}'s algorithm, Biedl and Brandenburg prove that no bipartite plane graph ``contains a star'' and that Miura \emph{et al.}'s algorithm works more in general for plane graphs that become $4$-connected if an edge is added to them. The upper bound of Biedl and Brandenburg is tight as the authors show a bipartite plane graph requiring $(\lceil{\frac{n}{2}}\rceil-1) \times (\lfloor{\frac{n}{2}}\rfloor)$ area in any straight-line planar drawing (see Fig.~\ref{fig:straight-line-fourconnected}(b)).
\subsection{Series-Parallel Graphs} \label{se:straight-series-parallel}
In this section, we discuss algorithms and bounds for constructing small-area planar straight-line drawings of series-parallel graphs.
No sub-quadratic area upper bound is known for constructing small-area planar straight-line drawings of series-parallel graphs. The best known quadratic upper bound for straight-line drawings is provided in~\cite{zhn-sgdpgbb-10}.
In~\cite{f-lbarspg-j10} Frati proved that there exist series-parallel graphs requiring $\Omega(n 2^{\sqrt{\log n}})$ area in any straight-line or poly-line grid drawing. Such a result is achieved in two steps. In the first one, an $\Omega(n)$ lower bound for the maximum between the height and the width of any straight-line or poly-line grid drawing of $K_{2,n}$ is proved, thus answering a question of Felsner \emph{et~al.}~\cite{journals/jgaa/FelsnerLW03} and improving upon previous results of Biedl \emph{et~al.}~\cite{journals/ipl/BiedlCL03}. In the second one, an $\Omega(2^{\sqrt{\log n}})$ lower bound for the minimum between the height and the width of any straight-line or poly-line grid drawing of certain series-parallel graphs is proved.
The proof that $K_{2,n}$ requires $\Omega(n)$ height or width in any straight-line or poly-line drawing has several ingredients. First, a simple ``optimal'' drawing algorithm for $K_{2,n}$ is exhibited, that is, an algorithm is presented that computes a drawing of $K_{2,n}$ inside an arbitrary convex polygon if such a drawing exists. Second, the drawings constructed by the mentioned algorithm inside a rectangle are studied. Such a study reveals that the slopes of the segments representing the edges of $K_{2,n}$ have a strong relationship with the relatively prime numbers as ordered in the \emph{Stern-Brocot} tree (see~\cite{s-uzf-58,b-cranm-60} and Fig.~\ref{fig:sternbrocot}). Such a relationship leads to derive some arithmetical properties of the lines passing through infinite grid points in the plane and to achieve the $\Omega(n)$ lower bound.
\begin{figure}[htb]
\centering{
\mbox{\epsfig{figure=SternBrocot.eps,scale =0.5,clip=}}}
\caption{The Stern-Brocot tree is a tree containing all the pairs of relatively prime numbers.}
\label{fig:sternbrocot}
\end{figure}
The results on the area requirements of $K_{2,n}$ are then used to construct series-parallel graphs (shown in Fig.~\ref{fig:straight-line-sp-improvedgraphs}) out of several copies of $K_{2,2^{\sqrt{\log n}}}$ and to prove that such a graph requires $\Omega(2^{\sqrt{\log n}})$ height and width in any straight-line or poly-line grid drawing.
\begin{figure}[htb]
\centering{
\begin{tabular} {c c c}
\mbox{\epsfig{figure=G1.eps,scale =0.5,clip=}} \hspace{3mm} &
\mbox{\epsfig{figure=G2.eps,scale =0.5,clip=}} \hspace{3mm} &
\mbox{\epsfig{figure=G3.eps,scale =0.5,clip=}} \\
(a) \hspace{3mm} & (b) \hspace{3mm} & (c)
\end{tabular}}
\caption{The inductive construction of series-parallel graphs requiring $\Omega(2^{\sqrt{\log n}})$ height and width in any straight-line or poly-line grid drawing.}
\label{fig:straight-line-sp-improvedgraphs}
\end{figure}
As no sub-quadratic area upper bound is known for straight-line planar drawings of series-parallel graphs the following is open.
\begin{problem}
Close the gap between the $O(n^2)$ upper bound and the $\Omega(n 2^{\sqrt{\log n}})$ lower bound for the area requirements of straight-line drawings of series-parallel graphs.
\end{problem}
Related to the above problem, Wood~\cite{w-08} conjectures the following: Let $p_1,\dots,p_k$ be positive integers. Let $G(p_1)$ be the graph obtained from $K_3$ by adding $p_1$ new vertices adjacent to $v$ and $w$ for each edge $(v,w)$ of $K_3$. For $k \geq 2$, let $G(p_1,p_2,\dots,p_k)$ be the graph obtained from $G(p_1,p_2,\dots,p_{k-1})$ by adding $p_k$ new vertices adjacent to $v$ and $w$ for each edge $(v,w)$ of $G(p_1,p_2,\dots,p_{k-1})$. Observe that $G(p_1,p_2,\dots,p_k)$ is a series-parallel graph.
\begin{conjecture} (D. R. Wood) Every straight-line grid drawing of $G(p_1,p_2,\dots,p_k)$ requires $\Omega(n^2)$ area for some choice of $k$ and $p_1,p_2,\dots,p_k$.
\end{conjecture}
\subsection{Outerplanar Graphs} \label{se:straight-outerplanar}
In this section, we discuss algorithms and bounds for constructing small-area planar straight-line drawings of outerplanar graphs.
The first non-trivial bound appeared in~\cite{gr-aepsdog-07}, where Garg and Rusu proved that every outerplanar graph with maximum degree $d$ has a straight-line drawing with $O(dn^{1.48})$ area. Such a result is achieved by means of an algorithm that works by induction on the dual tree $T$ of the outerplanar graph $G$. Namely, the algorithm finds a path $P$ in $T$, it removes from $G$ the subgraph $G_P$ that has $P$ as a dual tree, it inductively draws the outerplanar graphs that are disconnected by such a removal, and it puts all the drawings of such outerplanar graphs together with a drawing of $G_P$, obtaining a drawing of the whole outerplanar graph.
The first sub-quadratic area upper bound for straight-line drawings of outerplanar graphs has been proved by Di Battista and Frati in~\cite{df-sadog-j09}. The result in~\cite{df-sadog-j09} uses the following ingredients. First, it is shown that the dual binary tree $T$ of a maximal outerplanar graph $G$ is a subgraph of $G$ itself. Second, a restricted class of straight-line drawings of binary trees, called \emph{star-shaped drawings}, is defined. Star-shaped drawings are straight-line drawings in which special visibility properties among the nodes of the tree are satisfied (see Fig.~\ref{fig:straight-starshaped}).
\begin{figure}[htb]
\centering{
\mbox{\epsfig{figure=Star-shaped.eps,scale =0.5,clip=}}}
\caption{A star-shaped drawing $\Gamma$ of a binary tree $T$ (with thick edges and black vertices). The dashed edges and white vertices augment $\Gamma$ into a straight-line drawing of the outerplanar graph $T$ is dual to.}
\label{fig:straight-starshaped}
\end{figure}
Namely, if a tree $T$ admits a star-shaped drawing $\Gamma$, then the edges that augment $T$ into $G$ can be drawn in $\Gamma$ without creating crossings, thus resulting in a straight-line planar drawing of $G$. Third, an algorithm is shown to construct a star-shaped drawing of any binary tree $T$ in $O(n^{1.48})$ area. Such an algorithm works by induction on the number of nodes of $T$ (Fig.~\ref{fig:straight-line-outerplanar} depicts two inductive cases of such a construction), making use of a strong combinatorial decomposition of ordered binary trees introduced by Chan in~\cite{c-nlabdbt-02} (discussed in Sect.~\ref{se:straight-trees}).
\begin{figure}[htb]
\centering{
\begin{tabular} {c c}
\mbox{\epsfig{figure=OuterplanarComposition.eps,scale =0.4,clip=}} \hspace{4mm} &
\mbox{\epsfig{figure=OuterplanarComposition2.eps,scale =0.4,clip=}}
\end{tabular}}
\caption{Two inductive cases of the algorithm to construct star-shaped drawing of binary trees yielding an $O(n^{1.48})$ upper bound for straight-line drawings of outerplanar graphs. The rectangles and the half-circles represent subtrees recursively drawn by the construction on the right and on the left part of the figure, respectively.}
\label{fig:straight-line-outerplanar}
\end{figure}
Frati used in~\cite{f-sdogdnlogna-07} the same approach of~\cite{df-sadog-j09}, together with a different geometric construction (shown in Fig.~\ref{fig:straight-outerplanar-degreed}), to prove that every outerplanar graph with degree $d$ has a straight-line drawing with $O(dn\log n)$ area.
\begin{figure}[htb]
\centering{
\mbox{\epsfig{figure=OuterplanarComposition3.eps,scale =0.5,clip=}}}
\caption{The inductive construction of star-shaped drawings of binary trees yielding an $O(dn\log n)$ upper bound for straight-line drawings of outerplanar graphs with degree $d$. The rectangles represent recursively constructed star-shaped drawings of subtrees.}
\label{fig:straight-outerplanar-degreed}
\end{figure}
As far as we know, no super-linear area lower bound is known for straight-line drawings of outerplanar graphs. In~\cite{b-dopgia-02} Biedl defined a class of outerplanar graphs, called \emph{snowflake graphs}, and conjectured that such graphs require $\Omega(n \log n)$ area in any straight-line or poly-line drawing. However, Frati disproved such a conjecture in~\cite{f-sdogdnlogna-07} by exhibiting $O(n)$ area straight-line drawings of snowflake graphs. In the same paper, he conjectured that an $O(n\log n)$ area upper bound for straight-line drawings of outerplanar graphs can not be achieved by squeezing the drawing along one coordinate direction, as stated in the following.
\begin{conjecture} (F. Frati) There exist $n$-vertex outerplanar graphs for which, for any straight-line drawing in which the longest side of the bounding-box is $O(n)$, the smallest side of the bounding-box is $\omega(\log n)$.
\end{conjecture}
The following problem remains wide open.
\begin{problem}
Close the gap between the $O(n^{1.48})$ upper bound and the $\Omega(n)$ lower bound for the area requirements of straight-line drawings of outerplanar graphs.
\end{problem}
\subsection{Trees} \label{se:straight-trees}
In this section, we present algorithms and bounds for constructing planar straight-line drawings of trees.
The best bound for constructing general trees is, as far as we know, the $O(n \log n)$ area upper bound provided by a simple modification of the $hv$-drawing algorithm of Crescenzi, Di Battista, and Piperno~\cite{cdp-noad-92}. Such an algorithm proves that a straight-line drawing of any tree $T$ in $O(n) \times O(\log n)$ area can be constructed with the further constraint that the root of $T$ is placed at the bottom-left corner of the bounding box of the drawing. If $T$ has one node, such a drawing is trivially constructed. If $T$ has more than one node, then let $T_1, \dots, T_k$ be the subtrees of $T$, where we assume, w.l.o.g., that $T_k$ is the subtree of $T$ with the greatest number of nodes. Then, the root of $T$ is placed at $(0,0)$, the subtrees $T_1, \dots, T_{k-1}$ are placed one besides the other, with the bottom side of their bounding boxes on the line $y=1$, and $T_{k}$ is placed besides the other subtrees, with the bottom side of its bounding box on the line $y=0$. The width of the drawing is clearly $O(n)$, while its height is $h(n)=\max\{h(n-1),1+h(n/2)\}=O(\log n)$, where $h(n)$ denotes the maximum height of a drawing of an $n$-node tree constructed by the algorithm. See Fig.~\ref{fi:trees-hvdrawing} for an illustration of such an algorithm. Interestingly, no super-linear area lower bound is known for the area requirements of straight-line drawings of trees.
\begin{figure}[htb]
\centering{
\mbox{\epsfig{figure=HV.eps,scale =0.5,clip=}}}
\caption{Inductive construction of a straight-line drawing of a tree in $O(n\log n)$ area.}
\label{fi:trees-hvdrawing}
\end{figure}
For the special case of bounded-degree trees linear area bounds have been achieved. In fact, Garg and Rusu presented an algorithm to construct straight-line drawings of binary trees in $O(n)$ area~\cite{gr-sdbtlaaar-04} and an algorithm to construct straight-line drawings of trees with degree $O(\sqrt n)$ in $O(n)$ area~\cite{iccsa/GargR03}. Both algorithms rely on the existence of simple \emph{separators} for bounded degree trees. Namely, every binary tree $T$ has a \emph{separator edge}, that is an edge whose removal disconnects $T$ into two trees both having at most $2n/3$ vertices~\cite{Val81} and every degree-$d$ tree $T$ has a vertex whose removal disconnects $T$ into at most $d$ trees, each having at most $n/2$ nodes~\cite{iccsa/GargR03}. Such separators are exploited by Garg and Rusu to design quite complex inductive algorithms that achieve linear area bounds and optimal aspect ratio.
The following problem remains open:
\begin{problem}
Close the gap between the $O(n \log n)$ upper bound and the $\Omega(n)$ lower bound for the area requirements of straight-line drawings of trees.
\end{problem}
A lot of attention has been devoted to studying the area requirements of straight-line drawings of trees satisfying additional constraints. Table~\ref{ta:trees-straight-line} summarizes the best known area bounds for various kinds of straight-line drawings of trees.
\begin{table}[!htb]\centering\footnotesize
\linespread{1.2}
\selectfont
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\cline{2-9}
\multicolumn{1}{c|}{} & \emph{Ord. Pres.} & \emph{Upw.} & \emph{Str. Upw.} & \emph{Orth.} & \emph{Upper Bound} & \emph{Refs.} & \emph{Lower Bound} & \emph{Refs.} \\
\cline{2-9}
\hline
\emph{Binary} & & & & & $O(n)$ & \cite{gr-sdbtlaaar-04} & $\Omega(n)$ & \emph{trivial} \\
\hline
\emph{Binary} & \checkmark & & & & $O(n\log \log n)$ & \cite{gr-aeoppsdot-03} & $\Omega(n)$ & \emph{trivial} \\
\hline
\emph{Binary} & & \checkmark & & & $O(n\log \log n)$ & \cite{skc-aeastd-00} & $\Omega(n)$ & \emph{trivial} \\
\hline
\emph{Binary} & & & \checkmark & & $O(n\log n)$ & \cite{cdp-noad-92} & $\Omega(n \log n)$ & \cite{cdp-noad-92} \\
\hline
\emph{Binary} & \checkmark & & \checkmark & & $O(n \log n)$ & \cite{gr-aeoppsdot-03} & $\Omega(n\log n)$ & \cite{cdp-noad-92}\\
\hline
\emph{Binary} & & & & \checkmark & $O(n \log \log n)$ & \cite{cgkt-oaarsod-02,skc-aeastd-00} & $\Omega(n)$ & \emph{trivial}\\
\hline
\emph{Binary} & & \checkmark & & \checkmark & $O(n \log n)$ & \cite{cdp-noad-92,cgkt-oaarsod-02} & $\Omega(n\log n)$ & \cite{cgkt-oaarsod-02}\\
\hline
\emph{Binary} & \checkmark & & & \checkmark & $O(n^{1.5})$ & \cite{f-sodbtt-06} & $\Omega(n)$ & \emph{trivial}\\
\hline
\emph{Ternary} & & & & \checkmark & $O(n^{1.631})$ & \cite{f-sodbtt-06} & $\Omega(n)$ & \emph{trivial}\\
\hline
\emph{Ternary} & \checkmark & & & \checkmark & $O(n^2)$ & \cite{f-sodbtt-06} & $\Omega(n^2)$ & \cite{f-sodbtt-06}\\
\hline
\emph{General} & & & & & $O(n \log n)$ & \cite{cdp-noad-92} & $\Omega(n)$ & \emph{trivial} \\
\hline
\emph{General} & \checkmark & & & & $O(n\log n)$ & \cite{gr-aeoppsdot-03} & $\Omega(n)$ & \emph{trivial} \\
\hline
\emph{General} & & \checkmark & & & $O(n\log n)$ & \cite{cdp-noad-92} & $\Omega(n)$ & \emph{trivial} \\
\hline
\emph{General} & & & \checkmark & & $O(n\log n)$ & \cite{cdp-noad-92} & $\Omega(n \log n)$ & \cite{cdp-noad-92} \\
\hline
\emph{General} & \checkmark & & \checkmark & & $O(n 4^{\sqrt{2\log n}})$ & \cite{c-nlabdbt-02} & $\Omega(n\log n)$ & \cite{cdp-noad-92}\\
\hline
\end{tabular}
\vspace{2mm}
\caption{\small Summary of the best known area bounds for straight-line drawings of trees. ``Ord. Pres.'', ``Upw.'', ``Str. Upw.'', and ``Orth.'' stand for order-preserving, upward, strictly-upward, and orthogonal, respectively.}
\label{ta:trees-straight-line}
\end{table}
Concerning \emph{straight-line upward drawings}, the illustrated algorithm of Crescenzi {\em et al.}~\cite{cdp-noad-92} achieves the best known upper bound of $O(n\log n)$. For trees with constant degree, Shin \emph{et al.}~prove in~\cite{skc-aeastd-00} that upward straight-line drawings in $O(n\log \log n)$ area can be constructed. Their algorithm is based on nice inductive geometric constructions and suitable tree decompositions. No super-linear area lower bound is known, neither for binary nor for general trees, hence the following are open:
\begin{problem}
Close the gap between the $O(n \log n)$ upper bound and the $\Omega(n)$ lower bound for the area requirements of upward straight-line drawings of trees.
\end{problem}
\begin{problem}
Close the gap between the $O(n \log \log n)$ upper bound and the $\Omega(n)$ lower bound for the area requirements of upward straight-line drawings of binary trees.
\end{problem}
Concerning \emph{straight-line strictly-upward drawings}, tight bounds are known. In fact, the algorithm of Crescenzi {\em et al.}~\cite{cdp-noad-92} can be suitably modified in order to obtain strictly-upward drawings (instead of aligning the subtrees of the root with their bottom sides on the same horizontal line, it is sufficient to align them with their left sides on the same vertical line). The same authors also showed a binary tree $T^*$ requiring $\Omega(n \log n)$ area in any strictly-upward drawing, hence their bound is tight. The tree $T^*$, that is shown in Fig.~\ref{fig:trees-upward-LB}, is composed of a path with $\Omega(n)$ nodes (forcing the height of the drawing to be $\Omega(n)$) and of a complete binary tree with $\Omega(n)$ nodes (forcing the width of the tree to be $\Omega(\log n)$).
\begin{figure}[htb]
\centering{
\mbox{\epsfig{figure=Trees-upwardLB.eps,scale =0.5,clip=}}}
\caption{A binary tree $T^*$ requiring $\Omega(n \log n)$ area in any strictly-upward drawing.}
\label{fig:trees-upward-LB}
\end{figure}
Concerning \emph{straight-line order-preserving drawings}, Garg and Rusu have shown in~\cite{gr-aeoppsdot-03} how to obtain an $O(n\log n)$ area upper bound for general trees. The algorithm of Garg and Rusu inductively assumes that an \emph{$\alpha$-drawing} of a tree $T$ can be constructed, that is, a straight-line order-preserving drawing of $T$ can be constructed with the further constraints that the root $r$ of $T$ is on the upper left corner of the bounding-box of the drawing, that the children of $r$ are placed on the vertical line one unit to the right of $r$, and that the vertical distance between $r$ and any other node of $T$ is at least $\alpha$. Refer to Fig.~\ref{fig:gargrusu-ordered}(a). To construct a drawing of $T$, the algorithm considers inductively constructed drawings of all the subtrees rooted at the children of $r$, except for the node $u$ that is the root of the subtree of $r$ with the greatest number of nodes, and place such drawings one unit to the right of $r$, with their left side aligned. Further, the algorithm considers inductively constructed drawings of all the subtrees rooted at the children of $u$, except for the node $v$ that is the root of the subtree of $u$ with the greatest number of nodes, and place such drawings two units to the right of $r$, with their left side aligned. Finally, the subtree rooted at $v$ is inductively drawn, the drawing is reflected and placed with its left side on the same vertical line as $r$. Thus, the height of the drawing is clearly $O(n)$, while its width is $w(n)=\max\{w(n-1),3+w(n/2)\}=O(\log n)$, where $w(n)$ denotes the maximum width of a drawing of an $n$-node tree constructed by the algorithm. Garg and Rusu also show how to combine their described result with a decomposition scheme of binary trees due to Chan \emph{et al.}~\cite{cgkt-oaarsod-02} to obtain $O(n\log \log n)$ area straight-line order-preserving drawings of binary trees. As no super-linear lower bound is known for the area requirements of straight-line order-preserving drawings of trees, the following problems remain open:
\begin{figure}[htb]
\centering{
\begin{tabular} {c c c c}
\mbox{\epsfig{figure=GargRusu-Ordered.eps,scale =0.45,clip=}} \hspace{4mm} &
\mbox{\epsfig{figure=GargRusu-Binary1.eps,scale =0.45,clip=}} \hspace{4mm} &
\mbox{\epsfig{figure=GargRusu-Binary2.eps,scale =0.45,clip=}} \hspace{4mm} &
\mbox{\epsfig{figure=Chan-Binary.eps,scale =0.45,clip=}} \\
(a) \hspace{4mm} & (b) \hspace{4mm} & (c) \hspace{4mm} & (d)
\end{tabular}}
\caption{(a) The inductive construction of a straight-line order-preserving drawing of a tree in $O(n\log n)$ area. (b)--(c) The inductive construction of a straight-line strictly-upward order-preserving drawing of a binary tree in $O(n\log n)$ area. The construction in (b) (resp. in (c)) refers to the case in which the left (resp. the right) subtree of $r$ contains more nodes than the right (resp. the left) subtree of $r$. (d) The geometric construction of the algorithm of Chan.}
\label{fig:gargrusu-ordered}
\end{figure}
\begin{problem}
Close the gap between the $O(n \log n)$ upper bound and the $\Omega(n)$ lower bound for the area requirements of straight-line order-preserving drawings of trees.
\end{problem}
\begin{problem}
Close the gap between the $O(n \log \log n)$ upper bound and the $\Omega(n)$ lower bound for the area requirements of straight-line order-preserving drawings of binary trees.
\end{problem}
Concerning \emph{straight-line strictly-upward order-preserving drawings}, Garg and Rusu have shown in~\cite{gr-aeoppsdot-03} how to obtain an $O(n\log n)$ area upper bound for binary trees. Observe that such an upper bound is still matched by the described $\Omega(n\log n)$ lower bound of Crescenzi {\em et al.}~\cite{cdp-noad-92}. The algorithm of Garg and Rusu, shown in Figs.~\ref{fig:gargrusu-ordered}(b)--(c), is similar to their described algorithm for constructing straight-line order-preserving drawings of trees. The results of Garg and Rusu improved upon previous results by Chan in~\cite{c-nlabdbt-02}. In~\cite{c-nlabdbt-02}, the author proved that every binary tree admits a straight-line strictly-upward order-preserving drawing in $O(n^{1+\epsilon})$ area, for any constant $\epsilon>0$. In the same paper, the author proved the best known upper bound for the area requirements of straight-line strictly-upward order-preserving drawings of trees, namely $O(n 4^{\sqrt{2\log n}})$. The approach of Chan consists of using very simple geometric constructions together with non-trivial tree decompositions. The simplest geometric construction discussed by Chan consists of selecting a path $P$ in the input tree $T$, of drawing $P$ on a vertical line $l$, and of inductively constructing drawings of the subtrees of $P$ to be placed to the left and right of $l$ (see Fig.~\ref{fig:gargrusu-ordered}(d)). Thus, denoting by $w(n)$ the maximum width of a drawing constructed by the algorithm, it holds $w(n)=1+w(n_1)+w(n_2)$, where $n_1$ and $n_2$ are the maximum number of nodes in a left subtree of $P$ and in a right subtree of $P$, respectively (assuming that $w(n)$ is monotone with $n$). Thus, depending on the way in which $P$ is chosen, different upper bounds on the asymptotic behavior of $w(n)$ can be achieved. Chan proves that $P$ can be chosen so that $w(n)=O(n^{0.48})$. Such a bound is at the base of the best upper bound for constructing straight-line drawings of outerplanar graphs (see~\cite{df-sadog-j09} and Sect.~\ref{se:straight-outerplanar}). An improvement on the following problem would be likely to improve the area upper bound on straight-line drawings of outerplanar graphs:
\begin{problem}
Let $w(n)$ be the function inductively defined as follows: $w(0)=0$, $w(1)=1$, and, for any $n>1$, let $w(n)=\max_T \{\min_P \{1 + w(n_1) + w(n_2)\}\}$, where the maximum is among all ordered rooted trees $T$ with $n$ vertices, the minimum is among all the root-to-leaf paths $P$ in $T$, where $n_1$ denotes the largest number of nodes in a left subtree of $P$, and where $n_2$ denotes the largest number of nodes in a right subtree of $P$. What is the asymptotic behavior of $w(n)$?
\end{problem}
It is easy to observe an $\Omega(\log n)$ lower bound for $w(n)$. We believe that in fact $w(n)=\Omega(2^{\sqrt{\log n}})$, but it is not clear to us whether the same bound can be achieved from above.
Turning the attention back to straight-line strictly-upward order-preserving drawings, the following problem remains open:
\begin{problem}
Close the gap between the $O(n 4^{\sqrt{2\log n}})$ upper bound and the $\Omega(n \log n)$ lower bound for the area requirements of straight-line strictly-upward order-preserving drawings of trees.
\end{problem}
Concerning \emph{straight-line orthogonal drawings}, Chan \emph{et al.}~in~\cite{cgkt-oaarsod-02} and Shin \emph{et al.}~in~\cite{skc-aeastd-00} have independently shown that $O(n \log \log n)$ area suffices for binary trees. Both algorithms are based on nice inductive geometric constructions and on non-trivial tree decompositions. Frati proved in~\cite{f-sodbtt-06} that every ternary tree admits a straight-line orthogonal drawing in $O(n^{1.631})$ area. The following problems are still open:
\begin{problem}
Close the gap between the $O(n \log \log n)$ upper bound and the $\Omega(n)$ lower bound for the area requirements of straight-line orthogonal drawings of binary trees.
\end{problem}
\begin{problem}
Close the gap between the $O(n^{1.631})$ upper bound and the $\Omega(n)$ lower bound for the area requirements of straight-line orthogonal drawings of ternary trees.
\end{problem}
Concerning \emph{straight-line upward orthogonal drawings}, Crescenzi \emph{et al.}~\cite{cdp-noad-92} and Chan \emph{et al.} in~\cite{cgkt-oaarsod-02} have shown that $O(n \log n)$ area suffices for binary trees. Such an area bound is worst-case optimal, as proved in~\cite{cgkt-oaarsod-02}. The tree providing the lower bound, shown in Fig.~\ref{fig:trees-orthogonal-LB}, consists of a path to which some complete binary trees are attached.
\begin{figure}[htb]
\centering{
\mbox{\epsfig{figure=Trees-orthogonalLB.eps,scale =0.5,clip=}}}
\caption{A binary tree requiring $\Omega(n \log n)$ area in any straight-line upward orthogonal drawing. The tree is composed of a path $P$ and of complete binary trees with size $n^{\alpha/2}$, where $\alpha$ is some constant greater than $0$, attached to the $i$-th node of $P$, for each $i$ multiple of $n^{\alpha/2}$.}
\label{fig:trees-orthogonal-LB}
\end{figure}
Concerning \emph{straight-line order-preserving orthogonal drawings}, $O(n^{1.5})$ and $O(n^2)$ area upper bounds are known~\cite{f-sodbtt-06} for binary and ternary trees, respectively. Once again such algorithms are based on simple inductive geometric constructions. While the bound for ternary trees is tight, no super-linear lower bound is known for straight-line order-preserving orthogonal drawings of binary trees, hence the following is open:
\begin{problem}
Close the gap between the $O(n^{1.5})$ upper bound and the $\Omega(n)$ lower bound for the area requirements of straight-line order-preserving orthogonal drawings of binary trees.
\end{problem}
\section{Poly-line Drawings}\label{se:poly-line}
In this section, we discuss algorithms and bounds for constructing small-area planar poly-line drawings of planar graphs and their subclasses.
In Sect.~\ref{se:poly-planar} we deal with general planar graphs, in Sect.~\ref{se:poly-sp} we deal with series-parallel and outerplanar graphs, and in Sect.~\ref{se:poly-trees} we deal with trees. Table~\ref{ta:poly-line} summarizes the best known area bounds for poly-line planar drawings of planar graphs and their subclasses. Observe that the lower bound of the table referring to general planar graphs hold true for {\em plane} graphs.
\begin{table}[!htb]\footnotesize
\centering
\linespread{1.2}
\selectfont
\begin{tabular}{|c|c|c|c|c|}
\cline{2-5}
\multicolumn{1}{c|}{} & \emph{Upper Bound} & \emph{Refs.} & \emph{Lower Bound} & \emph{Refs.} \\
\hline
{\em General Planar Graphs} & $\frac{4(n-1)^2}{9}$ & \cite{bsm-wtr-02} & $\frac{4(n-1)^2}{9}$ & \cite{fpp-hdpgg-90}\\
\hline
{\em Series-Parallel Graphs} & $O(n^{1.5})$ & \cite{b-sdogspgopg-11} & $\Omega(n 2^{\sqrt{\log n}})$ & \cite{f-lbarspg-j10}\\
\hline
{\em Outerplanar Graphs} & $O(n \log n)$ & \cite{b-dopgia-02,b-sdogspgopg-11} & $\Omega(n)$ & \emph{trivial}\\
\hline
{\em Trees} & $O(n \log n)$ & \cite{cdp-noad-92} & $\Omega(n)$ & \emph{trivial}\\
\hline
\end{tabular}
\vspace{2mm}
\caption{\small A table summarizing the area requirements for poly-line planar drawings of several classes of planar graphs.}
\label{ta:poly-line}
\end{table}
\subsection{General Planar Graphs} \label{se:poly-planar}
Every $n$-vertex plane graph admits a planar poly-line drawing on a grid with $O(n^2)$ area. In fact, this has been known since the beginning of the 80's~\cite{w-dpgt-82}. Tamassia and Tollis introduced in~\cite{tt-uavrpg-86} a technique that has later become pretty much a standard for constructing planar poly-line drawings. Namely, the authors showed that a poly-line drawing $\Gamma$ of a plane graph $G$ can be easily obtained from a visibility representation $R$ of $G$; moreover, $\Gamma$ and $R$ have asymptotically the same area. In order to obtain a visibility representation $R$ of $G$, Tamassia and Tollis design a very nice algorithm (an application is shown in Fig.~\ref{fig:tamassia-vr}). The algorithm assumes that $G$ is biconnected (if it is not, it suffices to augment $G$ to biconnected by inserting dummy edges, apply the algorithm, and then remove the inserted dummy edges to obtain a visibility representation of $G$). The algorithm consists of the following steps:
\begin{figure}[htb]
\centering{
\begin{tabular} {c c}
\mbox{\epsfig{figure=Visibility1.eps,scale =0.4,clip=}} \hspace{7mm} &
\mbox{\epsfig{figure=Visibility2.eps,scale =0.4,clip=}} \\
(a) \hspace{7mm} & (b)
\end{tabular}
}
\caption{An illustration for the algorithm of Tamassia and Tollis~\cite{tt-uavrpg-86}. (a) White circles and solid edges represent $G$. Black circles and dashed edges represent $G^*$. An st-numbering of $G$ (and the corresponding orientation) is shown. An orientation of $G^*$ and the number $2\psi(f)$ for each each face $f$ of $G$ is shown. (b) A visibility representation of $G$.}
\label{fig:tamassia-vr}
\end{figure}
(1) Consider an orientation of $G$ induced by an \emph{st-numbering} of $G$, that is a bijective mapping $\phi : V(G)\rightarrow \{1,\dots,n\}$ such that, for a given edge $(s,t)$ incident to the outer face of $G$, $\phi(s)=1$, $\phi(t)=n$, and for each $u\in V(G)$ with $u \neq s,t$, there exist two neighbors of $u$, say $v$ and $w$, such that $\phi(v)<\phi(u)<\phi(w)$; (2) consider the orientation of the dual graph $G^*$ of $G$ induced by the orientation of $G$; (3) the $y$-coordinate of each vertex-segment $u$ is given by $\phi(u)$; (4) the $y$-coordinates of the endpoints of each edge-segment $(u,v)$ are given by $\phi(u)$ and $\phi(v)$; (5) the $x$-coordinate of edge-segment $(s,t)$ is set equal to $-1$; (6) the $x$-coordinate of each edge-segment $(u,v)$ is chosen to be any number strictly between $2\psi(f)$ and $2\psi(g)$, where $f$ and $g$ are the faces adjacent to $(u,v)$ in $G$ and $\psi(f)$ denotes the length of the longest path from the source to $f$ in $G^*$; (7) finally, the $x$-coordinates of the endpoints of each vertex-segment $u$ is set equal to the smallest and largest $x$-coordinates of its incident edges.
After the algorithm of Tamassia and Tollis, a large number of algorithms have been proposed to construct poly-line drawings of planar graphs (see, e.g.,~\cite{gm-ppdgar-98,gw-fdpgcp-00,cdgk-dpgca-01,zs-ppd-08,z-ppdgt-10}), proposing several tradeoffs between area requirements, number of bends, and angular resolution. Here we briefly discuss an algorithm proposed by Bonichon \emph{et al.}~in~\cite{bsm-wtr-02}, the first one to achieve optimal area, namely $\frac{4(n-1)^2}{9}$. The algorithm consists of two steps. In the first one, a deep study of Schnyder realizers (see~\cite{s-epgg-90} and Sect.~\ref{se:straight-planar} for the definition of Schnyder realizers) leads to the definition of a \emph{weak-stratification} of a realizer. Namely, given a realizer $(T_0,T_1,T_2)$ of a triangulation $G$, a weak-stratification is a layering $L$ of the vertices of $G$ such that $T_0$ (which is rooted at the vertex incident to the outer face of $G$) is upward, while $T_1$ and $T_2$ (which are rooted at the vertices incident to the outer face of $G$) are downward and some further conditions are satisfied. Each vertex will get a $y$-coordinate which is equal to its layer in the weak stratification. In the second step $x$-coordinates for vertices and bends are computed. The conditions of the weak stratification ensure that a planar drawing can in fact be obtained.
\subsection{Series-Parallel and Outerplanar Graphs} \label{se:poly-sp}
Biedl proved in~\cite{b-sdogspgopg-11} that every series-parallel graph admits a poly-line drawing with $O(n^{1.5})$ area and a poly-line drawing with $O(fn \log n)$ area, where $f$ is the fan-out of the series-parallel graph. In particular, since outerplanar graphs are series-parallel graphs with fan-out two, the last result implies that outerplanar graphs admit poly-line drawings with $O(n \log n)$ area. Biedl's algorithm constructs a visibility representation $R$ of the input graph $G$ with $O(n^{1.5})$ area; a poly-line drawing $\Gamma$ with asymptotically the same area of $R$ can then be easily obtained from $R$.
\begin{figure}[htb]
\centering{
\begin{tabular} {c c c c}
\mbox{\epsfig{figure=Biedl-SP-1a.eps,scale =0.45,clip=}} \hspace{4mm} &
\mbox{\epsfig{figure=Biedl-SP-1.eps,scale =0.45,clip=}} \hspace{4mm} &
\mbox{\epsfig{figure=Biedl-SP-2a.eps,scale =0.33,clip=}} \hspace{4mm} &
\mbox{\epsfig{figure=Biedl-SP-2.eps,scale =0.33,clip=}} \\
(a) \hspace{4mm} & (b) \hspace{4mm} & (c) \hspace{4mm} & (d) \vspace{3mm}\\
\mbox{\epsfig{figure=Biedl-SP-3a.eps,scale =0.27,clip=}} \hspace{3mm} &
\mbox{\epsfig{figure=Biedl-SP-3.eps,scale =0.27,clip=}} \hspace{3mm} &
\mbox{\epsfig{figure=Biedl-SP-4a.eps,scale =0.27,clip=}} \hspace{3mm} &
\mbox{\epsfig{figure=Biedl-SP-4.eps,scale =0.27,clip=}} \\
(e) \hspace{3mm} & (f) \hspace{3mm} & (g) \hspace{3mm} & (h)
\end{tabular}
}
\caption{Biedl's algorithm for constructing visibility representations of series-parallel graphs. (a)--(b) The base case. (c)--(d) The parallel case. (e)--(h) The series case.}
\label{fig:biedl-sp}
\end{figure}
In order to construct a visibility representation $R$ of the input graph $G$, Biedl relies on a strong inductive hypothesis, namely that a small area visibility representation $R$ of $G$ can be constructed with the further constraint that the poles $s$ and $t$ of $G$ are placed at the top right corner and at the bottom right corner of the representation, respectively. Figs.~\ref{fig:biedl-sp}(a)--(b) show how this is accomplished in the base case. The parallel case is also pretty simple, as the visibility representations of the components of $G$ are just placed one besides the other (as in Figs.~\ref{fig:biedl-sp}(c)--(d)). The series case is much more involved. Namely, assuming w.l.o.g. that $G$ is the series of two components $H_1$ and $H_2$, where $H_1$ has poles $s$ and $x$ and $H_2$ has poles $x$ and $t$, and assuming w.l.o.g. that $H_2$ has more vertices than $H_1$, then if $H_2$ is the parallel composition of a ``small'' number of components, the composition shown in Figs.~\ref{fig:biedl-sp}(e)--(f) is applied, while if $H_2$ is the parallel composition of a ``large'' number of components, the composition shown in Figs.~\ref{fig:biedl-sp}(g)--(h) is applied. The rough idea behind these constructions is that if $H_2$ is the parallel composition of a small number of components, then a vertical unit can be spent for each of them without increasing much the height of the drawing; on the other hand, if $H_2$ is the parallel composition of a large number of components, then lots of such components have few vertices, hence two of them can be placed one above the other without increasing much the height of the drawing.
The following problems remain open:
\begin{problem}
Close the gap between the $O(n^{1.5})$ upper bound and the $\Omega(n 2^{\sqrt{\log n}})$ lower bound for the area requirements of poly-line drawings of series-parallel graphs.
\end{problem}
\begin{problem}
Close the gap between the $O(n \log n)$ upper bound and the $\Omega(n)$ lower bound for the area requirements of poly-line drawings of outerplanar graphs.
\end{problem}
\subsection{Trees} \label{se:poly-trees}
No algorithms are known exploiting the possibility of bending the edges of a tree to get area bounds better than the corresponding ones shown for straight-line drawings.
\begin{problem}
Close the gap between the $O(n \log n)$ upper bound and the $\Omega(n)$ lower bound for the area requirements of poly-line drawings of trees.
\end{problem}
However, better bounds can be achieved for poly-line drawings satisfying further constraints. Table~\ref{ta:trees-straight-line} summarizes the best known area bounds for various kinds of poly-line drawings of trees.
\begin{table}[!htb]\centering\footnotesize
\linespread{1.2}
\selectfont
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\cline{2-9}
\multicolumn{1}{c|}{} & \emph{Ord. Pres.} & \emph{Upw.} & \emph{Str. Upw.} & \emph{Orth.} & \emph{Upper Bound} & \emph{Refs.} & \emph{Lower Bound} & \emph{Refs.} \\
\cline{2-9}
\hline
\emph{Binary} & & & & & $O(n)$ & \cite{ggt-putdoa-96} & $\Omega(n)$ & \emph{trivial} \\
\hline
\emph{Binary} & \checkmark & & & & $O(n\log \log n)$ & \cite{gr-aeoppsdot-03} & $\Omega(n)$ & \emph{trivial} \\
\hline
\emph{Binary} & & \checkmark & & & $O(n)$ & \cite{ggt-putdoa-96} & $\Omega(n)$ & \emph{trivial} \\
\hline
\emph{Binary} & & & \checkmark & & $O(n\log n)$ & \cite{cdp-noad-92} & $\Omega(n \log n)$ & \cite{cdp-noad-92} \\
\hline
\emph{Binary} & \checkmark & & \checkmark & & $O(n \log n)$ & \cite{ggt-putdoa-96} & $\Omega(n\log n)$ & \cite{cdp-noad-92}\\
\hline
\emph{Binary} & & & & \checkmark & $O(n)$ & \cite{Val81} & $\Omega(n)$ & \emph{trivial}\\
\hline
\emph{Binary} & & \checkmark & & \checkmark & $O(n \log \log n)$ & \cite{ggt-putdoa-96} & $\Omega(n \log \log n)$ & \cite{ggt-putdoa-96}\\
\hline
\emph{Binary} & \checkmark & & & \checkmark & $O(n)$ & \cite{dt-laepg-81} & $\Omega(n)$ & \emph{trivial}\\
\hline
\emph{Binary} & \checkmark & \checkmark & & \checkmark & $O(n \log n)$ & \cite{k-saoudbtt-95} & $\Omega(n \log n)$ & \cite{ggt-putdoa-96}\\
\hline
\emph{Ternary} & & & & \checkmark & $O(n)$ & \cite{Val81} & $\Omega(n)$ & \emph{trivial}\\
\hline
\emph{Ternary} & & \checkmark & & \checkmark & $O(n \log n)$ & \cite{k-saoudbtt-95} & $\Omega(n \log n)$ & \cite{k-saoudbtt-95}\\
\hline
\emph{Ternary} & \checkmark & & & \checkmark & $O(n)$ & \cite{dt-laepg-81} & $\Omega(n)$ & \emph{trivial}\\
\hline
\emph{Ternary} & \checkmark & \checkmark & & \checkmark & $O(n \log n)$ & \cite{k-saoudbtt-95} & $\Omega(n \log n)$ & \cite{ggt-putdoa-96}\\
\hline
\emph{General} & & & & & $O(n \log n)$ & \cite{cdp-noad-92} & $\Omega(n)$ & \emph{trivial} \\
\hline
\emph{General} & \checkmark & & & & $O(n\log n)$ & \cite{gr-aeoppsdot-03} & $\Omega(n)$ & \emph{trivial} \\
\hline
\emph{General} & & \checkmark & & & $O(n\log n)$ & \cite{cdp-noad-92} & $\Omega(n)$ & \emph{trivial} \\
\hline
\emph{General} & & & \checkmark & & $O(n\log n)$ & \cite{cdp-noad-92} & $\Omega(n \log n)$ & \cite{cdp-noad-92} \\
\hline
\emph{General} & \checkmark & & \checkmark & & $O(n 4^{\sqrt{2\log n}})$ & \cite{c-nlabdbt-02} & $\Omega(n\log n)$ & \cite{cdp-noad-92}\\
\hline
\end{tabular}
\vspace{2mm}
\caption{\small Summary of the best known area bounds for poly-line drawings of trees. ``Ord. Pres.'', ``Upw.'', ``Str. Upw.'', and ``Orth.'' stand for order-preserving, upward, strictly-upward, and orthogonal, respectively.}
\label{ta:trees-straight-line}
\end{table}
Concerning \emph{poly-line upward drawings}, a linear area bound is known, due to Garg {\em et al.}~\cite{ggt-putdoa-96}, for all trees whose degree is $O(n^{\delta})$, where $\delta$ is \emph{any} constant less than $1$. The algorithm of Garg {\em et al.}~first constructs a layering $\gamma(T)$ of the input tree $T$; in $\gamma(T)$ each node $u$ is assigned a layer smaller than or equal to the layer of the leftmost child of $u$ and smaller than the layer of any other child of $u$; second, the authors show that $\gamma(T)$ can be converted into an upward poly-line drawing whose height is the number of layers and whose width is the maximum \emph{width of a layer}, that is the number of nodes of the layer plus the number of edges crossing the layer; third, the authors show how to construct a layering of every tree whose degree is $O(n^{\delta})$ so that the number of layers times the maximum width of a layer is $O(n)$. No upper bound better than $O(n \log n)$ (from the results on straight-line drawings, see~\cite{cdp-noad-92} and Sect.~\ref{se:straight-trees}) and no super-linear lower bound is known for trees with unbounded degree.
\begin{problem}
Close the gap between the $O(n \log n)$ upper bound and the $\Omega(n)$ lower bound for the area requirements of poly-line upward drawings of trees.
\end{problem}
Concerning \emph{poly-line order-preserving strictly-upward drawings}, Garg {\em et al.}~\cite{ggt-putdoa-96} show a simple algorithm to achieve $O(n\log n)$ area for bounded-degree trees. The algorithm, whose construction is shown in Fig.~\ref{fig:changood}(a), consists of stacking inductively constructed drawings of the subtrees of the root of the input tree $T$, in such a way that the tree with the greatest number of nodes is the bottommost in the drawing. The edges connecting the root to its subtrees are then routed besides the subtrees. The $O(n\log n)$ area upper bound is tight. Namely, there exist binary trees requiring $\Omega(n \log n)$ area in any strictly-upward order-preserving drawing~\cite{cdp-noad-92} and binary trees requiring $\Omega(n \log n)$ area in any (even non-strictly) upward order-preserving drawing~\cite{ggt-putdoa-96}. The lower bound tree of Garg {\em et al.}~is shown in Fig.~\ref{fig:changood}(b). As far as we know, no area bounds better than the ones for straight-line drawings have been proved for general trees, hence the following are open:
\begin{problem}
Close the gap between the $O(n \log n)$ upper bound and the $\Omega(n)$ lower bound for the area requirements of poly-line order-preserving drawings of trees.
\end{problem}
\begin{problem}
Close the gap between the $O(n 4^{\sqrt{\log n}})$ upper bound and the $\Omega(n \log n)$ lower bound for the area requirements of poly-line order-preserving strictly-upward drawings of trees.
\end{problem}
\begin{figure}[htb]
\centering{
\begin{tabular} {c c}
\mbox{\epsfig{figure=ChanGood.eps,scale =0.45,clip=}} \hspace{7mm} &
\mbox{\epsfig{figure=ChanGoodLower.eps,scale =0.45,clip=}} \\
(a) \hspace{7mm} & (b)
\end{tabular}
}
\caption{(a) The construction of Garg {\em et al.}~\cite{ggt-putdoa-96} to obtain $O(n\log n)$ area poly-line order-preserving strictly-upward drawings of bounded-degree trees. (b) A tree requiring $\Omega(n \log n)$ area in any upward order-preserving drawing. The triangle represents a complete binary tree with $n/3$ nodes.}
\label{fig:changood}
\end{figure}
Concerning \emph{orthogonal drawings}, Valiant proved in~\cite{Val81} that every $n$-node ternary tree (and every $n$-node binary tree) admits a $\Theta(n)$ area orthogonal drawing. Such a result was strengthened by Dolev and Trickey in ~\cite{dt-laepg-81}, who proved that ternary trees (and binary trees) admit $\Theta(n)$ area order-preserving orthogonal drawings. The technique of Valiant is based on the use of separator edges (see~\cite{Val81} and Sect.~\ref{se:straight-trees}). The result of Dolev and Trickey is a consequence of a more general result on the construction of linear area embeddings of degree-$4$ outerplanar graphs.
Concerning \emph{orthogonal upward drawings}, an $O(n \log \log n)$ area bound for binary trees was proved by Garg \emph{et al.} in~\cite{ggt-putdoa-96}. The algorithm has several ingredients. (1) A simple algorithm is shown to construct orthogonal upward drawings in $O(n \log n)$ area; such drawings exhibit the further property that no vertical line through a node of degree at most two intersects the drawing below such a node. (2) The \emph{separator tree} $S$ of the input tree $T$ is constructed; such a tree represents the recursive decomposition of a tree via separator edges; namely, $S$ is a binary tree that is recursively constructed as follows: The root $r$ of $S$ is associated with tree $T$ and with a separator edge of $T$, that splits $T$ into subtrees $T_1$ and $T_2$; the subtrees of $r$ are the separator trees associated with $T_1$ and $T_2$; observe that the leaves of $S$ are the nodes of $T$. (3) A \emph{truncated separator tree} $S'$ is obtained from $S$ by removing all the nodes of $S$ associated with subtrees of $T$ with less than $\log n$ nodes. (4) Drawings of the subtrees of $T$ associated with the leaves of $S'$ are constructed via the $O(n \log n)$ area algorithm. (5) Such drawings are stacked one on top of the other and the separator edges connecting them are routed (see Fig.~\ref{fig:changoodorthogonal}(a)). The authors prove that the constructed drawings have $O(\frac{n \log \log n}{\log n})$ height and $O(\log n)$ width, thus obtaining the claimed upper bound. The same authors also proved that the $O(n \log \log n)$ bound is tight, by exhibiting the class of trees shown in Fig.~\ref{fig:changoodorthogonal}(b). In~\cite{k-saoudbtt-95} Kim showed that $\Theta(n \log n)$ area is an optimal bound for upward orthogonal drawings of ternary trees. The upper bound comes from a stronger result on orthogonal order-preserving upward drawings cited below, while the lower bound comes from the tree shown in Fig.~\ref{fig:changoodorthogonal}(c).
\begin{figure}[htb]
\centering{
\begin{tabular} {c c c}
\mbox{\epsfig{figure=ChanGoodOrthogonal.eps,scale =0.45,clip=}} \hspace{7mm} &
\mbox{\epsfig{figure=ChanGoodLower2.eps,scale =0.45,clip=}} \hspace{7mm} &
\mbox{\epsfig{figure=KimLower.eps,scale =0.45,clip=}} \\
(a) \hspace{7mm} & (b) \hspace{7mm} & (c)
\end{tabular}
}
\caption{(a) The construction of Garg {\em et al.}~\cite{ggt-putdoa-96} to obtain $O(n\log \log n)$ area orthogonal upward drawings of binary trees. Rectangles represent drawings of small subtrees constructed via an $O(n \log n)$ area algorithm. (b) A binary tree requiring $\Omega(n \log \log n)$ area in any upward orthogonal drawing. The tree is composed of a chain with $n/3$ nodes, a complete binary tree with $n/3$ nodes (the large triangle in the figure), and $\frac{n}{3\sqrt{\log n}}$ subtrees (the small triangles in the figure) with $\sqrt{\log n}$ nodes rooted at the child of each $\sqrt{\log n}$-th node of the chain. (c) A ternary tree requiring $\Omega(n \log n)$ area in any upward orthogonal drawing. The tree is composed of a chain with $n/4$ nodes, two other children for each node of the chain, and a complete binary tree with $n/4$ nodes (the large triangle in the figure)}
\label{fig:changoodorthogonal}
\end{figure}
Concerning \emph{orthogonal order-preserving upward drawings}, $\Theta(n \log n)$ is an optimal bound both for binary and ternary trees. In fact, Kim~\cite{k-saoudbtt-95} proved the upper bound for ternary trees (such a bound can be immediately extended to binary trees). The simple construction of Kim is presented in Fig.~\ref{fig:kim}. The lower bound directly comes from the results of Garg \emph{et al.}~on order-preserving upward (non-orthogonal) drawings~\cite{ggt-putdoa-96}.
\begin{figure}[htb]
\centering{
\begin{tabular} {c c c}
\mbox{\epsfig{figure=Kim1.eps,scale =0.3,clip=}} \hspace{7mm} &
\mbox{\epsfig{figure=Kim2.eps,scale =0.3,clip=}} \hspace{7mm} &
\mbox{\epsfig{figure=Kim3.eps,scale =0.3,clip=}} \\
(a) \hspace{7mm} & (b) \hspace{7mm} & (c)
\end{tabular}
}
\caption{An algorithm to construct $O(n \log n)$ area orthogonal order-preserving upward drawings of ternary trees. The figures illustrate the cases in which: (a) The right subtree has the greatest number of nodes; (b) the middle subtree has the greatest number of nodes; and (b) the left subtree has the greatest number of nodes.}
\label{fig:kim}
\end{figure}
\section{Upward Drawings}\label{se:upward}
In this section, we discuss algorithms and bounds for constructing small-area planar straight-line/poly-line upward drawings of upward planar directed acyclic graphs. Table~\ref{ta:upward} summarizes the best known area bounds for straight-line upward planar drawings of upward planar DAGs and their subclasses.
\begin{table}[!htb]\footnotesize
\centering
\linespread{1.2}
\selectfont
\begin{tabular}{|c|c|c|c|c|}
\cline{2-5}
\multicolumn{1}{c|}{} & \emph{Upper Bound} & \emph{Refs.} & \emph{Lower Bound} & \emph{Refs.} \\
\hline
{\em General Upward Planar DAGs} & $O(c^n)$ & \cite{GargTam93} & $\Omega(b^n)$ & \cite{BattistaTT92}\\
\hline
{\em Fixed-Embedding Series-Parallel DAGs} & $O(c^n)$ & \cite{GargTam93} & $\Omega(b^n)$ & \cite{BertolazziCBTT94}\\
\hline
{\em Series-Parallel DAGs} & $O(n^2)$ & \cite{BertolazziCBTT94} & $\Omega(n^2)$ & \emph{trivial}\\
\hline
{\em Bipartite DAGs} & $O(c^n)$ & \cite{GargTam93} & $\Omega(b^n)$ & \cite{f-mapudt-08}\\
\hline
{\em Fixed-Embedding Directed Trees} & $O(c^n)$ & \cite{GargTam93} & $\Omega(b^n)$ & \cite{f-mapudt-08}\\
\hline
{\em Directed Trees} & $O(n \log n)$ & \cite{f-mapudt-08} & $\Omega(n \log n)$ & \cite{f-mapudt-08}\\
\hline
\end{tabular}
\vspace{2mm}
\caption{\small A table summarizing the area requirements for straight-line upward planar drawings of upward planar DAGs; $b$ and $c$ denote constants greater than $1$.}
\label{ta:upward}
\end{table}
It is known that testing the upward planarity of a DAG is an NP-complete problem if the DAG has a variable embedding~\cite{GargT01}, while it is polynomial-time solvable if the embedding of the DAG is fixed~\cite{BertolazziBLM94}, if the underlying graph is an outerplanar graph~\cite{Papakostas94}, if the DAG has a single source~\cite{HuttonL96}, or if it is bipartite~\cite{BattistaLR90}. Di Battista and Tamassia~\cite{BattistaT88} showed that a DAG is upward planar if and only if it is a subgraph of an st-planar DAG. Some families of DAGs are always upward planar, like the series-parallel DAGs and the directed trees.
Di Battista and Tamassia proved in~\cite{BattistaT88} that every upward planar DAG admits an upward straight-line drawing. Such a result is achieved by means of an algorithm similar to F\'ary's algorithm for constructing planar straight-line drawings of undirected planar graphs (see Sect.~\ref{se:straight-planar}). However, while planar straight-line drawings of undirected planar graphs can be constructed in polynomial area, Di Battista \emph{et al.}~proved in~\cite{BattistaTT92} that there exist upward planar DAGs that require exponential area in any planar straight-line upward drawing. Such a result is achieved by considering the class $G_n$ of DAGs whose inductive construction is shown in Fig.~\ref{fig:upward}(a)--(b) and by using some geometric considerations to prove that the area of the smallest region containing an upward planar straight-line drawing of $G_{n}$ is a constant number of times larger than the area of a region containing an upward planar straight-line drawing of $G_{n-1}$. The techniques introduced by Di Battista \emph{et al.}~in~\cite{BattistaTT92} to prove the exponential lower bound for the area requirements of upward planar straight-line drawings of upward planar DAGs have later been strengthened by Bertolazzi \emph{et al.}~in~\cite{BertolazziCBTT94} and by Frati in~\cite{f-mapudt-08} to prove, respectively, that there exist series-parallel DAGs with fixed embedding (see Fig.~\ref{fig:upward}(c)) and there exist directed trees with fixed embedding (see Fig.~\ref{fig:upward}(d)) requiring exponential area in any upward planar straight-line drawing. Similar lower bound techniques have also been used to deal with straight-line drawings of clustered graphs (see Sect.~\ref{se:clustered}).
\begin{figure}[htb]
\centering
\begin{tabular}{c c c c}
\mbox{\epsfig{figure=upward-lowerbound1.eps,scale=0.55,clip=}} \hspace{2mm} &
\mbox{\epsfig{figure=upward-lowerbound2.eps,scale=0.55,clip=}} \hspace{2mm} &
\mbox{\epsfig{figure=upward-lowerbound3.eps,scale=0.55,clip=}} \hspace{2mm} &
\mbox{\epsfig{figure=upward-lowerbound4.eps,scale=0.55,clip=}}\\
(a) \hspace{2mm} & (b) \hspace{2mm} & (c) \hspace{2mm} & (d) \\
\end{tabular}
\caption{(a)-(b) Inductive construction of a class $G_n$ of upward planar DAGs requiring exponential area in any planar straight-line upward drawing. (c) Inductive construction of a class of series-parallel DAGs requiring exponential area in any planar straight-line upward drawing respecting a fixed embedding. (d) A class of directed trees requiring exponential area in any planar straight-line upward drawing respecting a fixed embedding.}
\label{fig:upward}
\end{figure}
On the positive side, area-efficient algorithms exist for constructing upward planar straight-line drawings for restricted classes of upward planar DAGs. Namely, Bertolazzi \emph{et al.}~in~\cite{BertolazziCBTT94} have shown how to construct upward planar straight-line drawings of series-parallel DAGs in optimal $\Theta(n^2)$ area, and Frati~\cite{f-mapudt-08} has shown how to construct upward planar straight-line drawings of directed trees in optimal $\Theta(n \log n)$ area. Both algorithms are based on the inductive construction of upward planar straight-line drawings satisfying some additional geometric constraints. We remark that for upward planar DAGs whose underlying graph is a series-parallel graph neither an exponential lower bound nor a polynomial upper bound is known for the area requirements of straight-line upward planar drawings. Observe that testing upward planarity for this family of graphs can be done in polynomial time~\cite{dgl-usupt-05}.
\begin{problem}
What are the area requirements of straight-line upward planar drawings of upward planar DAGs whose underlying graph is a series-parallel graph?
\end{problem}
Algorithms have been provided to construct upward planar poly-line drawings of upward planar DAGs. The first $\Theta(n^2)$ optimal area upper bound for such drawings has been established by Di Battista and Tamassia in~\cite{BattistaT88}. Their algorithm consists of first constructing an upward visibility representation of the given upward planar DAG and then of turning such a representation into an upward poly-line drawing. Such a technique has been discussed in Sect.~\ref{se:poly-line}.
\section{Convex Drawings}\label{se:convex}
In this section, we discuss algorithms and bounds for constructing small-area convex and strictly-convex drawings of planar graphs. Table~\ref{ta:convex} summarizes the best known area bounds for convex and strictly-convex drawings of planar graphs.
\begin{table}[!htb]\footnotesize
\centering
\linespread{1.2}
\selectfont
\begin{tabular}{|c|c|c|c|c|}
\cline{2-5}
\multicolumn{1}{c|}{} & \emph{Upper Bound} & \emph{Refs.} & \emph{Lower Bound} & \emph{Refs.} \\
\hline
{\em Convex} & $n^2+O(n)$ & \cite{ChrobakK97,st-cdpg-92,BattistaTV99,BonichonFM07} & $\frac{4n^2}{9} - O(n)$ & \cite{Val81,fpp-hdpgg-90,FratiP07,mnra-madp3t-10}\\
\hline
{\em Strictly-Convex} & $O(n^4)$ & \cite{br-scdpg-06} & $\Omega(n^3)$ & \cite{a-lbvscb-63,r-baclp-88,BaranyP92,BaranyT04} \\
\hline
\end{tabular}
\vspace{2mm}
\caption{\small A table summarizing the area requirements for convex and strictly-convex drawings of triconnected plane graphs.}
\label{ta:convex}
\end{table}
Not every planar graph admits a convex drawing. Tutte~\cite{t-crg-60,t-hdg-63} proved that every triconnected planar graph $G$ admits a strictly-convex drawing in which its outer face is drawn as an arbitrary strictly-convex polygon $P$. His algorithm consists of first drawing the outer face of $G$ as $P$ and then placing each vertex at the barycenter of the positions of its adjacent vertices. This results in a set of linear equations that always admits a unique solution.
Characterizations of the plane graphs admitting convex drawings were given by Tutte in~\cite{t-crg-60,t-hdg-63}, by Thomassen in~\cite{Thomassen80,t-prg-84}, by Chiba, Yamanouchi, and Nishizeki in~\cite{cyn-lacdp-84}, by Nishizeki and Chiba in~\cite{nc-pgta-88}, by Di Battista, Tamassia, and Vismara in~\cite{BattistaTV01}. Roughly speaking, the plane graphs admitting convex drawings are biconnected, their separation pairs are composed of vertices both incident to the outer face, and distinct separation pairs do not ``nest''. Chiba, Yamanouchi, and Nishizeki presented in~\cite{cyn-lacdp-84} a linear-time algorithm for testing whether a graph admits a convex drawing and producing a convex drawing if the graph allows for one. The area requirements of convex and strictly-convex grid drawings have been widely studied, especially for triconnected plane graphs.
Convex grid drawings of triconnected plane graphs can be realized on a quadratic-size grid. This was first shown by Kant in~\cite{Kant96}. In fact, Kant proved that such drawings can always be realized on a $(2n-4)\times(n-2)$ grid. The result is achieved by defining a stronger notion of canonical ordering of a plane graph (see Sect.~\ref{se:straight-planar}). Such a strengthened canonical ordering allows to construct every triconnected plane graph $G$ starting from a cycle delimiting an internal face of $G$ and repeatedly adding to the previously constructed biconnected graph $G_{k}$ a vertex or a path in the outer face of $G_{k}$ so that the newly formed graph $G_{k+1}$ is also biconnected (see Fig.~\ref{fig:kant}). Observe that this generalization of the canonical ordering allows to deal with plane graphs containing non-triangular faces. Similarly to de Fraysseix \emph{et al.}'s algorithm~\cite{fpp-hdpgg-90}, Kant's algorithm exploits a canonical ordering of $G$ to incrementally construct a convex drawing of $G$ in which the outer face of the currently considered graph $G_k$ is composed of segments whose slopes are either $-45^{\degree}$, or $0^{\degree}$, or $45^{\degree}$.
\begin{figure}[htb]
\centering
\mbox{\epsfig{figure=KantCanonical.eps,scale=0.35,clip=}}
\caption{An illustration of the canonical ordering of a triconnected plane graph.}
\label{fig:kant}
\end{figure}
The bound of Kant was later improved down to $(n-2)\times(n-2)$ by Chrobak and Kant~\cite{ChrobakK97}, and independently by Schnyder and Trotter~\cite{st-cdpg-92}. The result of Chrobak and Kant again relies on a canonical ordering. On the other hand, the result of Schnyder and Trotter relies on a generalization of the Schnyder realizers (see Sect.~\ref{se:straight-planar}) in order to deal with triconnected plane graphs. Such an extension was independently shown by Di~Battista, Tamassia, and Vismara~\cite{BattistaTV99}, who proved that every triconnected plane graph has a convex drawing on a $(f-2)\times(f-2)$ grid, where $f$ is the number of faces of the graph. The best bound is currently, as far as we know, an $(n-2-\Delta)\times(n-2-\Delta)$ bound achieved by Bonichon, Felsner, and Mosbah in~\cite{BonichonFM07}. The bound is again achieved using Schnyder realizers. The parameter $\Delta$ is dependent of the Schnyder realizers, and can vary among $0$ and $\frac{n}{2}-2$. The following remains open:
\begin{problem}
Close the gap between the $(n-2-\Delta)\times(n-2-\Delta)$ upper bound and the $\frac{4n^2}{9}-O(n)$ lower bound for the area requirements of convex drawings of triconnected plane graphs.
\end{problem}
Strictly-convex drawings of triconnected plane graphs might require $\Omega(n^3)$ area. In fact, an $n$-vertex cycle needs $\Omega(n^3)$ area in any grid realization (see, e.g.,~\cite{a-lbvscb-63,BaranyP92,BaranyT04}). The currently best lower bound for the area requirements of a strictly-convex polygon drawn on the grid, which has been proved by Rabinowitz in~\cite{r-baclp-88}, is $\frac{n^3}{8\pi^2}$. The first polynomial upper bound for strictly-convex drawings of triconnected plane graphs has been proved by Chrobak, Goodrich, and Tamassia in~\cite{ChrobakGT96}. The authors showed that every triconnected plane graph admits a strictly-convex drawing in an $O(n^3)\times O(n^3)$ grid. Their idea consists of first constructing a (non-strictly-) convex drawing of the input graph, and of then perturbing the positions of the vertices in order to achieve strict convexity. A more elaborated technique relying on the same idea allowed Rote to achieve an $O(n^{7/3})\times O(n^{7/3})$ area upper bound in~\cite{Rote05}, which was further improved by B\'ar\'any and Rote to $O(n^2)\times O(n^2)$ and to $O(n)\times O(n^3)$ in~\cite{br-scdpg-06}. The last ones are, as far as we know, the best known upper bounds. One of the main differences between the Chrobak \emph{et al.}'s algorithm, and the B\'ar\'any and Rote's ones is that the former one constructs the intermediate non-strictly-convex drawing by making use of a canonical ordering of the graph, while the latter ones by making use of the Schnyder realizers. The following is, in our opinion, a very nice open problem:
\begin{problem}
Close the gap between the $O(n^4)$ upper bound and the $\Omega(n^3)$ lower bound for the area requirements of strictly-convex drawings of triconnected plane graphs.
\end{problem}
\section{Proximity Drawings}\label{se:proximity}
In this section, we discuss algorithms and bounds for constructing small-area proximity drawings of planar graphs.
Characterizing the graphs that admit a proximity drawing, for a certain definition of proximity, is a difficult problem. For example, despite several research efforts (see, e.g.,~\cite{d-rdt-90,ll-pdog-96,ds-gtcidr-96}), characterizing the graphs that admit a \emph{realization} (word which often substitutes \emph{drawing} in the context of proximity graphs) as Delaunay triangulations is still an intriguing open problem. Dillencourt showed that every maximal outerplanar graph can be realized as a Delaunay triangulation~\cite{d-rdt-90} and provided examples of small triangulations which can not. The decision version of several realizability problems (that is, given a graph $G$ and a definition of proximity, can $G$ be realized as a proximity graph?) is $\mathcal{NP}$-hard. For example, Eades and Whitesides proved that deciding whether a tree can be realized as a minimum spanning tree is an $\mathcal{NP}$-hard problem~\cite{EadesW96}, and that deciding whether a graph can be realized as a nearest neighbor graph is an $\mathcal{NP}$-hard problem~\cite{ew-lerpnng-96}, as well. Both proofs rely on a mechanism for providing the hardness of graph drawing problems, called \emph{logic engine}, which is interesting by itself. On the other hand, for several definitions of proximity graphs (such as Gabriel graphs and relative neighborhood graphs), the realizability problem is polynomial-time solvable for trees, as shown by Bose, Lenhart, and Liotta~\cite{bll-cpt-96}; further, Lubiw and Sleumer proved that maximal outerplanar graphs can be realized as relative neighborhood graphs and Gabriel graphs~\cite{ls-mogrng-93}, a result later extended by Lenhart and Liotta to all biconnected outerplanar graphs~\cite{ll-pdog-96}. For more results about proximity drawings, see~\cite{gll-pds-94,l-cpdg-95,gd-handbook}.
Most of the known algorithms to construct proximity drawings produce representations whose size increases exponentially with the number of vertices (see, e.g.,~\cite{ls-mogrng-93,bll-cpt-96,ll-pdog-96,dlw-swp-06}). This seems to be unavoidable for most kinds of proximity drawings, although few exponential area lower bounds are known. Liotta \emph{et al.}~\cite{ltt-argd-97} showed a class of graphs (whose inductive construction is shown in Fig.~\ref{fig:gabriel}) requiring exponential area in any Gabriel drawing, in any weak Gabriel drawing, and in any $\beta$-drawing.
\begin{figure}[htb]
\centering
\begin{tabular}{c c}
\mbox{\epsfig{figure=Gabrielgraphs.eps,scale=0.55,clip=}} \hspace{5mm} &
\mbox{\epsfig{figure=Gabrielgraphs2.eps,scale=0.55,clip=}}
\end{tabular}
\caption{Inductive construction of a class $G_n$ of graphs requiring exponential area in any Gabriel drawing, in any weak Gabriel drawing, and in any $\beta$-drawing.}
\label{fig:gabriel}
\end{figure}
Their proof is based on the observation that the circles whose diameters are the segments representing the edges incident to the outer face of $G_n$ can not contain any point in their interior. Consequently, the vertices of $G_{n-1}$ are allowed only to be placed in a region whose area is a constant number of times smaller than the area of $G_{n}$. On the other hand, Penna and Vocca~\cite{pv-pdpav-04} showed algorithms to construct polynomial-area weak Gabriel drawings and weak $\beta$-drawings of binary and ternary trees.
A particular attention has been devoted to the area requirements of Euclidean minimum spanning trees. In their seminal paper on Euclidean minimum spanning trees, Monma and Suri~\cite{MonmaS92} proved that any tree of maximum degree $5$ admits a planar embedding as a Euclidean minimum spanning tree. Their algorithm, whose inductive construction is shown in Fig.~\ref{fig:monmasuri}, consists of placing the neighbors $r_i$ of the root $r$ of the tree on a circumference centered at $r$, of placing the neighbors of $r_i$ on a much smaller circumference centered at $r_i$, and so on. Monma and Suri~\cite{MonmaS92} proved that the area of the realizations constructed by their algorithm is $2^{\Omega(n^2)}$ and conjectured that exponential area is sometimes required to construct realizations of degree-$5$ trees as Euclidean minimum spanning trees.
\begin{figure}[htb]
\centering
\mbox{\epsfig{figure=MonmaSuri.eps,scale=0.35,clip=}}
\caption{An illustration of the algorithm of Monma and Suri to construct realizations of degree-$5$ trees as Euclidean minimum spanning trees.}
\label{fig:monmasuri}
\end{figure}
Frati and Kaufmann~\cite{fk-pabmemstrt-08} showed how to construct polynomial area realizations of degree-$4$ trees as Euclidean minimum spanning trees. Their technique consists of using a decomposition of the input tree $T$ (similar to the ones presented in Sect.s~\ref{se:straight-outerplanar} and~\ref{se:straight-trees}) in which a path $P$ is selected such that every subtree of $P$ has at most $n/2$ nodes. Euclidean minimum spanning tree realizations of such subtrees are then inductively constructed and placed together with a drawing of $P$ to get a drawing of $T$. Suitable angles and lengths for the edges in $P$ have to be chosen to ensure that the resulting drawing is a Euclidean minimum spanning tree realization of $T$. The sketched geometric construction is shown in Fig.~\ref{fig:mstupper}.
\begin{figure}[htb]
\centering
\mbox{\epsfig{figure=mstupper.eps,scale=0.4,clip=}}
\caption{An illustration of the algorithm of Frati and Kaufmann to construct polynomial-area realizations of degree-$4$ trees as Euclidean minimum spanning tree realizations.}
\label{fig:mstupper}
\end{figure}
Very recently, Angelini~\emph{et al.}~proved in~\cite{abcfks-aremst-11} that in fact there exist degree-$5$ trees requiring exponential area in any realization as a Euclidean minimum spanning tree. The tree $T^*$ exhibited by Angelini~\emph{et al.}, which is shown in Fig.~\ref{fig:treelowerbound}, consists of a degree-$5$ complete tree $T_c$ with a constant number of vertices and of a set of degree-$5$ caterpillars, each one attached to a distinct leaf of $T_c$. The complete tree $T_c$ forces the angles incident to an end-vertex of the backbone of at least one of the caterpillars to be very small, that is, between $60^{\degree}$ and $61^{\degree}$. Using this as a starting point, Angelini~\emph{et al.}~prove that each angle incident to a vertex of the caterpillar is either very small, that is, between $60^{\degree}$ and $61^{\degree}$, or is very large, that is, between $89.5^{\degree}$ and $90.5^{\degree}$. As a consequence, the lengths of the edges of the backbone of the caterpillar decrease exponentially along the caterpillar, thus obtaining the area bound. There is still some distance between the best known lower and upper bounds, hence the following is open:
\begin{figure}[htb]
\centering{
\mbox{\epsfig{figure=TreeLowerBound.eps,width=0.5\textwidth,height=5cm,clip=}}}
\caption{A tree $T^*$ requiring $2^{\Omega(n)}$ area in any Euclidean minimum spanning tree realization.}
\label{fig:treelowerbound}
\end{figure}
\begin{problem}
Close the gap between the $2^{O(n^2)}$ upper bound and the $2^{\Omega(n)}$ lower bound for the area requirements of Euclidean minimum spanning tree realizations.
\end{problem}
Greedy drawings are a kind of proximity drawings that recently attracted lot of attention, due to their application to network routing. Namely, consider a network in which each node $a$ that has to send a packet to some node $b$ forwards the packet to any node $c$ that is closer to $b$ than $a$ itself. If the position of any node $u$ is not its real geographic location, but rather the pair of coordinates of $u$ in a drawing $\Gamma$ of the network, it is easy to see that routing protocol never gets stuck if and only if $\Gamma$ is a greedy drawing. Greedy drawings were introduced by Rao \emph{et al.}~in~\cite{rpss-grwli-03}. A lot of attention has been devoted to a conjecture of~\cite{pr-crgr-05} stating that every triconnected planar graph has a greedy drawing. Dhandapani verified the conjecture for triangulations in~\cite{d-gdt-10}, and later Leighton and Moitra~\cite{lm-srgems-44} and independently Angelini~\emph{et al.}~\cite{afg-acgdt-10} completely settled the conjecture in the positive. The approach of Leighton and Moitra (the one of Angelini~\emph{et al.}~is amazingly similar) consists of finding a certain subgraph of the input triconnected planar graph, called a \emph{cactus graph}, and of constructing a drawing of the cactus by induction. Greedy drawings have been proved to exist for every graph if the coordinates are chosen in the hyperbolic plane~\cite{k-gruhs-07}. Research efforts have also been devoted to construct greedy drawings in small area. More precisely, because of the routing applications, attention has been devoted to the possibility of encoding the coordinates of a greedy drawing with a small number of bits. When this is possible, the drawing is called \emph{succinct}. Eppstein and Goodrich~\cite{eg-sgghp-09} and Goodrich and Strash~\cite{gs-sggrep-09} showed how to modify the algorithm of Kleinberg~\cite{k-gruhs-07} and the algorithm of Leighton and Moitra~\cite{lm-srgems-44}, respectively, in order to construct drawings in which the vertex coordinates are represented by a logarithmic number of bits. On the other hand, Angelini~\emph{et al.}~\cite{adf-sgddae-10} proved that there exist trees requiring exponential area in any greedy drawing (or equivalently requiring a polynomial number of bits to represent their Cartesian coordinates in the Euclidean plane). The following is however open:
\begin{problem}
Is it possible to construct greedy drawings of triconnected planar graphs in the Euclidean plane in polynomial area?
\end{problem}
Partially positive results on the mentioned open problem were achieved by He and Zhang, who proved in~\cite{hz-scgd3pg-11} that succinct convex \emph{weekly greedy} drawings exist for all triconnected planar graphs, where weekly greedy means that the distance between two vertices $u$ and $v$ in the drawing is not the usual Euclidean distance $D(u,v)$ but a function $H(u,v)$ such that $D(u,v) \leq H(u,v) \leq 2 \sqrt 2 D(u,v)$. On the other hand, Cao et al.~proved in~\cite{csz-sggrep-09} that there exist triconnected planar graphs requiring exponential area in any \emph{convex} greedy drawing in the Euclidean plane.
\section{Clustered Graph Drawings}\label{se:clustered}
In this section, we discuss algorithms and bounds for constructing small-area $c$-planar drawings of clustered graphs. Table~\ref{ta:clustered} summarizes the best known area bounds for $c$-planar straight-line drawings of clustered graphs.
\begin{table}[!htb]\footnotesize
\centering
\linespread{1.2}
\selectfont
\begin{tabular}{|c|c|c|c|c|}
\cline{2-5}
\multicolumn{1}{c|}{} & \emph{Upper Bound} & \emph{Refs.} & \emph{Lower Bound} & \emph{Refs.}\\
\hline
{\em Clustered Graphs} & $O(c^n)$ & \cite{EadesFLN06,afk-srdcg-11} & $\Omega(b^n)$ & \cite{cocoon/FengCE95}\\
\hline
{\em $c$-Connected Trees} & $O(n^2)$ & \cite{BattistaDF07} & $\Omega(n^2)$ & \cite{BattistaDF07} \\
\hline
{\em Non-$c$-Connected Trees} & $O(c^n)$ & \cite{EadesFLN06,afk-srdcg-11} & $\Omega(b^n)$ & \cite{BattistaDF07} \\
\hline
\end{tabular}
\vspace{2mm}
\caption{\small A table summarizing the area requirements for $c$-planar straight-line drawings of clustered graphs in which clusters are convex regions; $b$ and $c$ denote constants greater than $1$.}
\label{ta:clustered}
\end{table}
Given a clustered graph, testing whether it admits a $c$-planar drawing is a problem of unknown complexity, and is perhaps the most studied problem in the Graph Drawing community during the last ten years~\cite{FengCE95,feng97algorithms,Dahlhaus98,GutwengerJLMPW02,GoodrichLS05,CorteseBPP05,cdpp-cccc-05,CornelsenW06,JelinkovaKKPSV07,df-ectefcgsf-07,cdfpp-cccg-j-08,JelinekSTV08,JelinekJKL08,afp-scgcp-09}.
Suppose that a $c$-planar clustered graph $C$ is given together with a $c$-planar embedding. How can the graph be drawn? Such a problem has been intensively studied in the literature and a number of papers have been presented for constructing $c$-planar drawings of $c$-planar clustered graphs within many drawing conventions.
Eades \emph{et al.}~show in~\cite{EadesFLN06} an algorithm for constructing $c$-planar straight-line drawings of $c$-planar clustered graphs in which each cluster is drawn as a convex region. Such a result is achieved by first studying how to construct planar straight-line drawings of hierarchical graphs. A \emph{hierarchical graph} is a graph such that each vertex $v$ is assigned a number $y(v)$, called the \emph{layer} of $v$; a drawing of a hierarchical graph has to place each vertex $v$ on the horizontal line $y=y(v)$. Eades \emph{et al.}~show an inductive algorithm to construct a planar straight-line drawing of any hierarchical-planar graph. Second, Eades \emph{et al.}~show how to turn a $c$-planar clustered graph $C$ into a hierarchical graph $H$ such that, for each cluster $\mu$ in $C$, all the vertices in $\mu$ appear in consecutive layers of the hierarchy. This implies that, once a planar straight-line drawing of $H$ has been constructed, as in Fig.~\ref{fig:clustered-convex}(a), each cluster $\mu$ can be drawn as a region surrounding the convex hull of the vertices in $\mu$, resulting in a straight-line $c$-planar drawing of $C$ in which each cluster is drawn as a convex region, as in Fig.~\ref{fig:clustered-convex}(b).
\begin{figure}[htb]
\centering
\begin{tabular}{c c}
\mbox{\epsfig{figure=Layered.eps,scale=0.3,clip=}} \hspace{8mm} &
\mbox{\epsfig{figure=LayeredToClustered.eps,scale=0.3,clip=}}\\
(a) \hspace{8mm} & (b)\\
\end{tabular}
\caption{(a) A planar straight-line drawing of a hierarchical graph $H$. Graph $H$ is obtained from a clustered graph $C$ by assigning consecutive layers to vertices of the same cluster. (b) A straight-line $c$-planar drawing of $C$.}
\label{fig:clustered-convex}
\end{figure}
Angelini \emph{et al.}, improving upon the described result of Eades \emph{et al.}~in~\cite{EadesFLN06} and answering a question posed in~\cite{EadesFLN06}, show in~\cite{afk-srdcg-11} an algorithm for constructing a \emph{straight-line rectangular drawing} of any clustered graph $C$, that is, a $c$-planar straight-line drawing of $C$ in which each cluster is drawn as an axis-parallel rectangle (more in general, the algorithm of Angelini \emph{et al.}~constructs straight-line $c$-planar drawings in which each cluster is an arbitrary convex shape). The algorithm of Angelini \emph{et al.}~is reminiscent of F\'ary's algorithm (see \cite{f-srpg-48} and Sect.~\ref{se:straight-planar}). Namely, the algorithm turns a clustered graph $C$ into a smaller clustered graph $C'$ by either removing a cluster, or splitting $C$ in correspondence of a separating $3$-cycle, or contracting an edge of $C$. A straight-line rectangular drawing of $C'$ can then be inductively constructed and easily augmented to a straight-line rectangular drawing of $C$. When none of the inductive cases applies, the clustered graph is an \emph{outerclustered graph}, that is, every cluster contains a vertex incident to the outer face (see Fig.~\ref{fig:clustered-rectangular}(a)). In order to draw an outerclustered graph $C$, Angelini \emph{et al.} show how to split $C$ into three \emph{linearly-ordered outerclustered graphs}, that are outerclustered graphs such that the graph induced by the ``direct containment'' relationship among clusters is a path (see Fig.~\ref{fig:clustered-rectangular}(b)), where a cluster $\mu$ \emph{directly contains} a cluster $\nu$ if $\mu$ contains $\nu$ and $\mu$ contains no cluster $\rho$ containing $\nu$. Moreover, they show how to combine the drawings of such graphs to get a straight-line rectangular drawing of $C$. Finally, Angelini \emph{et al.}~show an inductive algorithm for constructing a straight-line rectangular drawing of any linearly-ordered outerclustered graphs $C$. Such an algorithm finds a subgraph of $C$ (a path plus an edge) that splits $G$ into smaller linearly-ordered outerclustered graphs, inductively draws such subgraphs and combines their drawings to get a straight-line rectangular drawing of $C$.
\begin{figure}[htb]
\centering
\begin{tabular}{c c}
\mbox{\epsfig{figure=outerclusteredSimple.eps,scale=0.38,clip=}} \hspace{8mm} &
\mbox{\epsfig{figure=linearly-ordered.eps,scale=0.25,clip=}}\\
(a) \hspace{8mm} & (b)\\
\end{tabular}
\caption{(a) An outerclustered graph. (b) A linearly-ordered outerclustered graph. Any two consecutive clusters in the sequence $\mu_1,\dots,\mu_{12}$ are one the parent of the other.}
\label{fig:clustered-rectangular}
\end{figure}
Both the algorithm of Eades \emph{et al.}~and the algorithm of Angelini \emph{et al.}~construct drawings requiring, in general, exponential area. However, Feng \emph{et al.}~proved in~\cite{cocoon/FengCE95} that there exists a clustered graph $C$ requiring exponential area in any straight-line $c$-planar drawing in which the clusters are represented by convex regions. The proof of such a lower bound is strongly based on the proof of Di Battista \emph{et al.}~that there exist directed graphs requiring exponential area in any upward straight-line drawing (see~\cite{BattistaTT92} and Sect.~\ref{se:upward}). Eades \emph{et al.}~showed in~\cite{EadesFN99} how to construct $O(n^2)$ area $c$-planar orthogonal drawings of clustered graphs with maximum degree $4$; the authors first construct a visibility representation of the given clustered graph and then turn such a representation into an orthogonal drawing. Di Battista \emph{et al.}~\cite{BattistaDF07} show algorithms for drawing clustered trees in small area. In particular, they show an inductive algorithm to construct straight-line rectangular drawings of $c$-connected clustered trees in $O(n^2)$ area; however, they prove that there exist non-$c$-connected trees requiring exponential area in any straight-line drawing in which the clusters are represented by convex regions, again using the tools designed by Di Battista \emph{et al.}~in~\cite{BattistaTT92}. The following problem has been left open by Di Battista \emph{et al.}~\cite{BattistaTT92}.
\begin{problem}
What are the area requirements of order-preserving straight-line $c$-planar drawings of clustered trees in which clusters are represented by convex regions?
\end{problem}
|
\section{Introduction}
Multicomponent quantum gases are fascinating systems~\cite{multicomponent1}.
Basic research in this area has enormously grown in the last few years~\cite{multicomponent2}. Due to the ability of optically trapping and cooling
gases to extremely low temperatures, it is possible to study different
phenomena in bosonic~\cite{Andrews,bosonicmixtures} as well as fermionic
mixtures~\cite{fermions}. Important quantum effects like Bose-Einstein
condensation (BEC) and superconductivity can now be studied in a very
controlled way in multicomponent atomic systems.
Interesting experiments with mixed bosonic quantum fluids have been done by
simultaneously trapping $^{87}Rb$ atoms in two different hyperfine states~\cite{Myatt,Hall-1, Hall-2, Lin}. The relative population is reached by
applying a coupling field characterized by a Rabi frequency $\Omega_R$ and a
detuning $\nu$ with respect to the spacing between the energy levels of the
two hyperfine states. In this way, it is possible to transfer atoms from one
hyperfine state to the other, producing a Josephson-type interaction between
species~\cite{Joseph-Exp.1,Legget-Sols-1,Legget-Sols-2}.
In general, the name \textquotedblleft Josephson
interaction\textquotedblright\ refers to the interaction of a large number
of bosonic degrees of freedom allowed to occupy two different quantum
states. Although it was originally proposed in superconductor systems~\cite
{Josephson}, where the bosons are Cooper pairs, there are many other systems
where this effect shows up. A review covering different physical systems can
be found in Ref.~[\cite{Barone}]. We can distinguish two types of
Josephson effects~\cite{Leggett-Review}: the so-called \textquotedblleft
external\textquotedblright , where the two states are spatially separated,
like, for instance, in BEC trapped in a double-well potential~\cite
{Albiez,doublewell,Gati,Levy}, or the \textquotedblleft
internal\textquotedblright , where the two bosonic states are
interpenetrated, without geometrical distinction, like, for instance, the
experiments in Refs.~\cite{Hall-1, Hall-2}. In this paper, we are mainly
interested in the latter case of internal Josephson-type interactions.
Static and dynamical properties of binary bosonic mixtures in different trap
geometries have been studied theoretically by essentially using
Gross-Pitaevskii equations~\cite{GPMixtures,Smerzi-1,Smerzi-2,Salerno,Bruno-1,Bruno-2,Bruno-3}.
Moreover, to study properties of uniform condensates, especially those
issues related with fluctuations, such as symmetry restoration, reentrances,
etc., quantum field theory at finite density and
temperature~\cite{Rudnei-1,Rudnei-2,Rudnei-3,BarciFragaRamos}
is a useful technique. Related models, such as $O(N)$ models, have also been
extensively studied by using large-$N$ approximation and
renormalization-group techniques~\cite{largeN-1,largeN-2}. These papers are
mostly concentrated in multicomponent systems which conserve the particle
number of each species independently.
Motivated by these results, we decided to address the effect of
Josephson-like interactions in uniform bosonic mixtures. For simplicity, we
have considered an $O(2)$ model, perturbed with an explicit
symmetry-breaking term parametrized by the Rabi frequency $\Omega _{R}$ and
the detuning term $\nu $. This model is analyzed in mean-field approximation
plus Gaussian fluctuations.
In the absence of Josephson interactions, this model is at the onset of phase separation, since the two species are not physically distinguishable. However, the presence of Josephson interactions changes this scenario since it explicitly breaks $O(2)$ symmetry. There is a temperature regime where
the two atomic species uniformly condensate at the same critical temperature $T_{c}$
and their relative phase is locked by the phase of the applied electromagnetic field responsible for the Rabi coupling and the detuning. The relative population of each condensate strongly depends on the ratio $\Omega _{R}/|\nu |$. The main results of this paper are shown in Figures~(\ref{fig.nu}) and~(\ref{fig.Omega}) where we depict the condensate fraction of the two species as a function of temperature for different values of the parameter $\Omega _{R}/|\nu |$. Thus, controlling the external laser parameters, {\em i.e.}, the Rabi coupling, the laser frequency (essentially the detuning) and
the phase, it is possible to control each one of the condensate fractions as well as its phase difference.
An important result is that, due
to the original $O(2)$ symmetry, the effective Rabi frequency, given by
$\Omega _{\mathrm{eff}}=\sqrt{\Omega _{R}^{2}+|\nu |^{2}}$ is strongly
renormalized by thermal fluctuations. On the other hand, the ratio $\Omega _{R}/|\nu |$, that controls the bosonic mixture,
remains unaffected by quantum as well as thermal fluctuations. Thus, the ratio between both condensates are temperature independent, allowing the possibility of control the relative condensate fractions with high accuracy.
The paper is organized as follows. In section \ref{QFT}, we describe a
general model for a binary mixture using quantum field theory language. In
section~\ref{O2}, we concentrate on the $O(2)$ model perturbed with
Josephson interactions. In section \ref{MF}, we present the mean-field
solution, while in section~\ref{Fluctuations} we analyze the effect of
fluctuations. Numerical results are presented in section \ref{Numerics} and,
finally, we discuss our results in section~\ref{discussion}. We reserve a brief appendix
\ref{app} to describe the definitions of Rabi frequency and detuning parameter used to built our
model.
\section{A quantum field theory for binary bosonic mixtures}
\label{QFT}
We will consider two bosonic species described by two complex fields, $\phi (\vec{x},t)$ and $\psi (\vec{x},t)$. The model is defined by the action
\begin{equation}
S=\int d^{3}xdt\;\left\{ \mathcal{L}_{\psi }+\mathcal{L}_{\phi }+\mathcal{L}
_{I}\right\} , \label{S}
\end{equation}
where $\mathcal{L_{\psi }}$ and $\mathcal{L_{\phi }}$ are the
non-relativistic quadratic Lagrangian densities
\begin{eqnarray}
\mathcal{L_{\psi }} &=&\psi ^{\ast }\left( i\partial _{t}+\frac{\nabla ^{2}}{
2m}+\mu _{\psi }\right) \psi \;, \label{Lpsi} \\
\mathcal{L_{\phi }} &=&\phi ^{\ast }\left( i\partial _{t}+\frac{\nabla ^{2}}{
2m}+\mu _{\phi }\right) \phi \;. \label{Lphi}
\end{eqnarray}
$\mu_\psi$ and $\mu_\phi$ are the chemical potentials for the $\psi$ and $\phi$ species, respectively. We choose the same mass $m$ for both species, since
we are interested in mixtures composed by a single element in two different
hyperfine states.
It is convenient to parametrize the chemical potentials as
\begin{eqnarray}
\mu _{\phi }&=&\mu +\Omega _{R}
\label{muphi} \\
\mu _{\psi }&=&\mu -\Omega _{R}
\label{mupsi}
\end{eqnarray}
The parameter $\mu$ controls the overall particle
density at the time that the Rabi frequency $\Omega _{R}$ controls the
population imbalance (see Appendix~\ref{app} for the microscopic physical meaning
of $\Omega_R$).
Throughout the paper, we have used a unit system in
which $\hbar =1$.
The interaction Lagrangian density $\mathcal{L}_{I}$ can be split into two
terms,
\begin{equation}
\mathcal{L}_{I}=\mathcal{L}_{c}+\mathcal{L}_{J}\;. \label{LI}
\end{equation}
The first term, $\mathcal{L}_{c}$, contains two-body interactions that
preserve the particle number of each species individually. For diluted
gases, it can be approximated as a local quartic polynomial of the form
\begin{equation}
\mathcal{L}_{c}=-\frac{g_{\psi }}{2}\left( \psi ^{\ast }\psi \right) ^{2}-
\frac{g_{\phi }}{2}\left( \phi ^{\ast }\phi \right) ^{2}-g_{\phi \psi }\psi
^{\ast }\psi \phi ^{\ast }\phi , \label{Lc}
\end{equation}
where the coupling constants $g_{\psi }=4\pi a_{\psi }/m$, $g_{\phi }=4\pi
a_{\phi }/m$ and $g_{\phi \psi }=8\pi a_{\phi \psi }/m$ are written in terms
of the intraspecies s-wave scattering lengths $a_{\psi }$, $a_{\phi }$ and
the interspecies s-wave scattering length $a_{\phi \psi }$. Note that this
interaction term is invariant under $U(1)_{\phi }\otimes U(1)_{\psi }$
transformations.
The second term of Eq.~(\ref{LI}) does not conserve the particle number of
each species individually. It conserves, however, the total particle number.
This term explicitly breaks the symmetry of Eq.~(\ref{Lc}) as $U(1)_{\phi
}\otimes U(1)_{\psi }\rightarrow U(1)_{\phi +\psi }$. We generally call
these terms as Josephson interactions, since they couple the phases of each
bosonic component. The simplest terms can be written as
\begin{equation}
\mathcal{L}_{J}=\nu \psi ^{\ast }\phi +\nu ^{\ast }\phi ^{\ast }\psi -\frac{
g_{J}}{2}\left( \psi ^{\ast }\psi ^{\ast }\phi \phi +\phi ^{\ast }\phi
^{\ast }\psi \psi \right) . \label{LJ}
\end{equation}
The quadratic term, proportional to $\nu $, and the quartic two-body
interaction term have, in general, very different origins. The one-particle
term is proportional to the detuning $\nu $, where
we have considered a complex parameter in such a way to control the relative phases
of the condensates (see Appendix~\ref{app} for its definition). Considering the two species as components of an isospin
doublet, this term arises like an effective spin-orbit interaction~\cite{Garcia-March,DanWei}.
We could also consider one-body terms of this type with derivative
couplings. However, to keep matters as simple as
possible, we will consider only this term. The second term in Eq.~(\ref{LJ})
represents scattering processes in which the internal hyperfine state of the
atoms is not conserved. In the absence of $\nu $, these processes are
unlikely to occur, since both hyperfine states are energetically well
separated. However, in the presence of a laser with small detuning between
the frequency differences, a very small coupling constant $g_{J}$ could
produce qualitatively different results.
Some aspects of the phase diagram of the model of Eqs.~(\ref{Lpsi}), (\ref{Lphi}) and~(\ref{Lc}), \emph{without Josephson couplings} ($\mathcal{L}_{J}=0$),
have been previously studied. The zero-temperature
mean-field analysis clearly establishes three different regimes, depending on relations between intra and inter species coupling constants. If
\begin{equation}
g_{\phi }g_{\psi }-g_{\psi \phi }^{2}>0 \ ,
\label{constraint}
\end{equation}
it is possible to have two coexisting condensates~\cite{Rudnei-1}.
Conversely, if
\begin{equation}
g_{\phi }g_{\psi }-g_{\psi \phi }^{2}<0
\label{constraint}
\end{equation}
both condensates cannot coexist and they
tend to spatially separate, producing an inhomogeneous state~\cite{exp-PS}.
In addition, there is a special intermediate regime,
\begin{equation}
g_{\phi }g_{\psi }-g_{\psi \phi }^{2}=0\; ,
\label{eq:intermediate}
\end{equation}
that could be considered as the onset of homogeneous instability, since it is a fine tune region at the transition between the homogeneous and the inhomogeneous ground states.
Although it could be very difficult to experimentally reach this regime, it is a very interesting one due to its symmetry properties, as we will describe in the next section.
\section{$O(2)$ model with Josephson anisotropy}
\label{O2}
The model described in the preceding section has a very rich phase diagram
depending on the relative values of the coupling constants and of the
temperature. However, there is a special point of maximum symmetry where the
analysis gets simpler. Let us analyze model~(\ref{S}-\ref{LJ}) in its
maximum symmetry point given by $g_{\phi ,\psi }=g_{\phi }=g_{\psi }=g$, $\Omega_{R}=0$, $\nu =0$ and $g_{J}=0$.
This point is at the intermediate regime described by Eq.~(\ref{eq:intermediate}).
The interaction term,
Eq.~(\ref{Lc}), takes the simpler form
\begin{equation}
\mathcal{L}_{c}=-\frac{g}{2}\left( \psi ^{\ast }\psi +\phi ^{\ast }\phi
\right) ^{2}.
\label{Lg}
\end{equation}
In addition to the $U(1)_{\phi }\otimes U(1)_{\psi }$ phase symmetry,
there is an emergent $O(2)$ symmetry, corresponding with rotations in the
isospin space $(\phi ,\psi )$. Thus, on the one hand, the particle number of each
species is independently conserved. On the other hand, the two species are physically
indistinguishable since any isospin rotation mixing the two species has
exactly the same action. Thus, the question of the difference between homogeneous and inhomogeneous phases has no real meaning at this point. However, an infinitesimal deviation of the coupling constants leads the systems to one or to the other phase, depending on whether $g_{\phi\psi}<g$ or $g_{\phi\psi}>g$. It is in this sense that we say that the $O(2)$ model, $g_{\phi\psi}=g$, is at the onset of the homogeneous instability. It is interesting to note that we can rewrite the model in terms of the real and imaginary components of the fields ($\Re\psi,\Im\psi, \Re\phi,\Im\phi$). In this representation, it is completely equivalent to a four-vector model with $O(4)$ symmetry, which has been extensively studied in the literature related with the Chiral QCD phase transition~\cite{Wilczek} and, more recently, in the context of Bose-Einstein condensates~\cite{Kleinert}.
Next, we minimally break the $O(2)$ symmetry unbalancing the chemical potentials
with a term proportional to $\Omega _{R}$ and a Josephson term of the form
\begin{equation}
\mathcal{L}_{\nu }=\nu \psi ^{\ast }\phi +\nu ^{\ast }\phi ^{\ast }\psi .
\label{Lnu}
\end{equation}
For simplicity, we ignore two-body Josephson interactions (given by the term
proportional to $g_{J}$ in Eq.~(\ref{LJ})), since, in principle, it is of
higher order than the one-body interaction term we are considering.
The structure of this model is clearly visualized by defining new fields $(\varphi _{1},\varphi _{2})$ obtained by an isospin rotation of the original
fields $(\phi ,\psi )$,
\begin{equation}
\left(
\begin{array}{c}
\varphi _{1} \\
\varphi _{2}
\end{array}
\right) =M\left(
\begin{array}{c}
\phi \\
\psi
\end{array}
\right) ,
\end{equation}
where the rotation matrix is
\begin{equation}
M=\frac{1}{D}\left(
\begin{array}{cc}
\Omega _{\mathrm{eff}}-\Omega _{R} & -\nu \\
& \\
\nu ^{\ast } & \Omega _{\mathrm{eff}}-\Omega _{R}
\end{array}
\right) , \label{M}
\end{equation}
with
\begin{equation}
D=\sqrt{\left( \Omega _{\mathrm{eff}}-\Omega _{R}\right) ^{2}+|\nu |^{2}}\; ,
\end{equation}
and $\Omega _{\mathrm{eff}}=\sqrt{\Omega _{R}^{2}+|\nu |^{2}}$ is called the \emph{effective Rabi frequency} (see appendix~{\ref{app}). Of course,
one can immediately check that $\det (M)=1$. With this transformation, the
Lagrangian density takes the form
\begin{eqnarray}
\mathcal{L} &=&\varphi _{1}^{\ast }\left( i\partial _{t}+\frac{\nabla ^{2}}{2m}+\mu _{+}\right) \varphi _{1} + \varphi _{2}^{\ast }\left( i\partial _{t}+\frac{\nabla ^{2}}{2m}+\mu_{-}\right) \varphi _{2} \nonumber \\
& & - \ \frac{g}{2}\left( \varphi _{1}^{\ast }\varphi _{1}+\varphi _{2}^{\ast}\varphi _{2}\right) ^{2},
\label{Lvarphi}
\end{eqnarray}
where
\begin{eqnarray}
\mu _{+}&=&\mu +\Omega _{\mathrm{eff}},
\label{eq:mu+}\\
\mu _{-}&=&\mu -\Omega _{\mathrm{eff}}.
\label{eq:mu-}
\end{eqnarray}
We see that, while
terms proportional to $\nu $ and $\Omega _{R}$ break the
$O(2)$ and $U(1)_{\phi }\otimes U(1)_{\psi }$ symmetries, the system still
has an $U(1)_{\varphi _{1}}\otimes U(1)_{\varphi _{2}}$ symmetry in the new
variables. Thus, there is a direction in isospin space in which the particle
number of both species is still conserved independently. Equations~(\ref{eq:mu+}) and~(\ref{eq:mu-}), that define the chemical potentials in the new basis, are quite similar with
Eqs.~(\ref{muphi}) and~(\ref{mupsi}) for the chemical potentials of $\psi$ and $\phi$, with the difference that the Rabi frequency, $\Omega_R$ in the former case, should be substituted by the effective Rabi frequency, $\Omega_{\rm eff}$, in the latter. This simple behavior is a consequence of the $O(2)$ symmetry of the two-body interaction term, Eq.~(\ref{Lg}). It is not difficult to realize that, if we fix the coupling constants slightly away from the maximal symmetry point, $g_{\psi \phi }\neq g$, a term proportional to $\varphi _{1}\varphi _{1}\varphi _{2}^{\ast }\varphi _{2}^{\ast }$ would be
generated upon an isospin rotation, breaking in this way $U(1)_{\varphi_{1}}\otimes U(1)_{\varphi _{2}}\rightarrow U(1)$. In this sense, the model
of Eq.~(\ref{Lvarphi}) implements a minimal perturbation of the complex $O(2)$ model.
Interestingly, Eq.~(\ref{Lvarphi}) does not depend on $\Omega _{R}$ and $\nu$ independently, but only depends on the effective Rabi frequency $\Omega _{\mathrm{eff}}=\sqrt{\Omega _{R}^{2}+|\nu |^{2}}$ (see Appendix~\ref{app} to see the relevance of the effective Rabi frequency in a simpler case of a two-level system). On the other hand, the
rotation matrix of Eq.~(\ref{M}) depends only on the ratio $\Omega _{R}/|\nu|$. It is instructive to see the form of the rotation matrix in two
different limits.
Let us consider, for instance, $|\nu |\ll \Omega _{R}$. In this case,
\begin{equation}
M=\left(
\begin{array}{cc}
\frac{|\nu |}{2\Omega _{R}} & -e^{i\alpha} \\
& \\
e^{-i\alpha} & \frac{|\nu |}{2\Omega _{R}}
\end{array}
\right) ,
\end{equation}
where we have defined $\nu =|\nu |\exp (i\alpha)$. In the extreme
limit of $\nu \rightarrow 0$, both species are decoupled, as expected,
and the mixture is proportional to $\frac{|\nu |}{2\Omega _{R}}+O((|\nu|/2\Omega _{R})^{2})$.
In the opposite limit, $|\nu |\gg \Omega _{R}$,
\begin{equation}
M=\frac{1}{\sqrt{2}}\left(
\begin{array}{cc}
1-\frac{\Omega _{R}}{2|\nu |} & -e^{i\alpha}\left( 1+\frac{\Omega
_{R}}{2|\nu |}\right) \\
& \\
e^{-i\alpha}\left( 1+\frac{\Omega _{R}}{2|\nu |}\right) & 1-\frac{\Omega _{R}}{2|\nu |}
\end{array}
\right) .
\end{equation}
In the extreme limit, $\Omega _{R}\rightarrow 0$, the fields are symmetrically
superposed, depending just on the phase of the detuning parameter,
\begin{eqnarray}
\varphi _{1} &=&\frac{1}{\sqrt{2}}\left( \phi -e^{i\alpha}\psi\right) , \\
\varphi _{2} &=&\frac{1}{\sqrt{2}}\left( e^{-i\alpha}\phi +\psi\right) .
\end{eqnarray}
Small values of $\Omega _{R}$ produce corrections of order $\Omega _{R}/|\nu|$.
\section{Mean-field approximation}
\label{MF}
Let us analyze the model of Eq.~(\ref{Lvarphi}) in the mean-field
approximation. Minimizing the action $S=\int dtd^3x {\cal L}$ with ${\cal L}$ given by Eq.~(\ref{Lvarphi}), we obtain the equations of motion analogous to the Gross-Pitaevskii equations
\begin{eqnarray}
\left( i\partial _{t}+\frac{\nabla ^{2}}{2m}+\mu _{+}-g\left( \varphi
_{1}^{\ast }\varphi _{1}+\varphi _{2}^{\ast }\varphi _{2}\right) \right)
\varphi _{1} &=&0,
\label{meanfield1} \\
\left( i\partial _{t}+\frac{\nabla ^{2}}{2m}+\mu _{-}-g\left( \varphi
_{1}^{\ast }\varphi _{1}+\varphi _{2}^{\ast }\varphi _{2}\right) \right)
\varphi _{2} &=&0.
\label{meanfield2}
\end{eqnarray}
Looking for uniform and static solutions $\varphi _{1,2}(x,t)\equiv \varphi
_{1,2}^{0}$ we have
\begin{eqnarray}
\left( \mu _{+}-g\left[ |\varphi _{1}^{0}|^{2}+|\varphi _{2}^{0}|^{2}\right]
\right) \varphi _{1}^{0} &=&0, \label{MF1} \\
\left( \mu _{-}-g\left[ |\varphi _{1}^{0}|^{2}+|\varphi _{2}^{0}|^{2}\right]
\right) \varphi_{2}^{0} &=&0. \label{MF2}
\end{eqnarray}
Assuming that $\varphi_{1,2}^{0}\neq 0$, we can subtract Eq.~(\ref{MF2})
from Eq.~(\ref{MF1}), obtaining $\Delta \mu =\mu_{+}-\mu_{-}=0$.
Therefore, the two fields $\varphi _{1,2}$ cannot condensate simultaneously,
since a solution $\varphi_{1,2}^{0}\neq 0$ does not exist, except in the
case $\Delta \mu =2\Omega_{\mathrm{eff}}=0$. Instead, we have two possible
solutions,
\begin{equation}
\varphi_{1}^{0}=0\;\;\;\;\;\;\;\;,\;\;\;\;\;\;\;|\varphi_{2}^{0}|^{2}=\mu_{-}/g
\end{equation}
or
\begin{equation}
|\varphi_{1}^{0}|^{2}=\mu_{+}/g\;\;\;\;\;\;\;\;,\;\;\;\;\;\;\;\varphi_{2}^{0}=0 \ .
\label{phi2=0}
\end{equation}
Let us consider the solution $\varphi _{2}^{0}=0$, Eq.~(\ref{phi2=0}). Using
the matrix $M^{-1}$, given by the inverse of Eq.~(\ref{M}), it is simple to
turn back to the original fields, obtaining
\begin{eqnarray}
\phi _{0} &=&\frac{\Omega_{\mathrm{eff}}-\Omega_{R}}{\sqrt{\left( \Omega_{\mathrm{eff}}-\Omega _{R}\right) ^{2}+|\nu |^{2}}}\;\varphi _{1}^{0}\;,
\label{phi0} \\
\psi _{0} &=&-\frac{\nu ^{\ast }}{\sqrt{\left( \Omega _{\mathrm{eff}}-\Omega_{R}\right) ^{2}+|\nu |^{2}}}\;\varphi _{1}^{0}\;,
\label{psi0}
\end{eqnarray}
where $\phi_{0}$ and $\psi_{0}$ are the condensate amplitudes of the
fields $\phi (x)$ and $\psi (x)$, respectively. The first observation is
that the two original species $\phi $ and $\psi $ condense simultaneously
and the relative phase between these condensates, $\Delta \alpha $, is fixed
by the phase of the parameter $\nu $,
\begin{equation}
\Delta \alpha =\alpha+\pi \ .
\end{equation}
At this point, it is important to emphasize this mean-field result. In the absence of
Josephson interactions, the two species $\psi$ and $\phi$
cannot be distinguished from each other. In fact, the order parameter in this case is
$|\varphi_1^0|^2+|\varphi_2^0|^2=|\psi_0|^2+|\phi_0|^2$, which is invariant under $O(2)$ transformations.
The presence of Josephson interactions changes this situation since it breaks the $O(2)$ symmetry.
\begin{figure}[tbp]
\includegraphics[height=5.5 cm]{Rabi.eps}
\caption{Effective Rabi frequency $\Omega_{\rm eff}=\sqrt{\Omega_R^2+|\nu|^2}$ as parametrized by equations~(\ref{Omegatheta}) and~(\ref{nutheta}). While $\Omega_{\rm eff}$ is strongly temperature dependent, as shown in Eq.~(\ref{mu+mu-}), the angle given by $\tan\theta=\Omega_R/|\nu|$ is not affected by thermal fluctuations.}
\label{fig.Rabi}
\end{figure}
Moreover, the condensate fraction of both species depends on the ratio
$\Omega _{R}/|\nu |$. It is instructive to parametrize $\Omega _{R}$ and
$|\nu |$ in the following way (as shown in Fig.~(\ref{fig.Rabi})),
\begin{eqnarray}
\Omega _{R} &=&\Omega_{\rm eff}\sin \theta , \label{Omegatheta} \\
|\nu | &=&\Omega_{\rm eff}\cos \theta ,
\label{nutheta}
\end{eqnarray}
with $0\leq \theta \leq \pi /2$. In terms of this parametrization, the ratio
between the condensate densities takes the form
\begin{equation}
\frac{|\phi _{0}|^{2}}{|\psi _{0}|^{2}}=\sec ^{2}\theta \left( 1-\sin \theta
\right) ^{2}, \label{phi/psi}
\end{equation}
which does not depend on $\Omega_{\rm eff} $ but only on $\tan \theta =\Omega
_{R}/|\nu |$. We depict this function in Fig.~(\ref{fig.theta}).
\begin{figure}[ht]
\includegraphics[height=5.5 cm]{fig1.eps}
\caption{$|\protect\phi _{0}|^{2}/|\protect\psi _{0}|^{2}$ given by
Eq.~(\protect\ref{phi/psi}) as a function of $\protect\theta $, where $\tan\theta=\Omega_R/|\nu|$.}
\label{fig.theta}
\end{figure}
For $\theta \rightarrow 0$ or $\Omega _{R}\rightarrow 0$ with $|\nu |\neq 0$, both condensates have essentially the same fraction. On the other hand, for $\theta \rightarrow \pi /2$ or $|\nu |\rightarrow 0$ with $\Omega
_{R}\neq 0$, only one of the fields condensates. We will show in the next section that, while $\Omega_{\rm eff}$ is renormalized by temperature, the present result is temperature independent.
\section{Effect of Fluctuations}
\label{Fluctuations}
To study thermal as well as quantum fluctuations, we start by considering
the following Euclidean ($\tau=i t$) finite temperature field theory:
\begin{eqnarray}
& & S_{\mathrm{E}}(\beta ) =\int_{0}^{\beta }d\tau \int d^{3}x\left[ \varphi_{1}^{\ast }\left( \partial _{\tau }-\frac{\nabla ^{2}}{2m}-\mu _{+}\right) \varphi _{1}\right. \nonumber \\
&+& \left. \varphi _{2}^{\ast }\left( \partial _{\tau }-\frac{\nabla ^{2}}{2m}-\mu_{-}\right) \varphi _{2} + \frac{g}{2}\left( \varphi _{1}^{\ast }\varphi _{1}+\varphi_{2}^{\ast }\varphi _{2}\right) ^{2}\right]
\label{SE}
\end{eqnarray}
with $\beta =1/T$. The partition function reads
\begin{eqnarray}
Z(\beta ,\vec{J}) &=&\int \mathcal{D}\varphi _{1}\mathcal{D}\varphi_{1}^{\ast }\mathcal{D}\varphi _{2}\mathcal{D}\varphi _{2}^{\ast }\;e^{-S_{\mathrm{E}}+\int d^{3}xd\tau \vec{J}\cdot \vec{\varphi}} \nonumber \\
&=&e^{-\beta VW[\beta ,J]}, \label{Z}
\end{eqnarray}
where we have introduced a source $\vec{J}$ in order to compute field
correlation functions. The functional integration measure implicitly
contains the cyclic bosonic boundary condition in Euclidean time, $\varphi
_{1,2}(0,x)=\varphi _{1,2}(\beta ,x)$. $W[\beta ,\vec{J}]=-\frac{1}{\beta V}
\ln Z$ is the Helmholtz free energy density.
The main purpose of this section is to compute $W[\beta ,J]$ in mean-field
approximation plus Gaussian fluctuations.
We expect that, at least in a certain temperature range, fluctuations will not change the general mean-field structure. With
this in mind, in order to compute $W[\beta ,J]$ we replace in Eq.~(\ref{Z})
the following decomposition
\begin{eqnarray}
\varphi _{1}(x,\tau ) &=&\varphi _{1}^{0}+\tilde{\varphi}_{1}(x,\tau ) \label{decomposition1} \\
\varphi _{2}(x,\tau ) &=&\tilde{\varphi}_{2}(x,\tau ) \label{decomposition2}
\end{eqnarray}
in which $\int d^{3}x\;\tilde{\varphi}_{1,2}=0$ and $\varphi _{1}^{0}(J)$ is
a solution of the mean field equations,
\begin{eqnarray}
\left. \frac{\delta S_{E}}{\delta \varphi _{1}}\right\vert _{\varphi_{1}=\varphi _{1}^{0},\varphi _{2}=0} &=&J\;, \\
\left. \frac{\delta S_{E}}{\delta \varphi _{2}}\right\vert _{\varphi_{1}=\varphi _{1}^{0},\varphi _{2}=0} &=&0\;,
\end{eqnarray}
where we have chosen a constant source $\vec{J}$, pointing in the $\varphi _{1}$ direction.
Retaining up to second-order terms in the fluctuations we obtain
\begin{equation}
Z(\beta )=e^{-\beta VU_{0}(\varphi _{1}^{0})}\int \left[ \mathcal{D}\tilde{\varphi}\right] \;e^{-\int d\tau d^{3}x\sum_{ij}\tilde{\varphi}_{i}^{\ast}S_{ij}^{(2)}\tilde{\varphi}_{j}}\;,
\end{equation}
where
\begin{equation}
U_{0}=-\mu _{+}|\varphi _{1}^{0}|^{2}+\frac{g}{2}|\varphi _{1}^{0}|^{4}\;\;.
\end{equation}
The integration measure is
\begin{equation}
\lbrack \mathcal{D}\tilde{\varphi}]=\mathcal{D}\tilde{\varphi}_{1}\mathcal{D} \tilde{\varphi}_{1}^{\ast }\mathcal{D}\tilde{\varphi}_{2}\mathcal{D}\tilde{\varphi}_{2}^{\ast}
\end{equation}
and the quadratic kernel,
\begin{equation}
S_{ij}^{(2)}=\left. \frac{\delta ^{2}S_{\mathrm{E}}}{\delta \varphi_{j}^{\ast}\delta\varphi _{i}}\right\vert _{\varphi _{1}=\varphi_{1}^{0},\varphi _{2}=0}\;\;,
\end{equation}
with $i,j=1,2$.
Integrating out quadratic fluctuations, we find an expression for the free
energy density,
\begin{equation}
W[J,\beta ]=U_{0}+\Delta W \label{W}
\end{equation}
with
\begin{equation}
\Delta W[J,\beta ]=\frac{1}{2}\ln \det \hat{S}^{(2)}=\frac{1}{2}{\rm Tr}\ln \hat{S}^{(2)}\;. \label{DeltaW}
\end{equation}
The matrix $\hat{S}^{(2)}$ in the \{Re$(\tilde{\varphi}_{1})$, Im$(\tilde{\varphi}_{1})$, Re$(\tilde{\varphi}_{2})$, Im$(\tilde{\varphi}_{2})$\} basis
decouples into two independent $2\times 2$ blocks,
\begin{equation}
\hat{S}^{(2)}=\left(
\begin{array}{cc}
\hat{S}_{a}^{(2)} & 0 \\
0 & \hat{S}_{b}^{(2)}
\end{array}
\right) ,
\end{equation}
with
\begin{equation}
\hat{S}_{a}^{(2)}=\left(
\begin{array}{cc}
-\frac{\nabla ^{2}}{2m}-\mu _{+}+3g|\varphi _{1}^{0}|^{2} & i\partial _{\tau} \\
-i\partial _{\tau } & -\frac{\nabla ^{2}}{2m}-\mu _{+}+g|\varphi_{1}^{0}|^{2}
\end{array}
\right)
\end{equation}
and
\begin{equation}
\hat{S}_{b}^{(2)}=\left(
\begin{array}{cc}
-\frac{\nabla ^{2}}{2m}-\mu _{-}+g|\varphi _{1}^{0}|^{2} & i\partial_{\tau } \\
-i\partial _{\tau } & -\frac{\nabla ^{2}}{2m}-\mu _{-}+g|\varphi_{1}^{0}|^{2}
\end{array}
\right) .
\end{equation}
It is not difficult to compute the trace in Fourier space, obtaining
\begin{equation}
\Delta W=\frac{1}{2\beta }\sum_{n=-\infty }^{+\infty }\int \frac{d^{3}q}{(2\pi )^{3}}\ln \left\{ \left( \omega _{n}^{2}+E_{1}^{2}\right) \left(\omega _{n}^{2}+E_{2}^{2}\right) \right\} ,
\end{equation}
where $\omega _{n}=2\pi n/\beta $ are the Matsubara frequencies,
\begin{equation}
E_{1}=\sqrt{\left( \frac{q^{2}}{2m}-\mu _{+}+3g|\varphi _{1}^{0}|^{2}\right)
\left( \frac{q^{2}}{2m}-\mu _{+}+g|\varphi _{1}^{0}|^{2}\right) } \label{E1}
\end{equation}
and
\begin{equation}
E_{2}=\frac{q^{2}}{2m}-\mu _{-}+g|\varphi _{1}^{0}|^{2}\;. \label{E2}
\end{equation}
Summing up the Matsubara frequencies, using
\begin{equation}
\frac{1}{\beta }\sum_{n}\ln (\omega _{n}^{2}+E_{i}^{2})=E_{i}+\frac{2}{\beta}\ln \left( 1-e^{-\beta E_{i}}\right) ,
\end{equation}
we obtain
\begin{equation}
\Delta W=\frac{1}{2}\int \frac{d^{3}q}{(2\pi )^{3}}\sum_{i}\left\{ E_{i}+
\frac{2}{\beta }\ln \left( 1-e^{-\beta E_{i}}\right) \right\} .
\end{equation}
It is interesting to note that, if we substitute the mean-field value for $\varphi _{1}^{0}$, given by Eq.~(\ref{phi2=0}), into Eqs.~(\ref{E1}) and~(\ref{E2}), we immediately obtain
\begin{equation}
\tilde{E}_{1}=\sqrt{\left( \frac{q^{2}}{2m}\right) \left( \frac{q^{2}}{2m} + 2g|\varphi _{1}^{0}|^{2}\right) }
\label{Goldstone}
\end{equation}
and
\begin{equation}
\tilde{E}_{2}=\frac{q^{2}}{2m}+2 \Omega _{\rm eff}\;.
\label{Gap}
\end{equation}
Equations.~(\ref{Goldstone}) and~(\ref{Gap}) are the usual energy excitations
computed in the Bogoliubov approximation. Note that $\lim_{q\rightarrow 0}
\tilde{E}_{1}=0$, corresponding with the Goldstone mode associated with the
spontaneous breakdown of the $U_{\varphi _{1}}(1)$ symmetry, while Eq.~(\ref{Gap}) is a gapped mode corresponding to non-condensate fluctuations.
It is useful to express the free energy $W(\beta ,J)$ in terms of the order parameter:
\begin{equation}
\bar{\varphi}=\delta W/\delta J=\varphi _{1}^{0}+\frac{1}{2}\mathrm{Tr}\left[\frac{1}{\hat{S}^{(2)}}\frac{\delta \hat{S}^{(2)}}{\delta \varphi _{1}^{0}} \frac{\delta \varphi _{1}^{0}}{\delta J}\right] \;.
\label{OP}
\end{equation}
At mean-field level, the order parameter is exactly the mean-field solution $\varphi _{1}^{0}$. However, when fluctuations are taken into account, the result given by Eq.~(\ref{OP}) is more involved.
We define the Gibbs free energy as a functional of the order parameter $\bar\varphi$ by making a Legendre transformation
\begin{equation}
\Gamma[\beta, \bar\varphi]= \bar\varphi J- W \; , \label{Gibbs}
\end{equation}
where $\delta \Gamma/\delta\bar\varphi=J$. In Eq.~(\ref{Gibbs}), $J$ is a
function of the order parameter $\bar\varphi$ obtained by inverting Eq.~(\ref{OP}). To leading order in the fluctuations the result is
\begin{eqnarray}
\Gamma[\beta, \bar\varphi]&=& \mu_+ |\bar\varphi|^2-\frac{g}{2}|\bar\varphi|^4 \nonumber \\
&-&\frac{1}{2}\int \frac{d^3q}{(2\pi)^3} \sum_i \left\{E_i+\frac{2}{\beta}\ln\left(1-e^{-\beta E_i} \right) \right\}.
\label{Gamma}
\end{eqnarray}
This is the Gibbs free energy computed at mean field plus Gaussian
fluctuations or, in the language of quantum field theory, the finite
temperature one-loop effective action.
The actual condensate amplitude $\bar{\varphi}_{m}$ is computed by
minimizing the free energy,
\begin{equation}
\left. \frac{\partial \Gamma \lbrack \beta ,\bar{\varphi}]}{\partial \bar{\varphi}}\right\vert _{\bar{\varphi}=\bar{\varphi}_{m}}=0.
\label{minimizing}
\end{equation}
By analogy with the mean field solution $|\varphi_1^0|^2=\mu_+/g$ we can define an
effective chemical potential in the following way,
\begin{equation}
|\bar{\varphi}_{m}|^{2}(T)=\frac{1}{g}\;\bar\mu_{+}(T),
\label{eq:barmu+}
\end{equation}
where now $\bar\mu_+(T)$ is the effective chemical potential for the $\varphi_1$ component, renormalized by quantum as well as thermal fluctuations.
Using Eqs.~(\ref{Gamma}) and~(\ref{minimizing}), we obtain an expression for $\bar{\mu}_{+}(T)$
in terms of the original bare $\mu _{+}$,
\begin{eqnarray}
\lefteqn{\bar{\mu}_{+}=\mu _{+}-\frac{1}{2}\int \frac{d^{3}q}{(2\pi )^{3}} \times } \nonumber \\
&\times &\left\{ \frac{(2\frac{q^{2}}{2m}+\bar{\mu}_{+})(1+2n(E_{1}))}{\sqrt{\left( \frac{q^{2}}{2m}\right) \left( \frac{q^{2}}{2m}+2\bar{\mu}_{+}\right)}}+1+2n(E_{2})\right\} ,
\label{mu+}
\end{eqnarray}
where $n(E_{i})$ is the usual Bose distribution
\begin{equation}
n(E_{i})=\frac{1}{e^{\beta E_{i}}-1} \label{Bose}
\end{equation}
with $i=1,2$.
The total particle density of each species can be computed as
\begin{eqnarray}
\rho_{\varphi_1}&=&\left.\frac{\partial \Gamma}{\partial\mu_+}\right|_{\bar\varphi_m}= \frac{\mu_+}{g}-\frac{1}{2} \int \frac{d^3q}{(2\pi)^3} \frac{q^2/2m}{E_1} (1+2n(E_1)) \nonumber \\
&&\mbox{\hspace*{1.8 cm}}-\frac{1}{2}\int \frac{d^3q}{(2\pi)^3} (1+2n(E_2)) \;, \\
\rho_{\varphi_2}&=&\left.\frac{\partial \Gamma}{\partial\mu_-}
\right|_{\bar\varphi_m}=\frac{1}{2}\int \frac{d^3q}{(2\pi)^3} (1+2n(E_2))\;.
\end{eqnarray}
Using the relation between $\mu_+$ and $\bar\mu_+$ given by Eq.~(\ref{mu+}),
we finally get
\begin{eqnarray}
\rho_{\varphi_1}&=& \frac{\bar\mu_+}{g}+\frac{1}{2} \int \frac{d^3q}{(2\pi)^3} \frac{\frac{q^2}{2m}+\bar\mu_+}{E_+}\coth(\beta E_+/2), \label{rho1} \\
\rho_{\varphi_2}&=&\frac{1}{2}\int \frac{d^3q}{(2\pi)^3} \coth(\beta E_-/2),
\label{rho2}
\end{eqnarray}
with
\begin{eqnarray}
E_+&=&\sqrt{\left(\frac{q^2}{2m}\right)\left(\frac{q^2}{2m}+2\bar\mu_+(T)
\right)}, \label{E+} \\
E_-&=&\frac{q^2}{2m}+2\bar\Omega_{\rm eff}(T)\;,
\label{E-}
\end{eqnarray}
where we have defined the renormalized effective Rabi frequency $\bar\Omega_{\rm eff}(T)=(\bar\mu_+-\bar\mu_-)/2 $ as a difference between the renormalized chemical potentials, in analogy with the bare effective Rabi frequency $\Omega_{\rm eff}=(\mu_+-\mu_-)/2 $. Notice that, while Eq.~(\ref{rho1}) completely determines $\bar\mu_+$, Eq.~(\ref{rho2}) is the definition of the renormalized chemical potential $\bar\mu_-$, through the expression for the excitation energy $E_-$ (Eq.~(\ref{E-})). In terms of these variables, Eqs.~(\ref{rho1}) and~(\ref{rho2}) are coupled equations. However, it is more convenient to work with $\bar\mu_+$ and $\bar\Omega_{\rm eff}$ as independent variables, in such a way that Eqs.~(\ref{rho1}) and~(\ref{rho2}) are now decoupled equations. In terms of these variables, all other chemical potentials are linear combinations of the former, such as, $\bar\mu_-=\bar\mu_+-2\bar\Omega_{\rm eff}$ and $\bar\mu=\bar\mu_+-\bar\Omega_{\rm eff}$.
Expressions~(\ref{rho1}) and~(\ref{rho2}) have the usual ultraviolet
divergences of a field theory at $T=0$. As is well known, temperature
fluctuations are always convergent. The usual way to deal with this
divergence is to regularize the integral and then renormalize the bare
constants $\mu _{+}$, $\mu _{-}$ and $g$, in order to obtain finite
results. A convenient procedure, in the non-relativistic scalar case, is the
cut-off technique. If we simply limit the momentum integrals using an
ultraviolet cut-off, $0\leq |\vec{q}|\leq \Lambda $, the results are
obviously $\Lambda $-dependent. However, if we begin the calculations with
renormalized constants, $\mu _{\pm }^{R}=\mu _{\pm }+\delta \mu _{\pm
}(\Lambda )$, we can adjust $\delta \mu _{\pm }(\Lambda )$ to make the
result independent of $\Lambda$. At the end, we can safely take the limit $\Lambda \rightarrow \infty $. After this procedure, the renormalized expressions read
\begin{eqnarray}
\rho _{\varphi _{1}} &=&\frac{\bar{\mu}_{+}}{g}+\frac{(m\bar{\mu}_{+})^{3/2}}{3\pi ^{2}}+\int \frac{d^{3}q}{(2\pi )^{3}}\frac{\frac{q^{2}}{2m}+\bar{\mu}_{+}}{E_{+}(e^{\beta E_{+}}-1)}, \label{rho1R} \\
\rho _{\varphi _{2}} &=&\int \frac{d^{3}q}{(2\pi )^{3}}\frac{1}{e^{\beta E_{-}}-1} \ .
\label{rho2R}
\end{eqnarray}
Equation~(\ref{rho1R}) implicitly defines the condensate density $\bar{\varphi}_{m}(T)$
or, equivalently, the effective chemical potential $\bar{\mu}_{+}(T)$, given by Eq.~(\ref{eq:barmu+}). This equation coincides with that derived from a one-loop effective
potential of a single self-interacting field~\cite{Rudnei-1}. Moreover,
eq.~(\ref{rho2R}) determines the effective Rabi frequency,
$\bar\Omega _{\mathrm{eff}}(T)$ through the expression for $E_-$, Eq.~(\ref{E-}).
In Eqs.~(\ref{rho1R}) and~(\ref{rho2R}), $\rho_{\varphi_1}$ and $\rho_{\varphi_2}$ are two independent constants, since the particle number of each species is conserved independently, due to the symmetry
$U_{\varphi_1}(1)\otimes U_{\varphi_2}(1)$.
The critical
temperature, $T_{c}$, is easily computed by fixing $\bar{\mu}_{+}(T_c)=0$ in
Eq.~(\ref{rho1R}), obtaining the usual expression for an ideal gas,
\begin{equation}
T_{c}=\frac{2\pi }{m\zeta (3/2)^{2/3}}\rho _{\varphi _{1}}^{2/3}, \label{Tc}
\end{equation}
with $\zeta (3/2)\sim 2.612$. We expect corrections of $T_{c}$ only at a
two-loop approximation~\cite{Tc1,Tc2}.
Since $E_-$ are gapped energy excitations, the integral in Eq.~(\ref{rho2R}) can be safely done in the classical limit. Solving for $\bar\Omega_{\rm eff}(T)$ we obtain,
\begin{equation}
\bar\Omega _{\mathrm{eff}}(T)=\frac{T}{2}\ln \left[ \left( \frac{\rho
_{\varphi _{1}}}{\rho _{\varphi _{2}}}\right) \left( \frac{T}{T_{c}}\right)
^{3/2}\right] \;. \label{mu+mu-}
\end{equation}
Note that there is a minimum temperature for which $\bar\Omega_{\rm eff}(T_r)=0$, given by
$T_{r}=(\rho _{\varphi_{2}}/\rho _{\varphi _{1}})^{2/3}T_{c}$.
At this temperature, the $O(2)$ symmetry is restored. This reentrance transition makes the excitation energy $E_-$ (Eq.~(\ref{E-})) gapless, producing an instability of the mean-field solution. Then, at this temperature, the chosen mean-field solution is unstable under Gaussian fluctuations.
In order to have the condensate structure given by Eqs.~(\ref{phi0}) and~(\ref{psi0}), we need to
fix $\rho _{\varphi _{2}}/\rho _{\varphi _{1}}<1$ and $T_{r}<T<T_{c}$.
In the next section, we numerically compute the condensate fractions as
functions of temperature for different values of the parameters.
\section{Numerical results}
\label{Numerics}
To compute the condensate density profile we rewrite Eq.~(\ref{rho1R}) in
dimensionless form. For this, we define the condensate fraction $\rho _{c}=\bar{\mu}_{+}/(g\rho _{\varphi _{1}})$. The dimensionless temperature is
defined as $\bar{T}=T/T_{c}$ and we introduce the diluteness parameter $n_{\varphi _{1}}=\rho _{\varphi _{1}}a^{3}$, where $a$ is the s-wave
scattering length. Using these definitions, we can write Eq.~(\ref{rho1R})
in the following form:
\begin{eqnarray}
1 &=&\rho _{c}+\frac{8}{3\pi ^{1/2}}n_{\varphi _{1}}^{1/2}\rho _{c}^{3/2} \nonumber \\
&+&\frac{4}{\pi ^{1/2}\zeta (3/2)}\bar{T}^{3/2}\int_{0}^{\infty }dyy\frac{y^{2}+2\zeta (3/2)^{2/3}n_{\varphi _{1}}^{1/3}\rho _{c}\bar{T}^{-1}}{\sqrt{y^{2}+4\zeta (3/2)^{2/3}n_{\varphi _{1}}^{1/3}\rho _{c}\bar{T}^{-1}}} \nonumber \\
&\times &\left( e^{y\sqrt{y^{2}+4\zeta (3/2)^{2/3}n_{\varphi _{1}}^{1/3}\rho_{c}\bar{T}^{-1}}}-1\right) ^{-1}\;\;.
\label{rhoca}
\end{eqnarray}
It is simple to check that the limit $n_{\varphi _{1}}\rightarrow 0$ leads
to the ideal gas result $\rho _{c}=1-\bar{T}^{3/2}$. The second term of the
r.h.s. of eq.~(\ref{rhoca}) gives the quantum depletion of the condensate,
while the third term represents the temperature dependence. Numerically
solving Eq.~(\ref{rhoca}), we can obtain the condensate fraction $\rho _{c}(\bar{T})$ for different values of the diluteness parameter $n_{\varphi _{1}}$.
From this result, it is simple to compute the condensate fractions for the
original fields $\phi $ and $\psi $, using Eqs.~(\ref{phi0}) and~(\ref{psi0}).
We define the condensate fractions for the fields $\phi$ and $\psi$ as
$\rho^c_{\phi}=|\phi_0|^2/(\rho_{\varphi_1}+\rho_{\varphi_2})$ and
$\rho^c_{\psi}=|\psi_0|^2/(\rho_{\varphi_1}+\rho_{\varphi_2})$, where we
chose the total particle density $\rho_{\phi}+\rho_{\psi}=\rho_{\varphi_1} + \rho_{\varphi_2}$ to normalize the fractions. Then, we use Eqs.~(\ref{phi0})
and~(\ref{psi0}) to relate $\rho^c_{\phi}$ and $\rho^c_{\psi}$ with $\rho_c$, given by Eq.~(\ref{rhoca}).
There are two interesting regimes to focus on. For $|\nu |/\Omega _{R}\ll 1$, the condensate fractions become
\begin{eqnarray}
\rho _{\phi }^{c} &\sim &\left( 1-\frac{\rho _{\varphi _{2}}}{\rho _{\varphi
_{1}}}\right) \left( \frac{|\nu |}{2\Omega _{R}}\right) ^{2}\rho _{c}\;, \label{numinorOmega1} \\
\rho _{\psi }^{c} &\sim &\left( 1-\frac{\rho _{\varphi _{2}}}{\rho _{\varphi
_{1}}}\right) \rho _{c}\;.
\label{numinorOmega2}
\end{eqnarray}
The first factor compensates the normalizations of $\rho _{\phi ,\psi }^{c}$
and $\rho _{c}$. To obtain it, we have considered $\rho _{\varphi _{2}}/\rho
_{\varphi _{1}}<1$ and we have dropped terms proportional to $(\rho
_{\varphi _{2}}/\rho _{\varphi _{1}})^{2}$. The condensate fraction is
determined by the factor $(|\nu |/2\Omega _{R})^{2}$ and the next
corrections to eqs.~(\ref{numinorOmega1}) and (\ref{numinorOmega2}) are
proportional to $(|\nu |/2\Omega _{R})^{4}$. In Fig.~(\ref{fig.nu}) we
show the typical profile of both condensates, where we have fixed $n_{\varphi _{1}}=10^{-5}$,
$\rho _{\varphi _{2}}/\rho_{\varphi _{1}}=10^{-1}$ and $|\nu |/2\Omega _{R}=0.24$. Note that $\rho _{\phi }^{c}$ is strongly
suppressed by the factor $\nu /\Omega _{R}$ and tends to disappear in the
limit $|\nu |\rightarrow 0$. An interesting observation is that the factor $|\nu |/\Omega _{R}$ is not corrected by temperature fluctuations. This is a
direct consequence of the $O(2)$ symmetry of the two-body interaction.
\begin{figure}[tbp]
\includegraphics[height=6 cm]{fig2.eps}
\caption{Condensate fractions as functions of the dimensionless temperature $\bar{T}$ in the limit $|\protect\nu |/\Omega _{R}<1$. The solid line
represents $\protect\rho _{\protect\psi }^{c}$, given by eq.~(\protect\ref{numinorOmega2}), while the dashed line is $\protect\rho _{\protect\phi }^{c}$, given by eq. (\protect\ref{numinorOmega1}). We have fixed
$n_{\protect\varphi _{1}}=10^{-5}$, $\protect\rho _{\protect\varphi _{2}}/\protect\rho _{\protect\varphi _{1}}=10^{-1}$
and $|\protect\nu |/2\Omega _{R}=0.24$.}
\label{fig.nu}
\end{figure}
In the opposite regime $\Omega_R/ |\nu|\ll1$, the condensate densities of
both species are essentially equal, with small corrections, given by
\begin{eqnarray} \label{Omegaminornu1}
\rho^c_{\phi_0}&\sim& \frac{1}{2}\left(1-\frac{\rho_{\varphi_2}}{\rho_{\varphi_1}}\right) \left(1-\frac{\Omega_R}{|\nu|}\right) \rho_c \; , \\
\rho^c_{\psi_0}&\sim& \frac{1}{2}\left(1-\frac{\rho_{\varphi_2}}{\rho_{\varphi_1}}\right) \left(1+\frac{\Omega_R}{|\nu|}\right) \rho_c \;,
\label{Omegaminornu2}
\end{eqnarray}
where we have discarded corrections of order $(\Omega_R/|\nu|)^2$. We show
these curves in Fig.~(\ref{fig.Omega}) for $n_{\varphi_1}=10^{-5}$, $\rho_{\varphi_2}/\rho_{\varphi_1}=10^{-1}$ and $\Omega_R/|\nu|=0.02$.
\begin{figure}[tbp]
\includegraphics[height=6 cm]{fig3.eps}
\caption{Condensate fractions as functions of the dimensionless temperature $\bar{T}$ in the limit $\Omega _{R}/|\protect\nu |<1$. The solid line represents $\protect\rho _{\protect\psi }^{c}$, given by eq.~(\protect\ref{Omegaminornu2}), while the dashed line is $\protect\rho _{\protect\phi }^{c}$, given by eq. (\protect\ref{Omegaminornu1}). We have fixed $n_{\protect\varphi _{1}}=10^{-5}$, $\protect\rho _{\protect\varphi _{2}}/\protect\rho _{\protect\varphi _{1}}=10^{-1}$ and $\Omega _{R}/|\protect\nu |=0.02$.}
\label{fig.Omega}
\end{figure}
\section{Discussion}
\label{discussion}
We have addressed the problem of equilibrium properties of a uniform mixture
of two bosonic fields in the presence of Josephson-type interactions. We
have considered a quantum field theory built by two non-relativistic complex
bosonic fields with general two-body local interactions. We have focused on
a particular symmetry point, in which, in addition to the $U(1)\otimes U(1)$
phase symmetry, there is an emergent $O(2)$ symmetry, related with
rotations in the isospin space $(\phi ,\psi )$.
We have minimally perturbed this model by considering the effect of
Josephson couplings that unbalance the species population by transferring
charge from one species to the other. These interactions are parametrized by
the Rabi frequency $\Omega _{R}$ and the detuning $\nu $. By making a
rotation in the isospin space, $(\phi ,\psi )\rightarrow (\varphi
_{1},\varphi _{2})$, we have shown that there is a special direction for
which the $U(1)\otimes U(1)$ phase symmetry is recovered and only one of the
bosonic species (say $\varphi _{1}$) could eventually condensate in this
framework. In this basis, the density of each bosonic species $\rho
_{\varphi _{1}}$ and $\rho _{\varphi _{2}}$ is conserved independently. Of
course, the $O(2)$ symmetry is still broken, provided the difference between
chemical potentials $\Delta \mu =\mu _{+}-\mu _{-}=2\Omega _{\mathrm{eff}}\neq 0$.
In the $(\varphi _{1},\varphi _{2})$ basis, it is simpler to compute
fluctuations. Specifically, we have computed finite temperature one-loop
effective action (the Gibbs free energy) as a function of the order
parameter and the temperature. In this way, by minimizing the free
energy, we have obtained the condensate fraction. Since the total density of each
species is conserved in this basis, the constant values of $\rho_{\varphi_1}$ and $\rho_{\varphi_2}$ completely determine the two chemical potentials $\bar\mu_+$ and $\bar\mu_-$. Alternatively, there is an interesting decoupling if we work in terms of the parameters $\bar\mu_+$ and $\bar\Omega_{\rm eff}$. While the density $\rho_{\varphi_1}$ fixes the value of
$\bar\mu_+(T)$, the value of $\rho_{\varphi_2}$ determines the value of $\bar\Omega_{\rm eff}(T)$. In this way, we can explicitly compute two limiting temperatures given by $\bar\mu_+(T_c)=0$ and
$\Omega_{\rm eff}(T_r)=0$. $T_c$ is the critical temperature for the Bose-Einstein condensation and $T_r$ is a reentrance temperature where the $O(2)$ symmetry is recovered. Below this temperature, the
mean-field solution is unstable under thermal fluctuations. Thus, our
results are only valid for $T_{r}<T<T_{c}$. To compute the condensate fractions below $T_r$, it is necessary to assume that both species in the rotated frame ($\varphi_{1}, \varphi_{2}$) could condensate, making the computation of fluctuations more involved.
To obtain the condensate profiles of the original fields, we rotated back to
the original basis $(\phi ,\psi )$. This rotation only depends on the ratio $\Omega _{R}/|\nu |$. It is interesting to note that, due to the $O(2)$
symmetry of the two-body interaction, fluctuations only renormalize the
effective Rabi frequency $\Omega _{\mathrm{eff}}=\sqrt{\Omega _{R}^{2}+|\nu
|^{2}}$, while the ratio $\Omega _{R}/|\nu |$ remains unaffected. Thus, the
isospin rotation coefficients are temperature independent.
In figures~(\ref{fig.nu}) and~(\ref{fig.Omega}) we show the condensate
profiles of the $\psi $ and $\phi $ species as functions of the temperature
for different values of the parameter $\Omega _{R}/|\nu |$. We have shown
that, for a temperature interval $T_{r}<T<T_{c}$, both bosonic species
condensate and the relatives phases are locked by the laser phase $\alpha$. We also have shown that the ratio between the condensates
essentially depends on the temperature-independent parameter $\Omega
_{R}/|\nu |$. We clearly see that, for $|\nu |/\Omega _{R}\rightarrow 0$,
only one condensate survives, while in the opposite limit $\Omega _{R}/|\nu
|\rightarrow 0$, both condensates are essentially equal, with small
corrections of order $\Omega _{R}/|\nu |$.
The results presented in this paper are valid, provided the two-body
interaction is invariant under isospin rotations. Consider, for instance, a
small deviation from the $O(2)$ model, $g_{\phi }=g_{\psi }=g$, but $g_{\psi
\phi }=g+\Delta g$. Upon rotation to the $(\varphi _{1},\varphi _{2})$
basis, a term proportional to $\Delta g(\varphi _{1}^{\ast }\varphi
_{1}^{\ast }\varphi _{2}\varphi _{2})$ will be generated. Thus, even though
we have ignored this type of terms in the original model, they will be
generated in a more general two-body interaction case. Thus, for $\Delta
g\neq 0$, there is no isospin direction in which the $U(1)\otimes U(1)$
symmetry is recovered. This fact makes the study of quantum and thermal
fluctuations more involved. We hope to report on this issue shortly.
\section*{Acknowledgments}
The Brazilian agencies \emph{Conselho Nacional de Desenvolvimento
Cient\'{\i}fico e Tecnol\'{o}gico (CNPq)} , \emph{Funda{\c{c}}{\~{a}}o de Amparo {\`{a}}
Pesquisa do Estado do Rio de Janeiro (FAPERJ)} and \emph{ Coordena\c c\~ao de
Aperfei\c coamento de Pessoal de N\'\i vel Superior (CAPES)} are acknowledged for
partial financial support. V.C.S. was financed by a doctoral fellowship by
CAPES. Z.G.A. was partially financed by a post-doctoral fellowship by CAPES.
D.G.B. acknowledge support from the Abdus Salam International Centre for Theoretical
Physics, ICTP, Trieste as a senior associated.
|
\section{ The Functionalized Cahn-Hilliard equation}
Amphiphilic materials are typically small molecules which contain both hydrophilic and hydrophobic components. This class of materials includes
surfactants, lipids, and block copolymers. Their propensity to spontaneously assemble network morphologies has drawn scientific attention for more than a
century, \cite{amphiphilic2000}. While amphiphilic materials are ubiquitous in organic settings, where lipid bilayers form cell membranes and many organelles,
their widespread use as charge separators in energy conversion devices is more recent. Network morphologies must be distinguished from single layer interfaces that are
typical of binary metals and other purely hydrophobic blends. While single layer interfaces separate a phase $A$ from a phase $B$, network morphologies are
comprised of thin regions of a phase $B$ which interpenetrate, and typically percolate through, a domain dominated by phase $A$. The Cahn-Hilliard
free energy, proposed in 1958, \cite{CH}, has been very successfully employed as a model of single layer morphology in hydrophobic blends, and
its gradient flows accurately describe their evolution. Models of amphiphilic mixtures, such as \cite{TS-87} and \cite{GS-90}, have been proposed.
The functionalized Cahn-Hilliard free energy; see \cite{KeithBrian-09, KeithNiGr-11, keithdai_2013},
is a special case of these earlier models that supports stable network morphologies including co-dimension one bilayers and co-dimension
two pores as well as pearled morphologies and defects such as end-caps and junctions. Rigorous results for the FCH free energy include the existence of bilayer structures, \cite{doelmanpromislow_2013}, and an analysis of their bifurcation structure, \cite{keithgreg_2014}, in particular the pearling bifurcation
which initiates changes in the co-dimension of the underlying morphology, and is commonly observed in amphiphilic polymer blends; see \cite{szostak,hayward}. The goal of this paper is to rigorously establish the existence of pearled bilayers, as modulations to stationary bilayers,
in the planar FCH equation.
Amphiphillic mixtures, such as emulsions formed by adding a minority fraction of an oil and soap mixture
to water, form network morphologies due to the tendency of the surfactant phase, e.g. soap, to enhance the formation of interfaces.
To model the network formation, the authors of \cite{TS-87} and \cite{GS-90} were motivated by small-angle X-ray scattering (SAXS) data to include
a higher-order term in the usual Cahn-Hilliard expansion for the free energy. Viewing the mixture as a binary phase, where $u\in H^2(\Omega)$
denotes the volume fraction of surfactant contained within the bounded material domain $\Omega\subset\mathbb{R}^3$, they proposed a free energy of the form
\vskip -0.2in
\begin{equation}
{\cal F}(u) := \int_\Omega f(u) + \epsilon^2A(u)|\nabla u|^2 + \epsilon^2B(u) \Delta u+ C(u) (\epsilon^2\Delta u)^2\,dx,
\end{equation}
where for well-posedness $C>0$ and the dimensionless parameter $\varepsilon \ll 1$ dictates the ratio of the interfacial width to a
characteristic size of $\Omega$. Assuming zero-flux boundary conditions, integration by parts on the $A(u)$ term permits a re-writing of
the energy in the completed-square form
\begin{equation}
{\cal F}(u) = \int_\Omega C(u)\left(\epsilon^2\Delta u- \Frac{\overline{A}-B}{2C} \right)^2 +
f(u) -\Frac{(\overline{A}-B)^2}{4C(u)}\, dx,
\end{equation}
where $\overline{A}$ is a primitive of $A$. To simplify the form we replace $C(u)$ with $\frac12$, relabel the potential within the
squared term by $W'(u)$, and scale the potential outside the squared term as $\delta P(u)$ with $\delta\ll1$, yielding
\begin{equation}
{\cal F}(u) = \int_\Omega \frac{1}{2}\left( \epsilon^2 \Delta u-W'(u)\right)^2 + \delta P(u)\, dx.
\end{equation}
The first term is the square of the variational derivative of a Cahn-Hilliard type free energy, and
the strongly degenerate case $\delta=0$, has the special property that its global minimizers are precisely the
{\sl critical points} of the corresponding Cahn-Hilliard energy. A variant of this case was proposed
as a target for $\Gamma-$convergence analysis by De Giorgi; see \cite{DeGiorgi}.
The strong functionalized Cahn-Hilliard free energy corresponds to the distinguished limit $\delta=\varepsilon$,
a choice of potential $P$ which incorporates the functionalization parameters $\eta_1>0$, $\eta_2\in \mathbb{R}$ in the form
\begin{equation}
\label{FCHE}
\mathcal{F}(u)=\int_{\Omega}\frac{1}{2}\left(\varepsilon^2 \Delta u- W^\prime(u)\right)^2 - \varepsilon \left(\eta_1 \varepsilon^2|\nabla u|^2 + \eta_2 W(u) \right)\mathrm{d} x,
\end{equation}
and require the $C^\infty$-smooth potential $W: \mathbb{R}\rightarrow\mathbb{R}$ to be a double well potential with two minima at $u=-1$ and $u=m>0$ and one local maximum
at $u=0$. The minima have unequal depths, normalized so that $W(-1)=0>W(m)$ and the well is non-degenerate in the sense that
$\mu_-:=W^{\prime\prime}(-1)>0$, $\mu_+:=W^{\prime\prime}(m)>0$, and $\mu_0:=W^{\prime\prime}(0)<0$.
With these assumptions $u=-1$ is associated to a bulk solvent phase, while the value of $u+1>0$ is proportional to the density of the amphiphilic phase.
The strong FCH equation is the $H^{-1}$ gradient flow of the FCH energy \eqref{FCHE}, which takes the form
\begin{equation}
\label{e:FCH}
u_t=\Delta\frac{\delta \mathcal{F}}{\delta u}=\Delta\left((\varepsilon^2\Delta - W^{\prime\prime}(u) + \varepsilon \eta_1)
(\varepsilon^2\Delta u - W^{\prime}(u))+\varepsilon\eta_d W^\prime(u)\right),
\end{equation}
where $\eta_d:=\eta_1-\eta_2$. The gradient flow is mass-preserving when subject to zero-flux boundary conditions; see \cite{doelmanpromislow_2013} for details.
We focus on the stationary strong-FCH equation which takes the form
\begin{equation}
\label{e:sFCH}
(\varepsilon^2\Delta - W^{\prime\prime}(u) + \varepsilon \eta_1)
(\varepsilon^2\Delta u - W^{\prime}(u))+\varepsilon\eta_d W^\prime(u)=\varepsilon \gamma,
\end{equation}
subject to zero-flux boundary conditions. The constant $\gamma$ can be thought of as a Lagrange multiplier arising from mass conservation.
The FCH equation is known to support families of bilayer solutions, \cite{doelmanpromislow_2013}, which can be unstable to either pearling or meandering bifurcations.
Pearling refers to periodic modulations of the thickness of the bilayer, while the meander modes are associated with the curvature driven motion of the
underlying bilayer interface. In this work, we provide a fully rigorous proof of the existence of spatially periodic patterns which arise after the onset of the pearling
bifurcation. We restrict our attention to planar domains $\Omega\subseteq\mathbb{R}^2$, proving the major existence results
in the spatially extended case $\Omega=\mathbb{R}^2$. The construction of a bilayer morphology requires a choice of a smooth,
closed, co-dimension one interface $\Gamma\subset\Omega$ that is far from self intersection. We address
two simple choices of interface: the extended flat bilayer, corresponding to $\Gamma_f=\{(s,0)\,\bigl |\, s\in\mathbb{R}\}$, and the circular bilayer of radius $R_0>0$,
corresponding to $\Gamma_{R_0}:=\{(R_0\cos{\theta}, R_0\sin{\theta})\, \big|\, \theta\in[0,2\pi)\}.$ Our construction applies spatial dynamics techniques,
a center-manifold-reduction argument, and a normal form transformation to the stationary, strong-FCH equation, yielding an 8th order ODE system,
which weakly couples the four dimensional pearling subspace and the four dimensional meander subspace. To prove the existence, we
restrict to the pearling subspace, yielding a four-dimensional reduced system, called the pearling normal form (PNF), \eqref{e:PNF},
\begin{equation*}
\begin{cases}
\dot{C_1}&=\mathrm{i} (1+\omega_1\varepsilon) C_1 + C_2+\mathrm{i} C_1\big[\alpha_7C_1\bar{C}_1+\alpha_8\mathrm{i}(C_1\bar{C}_2-\bar{C}_1C_2)\big],\\
\dot{C_2}&=\mathrm{i} (1+\omega_1\varepsilon) C_2 +\mathrm{i} C_2\big[\alpha_7C_1\bar{C}_1+\alpha_8\mathrm{i}(C_1\bar{C}_2-\bar{C}_1C_2)\big]
+ C_1\left[-\alpha_0\varepsilon+\mathrm{i} \alpha_2(C_1\bar{C_2}-\bar{C_1}C_2)\right],
\end{cases}
\end{equation*}
where $C_1$, $C_2\in\mathbb{C}$, the constants $\omega_1, \alpha_j\in\mathbb{R}$, and the conjugate equations are omitted. It is at this level that the structure of the
pearling bifurcation is made clear: the PNF admits a \textit{degenerate $1:1$ resonance}, related to the $1:1$ resonances extensively investigated in \cite{imd_1989, iooper_1993, harioo}. As in the $1:1$ resonance case, the PNF has two first integrals
\[
K:=\frac{\mathrm{i}}{2}(C_1\bar{C}_2-\bar{C}_1C_2), \quad H:=|C_2|^2+\left(-\alpha_0\varepsilon+2\alpha_2K\right)|C_1|^2.
\]
Imposing consistency conditions to the solutions of the PNF slaves $H$ to the scaled parameter $\kappa:=\varepsilon^{-3/2} K$, which remains as a free parameter in the
construction of the pearled solutions. More importantly, the parameter $\alpha_0$ in the PNF, given in \eqref{def-alpha0}, is precisely the critical bifurcation parameter whose sign characterizes the onset of the pearling bifurcation. For $\alpha_0>0$ we characterize the pearled solutions of the PNF and establish their existence in
the full system through a persistence argument. While the persistence argument is based upon \cite{iooper_1993}, the analysis in this case is more delicate as
the degeneracy corresponds to a distinct singularity requiring different scalings. Moreover the coupling between the pearling modes and the meander modes
requires the analysis of an eight dimensional problem. In the remainder of this section we make a rigorous statement of these results.
\subsection{Pearling of Extended Flat Bilayers}
The existence of a one-dimensional family of flat bilayer solutions, $u_h$,
parameterized by the Lagrange multiplier, $\gamma$, was established in \cite{doelmanpromislow_2013}.
Their construction is based upon new coordinates, corresponding to the $\varepsilon$-scaled distance $r$ to $\Gamma_f$ and a tangential variable $\tau$
for which the Laplacian takes the form
\begin{equation}
\label{e:slap}
\varepsilon^2\Delta=\partial_r^2+\varepsilon^2\partial_\tau^2,
\end{equation}
and the stationary equation (\ref{e:sFCH}) is rewritten as
\begin{equation}
\label{e:2sFCH}
\left(\partial_r^2-W^{\prime\prime}(u)+\varepsilon^2\partial_\tau^2+\varepsilon\eta_1\right)
\left(\partial_r^2u-W^\prime(u)+\varepsilon^2\partial_\tau^2u\right)+\varepsilon\eta_d W^\prime(u)=\varepsilon\gamma.
\end{equation}
For the flat interface, the bilayer profile is independent of the tangential variable, $\tau$, and hence is captured as the first component of a homoclinic solution of the
$4$-th order extended flat-bilayer ODE system in $r\in\mathbb{R}$,
\begin{equation}
\label{e:rODE}
\begin{cases}
\partial_r u=p, \\
\partial_r p=W^\prime(u)+\varepsilon v, \\
\partial_r v=q, \\
\partial_r q=W^{\prime\prime}(u)v+\left(\gamma-\eta_dW^\prime(u)\right)-\varepsilon\eta_1v,
\end{cases}
\end{equation}
For sufficiently small $\varepsilon$, this extended flat-bilayer ODE system \eqref{e:rODE} contains $3$ critical points, among which we consider the one with leading order
$(-1,0,-\frac{\gamma}{\mu_-},0)$, which we denote as
\[
P_-(\varepsilon)=\left(u_-(\varepsilon),0,v_-(\varepsilon),0\right).
\]
Indeed, via \eqref{e:rODE}, it is straightforward to see that the parameter $\gamma$ relates linearly, at leading order, to the far-field density of amphiphilic
material, $1+u_-(\varepsilon)$, via the expansion
\[
1+ u_-(\varepsilon;\gamma)=\frac{\gamma}{\mu_-^2}\varepsilon+\mathcal{O}(\varepsilon^2).
\]
In \cite{doelmanpromislow_2013} the existence of the flat homoclinic solution $U_h=(u_h,p_h, v_h,q_h)^T$ is established for $\varepsilon>0$ sufficiently small,
but independent of $\eta_1, \eta_2,$ and $\gamma$. The construction follows by perturbation off of the $\varepsilon=0$ case, in which case the first component $u_0$ is the solution
of the two-dimensional ODE
\begin{equation}
\label{2d-ODE}
\partial_r^2 u_0 = W'(u_0),
\end{equation}
which is homoclinic to $u_-(0)$. The linearization of (\ref{2d-ODE}) about $u_0$, yields the operator
\begin{equation}
\label{cL_0-def}
\mathcal{L}_0:=\partial_r^2-W''(u_0),
\end{equation}
which, acting on $L^2(\mathbb{R})$, has a single positive eigenvalue, $\lambda_0>0$, and a zero eigenvalue, $\lambda_1=0$, with the remainder of the spectrum strictly negative. Denoting
the associated eigenfunctions by $\psi_0$ and $\psi_1$ and introducing, $v_0\in L^\infty(\mathbb{R})$, the unique, even solution of
\begin{equation}
v_0 = \gamma \mathcal{L}_0^{-1} 1 -\eta_d \mathcal{L}_0^{-1}W'(u_0),
\end{equation}
the pearling bifurcation of the bilayer $u_h$ is characterized in terms of the functionalization parameters $\eta_1$ and $\eta_2$ via the sign of the quantity
\begin{equation}
\label{def-alpha0}
\alpha_0=\frac{1}{4\lambda_0^2}\int_{\mathbb{R}}\left(W^{\prime\prime\prime}(u_0)v_0-\eta_dW^{\prime\prime}(u_0)\right)\psi_0^2\mathrm{d} r= \alpha_{01} \gamma - \alpha_{02} \eta_d, \end{equation}
where the constants
\begin{equation}
\begin{aligned}
\alpha_{01} &= \frac{1}{4\lambda_0^2} \int_{\mathbb{R}} W^{\prime\prime\prime}(u_0)(\mathcal{L}^{-1} 1)\psi_0^2\mathrm{d} r, \\
\alpha_{02} & := \int_{\mathbb{R}} \left((\mathcal{L}_0^{-1}W'(u_0)+ W^{\prime\prime}(u_0)\right)\psi_0^2\mathrm{d} r,
\end{aligned}
\end{equation}
depend only upon the shape of the double well potential, $W$.
Our main result for flat bilayers establishes that a one parameter family of pearled solutions of (\ref{e:2sFCH}) generically bifurcates out
of each stationary flat bilayer for $\alpha_0>0$.
\begin{Theorem}[existence of extended pearled flat bilayers]
\label{t:main}
Fix $\eta_1, \eta_2, \gamma \in\mathbb{R}$. Assume that $W$ is a non-degenerate double well potential and that
$\alpha_0$ defined in (\ref{def-alpha0}) is strictly positive and
\begin{equation}
\label{def-beta0}
\beta_0:=\frac{1}{4\lambda_0^2}\int_\mathbb{R}\left(W^{\prime\prime\prime}(u_0)v_0-\eta_dW^{\prime\prime}(u_0)\right)\psi_1^2\mathrm{d} r\neq 0,\\
\end{equation}
Then there exist positive constants $\varepsilon_0>0$ and $\kappa_0>0$ such that, for any $\varepsilon\in(0,\varepsilon_0]$, up to translation,
the extended stationary strong-FCH \eqref{e:2sFCH}
admits a smooth one-parameter family of extended pearled solutions,
$u_\mathrm{p}(\tau,r;\sqrt[4]{\varepsilon},\sqrt{|\kappa|})$ with period $T_\mathrm{p}(\sqrt[4]{\varepsilon},\sqrt{|\kappa|})$, parameterized by $\kappa\in[-\kappa_0,\kappa_0]$. More specifically, $u_{\mathrm{p}}$ and $T_{\mathrm{p}}$ are smooth with respect to their arguments within the domains expect at $\kappa=0$. The extended pearled solution $u_\mathrm{p}$ admits the asymptotic form
\begin{equation}
u_\mathrm{p}(\tau, r)=u_h(r)+2\frac{\sqrt{\varepsilon|\kappa|}}{\sqrt[4]{\alpha_0}}\cos\left(\frac{2\pi}{T_\mathrm{p}} \tau\right)\psi_0(r)+
\mathcal{O}\left(\varepsilon(\sqrt{\varepsilon}+\sqrt{|\kappa|})\right),
\end{equation}
where the error is measured in the $L^\infty(\mathbb{R}^2)$-norm and
\begin{equation}
T_\mathrm{p}=\frac{2\pi\varepsilon}{\sqrt{\lambda_0}}\left[1-\sqrt{\alpha_0\varepsilon}+\mathcal{O}\left(\varepsilon(1+\sqrt{|\kappa|})\right)\right] .
\end{equation}
Moreover,
the far-field limit of the extended pearled solution is
\begin{equation}
\lim_{r\rightarrow \infty}u_\mathrm{p}(\tau,r)=\lim_{r\rightarrow \infty}u_h(r)=u_-(\varepsilon).
\end{equation}
\end{Theorem}
\subsection{Pearling of extended Circular Bilayers}
For a circular co-dimension one interface $\Gamma_{R_0}$ we take the tangential coordinate $s$ to represent the direction with constant curvature $k=-R_0$, and rescale the corresponding independent variable
as $\theta=s/R_0$ which lies in $[0,2\pi]$.
The Laplacian admits the expression
\begin{equation}
\label{e:plap}
\varepsilon^2\Delta=\partial_r^2+\frac{\varepsilon}{R_0+\varepsilon r}\partial_r+\frac{\varepsilon^2}{(R_0+\varepsilon r)^2}\partial_\theta^2,
\end{equation}
and the stationary strong-FCH \eqref{e:sFCH} in $(r, \theta)$ takes the form
\begin{equation}
\label{e:2sFCH2}
\Big(\partial_r^2-W^{\prime\prime}(u)+\frac{\varepsilon\partial_r}{R_0+\varepsilon r}+\frac{\varepsilon^2\partial_\theta^2}{(R_0+\varepsilon r)^2}+\varepsilon\eta_1\Big)
\Big(\partial_r^2u-W^\prime(u)+\frac{\varepsilon\partial_ru}{R_0+\varepsilon r}+\frac{\varepsilon^2\partial_\theta^2u}{(R_0+\varepsilon r)^2}\Big)+\varepsilon\eta_d W^\prime(u)=\varepsilon\gamma.
\end{equation}
Suppressing the tangential variable $\theta$, the stationary strong-FCH \eqref{e:2sFCH2} reduces to the extended circular-bilayer ODE system in $r\in\mathbb{R}$,
\begin{equation}
\label{e:rODE2}
\begin{cases}
\partial_r u=p, \\
\partial_r p=W^\prime(u)+\varepsilon v, \\
\partial_r v=q, \\
\partial_r q=W^{\prime\prime}(u)v+[\gamma_1-\eta_dW^\prime(u)]+\varepsilon[\gamma_2-\frac{2}{R_0}q+\frac{1}{R_0^2}W^\prime(u)-\eta_1v-
\frac{1}{R_0}\eta_1p]+\mathcal{O}(\varepsilon^2),
\end{cases}
\end{equation}
where $\gamma$ has been expanded as,
\[\gamma=\gamma_1+\varepsilon\gamma_2+\mathcal{O}(\varepsilon^2).\]
Like the flat-bilayer system, the extended circular-bilayer ODE system \eqref{e:rODE2} possesses $3$ critical points, of which we single out
the critical point
\[P_-(\varepsilon)=\left(u_-(\varepsilon),0,v_-(\varepsilon),0\right),\]
which satisfies $P_-(\varepsilon)\to (-1,0,-\frac{\gamma_1}{\mu_-},0)$, as $\varepsilon\to0.$ In \cite{doelmanpromislow_2013}, it was shown
that for fixed $\eta_1, \eta_2$ and $R_0>0$ there exists a unique function
$\gamma_h=\gamma_1+\mathcal{O}(\varepsilon)$ for which
\begin{equation}
\label{def-gamma_1}
\gamma_1=(\eta_d-2\eta_1)\frac{\int_\mathbb{R} (u^{\prime }_0 )^2\mathrm{d} r}{2\int_\mathbb{R}(u_0+1)\mathrm{d} r},
\end{equation}
such that for the choice $\gamma=\gamma_h(\varepsilon)$ there exists a nontrivial orbit of (\ref{e:rODE2}) which is homoclinic to $P_-(\varepsilon).$
\begin{Remark}
The parameter $\gamma$ is free for flat bilayers while it is prescribed for circular bilayers because
the flat-bilayer ODE system \eqref{e:rODE} is Hamiltonian while the circular-bilayer ODE system \eqref{e:rODE2} is not.
\end{Remark}
Our main result for circular bilayers provides the existence of discrete families of one-parameter, pearled, bilayer solutions of the stationary strong-FCH equation (\ref{e:2sFCH2}); see Figure \ref{f:cirpearling}. Both their radii $R_{0,n}=R_{0,n}(\varepsilon, \kappa)$ and pearling amplitudes are parameterized
by the value of the scaled first-integral $\kappa$ of the Pearling Normal Form equation.
\begin{Theorem}[existence of extended pearled circular bilayers]
\label{t:main2}
Fix $\eta_1, \eta_2\in\mathbb{R}$ and $R_->0$. Assume that $W$ is a non-degenerate double well potential and that $\alpha_0$ and $\beta_0$, defined in
(\ref{def-alpha0}) and (\ref{def-beta0}) respectively, satisfy $\alpha_0>0$, $\beta_0\neq 0$.
Then there exist constants $\varepsilon_0, \kappa_0>0$ and $\mathrm{n_-}>0$ such that,
for all $\varepsilon\in(0,\varepsilon_0]$ and each $\mathrm{n}\in \mathbb{Z}^+\cap[\frac{\mathrm{n_-}}{\varepsilon},+\infty)$,
the stationary, strong-FCH equation \eqref{e:2sFCH2} in the infinite strip $(\theta,r)\in(\mathbb{R}/2\pi\mathbb{Z})\times\mathbb{R}$, subject to the choice $\gamma=\gamma_h(\epsilon)$, with $\gamma_h$ defined by (\ref{def-gamma_1}),
admits, up to translation, a finite family of one-parameter pearled solutions
$u_{\mathrm{p,n}}(\theta,r;\sqrt[4]{\varepsilon},\sqrt{|\kappa|})$ with period $\frac{2\pi}{n}$ and
radius $R_{\mathrm{0,n}}(\sqrt[4]{\varepsilon},\sqrt{|\kappa|})\geq R_-$.
Each solution is parameterized by $\kappa\in[-\kappa_0,\kappa_0]$, and
is smooth with respect to its arguments except at $\kappa=0$. The extended pearled solution $u_{\mathrm{p,n}}$ admits the asymptotic form
\begin{equation}
\label{circ-bilayer}
u_{\mathrm{p,n}}(\theta, r;\sqrt[4]{\varepsilon},\sqrt{|\kappa|})=u_h(r)+2\frac{\sqrt{\varepsilon|\kappa|}}{\sqrt[4]{\alpha_0}}\cos(\mathrm{n}\theta)\psi_0(r)+\mathcal{O}\left(\varepsilon(\sqrt{\varepsilon}+\sqrt{|\kappa|})\right),
\end{equation}
where the radius of the circular bilayer
\begin{equation}
R_{\mathrm{0,n}} = \Frac{\mathrm{n}\varepsilon}{\sqrt{\lambda_0}} \left[1-\sqrt{\alpha_0\varepsilon}+\mathcal{O}\left(\varepsilon(1+\sqrt{|\kappa|})\right)\right].
\end{equation}
depends only weakly upon $\kappa$. The far-field limit of the extended pearled solution
\begin{equation}
\lim_{r\rightarrow \infty}u_{\mathrm{p,n}}(\theta,r)=\lim_{r\rightarrow \infty}u_h(r)=u_-(\varepsilon),
\end{equation}
is independent of $\mathrm{n}$.
\end{Theorem}
\begin{figure}[H]
\centering
\includegraphics[width=0.6\textwidth]{EQPL14.png}
\caption{ Quarter-plane views of equilibrium of the strong FCH equation \eqref{e:FCH} corresponding to radially symmetric bilayer initialdata with $\varepsilon=0.1$
and double well potential $W$ as given in Section 5 of \cite{doelmanpromislow_2013}. (left) For $\eta_1= 1$ and $\eta_2= 2$, we have $\alpha_0<0$, and the $t=3000$
evolution is a circular bilayer equilibrium. (right) For $\eta_1=2$ and $\eta_2=2$, we have $\alpha_0>0$, and the $t=500$ evolution of the initial data yields a circular pearled bilayer.}
\label{f:cirpearling}
\end{figure}
\begin{Remark}
The number $n$ can be interpreted as the number of ``beads'' within a pearled circular bilayer. The size of each bead--the periodicity in the physical variables-- is
\[
T_{\mathrm{p},n}:=\frac{2\pi R_{0,n}}{n}=\frac{2\pi\varepsilon}{\sqrt{\lambda_0}}\left[1-\sqrt{\alpha_0\varepsilon}+\mathcal{O}\left(\varepsilon(1+\sqrt{|\kappa|})\right)\right],
\]
depends only weakly upon $\kappa$, at order $\mathcal{O}(\varepsilon^2\sqrt{|\kappa|})$, while the leading order amplitude of each bead,
\begin{equation}
\label{Ap-def}
A_\mathrm{p}:=2\frac{\sqrt{\varepsilon|\kappa|}}{\sqrt[4]{\alpha_0}},
\end{equation}
scales with $(\sqrt{\varepsilon|\kappa|})$.
\end{Remark}
For both the flat and circular interfaces, the form of the amplitude of the pearled pattern suggests a divergence as
$\alpha_0\to0^+$, however this is an anomaly arising from the degeneracy of the $1:1$ resonance in the PNF system, \eqref{e:PNF}. Indeed an analysis
of Lemma \ref{l:29} shows that a necessary condition for the existence of periodic patterns is
\begin{equation}\label{e:super}
\sqrt{\varepsilon_0}\kappa_0< \frac{\alpha_0}{2|\alpha_2|},
\end{equation}
from which we deduce that the pearling bifurcation, while degenerate, retains some supercritical characteristics.
\begin{Proposition}[super-criticality of pearled bilayers]
In addition to the assumptions of either Theorem~\ref{t:main} or \ref{t:main2}, assume that
$\alpha_2$, defined in \eqref{e:alphas}, satisfies $\alpha_2\neq 0$. Fix
$\varepsilon\in(0,\varepsilon_0)$ and tune $\eta_1$ and $\eta_2$ so that $\alpha_0$ goes to $0$; then,
under this limit, the pearling amplitude, defined in \eqref{Ap-def}, satisfies
\[
\lim\limits_{\alpha_0\to0}\sup\limits_{\kappa\in[-\kappa_0,\kappa_0]}\Frac{A_p(\kappa)}{\sqrt[4]{\alpha_0}}\leq C,
\]
for some constant $C>0.$
\end{Proposition}
\subsection{Pearling and Degeneracy in Bounded Domains}
The existence results for both bilayers and pearled bilayers naturally extend to a bounded domain, $\Omega\subset \mathbb{R}^2$
so long as the domain possesses the same symmetry as the bilayer interface. Indeed, for typical homogeneous boundary conditions,
such as discussed in \cite{PZ-2013}, and for a bilayer interface $\Gamma$ that is an $\mathcal{O}(1)$ distance from
$\partial\Omega$ in the unscaled coordinates, then the exponential decay of the extended pearled patterns in $r$ leads to an $\mathcal{O}(\varepsilon^{-1})$ exponential decay
in the unscaled coordinates, and a standard matching argument; such as in \cite{PB_2011}, permits an extension of the existence result. This
is particularly relevant for the circular bilayers within a concentric circular domain. The adaptation of the extended flat bilayer to a flat bilayer
within a rectangular domain subject to periodic boundary conditions is trivial so long as the flat interface intersects the domain boundary at a right angle; see Figure \ref{f:domain} for an illustration.
The construction of the associated pearled solutions requires a tuning of the periodicity of the pearled pattern, as in the case of
the circular bilayer.
For the gradient flow (\ref{e:FCH}), the total mass $\int_\Omega u(x)\mathrm{d}x$ is conserved under time evolution, and as such it is natural to search
for equilibria with prescribed total mass. For circular bilayers; see Figure \ref{f:domain}, the far-field value of $u$ is prescribed,
and the mass of a circular bilayer is an increasing function of the radius $R_0$. Moreover the mass is independent of the pearling correction,
at least to leading order, thus the total mass of the circular bilayer $u_{p,n}$ in \eqref{circ-bilayer} increases monotonically with its
radius $R_{0,n}$; however the admissible radii
$$\left \{ R_{\mathrm{0,n}}(\kappa) \, \Bigl|\, \mathrm{n}\in \mathbb{Z}^+\cap[\frac{\mathrm{n_-}}{\varepsilon},+\infty), \, \kappa\in[-\kappa_0,\kappa_0]\right\}.$$
depend only weakly upon the internal parameter $\kappa$. Indeed the gaps between consecutive radii satisfy
\[R_{\mathrm{0,n}}(\kappa)-R_{\mathrm{0,n+1}}(\kappa)= \frac{\varepsilon}{\sqrt{\lambda_0}}+\mathcal{O}\left(\varepsilon^{\frac32}\right),\]
while the range of the radii over the values of $\kappa$ is bounded by
$|R_{\mathrm{0,n}}(\kappa_0)-R_{\mathrm{0,n}}(0)|\leq \mathcal{O}(\varepsilon^2).$
While we have established the existence of radii $R_0$ which support pearled bilayers, there also may exist radii, and corresponding
total masses,for which no pearled circular bilayer solutions exist local to the associated circular bilayer; see Figure \ref{f:R0Kappa}.
\begin{minipage}{0.58\textwidth}
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{alpha.png}
\caption{(left) A circular bilayer with interface $\Gamma_{R_0}$ in a concentric domain of radius $R_b$. (right) A flat
bilayer with interface $\Gamma_f$ which intersects the rectangular domain at a right angle.}
\label{f:domain}
\end{figure}
\end{minipage}\qquad
\begin{minipage}{0.36\textwidth}
\begin{figure}[H]
\includegraphics[width=\textwidth]{R0Kappa.png}
\caption{The admissible radii $\{R_{0,\textrm{n}}\}$ graphed verses $\kappa$ for fixed $\varepsilon$. The gaps between
successive radii are $\mathcal{O}(\epsilon)$ while the variation in $R_{0,{\textrm{n}}}$ with $\kappa$ is $\mathcal{O}(\varepsilon^2).$}
\label{f:R0Kappa}
\end{figure}
\end{minipage}
As an existence problem, these scalings imply that an $\mathcal{O}(\varepsilon^3)$ change in the mass faction, which corresponds
to an $\mathcal{O}(\varepsilon^2)$ change in the bilayer radius $R_0$, can induce an $\mathcal{O}(1)$ impact on $\kappa$, and hence an
$\mathcal{O}(\sqrt{\varepsilon})$ influence on the pearling amplitude of the associated equilibrium. This sensitivity of the pearling amplitude
to the mass fraction exemplifies the degeneracy of the pearled morphologies. The size of the pearled ``beads" is fixed, but the amplitude
of the pearling pattern couples sensitively to the full system. In particular for the strong FCH gradient flow, \eqref{e:FCH}, the possibility of
non-existence of pearled morphologies at particular mass fractions and the delicate interaction between the radius of a circular bilayer and the
amplitude of the high-frequency pearled morphology suggest a complex problem whose resolution may be quite sensitive to
numerical truncation error.
\section{Pearling of the Flat Planar Bilayer}\label{s:2}
This section presents the construction of the pearled solutions $u_\mathrm{p}$ to the stationary strong-FCH \eqref{e:2sFCH} about
an infinite, flat, co-dimension one interface, $\Gamma_f$ embedded in $\mathbb{R}^2.$
The extended pearled solutions $u_\mathrm{p}$ are small-amplitude modulations of the extended flat bilayers $u_h$,
periodic in the flat direction $\tau$. The construction is organized as follows: In Section \ref{ss:21}, the application of spatial dynamics techniques, together with a center manifold reduction, reduces the FCH equation to an 8th order ODE system; the derivation of the leading-order terms of the reduced ODE system are summarized in Section \ref{ss:22} with the details relegated to the Appendix. A normal form analysis presented in Section \ref{ss:23} reveals the pearling bifurcation structure; and in section \ref{ss:24}, it is shown that the pearling norm form admits a family of periodic orbits, which persist as solutions of the full reduced ODE system, yielding the extended pearled solutions $u_\mathrm{p}$ of Theorem~\ref{t:main}.
\subsection{Spatial dynamics and center manifold reduction}\label{ss:21}
The spatial dynamics analysis begins by re-writing equation \eqref{e:EPP} as an infinite-dimension dynamical system in the rescaled
$\tau$ variable followed by a normal form reduction on the associated center manifold.
To this end, we rescale $\tau$ by $t=\frac{\sqrt{\lambda_0}}{\varepsilon}\tau$ and search for extended pearled solutions
$u_\mathrm{rp}$ of
\begin{equation}
\label{e:EPP}
\left(\partial_r^2-W^{\prime\prime}(u)+\lambda_0\partial_t^2+\varepsilon\eta_1\right)
\left(\partial_r^2u-W^\prime(u)+\lambda_0\partial_t^2u\right)+\varepsilon\eta_d W^\prime(u)-\varepsilon\gamma=0,
\end{equation}
which satisfy boundary conditions at infinity,
\begin{equation}
\lim_{r\rightarrow \pm\infty}|u_{\mathrm{rp}}(t,r)-u_-(\varepsilon)|=0, \text{ for all }t\in\mathbb{R},
\end{equation}
and are even and $T_\mathrm{rp}$-periodic in $t$,
\begin{equation}\label{e:pc}
u_\mathrm{rp}(-t,r)=u_\mathrm{rp}(t,r), \quad u_\mathrm{rp}(t+T_\mathrm{rp},r)=u_\mathrm{rp}(t,r), \text{ for all }(t,r)\in\mathbb{R}^2,
\end{equation}
where $T_\mathrm{rp}$ is to be determined.
We replace $u$ with $u_h+\delta u$ in \eqref{e:EPP} and consider the equation of the perturbation $\delta u$. For brevity, we denote the
the perturbation by ``$u$'', instead of ``$\delta u$''. The perturbation solves the system
\begin{equation}
\label{e:EPU}
\mathcal{L}u+\mathcal{F}(u)=0,
\end{equation}
where the linear operator
\begin{equation}
\mathcal{L}:=\left(\mathcal{L}_h+\lambda_0\partial_t^2+\varepsilon\eta_1\right)\left(\mathcal{L}_h+\lambda_0\partial_t^2\right)+\mathcal{M},
\end{equation}
is expressed in terms of the second order operator, $\mathcal{L}_h:=\partial_r^2-W^{{\prime\prime}}(u_h)$ and the potential
\[\mathcal{M}:=\varepsilon\eta_d W^{{\prime\prime}}(u_h)-\left(\partial_r^2u_h-W^\prime(u_h)\right)W^{{\prime\prime\prime}}(u_h),\]
while the nonlinearity given by
\begin{equation}
\label{e:F}
\begin{aligned}
\mathcal{F}(u,\varepsilon):=&-\lambda_0W^{{\prime\prime\prime}}(u_h+u)\left(\partial_tu\right)^2-2\lambda_0\left(W^{\prime\prime}(u_h+u)-W^{\prime\prime}(u_h)\right)\partial_t^2u-\\
&\left[\mathcal{L}_h+\varepsilon(\eta_1-\eta_d)-\left(W^{{\prime\prime}}(u_h+u)-W^{{\prime\prime}}(u_h)\right)\right]\left(W^\prime(u_h+u)-W^\prime(u_h)-W^{{\prime\prime}}(u_h)u\right)-\\
&\left(W^{{\prime\prime}}(u_h+u)-W^{{\prime\prime}}(u_h)\right)\mathcal{L}_h u-
\left(\partial_r^2u_h-W^\prime(u_h)\right)\left(W^{{\prime\prime}}(u_h+u)-W^{\prime\prime}(u_h)-W^{\prime\prime\prime}(u_h)u\right).
\end{aligned}
\end{equation}
We recast the system \eqref{e:EPU} in the vector form
\begin{equation}
\label{e:2ddFCH}
\dot{U}=\mathbb{L}(\varepsilon)U+\mathbb{F}(U,\varepsilon),
\end{equation}
using the transformation $U_1=u$, $U_2=u_t$, $U_3=\lambda_0u_{tt}+\mathcal{L}_h u$, $U_4=\partial_t\left(\lambda_0u_{tt}+\mathcal{L}_h u\right)$ and introducing
\[
U = \begin{pmatrix}U_1\\U_2\\U_3\\U_4\end{pmatrix},\quad
\mathbb{L}(\varepsilon) = \begin{pmatrix}
0&1&0&0\\ -\frac{1}{\lambda_0}\mathcal{L}_h&0&\frac{1}{\lambda_0}&0\\ 0&0&0&1\\
-\frac{1}{\lambda_0}\mathcal{M}& 0&
-\frac{1}{\lambda_0}(\mathcal{L}_h+\varepsilon\eta_1)&0\end{pmatrix},\quad
\mathbb{F}(U,\varepsilon) = \begin{pmatrix}0\\0\\0\\-\frac{1}{\lambda_0}\mathcal{F}\end{pmatrix}.
\]
\begin{Remark}
To avoid technicalities we search for $u_p$ for a fixed value of $\gamma$. It is straightforward to recover the smooth dependence
of $u_p$ with respect to $\gamma$.
\end{Remark}
We observe that, for given small $\varepsilon$,
$\mathbb{L}(\varepsilon): \mathcal{D}(\mathbb{L})\rightarrow \mathcal{X}$ is a closed operator defined in the Hilbert space $\mathcal{X}$ with
its domain $\mathcal{D}(\mathbb{L})=\mathcal{Y}$, where
\[
\mathcal{X}=H^3(\mathbb{R})\times H^2(\mathbb{R}) \times H^1(\mathbb{R}) \times L^2(\mathbb{R}), \quad
\mathcal{Y}=H^4(\mathbb{R}) \times H^3(\mathbb{R})\times H^2(\mathbb{R}) \times H^1(\mathbb{R}).
\]
In the sequel we replace $\partial_t u$ and $\partial_t^2 u$ with $U_2$ and
$\frac{1}{\lambda_0}(U_3-\mathcal{L}_h U_1)$, respectively, in equation \eqref{e:F} for $\mathcal{F}$.
The map $\mathbb{F}: \mathcal{Y}\times [-\varepsilon_0,\varepsilon_0]\rightarrow\mathcal{Y}$ is smooth, for
$\varepsilon_0>0$ is sufficiently small.
\begin{Lemma}
\label{l:spec}
The spectrum of $\mathbb{L}_*:=\mathbb{L}(0,0)$, $\sigma(\mathbb{L}_*)$, as shown in Figure \ref{f:specL}, satisfies
\begin{itemize}
\item[(\rmnum{1})]$\sigma_c(\mathbb{L}_*):=\sigma(\mathbb{L}_*)\cap i\mathbb{R}=\{0, \pm i\}$, where eigenvalue $0$ has geometric multiplicity 1 and algebraic multiplicity 4, and eignvalues
$\pm \mathrm{i}$ have geometric multiplicity 1 and algebraic multiplicity 2.
\item[(\rmnum{2})]There exists $\eta>0$ such that $\sigma(\mathbb{L}_*)\cap \{|\mathop{\mathrm{Re}}\lambda|\leq \eta\}=\sigma_c(\mathbb{L}_*)$.
\end{itemize}
\end{Lemma}
\begin{figure}[H]
\centering
\includegraphics[width=0.40\textwidth]{specL.png}
\caption{The spectrum of $\mathbb{L}_*$ indicating the center eigenvalues and their multiplicity.}
\label{f:specL}
\end{figure}
\begin{Proof}
We first introduce the operator
\begin{equation*}
\begin{matrix}
\mathcal{L}^\lambda: & H^4(\mathbb{R}) & \longrightarrow & L^2(\mathbb{R}) \\
& u & \longmapsto & \left(\mathcal{L}_0+\lambda_0\lambda^2\right)^2u
\end{matrix}
\end{equation*}
which, for any $\lambda\in\mathbb{C}$, has the same Fredholm properties as the operator $\mathbb{L}_*-\lambda\mathrm{\,Id}\,$; see a similar case in \cite{ssmorse_2008} for
a detailed proof. More specifically, $\mathbb{L}_*-\lambda\mathrm{\,Id}\,$ is Fredholm if and only if $\mathcal{L}^\lambda$ is Fredholm. In addition, if Fredholm, then $\mathbb{L}_*-\lambda\mathrm{\,Id}\,$ and $\mathcal{L}^\lambda$ have the same Fredholm index. We omit the technical details required to establish that
$\mathop{\mathrm{dim}}\mathop{\mathrm{CoKer}} (\mathbb{L}_*-\lambda\mathrm{\,Id}\,)=\mathop{\mathrm{dim}}\mathop{\mathrm{CoKer}} \mathcal{L}^\lambda$; however it is straightforward to see that
$$\mathop{\mathrm{dim}}\ker (\mathbb{L}_*-\lambda\mathrm{\,Id}\,)=\mathop{\mathrm{dim}}\ker \mathcal{L}^\lambda,$$
since
\[
(\mathbb{L}_*-\lambda\mathrm{\,Id}\,)\begin{pmatrix}U_1\\ U_2 \\ U_3 \\ U_4 \end{pmatrix}=0 \Longleftrightarrow \mathcal{L}^\lambda U_1=0.
\]
To obtain the spectral properties of $\mathbb{L}_*$, the dispersion relation of $\mathcal{L}^\lambda$ implies that
\[
\sigma(\mathbb{L}_*)=\{\lambda\in\mathbb{C}\mid (\mu+\lambda_0\lambda^2)^2=0, \text{ for some }\mu\in\sigma(\mathcal{L}_0)\},
\]
where $\mathcal{L}_0$, defined in \eqref{cL_0-def}, is of Sturm-Liouville type with simple, real spectrum thats satisfies
\[
\sigma(\mathcal{L}_0)\cap \{\mathop{\mathrm{Re}}\lambda\geq 0\}=\{0,\lambda_0\}, \quad
\sigma(\mathcal{L}_0)\cap \{\mathop{\mathrm{Re}}\lambda< 0\}\subset (\infty,-c), \text{ for some }c>0.
\]
These observations conclude the proof.
\end{Proof}
The center space $\mathcal{X}_c$ of $\mathbb{L}_*$, that is, the spectral subspace associated to $\sigma_c(\mathbb{L}_*)$, is 8-dimensional
and spanned by the eigenfunctions $\{E_1, E_2,\bar{E}_1,\bar{E}_2, F_1, F_2, F_3, F_4\}$, where
\begin{equation}
\begin{aligned}
E_1=\begin{pmatrix}1\\ \mathrm{i} \\ 0 \\ 0 \end{pmatrix}\psi_0,\quad
E_2=\begin{pmatrix}\mathrm{i}\\ 0 \\ 2\lambda_0\mathrm{i} \\ -2\lambda_0 \end{pmatrix}\psi_0,\quad
F_1=\begin{pmatrix}1 \\ 0 \\ 0 \\ 0 \end{pmatrix}\psi_1, \\
F_2=\begin{pmatrix}0 \\ 1 \\ 0 \\ 0 \end{pmatrix}\psi_1,\quad
F_3=\begin{pmatrix}0 \\ 0 \\ \lambda_0 \\ 0 \end{pmatrix}\psi_1,\quad
F_4=\begin{pmatrix}0 \\ 0 \\ 0 \\ \lambda_0 \end{pmatrix}\psi_1.
\end{aligned}
\end{equation}
Moreover, these generalized eigenfunctions of $\mathbb{L}_*$ satisfies
\begin{equation}\label{e:eigen}
\begin{aligned}
(\mathbb{L}_*-\mathrm{i})E_1=0, \quad (\mathbb{L}_*-\mathrm{i})E_2=E_1, \quad \mathbb{L}_*F_1=0, \quad \mathbb{L}_*F_2=F_1,\text{ }\\
(\mathbb{L}_*+\mathrm{i})\bar{E}_1=0, \quad (\mathbb{L}_*+\mathrm{i})\bar{E}_2=\bar{E}_1, \quad \mathbb{L}_*F_3=F_2, \quad \mathbb{L}_*F_4=F_3,\\
S_1^2=\mathrm{\,Id}\,,\quad S_1E_1=\bar{E}_1,\quad S_1E_2=-\bar{E}_2,\quad S_1F_j=F_j, \quad S_1F_k=-F_k, j=1,3; k=2,4,\\
S_2^2=\mathrm{\,Id}\,,\quad S_2E_j=E_j,\quad S_2\bar{E}_j=\bar{E}_j,\quad S_2F_k=-F_k, j=1,2; k=1,2,3,4.
\end{aligned}
\end{equation}
where $S_1$ and $S_2$ are the symmetries inherited from the $t\rightarrow-t$ and $r\rightarrow -r$ symmetries of the original PDE \eqref{e:EPP}. Here $S_1$ is a reversible symmetry and plays a crucial role in the subsequent bifurcation analysis. From \eqref{e:eigen} we develop an explicit expression of the spectral projection $\mathbb{P}_c:\mathcal{X}\rightarrow \mathcal{X}_c$,
\begin{equation}\label{e:uc0}
\begin{aligned}
U_c:=\mathbb{P}_c U=& \langle U, E_1^{\mathrm{ad}}\rangle E_1+ \langle U,E_2^{\mathrm{ad}}\rangle E_2+ \langle U,\bar{E}_1^{\mathrm{ad}}\rangle \bar{E}_1+ \langle U,\bar{E}_2^{\mathrm{ad}}\rangle \bar{E}_2+\\
& \langle U, F_1^{\mathrm{ad}}\rangle F_1+ \langle U,F_2^{\mathrm{ad}}\rangle F_2+ \langle U,F_3^{\mathrm{ad}}\rangle F_3+ \langle U,F_4^{\mathrm{ad}}\rangle F_4,
\end{aligned}
\end{equation}
where
\begin{equation}
\begin{aligned}
E_1^{\mathrm{ad}}=\begin{pmatrix} \frac{1}{2} \\ \frac{\mathrm{i}}{2} \\ -\frac{1}{4\lambda_0} \\ 0 \end{pmatrix}\psi_0,\quad
E_2^{\mathrm{ad}}=\begin{pmatrix} 0 \\ 0 \\ \frac{\mathrm{i}}{4\lambda_0} \\ -\frac{1}{4\lambda_0} \end{pmatrix}\psi_0,\quad
F_1^{\mathrm{ad}}=\begin{pmatrix}1 \\ 0 \\ 0 \\ 0 \end{pmatrix}\psi_1, \\
F_2^{\mathrm{ad}}=\begin{pmatrix}0 \\ 1 \\ 0 \\ 0 \end{pmatrix}\psi_1,\quad
F_3^{\mathrm{ad}}=\begin{pmatrix}0 \\ 0 \\ \frac{1}{\lambda_0} \\ 0 \end{pmatrix}\psi_1,\quad
F_4^{\mathrm{ad}}=\begin{pmatrix}0 \\ 0 \\ 0 \\ \frac{1}{\lambda_0} \end{pmatrix}\psi_1.
\end{aligned}
\end{equation}
These vector functions with superscript ``$\mathrm{ad}$'' are generalized eigenfunctions of the adjoint operator $\mathbb{L}_*^\mathrm{ad}$ associated to $0$ and $\pm\mathrm{i}$
in $(L^2(\mathbb{R}))^4$ with canonical inner product $\langle\cdot,\cdot\rangle$.
Moreover, a standard calculation \cite{kappro} shows that, for any given $w_0>1$, there exists $C\geq1$ such that
\begin{equation}
\label{e:nest}
\|(\mathrm{i} w - \mathbb{L}_*)^{-1}U\|_{\mathcal{X}}\leq \frac{C}{|w|} \|U\|_{\mathcal{X}}, \text{ for all } |w|\geq w_0, w\in\mathbb{R}, U\in (\mathrm{\,Id}\,-\mathbb{P}_c)\mathcal{X}.
\end{equation}
Therefore, based on Lemma \ref{l:spec} and the norm estimate \eqref{e:nest} on $\mathbb{L}_*|_{(\mathrm{\,Id}\,-\mathbb{P}_c)\mathcal{X}}$,
we can apply the center manifold reduction theorem to the system \eqref{e:2ddFCH} and obtain the following proposition (see \cite[Theorem 2.9]{harioo}).
\begin{Proposition}
Given any fixed $\gamma$ and $k\in\mathbb{Z}^+$, there exist open sets containing the origin $\mathcal{U}\subset \mathcal{X}_c$, $\mathcal{V}\subset (\mathrm{\,Id}\,-\mathbb{P}_c)\mathcal{Y}$, $\mathcal{W}\in\mathbb{R}$,
and a $C^k$-smooth map $\Psi:\mathcal{U}\times\mathcal{W}\rightarrow\mathcal{V}$, for any fixed nonnegative integer $k$, such that the center manifold $\mathcal{M}_c$,
that is, the graph of the map $\Psi$, has the following properties.
\begin{itemize}
\item[(\rmnum{1})]The center manifold $\mathcal{M}_c$ is tangent to the center eigenspace $\mathcal{X}_c$,
\begin{equation}
\|\Psi(U_c,\varepsilon)\|_\mathcal{Y}=\mathcal{O}(|\varepsilon|\|U_c\|+\|U_c\|^2).
\end{equation}
\item[(\rmnum{2})]The center manifold $\mathcal{M}_c$ is locally invariant, that is,
if $U$ is a solution to \eqref{e:2ddFCH} with $U(0)\in\mathcal{M}_c$ and
$U(t)\in \mathcal{U}\times\mathcal{V}$ for $t\in[0, T]$, then $U(t)\in\mathcal{M}_c$ for all $t\in[0,T]$.
\item[(\rmnum{2})]The center manifold $\mathcal{M}_c$ contains all bounded solutions to \eqref{e:2ddFCH} with $\mathbb{R}$ as the existence interval, that is,
if $U$ is a solution to \eqref{e:2ddFCH} satisfying $\{U(t)\mid t\in\mathbb{R}\}\subset \mathcal{U}\times\mathcal{V}$, then $\{U(t)\mid t\in\mathbb{R}\}\subset\mathcal{M}_c$.
\end{itemize}
\end{Proposition}
\subsection{Reduced center manifold ODE}\label{ss:22}
In this section we calculate the reduced ODE system obtained by restricting \eqref{e:2ddFCH} to the center manifold. From the analysis presented
in Section~\ref{ss:21} and summarized in Figure~\ref{f:specL} it follows that the reduced ODE system is of $8$-th order which can be viewed as a
coupling of two four-dimensional systems which exhibit the so-called ``reversible-Hopf bifurcation'' and the ``reversible $0^{4+}$ bifurcation''.
Moreover, the coupling occurs at the nonlinear level and is weak.
On the linear level, the $S_1$-reversibility of the reduction to the $\pm\mathrm{i}$-eigenspace gives rise to the ``reversible-Hopf bifurcation'', which is
well-studied; see \cite{glebsky_1995,iooper_1993}; while the $S_1$-reversibility of the $0$-eigenspace gives rise to the ``reversible $0^{4+}$ bifurcation'', whose study is quite open; see \cite{harioo}. Fortunately, extended pearled solutions result from the ``reversible-Hopf bifurcation''.
Moreover, it is known that the analysis of this bifurcation relies on the coefficients of the cubic terms in the norm form \cite{iooper_1993}.
Therefore, all the necessary terms of the reduced ODE system, up to cubic order, are explicitly determined in this section.
To restrict the system \eqref{e:2ddFCH} to the center manifold we consider $U$ in the form
\begin{equation}
\label{e:sch}
U=U_c+\Psi(U_c,\varepsilon).
\end{equation}
Substituting this form \eqref{e:sch} into \eqref{e:2ddFCH} and applying the projection $\mathbb{P}_c$, we obtain the reduced equation,
\begin{equation}
\label{e:2dred}
\dot{U}_c=\mathbb{L}_*U_c+\mathbb{P}_c\Big(\mathbb{M}(\varepsilon)\big(U_c+\Psi(U_c,\varepsilon)\big)+\mathbb{F}\big(U_c+\Psi(U_c,\varepsilon),\varepsilon\big)\Big),
\end{equation}
where $\mathbb{M}(\varepsilon):=\mathbb{L}(\varepsilon)-\mathbb{L}_*$.
Moreover, from \eqref{e:uc0}, we note that $U_c$ admits the general expression
\begin{equation}
\label{e:ucg}
U_c(t)=\sum_{j=1}^2\left(A_j(t)E_j+\bar{A}_j(t)\bar{E}_j\right)+\sum_{k=1}^4B_k(t)F_k,
\end{equation}
Using this expression of $U_c$, we rewrite the reduced system \eqref{e:2dred} explicitly in terms of
\begin{equation}\label{e:bfA}
\mathbf{A}:=(A_1,A_2,\bar{A}_1,\bar{A}_2,B_1,B_2,B_3,B_4).
\end{equation}
We summarize the essential result into Lemma \ref{l:sim}, relegating the detailed results and concomitant calculations to Appendix \ref{ss:31}.
The principle technicality in the calculation lies in finding the explicit expression of
$\Psi_{(2,0,0)}(U_c,U_c)$ in terms of $\mathbf{A}$; see Lemma \ref{l:31} for details.
\begin{Lemma}\label{l:sim}
The reduced system \eqref{e:2dred}, in terms of ${\bf A}$, called the reduced ODE system, admits the expression
\begin{equation}
\label{e:2dcredA}
\dot{\mathbf{A}}=\mathbf{L}(\varepsilon)\mathbf{A}+\mathbf{R}_2(\mathbf{A})
+\mathbf{R}_3(\mathbf{A})+ \mathcal{O}\left(|\varepsilon|^2\|{\bf A}\|+|\varepsilon|\|{\bf A}\|^2+\|{\bf A}\|^4\right),
\end{equation}
where the linear term ${\bf L}$, the quadratic term ${\bf R}_2$, the cubic term ${\bf R}_3$ are of the following expressions.
\begin{equation}\label{e:bfL}
\mathbf{L}(\varepsilon)=\begin{pmatrix}
\mathrm{i}(1+\mu_1\varepsilon) & 1-\mu_1\varepsilon & \mathrm{i} \mu_1\varepsilon & \mu_1\varepsilon
& 0 & 0 & 0 &0 \\
\mu_2\varepsilon & \mathrm{i} \left(1+\mu_3\varepsilon \right) & \mu_2\varepsilon & -\mathrm{i} \mu_3 \varepsilon
& 0 & 0 & 0 &0 \\
-\mathrm{i} \mu_1\varepsilon & \mu_1 \varepsilon& -\mathrm{i}\left(1+\mu_1\varepsilon\right) & 1+\mu_1\varepsilon
& 0 & 0 & 0 &0 \\
\mu_2\varepsilon & \mathrm{i} \mu_3\varepsilon &\mu_2 \varepsilon& -\mathrm{i} \left(1+\mu_3 \varepsilon\right) & 0 & 0 & 0 &0 \\
0 & 0 & 0 &0 & 0 & 1 & 0 & 0\\
0 & 0 & 0 &0 & \mu_4 \varepsilon& 0 & 1 & 0\\
0 & 0 & 0 &0 & 0 &0 & 0 & 1\\
0 & 0 & 0 &0 & \mu_5\varepsilon & 0 & \mu_6\varepsilon & 0\\
\end{pmatrix},
\end{equation}
\begin{equation*}
\mathbf{R}_2(\mathbf{A})=\left(0,R_{2,2},0,\bar{R}_{2,2},0,0,0,R_{2,8}\right)^T, \quad
\mathbf{R}_3({\bf A})=\left(0,R_{3,2},0,\bar{R}_{3,2},0,0,0,R_{3,8}\right)^T,
\end{equation*}
where the expressions of every $\mu_j\in \mathbb{R}$ and $R_{2\backslash 3, 2\backslash 8}$ in terms of ${\bf A}$ can be found in Lemma \ref{l:reducedA}.
\end{Lemma}
\subsection{Norm forms}\label{ss:23}
We obtain a normal form of the leading-order-term reduced system via a composition of a linear versal transformation and a near-identity nonlinear transformation. The versal transformation allows a Jordan-form type decomposition which is smooth in the parameters, see \cite{viarnold} for full details.
\begin{Lemma}\label{l:versal}
For sufficiently small $\varepsilon$, there exists a smooth linear map $\mathbf{T}(\varepsilon)$ with $\mathbf{T}(0)=\mathrm{\,Id}\,$ such that
under the transformation
\[
\mathbf{A}=\mathbf{T}(\varepsilon)\mathbf{C}, \quad \mathbf{C}=(C_1,C_2,\bar{C}_1,\bar{C}_2, D_1,D_2,D_3,D_4)^T,
\]
the linear part of \eqref{e:2dcredA} in $\mathbf{A}$, that is,
\begin{equation}
\label{e:l2dcred}
\dot{\mathbf{A}}=\mathbf{L}(\varepsilon)\mathbf{A},
\end{equation}
takes the versal normal form
\begin{equation}
\dot{\mathbf{C}}=\mathscr{L}(\varepsilon)\mathbf{C}+\mathcal{O}\left(|\varepsilon|^2\|{\bf C}\|\right),
\end{equation}
where
\begin{equation}
\label{e:ln2dcred}
\mathscr{L}(\varepsilon)=\begin{pmatrix}
\mathrm{i}\left(1+\omega_1\varepsilon\right) & 1 & 0 & 0 & 0 & 0 & 0 & 0 \\
\omega_2\varepsilon& \mathrm{i}\left(1+\omega_1\varepsilon\right) & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & -\mathrm{i}\left(1+\omega_1\varepsilon\right) & 1 & 0 & 0 & 0 & 0 \\
0 & 0 & \omega_2 \varepsilon & -\mathrm{i}\left(1+\omega_1\varepsilon\right) & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 \\
0 & 0 & 0 & 0 & \omega_3\varepsilon & 0 & \omega_4\varepsilon & 0 \\
\end{pmatrix}.
\end{equation}
Here we have introduced
\begin{equation}\label{e:ls}
\omega_1=\frac{1}{2}(\mu_1+\mu_3),\quad \omega_2=\mu_2,\quad \omega_3=\mu_5,\quad \omega_4=\mu_4+\mu_6,
\end{equation}
where the expression of each $\mu_j\in\mathbb{R}$ can be found in Lemma \ref{l:reducedA}.
\end{Lemma}
\begin{Proof}
We point out that $\mathbf{L}(\varepsilon)$ inherits the symmetries $\tau\rightarrow -\tau$ and $r\rightarrow -r$ of the original PDE \eqref{e:2sFCH}, that is,
\[
S_1\mathbf{L}(\varepsilon)=-\mathbf{L}(\varepsilon)S_1,\quad S_2\mathbf{L}(\varepsilon)=\mathbf{L}(\varepsilon)S_2,
\]
where
\begin{equation*}
\begin{aligned}
S_1(A_1,A_2,\bar{A}_1,\bar{A}_2, B_1,B_2,B_3,B_4)^T=(\bar{A}_1,-\bar{A}_2,A_1,-A_2, B_1,-B_2,B_3,-B_4)^T,\\
S_2(A_1,A_2,\bar{A}_1,\bar{A}_2, B_1,B_2,B_3,B_4)^T=(A_1,A_2,\bar{A}_1,\bar{A}_2, -B_1,-B_2,-B_3,-B_4)^T.
\end{aligned}
\end{equation*}
Then, according to \cite[Theorem 4.4]{viarnold}, a versal deformation of the Jordan normal form $\mathbf{L}$ keeping the symmetries can be chosen in the form
\begin{equation*}
\begin{pmatrix}
\mathrm{i}\left(1+\widetilde{\omega}_1\right) & 1 & 0 & 0 & 0 & 0 & 0 & 0 \\
\widetilde{\omega}_2& \mathrm{i}\left(1+\widetilde{\omega}_1\right) & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & -\mathrm{i}\left(1+\widetilde{\omega}_1\right) & 1 & 0 & 0 & 0 & 0 \\
0 & 0 & \widetilde{\omega}_2 & -\mathrm{i}\left(1+\widetilde{\omega}_1\right) & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 \\
0 & 0 & 0 & 0 & \widetilde{\omega}_3 & 0 & \widetilde{\omega}_4 & 0 \\
\end{pmatrix}.
\end{equation*}
where $\widetilde{\omega}_j(\varepsilon)\in\mathbb{R}$ with $\omega_j(0)=0$ for $j=1,2,3,4$.
Comparing the coefficients of the characteristic polynomials of the two $4\times 4$ diagonal blocks associated to $(A_1,A_2, \bar{A}_1, \bar{A}_2)$ in \eqref{e:bfL} and \eqref{e:ln2dcred}, we have
\begin{equation*}
\begin{cases}
(1+\widetilde{\omega}_1)^2-\widetilde{\omega}_2=1+(\mu_1-\mu_2+\mu_3)\varepsilon,\\
\left((1+\widetilde{\omega}_1)^2+\widetilde{\omega}_2\right)^2=1+2(\mu_1+\mu_2+\mu_3)\varepsilon-4(\mu_2-\mu_3)\mu_1\varepsilon^2,
\end{cases}
\end{equation*}
from which we have
\begin{equation*}
\begin{aligned}
\widetilde{\omega}_1(\varepsilon)&=\frac{1}{2}(\mu_1+\mu_3)\varepsilon+\mathcal{O}\left(|\varepsilon|^2\right),
& \widetilde{\omega}_2(\varepsilon)&=\mu_2\varepsilon+\mathcal{O}\left(|\varepsilon|^2\right)
\end{aligned}
\end{equation*}
Similarly, we have
\begin{equation*}
\begin{aligned}
\widetilde{\omega}_3(\varepsilon)&=\mu_5\varepsilon+\mathcal{O}\left(|\varepsilon|^2\right)
& \widetilde{\omega}_4(\varepsilon)&=(\mu_4+\mu_6)\varepsilon+\mathcal{O}\left(|\varepsilon|^2\right)
\end{aligned}
\end{equation*}
We truncate this versal deformation up to linear terms in $\varepsilon$, denote it as $\mathscr{L}(\varepsilon)$ and conclude our proof.
\end{Proof}
On the other hand, we have the following nonlinear normal form.
\begin{Lemma}
There exist smooth families of degree-$2$ polynomials
\[
\Phi_2=(\Phi_{2,1},\Phi_{2,2},\Phi_{2,3},\Phi_{2,4},\Phi_{2,5},\Phi_{2,6},\Phi_{2,7},\Phi_{2,8})^T,
\]
and degree-$3$ polynomials
\[
\Phi_3=(\Phi_{3,1},\Phi_{3,2},\Phi_{3,3},\Phi_{3,4},\Phi_{3,5},\Phi_{3,6},\Phi_{3,7},\Phi_{3,8})^T,
\]
in terms of $\mathbf{C}$ such that
such that under the near-identity transformation
\begin{equation}\label{e:nonl}
\mathbf{A}=\mathbf{C}+\Phi_2(\mathbf{C})+\Phi_3(\mathbf{C}),
\end{equation}
the nonlinear part of \eqref{e:2dcredA}, that is,
\begin{equation}
\label{e:q2dcred}
\dot{\mathbf{A}}=\mathbf{L}(0)\mathbf{A}+\mathbf{R}_2(\mathbf{A},\mathbf{A})
+\mathbf{R}_3(\mathbf{A},\mathbf{A},\mathbf{A}),
\end{equation}
takes the normal form
\begin{equation}\label{e:nfnonl}
\dot{\mathbf{C}}=\mathbf{L}(0)\mathbf{C}+\mathscr{R}_2(\mathbf{C},\mathbf{C})
+\mathscr{R}_3(\mathbf{C},\mathbf{C},\mathbf{C})+\mathcal{O}(|\mathbf{C}|^4).
\end{equation}
Here $\mathscr{R}_2=0$ and
$\mathscr{R}_3=(\mathscr{R}_{3,1},\mathscr{R}_{3,2},\mathscr{R}_{3,3},\mathscr{R}_{3,4},\mathscr{R}_{3,5},\mathscr{R}_{3,6},\mathscr{R}_{3,7},\mathscr{R}_{3,8})^T$ is of the form
\begin{equation}
\label{e:nq2dcred}
\begin{aligned}
\mathscr{R}_{3,1}=&\mathrm{i} \Big\{C_1\big[\alpha_7C_1\bar{C}_1+\alpha_8\mathrm{i}(C_1\bar{C}_2-\bar{C}_1C_2)\big]+\alpha_9C_1D_1^2+
\alpha_{10}\mathrm{i} D_1(C_2D_1-C_1D_2)+\\
& \alpha_{11}C_1(2D_1D_3-D_2^2)+\alpha_{12}\mathrm{i}\big[C_1(D_2D_3-3D_1D_4)+C_2(2D_1D_3-D_2^2)\big]\Big\};\\
\mathscr{R}_{3,2}=&\Big\{C_1\big[\alpha_1C_1\bar{C}_1+\alpha_2\mathrm{i}(C_1\bar{C}_2-\bar{C}_1C_2)\big]+\alpha_3C_1D_1^2+
\alpha_{4}\mathrm{i} D_1(C_2D_1-C_1D_2)+\\
&\alpha_5C_1(2D_1D_3-D_2^2)+\alpha_6\mathrm{i}\big[C_1(D_2D_3-3D_1D_4)+C_2(2D_1D_3-D_2^2)\big]\Big\}+\\
&\mathrm{i}\Big\{C_2\big[\alpha_7C_1\bar{C}_1+\alpha_8\mathrm{i}(C_1\bar{C}_2-\bar{C}_1C_2)\big]+\alpha_9C_2D_1^2+
\alpha_{10}\mathrm{i} D_2(C_2D_1-C_1D_2)+\\
& \alpha_{11}C_1(3D_1D_4-D_2 D_3)+\alpha_{12}\mathrm{i}\big[2C_1(2D_3^2-3D_2D_4)+C_2(3D_1D_4-D_2D_3)\big]\Big\};\\
\mathscr{R}_{3,8}=&D_1(\beta_{1}C_1\bar{C}_1+\beta_{2}C_2\bar{C}_2)+\mathrm{i}(C_1\bar{C}_2-\bar{C}_1C_2)(\beta_3D_1+\beta_4D_3)+\beta_5C_1\bar{C}_1D_3+\\
&\beta_6D_2(C_1\bar{C}_2+\bar{C}_1C_2)+\beta_7[3(C_1\bar{C}_2+\bar{C}_1C_2)D_4-2C_2\bar{C}_2D_3]+\beta_8D_1D_2^2+\\
& D_1^2(\beta_9D_1+\beta_{10}D_3)+
\beta_{11}(D_2^2D_3-2D_1D_3^2)+\beta_{12}(D_2^2D_3-3D_1D_2D_4)+\\
&\beta_{13}(9D_2D_3D_4-9D_1D_4^2-4D_3^3);\\
\mathscr{R}_{3,{j+2}}=&\bar{\mathscr{R}}_{3,j}, \quad j=1,2; \quad \quad \mathscr{R}_{3,k}=0,\quad k= 5,6,7; \\
\end{aligned}
\end{equation}
where the explicit expressions of the coefficients $\alpha_j$,$\beta_k\in\mathbb{R}$, are given in Lemma \ref{l:32}. Moreover, the transformation preserves the reversibility $S_1$ and the symmetry $S_2$.
\end{Lemma}
\begin{Proof}
Following \cite[Chapter 3]{harioo}, we cast the normal form problem as a solvability issue on a space of polynomials in ${\bf C}$ which is expressed in terms of the Fredholm alternative of the operator
\begin{equation}
\mathcal{D}=\left(\mathrm{i} C_1+C_2\right)\frac{\partial}{\partial C_1}+\mathrm{i} C_2\frac{\partial}{\partial C_2}+
\left(-\mathrm{i} \bar{C}_1+\bar{C}_2\right)\frac{\partial}{\partial \bar{C}_1}+(-\mathrm{i} \bar{C}_2)\frac{\partial}{\partial \bar{C}_2}+
\sum_{j=1}^3D_{j+1}\frac{\partial}{\partial D_i}.
\end{equation}
For convenience, we introduce the polynomial space $\mathbf{P}_j$, $j=2,3$, which is the set of all degree-$j$ homogeneous polynomials in $\mathbf{C}$,
with the inner product
\[
\langle P\mid Q\rangle=P(\partial_\mathbf{C})\bar{Q}(\mathbf{C})|_{\mathbf{C}=0}.
\]
We point out here that the conjugacy $\bar{Q}$ only acts on the coefficients, in the sense that, for example, for $Q({\bf C})=\mathrm{i} C_1^2$, $\bar{Q}({\bf C})=-\mathrm{i} C_1^2$.
More specifically, plugging \eqref{e:nonl} and \eqref{e:nfnonl} into \eqref{e:q2dcred}, we obtain the following two equalities.
\begin{eqnarray}
\big(\mathcal{D}-\mathbf{L}(0)\big)\Phi_2&=&\mathbf{R}_2-\mathscr{R}_2,\qquad\qquad\qquad\qquad\qquad\quad\quad \label{e:nf2}\\
\big(\mathcal{D}-\mathbf{L}(0)\big)\Phi_3&=&\mathbf{R}_3+
2\mathbf{R}_2(\mathbf{C},\Phi_2)-\mathscr{R}_3-\big(D_{\mathbf{C}}\Phi_2\big)\mathscr{R}_2, \label{e:nf3}
\end{eqnarray}
From the Fredholm alternative, we may solve for $\Phi_{2}$ and $\mathscr{R}_{2}$ uniquely in \eqref{e:nf2} subject to
\[
\mathscr{R}_2\in\ker\left(\big(\mathcal{D}^{\mathrm{ad}}-\mathbf{L}^{\mathrm{ad}}(0)\big)|_{\mathbf{P}_2^8}\right),\quad \Phi_2\in \left(\ker\left((\mathcal{D}-\mathbf{L}(0))|_{\mathbf{P}_2^8}\right)\right)^\perp,
\]
where
\[
\mathcal{D}^{\mathrm{ad}}=(-\mathrm{i} C_1)\frac{\partial}{\partial C_1}+\left( C_1 - \mathrm{i} C_2\right)\frac{\partial}{\partial C_2}+
\mathrm{i} \bar{C}_1\frac{\partial}{\partial \bar{C}_1}+\left( \bar{C}_1+ \mathrm{i} \bar{C}_2\right)\frac{\partial}{\partial \bar{C}_2}+
\sum_{j=2}^4D_{j-1}\frac{\partial}{\partial D_i}.
\]
In fact, we claim that $\mathscr{R}_2=0$. To show this, we only need to verify that
\[
\mathbf{R}_2\in \mathop{\mathrm{Rg}}\left((\mathcal{D}-\mathbf{L}(0))|_{\mathbf{P}_2^8}\right)=\left(\ker\left((\mathcal{D}^{\mathrm{ad}}-\mathbf{L}^{ad}(0))|_{\mathbf{P}_2^8}\right)\right)^\perp,
\]
which follows from the expression of ${\bf R}_2$ in \eqref{e:Rjs} and the fact that
\begin{equation}\
\label{e:ker2}
\begin{aligned}
\ker\big(\mathcal{D}^{\mathrm{ad}}|_{\mathbf{P}_2}\big)=\mathop{\mathrm{span}}\{&C_1\bar{C}_1,C_1\bar{C}_2-\bar{C}_1C_2,D_1^2,2D_1D_3-D_2^2\},\\
\ker\big((\mathcal{D}^{\mathrm{ad}}+\mathrm{i})|_{\mathbf{P}_2}\big)=\mathop{\mathrm{span}}\{&C_1D_1,C_1D_2-C_2D_1\}.\\
\end{aligned}
\end{equation}
As a result, we obtain $\Phi_2\in \left(\ker\left((\mathcal{D}-\mathbf{L}(0))|_{\mathbf{P}_2^8}\right)\right)^\perp$ with coefficients,
\begin{equation}
\label{e:phi2}
\begin{aligned}
\Phi_{2,1}(\mathbf{C})=& \frac{2\nu_1}{9}\Big[\big(3 c_{1,+}^2-12 c_{1,-}^2 +14 c_{1,+}c_{2,-}-40 c_{1,-}c_{2,+}-9 c_{2,-}^2\big)\\
& +\mathrm{i}\big(12c_{1,+}c_{1,-} + 32 c_{1,+}c_{2,+} +4 c_{1,-}c_{2,-} \big)\Big]\\
& +\nu_2\Big[\big(-\frac{1}{2}D_1^2+D_1D_3+2D_2^2-2D_2D_4-3D_3^2\big) +2\mathrm{i}\big(D_1D_2-2D_2D_3+2D_3D_4\big)\Big],\\
\Phi_{2,2}(\mathbf{C})=& \frac{2\nu_1}{9}\Big[\big(-18c_{1,+}c_{1,-} + 12 c_{1,+}c_{2,+} + 6c_{1,-}c_{2,-}- 44 c_{2,+}c_{2,-}\big)\\
& +\mathrm{i}\big(9 c_{1,+}^2 -30 c_{1,+}c_{2,-}+24 c_{1,-}c_{2,+}+32 c_{2,+}^2+13 c_{2,-}^2\big)\Big]\\
& +\nu_2\Big[\big(D_1D_2+D_1D_4+D_2D_3-4D_3D_4\big)+\mathrm{i}\big(\frac{1}{2}D_1^2+D_1D_3-2D_2D_4-D_3^2+4D_4^2\big)\Big],\\
\Phi_{2,3}(\mathbf{C})=& \Phi_{2,1}(S_1\mathbf{C})=\bar{\Phi}_{2,1}(\mathbf{C}), \quad \quad
\Phi_{2,4}(\mathbf{C})= -\Phi_{2,2}(S_1\mathbf{C})=\bar{\Phi}_{2,2}(\mathbf{C}),\\
\Phi_{2,5}(\mathbf{C})=& 8\nu_2\left[(c_{1,+}-c_{2,-})D_1-2c_{1,-}D_2-4c_{1,+}D_3+8(c_{1,-}+c_{2,+})D_4 \right],\\
\Phi_{2,6}(\mathbf{C})=& 8\nu_2\left[-c_{1,-}D_1-(c_{1,+}+3c_{2,-})D_2+2(c_{1,-}-2c_{2,+})D_3+4c_{1,+}D_4 \right],\\
\Phi_{2,7}(\mathbf{C})=& 8\nu_2\left[-(c_{1,+}+c_{2,-})D_1-4c_{2,+}D_2+(c_{1,+}+3c_{2,-})D_3-2c_{1,-}D_4 \right],\\
\Phi_{2,8}(\mathbf{C})=& 8\nu_2\left[(c_{1,-}-2c_{2,+})D_1-(c_{1,+}-3c_{2,-})D_2-c_{1,-}D_3-(c_{1,+}-c_{2,-})D_4 \right].\\
\end{aligned}
\end{equation}
Conversely, it is less straightforward to obtain the explicit expression of $\mathscr{R}_3$.
We start by determining a representative form for $\mathscr{R}_3$.
Similar to the quadratic case, from the Fredholm alternative, we solve \eqref{e:nf3} uniquely for $\Phi_{3}$ and $\mathscr{R}_{3}$
subject to
\begin{equation*}
\begin{aligned}
&\begin{pmatrix} \mathscr{R}_{3,1} \\ \mathscr{R}_{3,2}\end{pmatrix}\in\ker\left(\begin{pmatrix}\mathcal{D}^{\mathrm{ad}}+\mathrm{i}&0\\-1&\mathcal{D}^{\mathrm{ad}}+\mathrm{i}\end{pmatrix}\mid_{\mathbf{P}_3^2}\right), \quad \mathscr{R}_{3,3}=\bar{\mathscr{R}}_{3,1}, \quad \mathscr{R}_{3,4}=\bar{\mathscr{R}}_{3,2},\\
&
\mathscr{R}_{3,5}=\mathscr{R}_{3,6}=\mathscr{R}_{3,7}=0, \quad \mathscr{R}_{3,8}\in \ker\big((\mathcal{D}^{\mathrm{ad}})^4|_{\mathbf{P}_3}\big);\\
&\begin{pmatrix} \Phi_{3,1} \\ \Phi_{3,2}\end{pmatrix}\in\left(\ker\left(\begin{pmatrix}\mathcal{D}-\mathrm{i}&-1\\0&\mathcal{D}-\mathrm{i}\end{pmatrix}\mid_{\mathbf{P}_3^2}\right)\right)^\perp, \quad \Phi_{3,3}=\bar{\Phi}_{3,1}, \quad \Phi_{3,4}=\bar{\Phi}_{3,2},\\
&
\Phi_{3,6}=\mathcal{D}\Phi_{3,5}, \quad \Phi_{3,7}=\mathcal{D}^2\Phi_{3,5},\quad \Phi_{3,8}=\mathcal{D}^3\Phi_{3,5}, \quad \Phi_{3,5}\in \left( \ker\big((\mathcal{D})^4|_{\mathbf{P}_3}\big)\right)^\perp.
\end{aligned}
\end{equation*}
Similarly, we point out that
\begin{equation}\
\label{e:ker3}
\begin{aligned}
\ker\big(\mathcal{D}^{\mathrm{ad}}|_{\mathbf{P}_3}\big)=\mathop{\mathrm{span}}\{&C_1\bar{C}_1D_1,(C_1\bar{C}_2-\bar{C}_1C_2)D_1,D_1^3,\\
&D_1(2D_1D_3-D_2^2),3D_1(D_2D_3-D_1D_4)-D_2^3\},\\
\ker\big((\mathcal{D}^{\mathrm{ad}}+\mathrm{i})|_{\mathbf{P}_3}\big)=\mathop{\mathrm{span}}\{&C_1^2\bar{C}_1,C_1(C_1\bar{C}_2-\bar{C}_1C_2),C_1D_1^2,C_1(2D_1D_3-D_2^2),\\
&(C_1D_2-C_2D_1)D_1,C_1(D_2D_3-3D_1D_4)+C_2(2D_1D_3-D_2^2)\}.
\end{aligned}
\end{equation}
Based on \eqref{e:ker3} and the condition that $\mathscr{R}_{3,1\backslash 2}$ satisfies
\begin{equation*}
\begin{pmatrix} \mathscr{R}_{3,1} \\ \mathscr{R}_{3,2}\end{pmatrix}\in\ker\big(\begin{pmatrix}\mathcal{D}^{\mathrm{ad}}+\mathrm{i}&0\\-1&\mathcal{D}^{\mathrm{ad}}+\mathrm{i}\end{pmatrix}\mid_{\mathbf{P}_3^2}\big),
\end{equation*}
we obtain $\mathscr{R}_{3,1\backslash 2}$ in the following form,
\begin{equation}
\begin{aligned}
\mathscr{R}_{3,1}=&C_1\big[\widetilde{\alpha}_7C_1\bar{C}_1+\widetilde{\alpha}_8(C_1\bar{C}_2-\bar{C}_1C_2)\big]+\widetilde{\alpha}_9C_1D_1^2+
\widetilde{\alpha}_{10}D_1(C_2D_1-C_1D_2)+\\
& \widetilde{\alpha}_{11}C_1(2D_1D_3-D_2^2)+\widetilde{\alpha}_{12}\big[C_1(D_2D_3-3D_1D_4)+C_2(2D_1D_3-D_2^2)\big];\\
\mathscr{R}_{3,2}=&C_1\big[\widetilde{\alpha}_1C_1\bar{C}_1+\widetilde{\alpha}_2(C_1\bar{C}_2-\bar{C}_1C_2)\big]+\widetilde{\alpha}_3C_1D_1^2+
\widetilde{\alpha}_{4}D_1(C_2D_1-C_1D_2)+\\
&\widetilde{\alpha}_5C_1(2D_1D_3-D_2^2)+\widetilde{\alpha}_6\big[C_1(D_2D_3-3D_1D_4)+C_2(2D_1D_3-D_2^2)\big]+\\
&C_2\big[\widetilde{\alpha}_7C_1\bar{C}_1+\widetilde{\alpha}_8(C_1\bar{C}_2-\bar{C}_1C_2)\big]+\widetilde{\alpha}_9C_2D_1^2+
\widetilde{\alpha}_{10}D_2(C_2D_1-C_1D_2)+\\
& \widetilde{\alpha}_{11}C_1(3D_1D_4-D_2 D_3)+\widetilde{\alpha}_{12}\big[2C_1(2D_3^2-3D_2D_4)+C_2(3D_1D_4-D_2D_3)\big].\\
\end{aligned}
\end{equation}
Moreover, based on \eqref{e:ker3} and the condition that $\mathscr{R}_{3,5}=\mathscr{R}_{3,6}=\mathscr{R}_{3,7}=0$, and $ \mathscr{R}_{3,8}\in \ker\big((\mathcal{D}^{\mathrm{ad}})^4|_{\mathbf{P}_3}\big)$, $ \mathscr{R}_{3,8}$ takes the form
\begin{equation*}
\begin{aligned}
\mathscr{R}_{3,8}=&\sum_{j, k=1}^2\sum_{\ell=1}^4\widetilde{\beta}_{jk\ell}C_j\bar{C}_kD_\ell+D_1^2\sum_{j=1}^4\widetilde{\beta}_jD_j+
\widetilde{\beta}_5D_1D_2^2+ \widetilde{\beta}_6D_2^3+ \widetilde{\beta}_7D_1D_2D_3+\\
&\widetilde{\beta}_8(D_2^2D_3-2D_1D_3^2)+\widetilde{\beta}_9(D_2^2D_3-3D_1D_2D_4)+\widetilde{\beta}_{10}(2D_1D_3D_4-D_2^2D_4)+\\
&\widetilde{\beta}_{11}(3D_1D_3D_4-D_2D_3^2)+\widetilde{\beta}_{12}(9D_2D_3D_4-9D_1D_4^2-4D_3^3),\\
\end{aligned}
\end{equation*}
where $\widetilde{\beta}_{224}=0$ and $\widetilde{\beta}_{124}+\widetilde{\beta}_{214}+3\widetilde{\beta}_{223}=0$.
We point out that this normal form inherits the symmetries of the original reduced ODE system. More specifically, $\Psi_{2\backslash 3}$ and $\mathscr{R}_{2\backslash 3}$ commute
with $S_1$ and $S_2$; see \cite[3.3]{harioo} for details.
As a result, the preservation of the reversibility $S_1$ further simplifies the cubic term $\mathscr{R}_{3}$. In fact, we have
\begin{equation}
\widetilde{\alpha}_1,\widetilde{\alpha}_3,\widetilde{\alpha}_5, \widetilde{\alpha}_8,\widetilde{\alpha}_{10},\widetilde{\alpha}_{12}\in\mathbb{R},\quad \widetilde{\alpha}_2,\widetilde{\alpha}_4,\widetilde{\alpha}_6, \widetilde{\alpha}_7,\widetilde{\alpha}_{9},\widetilde{\alpha}_{11}\in\mathrm{i}\mathbb{R},
\end{equation}
and
\begin{equation}
\begin{aligned}
\mathscr{R}_{3,8}=&D_1(\beta_{1}C_1\bar{C}_1+\beta_{2}C_2\bar{C}_2)+\mathrm{i}(C_1\bar{C}_2-\bar{C}_1C_2)(\beta_3D_1+\beta_4D_3)+\beta_5C_1\bar{C}_1D_3+\\
&\beta_6D_2(C_1\bar{C}_2+\bar{C}_1C_2)+\beta_7[3(C_1\bar{C}_2+\bar{C}_1C_2)D_4-2C_2\bar{C}_2D_3]+\beta_8D_1D_2^2+\\
& D_1^2(\beta_9D_1+\beta_{10}D_3)+
\beta_{11}(D_2^2D_3-2D_1D_3^2)+\beta_{12}(D_2^2D_3-3D_1D_2D_4)+\\
&\beta_{13}(9D_2D_3D_4-9D_1D_4^2-4D_3^3),\\
\end{aligned}
\end{equation}
where all $\beta_j\in\mathbb{R}$, $j=1, 2, \dots, 13$.
The expression for $\Phi_3$ is not required in the sequel and omitted. \end{Proof}
Applying the composition of the linear and nonlinear normal form transformation, that is,
\[
\mathbf{A}=\mathbf{T}(\varepsilon)\left(\mathbf{C}+\Psi_2(\mathbf{C})+\Psi_3(\mathbf{C})\right),
\]
the system \eqref{e:2dcredA} admits the normal form
\begin{equation}
\label{e:2dnfeq}
\begin{aligned}
\dot{\mathbf{C}}=\mathscr{L}(\varepsilon)\mathbf{C}+\mathscr{R}_3(\mathbf{C})+
\mathcal{O}\Big(|\varepsilon|^2\|\mathbf{C}\|+|\varepsilon|\|\mathbf{C}\|^2+\|\mathbf{C}\|^4\Big).
\end{aligned}
\end{equation}
Truncating the normal form system at cubic terms in $\mathbf{C}$ and leading order in $\varepsilon$, yields
\begin{equation}\label{e:2dnfc}
\dot{\mathbf{C}}=\mathscr{L}(\varepsilon)\mathbf{C}+\mathscr{R}_3(\mathbf{C}).
\end{equation}
The truncated normal form gains an extra rotational symmetry $R_\theta$, given by,
\begin{equation}\label{e:symm}
R_\theta(\mathbf{C})=(\mathrm{e}^{\mathrm{i} \theta}C_1,\mathrm{e}^{\mathrm{i} \theta}C_2,\mathrm{e}^{-\mathrm{i} \theta}\bar{C}_1,\mathrm{e}^{-\mathrm{i} \theta}\bar{C}_2,D_1,D_2,D_3,D_4).
\end{equation}
\begin{Remark}
This additional symmetry $R_\theta$ results from the form of the linear term in the original 8th-order ODE \eqref{e:2dcredA} and our particular choice of the normal form transformation; see \cite[Chapter 3]{harioo} for details. Moreover, this additional symmetry $R_\theta$ fails to hold for the full normal form system \eqref{e:2dnfeq} while the reversibility $S_1$ and the symmetry $S_2$ hold.
\end{Remark}
\subsection{Construction of extended pearled solutions}\label{ss:24}
We adapt the techniques of \cite[Section 3.1, 4.1]{iooper_1993}, employing rescalings and the implicit function theorem to construct periodic solutions to the normal form system \eqref{e:2dnfeq}, which correspond to extended pearled solutions of the flat-bilayer system \eqref{e:2sFCH}.
Restricting the truncated normal form system \eqref{e:2dnfc} to the subspace
\[\widetilde{\mathbb{R}^4}:=\{(C_1, C_2, \bar{C}_1, \bar{C}_2, 0,0,0,0)\mid C_1, C_2\in\mathbb{C}\},\]
yields
the 1:1 resonant normal form; see \cite{imd_1989, iooper_1993, harioo},
\begin{equation}
\label{e:rhopf}
\begin{cases}
\dot{C_1}=\mathrm{i} (1+\omega_1\varepsilon) C_1 + C_2+\mathrm{i} C_1\big[\alpha_7C_1\bar{C}_1+\alpha_8\mathrm{i}(C_1\bar{C}_2-\bar{C}_1C_2)\big],\\
\dot{C_2}=\mathrm{i} (1+\omega_1\varepsilon) C_2 +\mathrm{i} C_2\big[\alpha_7C_1\bar{C}_1+\alpha_8\mathrm{i}(C_1\bar{C}_2-\bar{C}_1C_2)\big]
+ C_1\left[\omega_2\varepsilon+\alpha_1C_1\bar{C_1}+\mathrm{i} \alpha_2(C_1\bar{C_2}-\bar{C_1}C_2)\right].
\end{cases}
\end{equation}
\begin{Remark}
The 4th-order system in \cite{iooper_1993} admits a very general normal form in which the even-order terms automatically vanishes. We can not make this generalization here since the invariance of the pearling modes is not guaranteed when we push the normal form to high orders.
\end{Remark}
The construction of the extended pearled solutions relies crucially on two properties of the 1:1 resonant normal form. First,
the 1:1 resonant normal form \eqref{e:rhopf} admits two \textit{first integrals},
\[
K=\frac{\mathrm{i}}{2}(C_1\bar{C}_2-\bar{C}_1C_2), \quad H=|C_2|^2-\left[\frac{\alpha_1}{2}|C_1|^2-\left(\omega_2\varepsilon+2\alpha_2K\right)\right]|C_1|^2,
\]
and as a consequence may be reduced to a 2nd-order ODE in the variable $u_1:=|C_1|^2$. The pearled morphologies we seek correspond to periodic
solutions of \eqref{e:rhopf}, which are temporal equilibrium of the 2nd-order ODE for $u_1$. As a second point, the 1:1 resonant normal form is autonomous in the pearling
modes $(C_1, C_2, \bar{C}_1,\bar{C}_2)$ and thus the subspace $\widetilde{\mathbb{R}^4}$ is invariant under the truncated normal form flow \eqref{e:2dnfc}.
In this sense, the pearling modes
$(C_1, C_2, \bar{C}_1, \bar{C}_2)$ and the meandering modes $(D_1, D_2, D_3, D_4)$ exhibit a weak coupling. Accordingly, we
anticipate that structures in the 1:1 resonant normal form
will persist in the full normal form system.
A complication in the persistence argument arises through the degeneracy of the particular 1:1 resonant normal form studied here.
The two parameters, $\alpha_1$ and $\omega_2$, characterize the 1:1 resonant normal form, where $\alpha_1$ is the coefficient of $C_1^2\bar{C}_1$ in
the second entry of the cubic normal form. As shown in Lemma \ref{l:32}, for the pearling problem we have
\[
\alpha_1=0,
\]
which leads to a \textit{degenerate 1:1 resonance}. For uniformity of notation and the sign consistency with the linear stability condition in
\cite{doelmanpromislow_2013}, we also introduce
\begin{equation}\label{e:alpha_0}
\alpha_0:=-\omega_2=-\mu_2=\frac{1}{4\lambda_0^2}\int_\mathbb{R}\left(W^{\prime\prime\prime}(u_0)v_0-\eta_dW^{\prime\prime}(u_0)\right)\psi_0^2\mathrm{d} r.
\end{equation}
With these modifications, we rename the degenerate system \emph{the pearling normal form} (PNF) system,
\begin{equation}\label{e:PNF}
\hskip -0.2in
\begin{cases}
\dot{C_1}&=\mathrm{i} (1+\omega_1\varepsilon) C_1 + C_2+\mathrm{i} C_1\big[\alpha_7C_1\bar{C}_1+\alpha_8\mathrm{i}(C_1\bar{C}_2-\bar{C}_1C_2)\big],\\
\dot{C_2}&=\mathrm{i} (1+\omega_1\varepsilon) C_2 +\mathrm{i} C_2\big[\alpha_7C_1\bar{C}_1+\alpha_8\mathrm{i}(C_1\bar{C}_2-\bar{C}_1C_2)\big]
+ C_1\left[-\alpha_0\varepsilon+\mathrm{i} \alpha_2(C_1\bar{C_2}-\bar{C_1}C_2)\right],
\end{cases}
\end{equation}
For this degenerate case the persistence issue is a singular perturbation problem; removing the singularity
requires two novel proper rescalings. After the first scaling, we construct a Poincare map, which is well-defined for sufficiently
small system parameters, including the zeroes. However the base of the transverse hyper-plane in the Poincare map consists of eigenvectors. As the
system parameters approach zero, the degeneracy of eigenvalues results in the coalescence of the eigenvectors, which we overcome via
a second rescaling. The persistence follows from an implicit-function-theorem argument.
The existence results for periodic solutions of the PNF system are summarized in the following lemma, where
for convenience, we assume $\varepsilon>0$ and introduce the rescaled first integral $\kappa:=\varepsilon^{-3/2}K$.
\begin{Lemma}[degenerate 1:1 resonance]\label{l:29}
For fixed $\eta_1$, $\eta_2$, $\gamma\in\mathbb{R}$ and a non-degenerate double-well potential $W$, there exist $\varepsilon_0$, $\kappa_0>0$ such that, for every $\varepsilon\in(0,\varepsilon_0]$, the PNF system \eqref{e:PNF}
admits a \textit{degenerate 1:1 resonance}, characterized by $\alpha_0$, defined in \eqref{e:alpha_0}. More specifically, we have
\begin{enumerate}
\item[(\rmnum{1})] For $\alpha_0<0$ , the PNF system \eqref{e:PNF} has no periodic solutions except for the trivial equilibrium.
\item[(\rmnum{2})] For $\alpha_0>0$, the PNF system \eqref{e:PNF} possesses a family of periodic orbits $(C_1^p, C_2^p,\bar{C}_1^p,\bar{C}_2^p)$,
parameterized by $\kappa\in[-\kappa_0,\kappa_0]$. In fact, the family of periodic orbits is smooth in terms of small $\sqrt{\varepsilon}$ and $\sqrt{|\kappa|}$ except for $\kappa=0$, admitting the form
\begin{equation}\label{e:peri}
\begin{aligned}
C_1^p(t,\theta;\sqrt{\varepsilon}, \sqrt{|\kappa|})=&\sqrt{\varepsilon|\kappa|} r_1\mathrm{e}^{\mathrm{i}(\omega t+\theta)},\\
C_2^p(t,\theta; \sqrt{\varepsilon},\sqrt{|\kappa|})=&\mathop{\mathrm{sgn}}(\kappa)\mathrm{i} \varepsilon\sqrt{|\kappa|} r_2\mathrm{e}^{\mathrm{i}(\omega t+\theta)},
\end{aligned}
\end{equation}
where
\begin{equation}\label{e:r1}
r_1(\sqrt{\varepsilon},\sqrt{|\kappa|})=
(\alpha_0-2\alpha_2\sqrt{\varepsilon}\kappa)^{-1/4},
\end{equation}
and
\begin{equation}
r_2=\frac{1}{r_1},\quad \omega=1+\omega_1\varepsilon+\mathop{\mathrm{sgn}}(\kappa)\sqrt{\varepsilon} r_2^2+\alpha_7\varepsilon|\kappa| r_1^2+2\alpha_8\varepsilon^{3/2}\kappa,\quad \theta\in\mathbb{R}/[0,2\pi].
\end{equation}
\end{enumerate}
\end{Lemma}
\begin{Proof}
Under the polar coordinate change
\[
C_1=\widetilde{r}_1\mathrm{e}^{\mathrm{i}(1+\omega_1\varepsilon+\theta_1)}, \quad C_2=\widetilde{r}_2\mathrm{e}^{\mathrm{i}(1+\omega_1\varepsilon+\theta_2)}, \quad u_1=\widetilde{r}_1^2, \quad u_2=\widetilde{r}_2^2,
\]
the PNF system \eqref{e:PNF} becomes
\begin{equation}\label{e:rPNF}
\begin{cases}
&\left(\frac{\mathrm{d} u_1}{\mathrm{d} t}\right)^2=4f(u_1),\\
&\frac{\mathrm{d} (\theta_2-\theta_1)}{\mathrm{d} t}=-K(u_1u_2)^{-1}f^\prime(u_1),\\
&\frac{\mathrm{d} \theta_1}{\mathrm{d} t}=Ku_1^{-1}+\alpha_7u_1+2\alpha_8K,\\
&\frac{\mathrm{d} u_2}{\mathrm{d} t}=\left(\alpha_7u_1+2\alpha_8K\right)\frac{\mathrm{d} u_1}{\mathrm{d} t},
\end{cases}
\end{equation}
where $f(u_1)=(-\alpha_0\varepsilon+2\alpha_2K)u_1^2+Hu_1-K^2$. We observe that a double root of
\[
f(u_1)=0,
\]
corresponds to an equilibrium of the ODE
\begin{equation}\label{e:u1}
\left(\frac{\mathrm{d} u_1}{\mathrm{d} t}\right)^2=4f(u_1),
\end{equation}
which corresponds to a periodic solution in the PNF system \eqref{e:PNF}.
We apply the rescaling
\[
u_1=\varepsilon v_1, \quad K=\varepsilon^{3/2}\kappa, \quad H=\varepsilon^2 h,
\]
to $f(u_1)=0$ and have
\[
(-\alpha_0+2\alpha_2\kappa\sqrt{\varepsilon})v_1^2+hv_1-\kappa^2=0
\]
which admits a double root if and only if
\begin{equation}\label{e:G}
\left(2(-\alpha_0+2\alpha_2\kappa\sqrt{\varepsilon})v_1+h,(-\alpha_0+2\alpha_2\kappa\sqrt{\varepsilon})v_1^2+\kappa^2\right)^T=0.
\end{equation}
If $\alpha_0<0$, \eqref{e:G} admits only the trivial solution for small $\varepsilon$ and $k$. If $\alpha_0>0$, then we can solve $v_1$ and $h$ in terms of $\varepsilon$
and $\kappa$. In fact, we have, for sufficiently small $\varepsilon$ and $k$,
\begin{equation}
\begin{cases}
v_1(\varepsilon, \kappa)=\frac{|\kappa|}{\sqrt{\alpha_0-2\alpha_2\sqrt{\varepsilon}\kappa}},\\
h(\varepsilon, \kappa)=2(\alpha_0-2\alpha_2\sqrt{\varepsilon}\kappa)v_1.
\end{cases}
\end{equation}
We conclude our proof by letting $r_1=\sqrt{v_1/|\kappa|}$.
\end{Proof}
In the sequel we assume
$\alpha_0>0$ and $\kappa\geq 0$. The analysis of the case $\kappa<0$ differs only by a sign change. To demonstrate the persistence of the periodic solutions of the PNF system in the full normal form system \eqref{e:2dnfeq}, it is necessary to remove the singular nature of the bifurcation.
To this end we apply the rescaling
\begin{equation}
\label{e:res}
\mathbf{C}=\sqrt{\varepsilon\kappa}\widetilde{\mathbf{C}},
\end{equation}
to the normal form system \eqref{e:2dnfeq}, obtaining a new ODE system
\begin{equation}
\label{e:ODE}
\begin{aligned}
\dot{\mathbf{C}}=\mathscr{L}(\varepsilon)\mathbf{C}+\varepsilon\kappa\mathscr{R}_3(\mathbf{C})
+\varepsilon\mathcal{O}\big(\varepsilon\|\mathbf{C}\|+\sqrt{\varepsilon\kappa}\|\mathbf{C}\|^2+\sqrt{\varepsilon\kappa^3}\|\mathbf{C}\|^4\big),
\end{aligned}
\end{equation}
where we have dropped the ``tilde'' notation on ${\bf C}$.
To simplify the proof of the persistence we introduce the new small parameter $\zeta$ for which $\zeta=0$ corresponds to the cubic truncation, while $\zeta=\varepsilon$ corresponds to the full normal form. Specifically, we study the system
\begin{equation}
\label{e:gODE}
\begin{aligned}
\dot{\mathbf{C}}=\mathbf{F}(\mathbf{C})
+\zeta\mathcal{O}\big(\varepsilon\|\mathbf{C}\|+\sqrt{\varepsilon\kappa}\|\mathbf{C}\|^2+\sqrt{\varepsilon\kappa^3}\|\mathbf{C}\|^4\big),
\end{aligned}
\end{equation}
where $\mathbf{F}(\mathbf{C}):=
\mathscr{L}(\varepsilon)\mathbf{C}+\varepsilon\kappa\mathscr{R}_3(\mathbf{C})$.
The following Proposition, taken from \cite{iooper_1993}, greatly simplifies the construction.
\begin{Proposition}\label{p:tpt}
An orbit of an autonomous reversible system is periodic and reversible
if and only if there exist two different fixed points on this orbit
with respect to the reversibility.
\end{Proposition}
We lift the scalars $r_1$ and $r_2$, introduced in Lemma \ref{l:29}, which serve as the base point for the periodic solutions of the PNF system, to a vector
in the $8$ dimensional space, defining \textit{the base point} $\mathbf{r}$ as
\[
\mathbf{r}(\sqrt{\varepsilon},\sqrt{\kappa})=(r_1,\mathrm{i}\sqrt{\varepsilon} r_2,r_1,-\mathrm{i}\sqrt{\varepsilon}r_2,0,0,0,0)^T,
\]
and obeserve that the periodic solution $R_{\omega t}\mathbf{r}$ to the system \eqref{e:gODE} when $\zeta=0$
has two fixed points under reversibility, that is,
\[
S_1\mathbf{r}=\mathbf{r},\quad S_1R_{\pi}\mathbf{r}=R_{\pi}\mathbf{r}.
\]
Here we recall that
\begin{equation*}
\begin{aligned}
S_1(C_1,C_2,\bar{C}_1,\bar{C}_2, D_1,D_2,D_3, D_4)^T&=(\bar{C}_1,-\bar{C}_2, C_1, -C_2, D_1, -D_2, D_3, -D_4)^T,\\
R_{\theta}(C_1,C_2,\bar{C}_1,\bar{C}_2, D_1,D_2,D_3, D_4)^T&=(\mathrm{e}^{\mathrm{i}\theta}C_1, \mathrm{e}^{\mathrm{i}\theta}C_2, \mathrm{e}^{-\mathrm{i}\theta}\bar{C}_1, \mathrm{e}^{-\mathrm{i}\theta}\bar{C}_2, D_1,D_2,D_3, D_4)^T.
\end{aligned}
\end{equation*}
We assign two transversal hyper-planes, $H_1$ and $H_2$, respectively to $\mathbf{r}$ and $R_{\pi}\mathbf{r}$, given as follows.
\begin{equation*}
H_1=\{\mathbf{C}\in\widetilde{\mathbb{R}^4}\times\mathbb{R}^4\mid S_1\mathbf{C}={\bf C}\},\quad
H_2=\{\mathbf{C}\in\widetilde{\mathbb{R}^4}\times\mathbb{R}^4\mid (\mathbf{C}-R_\pi\mathbf{r})\cdot R_\pi\mathbf{G}\mathbf{r}=0\},
\end{equation*}
where $\mathbf{G}$ is the infinitesimal generator of the group $R_\theta$ and ``$\cdot$" represents the Euclidean inner product.
It is then not hard to see that, for the rescaled system \eqref{e:gODE}, there exists a smooth Poincar\'{e} map, denoted as $\Pi$, from an open neighborhood of the base point $\mathbf{r}$ in $H_1$,
$N(\mathbf{r},H_1)$, into one of $R_\pi\mathbf{r}$ in $H_2$, $N(R_\pi\mathbf{r}, H_2)$. More specifically, we have
\begin{equation}
\Pi(\mathbf{C},\sqrt{\varepsilon}, \sqrt{\kappa},\zeta): N(\mathbf{r},H_1)\times [0,\sqrt{\varepsilon_0}]\times[0,\sqrt{\kappa_0}]\times [-\zeta_0,\zeta_0]\rightarrow N(R_\pi\mathbf{r}, H_2).
\end{equation}
Meanwhile, there is also a smooth ``arrival time" map
\begin{equation}
T(\mathbf{C},\sqrt{\varepsilon}, \sqrt{\kappa},\zeta): N(\mathbf{r},H_1)\times [0,\sqrt{\varepsilon_0}]\times[0,\sqrt{\kappa_0}]\times [-\zeta_0,\zeta_0]\rightarrow \mathbb{R}.
\end{equation}
According to Proposition \ref{p:tpt}, any point in $H_1\cap\mathop{\mathrm{Rg}}(\Pi)$ corresponds to a periodic orbit of the system \eqref{e:gODE}, vice versa.
To further analyze the Poincar\'{e} map, we first linearize the system \eqref{e:gODE} around the periodic orbit $R_{\omega t}\mathbf{r}$.
We introduce the change of variables local to $R_{\omega t}\mathbf{r}$,
\[
\mathbf{C}=R_{\omega t}(\mathbf{r}+\mathbf{q}),
\]
and study the flow of $\mathbf{q}$ instead, that is,
\begin{equation}\label{e:gODE1}
\frac{\mathrm{d}\mathbf{q}}{\mathrm{d}t}=\mathbf{F}(\mathbf{r}+\mathbf{q})-\mathbf{F}(\mathbf{r})-\omega \mathbf{G}\mathbf{q}
+\mathcal{O}\big(\sqrt{\varepsilon}(\sqrt{\varepsilon}+\sqrt{\kappa})|\zeta|\big).
\end{equation}
Linearizing the system \eqref{e:gODE1} at $\mathbf{q}=0$ yields the following system
\begin{equation}\label{e:gODEn}
\dot{\mathbf{q}}=\mathbf{H}\mathbf{q}
+\mathcal{O}\big(\varepsilon\kappa\|\mathbf{q}\|^2
+\sqrt{\varepsilon}(\sqrt{\varepsilon}+\sqrt{\kappa})|\zeta|\big),
\end{equation}
where $\mathbf{H}:=\nabla_{\mathbf{C}}\mathbf{F}(\mathbf{r})-\omega \mathbf{G}$.
\begin{Remark}
The reversibility holds within the truncated system
$ \dot{\mathbf{q}}=\mathbf{F}(\mathbf{r}+\mathbf{q})-\mathbf{F}(\mathbf{r})-\omega \mathbf{G}\mathbf{q}$,
but not within the full ODE system about $\mathbf{q}$, since the rotational symmetry $R_{\omega t}$ and the reversibility $S_1$ do not commute.
As a result, we have
\[
S_1{\bf H}=-{\bf H}S_1.
\]
\end{Remark}
The next step is to obtain the eigenvalues and corresponding eigenmodes of $\mathbf{H}$.
We note that $\mathbf{H}$ is block diagonal.
The upper diagonal block ${\bf H}_1$ of $\mathbf{H}$ is of the form
\begin{equation*}
\begin{aligned}
{\bf H}_1 =& \begin{pmatrix}
0 & 1 & 0 & 0 \\
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 \\
0 & 0 & 0 & 0
\end{pmatrix}+\sqrt{\varepsilon}r_2^2
\begin{pmatrix}
-\mathrm{i} & 0 & 0 & 0 \\
0 & -\mathrm{i} & 0 & 0 \\
0 & 0 & \mathrm{i} & 0 \\
0 & 0 & 0 & \mathrm{i}
\end{pmatrix}-\varepsilon
\begin{pmatrix}
0 & 0 & 0 & 0 \\
\alpha_0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 \\
0 & 0 & \alpha_0 & 0
\end{pmatrix}-\varepsilon^2\kappa r_2^2
\begin{pmatrix}
0 & 0 & 0 & 0 \\
\alpha_8 & 0 & \alpha_8 & 0 \\
0 & 0 & 0 & 0 \\
\alpha_8 & 0 & \alpha_8 & 0
\end{pmatrix}+
\\
&\varepsilon \kappa r_1^2\begin{pmatrix}
\mathrm{i}\alpha_7 & \alpha_8 & \mathrm{i}\alpha_7 & -\alpha_8 \\
0 & -\mathrm{i}\alpha_2 & 0 & \mathrm{i}\alpha_2 \\
-\mathrm{i}\alpha_7 & -\alpha_8 & -\mathrm{i}\alpha_7 & \alpha_8 \\
0 & -\mathrm{i}\alpha_2 & 0 & \mathrm{i}\alpha_2
\end{pmatrix}+\varepsilon^{3/2}\kappa
\begin{pmatrix}
\mathrm{i}\alpha_8 & 0 & \mathrm{i}\alpha_8 &0 \\
3\alpha_2-\alpha_7 &\mathrm{i}\alpha_8 & \alpha_2-\alpha_7 & -\mathrm{i} \alpha_8\\
-\mathrm{i}\alpha_8 & 0 & -\mathrm{i}\alpha_8 & 0 \\
\alpha_2-\alpha_7 & \mathrm{i} \alpha_8 & \alpha_2-\alpha_7 &-\mathrm{i}\alpha_8
\end{pmatrix}\\
=& \begin{pmatrix}
0 & 1 & 0 & 0 \\
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 \\
0 & 0 & 0 & 0
\end{pmatrix}+\sqrt{\alpha_0\varepsilon}
\begin{pmatrix}
-\mathrm{i} & 0 & 0 & 0 \\
0 & -\mathrm{i} & 0 & 0 \\
0 & 0 & \mathrm{i} & 0 \\
0 & 0 & 0 & \mathrm{i}
\end{pmatrix}+\mathcal{O}(\varepsilon).
\end{aligned}
\end{equation*}
It is straightforward to see that
\[
\mathbf{H}\mathbf{G}\mathbf{r}=0,\quad \mathbf{H}\frac{\partial \mathbf{r}}{\partial \sqrt{\kappa}}=\frac{\partial \omega}{\partial \sqrt{\kappa}}\mathbf{G}\mathbf{r},
\]
where $\mathbf{r}:=\mathbf{r}(\sqrt{\varepsilon},\sqrt{\kappa})$ and $\omega:=\omega(\sqrt{\varepsilon},\sqrt{\kappa})$. As a result, $0$ is an eigenvalue to the upper diagonal block ${\bf H}_1$ with algebraic multiplicity $2$. A direct calculation then shows that
the determinant of ${\bf H}_1$ is
\[
\det(\lambda-{\bf H}_1)=\lambda^4+4\varepsilon r_2^4\lambda^2,
\]
which indicates that the other two eigenvalues of ${\bf H}_1$ are
\[
\pm\lambda_{1}=\pm2\mathrm{i}\sqrt{\varepsilon}r_2^2
=\pm2\mathrm{i}\sqrt{\alpha_0\varepsilon-2\alpha_2\varepsilon^{3/2}\kappa}=\pm2\mathrm{i}\sqrt{\alpha_0\varepsilon}+\mathcal{O}(\varepsilon\kappa),
\]
with associated eigenvectors ${\bf r}^{\pm}_1$ satisfying
\[
\mathbf{H}{\bf r}_1^+=\lambda_1{\bf r}_1^+,\quad S_1\mathbf{r}_1^+=\mathbf{r}_1^-.
\]
More specifically, a nonzero vector of cofactors of any row of ${\bf H}_1-\lambda_1$ is an eigenvector with respect to $\lambda_1$ since the algebraic multiplicity of $\lambda_1$ is 1. We then let $\mathbf{r}_1^+=({\bf r}_{1,1}^+,0,0,0,0)^T$, where ${\bf r}_{1,1}^+$ is the vector of cofactors of the second row of ${\bf H}_1-\lambda_1$ after an $\varepsilon^{3/2}\kappa$-rescaling, that is,
\[
\mathbf{r}_{1,1}^+=\begin{pmatrix} \alpha_7\sqrt{\varepsilon}\kappa r_1^4\\ \mathrm{i}\alpha_7\varepsilon\kappa r_1^2 \\
2-\alpha_7\sqrt{\varepsilon}\kappa r_1^4 \\2\mathrm{i}\sqrt{\varepsilon} r_2^{2}+\mathrm{i}\alpha_7\varepsilon\kappa r_1^2\end{pmatrix}=\begin{pmatrix} 0\\ 0 \\ 2 \\0\end{pmatrix}+\mathcal{O}(\sqrt{\varepsilon}).
\]
The lower block ${\bf H}_2$ of ${\bf H}$ is of the form
\begin{equation*}
{\bf H}_2=\begin{pmatrix}
0 & 1 & 0 & 0 \\
0 & 0 & 1 & 0 \\
0 & 0 & 0 & 1 \\
c & 0 & b & 0
\end{pmatrix},
\end{equation*}
where
\[
b=\varepsilon \left[\omega_4+\beta_5 \kappa r_1^2-2\beta_7\varepsilon\kappa r_2^2+2\beta_4\sqrt{\varepsilon}\kappa\right] , \quad
c=\varepsilon\left(\beta_0+\beta_1\kappa r_1^2+\beta_2\varepsilon\kappa r_2^2+ 2\beta_3\sqrt{\varepsilon}\kappa\right).
\]
Here we use the fact that $\beta_0=\omega_3$.
Noting that the characteristic polynomial of ${\bf H}_2$ is
\[
\lambda^4-b\lambda^2-c=0,
\]
we conclude that ${\bf H}_2$ has nonzero eigenvalues if and only if $c\neq0$, which can be guaranteed
by further assuming that $\beta_0\neq 0$ for $\varepsilon$ and $\kappa$ sufficiently small.
We summarize our assumptions on system parameters in the following hypothesis.
\begin{Hypothesis}[generic and non-degeneracy condition]\label{h:gen}
We assume that
\begin{equation}\label{e:gen}
\alpha_0 >0,\quad \beta_0\neq 0, \quad \varepsilon>0, \quad \kappa>0.
\end{equation}
\end{Hypothesis}
Under this non-degeneracy assumption, we have, for sufficiently small $\kappa$ and $\varepsilon$,
\[
b^2+4c\neq 0,
\]
which implies that ${\bf H}_2$ admits four distinct nonzero eigenvalues $\pm\lambda_2$ and $\pm\lambda_3$ of order $\sqrt[4]{\varepsilon}$ and with associated eigenvectors $\mathbf{r}_2^\pm$ and $\mathbf{r}_3^\pm$ satisfying
\[
\mathbf{H}\mathbf{r}_2^+=\lambda_2\mathbf{r}_2^+,\quad \mathbf{H}\mathbf{r}_3^+=\lambda_3\mathbf{r}_3^+,\quad
S_1\mathbf{r}_2^+=\mathbf{r}_2^-,\quad S_1\mathbf{r}_3^+=\mathbf{r}_3^-.
\]
More specifically, we choose
\begin{equation*}
\begin{aligned}
& \lambda_2=\left(\frac{b+\sqrt{b^2+4c}}{2}\right)^{1/2}=\sqrt[4]{\beta_0\varepsilon}+\mathcal{O}(\varepsilon^{3/4}+\sqrt[4]{\varepsilon}\kappa), \qquad \mathbf{r}_2^+=(0,0,0,0,1, \lambda_2,\lambda_2^2,\lambda_2^3)^T,\\
& \lambda_3=\left(\frac{b-\sqrt{b^2+4c}}{2}\right)^{1/2}=\mathrm{i}\sqrt[4]{\beta_0\varepsilon}+\mathcal{O}(\varepsilon^{3/4}+\sqrt[4]{\varepsilon}\kappa), \qquad \mathbf{r}_3^+=(0,0,0,0,1, \lambda_3,\lambda_3^2,\lambda_3^3)^T.
\end{aligned}
\end{equation*}
Based on the spectral information about $\mathbf{H}$ we collected, we denote
\begin{equation*}
\begin{aligned}
\mathbf{r}_0(\sqrt{\varepsilon},\sqrt{\kappa})&=\sqrt[4]{\alpha_0}\frac{\partial \mathbf{r}}{\partial \sqrt{\kappa}}=(1,0,1,0,0,0,0,0)^T+\mathcal{O}(\sqrt{\varepsilon}),\\
\mathbf{r}_j(\sqrt{\varepsilon},\sqrt{\kappa})&=\mathbf{r}_j^++\mathbf{r}_j^-, \quad j=1,2,3,\\
\widetilde{\mathbf{r}}_1(\sqrt{\varepsilon},\sqrt{\kappa})&=\mathbf{r}_1^+-\mathbf{r}_1^-,\\
\widetilde{\mathbf{r}}_j(\sqrt[4]{\varepsilon},\sqrt{\kappa})&=\mathbf{r}_j^+-\mathbf{r}_j^-, \quad j=2,3.\\
\end{aligned}
\end{equation*}
We note that every ${\bf r}_j$, $j=0,1,2,3$ is a smooth with respect to its arguments. In particular, even though $\lambda_2$ and $\lambda_3$ are of order $\sqrt[4]{\varepsilon}$,
\begin{equation*}
\begin{aligned}
{\bf r}_j={\bf r}_j^++{\bf r}_j^-=2(0,0,0,0,1,0,\lambda_j^2,0)^T=2(0,0,0,0,1,0,0,0)^T+\mathcal{O}(\sqrt{\varepsilon}), \quad j=2,3,\\
\end{aligned}
\end{equation*}
is smooth in terms of $\sqrt{\varepsilon}$.
We characterize the two transversal hyperplanes, $H_1$ and $H_2$, by the eigenvectors, that is,
$H_1=\mathbf{r}+\widetilde{H}_1$ and $H_2=R_\pi \mathbf{r}+\widetilde{H}_2$, where
\begin{equation*}
\widetilde{H}_1=\mathop{\mathrm{span}}\{\mathbf{r}_0,\mathbf{r}_j\mid j=1,2,3\}, \quad
\widetilde{H}_2=\mathop{\mathrm{span}}\{R_\pi \mathbf{r}_0,R_\pi\mathbf{r}_j^\pm\mid j=1,2,3\}.
\end{equation*}
We also parameterize $\mathbf{q}_1\in \widetilde{H}_1$ and $\mathbf{q}_2\in \widetilde{H}_2$ by
\begin{equation}\label{e:loor}
\begin{aligned}
\mathbf{q}_1&=\sum_{j=0}^3q_{1,j}\mathbf{r}_j,\\
\mathbf{q}_2&=\sum_{j=0}^3q_{2,j}R_\pi\mathbf{r}_j+\sum_{j=1}^3\widetilde{q}_{2,j}R_\pi\widetilde{\mathbf{r}}_j
\end{aligned}
\end{equation}
where we denote $q_1=(q_{1,0},q_{1,1},q_{1,2},q_{1,3})$ and $q_2=(q_{2,0},q_{2,1},q_{2,2},q_{2,3},\widetilde{q}_{2,1},\widetilde{q}_{2,2},\widetilde{q}_{2,3})$.
\begin{Remark}\label{r:coal}
The parameterization \eqref{e:loor} is singular at $\varepsilon=0$, since the eigenvalues coalesce.
More specifically, when $\varepsilon=0$, multiple eigenvectors collapse into one, that is,
\[2{\bf r}_0(0,\sqrt{\kappa})=\mathbf{r}_1(0,\sqrt{\kappa}), \quad \mathbf{r}_2^+(0,\sqrt{\kappa})=\mathbf{r}_2^-(0,\sqrt{\kappa})=\mathbf{r}_3^+(0,\sqrt{\kappa})=\mathbf{r}_3^-(0,\sqrt{\kappa}).\]
\end{Remark}
Therefore, with this singular parameterization \eqref{e:loor}, we rewrite the Poincar\'{e} map and the arrival time map as follows.
\begin{equation*}
\begin{aligned}
&\begin{matrix}
\widetilde{\Pi}: &[-q_0, q_0]^4\times [0,\sqrt{\varepsilon}_0]\times[0,\sqrt{\kappa}_0]\times [-\zeta_0,\zeta_0]&\longrightarrow &N(R_\pi\mathbf{r}, H_2),\\
&(q_1,\sqrt{\varepsilon},\sqrt{\kappa},\zeta)
&\longmapsto&\Pi(\mathbf{r}+\sum_{j=0}^3q_{1,j}\mathbf{r}_j,\sqrt{\varepsilon}, \sqrt{\kappa},\zeta),\\
\end{matrix}\\
&\begin{matrix}
\widetilde{T}: &[-q_0, q_0]^4\times [0,\sqrt{\varepsilon}_0]\times[0,\sqrt{\kappa}_0]\times [-\zeta_0,\zeta_0]&\longrightarrow &N(R_\pi\mathbf{r}, H_2),\\
&(q_1,\sqrt{\varepsilon},\sqrt{\kappa},\zeta)
&\longmapsto&T(\mathbf{r}+\sum_{j=0}^3q_{1,j}\mathbf{r}_j,\sqrt{\varepsilon}, \sqrt{\kappa},\zeta).\\
\end{matrix}\\
\end{aligned}
\end{equation*}
Note that $\widetilde{\Pi}$ and $\widetilde{T}$ are smooth in terms of their arguments in the domain due to the fact that every ${\bf r}_j$ is smooth in terms of $\sqrt{\varepsilon}$ and $\sqrt{\kappa}$.
Moreover, according to the coalescence of eigenvectors when $\varepsilon=0$ in Remark \ref{r:coal}, it is not hard to verify that
\begin{equation}\label{e:Tep}
\widetilde{T}(q_1,\sqrt{\varepsilon},\sqrt{\kappa},\zeta)=\pi+\mathcal{O}\left(\sqrt{\varepsilon}(1+\sqrt{\kappa}+|\zeta|+\|q_1\|)\right)=\frac{\pi}{\omega}+\mathcal{O}\left(\sqrt{\varepsilon}(|\zeta|+\|q_1\|)\right).
\end{equation}
Applying the variation of constant formula to \eqref{e:gODEn} and the parameterization of ${\bf q}_1$ in \eqref{e:loor}, together with the equality \eqref{e:Tep}, we have that
\begin{equation}\label{e:poex}
\begin{aligned}
\widetilde{ \Pi}(q_1,\sqrt{\varepsilon},\sqrt{\kappa},\zeta)=&R_{\omega \widetilde{T}}\left(\mathbf{r}+
\mathrm{e}^{\mathbf{H} \widetilde{T}}\mathbf{q}_1\right)+\mathcal{O}(\varepsilon\kappa\|q_1\|^2+\sqrt{\varepsilon}(\sqrt{\varepsilon}+\sqrt{\kappa})|\zeta|)\\
=&R_{\pi}\mathbf{r}+R_{\pi}\exp(\mathbf{H} \frac{\pi}{\omega})\mathbf{q}_1+
\omega\left(\sum_{j=0}^3\widetilde{T}_jq_{1,j}\right)R_{\pi}\mathbf{G}\mathbf{r}+
\mathcal{O}\left(\sqrt{\varepsilon}(|\zeta|+\|q_1\|^2)\right)\\
=&R_{\pi}\mathbf{r}+q_{1,0}R_{\pi}\mathbf{r}_0+\sum_{j=1}^3q_{1,j}\left[
\cosh(\lambda_j\frac{\pi}{\omega})R_\pi\mathbf{r}_j+\sinh(\lambda_j\frac{\pi}{\omega})
R_\pi\widetilde{\mathbf{r}}_j\right]+\\
&\omega\left[\left(\widetilde{T}_0 +
\alpha_0^{1/4}\frac{\pi}{\omega^2}\frac{\partial \omega}{\partial \sqrt{\kappa}}\right)q_{1,0}+
\sum_{j=1}^3\widetilde{T}_jq_{1,j}\right]R_{\pi}\mathbf{G}\mathbf{r}+\mathcal{O}\left(\sqrt{\varepsilon}(|\zeta|+\|q_1\|^2)\right),
\end{aligned}
\end{equation}
where $\cosh$ and $\sinh$ take their natural analytic extension onto $\mathbb{C}$. Moreover, we have
\[
\widetilde{T}_0:=\frac{\partial \widetilde{T}}{\partial q_{1,0}}(0,\sqrt{\varepsilon},\sqrt{\kappa},0), \quad
\widetilde{T}_j:=\frac{\partial \widetilde{T}}{\partial q_{1,j}}(0,\sqrt{\varepsilon},\sqrt{\kappa},0), j=1,2,3.
\]
Noting that $R_{\pi}\mathbf{G}\mathbf{r}$ is transverse to $H_2$ and $\Pi(\mathbf{r}+\mathbf{q}_1,\zeta)\in H_2$,
we conclude that the coefficient of $R_{\pi}\mathbf{G}\mathbf{r}$ is zero, that is, in leading order,
\begin{equation}\label{e:tildeT}
\widetilde{T}_0=-\sqrt[4]{\alpha_0}\frac{\pi}{\omega^2}\frac{\partial \omega}{\partial \sqrt{\kappa}},\quad
\widetilde{T}_j=0, \quad j=1,2,3.
\end{equation}
Expressing the expansion of $\widetilde{\Pi}$ in \eqref{e:poex} in terms of ${\bf q}_2$ as in \eqref{e:loor}, we have
\begin{equation}
\begin{aligned}
q_{2,0}=&q_{1,0}+\mathcal{O}\left(\sqrt{\varepsilon}(|\zeta|+\|q_1\|^2)\right),\\
q_{2,j}=&\cosh(\lambda_j\frac{\pi}{\omega})q_{1,j}+\mathcal{O}\left(\sqrt{\varepsilon}(|\zeta|+\|q_1\|^2)\right),\quad j=1,2,3,\\
\widetilde{q}_{2,j}=&\sinh(\lambda_j\frac{\pi}{\omega})q_{1,j}+\mathcal{O}\left(\sqrt{\varepsilon}(|\zeta|+\|q_1\|^2)\right),\quad j=1,2,3.
\end{aligned}
\end{equation}
Therefore, $\mathbf{q}_2\in H_1\cap \mathop{\mathrm{Rg}}(\widetilde{\Pi})$ if and only if
\begin{equation}\label{e:imp}
\widetilde{q}_{2,j}(q_1,\sqrt[4]{\varepsilon}, \sqrt{\kappa}, \zeta)=0, \quad j=1,2,3.
\end{equation}
Moreover, noting that, under the assumption \eqref{e:gen} and the assumption that $\varepsilon$ and $\kappa$ are sufficiently small,
\begin{equation}\label{e:jacobian}
\begin{aligned}
&\varepsilon^{-1/2}\sinh(\lambda_1\frac{\pi}{\omega})
=\mathrm{i}\varepsilon^{-1/2}\sin\left(\frac{2\pi\sqrt{\varepsilon}r_2^2}{\omega}\right)
=2\sqrt{\alpha_0}\pi\mathrm{i}+\mathcal{O}(\sqrt{\varepsilon}),\\
&\varepsilon^{-1/4}\sinh(\lambda_2\frac{\pi}{\omega})
=\varepsilon^{-1/4}\lambda_2\frac{\pi}{\omega}+\mathcal{O}(\sqrt{\varepsilon})
=\sqrt[4]{\beta_0}\pi+\mathcal{O}(\sqrt{\varepsilon}+\kappa);\\
&\varepsilon^{-1/4}\sinh(\lambda_3\frac{\pi}{\omega})
=\varepsilon^{-1/4}\lambda_3\frac{\pi}{\omega}+\mathcal{O}(\sqrt{\varepsilon})
=\sqrt[4]{\beta_0}\pi\mathrm{i}+\mathcal{O}(\sqrt{\varepsilon}+\kappa),
\end{aligned}
\end{equation}
we apply the rescalings
\[
\widetilde{q}_{2,1}=\sqrt{\varepsilon}p_{2,1}, \qquad \widetilde{q}_{2,j}=\sqrt[4]{\varepsilon}p_{2,j}, \quad j=2,3,
\]
to the system \eqref{e:imp} and have
\begin{equation}\label{e:reimp}
\begin{cases}
&p_{2,1}(q_1,\sqrt[4]{\varepsilon}, \sqrt{\kappa},\zeta)
=2\sqrt{\alpha_0}\pi\mathrm{i} q_{1,1}+\mathcal{O}(\sqrt{\varepsilon}\|q_1\|+\|q_1\|^2+|\zeta|)=0,\\
&p_{2,2}(q_1,\sqrt[4]{\varepsilon}, \sqrt{\kappa},\zeta)
=\sqrt[4]{\beta_0}q_{1,2}+\mathcal{O}(\sqrt{\varepsilon}\|q_1\|+\kappa\|q_1\|+\sqrt[4]{\varepsilon}\|q_1\|^2+\sqrt[4]{\varepsilon}|\zeta|)=0,\\
& p_{2,3}(q_1,\sqrt[4]{\varepsilon}, \sqrt{\kappa},\zeta)
=\sqrt[4]{\beta_0}\mathrm{i} q_{1,3}+\mathcal{O}(\sqrt{\varepsilon}\|q_1\|+\kappa\|q_1\|+\sqrt[4]{\varepsilon}\|q_1\|^2+\sqrt[4]{\varepsilon}|\zeta|)=0.
\end{cases}
\end{equation}
Since the Jacobian of the rescaled system \eqref{e:reimp} with respect to $(q_{1,1},q_{1,2}, q_{1,3})$ at $(q_1,\sqrt[4]{\varepsilon}, \sqrt{\kappa},\zeta)=(0,0,0,0)$ is nonzero,
we may apply the implicit function theorem to the rescaled system \eqref{e:reimp}, determining that,
\begin{itemize}
\item[(\rmnum{1})]for fixed small $\varepsilon\in[0, \varepsilon_0]$, $\kappa\in[0, \kappa_0]$ and $\zeta\in[\zeta_0, \zeta_0]$,
there exists a one-parameter family of persistent reversible periodic orbits in \eqref{e:gODE}, parametrized by $q_{1,0}$. The periodic orbit is smooth with respect to $(q_{1,0}, \sqrt[4]{\varepsilon},\sqrt{\kappa}, \zeta)$. If we ignore both the cases $\varepsilon=0$ and $\kappa=0$, then the periodic orbit is smooth with respect to $(q_{1,0}, \varepsilon,\kappa,\zeta)$. In addition, we have
\begin{equation}\label{e:00}
q_{1,j}(0,\sqrt[4]{\varepsilon}, \sqrt{\kappa},0)=0, \quad j=1,2,3,
\end{equation}
due to the fact that, $p_{2,j}(0,\sqrt[4]{\varepsilon}, \sqrt{\kappa},0)=0$, for $j=1,2,3$.
\item[(\rmnum{2})]for fixed small $\varepsilon\in[0, \min\{\varepsilon_0, \zeta_0\}]$ and $\kappa\in[0, \kappa_0]$,
there exists a one-parameter family of persistent reversible periodic orbits in \eqref{e:ODE}, parametrized by $q_{1,0}$. The periodic orbit is smooth with respect to $(q_{1,0}, \sqrt[4]{\varepsilon},\sqrt{\kappa})$. If we ignore both of cases $\varepsilon=0$ and $\kappa=0$, then the periodic orbit is smooth with respect to $(q_{1,0}, \varepsilon,\kappa)$.
\end{itemize}
The fact that $\kappa$ is a free-parameter seems to contradict the uniqueness of the $q_{0,1}$-family; however, by its definition, $q_{0,1}$ is
effectively a shift of $\kappa$ and thus there is no contradiction. More specifically, for fixed $\zeta$ and $\varepsilon$, the uniqueness of the $q_{1,0}$-family in \eqref{e:gODE} implies that, for sufficiently small $\kappa>0,q_{1,0}\in\mathbb{R}$,
\begin{equation}\label{e:shift}
\mathbf{r}(\sqrt{\varepsilon},\sqrt{\kappa})+\sum_{j=1}^3q_{1,j}(0,\sqrt[4]{\varepsilon},\sqrt{\kappa},\zeta)\mathbf{r}_j(\sqrt{\varepsilon},\sqrt{\kappa})=\mathbf{r}(\sqrt{\varepsilon},0)+q_{1,0}\mathbf{r}_0(\sqrt{\varepsilon},0) +\sum_{j=1}^3q_{1,j}(q_{1,0},\sqrt[4]{\varepsilon},0,\zeta)\mathbf{r}_j(\sqrt{\varepsilon},0).
\end{equation}
Setting $\zeta=\varepsilon$ and using the left hand side of \eqref{e:shift} as the initial condition to the system \eqref{e:ODE}, the initial value problem
\begin{equation}\label{e:IVP}
\begin{cases}
& \dot{\mathbf{C}}=\mathscr{L}(\varepsilon)\mathbf{C}+\varepsilon\kappa\mathscr{R}_3(\mathbf{C})
+\varepsilon\mathcal{O}\big(\varepsilon\|\mathbf{C}\|+\sqrt{\varepsilon\kappa}\|\mathbf{C}\|^2+\sqrt{\varepsilon\kappa^3}\|\mathbf{C}\|^4\big),\\
& \mathbf{C}(0)={\bf r}(\sqrt{\varepsilon},\sqrt{\kappa})+\sum_{j=1}^3q_{1,j}(0, \sqrt[4]{\varepsilon},\sqrt{\kappa},\varepsilon){\bf r}_j(\sqrt{\varepsilon},\sqrt{\kappa}),
\end{cases}
\end{equation}
admits a periodic solution, denoted as ${\bf C}^\mathrm{rp}$, with the period
\[
T_\mathrm{rp}(\sqrt[4]{\varepsilon},\sqrt{\kappa})=2\widetilde{T}(0,q_{1,1}(0, \sqrt[4]{\varepsilon},\sqrt{\kappa},\varepsilon),q_{1,2}(0, \sqrt[4]{\varepsilon},\sqrt{\kappa},\varepsilon),q_{1,3}(0, \sqrt[4]{\varepsilon},\sqrt{\kappa},\varepsilon),\sqrt[4]{\varepsilon},\sqrt{\kappa},\varepsilon).
\]
According to \eqref{e:Tep}, \eqref{e:tildeT} and \eqref{e:00}, we have the estimate
\[
T_\mathrm{rp}(\sqrt[4]{\varepsilon},\sqrt{\kappa})=\frac{2\pi}{\omega}+\mathcal{O}\left(\sqrt{\varepsilon}(\varepsilon+\|q_1\|^2)\right)=
\frac{2\pi}{\omega}+\mathcal{O}\left(\sqrt{\varepsilon^3}\right).
\]
Using the transformation
\[
\mathbf{C}=R_{\omega t}(\mathbf{r}+\mathbf{q}),
\]
the initial value problem \eqref{e:IVP} becomes
\begin{equation}\label{e:IVP1}
\begin{cases}
& \dot{\mathbf{q}}=\mathbf{H}\mathbf{q}
+\mathcal{O}\big(\varepsilon\|\mathbf{q}\|^2
+\varepsilon^{1/2}|\zeta|\big),\\
& \mathbf{q}(0)=\sum_{j=1}^3q_{1,j}(0, \sqrt[4]{\varepsilon},\sqrt{\kappa},\varepsilon){\bf r}_j(\sqrt{\varepsilon},\sqrt{\kappa}),
\end{cases}
\end{equation}
which admits a bounded solution $\|{\bf q}(t)\|_{\infty}=\mathcal{O}(\varepsilon)$.
We summarize the results above in the following lemma.
\begin{Lemma}\label{l:pODE}
For fixed $\varepsilon\in[0,\varepsilon_0]$, up to translation,
the rescaled normal form ODE system \eqref{e:ODE},
\begin{equation*}
\begin{aligned}
\dot{\mathbf{C}}=\mathscr{L}(\varepsilon)\mathbf{C}+\varepsilon\kappa\mathscr{R}_3(\mathbf{C})
+\varepsilon\mathcal{O}\big(\varepsilon\|\mathbf{C}\|+\sqrt{\varepsilon\kappa}\|\mathbf{C}\|^2+\sqrt{\varepsilon\kappa^3}\|\mathbf{C}\|^4\big),
\end{aligned}
\end{equation*}
admits a one-parameter family of persistent reversible periodic orbits, $\mathbf{C}^\mathrm{rp}(t;\sqrt[4]{\varepsilon}, \sqrt{\kappa})$, parametrized by $\kappa\in[0,\kappa_0]$.
The periodic orbit $\mathbf{C}^\mathrm{rp}$ is smooth with respect to all parameters $(t;\sqrt[4]{\varepsilon}, \sqrt{\kappa})$. When neither $\varepsilon=0$ nor $\kappa=0$, then $\mathbf{C}^\mathrm{rp}$ is smooth with respect to $(\varepsilon,\kappa)$ and admits the form
\begin{equation}
{\bf C}^\mathrm{rp}(t;\sqrt[4]{\varepsilon},\sqrt{\kappa})=R_{\omega t}{\bf r}(\sqrt{\varepsilon},\sqrt{\kappa})+\mathcal{O}\left(\varepsilon\right),
\end{equation}
where the error is measured in the $L^\infty$ norm.
The period of $\mathbf{C}^\mathrm{rp}$, denoted by $T_\mathrm{rp}$, admits the expansion
\begin{equation}
T_\mathrm{rp}(\sqrt[4]{\varepsilon}, \sqrt{\kappa})=
\frac{2\pi}{\omega}+\mathcal{O}\left(\sqrt{\varepsilon^3}\right).
\end{equation}
\end{Lemma}
\begin{Remark}
We can also prove this lemma by using the right hand side of \eqref{e:shift} as the initial condition to \eqref{e:ODE}. But we then have to take a detour to find out the expressions of each $q_{1,j}$, $j=0,1,2,3$, in terms of $(\sqrt[4]{\varepsilon}, \sqrt{\kappa}, \zeta)$. In fact, a direct calculation using \eqref{e:shift} shows that
\begin{equation*}
\begin{aligned}
q_{1,0}(\sqrt[4]{\varepsilon},\sqrt{\kappa},\zeta)
&=\frac{\sqrt{\kappa}}{2}(r_1+\frac{r_2}{\sqrt{\alpha_0}})+q_{1,1}(0,\sqrt[4]{\varepsilon},\sqrt{\kappa},\zeta)(1-\frac{r_2^2}{\sqrt{\alpha_0}})=\frac{\sqrt{\kappa}}{\sqrt[4]{\alpha_0}}+\mathcal{O}\left(\varepsilon\kappa^{5/2}+\sqrt{\varepsilon}\kappa|\zeta|\right),\\
q_{1,1}(q_{1,0},\sqrt[4]{\varepsilon},0,\zeta)&=\frac{\sqrt{\kappa}}{4}(r_1-\frac{r_2}{\sqrt{\alpha_0}})+\frac{1}{2}q_{1,1}(0,\sqrt[4]{\varepsilon},\sqrt{\kappa},\zeta)(1+\frac{r_2^2}{\sqrt{\alpha_0}})=\frac{\alpha_2\sqrt{\varepsilon\kappa^3}}{2\sqrt[4]{\alpha_0^5}}+\mathcal{O}\left(\varepsilon\kappa^{5/2}+|\zeta|\right),\\
q_{1,2}(q_{1,0},\sqrt[4]{\varepsilon},0,\zeta)&=\frac{\sum_{j=2}^3q_{1,j}(0,\sqrt[4]{\varepsilon},\sqrt{\kappa},\zeta)(\lambda_j^2(\sqrt{\varepsilon},\sqrt{\kappa})-\lambda_3^2(\sqrt{\varepsilon},0))}{\lambda_2^2(\sqrt{\varepsilon},0)-\lambda_3^2(\sqrt{\varepsilon},0)}=\mathcal{O}\left(|\zeta|(1+\sqrt[4]{\varepsilon}+\sqrt{\kappa})\right),\\
q_{1,3}(q_{1,0},\sqrt[4]{\varepsilon},0,\zeta)&=\frac{\sum_{j=2}^3q_{1,j}(0,\sqrt[4]{\varepsilon},\sqrt{\kappa},\zeta)(\lambda_j^2(\sqrt{\varepsilon},\sqrt{\kappa})-\lambda_2^2(\sqrt{\varepsilon},0))}{\lambda_3^2(\sqrt{\varepsilon},0)-\lambda_2^2(\sqrt{\varepsilon},0)}=\mathcal{O}\left(|\zeta|(1+\sqrt[4]{\varepsilon}+\sqrt{\kappa})\right).\\
\end{aligned}
\end{equation*}
Therefore, in the system \eqref{e:ODE}, by setting $\zeta=\varepsilon$, we obtain that
\begin{equation*}
\begin{aligned}
q_{1,0}(\sqrt[4]{\varepsilon}, \sqrt{\kappa})&:=q_{1,0}(\sqrt[4]{\varepsilon}, \sqrt{\kappa},\varepsilon)=\frac{\sqrt{\kappa}}{\sqrt[4]{\alpha_0}}+\mathcal{O}\left(\varepsilon\kappa(\sqrt{\kappa^3}+\sqrt{\varepsilon})\right),\\
q_{1,1}(\sqrt[4]{\varepsilon}, \sqrt{\kappa})&:=q_{1,1}(q_{1,0}(\sqrt{\varepsilon},\sqrt{\kappa},\varepsilon),\sqrt[4]{\varepsilon}, 0,\varepsilon)=\mathcal{O}\left(\sqrt{\varepsilon}(\sqrt{\kappa^3}+\sqrt{\varepsilon})\right),\\
q_{1,j}(\sqrt[4]{\varepsilon},\sqrt{\kappa})&:=q_{1,j}(q_{1,0}(\sqrt{\varepsilon},\sqrt{\kappa},\varepsilon),\sqrt[4]{\varepsilon}, 0,\varepsilon)=\mathcal{O}(\varepsilon+\varepsilon\sqrt{\kappa}), \quad j=2,3.
\end{aligned}
\end{equation*}
\end{Remark}
Summarizing the results, we can now prove the main theorem--Theorem \ref{t:main}.
\begin{Proof}[Proof of Theorem \ref{t:main}]
The periodic solution ${\bf C}^{\mathrm{rp}}(t;\sqrt[4]{\varepsilon},\sqrt{\kappa})$ of the system \eqref{e:ODE} corresponds to a periodic solution $u_{\mathrm{rp}}(t,r; \sqrt[4]{\varepsilon},\sqrt{\kappa})$ of the PDE \eqref{e:EPP},
\begin{equation*}
\left(\partial_r^2-W^{\prime\prime}(u)+\lambda_0\partial_t^2+\varepsilon\eta_1\right)
\left(\partial_r^2u-W^\prime(u)+\lambda_0\partial_t^2u\right)+\varepsilon\eta_d W^\prime(u)-\varepsilon\gamma=0.
\end{equation*}
In fact, based on the center manifold reduction, the normal form transformation and the rescalings, especially Lemma \ref{l:pODE}, we have
\begin{equation}
\begin{aligned}
u_\mathrm{rp}(t, r)=&u_h(r)+\left[(A_1(t)+\bar{A}_1(t))+\mathrm{i}(A_2(t)-\bar{A}_2(t))\right]\psi_0(r)+\\
&B_1(t)\psi_1(r)+\mathcal{O}(\sqrt{\varepsilon}\|{\bf A}\|+\|{\bf A}\|^2)\\
=&u_h(r)+\sqrt{\varepsilon\kappa}\left[(C_1^\mathrm{rp}(t)+\bar{C}_1^\mathrm{rp}(t))+\mathrm{i}(C_2^\mathrm{rp}(t)-\bar{C}_2^\mathrm{rp}(t))\right]\psi_0(r)+\\
&\sqrt{\varepsilon\kappa}D_1^\mathrm{rp}(t)\psi_1(r)+\mathcal{O}(\varepsilon\sqrt{\kappa}\|{\bf C}^\mathrm{rp}\|)\\
=&u_h(r)+2\sqrt{\varepsilon\kappa}r_1\cos(\omega t)\psi_0(r)+\mathcal{O}\left(\varepsilon(\sqrt{\varepsilon}+\sqrt{\kappa})\right)\\
=&u_h(r)+2\frac{\sqrt{\varepsilon\kappa}}{\sqrt[4]{\alpha_0}}\cos(\omega t)\psi_0(r)+\mathcal{O}\left(\varepsilon(\sqrt{\varepsilon}+\sqrt{\kappa})\right),\\
\end{aligned}
\end{equation}
where we have the expression of $\omega$ from Lemma \ref{l:29}, that is,
\[
\omega=1+\omega_1\varepsilon+\sqrt{\varepsilon}r_2^2+\alpha_7\varepsilon\kappa r_1^2+2\alpha_8\varepsilon^{3/2}\kappa.
\]
Moreover, the period of $u_{\mathrm{rp}}$, denoted by $T_{\mathrm{rp}}$, admits the expansion
\[
T_\mathrm{rp}(\sqrt[4]{\varepsilon}, \sqrt{\kappa})=
\frac{2\pi}{\omega}+\mathcal{O}\left(\sqrt{\varepsilon^3}\right).
\]
Furthermore, since the PDE \eqref{e:EPP} is a rescaled version of the stationary FCH \eqref{e:2sFCH} with the rescaling
$t=\frac{\sqrt{\lambda_0}}{\varepsilon}\tau$, the periodic solution $u_{\mathrm{rp}}$ of the PDE \eqref{e:EPP} corresponds to
a periodic solution of the PDE \eqref{e:2sFCH}, denoted as $u_{\mathrm{p}}$ with a period $T_{\mathrm{p}}$. In fact,
\begin{equation}
\begin{aligned}
T_{\mathrm{p}}(\sqrt[4]{\varepsilon},\sqrt{\kappa})&=\frac{\varepsilon}{\sqrt{\lambda_0}}T_{\mathrm{rp}}(\sqrt[4]{\varepsilon},\sqrt{\kappa})=\frac{2\pi\varepsilon}{\sqrt{\lambda_0}}\left[1-\sqrt{\alpha_0\varepsilon}+\mathcal{O}\left(\varepsilon(1+\sqrt{\kappa})\right)\right],\\
u_\mathrm{p}(\tau, r;\sqrt[4]{\varepsilon},\sqrt{\kappa})&=u_\mathrm{rp}(\frac{\sqrt{\lambda_0}}{\varepsilon}\tau, r;\sqrt[4]{\varepsilon},\sqrt{\kappa})
=u_h(r)+2\frac{\sqrt{\varepsilon\kappa}}{\sqrt[4]{\alpha_0}}\cos(\frac{2\pi}{T_{\mathrm{p}}} \tau)\psi_0(r)+\mathcal{O}\left(\varepsilon(\sqrt{\varepsilon}+\sqrt{\kappa})\right),
\end{aligned}
\end{equation}
which concludes the proof of Theorem \ref{t:main}.
\end{Proof}
\section{Pearling of the Circular Planar Bilayer}\label{s:3}
In this section we consider the case in which the bilayer sinterface $\Gamma_{R_0}$ is a circle in $\mathbb{R}^2$,
and construct the extended pearled solutions to the extended stationary strong FCH equation \eqref{e:2sFCH2} in $(r, \theta)\in\mathbb{R}\times\mathbb{R}/2\pi\mathbb{Z}$,
\begin{equation*}
\Big(\partial_r^2-W^{\prime\prime}(u)+\frac{\varepsilon\partial_r}{R_0+\varepsilon r}+\frac{\varepsilon^2\partial_\theta^2}{(R_0+\varepsilon r)^2}+\varepsilon\eta_1\Big)
\Big(\partial_r^2u-W^\prime(u)+\frac{\varepsilon\partial_ru}{R_0+\varepsilon r}+\frac{\varepsilon^2\partial_\theta^2u}{(R_0+\varepsilon r)^2}\Big)+\varepsilon\eta_d W^\prime(u)=\varepsilon\gamma.
\end{equation*}
To exploit the analysis in the Section~\ref{s:2}, we rescale $\theta$ by $\vartheta=\frac{R_0\sqrt{\lambda_0}}{\varepsilon}\theta$ and search for extended pearled solutions
$u_\mathrm{rp}$ of
\begin{equation}
\label{e:EPP2}
\big(\partial_r^2-W^{\prime\prime}(u)+\frac{\varepsilon\partial_r}{R_0+\varepsilon r}+\frac{R_0^2\lambda_0\partial_\vartheta^2}{(R_0+\varepsilon r)^2}+\varepsilon\eta_1\big)
\big(\partial_r^2u-W^\prime(u)+\frac{\varepsilon\partial_ru}{R_0+\varepsilon r}+\frac{R_0^2\lambda_0\partial_\vartheta^2u}{(R_0+\varepsilon r)^2}\big)+\varepsilon\eta_d W^\prime(u)=\varepsilon\gamma.
\end{equation}
which satisfy the boundary conditions at infinity,
\begin{equation}
\lim_{r\rightarrow \pm\infty}|u_{\mathrm{rp}}(\vartheta,r)-u_\infty|=0, \text{ for all }\vartheta\in\mathbb{R},
\end{equation}
and an even and periodic in $\vartheta$,
\begin{equation}\label{e:pc2}
u_\mathrm{rp}(-\vartheta,r)=u_\mathrm{rp}(\vartheta,r), \quad u_\mathrm{rp}(\vartheta+T_{\mathrm{rp}},r)=u_\mathrm{rp}(\vartheta,r), \text{ for all }(\vartheta,r)\in\mathbb{R}^2,
\end{equation}
where $T_{\mathrm{rp}}$ and $u_\infty$ are constants to be determined.
We first prove the following proposition, which is similar to the Theorem \ref{t:main}.
\begin{Proposition}\label{p:pcb}
Fix $\eta_1, \eta_2\in\mathbb{R}$ and $R_0>0$. Assume that $W$ is a non-degenerate double well potential and $\alpha_0>0$, $\beta_0\neq 0$. Then there exist positive constants $\varepsilon_0>0$ and $\kappa_0>0$ such that, for any $\varepsilon\in(0,\varepsilon_0]$, up to translation, the extended stationary FCH \eqref{e:2sFCH2} in the plane $(\theta,r)\in\mathbb{R}^2$,
\begin{equation*}
\Big(\partial_r^2-W^{\prime\prime}(u)+\frac{\varepsilon\partial_r}{R_0+\varepsilon r}+\frac{\varepsilon^2\partial_\theta^2}{(R_0+\varepsilon r)^2}+\varepsilon\eta_1\Big)
\Big(\partial_r^2u-W^\prime(u)+\frac{\varepsilon\partial_ru}{R_0+\varepsilon r}+\frac{\varepsilon^2\partial_\theta^2u}{(R_0+\varepsilon r)^2}\Big)+\varepsilon\eta_d W^\prime(u)=\varepsilon\gamma.
\end{equation*}
admits a smooth one-parameter family of extended pearled solutions,
$u_\mathrm{p}(\theta,r;\sqrt[4]{\varepsilon},\sqrt{|\kappa|})$ with period $T_\mathrm{p}(\sqrt[4]{\varepsilon},\sqrt{|\kappa|})$, parameterized by $\kappa\in[-\kappa_0,\kappa_0]$. In fact, $u_{\mathrm{p}}$ and $T_{\mathrm{p}}$ are smooth with respect to their arguments within the domains expect at $\kappa=0$. The extended pearled solution $u_\mathrm{p}$ admits the asymptotic form
\begin{equation}
u_\mathrm{p}(\theta, r;\sqrt[4]{\varepsilon},\sqrt{|\kappa|})=u_h(r)+2\frac{\sqrt{\varepsilon\kappa}}{\sqrt[4]{\alpha_0}}\cos(\frac{2\pi}{T_{\mathrm{p}}} \theta)\psi_0(r)+\mathcal{O}\left(\varepsilon(\sqrt{\varepsilon}+\sqrt{\kappa})\right),
\end{equation}
where
\begin{equation}
T_{\mathrm{p}}(\sqrt[4]{\varepsilon},\sqrt{\kappa})=\frac{2\pi\varepsilon}{R_0\sqrt{\lambda_0}}\left[1-\sqrt{\alpha_0\varepsilon}+\mathcal{O}\left(\varepsilon(1+\sqrt{\kappa})\right)\right].
\end{equation}
The far-field limit of the extended pearled solution is
\begin{equation}
\lim_{r\rightarrow \infty}u_\mathrm{p}(\theta,r)=\lim_{r\rightarrow \infty}u_h(r)=u_-(\varepsilon).
\end{equation}
Moreover, for any $\varepsilon\in(0,\varepsilon_0]$, the extended stationary FCH \eqref{e:2sFCH2} in the infinite periodic strip $(\theta,r)\in\left(\mathbb{R}/2\pi\mathbb{Z}\right)\times \mathbb{R}$, admits a discrete family of extended pearled solutions,
$u_\mathrm{p}(\theta,r;\sqrt[4]{\varepsilon},\sqrt{|\kappa_j|})$with period $T_\mathrm{p}(\sqrt[4]{\varepsilon},\sqrt{|\kappa_j|})$, where
\[
\kappa_j\in\{\kappa\in[-\kappa_0,\kappa_0]\backslash\{0\}\mid
\frac{2\pi}{T_{\mathrm{p}}(\sqrt[4]{\varepsilon},\sqrt{\kappa})}\in\mathbb{Z}^+\}.
\]
\end{Proposition}
\begin{Proof}
The analysis of the circular interface system \eqref{e:EPP2} differs from that of the interface flat system \eqref{e:EPP}
in two major points:
\begin{itemize}
\item[(\rmnum{1})] The circular system \eqref{e:EPP2} has different linear terms in $\varepsilon$ than the flat system \eqref{e:EPP}.
\item[(\rmnum{2})] The $S_2$ symmetry does not hold for the extended circular bilayers as it does for the flat case.
\end{itemize}
These differences only require that we recompute the versal normal form. More specifically, we replace $u$ with $u_h+\delta u$ in \eqref{e:EPP2} and consider the equation of the perturbation $\delta u$ (again repurposing ``$u$'' to denote the perturbation).
\begin{equation}
\label{e:EPU2}
\widetilde{\mathcal{L}}u+\widetilde{\mathcal{F}}(u)=0,
\end{equation}
where
\begin{equation}
\widetilde{\mathcal{L}}:=\left(\widetilde{\mathcal{L}}_h+\frac{R_0^2\lambda_0\partial_\vartheta^2}{(R_0+\varepsilon r)^2}+\varepsilon\eta_1\right)\left(\widetilde{\mathcal{L}}_h+\frac{R_0^2\lambda_0\partial_\vartheta^2}{(R_0+\varepsilon r)^2}\right)+\widetilde{\mathcal{M}},
\end{equation}
with $\widetilde{\mathcal{L}}_h:=\partial_r^2-W^{{\prime\prime}}(u_h)+\frac{\varepsilon\partial_r}{R_0+\varepsilon r}$,
$\widetilde{\mathcal{M}}:=\varepsilon\eta_d W^{{\prime\prime}}(u_h)-\left(\partial_r^2u_h-W^\prime(u_h)+\frac{\varepsilon\partial_ru_h}{R_0+\varepsilon r}\right)W^{{\prime\prime\prime}}(u_h)$, and
\begin{equation}
\label{e:F2}
\begin{aligned}
\widetilde{ \mathcal{F}}(u,\varepsilon):=&-\frac{R_0^2\lambda_0}{(R_0+\varepsilon r)^2}W^{{\prime\prime\prime}}(u_h+u)\left(\partial_\vartheta u\right)^2-2\frac{R_0^2\lambda_0}{(R_0+\varepsilon r)^2}\left(W^{\prime\prime}(u_h+u)-W^{\prime\prime}(u_h)\right)\partial_\vartheta^2u-\\
&\left[\mathcal{L}_h+\varepsilon(\eta_1-\eta_d)-\left(W^{{\prime\prime}}(u_h+u)-W^{{\prime\prime}}(u_h)\right)\right]\left(W^\prime(u_h+u)-W^\prime(u_h)-W^{{\prime\prime}}(u_h)u\right)-\\
& \left(\partial_r^2u_h-W^\prime(u_h)+\frac{\varepsilon\partial_ru_h}{R_0+\varepsilon r}\right)\left(W^{{\prime\prime}}(u_h+u)-W^{\prime\prime}(u_h)-W^{\prime\prime\prime}(u_h)u\right)-\\
&\left(W^{{\prime\prime}}(u_h+u)-W^{{\prime\prime}}(u_h)\right)\mathcal{L}_h u.
\end{aligned}
\end{equation}
we recast the system as
\begin{equation}
\label{e:2ddFCH2}
\dot{U}=\widetilde{\mathbb{L}}(\varepsilon)U+\widetilde{\mathbb{F}}(U,\varepsilon),
\end{equation}
where
\[
\widetilde{\mathbb{L}}(\varepsilon) = \begin{pmatrix}
0&1&0&0\\ -\frac{(R_0+\varepsilon r)^2}{R_0^2\lambda_0}\widetilde{\mathcal{L}}_h&0&\frac{(R_0+\varepsilon r)^2}{R_0^2\lambda_0}&0\\ 0&0&0&1\\
-\frac{(R_0+\varepsilon r)^2}{R_0^2\lambda_0}\widetilde{\mathcal{M}}& 0&
-\frac{(R_0+\varepsilon r)^2}{R_0^2\lambda_0}(\widetilde{\mathcal{L}}_h+\varepsilon\eta_1)&0\end{pmatrix},\quad
\widetilde{\mathbb{F}}(U,\varepsilon) = \begin{pmatrix}0\\0\\0\\-\frac{(R_0+\varepsilon r)^2}{R_0^2\lambda_0}\widetilde{\mathcal{F}}\end{pmatrix}.
\]
We then have
\[
\frac{\partial \widetilde{\mathbb{L}}}{\partial \varepsilon}(0)=\frac{1}{\lambda_0}
\begin{pmatrix}
0 & 0 & 0 & 0 \\
-\frac{2r}{R_0}\mathcal{L}_0 -\frac{1}{R_0}\partial_r+W^{\prime\prime\prime}(u_0)u_1 & 0 & \frac{2r}{R_0} & 0\\
0 & 0 & 0 & 0 \\
-\eta_dW^{\prime\prime}(u_0)+W^{\prime\prime\prime}(u_0)(\mathcal{L}_0u_1+\frac{\partial_r u_0}{R_0}) & 0 & -\frac{2r}{R_0}\mathcal{L}_0 -\frac{1}{R_0}\partial_r+W^{\prime\prime\prime}(u_0)u_1-\eta_1 & 0
\end{pmatrix},
\]
which, after direct computation, leads to that the linear part in the reduced system in terms of ${\bf A}$, denoted as $ \widetilde{{\bf L}}(\varepsilon)$ just like its counterpart in \eqref{e:2dcredA}, is of a more complicated form
\begin{equation*}
\widetilde{\mathbf{L}}(\varepsilon)=\begin{pmatrix}
\mathrm{i}(1+\mu_1\varepsilon) & 1-\mu_1\varepsilon & \mathrm{i} \mu_1\varepsilon & \mu_1\varepsilon
& \mathrm{i} \widetilde{\mu}_1\varepsilon & 0 & \mathrm{i} \widetilde{\mu}_2\varepsilon &0 \\
\mu_2\varepsilon& \mathrm{i} \left(1+\mu_3\varepsilon \right) & \mu_2\varepsilon & -\mathrm{i} \mu_3\varepsilon
& \widetilde{\mu}_3\varepsilon & 0 & \widetilde{\mu}_4\varepsilon &0 \\
-\mathrm{i} \mu_1\varepsilon & \mu_1\varepsilon & -\mathrm{i}\left(1+\mu_1\varepsilon\right) & 1-\mu_1\varepsilon
& -\mathrm{i} \widetilde{\mu}_1\varepsilon & 0 & -\mathrm{i} \widetilde{\mu}_2\varepsilon &0 \\
\mu_2\varepsilon & \mathrm{i} \mu_3\varepsilon &\mu_2\varepsilon & -\mathrm{i} \left(1+\mu_3\varepsilon \right)
& \widetilde{\mu}_3\varepsilon & 0 & \widetilde{\mu}_4\varepsilon &0 \\
0 & 0 & 0 &0 & 0 & 1 & 0 & 0\\
\widetilde{\mu}_5\varepsilon & \mathrm{i} \widetilde{\mu}_6\varepsilon & \widetilde{\mu}_5\varepsilon & -\mathrm{i}\widetilde{\mu}_6\varepsilon & \mu_4\varepsilon & 0 & 1 & 0\\
0 & 0 & 0 &0 & 0 &0 & 0 & 1\\
\widetilde{\mu}_7\varepsilon & \mathrm{i} \widetilde{\mu}_8\varepsilon & \widetilde{\mu}_7\varepsilon & -\mathrm{i}\widetilde{\mu}_8\varepsilon & \mu_5\varepsilon & 0 & \mu_6\varepsilon & 0\\
\end{pmatrix},
\end{equation*}
Nevertheless, up to linear terms in $\varepsilon$, there exists a versal normal form of $\widetilde{{\bf L}}(\varepsilon)$ preserving the reversibility $S_1$, which takes the exact expression as its counterpart \eqref{e:ln2dcred} in the flat case,
that is,
\begin{equation}
\label{e:ln2dcred2}
\mathscr{L}(\varepsilon)=\begin{pmatrix}
\mathrm{i}\left(1+\omega_1\varepsilon\right) & 1 & 0 & 0 & 0 & 0 & 0 & 0 \\
\omega_2\varepsilon& \mathrm{i}\left(1+\omega_1\varepsilon\right) & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & -\mathrm{i}\left(1+\omega_1\varepsilon\right) & 1 & 0 & 0 & 0 & 0 \\
0 & 0 & \omega_2\varepsilon & -\mathrm{i}\left(1+\omega_1\varepsilon\right) & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 \\
0 & 0 & 0 & 0 & \omega_3\varepsilon & 0 & \omega_4\varepsilon & 0 \\
\end{pmatrix},
\end{equation}
where $\omega_1=\frac{1}{2}(\mu_1+\mu_3)$, $\omega_2=\mu_2$, $\omega_3=\mu_5$, $\omega_4=\mu_4+\mu_6$. The rest of the proof is the same as the flat case.
\end{Proof}
Theorem \ref{t:main2} is drived from Proposition \ref{p:pcb} by rescaling and inverting the relation between the radius and $\kappa.$
\begin{Proof}[Proof of the Theorem \ref{t:main2}]
The stationary FCH \eqref{e:2sFCH2} on the extended plane $(\theta,r)\in\mathbb{R}^2$ admits a pearled solution for any $R_0\in[R_-, \infty]$, $\varepsilon\in(0,\varepsilon_0]$ and $\kappa\in[-\kappa_0,\kappa_0]$. On the other hand, the stationary FCH \eqref{e:2sFCH2} on the infinite strip $(\theta,r)\in(\mathbb{R}/2\pi\mathbb{Z})\times\mathbb{R}$ requires that
\[
T_\mathrm{p}=\frac{2\pi}{\textrm{n}}, \text{ for some }\textrm{n}\in\mathbb{Z}^+.
\]
Therefore, we have
\[
\textrm{n}=\frac{R_0\sqrt{\lambda_0}}{\varepsilon}\left[1+\sqrt{\alpha_0\varepsilon}+\mathcal{O}\left(\varepsilon(1+\sqrt{\kappa})\right)\right],
\]
which indicates that there exists $\mathrm{n_-}>0$ so that $\textrm{n}\in[\frac{\mathrm{n_-}}{\varepsilon},\infty)$.
\end{Proof}
\section{Appendix}
We perform the calculations omitted in Section \ref{ss:22} and Section \ref{ss:23}. We begin by computing the leading order terms of the reduced
8th-order ODE system in Appendix \ref{ss:31}. The calculation of explicit expressions for $\alpha_1$ and $\alpha_2$ follows in Appendix \ref{ss:32}.
\subsection{The reduced-ODE system in terms of ${\bf A}$}\label{ss:31}
\begin{Lemma}\label{l:reducedA}
The reduced system \eqref{e:2dred},
\begin{equation*}
\dot{U}_c=\mathbb{L}_*U_c+\mathbb{P}_c\Big(\mathbb{M}(\varepsilon)\big(U_c+\Psi(U_c,\varepsilon)\big)+\mathbb{F}\big(U_c+\Psi(U_c,\varepsilon),\varepsilon,\big)\Big),
\end{equation*}
in terms of ${\bf A}:=(A_1, A_2,\bar{A}_1,\bar{A}_2, B_1, B_2, B_3, B_4)$, admits the expression
\begin{equation*}
\dot{\mathbf{A}}=\mathbf{L}(\varepsilon)\mathbf{A}+\mathbf{R}_2(\mathbf{A},\mathbf{A})
+\mathbf{R}_3(\mathbf{A},\mathbf{A},\mathbf{A})+ \mathcal{O}\left(|\varepsilon|^2\|{\bf A}\|+|\varepsilon|\|{\bf A}\|^2+\|{\bf A}\|^4\right),
\end{equation*}
where the linear term ${\bf L}$, the quadratic term ${\bf R}_2$, the cubic term ${\bf R}_3$ are of the following expressions.
\begin{equation*}
\mathbf{L}(\varepsilon)=\begin{pmatrix}
\mathrm{i}(1+\mu_1\varepsilon) & 1-\mu_1\varepsilon & \mathrm{i} \mu_1\varepsilon & \mu_1\varepsilon
& 0 & 0 & 0 &0 \\
\mu_2\varepsilon & \mathrm{i} \left(1+\mu_3\varepsilon \right) & \mu_2\varepsilon & -\mathrm{i} \mu_3 \varepsilon
& 0 & 0 & 0 &0 \\
-\mathrm{i} \mu_1\varepsilon & \mu_1 \varepsilon& -\mathrm{i}\left(1+\mu_1\varepsilon\right) & 1+\mu_1\varepsilon
& 0 & 0 & 0 &0 \\
\mu_2\varepsilon & \mathrm{i} \mu_3\varepsilon &\mu_2 \varepsilon& -\mathrm{i} \left(1+\mu_3 \varepsilon\right) & 0 & 0 & 0 &0 \\
0 & 0 & 0 &0 & 0 & 1 & 0 & 0\\
0 & 0 & 0 &0 & \mu_4 \varepsilon& 0 & 1 & 0\\
0 & 0 & 0 &0 & 0 &0 & 0 & 1\\
0 & 0 & 0 &0 & \mu_5\varepsilon & 0 & \mu_6\varepsilon & 0\\
\end{pmatrix},
\end{equation*}
\begin{equation*}
\mathbf{R}_2(\mathbf{A},\mathbf{A})=\left(0,R_{2,2},0,\bar{R}_{2,2},0,0,0,R_{2,8}\right)^T, \quad
\mathbf{R}_3(\mathbf{A},\mathbf{A},{\bf A})=\left(0,R_{3,2},0,\bar{R}_{3,2},0,0,0,R_{3,8}\right)^T.
\end{equation*}
Here we have
\begin{equation}\label{e:linearco}
\begin{aligned}
&\mu_1=-\frac{1}{2\lambda_0}\int_\mathbb{R} W^{\prime\prime\prime}(u_0)u_1\psi_0^2\mathrm{d} r,\quad
\quad \mu_2=-\frac{1}{4\lambda_0^2}\int_\mathbb{R}\left(W^{\prime\prime\prime}(u_0)\mathcal{L}_0u_1-\eta_dW^{\prime\prime}(u_0)\right)\psi_0^2\mathrm{d} r,\\
&\mu_3=\frac{\eta_1}{2\lambda_0}-\frac{1}{4\lambda_0^2}\int_\mathbb{R}\left(W^{\prime\prime\prime}(u_0)(\mathcal{L}_0+2\lambda_0)u_1-\eta_dW^{\prime\prime}(u_0)\right)\psi_0^2\mathrm{d} r,\\
&\mu_4=\frac{1}{\lambda_0} \int_\mathbb{R} W^{\prime\prime\prime}(u_0)u_1\psi_1^2\mathrm{d} r,
\quad \mu_5=\frac{1}{\lambda_0^2}\int_\mathbb{R}\left(W^{\prime\prime\prime}(u_0)\mathcal{L}_0u_1-\eta_dW^{\prime\prime}(u_0)\right)\psi_1^2\mathrm{d} r,\\
& \mu_6=-\frac{\eta_1}{\lambda_0}+\frac{1}{\lambda_0} \int_\mathbb{R} W^{\prime\prime\prime}(u_0)u_1\psi_1^2\mathrm{d} r,\\
\end{aligned}
\end{equation}
and
\begin{equation}\label{e:Rjs}
\begin{aligned}
R_{2,2}=&2\nu_1\left(-a_{1,+}^2-6a_{1,+}a_{2,-}+2a_{1,-}^2+7a_{2,-}^2\right)+\nu_2(\frac{1}{2}B_1^2+B_2^2+2B_1B_3), \\
R_{2,8}=&8\nu_2\left[\left(a_{1,+} + 3 a_{2,-}\right)B_1+2 a_{1,-}B_2 -2(a_{1,+}-a_{2,-})B_3\right],\\
R_{3,2}=&\left(-\frac{2\nu_3}{3}+\nu_6\right)(a_{1,+}-a_{2,-})^3+2\nu_3\left[a_{1,-}^2(a_{1,+}-a_{2,-})- 2 a_{2,-}(a_{1,+}-a_{2,-})^2\right]+\\
&\left[\frac{3}{4}\nu_7(a_{1,+}-a_{2,-})-\nu_4 a_{2, -} \right]B_1^2+\nu_4\left[-a_{1,-}B_1B_2+2(a_{1,+}-a_{2,-})(2B_1B_3+ B_2^2)\right]+\rho({\bf A}),\\
R_{3,8}=&(a_{1,+}-a_{2,-})^2\left[4\nu_4(B_1-B_3)-6\nu_7B_1\right]+8\nu_4(a_{1,+}-a_{2,-})(2a_{2,-}B_1+a_{1,-}B_2)-\\
&4\nu_4a_{1,-}^2B_1-\left[\frac{1}{2}\nu_8B_1^3+\nu_5(B_1^2B_3+B_1B_2^2)\right]+\widetilde{\rho}({\bf A}),\\
\end{aligned}
\end{equation}
where
\begin{equation}\label{e:coco}
\begin{aligned}
&a_{1,+}=\frac{A_1+\bar{A}_1}{2},\quad a_{1,-}=\frac{A_1-\bar{A}_1}{2\mathrm{i}},\quad a_{2,+}=\frac{A_2+\bar{A}_2}{2},\quad a_{2,-}=\frac{A_2-\bar{A}_2}{2\mathrm{i}},\\
&\nu_1=-\frac{1}{4\lambda_0}\int_\mathbb{R} W^{\prime\prime\prime}(u_0)\psi_0^3\mathrm{d} r,\quad
\nu_2=-\frac{1}{4\lambda_0}\int_\mathbb{R} W^{\prime\prime\prime}(u_0)\psi_0\psi_1^2\mathrm{d} r,\\
&\nu_3=-\frac{1}{\lambda_0}\int_\mathbb{R} W^{\prime\prime\prime\prime}(u_0)\psi_0^4\mathrm{d} r, \quad
\nu_4=- \frac{1}{\lambda_0}\int_\mathbb{R} W^{\prime\prime\prime\prime}(u_0)\psi_0^2\psi_1^2\mathrm{d} r,\quad
\nu_5=- \frac{1}{\lambda_0}\int_\mathbb{R} W^{\prime\prime\prime\prime}(u_0)\psi_1^4\mathrm{d} r,\\
&\nu_6=\frac{1}{\lambda_0^2}\int_\mathbb{R} \big(W^{\prime\prime\prime}(u_0)\big)^2\psi_0^4\mathrm{d} r,\quad
\nu_7=\frac{1}{\lambda_0^2}\int_\mathbb{R} \big(W^{\prime\prime\prime}(u_0)\big)^2\psi_0^2\psi_1^2\mathrm{d} r,\quad
\nu_8=\frac{1}{\lambda_0^2}\int_\mathbb{R} \big(W^{\prime\prime\prime}(u_0)\big)^2\psi_1^4\mathrm{d} r,\\
&\rho({\bf A})=\int_\mathbb{R} Z({\bf A})\cdot ({\bf A}^T{\bf X}{\bf A})\mathrm{d} r,\quad
\widetilde{\rho}({\bf A})=\int_\mathbb{R} \widetilde{Z}({\bf A})\cdot \left( {\bf A}^T{\bf X}{\bf A}\right)\mathrm{d} r.\\
\end{aligned}
\end{equation}
In the last two expressions of $\rho$ and $\widetilde{\rho}$, the notation ``$\cdot$" denotes the Euclidean inner product in $\mathbb{R}^4$ and the expression of $X$ is as shown in \eqref{e:Xen} . Moreover, $Z({\bf A})$ and $\widetilde{Z}({\bf A})$ admits the forms of
\begin{equation}\label{e:Z}
\begin{aligned}
Z({\bf A})=&\frac{1}{2\lambda_0^2}W^{\prime\prime\prime}(u_0)\psi_0^2\left[\begin{pmatrix}\mathcal{L}_0\\0\\-2\\0\end{pmatrix}a_{1,+}+\begin{pmatrix}0\\2\lambda_0\\0\\0\end{pmatrix}a_{1,-}+\begin{pmatrix}4\lambda_0-\mathcal{L}_0\\0\\2\\0\end{pmatrix}a_{2,-}\right]+\\
&\frac{1}{4\lambda_0^2}W^{\prime\prime\prime}(u_0)\psi_0\psi_1\left[\begin{pmatrix}\mathcal{L}_0-\lambda_0\\0\\-2\\0\end{pmatrix}B_1+\begin{pmatrix}0\\-2\lambda_0\\0\\0\end{pmatrix}B_2+\begin{pmatrix}-2\lambda_0\\0\\0\\0\end{pmatrix}B_3\right],\\
\widetilde{Z}({\bf A})=&\frac{2}{\lambda_0^2}W^{\prime\prime\prime}(u_0)\psi_0\psi_1\left[\begin{pmatrix}\mathcal{L}_0-\lambda_0\\0\\2\\0\end{pmatrix}a_{1,+}+\begin{pmatrix}0\\-2\lambda_0\\0\\0\end{pmatrix}a_{1,-}+\begin{pmatrix}-3\lambda_0-\mathcal{L}_0\\0\\-2\\0\end{pmatrix}a_{2,-}\right]+\\
&\frac{1}{\lambda_0^2}W^{\prime\prime\prime}(u_0)\psi_1^2\left[\begin{pmatrix}-\mathcal{L}_0\\0\\2\\0\end{pmatrix}B_1+\begin{pmatrix}0\\2\lambda_0\\0\\0\end{pmatrix}B_2+\begin{pmatrix}2\lambda_0\\0\\0\\0\end{pmatrix}B_3\right],\\
\end{aligned}
\end{equation}
\end{Lemma}
\begin{Proof}
To simplify the calculation of the leading order terms of \eqref{e:2dred} in terms of ${\bf A}$ we introduce the following notation.
For any given integer $k\in\mathbb{Z}^+$, Banach spaces
$\{\mathcal{X}_j\}_{j=0}^k$ and a smooth map $F:\Pi_{j=1}^k\mathcal{X}_j\rightarrow \mathcal{X}_0$, we define
\[
F_p:=\big(\Pi_{j=1}^k(p_j)!\big)^{-1}\partial_{x_1}^{p_1}\cdot\ldots\cdot\partial_{x_k}^{p_k}F(0,\dots,0),
\]
where
\[
p=(p_1,\dots,p_k)\in \mathbb{Z}^k,\quad p_j\geq 0, \quad j=1,\dots,k.
\]
We note
\[
\mathbb{M}(0)=0,\quad \mathbb{F}(0,\varepsilon)\equiv 0, \quad \Psi(0,\varepsilon)\equiv 0, \quad \mathbb{F}_{(1,0)}=0,\quad \Psi_{(1,0)}=0,
\]
and conclude that the reduced system, up to cubic terms of $U_c$, is of the form
\begin{equation}
\label{e:2dred1}
\begin{aligned}
\dot{U}_c=&\mathbb{L}_*U_c+\varepsilon\mathbb{P}_c\mathbb{M}_{1} U_c+\mathbb{P}_c\mathbb{F}_{(2,0)}(U_c,U_c)+
\mathbb{P}_c\left(2\mathbb{F}_{(2,0)}(U_c,\Psi_{(2,0)}(U_c,U_c))+\mathbb{F}_{(3,0)}(U_c,U_c,U_c)\right),
\end{aligned}
\end{equation}
with the higher order terms in the form of
$ \mathcal{O}\left(|\varepsilon|^2\|U_c\|+|\varepsilon|\|U_c\|^2+\|U_c\|^4\right)$.
A direct calculation shows that
\begin{equation}
\label{e:coeffs}
\begin{aligned}
&\mathbb{M}_{1}=\frac{1}{\lambda_0}
\begin{pmatrix}
0 & 0 & 0 & 0 \\
W^{\prime\prime\prime}(u_0)u_1 & 0 & 0 & 0\\
0 & 0 & 0 & 0 \\
-\eta_dW^{\prime\prime}(u_0)+W^{\prime\prime\prime}(u_0)\mathcal{L}_0u_1 & 0 & W^{\prime\prime\prime}(u_0)u_1-\eta_1 & 0
\end{pmatrix},\\
& \mathbb{F}_{(2,0)}(U_c,U_c)=\frac{1}{\lambda_0}\left(
W^{\prime\prime\prime}(u_0)\left(2u_cv_c-u_c\mathcal{L}_0u_c+\lambda_0p_c^2\right)+\frac{1}{2}\mathcal{L}_0\left(W^{\prime\prime\prime}(u_0)u_c^2\right)\right)\mathcal{E}_4,\\
& \mathbb{F}_{(3,0)}(U_c,U_c,U_c)=\frac{1}{\lambda_0}\left(
W^{\prime\prime\prime\prime}(u_0)u_c\left(u_cv_c-\frac{1}{2}u_c\mathcal{L}_0u_c+\lambda_0p_c^2\right)+\frac{1}{6}\mathcal{L}_0\left(W^{\prime\prime\prime\prime}(u_0)u_c^3\right)
-\frac{1}{2}(W^{\prime\prime\prime}(u_0))^2u_c^3\right)\mathcal{E}_4,\\
&\mathbb{F}_{(2,0)}(U_c,\Psi_{(2,0)}(U_c,U_c))=
\frac{1}{2\lambda_0}\left( V_c\cdot \Psi_{(2,0,0)}(U_c,U_c) \right)\mathcal{E}_4,\\
\end{aligned}
\end{equation}
where $\mathcal{E}_4=(0,0,0,1)^T$, $u_1$ comes from the Taylor expansion,
\[
u_h(\varepsilon)=u_0+\varepsilon u_1+\mathcal{O}(\varepsilon^2),
\]
and
\begin{equation}
\label{e:vc}
V_c=\left(2W^{\prime\prime\prime}(u_0)v_c-W^{\prime\prime\prime}(u_0)\mathcal{L}_0u_c+[\mathcal{L}_0, W^{\prime\prime\prime}(u_0)u_c],2\lambda_0W^{\prime\prime\prime}(u_0)p_c, 2W^{\prime\prime\prime}(u_0)u_c, 0\right)^T.
\end{equation}
We also use the notation that $U_c=(u_c,p_c,v_c,q_c)^T$, where
\begin{equation}
\label{e:uc}
\begin{array}{ll}
u_c=2(a_{1,+}-a_{2,-})\psi_0+B_1\psi_1, & p_c=-2a_{1,-}\psi_0+B_2\psi_1, \\
v_c=-4\lambda_0a_{2,-}\psi_0+\lambda_0B_3\psi_1, & q_c=-4\lambda_0a_{2,+}\psi_0+\lambda_0B_4\psi_1.
\end{array}
\end{equation}
A direct calculation, using \eqref{e:coeffs}-\eqref{e:uc} and the expression of ${\bf X}$ \eqref{e:Xen}, leads to explicit expressions of the linear part ${\bf L}$, the quadratic part ${\bf R}_2$ and the cubic term ${\bf R}_3$. Relegating the calculation of $\Psi_{(2,0)}(U_c,U_c)$ into Lemma \ref{l:31}, we conclude our proof.
\end{Proof}
\begin{Lemma}\label{l:31}
The quadratic term of the center manifold, $\Psi_{(2,0)}(U_c,U_c)$, is a quadratic form of ${\bf A}$ and thus takes the form
\begin{equation}\label{e:qcm}
\Psi_{(2,0)}(U_c,U_c)=\mathbf{A}^\mathbf{T}{\bf X}{\bf A}
\end{equation}
where ${\bf X}=\{X_{jk}\}_{j,k=1}^8$ is symmetric
and every entry $X_{jk}\in \mathbb{P}_h\mathcal{Y}$.
More specifically, we have
we have
\begin{equation}\label{e:Xen}
\begin{aligned}
&\begin{cases}
&X_{11}=\bar{X}_{33}=\left(2\mathrm{i}-\mathbb{L}_*\right)^{-1}Y_{1}, \\
&X_{12}=\bar{X}_{34}=\left(2\mathrm{i}-\mathbb{L}_*\right)^{-1}\left(Y_{2}-X_{11}\right), \\
&X_{22}=\bar{X}_{44}=\left(2\mathrm{i}-\mathbb{L}_*\right)^{-1}\left(Y_{4}-2X_{12}\right) ,
\end{cases}
\begin{cases}
&X_{13}=-\mathbb{L}_*^{-1}Y_{3}, \\
&X_{14}=-\mathbb{L}_*^{-1}\left(-Y_{2}-X_{13}\right), \\
&X_{23}=-\mathbb{L}_*^{-1}\left(Y_{2}-X_{13}\right) ,\\
&X_{24}=-\mathbb{L}_*^{-1}\left(-Y_{4}-X_{14}-X_{23}\right),
\end{cases}
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
&\begin{cases}
&X_{15}=\bar{X}_{35}=\left(\mathrm{i}-\mathbb{L}_*\right)^{-1}Y_{5}, \\
&X_{16}=\bar{X}_{36}=\left(\mathrm{i}-\mathbb{L}_*\right)^{-1}\left(\mathrm{i} Y_{7}-X_{15}\right), \\
&X_{17}=\bar{X}_{37}=\left(\mathrm{i}-\mathbb{L}_*\right)^{-1}\left(Y_{7}-X_{16}\right) ,\\
&X_{18}=\bar{X}_{38}=\left(\mathrm{i}-\mathbb{L}_*\right)^{-1}\left(-X_{17}\right), \\
&X_{25}=\bar{X}_{45}=\left(\mathrm{i}-\mathbb{L}_*\right)^{-1}\left(Y_{6}-X_{15}\right), \\
&X_{26}=\bar{X}_{46}=\left(\mathrm{i}-\mathbb{L}_*\right)^{-1}\left(-X_{16}-X_{25}\right), \\
&X_{27}=\bar{X}_{47}=\left(\mathrm{i}-\mathbb{L}_*\right)^{-1}\left(\mathrm{i} Y_{7}-X_{17}-X_{26}\right) ,\\
&X_{28}=\bar{X}_{48}=\left(\mathrm{i}-\mathbb{L}_*\right)^{-1}\left(-X_{18}-X_{27}\right),
\end{cases}\quad
\begin{cases}
&X_{55}=-\mathbb{L}_*^{-1}Y_{8}, \\
&X_{56}=-\mathbb{L}_*^{-1}\left(-X_{55}\right), \\
&X_{57}=-\mathbb{L}_*^{-1}\left(Y_{9}-X_{56}\right) ,\\
&X_{58}=-\mathbb{L}_*^{-1}\left(-X_{57}\right),\\
&X_{66}=-\mathbb{L}_*^{-1}\left(Y_{9}-2X_{56}\right),\\
&X_{67}=-\mathbb{L}_*^{-1}\left(-X_{57}-X_{66}\right),\\
&X_{68}=-\mathbb{L}_*^{-1}\left(-X_{58}-X_{67}\right),\\
&X_{77}=-\mathbb{L}_*^{-1}\left(-2X_{67}\right),\\
&X_{78}=-\mathbb{L}_*^{-1}\left(-X_{68}-X_{77}\right),\\
&X_{88}=-\mathbb{L}_*^{-1}\left(-2X_{78}\right),
\end{cases}
\end{aligned}
\end{equation}
where, introducing $\mathcal{E}_4=(0,0,0,1)^T$ we have
\begin{equation}\label{e:Yen}
\begin{aligned}
&Y_1=\left(\frac{1}{2\lambda_0}\mathcal{L}_0-2\right) W^{{\prime\prime\prime}}(u_0)\psi_0^2 \mathcal{E}_4-
6\lambda_0\nu_1\psi_0 \mathcal{E}_4, \\
&Y_2=\mathrm{i}\left(\frac{1}{2\lambda_0}\mathcal{L}_0+1\right) W^{{\prime\prime\prime}}(u_0)\psi_0^2\mathcal{E}_4+
6\mathrm{i}\lambda_0\nu_1 \psi_0\mathcal{E}_4, \\
&Y_3=\frac{1}{2\lambda_0}\mathcal{L}_0 W^{{\prime\prime\prime}}(u_0)\psi_0^2\mathcal{E}_4+
2\lambda_0\nu_1 \psi_0\mathcal{E}_4,\\
&Y_4=-\left(\frac{1}{2\lambda_0}\mathcal{L}_0+3\right)W^{{\prime\prime\prime}}(u_0)\psi_0^2\mathcal{E}_4-
14\lambda_0\nu_1 \psi_0\mathcal{E}_4, \\
&Y_5=\left(\frac{1}{2\lambda_0}\mathcal{L}_0-\frac{1}{2}\right) W^{{\prime\prime\prime}}(u_0)\psi_0\psi_1\mathcal{E}_4-
2\lambda_0\nu_2 \psi_1,\mathcal{E}_4 \\
&Y_6=\mathrm{i}\left(\frac{1}{2\lambda_0}\mathcal{L}_0+\frac{3}{2}\right)W^{{\prime\prime\prime}}(u_0)\psi_0\psi_1\mathcal{E}_4+
6\mathrm{i}\lambda_0\nu_2 \psi_1\mathcal{E}_4, \\
&Y_7=W^{{\prime\prime\prime}}(u_0)\psi_0\psi_1\mathcal{E}_4+
4\lambda_0\nu_2 \psi_1\mathcal{E}_4, \\
&Y_8=\frac{1}{2\lambda_0}\mathcal{L}_0W^{{\prime\prime\prime}}(u_0)\psi_1^2\mathcal{E}_4+
2\lambda_0\nu_2 \psi_0\mathcal{E}_4, \\
&Y_9=W^{{\prime\prime\prime}}(u_0)\psi_1^2\mathcal{E}_4+
4\lambda_0\nu_2 \psi_0\mathcal{E}_4.
\end{aligned}
\end{equation}
\end{Lemma}
\begin{Proof}
To find the explicit expression of $\Psi_{(2,0)}(U_c,U_c)$ in terms of ${\bf A}$, we first recall \eqref{e:2ddFCH} and \eqref{e:sch} as follows.
\begin{equation*}
\begin{aligned}
\dot{U}=\mathbb{L}(\varepsilon)U+\mathbb{F}(U,\varepsilon), \quad U=U_c+\Psi(U_c,\varepsilon).
\end{aligned}
\end{equation*}
Plugging \eqref{e:sch} into \eqref{e:2ddFCH}, applying the projection $\mathbb{P}_h:=\mathrm{\,Id}\,-\mathbb{P}_c$ and setting $\varepsilon=0$, we obtain
\begin{equation}
\label{e:teh}
\dot{\Psi}(U_c,0)=\mathbb{L}_*\Psi(U_c,0)+\mathbb{P}_h\mathbb{F}\big(U_c+\Psi(U_c,0),0\big).
\end{equation}
For simplicity, we note that $\mathbb{P}_h\mathbb{F}_{(2,0)}(U_c,U_c)$ is a quadratic form of ${\bf A}$ and thus takes the form
\[
\mathbb{P}_h\mathbb{F}_{(2,0)}(U_c,U_c)=\mathbf{A}^\mathbf{T}{\bf Y}{\bf A},
\]
where ${\bf Y}=\{Y_{jk}\}_{j,k=1}^8$ is symmetric
and every entry $Y_{jk}\in \mathbb{P}_h\mathcal{Y}$.
Restricting \eqref{e:teh} to the quadratic terms of $U_c$ and plugging in \eqref{e:qcm}, we have, for all ${\bf A}$,
\begin{equation*}
{\bf A}^T\left({\bf L}(0)^T{\bf X}+{\bf X}{\bf L}(0)\right){\bf A}= {\bf A}^T(\mathbb{L}_*{\bf X}){\bf A}+\mathbf{A}^\mathbf{T}{\bf Y}{\bf A},
\end{equation*}
that is,
\begin{equation}\label{e:XY}
{\bf L}(0)^T{\bf X}+{\bf X}{\bf L}(0)-\mathbb{L}_*{\bf X}={\bf Y},
\end{equation}
from which it is not hard to compute all entries of ${\bf X}$ recursively.
More explicitly, ${\bf Y}$ admits the form
\begin{equation}\label{e:Y}
\begin{pmatrix}
Y_1 & Y_2 & Y_3 & -Y_2 & Y_5 & \mathrm{i} Y_7 &Y_7 &0 \\
Y_2 & Y_4 & Y_2 & -Y_4 & Y_6 & 0 & \mathrm{i} Y_7 &0 \\
Y_3 & Y_2 & Y_1 &-Y_2 & Y_5 & -\mathrm{i} Y_7 & Y_7 &0 \\
-Y_2 & -Y_4 & -Y_2 & Y_4 & -Y_6 & 0 & -\mathrm{i} Y_7 &0 \\
Y_5 & Y_6 & Y_5 & -Y_6 & Y_{8} & 0 & Y_{9} &0 \\
\mathrm{i} Y_7 & 0 & -\mathrm{i} Y_7 &0 & 0 & Y_{9} & 0 & 0\\
Y_7 & \mathrm{i} Y_7 & Y_7 & -\mathrm{i} Y_7 & Y_{9} & 0 & 0 & 0\\
0 & 0 & 0 &0 & 0 & 0 & 0 &0\\
\end{pmatrix},
\end{equation}
where $Y_j$'s admit the expressions as in \eqref{e:Yen}.
Plugging \eqref{e:Y} into \eqref{e:XY}, we obtain the expression of $X_{ij}$'s as in \eqref{e:Xen}.
\end{Proof}
\subsection{Explicit expressions of $\alpha_1$ and $\alpha_2$}\label{ss:32}
\begin{Lemma}\label{l:32}
Among the coefficients of cubic terms of the normal form system \eqref{e:2dnfeq}, we have
\begin{equation}\label{e:alphas}
\begin{cases}
&\alpha_1=0,\\
&\alpha_2=-\frac{\nu_3}{3}+\frac{80}{9}\nu_1^2+\int_\mathbb{R}\left(W^{{\prime\prime\prime}}(u_0)\psi_0^2+4\lambda_0\nu_1\psi_0\right)\widetilde{\mathcal{L}}\left(W^{{\prime\prime\prime}}(u_0)\psi_0^2+4\lambda_0\nu_1\psi_0\right)\mathrm{d} r,
\end{cases}
\end{equation}
where $\widetilde{\mathcal{L}}:=\frac{1}{3\lambda_0^2}\left(\frac{1}{2}+2\lambda_0\mathcal{L}_0^{-1}+2\lambda_0(\mathcal{L}_0-4\lambda_0)^{-1}-\lambda_0(\mathcal{L}_0-4\lambda_0)^{-2}\right)$ is a self-adjoint operator.
\end{Lemma}
\vskip -0.4in
\begin{Remark}
The techniques used in the proof of Lemma \ref{l:32} permit the calculation of explicit expressions for each $\alpha_j$ and $\beta_k$. Nevertheless,
we only present the calculations of $\alpha_1$ and $\alpha_2$, as the other coefficients are not needed in the sequel.
\end{Remark}
\vskip -0.4in
\begin{Proof}
To calculate all these coefficients, we first recall the equality \eqref{e:nf3} with $\mathscr{R}_2=0$,
\begin{equation*}
\big(\mathcal{D}-\mathbf{L}(0,0)\big)\Phi_3=\mathbf{R}_3+2\mathbf{R}_2(\mathbf{C},\Phi_2)-\mathscr{R}_3,
\end{equation*}
and the restrictions
\[
\begin{pmatrix}\mathbf{R}_{3,1}+2\mathbf{R}_{2,1}(\mathbf{C},\Phi_2)-\mathscr{R}_{3,1}\\
\mathbf{R}_{3,2}+2\mathbf{R}_{2,2}(\mathbf{C},\Phi_2)-\mathscr{R}_{3,2}\end{pmatrix}\in
\left(\ker\left(\begin{pmatrix}\mathcal{D}^{\mathrm{ad}}+\mathrm{i}&0\\-1&\mathcal{D}^{\mathrm{ad}}+\mathrm{i}\end{pmatrix}\mid_{\mathbf{P}_3^2}\right)\right)^\perp,
\]
we have that $\alpha_1$ is exactly the coefficient of $C_1^2\bar{C}_1$ in
\[\mathbf{R}_{3,2}+2\mathbf{R}_{2,2}(\mathbf{C},\Phi_2),\]
that is,
\begin{equation}\label{e:alpha0}
\alpha_1=\frac{1}{2}\langle \mathbf{R}_{3,2}+2\mathbf{R}_{2,2}(\mathbf{C},\Phi_2)\mid C_1^2\bar{C}_1\rangle,
\end{equation}
where we recall that this inner product is the one of polynomials, defined as $\langle P \mid Q \rangle=P(\partial_{{\bf C}})\bar{Q}(C)$.
According to \eqref{e:Rjs} and \eqref{e:phi2}, we have
\begin{equation*}
\begin{aligned}
\alpha_1=\frac{1}{2}\langle \mathbf{R}_{3,2}+2\mathbf{R}_{2,2}(\mathbf{C},\Phi_2)\mid C_1^2\bar{C}_1\rangle=-6\nu_1^2+\frac{3}{8}\nu_6+\frac{1}{2}\langle\rho({\bf C})\mid C_1^2\bar{C}_1\rangle.
\end{aligned}
\end{equation*}
From the expression of $\rho({\bf A})$ in \eqref{e:coco}, it is straight forward to see that
\begin{equation}
\label{e:rhoalpha1}
\begin{aligned}
\frac{1}{2}\langle\rho({\bf C})\mid C_1^2\bar{C}_1\rangle&=\frac{1}{2}\int_\mathbb{R} \langle Z({\bf C})\cdot(X_{11}C_1^2+2X_{13}C_1\bar{C}_1)\mid C_1^2\bar{C}_1\rangle\mathrm{d} r,
\end{aligned}
\end{equation}
Based on \eqref{e:Xen} and \eqref{e:Yen}, a direct calculation shows that
\begin{equation*}
\begin{aligned}
X_{13}=-\mathbb{L}_*^{-1}Y_3&=\frac{1}{2}\begin{pmatrix}\mathcal{L}_0^{-1}\\0\\1\\0\end{pmatrix}\left(W^{{\prime\prime\prime}}(u_0)\psi_0^2+4\lambda_0\nu_1\psi_0\right),\\
X_{11}=-(\mathbb{L}_*-2\mathrm{i})^{-1}Y_1&=\frac{1}{2}\begin{pmatrix}\left(\mathcal{L}_0-4\lambda_0\right)^{-1}\\2\mathrm{i}\left(\mathcal{L}_0-4\lambda_0\right)^{-1}\\1\\2\mathrm{i}\end{pmatrix}\left(W^{{\prime\prime\prime}}(u_0)\psi_0^2+4\lambda_0\nu_1\psi_0\right).
\end{aligned}
\end{equation*}
Plugging \eqref{e:Z} and the above expressions into \eqref{e:rhoalpha1}, we have
\begin{equation*}
\begin{aligned}
\frac{1}{2}\langle\rho({\bf C})\mid C_1^2\bar{C}_1\rangle&=-\frac{3}{8}\nu_6+6\nu_1^2.
\end{aligned}
\end{equation*}
Therefore, we deduce that $\alpha_1=0.$
A similar calculation shows that
\[
\alpha_2=-\frac{\nu_3}{3}+\frac{80}{9}\nu_1^2+\int_\mathbb{R}\left(W^{{\prime\prime\prime}}(u_0)\psi_0^2+4\lambda_0\nu_1\psi_0\right)\widetilde{\mathcal{L}}\left(W^{{\prime\prime\prime}}(u_0)\psi_0^2+4\lambda_0\nu_1\psi_0\right)\mathrm{d} r,
\]
where $\widetilde{\mathcal{L}}:=\frac{1}{3\lambda_0^2}\left(\frac{1}{2}+2\lambda_0\mathcal{L}_0^{-1}+2\lambda_0(\mathcal{L}_0-4\lambda_0)^{-1}-\lambda_0(\mathcal{L}_0-4\lambda_0)^{-2}\right)$ is a self-adjoint operator.
\end{Proof}
\subsection*{Acknowledgment}
The first author acknowledges support from NSF DMS grants 1109127 and 1409940.
\bibliographystyle{siam}
|
\section{Introduction}
The dynamics of a flashing ratchet can be translated into a
counterintuitive phenomenon in gambling games which has recently
attracted considerable attention. It is the so-called {\em
Parrondo's paradox} \cite{stat,nature,fnl} consisting of two
losing games, A and B, that yield, when alternated, a winning
game.
In game A, a player tosses a coin and makes a bet on the throw. He
wins or loses 1 euro depending on whether the coin falls heads or tails.
The probability $p_1$ of winning is
$p_1=1/2-\epsilon\text{ with }0\leq\epsilon \ll 1$; so game A is
fair when $\epsilon=0$ and losing when $\epsilon>0$. By losing,
winning, and fair games here we mean that the average capital is a
decreasing, increasing, and a constant function of the number of
turns, respectively.
The second game ---or game B--- consists of two coins. The player
must throw coin 2 if his capital is not a multiple of three, and
coin 3 otherwise. The probability of winning with coin 2 is
$p_2=3/4-\epsilon$ and with coin 3 is $p_3=1/10-\epsilon$. They
are called ``good'' and ``bad'' coins respectively. It can be
shown that game B is also a losing game if $\epsilon>0$ and that
$\epsilon=0$ makes B a fair game \cite{fnl,newpar}. The rules of
both game A and B are depicted in fig. \ref{fig:rules}
\begin{figure}
\begin{center}
\includegraphics[width=10cm]{fig1-rules.eps}
\caption{Rules of the two Parrondo games}
\label{fig:rules}
\end{center}
\end{figure}
Surprisingly, switching between games A and B in a random fashion
or following some periodic sequences produces a winning game, for
$\epsilon>0$ sufficiently small, i.e., the average of player
earnings grows with the number of turns. Therefore, from two
losing games we actually get a winning game. This indicates that
the alternation of stochastic dynamics can result in a new
dynamics, which differs qualitatively from the original ones.
Alternation is either periodic or random in the flashing rachet
and in the paradoxical games. On the other hand, we have recently
studied the case of a {\em controlled} alternation of games, where
information about the state of the system can be used to select
the game to be played with the goal of maximising the capital
\cite{dinis}. This problem is trivial for a single player: the
best strategy is to select game A when his capital is a multiple
of three and B otherwise. This yields higher returns than any
periodic or random alternation. Therefore, choosing the game as a
function of the current capital presents a considerable advantage
with respect to ``blind'' strategies, i.e., strategies that do not
make use of any information about the state of the system, as it
is the case of the periodic and random alternation. Also, in a
flashing ratchet, switching on and off the ratchet potential
depending on the location of the Brownian particle allows one to
extract energy from a single thermal bath, in apparent
contradiction with the second law of thermodynamics \cite{review}.
This is nothing but a Maxwell demon, who operates having at his
disposal information about the position of the particle; and it is
the acquisition or the subsequent erasure of this information what
has an unavoidable entropy cost \cite{leff}, preventing any
violation of the second law.
Whereas a controlled alternation of games is trivial for a single
player, interesting and counter-intuitive phenomena can be found
in {\em collective} games. We have recently considered a
collective version of the original Parrondo's paradox.
In this model, the game ---A or B--- that a
large number $N$ of individuals play can be selected at every
turn. It turns out that blind strategies are winning whereas a
strategy which chooses the game with the highest average return is
losing \cite{dinis}.
In this paper, we extend our investigation of controlled
collective games considering a new strategy based on a majority
rule, i.e., on voting. This type of rule is relevant in several
situations, such as the modelling of public opinion
\cite{galam,krap} or the design of multi-layer neural networks by
means of {\em committee machines} \cite{barkai,cris}. We will show
that, in controlled games, the rule is very inefficient: if every
player votes for the game that gives him the highest return, then
the total capital decreases, whereas blind strategies generate a
steady gain. The same effect can be found for the
capital-independent games introduced in \cite{newpar}. As
mentioned above, for a single player, the majority rule does
defeats the blind strategies. The inefficiency of voting is
consequently a purely collective effect.
The paper is organised as follows. In Section \ref{sec:model} we
present the model and the counter-intuitive performance of the
different strategies. In Section \ref{sec:analysis} we discuss and
provide an intuitive explanation of this behaviour. In Sec.
\ref{sec:finite}, we analyse how the effect depends on the number
of players. In Section \ref{sec:history}, we extend these ideas to
the capital-independent games introduced in \cite{newpar}.
Finally, in Sec. \ref{sec:conclusions} we present our main
conclusions.
\section{The model}
\label{sec:model}
The model consists of a large number $N$ of players. In every
turn, they have to choose one of the two original Parrondo games,
described in the Introduction and in fig. \ref{fig:rules}. Then
{\em every} individual plays the selected game against the casino.
We will consider three strategies to achieve the collective
decision. {\em a)} The {\em random strategy}, where the game is
chosen randomly with equal probability. {\em b)} The {\em
periodic strategy}, where the game is chosen following a given
periodic sequence. The sequence that we will use throughout the
paper is $ABBABB\dots$ since it is the one giving the highest
returns. {\em c)} The {\em majority rule (MR) strategy}, where
every player votes for the game giving her the highest probability
of winning, with the game obtaining the most votes being selected.
The model is related to other extensions of the original Parrondo
games played by an ensemble of players, such as those considered
by Toral \cite{toral,toralcapred}. However, in our model the only
interaction among players can occur when the collective decision
is made. Once the game has been selected, each individual plays,
in a completely independent way, against the casino. Moreover, in
the periodic and random strategies there is no interaction at all
among the players, the model being equivalent to the original
Parrondo's paradox with a single player.
The MR makes use of the information about the state of the system,
whereas the periodic and random strategies are blind, in the sense
defined above. One should then expect a better performance of the
MR strategy. However, it turns out that, for large $N$, these
blind strategies produce a systematic winning whereas the MR
strategy is losing. This is shown in figure \ref{capital}, where
the capital per player as a function of the number of turns is
depicted for the three strategies and an infinite number of
players (see Appendix \ref{app} for details on how to obtain fig.
\ref{capital}).
\begin{figure}
\begin{center}
\includegraphics[width=7cm]{fig2-cap.eps}
\caption{Evolution of the capital per player in an infinite
ensemble for $\epsilon=0.005$ and the three strategies discussed
in the text. } \label{capital}
\end{center}
\end{figure}
\section{Analysis}
\label{sec:analysis}
How many players vote for each game? The key magnitude to answer
this question and to explain the system's behaviour is
$\pi_0(t)$, the fraction of players whose money is a multiple of
three at turn $t$. This fraction $\pi_0(t)$ of players vote for
game A in order to avoid the bad coin in game B. On the other
hand, the remaining fraction $1-\pi_0(t)$ vote for game B to play
with the good coin. Therefore, if $\pi_0(t)\geq 1/2$, there are
more votes for game A and, if $\pi_0(t)< 1/2$, then game B is
preferred by the majority of the players.
Let us focus now on the behaviour of $\pi_0(t)$ for $\epsilon=0$
when playing both games separately. If game A is played a large
number of times, $\pi_0(t)$ tends to 1/3 because the capital is a
symmetric and homogenous random walk under the rules of game A. On
the other hand, if B is played repeatedly, $\pi_0(t)$ tends to
5/13. This can be proved by analyzing game B as a Markov chain
\cite{fnl,newpar}. It is also remarkable that, for
$\pi_0(t)=5/13$, the average return when game B is played is zero.
Figure \ref{scheme} represents schematically the evolution of
$\pi_0(t)$ under the action of each game, as well as the
prescription of the MR strategy explained above.
Now we are ready to
explain why the MR strategy yields worse results than
the periodic and random sequences.
\begin{figure}
\begin{center}
\includegraphics[width=7cm]{fig3-esq.eps}
\caption{Schematic representation of the evolution of $\pi_0(t)$
under the action of game A and game B. The prescription of the MR
is also represented.}\label{scheme}
\end{center}
\end{figure}
We see that, as long as
$\pi_0(t)$ does not exceed $1/2$, the MR strategy chooses game B.
However, playing B takes $\pi_0$ closer to $5/13$, well below
$1/2$, and thus more than half of the players vote for game B
again.
After a number
of runs, the MR strategy gets trapped playing game B forever. Then
$\pi_0$ asymptotically approaches 5/13, and as this happens, game
B turns into a fair game when $\epsilon=0$. As a consequence, the
MR will not produce earnings any more, as can be seen in figure
\ref{mayoriae0}.
\begin{figure}
\begin{center}
\includegraphics[width=13cm]{fig4-may.eps}
\caption{Evolution of $\pi_0(t)$ (left) and the capital per player
(right) for $N=\infty$, $\epsilon=0$ for the MR and random
strategies. The MR chooses game B when $\pi_0$ is below the
straight
line depicted at 1/2 and game A otherwise.
} \label{mayoriae0} \end{center}
\end{figure}
The introduction of $\epsilon>0$ turns game B into a losing game
if played repeatedly. Consequently, the MR strategy becomes a
losing one as in figure \ref{capital}. To overcome this losing
tendency, the players must sacrifice their short-range profits,
not only for the benefit of the whole community but also for their
own returns in the future. Hence, some kind of cooperation among
the players is needed
to prevent them from losing their capital. A similar effect has been found by Toral in
another version of collective Parrondo's games. There, sharing the capital among
players induces a steady gain \cite{toral}.
In our case, the striking result is that no complex cooperation is
necessary. It is enough that the players agree to vote at random.
\section{Finite number of players}
\label{sec:finite}
In the previous analysis an infinite number of players has been
considered. Remarkably, for just one player the
MR strategy trivially performs
better than any periodic or random sequence, since it completely
avoids the use of the bad coin. In this Section we analyse the
crossover between the winning behaviour for a small number of
players and the losing behaviour when this number is large.
Figure \ref{figura5} shows numerical results of the average
capital per player for an increasing number of players ranging
from 10 to 1000.
\begin{figure}
\begin{center}
\includegraphics[width=10cm]{fig5-nfin.eps}
\caption{Simulation results for the average capital per player for
$N=10$, 50, 100, and 1000 players, $\epsilon=0.005$, and the three
different strategies. The simulations have been made over a
variable number of realizations, ranging from 10000 realizations
for $N=1$ to 10 realizations for $N=1000$. Simulations for the
random and periodic strategies have been made with $N=100$ players
and averaging over 100 realizations. For these blind strategies,
the result does not depend on the number of players $N$.}
\label{figura5}
\end{center}
\end{figure}
One can observe that, the larger the number of players, the worse
the results for the MR strategy, becoming losing for a number of
players between 50 and 100.
The above discussion for an infinite ensemble allows us to give a
qualitative explanation. The difference between large and small
$N$ is the magnitude of the fluctuations of $\pi_0(t)$ around its
expected value. If game B is chosen a large number of times in a
row, then the expected value of $\pi_0(t)$ is 5/13. On the other
hand, the MR selects B unless $\pi_0(t)$ is above 1/2. Therefore,
for the MR to select A, fluctuations must be of order
$1/2-5/13=3/26\simeq 0.115$. For $N$ players, the fraction of
players with capital multiple of three, $\pi_0(t)$, will be a
random variable following a binomial distribution, at least if B
has been played a large number of times in a row. If the expected
value of $\pi_0(t)$ is 5/13, fluctuations of $\pi_0(t)$ around
this value are of order $\sqrt{5/13\times 8/13\times 1/N}$. Then,
fluctuations will allow the MR strategy to choose A if $N\simeq
20$. Far above this value, fluctuations that drive $\pi_0(t)$
above 1/2 are very rare, and MR chooses B at every turn. On the
other hand, for $N$ around or below 20, there is an alternation of
the games that can even beat the optimal periodic strategy.
We see that the MR strategy can take profit of fluctuations much
better than blind strategies, but it loses all its efficiency when
these fluctuations are small. We believe that this is closely
related to the second law of thermodynamics. The law prohibits any
decrease of entropy only in the thermodynamic limit or for average
values. On the other hand, when fluctuations are present, entropy
can indeed decrease momentarily and this decrease can be exploited
by a Maxwell demon.
\section{History dependent games}
\label{sec:history}
A similar phenomenon is exhibited by the games introduced in Ref.
\cite{newpar}, whose rules depend on the history rather than on
the capital of each player. Game A is still the same as above,
whereas game B is played with three coins according to the
following table:
\begin{center}
\begin{tabular}{c|c|c|c}
Before last & Last& Prob. of win & Prob. of loss \\
$t-2$ & $t-1$ & at $t$ & at $t$ \\\hline loss & loss &
$p_1$ & $1-p_1$\\ loss & win & $p_2$ & $1-p_2$
\\ win & loss & $p_2$ & $1-p_2$ \\ win & win &
$p_3$ & $1-p_3$ \\
\end{tabular}
\end{center}
with $p_1=9/10-\epsilon$, $p_2=1/4-\epsilon$, and
$p_3=7/10-\epsilon$.
Introducing a large number of players but allowing just a randomly
selected fraction $\gamma$ of them to vote and play, the same ``voting paradox''
is recovered for sufficiently small $\gamma$. Again, blind strategies achieve a
constant growth of the average capital with the number of turns while the MR
strategy returns a decreasing average capital, as it is shown in figure
\ref{histo}.
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{fig6-npg.eps}
\caption{Evolution of the average money of the players in the
history-dependent games for $\gamma=0.5$, $\epsilon=0.005$ and
three different
strategies.} \label{histo} \end{center}
\end{figure}
\section{Conclusions}
\label{sec:conclusions}
We have shown that the paradoxical games based on the flashing
ratchet exhibit a counterintuitive phenomenon when a large number
of players are considered. A majority rule based on selfish voting
turns to be very inefficient for large ensembles of players. We
have also discussed how the rule only works for a small number of
players, since in that case it is able to exploit capital
fluctuations.
The interest of the model presented here is threefold. First of
all, it shows that cooperation among individuals can be beneficial
for everybody. In this sense, the model is related to that
presented by Toral in Ref. \cite{toralcapred}. Since John Maynard
Smith first applied game theory to biological problems
\cite{maynard}, games have been used in ecology and social
sciences as models to explain social behaviour of individuals
inside a group. Some generalizations of the voting model might be
useful for this purpose. For instance, it could be interesting to
analyse the effect of mixing selfish and cooperative players or
the introduction of players who could change their behaviour
depending on the fraction of selfish voters in previous turns.
Secondly, the effect can also be relevant in random decision
theory or the theory of stochastic control \cite{white} since it
shows how periodic or random strategies can be better than some
kind of optimization. In this sense, there has been some work on
general adaptive strategies in games related with Parrondo's
paradox \cite{behrends,rahmann}.
Thirdly, this model and, in particular, the analysis for $N$
finite, prompts the problem of how information can be used to
improve the performance of a system. In the models presented here,
information about the fluctuations of the capital is useful only
for a small number of players, that is, when these fluctuations
are significant. It will be interesting to analyse this crossover
in further detail, not only in the case of the games but also for
Brownian ratchets. Work in this direction is in progress.
\begin{ack}
The authors gratefully acknowledge fruitful discussions with
Christian Van den Broeck, who suggested us the introduction of the
MR strategy. We also thank H. Leonardo Mart\'{\i}nez for valuable
comments on the manuscript.
This work has been financially supported by grant
BFM2001-0291-C02-02 from Ministerio de Ciencia y Tecnolog\'{\i}a
(Spain) and by a grant from the {\em New del Amo Program}
(Universidad Complutense).
\end{ack}
|
\section{Preliminaries}
Recall that a space is \emph{crowded} if it is non-empty and has no isolated points. We will write $X\approx Y$ to mean that the spaces $X$ and $Y$ are homeomorphic. Given a subset $X$ of a space $Z$ and a subgroup $\mathcal{H}$ of $\mathcal{H}(Z)$, we will let
$$
\mathcal{H}[X]=\bigcup_{h\in\mathcal{H}}h[X]
$$
be the closure of $X$ under the action of $\mathcal{H}$. Notice that $\mathcal{H}[\mathcal{H}[X]]=\mathcal{H}[X]$. For simplicity, we will let $\mathcal{H}(x)=\mathcal{H}[\{x\}]$. Furthermore, given a group $\mathcal{H}$ and a subset $\mathcal{S}$ of $\mathcal{H}$, we will denote by $\langle\mathcal{S}\rangle$ the subgroup of $\mathcal{H}$ generated by $\mathcal{S}$.
Given a surjection $\pi:X\longrightarrow Y$, we will say that a subset $S$ of $X$ is \emph{saturated} with respect to $\pi$ if $\pi^{-1}[\pi[S]]=S$. The proof of the following simple lemma is left to the reader.
\begin{lemma}\label{pi}
Let $\pi:X\longrightarrow Y$ be a continuous surjection between compact spaces. If $A\subseteq X$ is saturated with respect to $\pi$ then $\pi\upharpoonright A:A\longrightarrow\pi[A]$ is a closed continuous surjection.
\end{lemma}
The following lemma, which originally appeared as \cite[Theorem 2.3]{vanmillh} in a slightly different form, will be the key to making our example homogeneous. We need some notation, which will be used throughout the entire paper. Let
$$
Q=\{x\in 2^\omega:\text{there exists }m\in\omega\text{ such that }x_n=x_m\text{ whenever }n\geq m\}.
$$
Given $q\in Q$, let $h_q\in\mathcal{H}(2^\omega)$ be the homeomorphism defined by $h_q(x)=x+q$ for $x\in2^\omega$, where $+$ denotes addition modulo $2$. Let $\mathcal{V}=\{h_q:q\in Q\}$, and observe that $\mathcal{V}$ is a subgroup of $\mathcal{H}(2^\omega)$.
\begin{lemma}\label{jan}
Assume that $X$ is a subspace of $2^\omega$ such that $\mathcal{V}[X]=X$. Then $X$ is homogeneous.
\end{lemma}
\begin{proof}
Fix the usual metric $\mathsf{d}$ on $2^\omega$ defined by $\mathsf{d}(x,y)=\sum_{n\in\omega}(|x_n-y_n|/2^n)$. By \cite[Lemma 2.1]{vanmillh} (see also \cite[Corollary 1.9.2]{vanmilli}), it will be enough to show that if $x,y\in X$ then $x$ and $y$ have arbitrarily small homeomorphic neighborhoods. This is straightforward, using the fact that each $h_q$ is an isometry with respect to $\mathsf{d}$.
\end{proof}
The following classical result, which is a well-known tool for ``killing'' homeomorphisms (see \cite{vanmills} for other applications), will be the key to obtaining rigid spaces. For a proof of Lemma \ref{lav},
see \cite[Theorem 3.9 and Exercise 3.10]{kechris}.
\begin{lemma}[Lavrentiev]\label{lav}
Let $f:W\longrightarrow W$ be a homeomorphism, where $W$ is a subspace of some Polish space $Z$. Then there exists a $\mathsf{G_\delta}$ subset $T$ of $Z$ and a homeomorphism $g:T\longrightarrow T$ such that $f\subseteq g$.
\end{lemma}
Our reference for set theory will be \cite{kunen}. We will denote by $\mathsf{MA(\sigma\textrm{-}centered)}$ the statement that Martin's Axiom for $\sigma$-centered posets holds. Recall that $\mathsf{add(meager)}$ is the minimum cardinal $\kappa$ such that there exists a collection $\mathcal{C}$ of size $\kappa$ consisting of meager subsets of $2^\omega$ such that $\bigcup\mathcal{C}$ is non-meager in $2^\omega$.\footnote{\,In \cite[Definitions III.1.2 and III.1.6]{kunen}, Kunen uses $\mathbb{R}$ instead of $2^\omega$. Since $\mathbb{R}\setminus\mathbb{Q}\approx\omega^\omega\approx 2^\omega\setminus Q$, it is easy to see that this makes no difference.} It is clear that $\omega_1\leq\mathsf{add(meager)}\leq\mathfrak{c}$. Furthermore, it is well-known that $\mathsf{MA(\sigma\textrm{-}centered)}$ implies $\mathsf{add(meager)}=\mathfrak{c}$ (see for example \cite[Lemmas III.3.22, III.3.26 and III.1.25]{kunen}).
The following lemma, which first appeared as \cite[Lemma 3.2]{baldwinbeaudoin}, will be the key to obtaining our countable dense homogeneous example. See \cite[Corollary 2.2]{medini} for a simpler version of the proof. Given an infinite cardinal $\lambda\leq\mathfrak{c}$ and a subset $D$ of $2^\omega$, we will say that $D$ is \emph{$\lambda$-dense} if $|D\cap U|=\lambda$ for every non-empty open subset $U$ of $2^\omega$.
\begin{lemma}[Baldwin, Beaudoin]\label{bb}
Assume $\mathsf{MA(\sigma\textrm{-}centered)}$. Let $\kappa<\mathfrak{c}$ be a cardinal. Suppose that $A_\alpha$ and $B_\alpha$ are $\lambda_\alpha$-dense subsets of $2^\omega$ for $\alpha < \kappa$, where each $\lambda_\alpha<\mathfrak{c}$ is an infinite cardinal. Also assume that $A_\alpha\cap A_\beta=\varnothing$ and $B_\alpha\cap B_\beta=\varnothing$ whenever $\alpha\neq\beta$. Then there exists $f\in\mathcal{H}(2^\omega)$ such that $f[A_\alpha]=B_\alpha$ for every $\alpha < \kappa$.
\end{lemma}
\section{The general method}
The following theorem gives a general method for embedding suitable zero-dimensional spaces into $2^\omega$, so that their complement will be non-trivial and rigid. The strategy of its proof is to combine Lemma \ref{lav} with the idea of ``splitting points'' in a linearly ordered space, which dates back to the classical double arrow space of Alexandroff and Urysohn (see \cite{alexandroffurysohn}).
\begin{theorem}\label{general}
Assume that the following requirements are satisfied.
\begin{itemize}
\item $X$ is a subspace of $2^\omega$.
\item $Y=2^\omega\setminus X$.
\item $\mathcal{H}$ is a subgroup of $\mathcal{H}(2^\omega)$.
\item $D$ is a countable dense subset of $2^\omega$ such that $D\subseteq Y$ and $D\cap Q=\varnothing$.
\item $\mathcal{G}$ is the collection of all homeomorphisms $g$ such that $\mathsf{dom}(g)$ and $\mathsf{ran}(g)$ are $\mathsf{G_\delta}$ subspaces of $2^\omega$.
\end{itemize}
Furthermore, assume that the following conditions hold.
\begin{enumerate}
\item\label{preserve} $\mathcal{H}[X]=X$.
\item\label{pushaway} If $h\in\mathcal{H}\setminus\{\mathsf{id}\}$ then $h[D]\cap D=\varnothing$.
\item\label{kill} If $g\in\mathcal{G}$ and $|\{x\in\mathsf{dom}(g):g(x)\notin\mathcal{S}(x)\}|=\mathfrak{c}$ for every subgroup $\mathcal{S}$ of $\mathcal{H}$ such that $|\mathcal{S}|<\mathsf{add(meager)}$ then there exists $z\in\mathsf{dom}(g)\cap X$ such that $g(z)\notin X$.
\end{enumerate}
Then there exists a subspace $X^\ast\approx X$ of $2^\omega$ such that $2^\omega\setminus X^\ast$ is dense in $2^\omega$ and rigid.
\end{theorem}
\begin{proof}
Let $Z=2^\omega$. First, we will show that $Y\cap U$ is uncountable for every non-empty open subset $U$ of $Z$. Condition (\ref{pushaway}) implies that $h[D]\cap g[D]=\varnothing$ whenever $h,g\in\mathcal{H}$ and $h\neq g$. In particular, since $\mathcal{H}[D]\subseteq\mathcal{H}[Y]=Y$ by condition (\ref{preserve}), the desired conclusion holds if $\mathcal{H}$ is uncountable. Now assume that $\mathcal{H}$ is countable. We will actually show that $X$ is a Bernstein set.\footnote{\,Recall that a subset $X$ of $2^\omega$ is a \emph{Bernstein set} if $X\cap K\neq\varnothing$ and $(2^\omega\setminus X)\cap K\neq\varnothing$ for every perfect subset $K$ of $2^\omega$. Notice that the complement of a Bernstein set is also a Bernstein set. Since $2^\omega\approx 2^\omega\times 2^\omega$, every Bernstein set is $\mathfrak{c}$-dense in $2^\omega$.} Fix a nowhere dense perfect subset $K$ of $Z$. Notice that $Z\setminus\mathcal{H}[K]$ is a comeager subset of $Z$, hence it contains a perfect subset $K'$. Fix a homeomorphism $g:K\longrightarrow K'$, and notice that $g\in\mathcal{G}$. Since $|\{x\in\mathsf{dom}(g):g(x)\notin\mathcal{H}(x)\}|=|K|=\mathfrak{c}$ and $|\mathcal{H}|=\omega<\mathsf{add(meager)}$, condition (\ref{kill}) shows that $X\cap K\neq\varnothing$. The same reasoning, applied to $g^{-1}$, shows that $Y\cap K\neq\varnothing$.
Let $D^\ast=\{d^-:d\in D\}\cup\{d^+:d\in D\}$, where we use the notation $d^-=\langle d,-1\rangle$ and $d^+=\langle d,1\rangle$ for $d\in D$. Define
$$
Z^\ast=(Z\setminus D)\cup D^\ast.
$$
Consider the function $\pi:Z^\ast\longrightarrow Z$ defined by the following two conditions.
\begin{itemize}
\item $\pi(d^-)=\pi(d^+)=d$ for every $d\in D$.
\item $\pi\upharpoonright (Z\setminus D)=\mathsf{id}$.
\end{itemize}
Let $\prec$ denote the linear ordering on $Z^\ast$ defined by the following two conditions.
\begin{itemize}
\item $d^-$ is the immediate predecessor of $d^+$ for every $d\in D$.
\item $x\prec y$ whenever $x,y\in Z^\ast$ and $\pi(x)<\pi(y)$, where $<$ denotes the usual lexicographic order on $Z$.
\end{itemize}
Consider the order topology induced by $\prec$ on $Z^\ast$. Since $D\cap Q=\varnothing$, it is easy to check that $Z^\ast$ is a compact crowded zero-dimensional space with no isolated points. Therefore $Z^\ast\approx 2^\omega$. Furthermore, it is easy to check that $\pi$ is a continuous surjection. Let $X^\ast$ be the subspace of $Z^\ast$ whose underlying set is $X$, and notice that $X^\ast$ is saturated with respect to $\pi$. Observe that $\pi\upharpoonright X^\ast:X^\ast\longrightarrow X$ is a closed continuous surjection by Lemma \ref{pi}. Since it is also injective (in fact, it is the identity), it follows that $X^\ast\approx X$.
Let $Y^\ast= Z^\ast\setminus X^\ast$, and notice that $Y^\ast\supseteq D^\ast$ is dense in $Z^\ast$. Assume that $f^\ast:Y^\ast\longrightarrow Y^\ast$ is a homeomorphism. Let $B^\ast=\bigcup_{k\in\mathbb{Z}}(f^\ast)^k[D^\ast]$ and $B=\pi[B^\ast]$. Let $W=Y^\ast\setminus B^\ast=Y\setminus B\subseteq Z$, and notice that $f=f^\ast\upharpoonright W:W\longrightarrow W$ is a homeomorphism. By Lemma \ref{lav}, there exists a $\mathsf{G_\delta}$ subspace $T$ of $Z$ and a homeomorphism $g:T\longrightarrow T$ such that $f\subseteq g$. Notice that $g\in\mathcal{G}$. Furthermore, since $T'=T\setminus\bigcup_{k\in\mathbb{Z}}g^k[B]\supseteq W$ is still a $\mathsf{G_\delta}$ and $g'=g\upharpoonright T':T'\longrightarrow T'$ is still a homeomorphism, we can assume without loss of generality that $T\cap B=\varnothing$.
First assume that $|\{x\in T:g(x)\notin\mathcal{S}(x)\}|<\mathfrak{c}$ for some subgroup $\mathcal{S}$ of $\mathcal{H}$ such that $|\mathcal{S}|<\mathsf{add(meager)}$, and fix such a subgroup. Notice that $T$ is a crowded Polish space because it is a dense $\mathsf{G_\delta}$ subset of $Z$. Let $T_h=\{x\in T:g(x)=h(x)\}$ for $h\in\mathcal{S}$, and notice that each $T_h$ is closed in $T$. Since $|\mathcal{S}|<\mathsf{add(meager)}$, it follows that at least one $T_h$ has non-empty interior in $T$. Assume, in order to get a contradiction, that $T_h$ has non-empty interior for some $h\in\mathcal{S}\setminus\{\mathsf{id}\}$, and fix such an $h$. Fix $a,b\in Z$ such that $a<b$ and $\varnothing\neq (a,b)\cap T\subseteq T_h$.
Fix $d\in D\cap (a,b)$. Since $D\cap Q=\varnothing$ and $Y\cap U$ is uncountable for every non-empty open subset $U$ of $Z$, it is possible to fix a sequence $\langle a_n:n\in\omega\rangle$ consisting of elements of $(a,d)\cap W=(a,d)\cap (Y\setminus B)$ and a sequence $\langle b_n:n\in\omega\rangle$ consisting of elements of $(d,b)\cap W=(d,b)\cap (Y\setminus B)$ such that $a_n\to d$ and $b_n\to d$ in $Z$. Observe that $a_n\to d^-$ and $b_n\to d^+$ in $Y^\ast$. In particular, since $f^\ast$ is a homeomorphism, the sequences $\langle f^\ast(a_n):n\in\omega\rangle$ and $\langle f^\ast(b_n):n\in\omega\rangle$ should converge to different limits in $Y^\ast$, hence in $Z^\ast$.
Notice that $h(a_n)\to h(d)$ and $h(b_n)\to h(d)$ in $Z$ by the continuity of $h$. Since $h\neq\mathsf{id}$, condition (\ref{pushaway}) guarantees that $h(d)\notin D$. Furthermore, the fact that each $a_n\in (a,b)\cap W\subseteq (a,b)\cap T\subseteq T_h$ implies that
$h(a_n)=g(a_n)=f(a_n)=f^\ast(a_n)\notin D$. Therefore $f^\ast(a_n)=h(a_n)\to h(d)$ in $Z^\ast$ as well. Using a similar argument, one sees that $f^\ast(b_n)=h(b_n)\to h(d)$ in $Z^\ast$, which is a contradiction. In conclusion, $T_h$ has non-empty interior if and only if $h=\mathsf{id}$. As one can easily check, this implies that $T_\mathsf{id}$ is dense in $T$. Therefore, the function $g$ is the identity on $T$, which implies that $f^\ast$ is the identity on $Y^\ast$. This shows that $Y^\ast$ is rigid.
Now assume that $|\{x\in T:g(x)\notin\mathcal{S}(x)\}|=\mathfrak{c}$ for every subgroup $\mathcal{S}$ of $\mathcal{H}$ such that $|\mathcal{S}|<\mathsf{add(meager)}$. Then, by condition (\ref{kill}), we can fix $z\in T\cap X$ such that $g(z)\notin X$. Since $T\cap B=\varnothing$, it is clear that $T=\mathsf{dom}(g)=\mathsf{ran}(g)$ is the disjoint union of $T\cap X$ and $W$. It follows that $g(z)\in W$. This is a contradiction, because it implies that $z=g^{-1}(g(z))=f^{-1}(g(z))\in W$.
\end{proof}
\section{The homogeneous example}
\begin{theorem}\label{homogeneous}
There exists a homogeneous subspace of $2^\omega$ whose complement is dense in $2^\omega$ and rigid.
\end{theorem}
\begin{proof}
Our plan is to apply Theorem \ref{general} with $\mathcal{H}=\mathcal{V}$. Let $\mathcal{G}=\{g_\alpha:\alpha\in\mathfrak{c}\}$ be an enumeration. Using the fact that $\mathcal{V}$ is countable, it is easy to construct a countable dense subset $D$ of $2^\omega$ such that $D\cap Q=\varnothing$ and $h[D]\cap D=\varnothing$ for every $h\in\mathcal{V}\setminus\{\mathsf{id}\}$.
Using transfinite recursion, we will construct increasing sequences $\langle X_\alpha:\alpha\in\mathfrak{c}\rangle$ and $\langle Y_\alpha:\alpha\in\mathfrak{c}\rangle$ consisting of subsets of $2^\omega$ so that the following conditions are satisfied for every $\alpha\in\mathfrak{c}$.
\begin{enumerate}
\item[(I)] $|X_\alpha|<\mathfrak{c}$ and $|Y_\alpha|<\mathfrak{c}$.
\item[(II)] $X_\alpha\cap Y_\alpha=\varnothing$.
\item[(III)] $\mathcal{V}[X_\alpha]=X_\alpha$.
\item[(IV)] If $|\{x\in\mathsf{dom}(g_\alpha):g_\alpha(x)\notin\mathcal{V}(x)\}|=\mathfrak{c}$ then there exists $z\in\mathsf{dom}(g_\alpha)\cap X_{\alpha+1}$ such that $g(z)\in Y_{\alpha+1}$.
\end{enumerate}
Start by setting $X_0=\varnothing$ and $Y_0=D$. Take unions at limit stages. At a successor stage $\alpha+1$, suppose that $X_\alpha$ and $Y_\alpha$ have already been constructed. Assume that $|\{x\in\mathsf{dom}(g_\alpha):g_\alpha(x)\notin\mathcal{V}(x)\}|=\mathfrak{c}$. Then, it is possible to fix
$$
z\in\mathsf{dom}(g_\alpha)\setminus(\mathcal{V}[Y_\alpha]\cup g_\alpha^{-1}[X_\alpha])
$$
such that $g_\alpha(z)\notin\mathcal{V}(z)$. Set $X_{\alpha+1}=\mathcal{V}[X_\alpha\cup\{z\}]$ and $Y_{\alpha+1}=Y_{\alpha}\cup\{g_\alpha(z)\}$. Conclude the construction by setting $X=\bigcup_{\alpha\in\mathfrak{c}}X_\alpha$.
Since $\mathcal{V}[X]=X$ by condition (III), it follows from Lemma \ref{jan} that $X$ is homogeneous. It is clear that conditions (\ref{preserve}) and (\ref{pushaway}) of Theorem \ref{general} are satisfied. To see that condition (\ref{kill}) holds, assume that $g\in\mathcal{G}$ and $|\{x\in\mathsf{dom}(g):g(x)\notin\mathcal{S}(x)\}|=\mathfrak{c}$ for every subgroup $\mathcal{S}$ of $\mathcal{V}$ such that $|\mathcal{S}|<\mathsf{add(meager)}$. Since $\mathcal{V}$ is countable, this trivially implies that $|\{x\in\mathsf{dom}(g_\alpha):g_\alpha(x)\notin\mathcal{V}(x)\}|=\mathfrak{c}$, where $\alpha\in\mathfrak{c}$ is such that $g_\alpha=g$. It follows from conditions (IV) and (II) that there exists $z\in\mathsf{dom}(g)\cap X$ such that $g(z)\notin X$.
\end{proof}
\begin{corollary}\label{hnotrh}
There exists a homogeneous subspace of $2^\omega$ that is not relatively homogeneous.
\end{corollary}
\section{The countable dense homogeneous example}
\begin{theorem}\label{cdh}
Assume $\mathsf{MA(\sigma\textrm{-}centered)}$. Then there exists a countable dense homogeneous subspace of $2^\omega$ whose complement is dense in $2^\omega$ and rigid.
\end{theorem}
\begin{proof}
Once again, we plan to apply Theorem \ref{general}. Let $\mathcal{G}=\{g_\alpha:\alpha\in\mathfrak{c}\}$ be an enumeration. Enumerate as $\{(A_\alpha,B_\alpha):\alpha\in\mathfrak{c}\}$ all pairs of countable dense subsets of $2^\omega$, making sure to list each pair cofinally often. Fix a countable dense subset $D$ of $2^\omega$ such that $D\cap Q=\varnothing$.
Using transfinite recursion, we will construct an increasing sequence $\langle \mathcal{H}_\alpha:\alpha\in\mathfrak{c}\rangle$ consisting of subgroups of $\mathcal{H}(2^\omega)$, together with increasing sequences $\langle X_\alpha:\alpha\in\mathfrak{c}\rangle$ and $\langle Y_\alpha:\alpha\in\mathfrak{c}\rangle$ consisting of subsets of $2^\omega$, so that the following conditions are satisfied for every $\alpha\in\mathfrak{c}$.
\begin{enumerate}
\item[(I)] $|X_\alpha|<\mathfrak{c}$ and $|Y_\alpha|<\mathfrak{c}$.
\item[(II)] $X_\alpha\cap Y_\alpha=\varnothing$.
\item[(III)] $|\mathcal{H}_\alpha|<\mathfrak{c}$.
\item[(IV)] $\mathcal{H}_\alpha[X_\alpha]=X_\alpha$.
\item[(V)] If $h\in\mathcal{H}_\alpha\setminus\{\mathsf{id}\}$ then $h[D]\cap D=\varnothing$.
\item[(VI)] If $|\{x\in\mathsf{dom}(g_\alpha):g_\alpha(x)\notin\mathcal{H}_\alpha(x)\}|=\mathfrak{c}$ then there exists $z\in\mathsf{dom}(g_\alpha)\cap X_{\alpha+1}$ such that $g(z)\in Y_{\alpha+1}$.
\item[(VII)] If $A_\alpha\cup B_\alpha\subseteq X_\alpha$ then there exists $f\in\mathcal{H}_{\alpha+1}$ such that $f[A_\alpha]=B_\alpha$.
\end{enumerate}
Start by setting $\mathcal{H}_0=\{\mathsf{id}\}$, $X_0=\varnothing$ and $Y_0=D$. Take unions at limit stages. At a successor stage $\alpha+1$, suppose that $\mathcal{H}_\alpha$, $X_\alpha$ and $Y_\alpha$ have already been constructed. Assume that $|\{x\in\mathsf{dom}(g_\alpha):g_\alpha(x)\notin\mathcal{H}_\alpha(x)\}|=\mathfrak{c}$. Then, it is possible to fix
$$
z\in\mathsf{dom}(g_\alpha)\setminus (\mathcal{H}_\alpha[Y_\alpha]\cup g_\alpha^{-1}[X_\alpha])
$$
such that $g_\alpha(z)\notin\mathcal{H}_\alpha(z)$. First we will construct $\mathcal{H}_{\alpha+1}$. If $A_\alpha\cup B_\alpha\nsubseteq X_\alpha$, set $\mathcal{H}_{\alpha+1}=\mathcal{H}_\alpha$. If $A_\alpha\cup B_\alpha\subseteq X_\alpha$, let $f\in\mathcal{H}(2^\omega)$ be obtained by applying Lemma \ref{keylemma} with $\mathcal{H}=\mathcal{H}_\alpha$, $X=X_\alpha\cup\{z\}$, $Y=Y_\alpha\cup\{g_\alpha(z)\}$, $A=A_\alpha$ and $B=B_\alpha$, then set $\mathcal{H}_{\alpha+1}=\langle\mathcal{H}_\alpha\cup\{f\}\rangle$. Finally, set $X_{\alpha+1}=\mathcal{H}_{\alpha+1}[X_\alpha\cup\{z\}]$ and $Y_{\alpha+1}=Y_\alpha\cup\{g_\alpha(z)\}$. Conclude the construction by setting $X=\bigcup_{\alpha\in\mathfrak{c}}X_\alpha$ and $\mathcal{H}=\bigcup_{\alpha\in\mathfrak{c}}\mathcal{H}_\alpha$.
Notice that $\mathcal{H}[X]=X$ by condition (IV). Using condition (VII), it is straightforward to check that $X$ is countable dense homogeneous. It is clear that conditions (\ref{preserve}) and (\ref{pushaway}) of Theorem \ref{general} are satisfied. To see that condition (\ref{kill}) holds, assume that $g\in\mathcal{G}$ and $|\{x\in\mathsf{dom}(g):g(x)\notin\mathcal{S}(x)\}|=\mathfrak{c}$ for every subgroup $\mathcal{S}$ of $\mathcal{H}$ such that $|\mathcal{S}|<\mathsf{add(meager)}$. Fix $\alpha\in\mathfrak{c}$ such that $g_\alpha=g$, and notice that $|\{x\in\mathsf{dom}(g_\alpha):g_\alpha(x)\notin\mathcal{H}_\alpha(x)\}|=\mathfrak{c}$ because $|\mathcal{H}_\alpha|<\mathfrak{c}=\mathsf{add(meager)}$ by condition (III) and $\mathsf{MA(\sigma\textrm{-}centered)}$. It follows from conditions (VI) and (II) that there exists $z\in\mathsf{dom}(g)\cap X$ such that $g(z)\notin X$.
\end{proof}
\begin{lemma}\label{keylemma}
Assume $\mathsf{MA(\sigma\textrm{-}centered)}$. Furthermore, assume that the following requirements are satisfied.
\begin{itemize}
\item $X$ and $Y$ are subsets of $2^\omega$ of size less than $\mathfrak{c}$.
\item $\mathcal{H}$ is a subgroup of $\mathcal{H}(2^\omega)$ of size less than $\mathfrak{c}$.
\item $\mathcal{H}[X]\cap Y=\varnothing$.
\item $D$ is a countable dense subset of $2^\omega$ such that $D\subseteq Y$.
\item $h[D]\cap D=\varnothing$ for every $h\in\mathcal{H}\setminus\{\mathsf{id}\}$.
\item $A$ and $B$ are countable dense subsets of $2^\omega$ such that $A\cup B\subseteq X$.
\end{itemize}
\newpage
\noindent Then there exists $f\in\mathcal{H}(2^\omega)$ such that the following conditions hold.
\begin{enumerate}
\item $f[A]=B$.
\item $\langle\mathcal{H}\cup\{f\}\rangle[X]\cap Y=\varnothing$.
\item $h[D]\cap D=\varnothing$ for every $h\in\langle\mathcal{H}\cup\{f\}\rangle\setminus\{\mathsf{id}\}$.
\end{enumerate}
\end{lemma}
\begin{proof}
Let $\lambda=\max\{|X|,|\mathcal{H}|,\omega\}$, and notice that $\lambda<\mathfrak{c}$. Start by constructing $X^\ast\supseteq X$ so that $X^\ast\setminus (A\cup B)$ is $\lambda$-dense in $2^\omega$ and $\mathcal{H}[X^\ast]\cap Y=\varnothing$. Notice that $\mathcal{H}[X^\ast]\setminus (A\cup B)$ will still be $\lambda$-dense in $2^\omega$. We claim that any $f\in\mathcal{H}(2^\omega)$ that satisfies the following conditions will also satisfy conditions (1) and (2).
\begin{enumerate}
\item[(I)] $f[A]=B$.
\item[(II)] $f[\mathcal{H}[X^\ast]\setminus A]=f[\mathcal{H}[X^\ast]\setminus B]$.
\end{enumerate}
In fact, conditions (I) and (II) immediately imply that $f[\mathcal{H}[X^\ast]]=\mathcal{H}[X^\ast]$, and therefore
$$
\langle\mathcal{H}\cup\{f\}\rangle[X]\subseteq\langle\mathcal{H}\cup\{f\}\rangle[\mathcal{H}[X^\ast]]=\mathcal{H}[X^\ast],
$$
which is disjoint from $Y$ by construction.
Let $\mathcal{H}^\ast=\mathcal{H}\setminus\{\mathsf{id}\}$. Define $T_h=\{x\in 2^\omega:h(x)=x\}$ for $h\in\mathcal{H}$, and observe that each $T_h$ is closed. Furthermore, if $h\in\mathcal{H}^\ast$ then $T_h$ is nowhere dense, since $D\subseteq 2^\omega\setminus T_h$. Define
$$
C=\{x\in 2^\omega:h(x)\neq x\text{ for every }h\in\mathcal{H}^\ast\}.
$$
Since $|\mathcal{H}^\ast|<\mathfrak{c}=\mathsf{add(meager)}$ by $\mathsf{MA(\sigma\textrm{-}centered)}$, one sees that $C=2^\omega\setminus\bigcup_{h\in\mathcal{H}^\ast}T_h$ is comeager in $2^\omega$. In particular, $C$ is $\mathfrak{c}$-dense in $2^\omega$, so it is possible to fix a collection $\mathcal{D}$ of size $\lambda$ consisting of countable dense subsets of $2^\omega$ such that the following conditions hold.
\begin{itemize}
\item $\bigcup\mathcal{D}\subseteq C\setminus\mathcal{H}[X^\ast\cup Y]$.
\item $E\cap F=\varnothing$ whenever $E,F\in\mathcal{D}$ and $E\neq F$.
\item $\mathcal{H}(x)\cap\mathcal{H}(y)=\varnothing$ whenever $x,y\in\bigcup\mathcal{D}$ and $x\neq y$.
\end{itemize}
Given arbitrary $E,F\in\mathcal{D}$ and $h,g\in\mathcal{H}$, one can easily verify that $h[E]\cap g[F]=\varnothing$ unless $E=F$ and $h=g$.
Let $\mathcal{T}$ be the set of all $s\in (\mathcal{H}^\ast\cup\{-1,1\})^{<\omega}$ such that the following conditions hold whenever $\{k,k+1\}\subseteq\mathsf{dom}(s)$. For notational convenience, we will always assume that $s=\langle s_{n-1},\ldots,s_0\rangle$, where $n=\mathsf{dom}(s)$.
\begin{itemize}
\item If $s_k\in\mathcal{H}^\ast$ then $s_{k+1}\in\{-1,1\}$.
\item If $s_k\in\{-1,1\}$ then $s_{k+1}\in\{s_k\}\cup\mathcal{H}^\ast$.
\end{itemize}
Notice that $|\mathcal{T}|\leq\lambda<\mathfrak{c}$. Suppose that some $f\in\mathcal{H}(2^\omega)$ has been chosen. Given $s\in\mathcal{T}$ such that $\mathsf{dom}(s)=n$, we will say that $w\in\langle\mathcal{H}\cup\{f\}\rangle$ is of \emph{type} $s$ if it can be written as
$$
w=h_{n-1}\cdots h_1h_0,
$$
where $h_k=s_k$ if $s_k\in\mathcal{H}^\ast$ and $h_k=f^{s_k}$ if $s_k\in\{-1,1\}$. In particular, $\mathsf{id}$ is the only element of type $\varnothing$.
Define $D_s$ for $s\in\mathcal{T}$ so that the following conditions hold.
\begin{itemize}
\item $D_\varnothing=D$.
\item $D_{\langle h\rangle^\frown s}=h[D_s]$ whenever $h\in\mathcal{H}^\ast$ and $\langle h\rangle^\frown s\in\mathcal{T}$.
\item $\mathcal{D}=\{D_{\langle 1\rangle^\frown s}:s\in\mathcal{T}\}\cup\{D_{\langle -1\rangle^\frown s}:s\in\mathcal{T}\}$ is an injective enumeration.
\end{itemize}
Using the properties of $\mathcal{D}$, it is straightforward to verify that $D_s\cap D_t=\varnothing$ whenever $s,t\in\mathcal{T}$ and $s\neq t$.
Therefore, by Lemma \ref{bb}, we can fix $f\in\mathcal{H}(2^\omega)$ that satisfies the following conditions, together with conditions (I) and (II).
\begin{enumerate}
\item[(III)] $f[D_s]=D_{\langle 1\rangle^\frown s}$ whenever $\langle 1\rangle^\frown s\in\mathcal{T}$.
\item[(IV)] $f[D_{\langle -1\rangle^\frown s}]=D_s$ whenever $\langle -1\rangle^\frown s\in\mathcal{T}$.
\end{enumerate}
In order to see that condition (3) holds, we will use induction on $n=\mathsf{dom}(s)$ to prove that $w[D]=D_s$ whenever $w\in\langle\mathcal{H}\cup\{f\}\rangle$ is of type $s$. The case $n=0$ is trivial, so assume that $n>0$. First assume that $s=\langle h\rangle^\frown s'$ for some $h\in\mathcal{H}^\ast$ and $s'\in\mathcal{T}$. This means that $w=hw'$ for some $w'$ of type $s'$. Using the inductive assumption $w'[D]=D_{s'}$, one sees that
$$
w[D]=h[w'[D]]=h[D_{s'}]=D_{\langle h\rangle^\frown s'}=D_s.
$$
Now assume that $s=\langle 1\rangle^\frown s'$ for some $s'\in\mathcal{T}$. This means that $w=fw'$ for some $w'$ of type $s'$. Using condition (III) and the inductive assumption $w'[D]=D_{s'}$, one sees that
$$
w[D]=f[w'[D]]=f[D_{s'}]=D_{\langle 1\rangle^\frown s'}=D_s.
$$
The case in which $s=\langle -1\rangle^\frown s'$ for some $s'\in\mathcal{T}$ can be dealt with similarly, using the fact that $f^{-1}[D_{s'}]=D_{\langle -1\rangle^\frown s'}$ by condition (IV).
\end{proof}
\begin{corollary}\label{cdhnotrcdh}
Assume $\mathsf{MA(\sigma\textrm{-}centered)}$. Then there exists a countable dense homogeneous subspace of $2^\omega$ that is not relatively countable dense homogeneous.
\end{corollary}
\begin{question}
Is it possible to prove in $\mathsf{ZFC}$ that there exists a countable dense homogeneous subspace of $2^\omega$ whose complement is dense in $2^\omega$ and rigid?
\end{question}
|
\section{Introduction}
In a previous paper \cite{BC2014a} we identified certain new properties of
the $N$ zeros of the polynomials of order $N$ belonging to the Askey scheme.
The main one of these properties identifies an $N\times N$
matrix---explicitly defined in terms of these $N$ zeros and of the
parameters characterizing the polynomial under consideration---which
features $N$ eigenvalues given by neat, explicit formulas, having moreover a
Diophantine connotation when the parameters of the polynomial are suitably
restricted. In this paper we identify somewhat analogous properties of the
N $ zeros of the polynomials of order $N$ belonging to the \textit{q}-Askey
scheme. Above and hereafter $N$ is an \textit{arbitrary positive integer},
and $q$ an arbitrary number (possibly even \textit{complex}; of course the
results reported in this paper reproduce--via appropriate developments---the
results of \cite{BC2014a} in the $q\rightarrow 1$ limit).
Let us recall \cite{BC2014a} that \textquotedblleft the properties of the
zeros of polynomials are a core problem of mathematics to which, over time,
an immense number of investigations have been devoted. Nevertheless new
findings in this area continue to emerge, see, for instance, \cit
{BCD1, BCD2, BCD3, BCD4,BCD5, CI2013,IR2013,BC2013,BC2014,CY1, CY2,BCY,C1,C2,C3},'' and see also the very recent
paper \cite{AT2014}.
The technique used to obtain the results reported below is somewhat
analogous to that employed in our previous paper \cite{BC2014a}, but there
are significant differences, due to the fact that the main tool of our
treatment are now Differential \textit{q}-Difference Equations (D\textit{q
DEs) instead of Differential Difference Equations (DDEs). Hence our
treatment below is patterned after that of \cite{BC2014a}; yet a previous
reading of \cite{BC2014a} is by no means mandatory to understand what
follows.
The main findings of this paper are reported in the following Section 2.
They detail properties of the zeros of the Askey-Wilson and \textit{q}-Racah
polynomials, which are the two ``highest'' classes of polynomials belonging
to the \textit{q}-Askey scheme \cite{KS}---so that these polynomials feature
$5$ arbitrary parameters (including $q\neq 1$), in addition to their degree
N$. Analogous properties can of course be obtained, from the results
reported below, for the zeros of the (variously ``named'' \cite{KS})
polynomials belonging to ``lower'' classes of the \textit{q}-Askey scheme,
via the reductions---corresponding to special assignments of the
parameters---that characterize the \textit{q}-Askey scheme \cite{KS}; we
leave this task to the interested reader.
Our findings are proven in Section 3. The definitions and some standard
properties of the Askey-Wilson and \textit{q}-Racah polynomials are reported
in the Appendix, for the convenience of the reader and also to specify our
notation; the reader is advised to glance through this Appendix before
reading the next section, and then to return to it whenever appropriate.
The results that follow are only a consequence of the \textit{explicit}
\textit{definitions} of the Askey-Wilson and \textit{q}-Racah polynomials
and of the \textit{q}-\textit{difference equations} they satisfy (see the
Appendix); the orthogonality properties that these polynomials also satisfy
play no role, so that the results reported below do not require the
restrictions on the parameters and arguments of these polynomials that are
instead mandatory for the validity of these orthogonality properties and of
other related properties \cite{KS}.
\section{Results}
To formulate our results we refer to the definitions and standard properties
of the Askey-Wilson and \textit{q}-Racah polynomials as reported in the
Appendix, to which the reader should also refer for the notation employed
hereafter. As indicated below, some of these results are immediate
consequences of known properties of the Askey-Wilson and \textit{q}-Racah
polynomials; our main results are proven in the following Section 3.
\subsection{Results for the zeros of the Askey-Wilson polynomials}
\textbf{Notation 2.1}. In this Section 2.1---as in the Askey-Wilson parts of
the Appendix and of the next Section 3---we often use the change of
variables
\begin{subequations}
\begin{equation}
z=x+\sqrt{x^{2}-1}~,~~~x=\frac{z^{2}+1}{2~z}~, \label{xz}
\end{equation
with $x$ being the argument of the Askey-Wilson polynomial $p_{N}\left(
x\right) \equiv p_{N}\left( a,b,c,d;q;x\right) $ and $z$ being the argument
of the corresponding rational function $P_{N}\left( z\right) \equiv
P_{N}\left( a,b,c,d;q;z\right) $, see (\ref{PpAW}). These relations
correspond to the assignment $x=\cos \theta $, $z=\exp \left(
\mathbf{i}\theta \right) $ (here and hereafter $\mathbf{i}$ denotes the
imaginary unit, $\mathbf{i}^{2}=-1)$; but of course in this paper the
argument $x$ \ of the Askey-Wilson polynomial is a \textit{complex variable
, not restricted to the interval $[-1,1]$ of the real line. And in
all our final formulas only even powers of the square-root $\sqrt{x^{2}-1}$
appear, so the determination of the square-root in the first formula (\re
{xz})---and in analogous formulas, see below---is not an issue. Likewise,
the $N$ zeros of the Askey-Wilson polynomial $p_{N}\left( x\right) $ are
denoted as $\bar{x}_{n},~p_{N}\left( \bar{x}_{n}\right) =0$, and the $N$
quantitie
\begin{equation}
\bar{z}_{n}=\bar{x}_{n}+\sqrt{\bar{x}_{n}^{2}-1}~,~~~\bar{x}_{n}=\frac{\bar{
}_{n}^{2}+1}{2~\bar{z}_{n}}~, \label{xznbar}
\end{equation
are $N$ zeros of the rational function $P_{N}\left( z\right) $, $P_{N}\left(
\bar{z}_{n}\right) =0$. $\square $
\textbf{Proposition 2.1}. Let $\bar{x}_{n}\equiv \bar{x}_{n}\left(
a,b,c,d;q;N\right) $ (with $n=1,2,...,N)$ be the $N$ zeros of the
Askey-Wilson polynomial $p_{N}\left( x\right) \equiv p_{N}\left(
a,b,c,d;q;x\right) $ of degree $N$ in $x$ (with $N$ an arbitrary positive
integer), so that (up to an irrelevant multiplicative constant $C_{N}
\textbf{) }$p_{N}\left( x\right) =C_{N}~\prod\nolimits_{n=1}^{N}\left( x
\bar{x}_{n}\right) $ ; and let $\bar{z}_{n}$ be related to $\bar{x}_{n}$by
\ref{xznbar}).
Then there hold the $N$\ algebraic equations
\end{subequations}
\begin{subequations}
\begin{equation}
A\left( \bar{z}_{n}\right) ~p_{N}\left( \frac{q^{2}~\bar{z}_{n}^{2}+1}{2~q
\bar{z}_{n}}\right) +A\left( \frac{1}{\bar{z}_{n}}\right) ~p_{N}\left( \frac
\bar{z}_{n}^{2}+q^{2}}{2~q~\bar{z}_{n}}\right) =0~,n=1,\ldots ,N~,
\label{EqA21a}
\end{equation
or, equivalently
\begin{equation}
A\left( \bar{z}_{n}\right) ~\prod\limits_{m=1}^{N}\left( q\bar{z}_{n}+\frac{
}{q\bar{z}_{n}}-\bar{z}_{m}-\frac{1}{\bar{z}_{m}}\right) +\left[ \left( \bar
z}_{s}\rightarrow \frac{1}{\bar{z}_{s}}\right) \right] =0~,~n=1,\ldots ,N~,
\label{EqA21b}
\end{equation
wher
\begin{equation}
A\left( z\right) \equiv A\left( a,b,c,d;q;z\right) =\frac{\left( 1-az\right)
~\left( 1-bz\right) ~\left( 1-cz\right) ~\left( 1-dz\right) }{\left(
1-z^{2}\right) ~\left( 1-qz^{2}\right) } \label{AA}
\end{equation
and the symbol $+\left[ \left( \bar{z}_{s}\rightarrow \frac{1}{\bar{z}_{s}
\right) \right] $ denotes the addition of everything that comes before it,
with $\bar{z}_{s}$ replaced by $\frac{1}{\bar{z}_{s}}$ for all $s=1,\ldots
,N $. $\square $
\textbf{Remark 2.1}\textit{.} It is evident that after the substitution
\bar{x}_{n}=\cos \bar{\theta}_{n}$ and $\bar{z}_{n}=\exp \left( \mathbf{i
\bar{\theta}_{n}\right) $, see~(\ref{xznbar}), the left-hand side of (\re
{EqA21b}) becomes an \textit{even} function of each $\bar{\theta}_{s}$,
s=1,\ldots ,N$, hence a function of $\bar{x}_{1},\ldots ,\bar{x}_{N}$. This
is why after the substitution $\bar{z}_{n}=\bar{x}_{n}+\sqrt{\bar{x
_{n}^{2}-1}$ in the left-hand sides of~(\ref{EqA21a}) and~(\ref{EqA21b}) all
the square roots disappear, regardless of the determination of each square
root~$\sqrt{\bar{x}_{n}^{2}-1}$, as long as it is consistent, for each $n$,
throughout the treatment. $\square $
This result is proved in the Appendix, see (\ref{Prop21}).
The following proposition is our main result for the $N$ zeros $\bar{x}_{n}$
of the Askey-Wilson polynomials.
\textbf{Proposition 2.2}. Let $\bar{x}_{n}\equiv \bar{x}_{n}\left(
a,b,c,d;q;N\right) $ (with $n=1,2,...,N)$ be the $N$ zeros of the
Askey-Wilson polynomial $p_{N}\left( x\right) \equiv p_{N}\left(
a,b,c,d;q;x\right) $ of degree $N$ in $x$ (with $N$ an arbitrary positive
integer), so that (up to an irrelevant multiplicative constant $C_{N}$)
p_{N}\left( x\right) =C_{N}~\prod\nolimits_{n=1}^{N}\left( x-\bar{x
_{n}\right) $ ; and let $\bar{z}_{n}$ be related to $\bar{x}_{n}$ by (\re
{xznbar}). Define the $N\times N$ matrix $\underline{M}\equiv \underline{M
\left( a,b,c,d;q;N;\underline{\bar{x}}\right) $, componentwise, as follows:
\end{subequations}
\begin{subequations}
\label{M}
\begin{eqnarray}
&&M_{nn}\equiv M_{nn}\left( a,b,c,d,;q;N;\underline{\bar{x}}\right) \notag
\\ \notag
&=&\frac{(q-1)}{2q^{N}}\Bigg\{\Bigg[\frac{2\bar{z}_{n}^{2}}{\bar{z}_{n}^{2}-
}G(\bar{z}_{n})\sum_{m=1,m\neq n}^{N}\Big(-\frac{q}{\bar{z}_{m}-q\bar{z}_{n}
+\frac{q\bar{z}_{m}}{q\bar{z}_{n}\bar{z}_{m}-1} \\
&&+\frac{1}{\bar{z}_{m}-\bar{z}_{n}}-\frac{\bar{z}_{m}}{\bar{z}_{n}\bar{z
_{m}-1}\Big)+\frac{2\bar{z}_{n}^{2}G^{\prime }(\bar{z}_{n})}{\bar{z
_{n}^{2}-1}\Bigg]\prod_{\ell =1,\ell \neq n}^{N}K(\bar{z}_{n},\bar{z}_{\ell
}) \notag \\
&&+\left[ \left( \bar{z}_{s}\rightarrow \frac{1}{\bar{z}_{s}}\right) \right]
\Bigg\}
\end{eqnarray
for $n=1,\ldots ,N,$ and
\begin{eqnarray}
&&M_{nm}\equiv M_{nm}\left( a,b,c,d,;q;N;\underline{\bar{x}}\right) \notag
\\ \notag
&=&\frac{(q-1)}{2q^{N}}\Bigg\{\frac{2\bar{z}_{m}^{2}}{\bar{z}_{m}^{2}-1}G
\bar{z}_{n})\Big[\frac{1}{\bar{z}_{m}-q\bar{z}_{n}}+\frac{q\bar{z}_{n}}{
\bar{z}_{n}\bar{z}_{m}-1} \\
&&-\frac{1}{\bar{z}_{m}-\bar{z}_{n}}-\frac{\bar{z}_{n}}{\bar{z}_{n}\bar{z
_{m}-1}\Big]\prod_{\ell =1,\ell \neq n}^{N}K(\bar{z}_{n},\bar{z}_{\ell })
\left[ \left( \bar{z}_{s}\rightarrow \frac{1}{\bar{z}_{s}}\right) \right]
\Bigg\}
\end{eqnarray
for $n,m=1,...,N~\ $with $n\neq m$. In these formula
\begin{equation}
G(\bar{z}_{n})=A(\bar{z}_{n})~\left( q\bar{z}_{n}-\frac{1}{\bar{z}_{n}
\right) ~, \label{eq:GG}
\end{equation
\begin{equation}
G^{\prime }(z)=\frac{d}{dz}G(z)~,
\end{equation
$A\equiv A\left( a,b,c,d;q;z\right) $ is defined by~(\ref{AA}),
\begin{equation}
K(\bar{z}_{n},\bar{z}_{m})=\frac{(\bar{z}_{m}-q\bar{z}_{n})~(q\bar{z}_{n
\bar{z}_{m}-1)}{(\bar{z}_{m}-\bar{z}_{n})~(\bar{z}_{n}\bar{z}_{m}-1)}
\label{eq:KK}
\end{equation
for $n,m=1,\ldots ,N$ with $n\neq m$, and the symbol $+\left[ \left( \bar{z
_{s}\rightarrow \frac{1}{\bar{z}_{s}}\right) \right] $ indicates addition of
everything that comes before it, within the curly brackets, with $\bar{z}_{s}
$ replaced by $\frac{1}{\bar{z}_{s}}$ for all $s=1,\ldots ,N$.
Then this $N\times N$ matrix $\underline{M}$ has the $N$ eigenvalue
\begin{eqnarray}
\mu _{n} &\equiv &\mu _{n}\left( abcd;q;N\right) =q^{-N}\left(
1-q^{n}\right) \left( 1-abcd~q^{2N-1-n}\right) \notag \\
&=&\left( q^{-N}+abcd~q^{N-1}-q^{n-N}-abcd~q^{N-1-n}\right) ~, \notag \\
n &=&1,2,...,N~.~\square
\end{eqnarray}
It is evident that the components of the matrix $M$ in \textbf{Proposition
2.2} are functions of $\bar{x}_{1},\ldots ,\bar{x}_{N}$, see \textbf{Remark
2.1}.
Some immediate corollaries of \textbf{Proposition 2.2} are worth a mention.
\textbf{Corollary 2.2.1}. If $abcd$ and $q$ are both \textit{rational}
numbers, the $N$ eigenvalues of the $N\times N$ matrix $\underline{M}$ (see
\ref{M})) are all \textit{rational} numbers.~$\square $.
This is a remarkable Diophantine property.
\textbf{Corollary 2.2.2}. The $N\times N$ matrix $\underline{M}$---which
depends of course on the $4$ \textit{a priori} arbitrary parameters $a,b,c,d
, explicitly via $A\left( z\right) ,$ see (\ref{AA}), and implicitly via the
dependence on these $4$ parameters of the $N$ zeros $\bar{x}_{n}\equiv \bar{
}_{n}\left( a,b,c,d;q;N\right) $ of the Askey-Wilson polynomials
p_{N}\left( a,b,c,d;q;x\right) $, see (\ref{M}), is \textit{isospectral}
under any variation of these $4$ parameters which does not change the value
of their product $abcd$.~$\square $
\textbf{Corollary 2.2.3}. Several identities satisfied by the $N$ zeros
\bar{x}_{n}\equiv \bar{x}_{n}\left( a,b,c,d;q;N\right) $ of the Askey-Wilson
polynomial $p_{N}\left( x\right) \equiv p_{N}\left( a,b,c,d;q;x\right) $ are
implied by the following standard consequences of \textbf{Proposition 2.2}:
\end{subequations}
\begin{subequations}
\begin{eqnarray}
\text{trace}\left[ \left( \underline{M}\right) ^{k}\right]
&=&\sum_{n=1}^{N}\left( q^{-N}+abcd~q^{N-1}-q^{n-N}-abcd~q^{N-1-n}\right)
^{k}~, \notag \\
k &=&1,2,3,...~,
\end{eqnarray}
\begin{equation}
\det \left[ \underline{M}\right] =q^{-N^{2}}~\left( q;q\right) _{N}~\left(
abcd~q^{N-1};q\right) _{N}~.
\end{equation
In particular, for $k=1,$ the first of these two formulas reads as follows:
\begin{equation}
\sum_{n=1}^{N}\left( M_{nn}\right) =N\left( q^{-N}+abcd~q^{N-1}\right)
+\left( \frac{1-q^{-N}}{1-q}\right) \left( q+abcd~q^{N-1}\right) ~.~\square
\end{equation}
\end{subequations}
\subsection{Results for the zeros of the \textit{q}-Racah polynomials}
Let $\bar{z}_{n}\equiv \bar{z}_{n}\left( \alpha ,\beta ,\gamma ,\delta
;q;N\right) $, where $n=1,2,...,N$, be the $N$ zeros of the \textit{q}-Racah
polynomial $R_{N}\left( x\right) \equiv R_{N}\left( \alpha ,\beta ,\gamma
,\delta ;q;z\right) $ of degree $N$ in $z$ (with $N$ an arbitrary positive
integer), where the variables $x$ and $z$ are related by formula~(\re
{qRacah}). Thus, up to an irrelevant multiplicative constant $C_{N}$,
R_{N}\left( z\right) =C_{N}~\prod\nolimits_{n=1}^{N}\left( z-\bar{z
_{n}\right) $, see~(\ref{FacRacah}). Let the $2N$ numbers $\bar{z
_{n}^{\left( \pm \right) }$ be defined as follows
\begin{equation}
\bar{z}_{n}^{\left( \pm \right) }=q^{\pm 1}\bar{z}_{n}\pm \left( \frac
1-q^{2}}{2q}\right) \left( \bar{z}_{n}-\sqrt{\bar{z}_{n}^{2}-4\gamma \delta
}\right) ,\;\;n=1,\ldots ,N~, \label{zn+-}
\end{equation
see (\ref{z+-}) (of course with the same determination of the square root as
in (\ref{Z})).
A result concerning the $N$ zeros of the \textit{q}-Racah polynomials of
degree $N$ reads then as follows.
\textbf{Proposition 2.3}. The $N$ zeros $\bar{z}_{n}\equiv \bar{z}_{n}\left(
\alpha ,\beta ,\gamma ,\delta ;q;N\right) $, where $n=1,\ldots ,N$, of the $N
$-th degree $q$-Racah polynomial $R_{N}\left( \alpha ,\beta ,\gamma ,\delta
;q;z\right) $ satisfy the following $N$ algebraic equations:
\begin{subequations}
\label{Prop23}
\begin{equation}
B\left( \bar{z}_{n}\right) ~R_{N}(\bar{z}_{n}^{\left( +\right) })+D\left(
\bar{z}_{n}\right) ~R_{N}(\bar{z}_{n}^{\left( -\right) })=0~,~~~n=1,...,N~,
\label{Prop23a}
\end{equation
or, equivalently
\begin{equation}
B\left( \bar{z}_{n}\right) ~\prod\limits_{m=1}^{N}\left( \bar{z}_{n}^{\left(
+\right) }-\bar{z}_{m}\right) +D\left( \bar{z}_{n}\right)
~\prod\limits_{m=1}^{N}\left( \bar{z}_{n}^{\left( -\right) }-\bar{z
_{m}\right) =0~,~~~n=1,...,N~, \label{Prop23b}
\end{equation
were the numbers $\bar{z_{n}}^{\left( \pm \right) }$ are defined by~(\re
{zn+-}) and the functions $B\left( z\right) $ respectively $D\left( z\right)
$ are defined by (\ref{B}) respectively~(\ref{D},~\ref{Z}). Note that this
result holds independently of which determination is taken for the square
root in the above definition of $\bar{z}_{n}^{\left( \pm \right) }$ and in
\ref{Z}), provided of course it is the same.~$\square $
This \textbf{Proposition 2.3} is proven in the Appendix, see (\ref{48a}).
The following proposition is our main result for the zeros of \textit{q
-Racah polynomials.
\textbf{Proposition 2.4}. Let $\bar{z}_{n}\equiv \bar{z}_{n}\left( \alpha
,\beta ,\gamma ,\delta ;q;N\right) $, where $n=1,2,...,N$, be the $N$ zeros
of the \textit{q}-Racah polynomial $R_{N}\left( \alpha ,\beta ,\gamma
,\delta ;q;z\right) $ of degree $N$ in $z$ (with $N$ an arbitrary positive
integer). Define the $N\times N$ matrix $\underline{L}\equiv \underline{L
\left( \alpha ,\beta ,\gamma ,\delta ;q;N;\underline{\bar{z}}\right) $,
componentwise, as follows:
\end{subequations}
\begin{subequations}
\label{L}
\begin{eqnarray}
&&L_{nn}=\Bigg[B^{\prime }(\bar{z}_{n})(\bar{z}_{n}^{(+)}-\bar{z}_{n})
\notag \\
&&+B(\bar{z}_{n})\left( C^{(+)}(\bar{z}_{n})-1+(\bar{z}_{n}^{(+)}-\bar{z
_{n})\sum_{m=1,m\neq n}^{N}W^{(+)}(\bar{z}_{n},\bar{z}_{m})\right) \Bigg
\prod_{\ell =1,\ell \neq n}^{N}\frac{\bar{z}_{n}^{(+)}-\bar{z}_{\ell }}{\bar
z}_{n}-\bar{z}_{\ell }} \notag \\
&&+\Bigg[D^{\prime }(\bar{z}_{n})(\bar{z}_{n}^{(-)}-\bar{z}_{n}) \notag \\
&&+D(\bar{z}_{n})\left( C^{(-)}(\bar{z}_{n})-1+(\bar{z}_{n}^{(-)}-\bar{z
_{n})\sum_{m=1,m\neq n}^{N}W^{(-)}(\bar{z}_{n},\bar{z}_{m})\right) \Bigg
\prod_{\ell =1,\ell \neq n}^{N}\frac{\bar{z}_{n}^{(-)}-\bar{z}_{\ell }}{\bar
z}_{n}-\bar{z}_{\ell }}, \notag \\
&&n=1,...,N~, \label{Lnn}
\end{eqnarray
\begin{eqnarray}
&&L_{nm}=B(\bar{z}_{n})\left( \frac{\bar{z}_{n}^{(+)}-\bar{z}_{n}}{\bar{z
_{n}-\bar{z}_{m}}\right) ^{2}\prod_{\ell =1\,\ell \neq n,m}^{N}\frac{\bar{z
_{n}^{(+)}-\bar{z}_{\ell }}{\bar{z}_{n}-\bar{z}_{\ell }} \notag \\
&&+D(\bar{z}_{n})\left( \frac{\bar{z}_{n}^{(-)}-\bar{z}_{n}}{\bar{z}_{n}
\bar{z}_{m}}\right) ^{2}\prod_{\ell =1\,\ell \neq n,m}^{N}\frac{\bar{z
_{n}^{(-)}-\bar{z}_{\ell }}{\bar{z}_{n}-\bar{z}_{\ell }
,~n,m=1,...,N,~n\neq m, \label{Lnm}
\end{eqnarray
where, as above, the functions $B\left( z\right) $ respectively $D\left(
z\right) $ are defined by (\ref{B}) respectively~(\ref{D},~\ref{Z}),
B^{\prime }(z)=\frac{d}{dz}B(z)$,
\begin{equation*}
C^{(\pm )}(\bar{z}_{n})=\frac{d\bar{z}_{n}^{(+)}}{d\bar{z}_{n}}=q^{\pm 1}\pm
\frac{1-q^{2}}{2q}\left( 1-\frac{\bar{z}_{n}}{\sqrt{\bar{z}_{n}^{2}-4\gamma
\delta q}}\right) ,
\end{equation*
\begin{equation*}
W^{(\pm )}(\bar{z}_{n},\bar{z}_{m})=\frac{C^{(\pm )}(\bar{z}_{n})(\bar{z
_{n}-\bar{z}_{m})-\bar{z}_{n}^{(\pm )}+\bar{z}_{m}}{(\bar{z}_{n}-\bar{z
_{m})(\bar{z}_{n}^{(\pm )}-\bar{z}_{m})},
\end{equation*
and the numbers $\bar{z}_{n}^{\left( \pm \right) }$ are defined by~(\re
{zn+-}). Then this $N\times N$ matrix $\underline{L}$ has the $N$ eigenvalue
\begin{eqnarray}
\lambda _{n} &\equiv &\lambda _{n}\left( \alpha \beta ;q;N\right)
=q^{-N}\left( 1-q^{m}\right) \left( 1-\alpha \beta ~q^{2N-m+1}\right) ~,
\notag \\
n &=&1,2,...,N~.~\square
\end{eqnarray}
Some immediate corollaries of \textbf{Proposition 2.4} are worth a mention.
\textbf{Corollary 2.4.1}. If $\alpha \beta $ and $q$ are both \textit
rational} numbers, the $N$ eigenvalues of the $N\times N$ matrix $\underline
L}$ (see (\ref{L})) are all \textit{rational} numbers.~$\square $.
This is a remarkable Diophantine property.
\textbf{Corollary 2.4.2}. The $N\times N$ matrix $\underline{L}$---which
depends of course on the $4$ \textit{a priori} arbitrary parameters $\alpha
,\beta ,\gamma ,\delta $, explicitly via $B\left( z\right) $ and $D\left(
z\right) $ (see (\ref{B}) and (\ref{D}) with (\ref{Z})), and implicitly via
the dependence on these $4$ parameters of the $N$ zeros $\bar{z}_{n}\equiv
\bar{z}_{n}\left( \alpha ,\beta ,\gamma ,\delta ;q;N\right) $ of the \textit
q}-Racah polynomial $R_{N}\left( \alpha ,\beta ,\gamma ,\delta ;q;z\right)
: see (\ref{L})---is \textit{isospectral} under any variation of these $4$
parameters which does not change the value of the product $\alpha \beta $.~
\square $
\textbf{Corollary 2.4.3}. Several identities satisfied by the $N$ zeros
\bar{z}_{n}\equiv \bar{z}_{n}\left( \alpha ,\beta ,\gamma ,\delta
;q;N\right) $ of the \textit{q}-Racah polynomial $R_{N}\left( \alpha ,\beta
,\gamma ,\delta ;q;z\right) $ are implied by the following standard
consequences of \textbf{Proposition 2.4}:
\end{subequations}
\begin{subequations}
\begin{equation}
\text{trace}\left[ \left( \underline{L}\right) ^{k}\right]
=q^{-kN}~\sum_{m=1}^{N}\left[ \left( 1-q^{m}\right) \left( 1-\alpha \beta
~q^{2N-m+1}\right) \right] ^{k}~,~~~k=1,2,3,...~,
\end{equation}
\begin{equation}
\det \left[ \underline{L}\right] =q^{-N^{2}}~\left( q;q\right) _{N}~\left(
\alpha \beta q^{N+1};q\right) _{N}~.
\end{equation
In particular, for $k=1,$ the first of these two formulas reads as follows:
\begin{equation}
\sum_{n=1}^{N}\left( L_{nn}\right) =N\left( q^{-N}+\alpha \beta
~q^{N+1}\right) +\frac{q\left( 1-q^{-N}\right) \left( 1+\alpha \beta
~q^{N}\right) }{1-q}~.~\square
\end{equation}
\section{Proof of the main results}
In this Section 3 we prove our main results for the zeros of the
Askey-Wilson and \textit{q}-Racah polynomials.
\subsection{Askey-Wilson polynomials}
Let $\Psi _{N}\left( z;t\right) \equiv \Psi _{N}\left( a,b,c,d;q;z;t\right) $
be a rational function of the (generally \textit{complex}) variable $z$,
which satisfies the Differential-\textit{q-}Difference Equation (D\textit{q
DE)
\end{subequations}
\begin{equation}
\frac{\partial ~\Psi _{N}\left( z;t\right) }{\partial ~t}=\left[ \left(
q^{-N}-1\right) ~\left( 1-abcd~q^{N-1}\right) -Q\right] ~\Psi _{N}\left(
z,t\right) ~, \label{DqDE}
\end{equation
with the ($t$-independent) \textit{q}-differential operator $Q$ acting on
the variable $z$ as defined by (\ref{Q}). And let us assume that this
rational function can be expressed (as confirmed below) by the following
linear superposition with $t$-dependent coefficients $c_{m}\left( t\right) $
of the $(N+1)$ rational functions $P_{N-m}\left( z\right) \equiv P_{N-m}\left(
a,b,c,d;q;z\right) $ defined by (\ref{PNz}), with $m=0,1,2,...,N$:
\begin{equation}
\Psi _{N}\left( z,t\right) =\sum_{m=0}^{N}\left[ c_{m}\left( t\right)
~P_{N-m}\left( z\right) \right] ~.
\end{equation}
It is then plain---see the eigenvalue equation (\ref{Qeigen}), of course
with $N$ replaced by $(N-m)$---that the $t$-evolution of the $(N+1)$
coefficients $c_{m}\left( t\right) $ is characterized by the following
system of ODEs,
\begin{eqnarray}
\dot{c}_{m}\left( t\right) &=&\left[ \left( q^{-N}-1\right) ~\left(
1-abcd~q^{N-1}\right) \right. \notag \\
&&\left. -\left( q^{m-N}-1\right) ~\left( 1-abcd~q^{N-m-1}\right) \right]
~c_{m}\left( t\right) \notag \\
&=&q^{-N}~\left( 1-q^{m}\right) \left( 1-abcd~q^{2N-1-m}\right) ~c_{m}\left(
t\right) ~, \notag \\
m &=&0,1,,2...,N~, \label{Eqcm}
\end{eqnarray
where the superimposed dot denotes of course a $t$-differentiation. Hence
\begin{subequations}
\label{cmt}
\begin{equation}
c_{0}\left( t\right) =c_{0}\left( 0\right) ~,
\end{equation
\begin{equation}
c_{m}\left( t\right) =\exp \left( \mu _{m}~t\right) ~c_{m}\left( 0\right)
~,~~~m=1,2,...,N~,
\end{equation
wher
\begin{eqnarray}
\mu _{m} &\equiv &\mu _{m}\left( abcd;q;N\right) =q^{-N}\left(
1-q^{m}\right) \left( 1-abcd~q^{2N-1-m}\right) ~, \notag \\
m &=&1,2,...,N~,
\end{eqnarray
implyin
\begin{equation}
\Psi \left( z;t\right) =c_{0}\left( 0\right) ~P_{N}\left( z\right)
+\sum_{m=1}^{N}\left[ \exp \left( \mu _{m}~t\right) ~c_{m}\left( 0\right)
~P_{N-m}\left( z\right) \right] ~. \label{Psizt}
\end{equation}
It is now clear (see (\ref{qDE})) that the \textquotedblleft
equilibrium''---i.e., $t$-independent---solution $\bar{\Psi}_{N}\left(
z\right) \equiv \bar{\Psi}_{N}\left( a,b,c,d;q;z\right) $ of D\textit{q}DE
\ref{DqDE}) reads
\end{subequations}
\begin{subequations}
\begin{equation}
\bar{\Psi}_{N}\left( z\right) =c_{0}\left( 0\right) ~P_{N}\left( z\right) ~,
\label{PsiEqui}
\end{equation
corresponding to the \textquotedblleft equilibrium'' (i. e., $t$-independent)
solution of the dynamical system (\ref{Eqcm}) which clearly read
\begin{equation}
\bar{c}_{0}\left( t\right) =\bar{c}_{0}\left( 0\right) ~;~~~\bar{c
_{m}\left( t\right) =0~,m=1,2,...,N~.
\end{equation}
It is at this stage convenient to perform---as in the Appendix---the change
of variables (\ref{xz}) from $z$ to $x$ and viceversa, implying
\end{subequations}
\begin{subequations}
\begin{equation}
\Psi _{N}\left( z;t\right) =\psi _{N}\left( \frac{z^{2}+1}{2~z};t\right)
~,~~~\psi _{N}\left( x;t\right) =\Psi _{N}\left( x+ \sqrt{x^{2}-1};t\right)
~, \label{Psipsi}
\end{equation
so that (\ref{Psizt}) becomes
\begin{equation}
\psi \left( x;t\right) =c_{0}\left( 0\right) ~p_{N}\left( x\right)
+\sum_{m=1}^{N}\left[ \exp \left( \mu _{m}~t\right) ~c_{m}\left( 0\right)
~p_{N-m}\left( x\right) \right] ~, \label{psixt}
\end{equation
where $p_{m}\left( x\right) $ is now the Askey-Wilson polynomial of degree
m $ (see (\ref{AWpolx}) and (\ref{Psipsi})). It is thus seen that $\psi
\left( x;t\right) $ is a ($t$-dependent) polynomial of degree $N$ in the
variable $x $, and the corresponding \textit{equilibrium} solution of (\ref{DqDE}) is, up to
the arbitrary multiplicative constant $c_{0}\left( 0\right) ,$ just the
Askey-Wilson polynomial of degree $N$,
\end{subequations}
\begin{equation}
\bar{\psi}_{N}\left( x\right) =\bar{c}_{0}\left( 0\right) ~p_{N}\left(
x\right)
\end{equation
(see (\ref{PsiEqui})).
Next, let us introduce the $N$ zeros $x_{n}\left( t\right) \equiv
x_{n}\left( a,b,c,d;q;N;t\right) $ of the polynomial $\psi \left( x;t\right)
$ of degree $N$ in $x$, by setting
\begin{equation}
\psi \left( x;t\right) =C_{N}~\prod\limits_{n=1}^{N}\left[ x-x_{n}\left(
t\right) \right] ~. \label{Defxnt}
\end{equation
Here $C_{N}$ is an arbitrary constant that plays no role in the following;
it is of course proportional to $c_{0}\left( 0\right) $, and the computation
(from (\ref{psixt}), (\ref{Defxnt}) and (\ref{AWpolx})) of the
proportionality factor can be left to the very diligent reader.
Let us now investigate the $t$-evolution of the $N$ zeros $x_{n}\left(
t\right) $, as implied by the Differential \textit{q}-Difference Equation
satisfied by $\psi _{N}\left( x;t\right) ,$ which obtains from the
D\textit{q}DE (\ref{DqDE}) satisfied by $\Psi \left( z;t\right) $ via the
change of variables (\ref{Psipsi}):
\begin{equation}
\frac{\partial ~\psi _{N}\left( x;t\right) }{\partial ~t}=\left[ \left(
q^{-N}-1\right) ~\left( 1-abcd~q^{N-1}\right) -Q\right] ~\psi _{N}\left(
x;t\right) ~. \label{DDEpsi}
\end{equation
The action of the operator $Q$ on functions of the variable $x$ is
given by the formula (see (\ref{Q}) and (\ref{Changexz}))
\begin{subequations}
\label{QQx}
\begin{equation}
Q~f\left( x\right) =\left[ A\left( z\right) ~\delta _{q}^{\left( +\right)
}+A\left( z^{-1}\right) ~\delta _{q}^{\left( -\right) }-A\left( z\right)
-A\left( z^{-1}\right) \right] ~f\left( x\right) \label{Qx}
\end{equation
wit
\begin{equation}
\delta _{q}^{\left( \sigma \right) }~f\left( x\right) =f\left( q^{\sigma
}x+\sigma \frac{1-q^{2}}{2qz}\right) ~,~~~\sigma =\pm 1~,
\end{equation
where the variable $z$ in (\ref{Qx}) is related to the variable $x$
via (\ref{xz}). It can be verified that the right-hand side of (\ref{Qx}) is
an even function of $\theta $ (defined by $z=\exp \left( i\theta \right) )$
hence a function of $x$, and so~--- loosely speaking~--- it makes no
difference to set $z=x+\sqrt{x^{2}-1}$ or $z=x-\sqrt{x^{2}-1}$ in this
definition of the two operators $\delta _{q}^{\left( \pm \right) }$,
provided of course the same convention is used throughout (see \textbf
Notation 2.1}).
It is plain from (\ref{Defxnt})---by logarithmic $t$-differentiation---that
\end{subequations}
\begin{eqnarray}
\frac{\partial ~\psi _{N}\left( x;t\right) }{\partial t}&=&-\psi _{N}\left(
x;t\right) ~\sum_{m=1}^{N}\frac{\dot{x}_{m}\left( t\right) }{x-x_{m}\left(
t\right) }\notag\\
&=&-C_{N}~\sum_{m=1}^{N}\left\{ \dot{x}_{m}\left( t\right)
\prod\limits_{\ell =1,~\ell \neq m}^{N}\left[ x-x_{\ell }\left( t\right)
\right] \right\} ~.
\end{eqnarray
Henc
\begin{equation}
\left. \frac{\partial ~\psi _{N}\left( x;t\right) }{\partial t}\right\vert
_{x=x_{n}\left( t\right) }=-C_{N}~\left\{ \dot{x}_{n}\left( t\right)
~\prod\limits_{\ell =1,~\ell \neq n}^{N}\left[ x_{n}\left( t\right) -x_{\ell
}\left( t\right) \right] \right\} ~,
\end{equation
and by setting $x=x_{n}\left( t\right) $ in (\ref{DDEpsi}) we get (via (\re
{QQx}) and (\ref{Changexz})---and (\ref{Defxnt}) implying of course $\psi
_{N}\left( x_{n}\left( t\right) ;t\right) =0$
\begin{eqnarray}
\dot{x}_{n} &=&\frac{(q-1)}{2q^{N}}\Bigg\{G(z_{n})\prod_{\ell =1,\ell \neq
n}^{N}K(z_{n},z_{\ell })+G\left( \frac{1}{z_{n}}\right) \prod_{\ell =1,\ell
\neq n}^{N}K\left( \frac{1}{z_{n}},\frac{1}{z_{\ell }}\right) \Bigg\}~,
\notag \label{xndot} \\
&\equiv &F_{n}(z_{1},\ldots ,z_{N})\equiv \hat{F}_{n}(x_{1},\ldots
,x_{N})~,~~~n=1,\ldots ,N~,
\end{eqnarray
where the functions $G$ and $K$ are defined by~(\ref{eq:GG}) and~(\ref{eq:KK
), respectively. Note that, for notational simplicity, we omitted to
indicate the $t$-dependence of $\dot{x}_{n}$, $x_{n}$, $x_{\ell }$, $z_{n}$.
Again, via $z_{s}=\exp \left( \mathbf{i}\theta _{s}\right) $ and $x_{s}=\cos
\theta _{s}$, it is clear that the right-hand side of~(\ref{xndot}) is an
even function of $\theta_s $, hence a function of $x_{s}$, where $s=1,\ldots ,N
$.
This is an interesting dynamical system, a complete investigation of which
is beyond the scope of the present paper.\ But before proceeding with our
task, let us pause and recall that the first idea to relate the zeros of
polynomials to a dynamical system goes back to Stieltjes \cite{S1885}, was
resuscitated in \cite{C1978} to identify \textquotedblleft solvable''
many-body problems (see also the extended treatment of this approach in \cit
{C2001}), and then extensively used to obtain results concerning the zeros
of the classical polynomials and of Bessel functions, see the paper \cit
{ABCOP1979} where several such findings are derived and reviewed. For more
recent developments along somewhat analogous lines see, for instance, \cit
{OS2004}, \cite{vD2005} and \cite{GVZ2014}.
Here we need to focus only on the behavior of this $t$-evolution in the
infinitesimal vicinity of the equilibrium configuration $x_{n}\left(
t\right) =\bar{x}_{n}\equiv \bar{x}_{n}\left( a,b,c,d;q;N\right) $, where
the $N$ numbers $\bar{x}_{n}$ are of course the $N$ zeros of the
Askey-Wilson polynomial $p_{N}\left( a,b,c,d;q;x\right) ,$ see (\ref{xnznbar
). To this end we set
\begin{subequations}
\begin{equation}
x_{n}\left( t\right) =\bar{x}_{n}+\varepsilon ~\xi _{n}\left( t\right)
~,~~~n=1,...,N~, \label{xksiepsi}
\end{equation
implying of cours
\begin{equation}
\dot{x}_{n}\left( t\right) =\varepsilon ~\dot{\xi}_{n}\left( t\right)
~,~~~n=1,...,N~, \label{xksidot}
\end{equation
with $\varepsilon $ infinitesimal.
It is plain that the insertion of this \textit{ansatz} in (\ref{xndot}) is
consistent to order $\varepsilon ^{0}=1$, thanks to \textbf{Proposition 2.1}.
The insertion of this \textit{ansatz} in (\ref{xndot}) yields, to order
\varepsilon $ (after a trivial but somewhat cumbersome computation) the
linear system of ODEs
\end{subequations}
\begin{equation}
\dot{\xi}_{n}\left( t\right) =\sum_{m=1}^{N}\left[ M_{nm}\left( \underline
\bar{x}}\right) ~\xi _{m}\left( t\right) \right] ~,~~~n=1,...,N~,
\label{ksidot}
\end{equation
with the $N\times N$ matrix $\underline{M}\left( \underline{\bar{x}}\right)
\equiv \underline{M}\left( a,b,c,d;q;N;\underline{\bar{x}}\right) $ given by
\begin{equation}
M_{nm}=\frac{\partial }{\partial x_{m}}\hat{F}_{n}(x_{1},\ldots ,x_{N})\Bigg
_{x_{s}=\bar{x}_{s}}=\frac{\partial }{\partial {z}_{m}}{F}_{n}(z_{1},\ldots
,z_{N})\frac{d{z}_{m}}{dx_{m}}\Bigg|_{z_{s}=\bar{z}_{s}},\;s=1,\ldots ,N,
\label{eq:MDefDerivFn}
\end{equation
which, after the substitution $\frac{dz_{m}}{dx_{m}}=\frac{2z_{m}^{2}}
z_{m}^{2}-1}$ (implied by $z_{m}=x_{m}+\sqrt{x_{m}^{2}-1}$) and
some trivial computations yields formula (\ref{M}). While by continuing the
expansion of the right-hand side of (\ref{xndot}) in powers of $\varepsilon $
and setting to zero the resulting coefficients of higher powers of
\varepsilon $ (since of course the left-hand side of (\ref{xndot}) contains
only a term of order $\varepsilon ,$ see (\ref{xksidot})), additional
formulas satisfied by the $N$ zeros $\bar{x}_{n}\equiv \bar{x}_{n}\left(
a,b,c,d;q;N\right) $ of the Askey-Wilson polynomial $P_{N}\left(
a,b,c,d;q;z\right) $ can be obtained.
The proof of \textbf{Proposition 2.2} is now a consequence of the fact that
the solution of the system of linear ODEs (\ref{ksidot}) is clearly a linear
superposition (with $t$-independent coefficients) of exponentials, $\exp
\left( \tilde{\mu}_{m}t\right) $, where the quantities $\tilde{\mu}_{m}$
(with $m=1,2,...,N$) are the $N$ eigenvalues of the $N\times N$ matrix
\underline{M}$; but---due to the simultaneous validity of the relations (\re
{psixt}), (\ref{Defxnt}) and (\ref{xksiepsi})---this solution must also be a
linear superposition (with $t$-independent coefficients) of the $t
-dependent quantities $c_{m}\left( t\right) .$ Hence (\ref{cmt}) implies
\tilde{\mu}_{m}=\mu _{m}=q^{-N}\left( 1-q^{m}\right) \left(
1-abcd~q^{2N-1-m}\right) $. \textbf{Proposition 2.2} is thereby proven.
Note that in our treatment above we implicitly assumed that the zeros
x_{n}\left( t\right) $ are---for all values of $t$---all different among
themselves. This is indeed the \textit{generic} situation. Any nongeneric
event---like the\ \textquotedblleft collision'' of two different zeros at
some special value of the parameter $t$---can be dealt with by appropriate
limits and in any case such possibilities---should they occur---do not
invalidate the proof of \textbf{Proposition 2.2}, as reported above.
\subsection{\textit{q}-Racah polynomials}
The following treatment is analogous---\textit{mutatis mutandis}---to that
of the previous Section 3.1, hence it shall be somewhat more terse.
Let $\Phi _{N}\left( z;t\right) \equiv \Phi _{N}\left( \alpha ,\beta ,\gamma
,\delta ;q;z;t\right) $ be a function of the variables $z$ and $t$, which
satisfies the Differential-\textit{q-}Difference Equation (D\textit{q}DE)
\begin{subequations}
\label{PDqDEqRacah}
\begin{eqnarray}
&&\frac{\partial ~\Phi _{N}\left( z;t\right) }{\partial ~t}=\left[ \left(
q^{-N}-1\right) \left( 1-\alpha \beta q^{N+1}\right) +B\left( z\right)
+D\left( z\right) \right] ~\Phi _{N}\left( z;t\right) \notag \\
&&-B\left( z\right) ~\Phi _{N}\left( z^{\left( +\right) };t\right) -D\left(
z\right) ~\Phi _{N}\left( z^{\left( -\right) };t\right) ~,
\end{eqnarray
where (see (\ref{z+-})
\begin{equation}
z^{\left( \pm \right) }\equiv z^{\left( \pm \right) }\left( \gamma \delta
;q;z\right) =q^{\pm 1}z\pm \left( \frac{1-q^{2}}{2q}\right) \left( z-\sqrt
z^{2}-4\gamma \delta q}\right)
\end{equation
and the functions $B\left( z\right) \equiv B\left( \alpha ,\beta ,\gamma
,\delta ;q;z\right) $ respectively $D\left( z\right) \equiv D\left( \alpha
,\beta ,\gamma ,\delta ;q;z\right) $ are defined by (\ref{B}) respectively
\ref{D}) with (see (\ref{Z})
\begin{equation}
Z\left( \gamma \delta q;z\right) =\frac{z+\sqrt{z^{2}-4\gamma \delta q}}
2\gamma \delta q}~. \label{ZZ}
\end{equation}
And let us assume that this function can be expressed (as confirmed below)
by the following linear superposition with $t$-dependent coefficients
c_{m}\left( t\right) $ of the $(N+1)$ \textit{q}-Racah polynomials
R_{N-m}\left( z\right) \equiv R_{N-m}\left( \alpha ,\beta ,\gamma ,\delta
;q;z\right) $ with $m=0,1,2,...,N$:
\end{subequations}
\begin{equation}
\Phi _{N}\left( z;t\right) =\sum_{m=0}^{N}\left[ c_{m}\left( t\right)
~R_{N-m}\left( z\right) \right] ~. \label{Phicmt}
\end{equation}
It is then plain---see (\ref{qdiffRacah}), of course with $N$ replaced by
(N-m)$---that the $t$-evolution of the $(N+1)$ coefficients $c_{m}\left(
t\right) $ is characterized by the following system of ODEs,
\begin{eqnarray}
\dot{c}_{m}\left( t\right) &=&\left[ \left( q^{-N}-1\right) \left( 1-\alpha
\beta q^{N+1}\right) \right. \notag \\
&&\left. -\left( q^{m-N}-1\right) ~\left( 1-\alpha \beta ~q^{N-m+1}\right)
\right] ~c_{m}\left( t\right) \notag \\
&=&q^{-N}~\left( 1-q^{m}\right) \left( 1-\alpha \beta ~q^{2N-m+1}\right)
~c_{m}\left( t\right) ~, \notag \\
m &=&0,1,,2...,N~,
\end{eqnarray
where again the superimposed dot denotes $t$-differentiation. Hence
\begin{subequations}
\label{cmtqRacah}
\begin{equation}
c_{0}\left( t\right) =c_{0}\left( 0\right) ~,
\end{equation
\begin{equation}
c_{m}\left( t\right) =\exp \left( \lambda _{m}~t\right) ~c_{m}\left(
0\right) ~,~~~m=1,2,...,N~, \label{cmtRacah}
\end{equation
where
\begin{eqnarray}
\lambda _{m} &\equiv &\lambda _{m}\left( \alpha \beta ;q;N\right)
=q^{-N}\left( 1-q^{m}\right) \left( 1-\alpha \beta ~q^{2N-m+1}\right) ~,
\notag \\
m &=&1,2,...,N~,
\end{eqnarray
implying (see (\ref{Phicmt})
\begin{equation}
\Phi _{N}\left( z;t\right) =c_{0}\left( 0\right) ~R_{N}\left( z\right)
+\sum_{m=1}^{N}\left[ \exp \left( \lambda _{m}~t\right) ~c_{m}\left(
0\right) ~R_{N-m}\left( z\right) \right] ~. \label{Phizt}
\end{equation}
Next, let us introduce the $N$ zeros $z_{n}\left( t\right) \equiv
z_{n}\left( \alpha ,\beta ,\gamma ,\delta ;q;N;t\right) $ of the polynomial
\Phi _{N}\left( z,t\right) $ of degree $N$ in $z$, by setting
\end{subequations}
\begin{equation}
\Phi _{N}\left( z;t\right) =C_{N}~\prod\limits_{n=1}^{N}\left[ z-z_{n}\left(
t\right) \right] ~. \label{FacPhi}
\end{equation
Here $C_{N}$ is an arbitrary constant that plays no role in the following.
Let us now investigate the $t$-evolution of the $N$ zeros $z_{n}\left(
t\right) $, as implied by the Differential \textit{q}-Difference Equation
\ref{PDqDEqRacah}) satisfied by $\Phi _{N}\left( z;t\right) $.
It is plain from (\ref{FacPhi})---by logarithmic $t$-differentiation---that
\begin{equation}
\frac{\partial ~\Phi _{N}\left( z;t\right) }{\partial t}=-\Phi _{N}\left(
z;t\right) ~\sum_{m=1}^{N}\frac{\dot{z}_{m}\left( t\right) }{z-z_{m}\left(
t\right) }=-C_{N}~\sum_{m=1}^{N}\left\{ \dot{z}_{m}\left( t\right)
\prod\limits_{\ell =1,~\ell \neq m}^{N}\left[ z-z_{\ell }\left( t\right)
\right] \right\} ~.
\end{equation
Henc
\begin{equation}
\left. \frac{\partial ~\Phi _{N}\left( z;t\right) }{\partial t}\right\vert
_{z=z_{n}\left( t\right) }=-C_{N}~\left\{ \dot{z}_{n}\left( t\right)
~\prod\limits_{\ell =1,~\ell \neq n}^{N}\left[ z_{n}\left( t\right) -z_{\ell
}\left( t\right) \right] \right\} ~,
\end{equation
and by setting $z=z_{n}\left( t\right) $ in (\ref{PDqDEqRacah}) we get (via
\ref{FacPhi}) implying $\Phi _{N}\left( z_{n}\left( t\right)
;t\right) =0$)
\begin{subequations}
\label{EvEqznRacah}
\begin{eqnarray}
&&\dot{z}_{n}=B\left( z_{n}\right) ~\left( z_{n}^{\left( +\right)
}-z_{n}\right) ~\prod\limits_{\ell =1,~\ell \neq n}^{N}\left( \frac
z_{n}^{\left( +\right) }-z_{\ell }}{z_{n}-z_{\ell }}\right) \notag \\
&&+D\left( z_{n}\right) ~\left( z_{n}^{\left( -\right) }-z_{n}\right)
~\prod\limits_{\ell =1,~\ell \neq n}^{N}\left( \frac{z_{n}^{\left( -\right)
}-z_{\ell }}{z_{n}-z_{\ell }}\right) ~, \notag \\
&&n=1,2,...,N~,
\end{eqnarray
where $B\left( z\right) \equiv B\left( \alpha ,\beta ,\gamma
,\delta ;q;z\right) $ respectively $D\left( z\right) \equiv D\left( \alpha
,\beta ,\gamma ,\delta ;q;z\right) $ are defined by (\ref{B}) respectively
\ref{D}) with (\ref{Z}), an
\begin{equation}
z_{n}^{\left( \pm \right) }=q^{\pm 1}z_{n}\pm \left( \frac{1-q^{2}}{2q
\right) \left( z_{n}-\sqrt{z_{n}^{2}-4\gamma \delta q}\right)
~,~~~n=1,...,N~.
\end{equation
Note that we omitted the arguments of $z_{n}\equiv
z_{n}\left( \alpha ,\beta ,\gamma ,\delta ;q;N;t\right) $ and, likewise,
z_{n}^{\left( \pm \right) }\equiv z_{n}^{\left( \pm \right) }\left( \alpha
,\beta ,\gamma ,\delta ;q;N;t\right) $.
Let us again mention that this set of nonlinear ODEs, (\ref{EvEqznRacah}),
is an interesting dynamical system, a complete investigation of which is
however beyond the scope of the present paper. Here we need to focus only on
its behavior in the infinitesimal vicinity of its equilibrium configuration
z_{n}\left( t\right) =\bar{z}_{n}\equiv \bar{z}_{n}\left( \alpha ,\beta
,\gamma ,\delta ;q;N\right) $, where the $N$ numbers $\bar{z}_{n}$ are now
of course the $N$ zeros of the \textit{q}-Racah polynomial $R_{N}\left(
\alpha ,\beta ,\gamma ,\delta ;q;z\right) $. To this end we set
\end{subequations}
\begin{subequations}
\begin{equation}
z_{n}\left( t\right) =\bar{z}_{n}+\varepsilon ~\zeta _{n}\left( t\right)
~,~~~n=1,...,N~, \label{zitaepsi}
\end{equation
implying of cours
\begin{equation}
\dot{z}_{n}\left( t\right) =\varepsilon ~\dot{\zeta}_{n}\left( t\right)
~,~~~n=1,...,N~,
\end{equation
with $\varepsilon $ infinitesimal.
It is plain that the insertion of this \textit{ansatz} in (\ref{EvEqznRacah
) is consistent to order $\varepsilon ^{0}=1$, thanks to \textbf{Proposition
2.3}.
The insertion of this \textit{ansatz} in (\ref{EvEqznRacah}) yields, to
order $\varepsilon $ (after a trivial but somewhat cumbersome computation)
the \textit{linear} system of ODEs
\end{subequations}
\begin{equation}
\dot{\zeta}_{n}\left( t\right) =\sum_{m=1}^{N}\left[ L_{nm}\left( \underline
\bar{z}}\right) ~\zeta _{m}\left( t\right) \right] ~,~~~n=1,...,N~,
\label{zittandot}
\end{equation
with the $N\times N$ matrix $\underline{L}\left( \underline{\bar{z}}\right)
\equiv \underline{L}\left( \alpha ,\beta ,\gamma ,\delta ;q;N;\underline
\bar{z}}\right) $ defined by (\ref{L}). While by continuing the expansion of
the right-hand side of (\ref{EvEqznRacah}) in powers of $\varepsilon $ and
setting to zero the resulting coefficients of higher powers of $\varepsilon
, additional formulas satisfied by the $N$ zeros $\bar{z}_{n}\equiv \bar{z
_{n}\left( \alpha ,\beta ,\gamma ,\delta ;q;N\right) $ of the \textit{q
-Racah polynomial $R_{N}\left( \alpha ,\beta ,\gamma ,\delta ;q;z\right) $
can be obtained.
The proof of \textbf{Proposition 2.4} is now a consequence of the fact that
the solution of the system of \textit{linear} ODEs (\ref{zittandot}) is
clearly a \textit{linear} superposition (with $t$-independent coefficients)
of exponentials, $\exp \left( \tilde{\lambda}_{m}t\right) $, where the
quantities $\tilde{\lambda}_{m}$ (with $m=1,2,...,N$) are the $N$
eigenvalues of the $N\times N$ matrix $\underline{L}$; but---due to the
simultaneous validity of the relations (\ref{Phizt}), (\ref{FacPhi}) and
\ref{zitaepsi})---this solution must also be a \textit{linear} superposition
(with $t$-independent coefficients) of the $t$-dependent quantities
c_{m}\left( t\right) .$ Hence (\ref{cmtqRacah}) implies $\tilde{\lambda
_{m}=\lambda _{m}=q^{-N}\left( 1-q^{m}\right) \left( 1-\alpha \beta
~q^{2N-m+1}\right) ,$ $m=1,2,...,N$. \textbf{Proposition 2.4} is thereby
proven.
\section{Appendix: Standard definitions and properties of the Askey-Wilson
and \textit{q}-Racah polynomials}
In this Appendix we report for the convenience of the reader, and also to
identify the notation used throughout this paper, a number of standard
formulas associated with the Askey-Wilson and \textit{q}-Racah polynomials.
We generally report these formulas from the standard compilation \cite{KS},
the formulas of which are identified by the notations (KS-X), where X stands
here for the notation appropriate to identify equations in this compilation;
in some cases we add certain immediate consequences of these formulas which
are not explicitly displayed in this compilation nor (to the best of our
knowledge) elsewhere.
The \textit{q}-Pochhammer symbol is defined as follows:
\begin{subequations}
\begin{equation}
\left( c;q\right) _{0}=1~;~~~\left( c;q\right) _{n}=\left( 1-c\right)
~\left( 1-cq\right) \cdot \cdot \cdot \left( 1-cq^{n-1}\right) ~~~\text
for~~~}n=1,2,3,...~, \label{q-Poch}
\end{equation
and we also use occasionally the synthetic notatio
\begin{equation}
\left( c_{1};q\right) _{k}\left( c_{2};q\right) _{k}\cdot \cdot \cdot \left(
c_{r};q\right) _{k}\equiv \left( c_{1},c_{2},...,c_{r};q\right) _{k}~,~\
\text{for~~~}r=0,1,2,...~. \label{Genq-Poch}
\end{equation}
The generalized basic hypergeometric function is defined as follows (see
(KS-0.4.2)):
\end{subequations}
\begin{eqnarray}
&&_{r+1}\phi _{s}\left( \left.
\begin{array}{c}
a_{0},a_{1},...,a_{r} \\
b_{1},b_{2},...,b_{s
\end{array
\right\vert q;z\right) \notag \\
&=&\sum_{k=0}^{\infty }\left[ \frac{\left( a_{0},a_{1},...,a_{r};q\right)
_{k}}{\left( b_{1},b_{2},...,b_{s};q\right) _{k}}\left( -1\right) ^{\left(
s-r\right) k}~q^{\left( s-r\right) k\left( k-1\right) /2}~\frac{z^{k}}
\left( q;q\right) _{k}}\right] ~. \label{BasicHypFunct}
\end{eqnarray
Here $r$ and $s$ are two arbitrary \textit{nonnegative} integers, but in the
following consideration shall be restricted to $r=s=3$.
The basic hypergeometric function (\ref{BasicHypFunct}) becomes a \textit
polynomial} in $z$ of degree $N$ if one of the parameters $a_{n}$ has the
values $q^{-N}$ with $N$ a positive integer---say, $a_{0}=$ $q^{-N}$
(without loss of generality, since its definition (\ref{BasicHypFunct})
implies that the basic generalized hypergeometric function is invariant
under permutations of the $r+1$ parameters $a_{j}$ as well as under
permutations of the $s$ parameters $b_{k}$)---provided no one of the other
parameters $a_{j}$ equals $q^{-\nu }$ with $\nu $ a negative integer smaller
in modulus than $N$, and no one of the parameters $b_{k}$ equals $q^{-\nu }$
with $\nu $ a negative integer (as we hereafter assume). This is of course a
simple consequence of the fact that, if $N$ is a positive integer, $\left(
q^{-N};q\right) _{n}$ vanishes for $n>N,$ see (\ref{q-Poch}). Note that in
this case the basic hypergeometric function is also a polynomial in each of
the $r$ parameters $a_{j}$ with $j=1,2,...,r.$
\subsection{Formulas for the Askey-Wilson polynomials}
Hereafter $\mathbf{i}$ denotes the imaginary unit, $\mathbf{i}^{2}=-1$.
The Askey-Wilson polynomial $p_{N}(a,b,c,d;q;x)$ with $x=\cos \theta$
is defined as follows (see (KS-3.1.1), and note some minor
notational changes):
\begin{subequations}
\label{AWpolx}
\begin{eqnarray}
&&p_{N}(a,b,c,d;q;\cos \theta )=\frac{\left(
ab,ac,ad;q\right) _{N}}{a^{N}}\cdot \notag \\
&&\cdot _{4}\phi _{3}\left( \left.
\begin{array}{c}
q^{-N},~abcd~q^{N-1},~a~\exp \left( \mathbf{i}\theta \right) ,~a~\exp \left(
-\mathbf{i}\theta \right) \\
ab,~ac,~a
\end{array
\right\vert q;q\right) ~,
\end{eqnarray
or, equivalently but more explicitly
\begin{eqnarray}
&&p_{N}(a,b,c,d;q;x)=\frac{\left( ab,ac,ad;q\right) _{N}~}{a^{N}}\cdot
\notag \\
&&\cdot \sum_{m=0}^{N}\left[ \frac{q^{m}\left( q^{-N};q\right) _{m}~\left(
abcd~q^{N-1};q\right) _{m}}{\left( q;q\right) _{m}~\left( ab;q\right)
_{m}~\left( ac;q\right) _{m}~\left( ad;q\right) _{m}}~\left\{ a;q;x\right\}
_{m}\right] ~, \notag \\
&& \label{AWPol}
\end{eqnarray
where we introduced the new (modified \textit{q}-Pochhammer) symbol defined
as follows:
\end{subequations}
\begin{eqnarray}
\left\{ a;q;x\right\} _{0} &=&1~; \notag \\
\left\{ a;q;x\right\} _{m} &=&\left( 1+a^{2}-2ax\right) ~\left(
1+q^{2}a^{2}-2aqx\right) \cdot \cdot \cdot \left(
1+a^{2}q^{2(m-1)}-2aq^{m-1}x\right) ~, \notag \\
m &=&1,2,3,...~. \label{ModqPoch}
\end{eqnarray
It is plain from this formula that $\left\{ a;q;x\right\} _{m}$ is a
polynomial of degree $m$ in $x$ (and also of degree $2m$ in $a$), hence that
the Askey-Wilson polynomials $p_{N}(a,b,c,d;q;x)$\textit{\ }are indeed
polynomials of degree $N$ in $x$ (see (\ref{AWPol})).
\textit{Notational remark}. For notational simplicity we often omit to
indicate explicitly the dependence on the $5$ parameters $a$, $b,$ $c,$ $d,$
$q$---or on some of them---provided this entails no ambiguity. $\square $
Let us also recall the\ related \textit{rational} function of $z$ defined as
follows:
\begin{eqnarray}
&&P_{N}\left( a,b,c,d;q;z\right) =\frac{\left( ab,ac,ad;q\right) _{N}}{a^{N}
\cdot \notag \\
&&\cdot _{4}\phi _{3}\left( \left.
\begin{array}{c}
q^{-N},~abcd~q^{N-1},~az,~a/z \\
ab,~ac,~a
\end{array
\right\vert q;q\right) ~, \label{PNz}
\end{eqnarray
see the formula in \cite{KS} preceding (KS-3.1.7). It is plain that this
amounts to the change of variable
\begin{equation}
x=\cos \theta =\frac{z^{2}+1}{2~z}~,~~\ z=e^{i\theta }=x+\sqrt{x^{2}-1}~,
\label{Changexz}
\end{equation
corresponding to the relations
\begin{subequations}
\label{PpAW}
\begin{equation}
P_{N}\left( a,b,c,d;q;z\right) =p_{N}\left( a,b,c,d;q;\frac{z^{2}+1}{2~z
\right) ~,
\end{equation
\begin{equation}
p_{N}\left( a,b,c,d;q;x\right) =P_{N}\left( a,b,c,d;q;x+\sqrt{x^{2}-1
\right) ~.
\end{equation
Note that although the square root $\sqrt{x^{2}-1}$ has two possible values,
the definition of the function $P_{N}\left( z\right) $, see (\ref{PNz}) and
\ref{q-Poch}), implies that the choice of the square root is irrelevant
because $\left( x+\sqrt{x^{2}-1}\right) \left( x-\sqrt{x^{2}-1}\right) =1$.
Hence if we denote as $\bar{x}_{n}$ the $N$ zeros of the Askey-Wilson
polynomial $p_{N}\left( a,b,c,d;q;x\right) ,
\end{subequations}
\begin{subequations}
\label{xnznbar}
\begin{equation}
p_{N}\left( a,b,c,d;q;\bar{x}_{n}\right) =0~,~\ ~n=1,2,...,N~,
\label{pNzeros}
\end{equation
it is plain that the rational function $P_{N}(a,b,c,d;q;z)$ features the $2N$
zeros $\bar{z}_{n}=\bar{x}_{n}+\sqrt{\bar{x}_{n}^{2}-1},$
\begin{equation}
P_{N}\left( a,b,c,d;q;\bar{z}_{n}\right) =0~,~\ ~n=1,2,...,N~.
\label{PNzeros}
\end{equation
In the following we will use whichever one of the two assignments of each square root
\sqrt{\bar{x}_{n}^{2}-1}$, $n=1,\ldots ,N$, is more convenient in order to
simplify the following formulas; of course the zeros $\bar{z}_{n}$ are
functions of the $6$ parameters $a$, $b$, $c$, $d$, $q$, $N$,
\begin{equation}
\bar{x}_{n}\equiv \bar{x}_{n}\left( a,b,c,d;q;N\right) ~,~~~\bar{z
_{n}\equiv \bar{z}_{n}\left( a,b,c,d;q;N\right) ~,~~~n=1,2,...,N~,
\end{equation
and they are related by the formulas
\begin{equation}
\bar{x}_{n}=\cos \bar{\theta}_{n}=\frac{\bar{z}_{n}^{2}+1}{2~\bar{z}_{n}
~,~~\ \bar{z}_{n}=e^{i\bar{\theta}_{n}}=\bar{x}_{n}+\sqrt{\bar{x}_{n}^{2}-1
~,~~~n=1,2,...,N~. \label{xnzn}
\end{equation
Note that occasionally we abuse language by referring both to the $N$
(generally \textit{complex}) numbers $\bar{x}_{n}$ and to the $2N$
(generally \textit{complex}) numbers $\bar{z}_{n}$ as \textit{zeros} of the
Askey-Wilson polynomial (of degree $N$), although of course only the $N$
numbers $\bar{x}_{n}$ are indeed $N$ zeros of the Askey-Wilson \textit
polynomial} $p_{N}\left( a,b,c,d;q;x\right) $, see (\ref{pNzeros}), while
the $2N$ numbers $\bar{z}_{n}$ defined by (\ref{xnzn}), in
terms of the $N$ numbers $\bar{x}_{n}$, are the zeros of the
\textit{rational} function $P_{N}\left( a,b,c,d;q;z\right) $, see (\re
{PNzeros}).
The rational function $P_{N}\left( a,b,c,d;q;z\right) $ satisfies the
following $q$-difference equation (see (KS-3.1.7)):
\end{subequations}
\begin{subequations}
\label{qDE}
\begin{equation}
Q~P_{N}\left( a,b,c,d;q;z\right) =\left( q^{-N}-1\right) ~\left(
1-abcd~q^{N-1}\right) ~P_{N}\left( a,b,c,d;q;z\right) ~, \label{Qeigen}
\end{equation
where the \textit{q}-difference operator $Q$ is defined as follows
\begin{equation}
Q~f\left( z\right) =\left[ A\left( z\right) ~\Delta _{q}^{\left( +\right)
}+A\left( z^{-1}\right) ~\Delta _{q}^{\left( -\right) }-A\left( z\right)
-A\left( z^{-1}\right) \right] ~f\left( z\right) ~. \label{Q}
\end{equation
Here $f\left( z\right) $ is an arbitrary function of the variable $z,$ the
operators $\Delta _{q}^{\left( \pm \right) }$ act as follows on $f\left(
z\right) ,
\begin{equation}
\Delta _{q}^{\left( \pm \right) }~f\left( z\right) =f\left( q^{\pm
1}~z\right) ~, \label{deltaplusminus}
\end{equation
and the function $A\left( z\right) $ is defined (here and throughout) as
follows
\begin{equation}
A\left( z\right) \equiv A\left( a,b,c,d;q;z\right) =\frac{\left( 1-az\right)
~\left( 1-bz\right) ~\left( 1-cz\right) ~\left( 1-dz\right) }{\left(
1-z^{2}\right) ~\left( 1-qz^{2}\right) }~. \label{A}
\end{equation
These formulas indicate that the operator $Q$---the definition of which
features (symmetrically) the $4$ parameters $a$, $b$, $c$, $d$ and moreover
the parameter $q$ (but not the parameter $N$)---has the rational functions
P_{N}(a,b,c,d;q;z)$ (for all positive integer values of $N$) as its
eigenfunctions, with corresponding eigenvalues $\left( q^{-N}-1\right)
~\left( 1-abcd~q^{N-1}\right) ,$ see (\ref{Qeigen}).
Note that, by setting $z=\bar{z}_{n}$ in (\ref{Qeigen}) one obtains the
following set of algebraic equations satisfied by the zeros of the
Askey-Wilson polynomial of degree $N$:
\end{subequations}
\begin{eqnarray}
&&A\left( \bar{z}_{n}\right) ~P_{N}(a,b,c,d;q;q~\bar{z}_{n}) +A\left( \frac{
}{\bar{z}_{n}}\right) ~P_{N}(a,b,c,d;q;\frac{\bar{z}_{n}}{q})=0~, \notag \\
n &=&1,...,N~. \label{Prop21}
\end{eqnarray
This observation is instrumental to prove \textbf{Proposition 2.2} (see
Section 3); it corresponds, via (\ref{A}) and (\ref{xnzn}), to \textbf
Proposition 2.1}---which has been displayed in Section 2 because we did not
find any previous mention of this rather trivial finding in the literature.
Although in fact---after this paper of ours was posted in the web as
arXiv:1410.0549---Jan Felipe van Diejen kindly brought to our attention that
essentially the same result was already reported in his paper \cite{vD2005},
see Theorem 2 and eq. (3.1b) there; and likewise our finding concerning the
zeros of the Wilson polynomials (see eq. (40) in our previous paper \cit
{BC2014a}) also appears there, see Theorem 3 in \cite{vD2005}.
\subsection{Formulas for the \textit{q}-Racah polynomials}
The \textit{q}-Racah polynomial $R_{N}(\alpha ,\beta ,\gamma ,\delta;q;z)$
is a polynomial of degree $N$ in $z$,
\begin{subequations}
\label{qRacah}
\begin{equation}
z\equiv z\left( \gamma \delta ;q;x\right) =q^{-x}+\gamma \delta q^{x+1}~,
\label{zqRacah}
\end{equation
defined as follows in terms of the generalized basic hypergeometric function
(see (KS-3.2.1), and note some minor notational changes):
\begin{equation}
R_{N}(\alpha ,\beta ,\gamma ,\delta ;q;z)=_{4}\phi _{3}\left( \left.
\begin{array}{c}
q^{-N},~\alpha \beta q^{N+1},~q^{-x},~\gamma \delta q^{x+1} \\
\alpha q,~\beta \delta q,~\gamma
\end{array
\right\vert q;q\right) ~, \label{qRacaha}
\end{equation
or, equivalently but more explicitly
\begin{eqnarray}
&&R_{N}(\alpha ,\beta ,\gamma ,\delta ;q;z)= \notag \\
&=&\sum_{m=0}^{N}\left[ \frac{q^{m}\left( q^{-N};q\right) _{m}\left( \alpha
\beta ~q^{N+1};q\right) _{m}\left( q^{-x};q\right) _{m}\left( \gamma \delta
q^{x+1};q\right) _{m}}{\left( q;q\right) _{m}\left( \alpha q;q\right)
_{m}\left( \beta \delta q;q\right) _{m}\left( \gamma q;q\right) _{m}}\right]
~. \label{qRacah2}
\end{eqnarray
The fact that $R_{N}(z)\equiv R_{N}(\alpha ,\beta ,\gamma ,\delta ;q;z)$ is
indeed a polynomial of degree $N$ in $z$ is implied by the last formula and
by the identity (valid via (\ref{zqRacah})),
\end{subequations}
\begin{equation}
\left( q^{-x};q\right) _{m}\left( \gamma \delta q^{x+1};q\right)
_{m}=\prod\limits_{s=0}^{m-1}\left( 1-zq^{s}+\gamma \delta q^{2s+1}\right) ~.
\end{equation}
\textbf{Remark A.1}. Hereafter the $5$ parameters $\alpha $, $\beta $,
\gamma $, $\delta $, $q\neq 1$ are arbitrary numbers (possibly \textit
complex}); note that this entails a somewhat more general definition of
\textit{q}-Racah polynomials than the standard one, see (KS-3.2.1), because
it does \textit{not} require the Diophantine restriction on one of the $3$
parameters, $\alpha q$ or $\beta \delta q$ or $\gamma q$, see the second
(unnumbered) equation after (KS-3.2.1) in \cite{KS}. Indeed, while this
restriction is required for the validity of many of the properties of
\textit{q}-Racah polynomials reported in \cite{KS}, it is not required for
the \textit{q}-difference equation satisfied by these polynomials, see
(KS-3.2.6) and immediately below, which is the only property of these
polynomials that we use in order to prove the properties of the zeros of
these polynomials reported in this paper. $\square $
The \textit{q}-Racah polynomial $R_{N}(z)\equiv R_{N}(\alpha ,\beta ,\gamma
,\delta ;q;z)$ satisfies the following \textit{q}-difference equation:
\begin{subequations}
\label{qdiffRacah}
\begin{eqnarray}
&&B\left( z\right) ~R_{N}(z^{\left( +\right) })-\left[ B\left( z\right)
+D\left( z\right) \right] ~R_{N}(z)+D\left( z\right) ~R_{N}(z^{\left(
-\right) }) \notag \\
&=&\left( q^{-N}-1\right) \left( 1-\alpha \beta q^{N+1}\right) ~R_{N}(z)~,
\label{qdiffRacaha}
\end{eqnarray
wher
\begin{equation}
z^{\left( \pm \right) }=z\left( x\pm 1\right) =q^{\pm 1}z\pm \left( \frac
1-q^{2}}{2q}\right) \left[ z-\sqrt{z^{2}-4\gamma \delta q}\right]
\label{z+-}
\end{equation
an
\begin{equation}
B\left( z\right) =\frac{\left[ 1-\alpha qZ\left( q;z\right) \right] ~\left[
1-\beta \delta qZ\left( q;z\right) \right] ~\left[ 1-\gamma qZ\left(
q;z\right) \right] ~\left[ 1-\gamma \delta qZ\left( q;z\right) \right] }
\left[ 1-\gamma \delta qZ^{2}\left( q;z\right) \right] ~\left[ 1-\gamma
\delta q^{2}Z^{2}\left( q;z\right) \right] }~, \label{B}
\end{equation
\begin{equation}
D\left( z\right) =\frac{q~\left[ 1-Z\left( \gamma \delta q;z\right) \right]
\left[ 1-\delta Z\left( \gamma \delta q;z\right) \right] ~\left[ \beta
-\gamma Z\left( \gamma \delta q;z\right) \right] ~\left[ \alpha -\gamma
\delta Z\left( \gamma \delta q;z\right) \right] }{\left[ 1-\gamma \delta
Z^{2}\left( \gamma \delta q;z\right) \right] ~\left[ 1-\gamma \delta
qZ^{2}\left( \gamma \delta q;z\right) \right] }~, \label{D}
\end{equation
wher
\begin{equation}
Z\left( \gamma \delta q;z\right) =q^{x}=\frac{z+\sqrt{z^{2}-4\gamma \delta q
}{2\gamma \delta q}~. \label{Z}
\end{equation
In (\ref{z+-}) and (\ref{Z}) the determination of the square root is
irrelevant; of course provided the \textit{same} determination is used
everywhere.
It is easily seen that these formulas correspond to (KS-3.2.6) via (\re
{zqRacah}).
It is plain from this formula that, if $\bar{z}_{n}$ are the $N$ zeros of
the \textit{q}-Racah polynomial of order $N$,
\end{subequations}
\begin{subequations}
\begin{equation}
R_{N}(\bar{z}_{n})=0~,~~~n=1,...,N~,
\end{equation
\begin{equation}
R_{N}(z)=C_{N}~\prod\limits_{n=1}^{N}\left( z-\bar{z}_{n}\right) ~,
\label{FacRacah}
\end{equation
where $C_N$ is a constant. Formula (\ref{qdiffRacaha}) implies the relation
\end{subequations}
\begin{subequations}
\begin{equation}
B\left( \bar{z}_{n}\right) ~R_{N}(\bar{z}_{n}^{\left( +\right) })+D\left(
\bar{z}_{n}\right) ~R_{N}(\bar{z}_{n}^{\left( -\right) })=0~,
\label{48a}
\end{equation
where of course (see (\ref{z+-})
\begin{equation}
\bar{z}_{n}^{\left( \pm \right) }=q^{\pm 1}\bar{z}_{n}\pm \left( \frac
1-q^{2}}{2q}\right) \left( \bar{z}_{n}-\sqrt{\bar{z}_{n}^{2}-4\gamma \delta
}\right) ~;\
\end{equation
and this formula coincides with (\ref{Prop23a}) and, via (\ref{FacRacah}),
with (\ref{Prop23b}). \textbf{Proposition 2.3} is thereby proven.
\section{Acknowledgements}
One of us (OB) would like to acknowledge with thanks the hospitality of the
Physics Department of the University of Rome ``La Sapienza'' on the occasion
of three two-week visits there in June 2012, May 2013 and June-July 2014;
the results reported in this paper were obtained during the last of these
visits. The other one (FC) would like to acknowledge with thanks the
hospitality of Concordia College for a one-week visit there in November 2013.
\bigskip
|
\section{Introduction}
\label{sec:intro}
In this article, we study regularity of periodic solutions to the initial boundary value problem
\begin{equation}
\label{eq:liouville}
\begin{cases}
\partial_{\alpha t}\ln u=f(\alpha)u,\,\,\,\,\,\,\,\,\,\,\,&\alpha\in(0,1),\,\,\,t>0,
\\
u(\alpha,0)=u_0(\alpha),\,\,\,\,\,\,\,\,\,\,\,\,&\alpha\in[0,1],
\\
u(0,t)=u(1,t)=g(t),\,\,\,\,\,\,\,\,\,\,\,\,\,&t\geq0
\end{cases}
\end{equation}
for given bounded continuous functions $f$ and $u_0>0$ with prescribed boundary data $g>0$. Moreover, replacing $u$ on the right-hand side of \eqref{eq:liouville}i) by an arbitrary nonnegative function $\mathcal{F}(u)\in C^1(0,+\infty)$ we also establish regularity criteria for the resulting generalization of \eqref{eq:liouville}.
Note that for $u>0$ to be $\alpha$-periodic, integration of (\ref{eq:liouville})i) (or of its generalization) over $(0, 1)$ requires $f(\alpha)$ to have at least one zero in $(0, 1)$ and to change sign if it is not identically zero.
We will refer to (\ref{eq:liouville}) as an initial periodic-boundary value problem for the sign-changing Liouville equation.
Two particular versions of Liouville's equation
\begin{equation}
\label{eq:hyp}
\square \psi (x,\tau) +\epsilon e^{\psi}=0,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\epsilon=\pm1
,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\square\equiv\partial^2_{\tau}-\partial^2_x\end{equation}
occur in various applications ranging from plasma physics and field theoretical modeling to fluid dynamics. This has made both versions of (\ref{eq:hyp}) a frequent topic of investigation. In particular cases, the equation can be interpreted as a model for a self-interacting scalar field in two-dimensional space-time, whose properties have been the subject of extensive study (see for instance \cite{Crowdy1}-\cite{P2}). Equation (\ref{eq:liouville})i) is obtained from (\ref{eq:hyp}) on changing to characteristic coordinates $\tau=\alpha+t$ and $x=\alpha-t$ and setting $\psi=\ln u$, with $f(\alpha)$ replacing $\epsilon$ in the resulting equation. Further, the elliptic Liouville equation
\begin{equation}
\label{eq:ell}
\Delta\ln\phi=-K\phi^2,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\Delta\equiv\partial^2_x+\partial^2_y,
\end{equation}
appears in the study of two-dimensional steady, incompressible Euler flow with $\phi=e^{\psi}$, for $\psi$ the stream function relating vorticity to the velocity field, (\cite{S1}, \cite{Crowdy2}). As in its hyperbolic counterpart, (\ref{eq:ell}) reduces to (\ref{eq:liouville})i) along curves $\alpha=x+iy$ and $t=x-iy$ for $f\equiv-\frac{K}{2}$ and $\phi^2=u$.
Amongst its many physical applications, (\ref{eq:liouville})i) is found in Riemannian geometry. On prescribing the Gaussian curvature $K(x,\tau)$ for a pseudo-metric $ds^2=g^{ij}dx_idx_j=e^{2v(x,\tau)}(dx^2-d\tau^2)$ in two-dimensional Minkowski space, the function $v$ satisfies the relation
$$K(x,\tau)=-e^{-2v(x,\tau)}\square v(x,\tau)$$
in isothermal coordinates $(x,\tau)$, (\cite{rogers}).
In the case of constant Gaussian curvature, the change to characteristic variables
then leads to Liouville's equation in the form, (\cite{L1}),
$$\partial_{\alpha t}v=Ke^{2v}.$$
In a related setting, the case of sign-changing $K$ has been studied recently for an elliptic version of the equation, (\cite{Ruf1}).
We note that the relevance of (\ref{eq:liouville})i) in the field of fluid dynamics is not limited to the hyperbolic and elliptic models (\ref{eq:hyp}) and (\ref{eq:ell}). Indeed, if we restrict ourselves to quantities $u$ having
constant spatial mean, then for a prescribed periodic function $f$ and initial data $u_0\equiv1$, (\ref{eq:liouville})i) appears in the study of classes of semi-bounded solutions to the three-dimensional incompressible Euler equations, (\cite{Okamoto,Wunsch,Sarria1}). In this context, $f$ controls the concavity of the components of the velocity field and $u$ represents the jacobian of the transformation associated to the particle trajectories in the fluid. Finally, we remark that the subsequent generalization \eqref{u}i) may also have applications in the study of bi-Hamiltonian equations such as the $\mu$Hunter-Saxton equation (\cite{Khesin}), which describes the orientation of highly inertial liquid crystal director fields in the presence of an external magnetic field.
The outline for the remainder of the paper is as follows. A general representation formula for periodic solutions to (\ref{eq:liouville}) is derived in \S\ref{sec:solution}, and certain aspects of the structure of discontinuous and singular solutions are examined in \S\ref{sec:singular}. In section \S\ref{subsec:smoothbdry}, $L^{\infty}$
boundedness is studied for smooth boundary data $g(t)$, with the effects of singular $g$ on interior regularity of solutions being considered in \S\ref{subsec:singularbdry}. Finite-time blowup in $L^p$ , $1\leq p<\infty$, is discussed in \S\ref{subsec:lpsmooth}. Lastly, regularity criteria for a generalization of \eqref{eq:liouville} (see \eqref{u}) is established in \S\ref{sec:generalization}. The reader may then refer to \S\ref{sec:examples} for specific examples.
\section{The Representation Formula}
\label{sec:solution}
In this section we derive a representation formula for solutions to (\ref{eq:liouville}). We begin by noticing that if $y$ satisfies the associated problem
\begin{equation}
\label{eq:p1}
\begin{cases}
\partial_{\alpha t}\ln y=f(\alpha)u_0(\alpha)y,\,\,\,\,\,\,&\,\,t>0,
\\
y(\alpha,0)=1,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,&\alpha\in[0,1],
\end{cases}
\end{equation}
then
\begin{equation}
\label{eq:truesol}
\begin{split}
u(\alpha,t)=u_0(\alpha)y(\alpha,t).
\end{split}
\end{equation}
Below, in formula (\ref{eq:ysoln}), we have used (\ref{eq:p1}) to establish a representation formula for $y(\alpha, t)$ in order to find $u(\alpha, t)$.
We first notice that on dividing (\ref{eq:p1})i) by $y$, differentiating in time, and using the calculus identity $\partial_t(z^{-1}\partial_{\alpha t}\ln z)=z^{-1}\partial_{\alpha}(\partial_{tt}\ln z - \frac{1}{2}(\partial_t\ln z)^2)$
this shows
\begin{equation}
\label{eq:eq0}
\partial_{\alpha}\mathcal{R}(\partial_t\ln y; t)=0
\end{equation}
where
\begin{equation}
\label{eq:eq2}
\mathcal{R}(v; t)\equiv\partial_t v-\frac{1}{2}v^2.
\end{equation}
Next, integrating (\ref{eq:eq0}) from $0$ to $\alpha$ (or from $1$ to $\alpha$) and using (\ref{eq:truesol}) with boundary data $u(0,t)=u(1, t)=g(t)$, it follows that
\begin{equation}
\label{eq:eq3}
\mathcal{R}(\partial_t\ln y; t)=\mathcal{R}\left(\frac{d\ln g}{dt}; t\right),
\end{equation}
where we adopt the (simplifying) assumption that $u_0(0)=1$ (and subsequently to be consistent, $g(0)=1$).
Now, the Schwarzian derivative of a function $w(t)$ may be defined in terms of $\mathcal{R}$ by
\begin{equation}
\mathcal{S}(w; t)=\mathcal{R}\left(\frac{d}{dt} \ln \frac{dw}{dt}; t\right)
\label{eq:eq5}
\end{equation}
and it has the property (\cite{CDO}) that
$\mathcal{S}(p; t)=\mathcal{S}(q; t)$ if and only if
$p$ and $q$ are related by a fractional linear (M\"{o}ebius) transformation,
$$q(t)=\frac{ap(t)+b}{cp(t)+d}.$$
It follows from \eqref{eq:eq3} and \eqref{eq:eq5} that if we set $y=\partial_tY$ and
\begin{equation}
\label{eq:G}
g(t)=\frac{dG}{dt},
\end{equation}
with $Y(\alpha, 0)=G(0)=0$, then
\begin{equation}
Y(\alpha, t)=\frac{a(\alpha)G(t)}{c(\alpha)G(t)+d(\alpha)}
\end{equation}
for some functions $a(\alpha), c(\alpha)$ and $d(\alpha)$. If we next set $\Delta(\alpha)=a(\alpha)d(\alpha)$, we find by successive differentiation that
\begin{equation}
\label{eq:y1}
y(\alpha, t)=\frac{\Delta(\alpha)g(t)}{(c(\alpha)G(t)+d(\alpha))^2}
\end{equation}
and
\begin{equation}
\label{eq:preceding}
\partial_ty(\alpha, t)=\Delta(\alpha)\frac{(c(\alpha)G(t)+d(\alpha))\dot{g}(t)-2c(\alpha)g^2(t)}{(c(\alpha)G(t)+d(\alpha))^3}.
\end{equation}
Consequently \eqref{eq:p1}ii) and \eqref{eq:y1} imply $\Delta(\alpha)=d^2(\alpha)$. Integrating \eqref{eq:p1}i) now gives
\begin{equation}
\partial_t\ln y(\alpha, t)=\frac{d\ln g}{dt}(t)+\psi(\alpha, t)
\end{equation}
where
\begin{equation}
\label{eq:psi}
\psi(\alpha, t)=\int_0^\alpha f(z)u_0(z)y(z, t)\,dz=\int_0^\alpha f(z)u(z, t)\,dz,
\end{equation}
and so at $t=0$,
\begin{equation}
\label{eq:psi0}
\partial_ty(\alpha, 0)=\dot{g}(0)+\psi_0(\alpha)
\end{equation}
where
\begin{equation}
\label{eq:Psi}
\psi_0(\alpha)=\int_0^\alpha f(z)u_0(z)\,dz.
\end{equation}
In contrast, equation \eqref{eq:preceding} implies
\begin{equation}
\label{eq:here}
\partial_ty(\alpha, 0)=\Delta (\alpha)\frac{d(\alpha)\dot{g}(0)-2c(\alpha)}{d^3(\alpha)}=\dot{g}(0)-2\frac{c(\alpha)}{d(\alpha)}
\end{equation}
which lets us combine equations \eqref{eq:y1}, \eqref{eq:psi0} and \eqref{eq:here} to write $y(\alpha, t)$ in terms of initial and boundary data as
\begin{equation}
\label{eq:ysoln}
y(\alpha, t)=\frac{g(t)}{(1-\frac{1}{2}\psi_0(\alpha)G(t))^2},
\end{equation}
giving
\begin{equation}
\label{eq:finalsolution}
u(\alpha, t)=\frac{u_0(\alpha)g(t)}{(1-\frac{1}{2}\psi_0(\alpha)G(t))^2}
\end{equation}
from \eqref{eq:truesol}.
\begin{rem}
We note that if we choose constant boundary data $g(t)\equiv1$ the final solution simplifies to
\begin{equation}
\label{eq:finalsolutionspecial}
u(\alpha,t)=\frac{u_0(\alpha)}{(1-\frac{t}{2}\psi_0(\alpha))^2},
\end{equation}
which, clearly, will persist for all time if $\psi_0(\alpha)\leq0$, for all $\alpha\in[0,1]$, but will become singular in finite time provided $\psi_0(\alpha)>0$ for some $\alpha\in (0, 1).$
More generally, since $g(t)>0$, monotonicity of $G(t)$ implies $G(t)\rightarrow G_\infty$ as $t\rightarrow\infty$, where $G_\infty>0$ may, or may not, be bounded. Equation \eqref{eq:finalsolution} then shows solutions persist for all time if ${max}_{\alpha\in (0, 1)}\psi_0(\alpha)<2/G_\infty$ while finite time blowup takes place if $\psi_0(\alpha)>2/G_\infty$ for some $\alpha\in(0, 1).$ \end{rem}
\section{Basic Properties of Singular Solutions}
\label{sec:singular}
Here we briefly examine some possible types of nonsmooth structure of solutions from the formula given by (\ref{eq:finalsolution}). If we allow jumps in $u(\alpha ,t)$ to be defined by
$$
[u(\alpha,\cdot)](t) = \lim_{\tau\downarrow t}u(\alpha, \tau) - \lim_{\tau\uparrow t}u(\alpha, \tau)
$$
and
$$
[u(\cdot , t)](\alpha) = \lim_{\beta\downarrow\alpha}u(\beta, t)- \lim_{\beta\uparrow\alpha}u(\beta ,t),
$$
then jump discontinuities resulting from jumps in the boundary or initial data functions $g(t)$ or $u_0(\alpha)$, propagate along characteristics (lines of constant $t$ or $\alpha$), since continuity of the primitive functions $G(t)$ and $\psi_0(\alpha)$ implies
$$
[u(\alpha,\cdot)](t)=\frac{u_0(\alpha)\, [g(\cdot)](t)}{(1-\frac{1}{2}G(t) \psi_0(\alpha))^2}
$$
and
$$
[u(\cdot , t)](\alpha)=\frac{[u_0(\cdot)](\alpha) \, g(t)}{(1-\frac{1}{2}G(t)\psi_0(\alpha))^2}.
$$
On requiring $u_0(\alpha)$ and $g(t)$ to be positive, jumps in $u(\alpha, t)$, which stem from initial or boundary data, remain nonzero along corresponding characteristics.
If we next denote the set,\, $\Sigma$\,, by
\begin{equation}
\label{eq:Sigma}
\Sigma = \{(\sigma, \tau)\in(0, 1)\times(0, \infty): \psi_0(\sigma)\,G(\tau)=2\},
\end{equation}
then, to be strictly valid, the solution formula \eqref{eq:finalsolution} requires that both a vertical and at least one horizontal characteristic avoid intersecting $\Sigma$ at any point $(\sigma, \tau)$ prior to reaching $(\alpha, t)$, (\cite{kichenassamy}). This is not always possible, but
the method used to construct formula \eqref{eq:ysoln} above remains {\em a posteriori} valid at $(\alpha, t)$
[using vertical characteristics together with horizontal characteristics coming either from the left (lines with constant $t> 0$ which meet $\alpha=0$) or from the right (through a similar construction\footnote[1]{For the latter construction one integrates instead from $1$ to $\alpha$ and reaches the same solution formula as before with $\psi_0(\alpha)$ replaced by $\int_1^\alpha f(\alpha)u_0(\alpha)d\alpha$. Compatibility of initial data with \eqref{eq:liouville}i) implies $\int_0^1 f(\alpha)u_0(\alpha)d\alpha$ = 0.} using periodicity of boundary data and meeting $\alpha=1$)] if one imposes appropriate conditions on the function $f(\alpha)$.
A condition sufficient for formula \eqref{eq:finalsolution} to hold everywhere beneath $\Sigma$ (both before and after solutions begin to develop singularities) is for $\psi_0(\alpha)$ to be positive on only a single, open, connected set in $(0, 1)$
on which it has a single maximum and no local minima.
If one assumes that $f'(\alpha_0)<0$ wherever $f(\alpha_0)=0$ and $\psi_0(\alpha_0)>0$, then this suffices since $\psi'_0(\alpha)=f(\alpha)u_0(\alpha)$ and $\psi_0(\alpha)$ is consequently convex down at its extrema.
In order to consider the properties of non-characteristic curves in $\Sigma$ further,
suppose in (\ref{eq:finalsolution}) that the functions $f(\alpha), \, g(t)>0$ and $u_0(\alpha)>0$ are continuous for $\alpha\in (0, 1)$ and set $\mathcal{F}(\alpha, t)=G(t)\psi_0(\alpha)-2$. Then $\mathcal{F}_t(\alpha, t)\neq 0$ wherever $\psi_0(\alpha)>0$ and, by the implicit function theorem, there exists a unique curve,
$t=\tilde{t}(\alpha)$, in the local neighborhood of any point $(\tilde{\alpha}, \tilde{t})$ where $\mathcal{F}(\tilde{\alpha}, \tilde{t})=0$, through which $\mathcal{F}(\alpha, \tilde{t}(\alpha))=0.$
For ${max}_{\alpha\in (0, 1)}\psi_0(\alpha)>2/G_\infty$,\, $\tilde{t}(\alpha)\in\Sigma$ then lies
in the region
\begin{equation}
\label{eq:A}
\mathcal{A}=\{(\alpha, t): \psi_0(\alpha)\geq 0, t\geq 0\}
\end{equation}
and is given by the formula
\begin{equation}
\label{eq:t}
\tilde{t}(\alpha)=G^{-1}(2/\psi_0(\alpha)).
\end{equation}
We define
\begin{equation}
\label{eq:dA}\partial\mathcal{A}=\{(\alpha, t): \psi_0(\alpha)=0, t> 0\}
\cup\{(\alpha, t):\psi_0(\alpha)>0, \,t=0\}
\end{equation}
and
\begin{equation}
\label{eq:A+-}\mathcal{A}_\pm=\{(\alpha, t)\in\mathcal{A}: f(\alpha)\gtrless 0\}\,\,\mbox{and}\,\,
\mathcal{A}_0=\{(\alpha, t)\in\mathcal{A}: f(\alpha)=0\}.
\end{equation}
On differentiating, $\tilde{t}(\alpha)$ in (\ref{eq:t}) gives, for $\alpha\in (0, 1)$,
\begin{equation}
\label{eq:slope}
\tilde{t}_\alpha(\alpha)=-\frac{2f(\alpha)u_0(\alpha)}{g\circ G^{-1}(2/\psi_0(\alpha))\,\psi_0^2(\alpha)}
\end{equation}
and so the slope of $\tilde{t}(\alpha)$ is negative in $\mathcal{A}_+$, positive in $\mathcal{A}_-$, and zero in $\mathcal{A}_0$.
We will be interested subsequently in singular curves, $t=\tilde{t}(\alpha)$, which may meet the boundary, $\partial\mathcal{A}$. In general, if $\psi_0(\alpha)$ is continuous on $[0, 1]$ and $G(t)$ is continuous and bounded on $[0, \infty)$, then no curve in $\Sigma$ can meet $\partial\mathcal{A}$.
Points on $\partial\mathcal{A}$ either take the form $(\alpha_\sharp, t_\sharp\, )$ where $\psi_0(\alpha_\sharp)=0$ and $G(t_\sharp)$ is unbounded, or $(\alpha_\flat, 0)$ where $\psi_0(\alpha_\flat)$ is unbounded and $G(0)=0$.
We will let $\mathcal{A}$ take the form of the set $\mathcal{C}=\{(\alpha, t):\alpha_l\leq\alpha\leq\alpha_r, t\geq 0\}$ where $\psi_0(\alpha)> 0\,\, $ for every
$\alpha\in(\alpha_l, \alpha_r)$ and $\psi_0(\alpha_l)=\psi_0(\alpha_r)=0$.
By assuming $f(\alpha)>0$ for $\alpha$ close to $0$, we can let $\alpha_l=0$, with $\alpha_r\leq1$. $\partial\mathcal{C}$ is defined in an analogous way to $\partial\mathcal{A}$.
Following from the definitions, we have that $\psi_0(0)=G(0)=0$, to which we add some further simplifying hypotheses, based on the choice of the coefficient $f(\alpha)$ and the data $u_0(\alpha)$ and $g(t)$, in that leading order behaviour is given by
$$\mbox{(H1)}\,\,\psi_0\sim \alpha^{a_0} \mbox{ as }\alpha\downarrow 0, \,\, \psi_0\sim |\alpha-\alpha_r|^{a_r} \mbox{ as }\alpha\uparrow\alpha_r ,
\mbox{ and }G\sim \, t^{b_0} \mbox{ as }t\downarrow 0$$
where $a_0, a_r, b_0 >0$ and the symbol $\sim$ will mean that the quotient of the two sides tends to a positive constant in the limit. Similarly, we will assume that in the limits of $\alpha$ approaching $\alpha_\flat \in(0, \alpha_r)$ or $t$~approaching $t_\sharp>0$,
$$\mbox{(H2)}\qquad\qquad \psi_0(\alpha)\sim |\alpha - \alpha_\flat |^a \mbox{ as }\alpha\rightarrow \tilde{\alpha}\mbox{ or }G(t)\sim |t- t_\sharp |^b \mbox{ as }t\rightarrow t_\sharp$$
where $a=a(\alpha_\flat),\, b=b(t_\sharp).$
If, under these assumptions, the curve $t=\tilde{t}(\alpha)$ connects to $\partial{\mathcal C}$ at $(\alpha_\flat, 0)$,
then (\ref{eq:Sigma}) implies that
$$|\tilde{t}(\alpha)|^{b_0}|\alpha-\alpha_\flat |^a\sim c\mbox{ as }\alpha\rightarrow\alpha_\flat $$
for some generic constant, $c>0.$
If $t=\tilde{t}(\alpha)$ connects to $\partial{\mathcal C}$ at~$(0, t_\sharp)$, then
$$|t_\sharp-\tilde{t}(\alpha)|^b\alpha^{a_0}\sim c\mbox{ as }\alpha\downarrow 0$$
As a result, $a_0, b_0 > 0$ require either $b<0$, or $a<0$, for the curve to meet $\partial\mathcal{C}$ at $t=0$, or $\alpha = 0$, respectively. In the latter case, one must clearly also have
$$|t_\sharp-\tilde{t}(\alpha)|^b|\alpha-\alpha_r|^{a_r}\sim c \mbox{ as }\alpha\uparrow\alpha_r .$$
In the event that $G(t)$ becomes unbounded only as $t\rightarrow\infty$, the branches of $\Sigma$ are asymptotic to $\alpha=0$ and $\alpha=\alpha_r$.
\section{Regularity Results}
\label{sec:regularity}
In this section we are concerned simply with finite-time blowup, or global existence in time, of (\ref{eq:finalsolution}). In \S\ref{subsec:smoothbdry}, we study the interior regularity of solutions that are smooth at the boundary for all time, while, in \S\ref{subsec:singularbdry}, the case of non-smooth $g(t)>0$ is considered. More particularly, for the former we establish criteria in terms of the sign-changing function $f$ and initial data $u_0$ leading to finite-time blowup or global-in-time solutions. Then, in \S\ref{subsec:singularbdry}, we examine the effects on the interior regularity of $u$ of boundary data $g(t)$ having a particular singular form, specifically, for some $0<t_b<+\infty$, $g(t)$ is taken to be smooth on $t\in[0,t_b)$ but $\lim_{t\uparrow t_b}g(t)=+\infty$. In this case, we find that under certain conditions the solution $u$ can in fact diverge somewhere in the interior at a time $0<t_*<t_b$.
First we define some terminology. Let $M_0$ denote the greatest value attained by $\psi_0(\alpha)$ at a finite number of locations $\overline\alpha_i\in[0,1],\, 1\leq i\leq n$, namely
\begin{equation}
\label{eq:M0}
M_0\equiv\max_{\alpha\in[0,1]}\psi_0(\alpha)=\psi_0(\overline\alpha_i),\qquad\quad 1\leq i\leq n.
\end{equation}
Notice that, since $\psi_0(0)=\psi_0(1)=0$, $M_0\in\mathbb{R}^+\cup\{0\}$.
\subsection{Smooth Boundary Data}
\label{subsec:smoothbdry}
Suppose $u_0(\alpha)>0$ and $f(\alpha)$ are bounded continuous functions for all $\alpha\in[0,1]$, and the boundary data $g(t)>0$ is smooth. In this section we examine $L^{\infty}(0,1)$ regularity of (\ref{eq:finalsolution}). We begin by establishing simple criteria leading to global-in-time solutions.
\subsubsection{Global-in-time Solutions}
\label{subsubsec:globalsmoothbdry}
\hfill
\noindent
Note that a solution to (\ref{eq:liouville}) will persist for all time as long as (\ref{eq:finalsolution}) remains both finite and positive for all $0<t<+\infty$. Suppose $u_0$ and $f$ are such that $M_0=0$, that is $\psi_0(\alpha)\leq0$ for all $\alpha\in[0,1]$. This implies that $1-\frac{1}{2}G(t)\psi_0(\alpha)\geq1$ or, from (\ref{eq:finalsolution}),
\begin{equation}
\label{eq:global1}
0<u(\alpha,t)\leq g(t)u_0(\alpha).
\end{equation}
Since $g(t)$ is smooth for all time and $u_0(\alpha)\in L^{\infty}[0,1]$, $u$ remains finite for all $\alpha\in[0,1]$ and $0<t<+\infty$. This leads to Theorem \ref{thm:globaltheorem} below.
\begin{thm}
\label{thm:globaltheorem}
Consider the initial boundary value problem (\ref{eq:liouville}) for smooth boundary data $g(t)>0$. Suppose both the initial data $u_0$ and sign-changing function $f$ are continuous and $u_0(\alpha)\in L^{\infty}(0,1)$. If $u_0$ and $f$ are such that $\psi_0(\alpha)\leq0$ for all $\alpha\in[0,1]$, then $0<\left\|u(\cdot,t)\right\|_{\infty}<+\infty$ for all time. Moreover, the result still holds in the case where $M_0>0$ as long as $g$ is such that
\begin{equation}
\label{eq:noblowcond}
\lim_{t\to+\infty}G(t)<\frac{2}{M_0}.
\end{equation}
\end{thm}
\begin{rem}
For $f(\alpha),u_0(\alpha)\in C^1[0,1]$, (\ref{eq:Psi}) implies that a sufficient condition for global solutions is that, for all $\alpha\in[0,1]$,
\begin{equation}
\label{eq:globalcond1}
f(\alpha)u_0'(\alpha)+f'(\alpha)u_0(\alpha)>0.
\end{equation}
Indeed, the above is equivalent to $\psi_0''>0$, which by $\psi_0(0)=\psi_0(1)=0$ implies that $M_0=0$. The reader may refer to \S\ref{sec:examples} where an example of a global solution is obtained for the case $f(\alpha)=2\alpha-1$ and $u_0(\alpha)\equiv1$. Also in \S\ref{sec:examples} we discuss a simple class of boundary data (a family of fractional linear maps) for which global solutions may be obtained for any choice of $u_0$ and $f$.
\end{rem}
\subsubsection{Finite-time $L^{\infty}$ Blowup}
\label{subsec:blowupsmoothbdry}
\hfill
\noindent
We now study finite-time blowup of solutions to \eqref{eq:liouville}. For smooth boundary data $g(t)>0$, suppose $u_0$ and $f$ are such that $M_0\in\mathbb{R}^+$, and assume $g$ is such that \eqref{eq:G} satisfies
\begin{equation}
\label{eq:blowcond}
\frac{2}{M_0}<\lim_{t\to+\infty}G(t)\leq +\infty.
\end{equation}
Since $g>0$ and $\dot G(t)=g(t)$, then by continuity \eqref{eq:blowcond} implies the existence of a finite $t_*>0$ such that
\begin{equation}
\label{eq:G*}
G(t_*)=\lim_{t\uparrow t_*}\int_0^{t}{g(s)\,ds}=\frac{2}{M_0}.
\end{equation}
More particularly, from \eqref{eq:finalsolution} we see that
\begin{equation}
\label{eq:blow1}
\lim_{t\uparrow t_*}u(\overline\alpha_i,t)=+\infty,\qquad\quad 1\leq i\leq n
\end{equation}
for
\begin{equation}
\label{eq:t*}
t_*\equiv G^{-1}\left(2/M_0\right),
\end{equation}
and where $\overline\alpha_i$ denote the finite number of locations where $M_0$ is attained. On the contrary, if $\alpha\neq\overline\alpha_i$ (so that $\psi_0(\alpha)<M_0$), $u((\alpha,t)$ will converge, as $t\uparrow t_*$, to a finite positive constant $C(\alpha)$ given by
\begin{equation}
\label{eq:blow2}
C(\alpha)=g(t_*)u_0(\alpha)\left(1-\frac{\psi_0(\alpha)}{M_0}\right)^{-2}.
\end{equation}
We summarize the above results in Theorem \ref{thm:blowuptheorem} below.
\begin{thm}
\label{thm:blowuptheorem}
Consider the initial boundary value problem (\ref{eq:liouville}) for smooth boundary data $g(t)>0$. Suppose both the initial data $u_0$ and sign-changing function $f$ are continuous and $u_0(\alpha)\in L^{\infty}(0,1)$. If $u_0$ and $f$ are such that $\psi_0$ attains its greatest value $M_0>0$ at a finite number of points $\overline\alpha_i\in(0,1),\, 1\leq i\leq n$, and if $g$ is such that \eqref{eq:blowcond} holds, then there exists a finite $t_*>0$ (given by \eqref{eq:t*}) at which $u(\overline\alpha_i,t)$ diverges as $t\uparrow t_*$. In contrast, for $\alpha\neq\overline\alpha_i$\,, $u(\alpha,t)$ converges to the finite, positive constant in (\ref{eq:blow2}).
\end{thm}
\begin{rem}
\label{rem:suff}
If there exist\, $0<\alpha_0<\alpha_1\leq\alpha_2\leq1$ such that
\begin{equation}
\label{suffblow1}
f(\alpha)u_0'(\alpha)\leq0,\quad f(\alpha_0)=0,\qquad \alpha\in[0,\alpha_1]
\end{equation}
and
\begin{equation}
\label{suffblow2}
(f(\alpha)u_0(\alpha))'\geq0,\quad f(\alpha_2)=0,\qquad \alpha\in[\alpha_1,1],
\end{equation}
then $M_0>0$. Consequently, \eqref{suffblow1}-\eqref{suffblow2} give a sufficient condition for $u(\overline\alpha_i,t)$ to blowup in finite time as long as \eqref{eq:blowcond} holds.
See \S\ref{sec:examples} for particular examples.
\end{rem}
\subsection{Singular Boundary Data}
\label{subsec:singularbdry}
In this section we study the effects of singular boundary data on the interior regularity of (\ref{eq:finalsolution}). More particularly, suppose (\ref{eq:G}) has the form
\begin{equation}
\label{eq:Gsingular}
G(t)=\frac{1}{\beta}\left(\frac{1}{(1-t)^{\beta}}-1\right),\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\beta>0,
\end{equation}
so that
\begin{equation}
\label{eq:blowcond2}
\lim_{t\uparrow 1}G(t)=+\infty.
\end{equation}
Since $\dot G(t)=g(t)$, (\ref{eq:Gsingular}) then yields
\begin{equation}
\label{eq:gsingular}
g(t)=\frac{1}{(1-t)^{1+\beta}},\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\beta>0
\end{equation}
as the induced boundary blowup rate with boundary blowup time\footnote[2]{A simple rescaling argument shows that the choice $t_b=1$ is without loss of generality.}
\begin{equation}
\label{eq:tbdry}
t_b=1.
\end{equation}
We find that if $u_0$ and $f$ are such that $M_0>0$ (e.g. they satisfy the conditions in Remark \ref{rem:suff} above), then for all $\beta>0$, $u(\overline\alpha_i,t)$ blows up at a finite time $t_*$ satisfying $0<t_*<t_b$; whereas, for $\alpha\neq\overline\alpha_i$, $u$ remains bounded and positive.
For the case $M_0=0$, define
\begin{equation}
\label{eq:omega}
\Omega\equiv\{\alpha\in[0,1]\,\,|\,\,\psi_0(\alpha)=0\},
\end{equation}
which note satisfies $\Omega\neq\emptyset$ due to $\psi_0(0)=\psi_0(1)=0$. Suppose $M_0=0$, so that $\Omega=\{\overline\alpha_i\}_{i=1}^n$. If $\alpha\in\Omega$, the induced boundary blowup time $t_b$ will represent the earliest blowup time for $u$; while, for $\alpha\notin\Omega$ (i.e. $\alpha\neq\overline\alpha_i$), the induced boundary blowup rate determines the behaviour of the solution (as well as its last configuration profile before blowup) as follows: If $\beta=1$, $u$ will stay both \textsl{finite} and \textsl{positive} as $t\uparrow t_b$, whereas, for $\beta\in(1,+\infty)$ or $\beta\in(0,1)$, $u$ will vanish or respectively diverge to $+\infty$ as $t\uparrow t_b$. Consequently, in the case $M_0=0$, we may refer to $\beta=1$ as a ``threshold'' exponent for the boundary singularity due to the drastic change in the last configuration profile of $u(\alpha,t)$ before blowup when $\beta=1\pm\epsilon$ for $\epsilon>0$ arbitrarily small. It turns out that the case $\beta=1$ corresponds to \eqref{eq:Gsingular} being fractional linear, so that its Schwarzian derivative vanishes identically, i.e. $\mathcal{S}(G;t)\equiv0$.
Now, using (\ref{eq:Gsingular}) and (\ref{eq:gsingular}) on (\ref{eq:finalsolution}), we obtain
\begin{equation}
\label{eq:usingular1.1}
u(\alpha,t)=\frac{4u_0(\alpha)}{\left(2-t(2+\psi_0(\alpha))\right)^2},\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\beta=1
\end{equation}
or
\begin{equation}
\label{eq:usingular1}
u(\alpha,t)=\frac{4\beta^2u_0(\alpha)\,(1-t)^{\beta-1}}{\left(2\beta(1-t)^{\beta}-\psi_0(\alpha)(1-(1-t)^{\beta})\right)^2},\quad\quad\beta\in\mathbb{R}^+\backslash\{1\},
\end{equation}
both of which imply
\begin{equation}
\label{eq:int2}
u(\alpha,t)=\frac{u_0(\alpha)}{(1-t)^{1+\beta}},\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\alpha\in\Omega,\,\,\,\,\,\,\,\beta\in\mathbb{R}^+,
\end{equation}
so that $u$ diverges on $\Omega$ as $t$ approaches the induced boundary blowup time $t_b=1$. However, below we see how under certain conditions, $u$ may still diverge at an earlier time somewhere on $[0,1]\backslash\Omega$.
\subsubsection{Boundary Singularity with $\beta=1$}
\hfill
\noindent
First we consider the simple case (\ref{eq:usingular1.1}), which corresponds to an induced boundary singularity of the form (\ref{eq:gsingular}) with $\beta=1$ . Since the earliest time $t_*$, satisfying $0<t_*\leq t_b$, for which (\ref{eq:usingular1.1}) diverges is obtained from
\begin{equation}
\label{eq:tsingular1}
t(\alpha)=\frac{2}{2+\psi_0(\alpha)},
\end{equation}
we find that
\begin{equation}
\label{eq:tsingular11}
t_*=
\begin{cases}
t_b=1,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,&M_0=0,
\\
\frac{2}{2+M_0}<t_b,\,\,\,\,\,\,\,\,&M_0>0.
\end{cases}
\end{equation}
Using the above on (\ref{eq:usingular1.1}) leads to the following result.
\begin{thm}
\label{thm:singular1}
Consider the initial boundary value problem (\ref{eq:liouville}) for initial data $u_0(\alpha)$ and sign-changing function $f(\alpha)$ both continuous and bounded. For $\beta=1$, suppose the boundary data $g(t)$ has the singular form (\ref{eq:gsingular}) with prescribed boundary blowup time $t_b=1$. Then for $u_0$ and $f$ such that $M_0>0$ (see e.g. \eqref{suffblow1}-\eqref{suffblow2}), there exists $0<t_*<t_b$, given by (\ref{eq:tsingular11})ii), such that
\begin{equation}
\label{eq:usingular11}
\lim_{t\uparrow t_*}u(\alpha,t)=
\begin{cases}
+\infty,\,\,\,\,\,\,\,\,\,\,\,\,\,&\alpha=\overline\alpha_i,
\\
u_0(\alpha)\left(\frac{2+M_0}{M_0-\psi_0(\alpha)}\right)^2,\,\,\,\,\,\,\,\,&\alpha\neq\overline\alpha_i.
\end{cases}
\end{equation}
On the contrary, if $u_0$ and $f$ are so that $M_0=0$, then the earliest blowup time for $u$ is the induced boundary blowup time $t_b$, and
\begin{equation}
\label{eq:usingular12}
\lim_{t\uparrow t_b}u(\alpha,t)=
\begin{cases}
+\infty,\,\,\,\,\,\,\,\,\,\,\,\,\,&\alpha=\overline\alpha_i,
\\
\frac{4u_0(\alpha)}{\psi_0(\alpha)^2},\,\,\,\,\,\,\,\,&\alpha\neq\overline\alpha_i.
\end{cases}
\end{equation}
\end{thm}
\subsubsection{Boundary Singularity with $\beta\in\mathbb{R}^+\backslash\{1\}$}
\hfill
\noindent
Next we examine the instance (\ref{eq:usingular1}), which corresponds to singular boundary data of the form (\ref{eq:gsingular}) for $\beta\in\mathbb{R}^+\backslash\{1\}$. Here the earliest time $0<t_*\leq t_b$ for which (\ref{eq:usingular1}) diverges is obtained from
\begin{equation}
\label{eq:tsingular25}
(1-t)^{\beta}=\frac{\psi_0(\alpha)}{2\beta+\psi_0(\alpha)}.
\end{equation}
This yields
\begin{equation}
\label{eq:tsingular2}
t(\alpha)=1-\left(\frac{\psi_0(\alpha)}{2\beta+\psi_0(\alpha)}\right)^{\frac{1}{\beta}}
\end{equation}
where, since we are only interested in times $0<t\leq t_b=1$, then $\alpha\in[0,1]$ in (\ref{eq:tsingular2}) must be such that
$$\frac{\psi_0(\alpha)}{2\beta+\psi_0(\alpha)}\geq0;$$
otherwise the denominator in (\ref{eq:usingular1}) either never vanishes or does so at a time greater than $t_b$.
As in the case $\beta=1$, (\ref{eq:tsingular2}) implies that $t_*=t_b=1$ whenever $\alpha\in\Omega$. Moreover, for $u_0$ and $f$ such that $M_0>0$, finite-time blowup occurs only in the interior (since $\psi_0(0)=\psi_0(1)=0$), more particularly,
\begin{equation}
\label{eq:usingular21}
\lim_{t\uparrow t_*}u(\alpha,t)=
\begin{cases}
+\infty,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,&\alpha=\overline\alpha_i,
\\
C(\alpha),\,\,\,\,\,\,\,\,&\alpha\neq\overline\alpha_i
\end{cases}
\end{equation}
for positive constants $C(\alpha)$ given by
$$C(\alpha)=\frac{u_0(\alpha)(2\beta+M_0)^{1+\frac{1}{\beta}}M_0^{1-\frac{1}{\beta}}}{(M_0-\psi_0(\alpha))^2}$$
and $t_*>0$ satisfying
\begin{equation}
\label{eq:tsingular221}
t_*=1-\left(\frac{M_0}{2\beta+M_0}\right)^{\frac{1}{\beta}}<t_b=1.
\end{equation}
Lastly, if $u_0$ and $f$ are such that $M_0=0$, then for $\alpha\in\Omega=\{\overline\alpha_i\}_{i=1}^n$, (\ref{eq:int2}) holds and
\begin{equation}
\label{eq:usingular00}
\lim_{t\uparrow t_b}u(\overline\alpha_i,t)\to+\infty,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\beta\in\mathbb{R}^+\backslash\{1\}.
\end{equation}
In this case the boundary blowup time $t_b=1$ is also the earliest blowup time; however, for $\alpha\notin\Omega$, the behaviour of the solution varies relative to the value of $\beta$. Indeed, for $\alpha\not\in\Omega$ we have that $\psi_0(\alpha)<0$, and either $2\beta+\psi_0(\alpha)<0$ or $2\beta+\psi_0(\alpha)>0$. If the former, (\ref{eq:tsingular2}) yields $t(\alpha)>t_b$, which implies that the denominator of (\ref{eq:usingular1}) remains positive and finite for all $0<t\leq t_b$. As a result, the time-dependent term in the numerator gives
\begin{equation}
\label{eq:usingular22}
\lim_{t\uparrow t_b}u(\alpha,t)=
\begin{cases}
0,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,&\beta\in(1,+\infty),\,\,\,\, \alpha\notin\Omega,
\\
+\infty,\,\,\,\,\,\,\,\,&\beta\in(0,1),\,\,\quad\,\, \alpha\notin\Omega.
\end{cases}
\end{equation}
If instead $2\beta+\psi_0(\alpha)>0$, so that $\frac{\psi_0}{2\beta+\psi_0}<0$, we use (\ref{eq:tsingular25}) and (\ref{eq:tsingular2}), as well as an argument similar to the one above, to show that $u$ diverges according to (\ref{eq:usingular00}) and (\ref{eq:usingular22}) above. We summarize these results in Theorem \ref{thm:singular2} below.
\begin{thm}
\label{thm:singular2}
Consider the initial boundary value problem (\ref{eq:liouville}) for continuous initial data $u_0(\alpha)\in L^{\infty}(0,1)$ and continuous sign-changing function $f(\alpha)$. Assume the boundary data $g(t)$ has the singular form (\ref{eq:gsingular}) for $\beta\in\mathbb{R}^+\backslash\{1\}$ and let $0<t_b<+\infty$ denote its induced blowup time. If $u_0$ and $f$ are such that $M_0>0$ (see e.g. \eqref{suffblow1}-\eqref{suffblow2}), then there exists $0<t_*<t_b$, given by (\ref{eq:tsingular221}), such that (\ref{eq:usingular21}) holds. In contrast, if $u_0$ and $f$ are so that $M_0=0$, then $u$ diverges according to (\ref{eq:usingular00})-(\ref{eq:usingular22}) as $t\uparrow t_*=t_b$ .
\end{thm}
The reader may refer to \S\ref{sec:examples} for specific examples.
\begin{rem}
From (\ref{eq:usingular00}) and (\ref{eq:usingular22}), both of which correspond to the case $M_0=0$, note that for $\beta\in(0,1)$, $u(\alpha,t)\to+\infty$ as $t\uparrow t_b$ everywhere in $[0,1]$. However, the singularity for $\alpha\in\Omega$ is ``stronger'' in the sense that, for $t_b-t>0$ small and $C\in\mathbb{R}^+$,
$$\frac{u(\alpha,t)}{u(\overline\alpha_i,t)}\sim C(t_b-t)^{2\beta},\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\alpha\neq\overline\alpha_i,$$
which vanishes as $t\uparrow t_b$.
\end{rem}
\subsection{Further $L^p(0,1)$ Regularity Results}
\label{subsec:lpsmooth}
In Theorem \ref{thm:blowuptheorem} of \S \ref{subsec:blowupsmoothbdry}, we established simple criteria, in terms of the initial and boundary data, as well as the sign-changing function $f$, for the existence of solutions to (\ref{eq:liouville}) which blowup in finite time in the $L^{\infty}(0,1)$ norm. In this section, we show that for $f$ in a large class of both smooth and non-smooth functions, $\left\|u(\cdot,t)\right\|_p\to+\infty$ as $t\uparrow t_*$ for all $1\leq p<+\infty$. More particularly, suppose $u_0$ and $f$ are such that $\psi_0(\alpha)$ attains a greatest positive value $M_0$ somewhere in $(0,1)$ and let $g(t)\in C^0[0,t_*]$ be such that \eqref{eq:blowcond} holds. Recall that $t_*>0$ denotes the finite $L^{\infty}(0,1)$ blowup time for $u$ satisfying \eqref{eq:G*}.
Now suppose there is $q\in\mathbb{R}^+$, $C_1\in\mathbb{R}^-$ and $r>0$ small, such that
\begin{equation}
\label{eq:local}
\psi_0(\alpha)\sim M_0+C_1\left|\alpha-\overline\alpha\right|^q
\end{equation}
for $0\leq\left|\alpha-\overline\alpha\right|\leq r$. To simplify subsequent arguments, we will assume that $M_0>0$ occurs at a single location $\overline\alpha\in(0,1)$. Moreover, in (\ref{eq:local}) we use the notation
\begin{equation}
\label{eq:simexplanation}
k(\alpha) \sim L + h(\alpha),
\end{equation}
valid for $0\leq|\alpha-\beta|\leq r$, to signify the existence of a function $l(\alpha)$ defined on $(\beta-r,\beta+r)$ such that
\begin{equation}
\label{eq:simexplanation2}
k(\alpha)-L=h(\alpha)(1+l(\alpha))\,\,\,\,\,\,\,\,\,\,\,\,\text{where}\,\,\,\,\,\,\,\,\,\,\,\,\,\,\lim_{\alpha\rightarrow\beta}l(\alpha)=0.
\end{equation}
Note that for $0<q<1$, (\ref{eq:local}) induces a ``cusp'' on the graph of $\psi_0$ at $\overline\alpha$, a ``kink'' if $q=1$, and various degrees of continuity on its derivatives at $\overline\alpha$ when $q>1$. Moreover, the boundedness of $u_0>0$, along with (\ref{eq:local}) and $\psi_0'=fu_0$, implies that
\begin{equation}
\label{eq:local2}
f(\alpha)\sim qC_1(\alpha-\overline\alpha)\left|\alpha-\overline\alpha\right|^{q-2}
\end{equation}
for $0\leq\left|\alpha-\overline\alpha\right|\leq r$. From (\ref{eq:local2}), observe that $f$ is continuous at $\overline\alpha$ if $q>1$, while a jump discontinuity of finite or infinite magnitude at this location will exist when $q=1$ or $0<q<1$, respectively. In any event, the solution formula (\ref{eq:finalsolution}) remains valid due to the integral term $\psi_0$ being continuous for all $\alpha\in[0,1]$ and $q>0$.
We now establish the following Theorem.
\begin{thm}
\label{thm:lp}
Consider the initial boundary value problem (\ref{eq:liouville}) for smooth initial data $u_0(\alpha)$ and let $t_*>0$ denote the finite $L^{\infty}(0,1)$ blowup time for $u$ established in Theorem \ref{thm:blowuptheorem}. Suppose the sign-changing function $f(\alpha)$ satisfies (\ref{eq:local2}) for $1/2<q<+\infty$, while \eqref{eq:blowcond} holds for the prescribed boundary data $g(t)\in C^0[0,t_*]$. Further, let $u_0$ and $f$ be such that $\psi_0$ in \eqref{eq:Psi} attains its greatest value $M_0>0$ at a finite number of points $\alpha_i\in(0,1)$, $1\leq i\leq n$. Then $\left\|u(\cdot,t)\right\|_p\to+\infty$ as $t\uparrow t_*$ for all $1\leq p\leq+\infty$.
\end{thm}
\begin{proof}
Applying Jensen's inequality to (\ref{eq:finalsolution}) implies that
\begin{equation}
\label{eq:jensen}
\left\|u(\cdot,t)\right\|_p\geq\int_0^1{\frac{u_0(\alpha)g(t)}{\left(1-\frac{1}{2}G(t)\psi_0(\alpha)\right)^2}d\alpha}
\end{equation}
for $1\leq p<+\infty$. But from (\ref{eq:local}),
$$\epsilon+M_0-\psi_0(\alpha)\sim \epsilon+\left|C_1\right|\left|\alpha-\overline\alpha\right|^q$$
for $0\leq\left|\alpha-\overline\alpha\right|\leq r$ and $\epsilon>0$ small. Consequently,
\begin{equation*}
\begin{split}
&\int_{\overline\alpha-r}^{\overline\alpha+r}{\frac{d\alpha}{(\epsilon+M_0-\psi_0(\alpha))^2}}\sim\int_{\overline\alpha-r}^{\overline\alpha+r}{\frac{d\alpha}{(\epsilon+\left|C_1\right|\left|\alpha-\overline\alpha\right|^q)^2}}
\\
&=\epsilon^{-2}\left[\int_{\overline\alpha-r}^{\overline\alpha}{\left(1+\frac{\left|C_1\right|}{\epsilon}\left(\overline\alpha-\alpha\right)^q\right)^{-2}d\alpha}+\int_{\overline\alpha}^{\overline\alpha+r}{\left(1+\frac{\left|C_1\right|}{\epsilon}\left(\alpha-\overline\alpha\right)^q\right)^{-2}d\alpha}\right].
\end{split}
\end{equation*}
Making the change of variables
$$\sqrt{\frac{\left|C_1\right|}{\epsilon}}(\overline\alpha-\alpha)^{\frac{q}{2}}=\tan\theta,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\sqrt{\frac{\left|C_1\right|}{\epsilon}}(\alpha-\overline\alpha)^{\frac{q}{2}}=\tan\theta$$
in the first and respectively second integral inside the bracket, we find that
\begin{equation}
\label{eq:general2}
\begin{split}
&\int_{\overline\alpha-r}^{\overline\alpha+r}{\frac{d\alpha}{(\epsilon+M_0-\psi_0(\alpha))^2}}\sim\frac{4\,\epsilon^{\frac{1}{q}-2}}{q\left|C_1\right|^{\frac{1}{q}}}\int_0^{\frac{\pi}{2}}{\frac{(\cos\theta)^{^{3-\frac{2}{q}}}}{(\sin\theta)^{^{1-\frac{2}{q}}}}d\theta}
\end{split}
\end{equation}
for small $\epsilon>0$. Suppose $q>1/2$. Then setting $\epsilon=\frac{2}{G}-M_0$ in (\ref{eq:general2}) implies that
\begin{equation}
\label{eq:general3}
\begin{split}
\int_{0}^{1}{\frac{d\alpha}{\left(1-\frac{1}{2}G(t)\psi_0(\alpha)\right)^2}}\sim C\left(G(t_*)-G(t)\right)^{^{\frac{1}{q}-2}}
\end{split}
\end{equation}
for $G(t_*)-G(t)>0$ small, $G(t_*)=\frac{2}{M_0}$ and $C\in\mathbb{R}^+$ given by
\begin{equation}
\label{eq:generalcst}
\begin{split}
C=\frac{8}{M_0^2}\left(\frac{M_0^2}{2\left|C_1\right|}\right)^{\frac{1}{q}}\Gamma\left(1+\frac{1}{q}\right)\Gamma\left(2-\frac{1}{q}\right)
\end{split}
\end{equation}
with $\Gamma\left(\cdot\right)$ the standard gamma function. We remark that the constant (\ref{eq:generalcst}) has been obtained via the identity
\begin{equation}
\label{eq:gammarel2}
\begin{split}
2\int_0^{\frac{\pi}{2}}{\frac{(\cos\theta)^{^{3-\frac{2}{q}}}}{(\sin\theta)^{^{1-\frac{2}{q}}}}d\theta}=q\,\Gamma\left(1+\frac{1}{q}\right)\Gamma\left(2-\frac{1}{q}\right),\,\,\,\,\,\,\,\,\,\,\,\,2>\frac{1}{q},
\end{split}
\end{equation}
which follows from well-known properties of the beta function. Then using (\ref{eq:general3}) on (\ref{eq:jensen}) yields
\begin{equation}
\label{eq:lower1}
\begin{split}
\left\|u(\cdot,t)\right\|_p\geq\int_0^1{\frac{u_0(\alpha)g(t)}{\left(1-\frac{1}{2}G(t)\psi_0(\alpha)\right)^2}d\alpha}\sim \frac{Cg(t_*)m_0}{\left(G(t_*)-G\right)^{2-\frac{1}{q}}}
\end{split}
\end{equation}
for $G(t_*)-G(t)>0$ small, $q>1/2$ and where, as a result of the boundedness and continuity of $u_0$ and $g$ for $\alpha\in[0,1]$ and respectively $t\in[0,t_*]$, both $m_0=\min_{\alpha\in[0,1]}u_0(\alpha)$ and $g(t_*)$ are finite, positive constants. Taking the limit as $t\uparrow t_*$ (so that by continuity $G(t)\uparrow G(t_*)$) in (\ref{eq:lower1}) yields
\begin{equation}
\label{eq:lower2}
\begin{split}
\lim_{t\uparrow t_*}\left\|u(\cdot,t)\right\|_p=+\infty
\end{split}
\end{equation}
for all $1\leq p\leq+\infty$.
\end{proof}
\section{A Generalized Sign-changing Liouville Equation}
\label{sec:generalization}
In this section we study regularity of solutions to the following generalization of \eqref{eq:liouville}:
\begin{equation}
\label{u}
\begin{cases}
\partial_{\alpha t}\ln u=f(\alpha)\mathcal{F}(u),\,\,\,\,\,\,\,\,\,\,\,&\alpha\in(0,1),\,\,\,t>0,
\\
u(\alpha,0)=u_0(\alpha),\,\,\,\,\,\,\,\,\,\,\,\,&\alpha\in[0,1],
\\
u(0,t)=u(1,t)=g(t),\,\,\,\,\,\,\,\,\,\,\,\,\,&t\geq0,
\end{cases}
\end{equation}
for $\mathcal{F}(z)$ an arbitrary differentiable function of $z$.
We will be particularly interested in the cases where the prescribed smooth boundary data $g(t)>0$ is either a non-decreasing function of time, $\dot g\geq0$, or a non-increasing one, $\dot g\leq0$. We begin by establishing the following blowup result for the former:
\begin{thm}
\label{general}
Consider the initial boundary value problem \eqref{u} for smooth initial data $u_0>0$ and smooth boundary data $g(t)>0$ satisfying $\dot g(t)\geq0$. Let $\alpha_0$ denote the first location in $(0,1)$ where the sign-changing function $f$ vanishes and assume there are positive constants $c$ and $d$ such that $\mathcal{F}(u)\in C^1(0,+\infty)$ satisfies
\begin{equation}
\label{ass}
0\leq\mathcal{F}(u),\quad\qquad c\mathcal{F}(u)\leq u\mathcal{F}'(u)\leq d\mathcal{F}(u)
\end{equation}
for $ '=\frac{d}{du}$. Then $u\to+\infty$ earliest at $\alpha=\alpha_0$ as $t$ approaches the finite time $t^*(\alpha_0)$ in \eqref{blowtime}.
\end{thm}
\begin{proof}
Let $\alpha_0$ be the first zero of $f(\alpha)$ in $(0,1)$. This is guaranteed to exist due to periodicity of $u$ and \eqref{ass}i), which imply that $\int_0^1{f(\alpha)\mathcal{F}(u)\,d\alpha}\equiv0$. More particularly, and without loss of generality, suppose
\begin{equation}
\label{fass}
f(\alpha)
\begin{cases}
>0,\qquad &\alpha\in[0,\alpha_0),
\\
=0,\qquad\qquad &\alpha=\alpha_0,
\\
<0,\qquad &\alpha\in(\alpha_0,1].
\end{cases}
\end{equation}
Define
\begin{equation}
\label{H}
H(\alpha,t)\equiv(\ln u)_t\big|_0^{\alpha}=\frac{\dot u(\alpha,t)}{u(\alpha,t)}-\frac{\dot g(t)}{g(t)}
\end{equation}
and note that, as a result of \eqref{u}i) and \eqref{fass},
\begin{equation}
\label{Hpos}
H(\alpha,t)>0\qquad\quad \alpha\in(0,\alpha_0].
\end{equation}
Now, a straight-forward computation shows that $H$ satisfies
\begin{equation}
\label{Heq}
H_t(\alpha,t) = \int_0^\alpha f(x)\mathcal{F}'(u)u(x,t)H(x,t)\,dx + \frac{\dot{g}(t)}{g(t)}\int_0^\alpha f(x)\mathcal{F}'(u)u(x,t)\,dx.
\end{equation}
Suppose the boundary data is such that $\dot g\geq0$. Then using \eqref{ass}, \eqref{H} and \eqref{Hpos}, and subsequently \eqref{u}i) on the right-hand side of \eqref{Heq}, we obtain
\begin{equation}
\label{Heq2}
H_t\geq\frac{c}{2}\left(\frac{\dot u}{u}\right)^2-\frac{c}{2}\left(\frac{\dot g}{g}\right)^2
\end{equation}
for $\alpha\in(0,\alpha_0]$. Since $\dot g/g\geq0$, \eqref{Heq2} then yields
\begin{equation}
\label{Heq3}
H_t\geq\frac{c}{2}H^2,\qquad\quad \alpha\in(0,\alpha_0],
\end{equation}
which we integrate to obtain
\begin{equation}
\label{Heq4}
0< \frac{1}{H(\alpha,t)}\leq\frac{1}{H_0(\alpha)}-\frac{c}{2}t,\qquad\quad \alpha\in(0,\alpha_0]
\end{equation}
where $H_0(\alpha)=H(0,\alpha)$. From \eqref{Heq4} and smoothness of $g$, we infer that $\dot u/u\to+\infty$ as $t$ approaches $t^*$ defined by
\begin{equation}
\label{tstar}
t^*\equiv\frac{2}{cH^*},\qquad\qquad H^*\equiv\max_{\alpha\in(0,\alpha_0]}H_0(\alpha).
\end{equation}
However, \eqref{u}i) implies that
\begin{equation}
\label{H0}
H_0(\alpha)=\int_0^{\alpha}{f(x)\mathcal{F}(u_0)\,dx}>0,\quad\qquad \alpha\in(0,\alpha_0],
\end{equation}
from which we conclude, by \eqref{fass}, that
\begin{equation}
\label{blowtime}
t^*=\frac{2}{cH_0(\alpha_0)}.
\end{equation}
Thus $\dot u/u$ will diverge earliest, as $t\uparrow t^*$, at $\alpha=\alpha_0$. Lastly, for $t\in[0,t^*)$ and $\alpha\in(0,\alpha_0]$, \eqref{Heq4} implies that
$$\partial_t\ln u\geq \partial_t\left[\ln g-\frac{2}{c}\ln\left(\frac{1}{H_0(\alpha)}-\frac{c}{2}t\right)\right],$$
which yields, upon integration and some simplification,
\begin{equation}
\label{blowincreasing}
u(\alpha,t)\geq\frac{g(t)u_0(\alpha)}{\left(1-\frac{c}{2}H_0(\alpha)t\right)^{2/c}}.
\end{equation}
From the above we infer that
$$\lim_{t\uparrow t^*}u(\alpha_0,t)=+\infty.$$
\end{proof}
Last we establish sufficient conditions for finite-time blowup or global-in-time existence of $u$ on $[0,\alpha_0]$ in the case $\dot g\leq0$.
\begin{thm}
\label{general2}
Consider the initial boundary value problem \eqref{u} for smooth initial data $u_0>0$ and smooth boundary data $g(t)>0$ satisfying $\dot g(t)\leq0$. Let $\alpha_0$ denote the first location in $(0,1)$ where the sign-changing function $f$ vanishes and assume there are positive constants $c$ and $d$ such that $\mathcal{F}(u)\in C^1(0,+\infty)$ satisfies \eqref{ass}. Then the following hold:
\begin{enumerate}
\item If $g$ is such that
\begin{equation}
\label{gcond2}
\begin{split}
\lim_{t\to+\infty}\int_0^t{g(s)^dds}>\frac{2}{cH_0(\alpha_0)},
\end{split}
\end{equation}
then there exists a finite $t_*>0$ such that $u\to+\infty$ earliest at $\alpha=\alpha_0$ as $t\uparrow t_*$.
\vspace{0.1in}
\item If $g$ satisfies
\begin{equation}
\label{gcond3}
\begin{split}
\lim_{t\to+\infty}\int_0^t{g(s)^{c}ds}\leq\frac{2}{dH_0(\alpha_0)},
\end{split}
\end{equation}
then $u$ persists globally in time.
\end{enumerate}
\end{thm}
\begin{proof}
Without loss of generality we once again assume $f$ satisfies \eqref{fass}. First we obtain an upper bound for $u$. Since $\dot g\leq0$, using \eqref{u}i), \eqref{ass}, \eqref{fass} and \eqref{H}, on \eqref{Heq}, we obtain
\begin{equation}
\label{global0}
H_t(\alpha,t)\leq
c\frac{\dot g}{g}H(\alpha,t)+\frac{d}{2}H(\alpha,t)^2
\end{equation}
for $\alpha\in[0,\alpha_0]$. Rewriting the above as
\begin{equation}
\label{global1}
\partial_t\left(\frac{g^c}{H}\right)\geq-\frac{d}{2}g^c
\end{equation}
and integrating the latter yields
\begin{equation}
\label{ineqlower}
\frac{g(t)^c}{H(\alpha,t)}\geq\frac{1}{H_0(\alpha)}-\frac{d}{2}\int_0^t{g(s)^cds}
\end{equation}
for $\alpha\in(0,\alpha_0]$ and $H_0$ as in \eqref{H0}. In the above we also used the simplifying assumption $g(0)=1$. Multiplying both sides of \eqref{ineqlower} by $H$ and using \eqref{H}, we may integrate the resulting inequality in time to obtain
\begin{equation}
\label{uppergen}
u(\alpha,t)\leq\frac{g(t)u_0(\alpha)}{\left(1-\frac{d}{2}H_0(\alpha)\int_0^t{g(s)^{c}ds}\right)^{2/d}}\,,
\end{equation}
which is valid on $[0,\alpha_0]$ and for as long as
\begin{equation}
\label{tinterval1}
\int_0^t{g(s)^{c}ds}<\frac{2}{dH_0(\alpha_0)}.
\end{equation}
Moreover, using \eqref{u}i), \eqref{ass}, \eqref{H} and $\dot g\leq0$ on \eqref{Heq}, leads to
\begin{equation}
\label{eqgen2}
\begin{split}
H_t(\alpha,t)\geq
d\frac{\dot g}{g}H(\alpha,t)+\frac{c}{2}H(\alpha,t)^2
\end{split}
\end{equation}
for all $\alpha\in[0,\alpha_0]$. Then proceeding as above we obtain
\begin{equation}
\label{eqgen3}
\begin{split}
0<\frac{g^d}{H(\alpha,t)}\leq\frac{1}{H_0(\alpha)}-\frac{c}{2}\int_0^t{g(s)^dds},
\end{split}
\end{equation}
from which we derive the lower-bound
\begin{equation}
\label{lowergen}
u(\alpha,t)\geq\frac{g(t)u_0(\alpha)}{\left(1-\frac{c}{2}H_0(\alpha)\int_0^t{g(s)^dds}\right)^{2/c}}.
\end{equation}
Inequality \eqref{lowergen} holds for all $\alpha\in[0,\alpha_0]$ and as long as
\begin{equation}
\label{tinterval2}
\int_0^t{g(s)^{d}ds}<\frac{2}{cH_0(\alpha_0)}.
\end{equation}
First suppose the smooth, non-increasing boundary data $g(t)>0$ is such that \eqref{gcond2} holds for some $d\in\mathbb{R}^+$. Then by continuity of $g$ there exists a finite $t_*>0$ such that
\begin{equation}
\label{reqblow}
\lim_{t\uparrow t_*}\int_0^t{g(s)^dds}=\frac{2}{cH_0(\alpha_0)}
\end{equation}
and, thus,
$$\lim_{t\uparrow t_*}u(\alpha_0,t)=+\infty$$
by \eqref{lowergen}. This establishes the first part of the Theorem. Note that no conflict arises between this blowup and the upper-bound in \eqref{uppergen}. Indeed, since $d>c$ and $\dot g\leq0$ we have that
\begin{equation}
\label{condition}
\int_0^t{g(s)^dds}\leq\int_0^t{g(s)^{c}ds}\quad\qquad\text{and}\qquad\quad \frac{2}{dH_0(\alpha_0)}<\frac{2}{cH_0(\alpha_0)}.
\end{equation}
Consequently, \eqref{reqblow}, \eqref{condition} and continuity of $g$ imply the existence of $0<t_1<t_*$ such that the right-hand side of \eqref{uppergen} diverges as $t\uparrow t_1$.
For the last part of the Theorem, suppose $g$ satisfies \eqref{gcond3}. Then \eqref{uppergen}, \eqref{lowergen} and \eqref{condition} imply that $u$ remains finite and positive for all $t\in\mathbb{R}^+$. This concludes the proof of the Theorem.
\end{proof}
\begin{rem}
A simple example representative of the blowup result in Theorem \ref{general2} is given by $g(t)\geq e^{-kt}$ for $k\in\mathbb{R}^+$ fixed. In this case the right-hand side of \eqref{eqgen3} is bounded above by $1/H_0 -c (1-e^{-kdt})/2kd$, which reaches zero in finite time provided $H_0>2kd/c$.
\end{rem}
\section{Examples}
\label{sec:examples}
\subsection{Example 1 - Global Solution and Smooth Boundary Data}
Let $u_0(\alpha)\equiv1$, $f(\alpha)=2\alpha-1$ and $g(t)=2t+1$. Then (\ref{eq:Psi}) and (\ref{eq:G}) give
$$\psi_0(\alpha)=\alpha^2-\alpha,\,\,\,\,\,\,\,\,\,\,\,\,G(t)=t^2+t.$$
Note that $\psi_0\leq0$, so that $M_0=0$. Using (\ref{eq:finalsolution}), we obtain
\begin{equation}
\label{eq:globalex1}
u(\alpha,t)=\frac{2t+1}{\left(1-\frac{\alpha t}{2}(t+1)(\alpha-1)\right)^2}.
\end{equation}
The solution remains finite, and positive, for all $\alpha\in[0,1]$ and $0\leq t<+\infty$, whereas
\begin{equation}
\label{eq:globalex2}
\lim_{t\to+\infty}u(\alpha,t)=
\begin{cases}
0,\,\,\,\,\,\,\,\,\,\,\,\,\,\,&\alpha\in(0,1),
\\
+\infty,\,\,\,\,\,\,\,\,\,\,&\alpha\in\{0,1\}.
\end{cases}
\end{equation}
See Figure (\ref{fig:ex})(A) below.
\subsection{Example 2 - Finite-time Blow-up and Smooth Boundary Data}
Consider the same initial and boundary data as in Example 1, but now take $f(\alpha)=1-2\alpha$. Then (\ref{eq:Psi}) becomes $\psi_0(\alpha)=\alpha-\alpha^2$ with $M_0=1/4$ attained at $\overline\alpha_1=1/2$. The solution (\ref{eq:finalsolution}) now reads
\begin{equation}
\label{eq:blowex2}
u(\alpha,t)=\frac{2t+1}{\left(1-\frac{\alpha t}{2}(t+1)(1-\alpha)\right)^2}.
\end{equation}
Since $G(t)=t^2+t$, we solve $G(t)=8$ and find that $t_*=\frac{1}{2}\left(\sqrt{33}-1\right)\sim2.37.$ Using (\ref{eq:blowex2}) we conclude that
\begin{equation}
\label{eq:blowex22}
\lim_{t\to t_*}u(\alpha,t)=
\begin{cases}
+\infty,\,\,\,\,\,\,\,\,\,\,\,\,\,\,&\alpha=\overline\alpha_1,
\\
\frac{\sqrt{33}}{(1-2\alpha)^4},\,\,\,\,\,\,\,\,\,\,&\alpha\neq\overline\alpha_1.
\end{cases}
\end{equation}
See Figure (\ref{fig:ex})(B) below.
\begin{center}
\begin{figure}[!ht]
\includegraphics[scale=0.34]{Global11.pdf}
\includegraphics[scale=0.36]{Blow1.pdf}
\caption{For Example 1 above, Figure A depicts the global-in-time behaviour of (\ref{eq:globalex1}), while B represents finite-time blowup of $u$ in (\ref{eq:blowex2}) as $t\uparrow t_*\sim2.37$.}
\label{fig:ex}
\end{figure}
\end{center}
The following two Examples are instances of Theorem \ref{thm:singular1}.
\subsection{Example 3 - Induced Boundary Blow-up for $M_0=0$}
For $u_0$ and $f$ as in Example 1, let $g(t)=(1-t)^{-2}$. This implies that $M_0=0$ occurs only at the boundary points. Then using (\ref{eq:usingular1.1}), we obtain
\begin{equation}
\label{eq:ex3}
u(\alpha,t)=\frac{4}{(2-t(\alpha^2-\alpha+2))^2},
\end{equation}
which diverges earliest, at the boundary points $\overline\alpha_i=\{0,1\}$, as $t\uparrow t_b=1$. In contrast, for $\alpha\in(0,1)$, $u$ remains finite (and positive) for all $t\in[0,1]$. In fact,
$$\lim_{t\uparrow 1}u(\alpha,t)=\frac{4}{(\alpha^2-\alpha)^2},\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\alpha\in(0,1).$$ See Figure \ref{fig:ex4}(A) below.
\subsection{Example 4 - Earlier Interior Blowup for $M_0>0$ with Singular $g(t)$}
Take $u_0$ and $f$ as in Example 2 and $g(t)=(1-t)^{-2}$. Then we now have $M_0=1/4$ occurring at $\overline\alpha_1=1/2$. From (\ref{eq:usingular1.1}), we obtain
\begin{equation}
\label{eq:ex4}
u(\alpha,t)=\frac{4}{(2+t(\alpha^2-\alpha-2))^2},
\end{equation}
which diverges earliest at $\alpha=\overline\alpha_1$ as
$$t\uparrow t_*=\frac{2}{2+M_0}=\frac{8}{9}<t_b=1.$$
In contrast, for $\alpha\neq\overline\alpha_1$, $u$ remains finite and positive for all $t\in[0,t_*]$. In this case, the final solution profile is given by
$$\lim_{t\uparrow t_*}u(\alpha,t)=\left(\frac{9}{1-4\alpha+4\alpha^2}\right)^2,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\alpha\in[0,1]\backslash\{\overline\alpha_1\}.$$ See Figure \ref{fig:ex4}(B) below.
\begin{center}
\begin{figure}[!ht]
\includegraphics[scale=0.32]{singglob1.pdf}
\includegraphics[scale=0.32]{singsing1.pdf}
\caption{Blowup profiles for Examples 4 and 5 with singular boundary data. Figure A represents earliest blowup of (\ref{eq:ex3}) at the boundary as $t\uparrow t_b=1$ in the case where $M_0=0$, whereas, for $M_0>0$, Figure B depicts earliest blowup of (\ref{eq:ex4}) in the interior as $t\uparrow t_*<t_b$.}
\label{fig:ex4}
\end{figure}
\end{center}
|
\section{Introduction}
The Survey of Income and Program Participation (SIPP) is the largest government survey of people
on public assistance in the United States. It includes longitudinal
data on income, labor force information, participation and eligibility
for governmental assistance programs, and general demographic
characteristics for individuals on public assistance; as such, it is used by a broad community of researchers and policy-makers \citep{kinnerreiterjos}.
In 2014, the Census Bureau redesigned the SIPP to utilize a longer
reference period (twelve months, instead of four) and a new instrument
that incorporates an event history calendar \citep{Moore2009}. The
Census Bureau made these changes with the hope of reducing costs and
respondent burden while improving accuracy.
To evaluate the redesign, the Census Bureau conducted a field test by giving the new survey to a non-overlapping sample of individuals drawn from the
same frame used to construct the 2008 production SIPP panel. The field test was restricted to individuals in low income strata in 20 states.
The Census Bureau also constructed a comparison dataset from the production SIPP panel comprising individuals from the same strata and states, with the intention of
assessing whether or not the change in collection instruments resulted in different distributions of key variables. Additional details
are available in \cite{USCensus2013}.
The data from the field test suffered from item nonresponse, as did the data
from the production SIPP. For example, among the sampled field
test individuals, approximately 16\% are missing employment status and, for those who reported participation in the Supplemental Nutrition Assistance Program (SNAP),
approximately 59\% are missing the benefit amounts. Unless the
missing data mechanisms are identical in both datasets,
e.g., both missing completely at random \citep{Rubin1976},
comparisons of available case analyses may result in inaccurate
conclusions about where estimates from the old and new designs differ.
Given the objective of comparing two datasets on many
analyses, a sensible approach is to create and utilize
multiply-imputed versions \citep{Rubin1987} of each sample. In
multiple imputation (MI) the imputer repeatedly samples values of the
missing data from their predictive distribution under an appropriate
model to create $m>1$ completed datasets. The analyst
then computes point and variance estimates in each of the $m$
datasets, and combines them using straightforward rules
\citep{Rubin1987, reiter:raghu:07}. These rules allow the analyst to account for
uncertainty due to the missing data when making inferences.
The SIPP data have distributional features that are
challenging to capture with imputation based on
standard (semi-)parametric models.
For example, some continuous variables have different variances and skewness at
different combinations of the categorical variables, and the
categorical variables have complex dependencies.
Thus, it is desirable to use imputation models that can capture such features in
each dataset with minimal tuning.
In this article, we introduce a nonparametric Bayesian joint
model for mixed continuous and categorical data suitable for use as a
flexible, fully coherent multiple imputation engine. The basic idea
is to
fuse two Dirichlet Process (DP) mixtures: a mixture of multinomial distributions (for the categorical data) and a mixture of multivariate normal regressions (for the continuous data, conditionally on the categorical variables).
We model dependence between the
categorical and continuous variables by (i)
specifying the means of the normal distributions
as component-specific functions of the categorical variables, and (ii) inducing dependence in the separate component assignments via a
hierarchical model. As we illustrate, the model includes
local dependence---i.e., dependence among variables within mixture components---between the categorical and continuous data;
thus, we call it a hierarchically coupled mixture model with local
dependence (HCMM-LD). Local dependence allows the model to more efficiently capture complex dependence structure in observed variables.
This ability to capture complex dependence, as well
as conform to different distributional shapes, is attractive for
multiple imputation contexts, as it helps the imputer to preserve
structure in the data that he or she may not have anticipated but may
be important to analysts. Mixture models have been suggested previously for multiple imputation of missing
categorical data \citep[e.g.,][]{Vermunt2008, Gebregziabher2010, Si2013,
Manrique-Vallier2012, Manrique-Vallier2012a, si:reiter:hillygus:14} and missing continuous data \citep[e.g.,][]{Bohning2007,
Elliott2007, Kim2013}. To our knowledge, mixture models have not been used to impute mixed categorical and continuous data.
The remainder of the article proceeds as follows. In Section
\ref{sec:motivation}, we illustrate some of the complex
distributional features of the variables in the SIPP, and we discuss
how existing multiple imputation routines could struggle to capture
such features. In Section \ref{sec:model}, we introduce the HCMM-LD\
including specification of prior distributions, and discuss related
nonparametric Bayesian models. %
In Section \ref{sec:sipp-sim}, we present results of a repeated sampling
simulation using complete cases from an existing SIPP panel,
illustrating the potential for improved performance of multiple imputation using the
HCMM-LD\ over a default application of MI by chained
equations \citep{Raghunathan2001,VanBuuren1999}. In Section
\ref{sec:msipp}, we apply the HCMM-LD\ to multiply-impute missing values
in the field test data as well as the representative subsample of the production SIPP. Some
conclusions about the comparability of the two designs change
after accounting for the missing data. Finally, in Section
\ref{sec:conclusion}, we conclude with a discussion of extensions and
future work.
\section{The Challenges of Imputing SIPP Data}\label{sec:motivation}
\begin{figure}
\centering
\includegraphics[width=.9\textwidth]{./fig/earn_by_hrs_edu.pdf}
%
\caption{Log monthly earnings from SIPP, by usual hours worked and education level}
\label{fig:earn_by_hrs_edu}
\end{figure}
SIPP is characteristic of many surveys in that it includes many
categorical variables and a smaller number of continuous variables, with
complicated dependence and nonstandard distributions. For example, using public-use data from the 2008 SIPP panel, Figure \ref{fig:earn_by_hrs_edu} displays
plots of (log) total earnings by usual hours worked and education level. The
distribution of income varies across levels of the discrete variables.
The earnings distribution is skewed right when usual hours $\leq 10$, whereas it
eventually becomes slightly skewed left as the number of hours increases. In the
first three panels, increasing education level is associated with increased
dispersion in the distribution of log earnings. In the last panel, increased
education is primarily associated with a location shift in earnings.
There is also evidence of higher-order dependence in the distributions
of the categorical variables. Table \ref{tab:loglm} shows analysis of deviance
tables for one, two, and three way loglinear models fit to a few
subsets of the categorical variables. All indicate some evidence of
interactions.
\begin{table}[t]
\caption{Analysis of deviance tables for loglinear models fit to subsets of the SIPP data.}\label{tab:loglm}
\centering
\begin{tabular}{lrrrrr}
\multicolumn{6}{c}{Race, sex, education level, hourly}\\
\hline
& Resid. Df & Resid. Dev & Df & Deviance & Pr($>$Chi) \\
\hline
\multicolumn{6}{l}{Race, sex, education level, hourly}\\
$\,\,\,$1 way & 108 & 10191.25 & & & \\
$\,\,\,$ 2 way & 69 & 211.50 & 39 & 9979.76 & $<10^{-6}$ \\
$\,\,\,$ 3 way & 20 & 39.92 & 49 & 171.58 & $<10^{-6}$ \\
& & & & & \\
\multicolumn{6}{l}{Marital status, usual hours, sex, no. own children}\\
$\,\,\,$1 way & 132 & 6350.60 & & & \\
$\,\,\,$ 2 way & 91 & 1383.23 & 41 & 4967.36 & $<10^{-6}$ \\
$\,\,\,$ 3 way & 30 & 168.83 & 61 & 1214.40 & $<10^{-6}$ \\
\hline
\end{tabular}
\end{table}
Given these complex distributional features, what sort of models might
one use for imputation of missing values? One possible approach is to use
a general location model (GLOM) \citep{Olkin1961,Little1985,Schafer1997}. For
continuous variables $Y$ and discrete variables $X$, the GLOM assumes that $(Y\mid X=x)\sim
N(\mu_x, \Sigma_x)$ and $X\sim \pi$ with $\pi\sim Dir(\alpha)$; see
also \cite{Liu1998} who generalize the $(Y\mid X)$ model to the class
of elliptically symmetric distributions. Estimation under this model
is infeasible unless each cell of the table implied by $X$ contains a large number of
observations. Thus, imputers typically impose further constraints,
most often that $\Sigma_x\equiv \Sigma$ for all $x$,
$\mu_x = D(x)B$ for a matrix of regression coefficients $B$ and design
vector $D(x)$, and $\pi$ satisfies loglinear constraints that include
interactions only up to a certain order. Multivariate
normality and common covariance structure seem unlikely to fit the
types of features apparent in Figure \ref{fig:earn_by_hrs_edu}.
Further, Table \ref{tab:loglm} suggests that it would be easy to
miss key interactions when selecting the loglinear model. Thus,
the GLOM seems overly restrictive for the SIPP data.
An alternative approach is to specify a sequence of univariate models
for each variable subject to missingness conditional on subsets of
the other variables, e.g., impute $a$ from $f(a \mid b,c)$, impute $b$
from $f(b \mid a,c)$, and impute $c$ from $f(c \mid a,b)$. This is
known as the ``chained equations'' or ``fully-conditional'' approach
\citep{Raghunathan2001,VanBuuren1999}.
While multiple imputation by chained equations (MICE) approaches have proven to be quite useful for
many datasets, they can be challenging to use
effectively for data with complex dependence like the
SIPP. For example, typical applications of MICE use
multinomial logistic regressions for the categorical
variables. Relationships between an outcome and the
remaining predictors may be nonlinear and involve interaction effects;
these can be difficult to find when the data have more than a handful
of variables (that may also be subject to missingness). Similar challenges arise when specifying models for continuous data, even with semiparametric extensions like predictive mean matching \citep{Little1988}.
Additionally, %
the selected conditional models may be incompatible; that is, there may not be any joint model with the specified conditionals \citep{Liu2014}.
This may result in imputation procedures with undesirable theoretical properties \citep{Si2013}.
A third and related approach is to specify a coherent joint distribution as a
sequence of conditional models \citep{ibrahimlips, ibrahim, ibrahim:chen:lip:herr}, for example $f(a, b,
c) = f(a) f(b \mid a) f(c \mid b, a)$.
Compared to typical chained equations
approaches, this has the advantage of resulting in a fully coherent joint
distribution. However, it still can be challenging to find and model complex
distributional features, particularly for models with many predictors. Additionally,
different conditioning sequences for the variables could result in different fits,
and the imputer may not have good information to help choose an order.
%
\section{Multiple Imputation via the HCMM-LD}\label{sec:model}
For $i=1,\dots,n$ sampled individuals, let
$X_i = (X_{i1},\dots, X_{ip})'$ be a vector of $p$
categorical variables for individual $i$, with each $X_{ij} \in
\{1,\dots, d_j\}$, and let $Y_i = (Y_{i1},\dots, Y_{iq})'$ be a vector of $q$ continuous responses taking values in $\mathbb{R}^q$.
We use $x_i$ and $y_i$ for specific values taken by $X_i$ and $Y_i$.
We also use superscript $\mathcal{X}$ and $\mathcal{Y}$ to signify that some parameter or latent variable is a component of the
model for $X$ or $Y$, respectively.
\subsection{The HCMM-LD for Imputing Mixed Data}
As noted in Section 1, mixture models have proven valuable for imputing multivariate missing data that are
strictly continuous or categorical.
The HCMM-LD\ fuses existing mixture models for strictly continuous or categorical data into a larger hierarchical model.
Thus, we begin with a brief
summary of these existing mixture models and discuss the shortcomings of various ``intuitive'' ways to combine them. We
present the HCMM-LD\ in Section \ref{sec:indexmodel}.
For imputing multivariate continuous data, \cite{Kim2013} use a truncated Dirichlet process (DP) mixture of normal distributions.
For $i=1,\dots,n$, let $H_i^{(\mathcal{Y})} \in \{1, \dots, {k^{\mathcal{(Y)}}}\}$ be the mixture component index for record $i$. This model assumes that
\begin{equation}
(Y_i\mid H_i^{(\mathcal{Y})}=\hy, \{(\mu_\hy, \Sigma_\hy) : 1\leq \hy\leq {k^{\mathcal{(Y)}}}\}) \sim N(\mu_\hy, \Sigma_\hy).\label{eq:ymodel}
\end{equation}
The prior distribution for $H_i^{(\mathcal{Y})}$ is a truncated version of the stick-breaking construction for the DP \citep{Sethuraman1994}, introduced in \cite{Ishwaran2001}:
\begin{gather}
\Pr(H_i^{(\mathcal{Y})}=\hy) = \byy{\phi_{\hy}}\\
\byy{\phi_{\hy}} = \byy{\xi_\hy}\prod_{l<\hy}(1-\byy{\xi_l}),\ \{\byy{\xi_\hy}: 1\leq \hy\leq {k^{\mathcal{(Y)}}}-1 \}\overset{iid}{\sim} Beta(1, \byy{\beta}),\
\byy{\xi_{{k^{\mathcal{(Y)}}}}}\equiv 1.\label{eq:Kip}
\end{gather}
For imputing multivariate categorical data, \cite{Si2013} adopt a truncated version of the DP mixture of product multinomials (MPMN) proposed
by \cite{Dunson2009}.
For $i=1,\dots,n$, let $H_i^{(\mathcal{X})} \in \{1, \dots, {k^{\mathcal{(X)}}}\}$ be the mixture component index for record $i$,
and let $\Pr(X_{ij} = x_{ij} \mid H_i^{(\mathcal{X})} = \hx) = \psi_{\hx x_{ij}}^{(j)}$. This model assumes that
\begin{align}
\Pr(X_i = x_i \mid H_i^{(\mathcal{X})} = \hx, \{\psi_\hx : 1\leq \hx\leq{k^{\mathcal{(X)}}}\}) &= \prod_{j=1}^{p}\psi_{\hx x_{ij}}^{(j)}\label{eq:dx},
\end{align}
where, for each $1\leq \hx\leq {k^{\mathcal{(X)}}}$, $\psi_\hx=\{ \psi_{\hx}^{(j)}: 1\leq j \leq p\}$ and each
$\psi^{(j)}_\hx = (\psi^{(j)}_{\hx 1}, \dots, \psi^{(j)}_{\hx d_j})'$ is a probability vector.
The prior on $\Pr(H_i^{(\mathcal{X})}=\hx)$ is another truncated stick breaking process:
\begin{gather}
\Pr(H_i^{(\mathcal{X})}=\hx) = \byx{\phi_\hx}\\
\byx{\phi_\hx} = \byx{\xi_\hx}\prod_{l<\hx}(1-\byx{\xi_l}),\ \{\byx{\xi_\hx}: 1\leq \hx\leq {k^{\mathcal{(X)}}}-1 \}\overset{iid}{\sim} Beta(1, \byx{\beta}),\ \byx{\xi_{k^{\mathcal{(X)}}}}\equiv 1.\label{eq:Hip}
\end{gather}
Given their success as imputation engines, it seems promising to fuse these two models into a coherent joint
distribution and MI engine for mixed data.
One approach is to assume the variables arise as in \eqref{eq:ymodel} and \eqref{eq:dx} with shared components
$H_i^{(\mathcal{X})}=H_i^{(\mathcal{Y})}\equiv H_i$, that is,
\begin{gather}
(Y_i\mid H_i=h,-) \sim N(\mu_h, \Sigma_h),\
\Pr(X_i = x_i \mid H_{i} = h, -) =\prod_{j=1}^{p}\psi_{hx_{ij}}^{(j)}.
\end{gather}
This model makes strong \emph{local independence} assumptions, namely that $Y \id X \mid H$. This puts a significant burden on
the mixture components. They must simultaneously capture non-normality in the distribution of $Y$, dependence between
$Y$ and $X$, and dependence within $X$. Doing so typically requires a large number of components and a commensurate
amount of data. For example, since $X$ is categorical the true mean function can be written as ${E(Y\mid X=x)=\tilde D(x)\tilde{B}}$, where $\tilde D(x)$ is the true design vector and $\tilde{B}$ is the matrix of true regression coefficients.
The
model has to include components at each distinct value of $\tilde D(x)\tilde{B}$---for all possible values of $x$---just to model the mean function, with further
components to capture non-Gaussian structure in $Y$ and dependence in $X$.
This burden can be alleviated somewhat by allowing the means to depend on $X$ as in the general location model: $\mu_h(x) = D(x)B_h$, with $D$ encoding main effects and possibly interactions. However, when $p>q$ as is common in survey data, the number of components required to adequately model dependence in
$X$ tends to be much larger than that required to model $Y$, particularly since this model allows for local dependence
in $Y$ (through the covariance matrices) but not in $X$.
An alternative approach is to use separate component indices and independent
prior distributions for $\Pr(H_i^{(\mathcal{X})}=\hx)$ and $\Pr(H_i^{(\mathcal{Y})}=\hy)$, as in \eqref{eq:ymodel}-\eqref{eq:Hip}, but add $(X,Y)$ dependence by setting $\mu_\hy(x) = D(x)B_{\hy}$.
We have
\begin{gather}
(Y_i\mid X_i = x_i, H_i^{(\mathcal{Y})}=\hy,-) \sim N(D(x_i)B_\hy, \Sigma_\hy),\label{eq:ymodelwithx}\\
\Pr(X_i = x_i \mid H_i^{(\mathcal{X})} = \hx, -) =\prod_{j=1}^{p}\psi_{\hx x_{ij}}^{(j)}.\label{eq}
\end{gather}
This model enforces restrictive assumptions about the relationship between $Y$ and $X$. For example, we would
have $E(Y\mid X=x) = D(x)\left[\sum_{\hy=1}^{{k^{\mathcal{(Y)}}}} B_\hy\byy{\phi_\hy} \right]$, so that the model is unable to capture interactions not
already coded in $D(x)$.
To construct the HCMM-LD, we use \eqref{eq:ymodelwithx} and \eqref{eq} as the data models. However, rather
than choose between common components or independent components, we use
a hierarchical prior distribution that maintains the desirable features of both, while
incorporating new forms of local dependence. We now outline this hierarchical prior distribution for $(H_i^{(\mathcal{X})}, H_i^{(\mathcal{Y})})$.
\subsubsection{Hierarchical prior for component indexes}\label{sec:indexmodel}
Let $Z_i$ be a third component index such that $1\leq Z_i\leq {k^{\mathcal{(Z)}}}$. We assume $H_i^{(\mathcal{X})}$ and $H_i^{(\mathcal{Y})}$ are independent given $Z_i$, so that
\begin{gather}
\Pr(H_i^{(\mathcal{X})}=\hx, H_i^{(\mathcal{Y})}=\hy \mid Z_i=z)
= \byx{\phi_{z\hx}}\byy{\phi_{z\hy}}\label{eq:hmodel}\\
\Pr(Z_i=z) = \lambda_z.\label{eq:zmodel}
\end{gather}
Here each $\byx{\phi_z} = \left(\byx{\phi_{z1}},\dots,\byx{\phi_{z{k^{\mathcal{(X)}}}}}\right)'$ and
$\byy{\phi_z} = \left(\byy{\phi_{z1}}, \dots,\byy{\phi_{z{k^{\mathcal{(Y)}}}}}\right)'$ are probability vectors.
Both are assigned independent truncated stick breaking priors. For $1\leq z\leq {k^{\mathcal{(Z)}}}$, we have
\begin{gather}
\byx{\phi_{z\hx}} = \byx{\xi_{z\hx}}\prod_{l<\hx}(1-\byx{\xi_{z\hx}}),\ \{\byx{\xi_{z\hx}}: 1\leq \hx\leq {k^{\mathcal{(X)}}}-1 \}\overset{iid}{\sim} Beta(1, \byx{\beta}),\ \byx{\xi_{k^{\mathcal{(X)}}}}\equiv 1.\\
\byy{\phi_{z\hy}} = \byy{\xi_{z\hy}}\prod_{l<\hy}(1-\byy{\xi_{z\hy}}),\ \{\byy{\xi_{z\hy}}: 1\leq \hy\leq {k^{\mathcal{(Y)}}}-1 \}\overset{iid}{\sim} Beta(1, \byy{\beta}),\ \byy{\xi_{k^{\mathcal{(Y)}}}}\equiv 1.
\end{gather}
Marginalizing over $Z$ gives $\Pr(H_i^{(\mathcal{X})}=\hx, H_i^{(\mathcal{Y})}=\hy) = \sum_{z=1}^{{k^{\mathcal{(Z)}}}} \lambda_z\byx{\phi_{z\hx}}\byy{\phi_{z\hy}}$, inducing dependence between the latent component membership indicators.
The top-level mixture probabilities $\lambda=(\lambda_1,\lambda_2,\dots,\lambda_{{k^{\mathcal{(Z)}}}})'$ are also assigned a truncated stick breaking process:
\begin{align}
\lambda_z = \byz{\xi}_z\prod_{l<z}(1-\byz{\xi}_l),\ \{\byz{\xi}_z: 1\leq z\leq {k^{\mathcal{(Z)}}}-1 \}\overset{iid}{\sim} Beta(1, \alpha),\ \byz{\xi}_{{k^{\mathcal{(Z)}}}}\equiv 1.
\end{align}
\cite{Banerjee2013} establish that as $({k^{\mathcal{(Z)}}},{k^{\mathcal{(X)}}},{k^{\mathcal{(Y)}}})$ all approach $\infty$, this is a well-defined prior distribution, which they call an infinite tensor factorization (ITF) prior.
We assign $\alpha, \byx{\beta}$, and $\byy{\beta}$ independent Gamma prior distributions with shape and rate parameters equal to 0.5. A convenient strategy for choosing the truncation levels $({k^{\mathcal{(Z)}}},{k^{\mathcal{(X)}}},{k^{\mathcal{(Y)}}})$ is to pick moderate initial values, increasing them if the number of occupied components approaches its upper bound. Appropriate values will depend on the dataset; for all the models fit in this paper we take ${k^{\mathcal{(Z)}}}=15,\ {k^{\mathcal{(X)}}}=90$, and ${k^{\mathcal{(Y)}}}=60$, which we found to be conservative {upper bounds. Generally, the HCMM-LD\ is insensitive to specific choices of $({k^{\mathcal{(Z)}}}, {k^{\mathcal{(X)}}}, {k^{\mathcal{(Y)}}})$ provided that they allow for unoccupied components. }
\subsubsection{Data model priors}
We next specify prior distributions for the parameters in \eqref{eq:ymodelwithx} and \eqref{eq}. For each $\psi_\hx^{(j)}$, we use independent Dirichlet distributions,
\begin{equation}
\psi^{(j)}_{\hx} \overset{iid}{\sim} Dir(\gamma_{\hx 1}^{(j)}, \dots, \gamma_{\hx d_j}^{(j)}).
\end{equation}
Reasonable default choices for the hyperparameters include setting $\gamma_{\hx}^{(j)} = (1,\dots,1)$ or
$\gamma_{\hx}^{(j)} = (1/d_j,\dots,1/d_j)$. Both represent relatively vague information about the within-component probabilities. In practice, we find that posterior predictive distributions are usually insensitive to this choice, and we
use the latter going forward.
For the parameters in each $Y$-component, we use hierarchical normal-inverse Wishart priors. The hierarchical priors are an alternative to more restrictive models, recognizing that many components will have a relatively small number of data points and that elements of $B_{\hy}$ in particular may be poorly estimated. We have
\begin{equation}
\{(B_{\hy}, \Sigma_{\hy})\}\overset{iid}{\sim} MatN( B_0, I, T_B)\times IW(v, \Sigma)
\end{equation}
\begin{equation}
(B_0, \Sigma)\sim MatN( 0, I, \sigma^2_{0\beta}I)\times W(w, \Sigma_0).
\end{equation}
Here, $MatN(M, \Phi, \Sigma)$ is the matrix normal distribution, i.e. the distribution of the $p^*\times q$ dimensional matrix $M + \Phi^{1/2}\Omega\Sigma^{1/2}$ when $\Omega$ is $p^*\times q$ with $\omega_{ij}\overset{iid}{\sim} N(0,1)$. We assume that $T_B =\diag(1/\tau_{1},\dots, 1/\tau_{q})$, and assign $\tau_1,\dots \tau_q$ independent $G(0.5, 0.5)$ priors. In applications we find the posterior predictive distributions to be insensitive to this choice.
To complete the hyperprior, we use the fact that $E(\Sigma_h) = \frac{v}{w-q-1}\Sigma_0$.
We center and scale each element of $Y$ marginally
and take $v=q+1$, $w=q+2$, and $\Sigma_0 = \frac{1}{q+1}I$ throughout. In sufficiently large samples, inferences are insensitive to the choice of $\sigma^2_{0\beta}$; we use $10$.
\subsection{Properties of the HCMM-LD\ }\label{sec:modelprop}
\begin{figure}
\centering
\includegraphics[trim=0cm 4cm 0cm 4cm, width=.45\textwidth]{./fig/plate.pdf}
%
\caption{Model structure of the HCMM-LD.}
\label{fig:plate}
\end{figure}
{Figure \ref{fig:plate} summarizes graphically the dependence structure of the HCMM-LD.} Marginally, $X$ has a latent class model {of the form
\begin{equation}
\Pr(X_i = x_i) = \sum_{\hx=1}^{{k^{\mathcal{(X)}}}} \left(\sum_{z=1}^{{k^{\mathcal{(Z)}}}} \lambda_z\byx{\phi_{z\hx}}\right) \prod_{j=1}^{p}\psi_{\hx x_{ij}}^{(j)},\label{eq:marx}
\end{equation}
where the term in parentheses gives the probability for class $s$}. This can capture any multivariate categorical distribution given sufficiently large ${k^{\mathcal{(X)}}}$ (unlike unsaturated loglinear models). Conditional on $Y=y$, $X$ still follows a latent class model but with class probabilities that are functions of $y$. The conditional distribution of $Y$ for any cell of the $X$ table is a mixture of multivariate normal distributions,
\begin{equation}
f(Y_i\mid X_i=x_i) = \sum_{\hy=1}^{{k^{\mathcal{(Y)}}}}
\frac{w_\hy(x_i)}{\sum_{l=1}^{{k^{\mathcal{(Y)}}}} w_l(x_i)}
N(Y_i; D(X_i)B_\hy, \Sigma_\hy)\label{eq:ygivenx}
\end{equation}
where $w_{\hy}(x_i) = \sum_{z=1}^{{k^{\mathcal{(Z)}}}} \lambda_z\byy{\phi_{z {\hy}}}
\sum_{\hx=1}^{{k^{\mathcal{(X)}}}}\byx{\phi_{z\hx}}\prod_{j=1}^{p}\psi_{\hx x_{ij}}^{(j)}$.
The marginal distribution of $Y$ is also a mixture of multivariate normals.
Thus, the HCMM-LD\ can represent a wide variety of shapes for the distribution of $Y$.
Since $w_{\hy}(x_i)$ also appears in the expression for the conditional mean of $Y$,
the HCMM-LD\ can
capture interactions not necessarily encoded in $D$.
%
%
The HCMM-LD\ {encodes} local dependence within and between $Y$ and $X$ in several ways{:} The $Y$-component specific regression
functions and covariance matrices allow the relationships between $Y$ and $X$, and within $Y$, to vary by component.
Further, the prior distribution in \eqref{eq:hmodel} - \eqref{eq:zmodel} implies that the HCMM-LD\ is a ``mixture of mixture models.''
Marginalizing over $H_i^{(\mathcal{X})}$ and $H_i^{(\mathcal{Y})}$, the density of $(Y_i,X_i)$ given $Z_i=z$ is
\begin{equation}
f(X_i, Y_i \mid Z_i=z) = \left(\sum_{\hy=1}^{k^{\mathcal{(Y)}}} \byy{\phi_{z\hy}}N(Y_i; D(X_i)B_{\hy}, \Sigma_{\hy}) \right)\left(\sum_{\hx=1}^{k^{\mathcal{(X)}}} \byx{\phi_{z\hx}}\prod_{j=1}^{p}\psi_{\hx X_{ij}}^{(j)}\right)\label{eq:xymidz},
\end{equation}
{so the joint density is
\begin{equation}
f(X_i, Y_i) = \sum_{z=1}^{{k^{\mathcal{(Z)}}}} \lambda_z\left(\sum_{\hy=1}^{k^{\mathcal{(Y)}}} \byy{\phi_{z\hy}}N(Y_i; D(X_i)B_{\hy}, \Sigma_{\hy}) \right)\left(\sum_{\hx=1}^{k^{\mathcal{(X)}}} \byx{\phi_{z\hx}}\prod_{j=1}^{p}\psi_{\hx X_{ij}}^{(j)}\right)\label{eq:xy}.
\end{equation}
}
For any $z$, $(X\mid Z=z)$ follows an MPMN model.
The distribution of $(Y\mid X, Z=z)$ is nearly the ANOVA-DDP model of \cite{DeIorio2004}, except that we relax
their common covariance assumption with the hierarchical prior. From \eqref{eq:xymidz}, we see that within top-level components $Z$
we have local \emph{dependence} within $X$, as well as between $X$ and $Y$ (and within $Y$). Note that $f(X_i, Y_i\mid Z_i=z)$ and $f(X_i, Y_i\mid Z_i=z')$ differ
only in their respective lower-level stick breaking weights, $(\phi^{(y)}_{z},\phi^{(x)}_{z})$ and $(\phi^{(y)}_{z'},\phi^{(x)}_{z'})$, and not in the lower-level parameters ($\{(B_\hy, \Sigma_\hy)\}$ and $\psi_\hx$). This is a parsimonious choice, somewhat akin to assuming common covariance structures across components in normal mixture models (but much more flexible).
\subsection{Related work}\label{sec:mixlit}
\cite{Dunson2010} extended Dunson's and Xing's (2009) MPMN to mixed data by assuming fully factorized (product) kernels in a DP mixture. This
model is a special case of the HCMM-LD\ with ${k^{\mathcal{(Y)}}}={k^{\mathcal{(X)}}}=1$, diagonal $\Sigma_\hy$, and $D(x_i)=1$. \cite{Dunson2010} note that when the number of variables grows the number of clusters also must grow to accommodate the dependence in the joint distribution. This is due to the local independence assumptions of the
product kernel, which forces all the dependence to be represented through a single cluster index. As described in Section \ref{sec:modelprop}, the HCMM-LD\ is able to avoid such strong local independence assumptions through the use of multivariate normal regression components with full covariance matrices, as well as the structure imposed by the hierarchical prior on the mixture component memberships.
A number of authors have proposed joint mixture models that
include limited local dependence to induce a
prior on one of the conditional distributions. These models decompose the joint kernel into a
conditional kernel for one variable given the others and a marginal product
kernel for predictors \citep{Shahbaba2009a,Molitor2010,Hannah2011}.
Some of these are special cases of the HCMM-LD\ obtained by restricting ${k^{\mathcal{(Y)}}}={k^{\mathcal{(X)}}}=1$ and imposing more structure on the local covariances. These models are less relevant for imputation, where the entire joint distribution is of interest. Moreover, the assumption of local marginal independence between the majority of the variables can lead to the same proliferation of clusters as in
latent class models \citep{Hannah2011}.
\cite{Banerjee2013} introduced the ITF prior to avoid the proliferation of components through \emph{dependent} component assignment in separate univariate mixtures. Compared to the HCMM-LD, this model encodes weaker local
dependence, arising strictly through shared lower-level components within top-level components.
\cite{Wade2011} proposed the enriched DP, which (like the ITF) models dependent cluster assignment.
The enriched DP separates a joint distribution into a conditional and a marginal,
assigning each a DP prior distribution where the base measure for the
conditional varies across the marginal. It lacks
the symmetry of the ITF, making it more difficult to interpret the
induced joint distribution and its margins. This is unappealing for our purposes.
The HCMM-LD\ has a number of benefits over existing alternatives. The hierarchical structure on component indices and other local dependence features allow the analyst to avoid the proliferation of clusters. At the same time, the induced marginal and conditional distributions are easy to derive and the HCMM-LD\ has appealing limiting forms (the MPMN model for $X$ and a multivariate regression or ANOVA-DDP for $Y\mid X$). Computation via MCMC is also straightforward, as detailed in the supplementary material.
\section{Repeated Sampling Simulation Studies}\label{sec:sipp-sim}
To evaluate the performance of HCMM-LD\ in multiple imputation (MI), we
conducted several repeated sampling simulation studies on a constructed population
taken from the first wave of the 2008 SIPP panel. We define the
population as individuals who reported positive income from work
during the reference period in Wave 1, excluding records with
missing entries.
The constructed population consists of $N$=30,507
respondents. With guidance from Census Bureau researchers, we select the two continuous
and eleven categorical variables displayed in
Table \ref{tab:sippvars}. We use a modest number of variables to
make a large repeated sampling study more efficient while keeping the problem
challenging. For example, the implied contingency table has over 7 million
cells and is very sparse. %
Below we present comparisons of the HCMM-LD\ versus a fully conditional approach, as implemented in the R package \texttt{mice} \citep{VanBuuren2011}. We compare against a fully conditional approach as these have been repeatedly shown to perform at least as well as existing joint models for MI \citep[e.g.,][]{VanBuuren2007,lee2010multiple,kropko2014multiple} and are in widespread use.
In the online supplement we also compare to a variant of the general location model; its performance is dominated by that of the HCMM-LD\ and MICE. The supplement also includes a simulation study comparing the HCMM-LD\ and MICE under missingness completely at random (MCAR). Overall conclusions under MCAR are similar to those below, but the performance difference is greater under MCAR since all variables were subject to missingnes in that case.
\begin{table}[t]
\caption{Variables in the repeated sampling simulation study}\label{tab:sippvars}
\begin{tabular}{|l|l|}
\hline
Variable & Levels \\ \hline
Total monthly earnings from employment & Continuous \\
Age & Continuous \\
Sex & 2 \\
Race & 5 \\
Marital Status & 6 \\
Born in the US & 2 \\
Number of own children in the home & 4 (0,1,2, or 3+) \\
Education level & 6 \\
Occupation & 23 \\
Worker Class & 3 (Private, Nonprofit, Government) \\
Union & 2 \\
Hourly & 2 \\
Usual Hours worked & 9 (0-80 in increments of 10 hours, 80+) \\ \hline
\end{tabular}
\end{table}
\subsection{MAR simulation study design}
We create 500 datasets by taking simple random samples of size
$n=6,000$. In each dataset, we let a random sample of 180 observations be complete cases. Age and sex are completely observed. Let $R_{i,\text{var}}=1$ when the variable ``var'' is missing for observation $i$ and $0$ otherwise. Define $U_{i} = \ind{\text{sex}_i=\text{``male''}}$. We sample
$R_{i,\text{earn}}$ and $R_{i,\text{child}}$ from Bernoulli distributions with probabilities derived from
\begin{align*}
\logit[\Pr(R_{i,\text{earn}}=1)] &= -0.25 + 0.5U_{i} - \left(\frac{\text{age}_i-25-25U_{i}}{25}\right)^2\\
\logit[\Pr(R_{i,\text{child}}=1)] &= -1.5U_{i} - \left(\frac{\text{age}_i - 40 + 10U_{i}}{30 + 10U_{i}}\right)^2.
\end{align*}
We partition the remaining variables into two blocks: demographic variables (race, marital status, born in US) and variables directly related to employment (education, occupation, worker class, union, hourly, hours worked). For each variable $j$ in the demographic block, we sample each $R_{ij}$ from Bernoulli distributions with probabilities derived from
\begin{equation}
\logit[\Pr(R_{ij}=1)] = -1 + 0.7R_{i,\text{child}} + 1.25\kappa_{ij},
\end{equation}
where $\kappa_i$ is a $3$-dimensional vector drawn from a normal distribution with mean 0, unit variances and all correlations equal to 0.3. For each variable $j'$ in the employment block, we sample each case's $R_{ij'}$ from Bernoulli distributions with probabilities depending on $R_{i,\text{earn}}$ instead of $R_{i,\text{child}}$:
\begin{equation}
\logit[\Pr(R_{ij'}=1)] = -1 + 0.7R_{i,\text{earn}} + 1.25\omega_{ij'},
\end{equation}
where $\omega_i$ is a $6$-dimensional vector drawn from a normal distribution with mean 0, unit variances and all correlations equal to 0.3. %
In this MAR design, approximately 1/3 of the entries are missing for each variable (except age and sex) and approximately 5\% of cases are complete. %
The resulting imputation problem is challenging, as highly correlated variables are more likely to be missing simultaneously.
The MAR mechanism results in biased available case estimates; for example, regressing log earnings on $U_m$ gives estimates of the coefficient around 0.25 (SE 0.02) in the available cases, compared to a true value of about 0.36.
\subsubsection{Generating imputations}
Within each of the 500 simulated datasets we create $M=10$ multiple imputations
with the HCMM-LD. We use the default prior distributions described
in Section \ref{sec:model}, after standardizing the continuous
variables, and include main effects for each categorical variable in $D(X)$.
We estimate the HCMM-LD\ for each dataset using 200,000 MCMC iterations from the
Gibbs sampler described in the supplemental material, discarding the
first $100,000$ iterations and keeping the imputations from every
$10,000^{th}$ iteration thereafter. This is very conservative;
examination of a handful of datasets suggests that these numbers
could be reduced by at least half without impacting the results. In
practice, of course, imputers should carefully examine MCMC diagnostics of
relevant identified parameters, such as marginal means, quantiles, and
variances or covariances in the completed datasets. We ran the
simulations in a heterogeneous cluster environment, so the run times
varied. As a reference, a 2014 MacBook Pro can complete 10,000 iterations of
the MCMC sampler in about 20 minutes. Our implementation could be made much more efficient; we discuss scalability in Section \ref{sec:conclusion}.
With each incomplete dataset, we also implement multiple imputation via
chained equations using the R package \texttt{mice}. Our goal is to compare default
applications of the software to a default application of the HCMM-LD, so we did not alter any of the options
to \texttt{mice} other than to set $M=10$. The default
procedure imputes continuous variables via predictive mean matching
\citep{Little1988} and uses logistic regressions to impute discrete
variables. Each conditional model includes a main effect for every
other variable. After imputing the data with both procedures, we obtain MI inferences for a number of estimands using the methods in
\citet{Rubin1987}. We compute completed-data estimates and standard errors using the
\texttt{survey} package in R \citep{Lumley2004},
incorporating a finite population correction.%
%
\subsubsection{Evaluation metrics}
{We evaluate the competing imputation methods based on the performance of the MI pooled estimate and associated confidence interval
for a range of estimands. For each estimate, the ``true'' value is the corresponding quantity computed in the population of $N$ individuals.
Ideally, under repeated sampling and realizations of the nonresponse process, and across a range of estimands, the imputation methods yield
completed datasets for which 1) pooled estimates are approximately unbiased for the corresponding population quantities, and 2)
pooled confidence intervals with level $\alpha$ contain their true population values at least $(1-\alpha)\%$ of the time \citep{Rubin1996}.
Thus, while the proposed imputation method is derived from a Bayesian model, the ultimate evaluations are frequentist;
see \cite{Rubin1987,Rubin1996} for justification of this perspective.}
\subsection{Results}
We begin by examining the means of log monthly earnings by age
(discretized into 10 year intervals except for $<18$, $18-25$, and
$65+$), sex and presence of own children.
We restrict to cells in the
table formed by the three categorical variables with expected counts
of at least 30. We work on the log scale rather than with untransformed incomes, as
the skewness of the income distribution makes normal approximations
more likely to hold.%
%
%
%
Figure \ref{fig:age-sex-kid-ci} shows the coverage rates and average
width of 95\% multiple imputation confidence intervals. For most cells the pooled confidence intervals have approximately the correct coverage (the cluster of points around (0.95, 0.95) on the left hand side of Fig. \ref{fig:age-sex-kid-ci}). However, there are three cells where the MICE-constructed imputations yield substantially lower coverage than under the HCMM-LD. On average the pooled confidence intervals have similar widths, suggesting that bias in MICE's imputations drive the poor coverage. This is confirmed by Figure \ref{fig:age-sex-kid-bias}, which shows that while neither method has uniformly lower bias, the range of bias under the HCMM-LD\ is much smaller.
\begin{figure}
{\centering \includegraphics[width=.4\linewidth]{fig/mar_res6000-mean_age_kid_sex_cover1}
\includegraphics[width=.4\linewidth]{fig/mar_res6000-mean_age_kid_sex_cover2}
}
\caption{(Left) Coverage rate of pooled nominal 95\% CI for mean log monthly earnings by age, sex, and own children in the home (Yes/No) (Right) Average CI width
of 95\% CI.}
\label{fig:age-sex-kid-ci}
\end{figure}
\begin{figure}
{\centering \includegraphics[trim={0 0.5cm 0 0}, width=.35\linewidth]{fig/mar_res6000-mean_age_kid_sex1}
\includegraphics[trim={0 0.5cm 0 0}, width=.35\linewidth]{fig/mar_res6000-mean_age_kid_sex2}
}
\caption{Standardized and percent bias of pooled
estimates of mean log monthly earnings by age, sex, and own children
in the home (Yes/No). Each line represents a cell mean, with the
left and right endpoints at the bias under MICE and HCMM-LD,
respectively.}
\label{fig:age-sex-kid-bias}
\end{figure}
These are not especially small cells, although they do have somewhat higher probabilities of missingness for the own child and income variables (around 0.5 and 0.4, respectively).
The same estimates are problematic in the MCAR simulation (see the supplementary material), so this is not merely a function of
the MAR mechanism used here.
Rather, it appears to be a function of complicated relationships among the variables involved.
Age and income have a relationship that the MICE imputations evidently
capture less effectively than the HCMM-LD\ imputations. For example, earnings tend to be lowest in the young (SIPP records earnings information on
respondents 15 or older), increasing during working years and falling
off again as those who can afford to retire do so. Additionally, the
variance in earnings is low in the younger cohort, roughly stable
through the working years, and increasing near and after retirement
age. Interactions also appear to be at play; the effect of
having their own child in the home varies across the respondent's age,
probably due in part to its high correlation with the age of the children, and
across the sexes as well. For example, the population difference in
log wages between those with children versus no children for 18-24 year old women is -0.159, whereas for 35-44 year
old women it is -0.076. In men the population differences are -0.064 for 18-24 year olds and 0.232 for 35-44 year olds.
\subsubsection{Regression Coefficients}
Next we consider linear regressions of log earnings on age, sex, usual
hours worked (recoded as $<30$, 30-60, and $60+$), and indicators for
marriage and own child under 18 in the
household. To begin we fit a model including an age squared term as
well as two- and three-way interactions between sex, own child, and
marital status. Figure \ref{fig:regms3sq} displays MI
estimates of the coefficients and the average width of their
confidence intervals. {The HCMM-LD\ imputations result in better
repeated sampling properties overall. Neither method was modifed to anticipate the nonlinear relationship between age and income, although both have some ability to capture such relationships. MICE's predictive mean matching effectively borrows residuals from cases with similar predicted means, providing some protection against model misspecification \citep{Little1988}, whereas the HCMM-LD\ specifically intends to capture complex relationships between the variables.} From Figure \ref{fig:regms3sq} it is clear that the HCMM-LD\ is more successful at capturing this structure, with coverage nearer the nominal level for age and age squared (and all other coefficients), despite age being completely observed.
\begin{figure}
{\centering \includegraphics[width=.4\linewidth]{fig/mar_res6000-regms3sq1}
\includegraphics[width=.4\linewidth]{fig/mar_res6000-regms3sq2}
}
\caption{(Left) Coverage rate of pooled nominal 95\% CI for the regression with three-way interaction and age squared, including fpc. (Right) Average width
of 95\% CI.}
\label{fig:regms3sq}
\end{figure}
We also considered the same model excluding the age squared term.
Figure \ref{fig:regms3} shows that the results are largely similar for the former, with coverage generally improved overall but both methods struggling on the coefficients for own child and its interaction with the indicator of being married. Bias and average widths of confidence intervals are generally similar between the two methods on all the coefficients. A notable exception is the coefficient on the indicator variable for working over 60 hours, for which MICE achieves 85\% coverage to HCMM-LD's 96\%. The difference is driven by bias; the true population coefficient is 0.23, and the average pooled estimate using the MICE imputations is 0.18 versus 0.23 under HCMM-LD. The effect persists even in the model including only main effects (coverage of 82\% under MICE, versus 97\% under HCMM-LD). Here the true coefficient is 0.25, and the average point estimate from MICE is 0.18 versus 0.24 in the HCMM-LD. Coverage, widths of CIs and bias were essentially identical under both methods for the other coefficients. We had expected MICE to dominate in the main effects model since this is a submodel of the linear model MICE used to impute income. The relatively poor performance appears to
be due to the small sample size of this group (807 in the
population) and large true effect,
which combine to make predictive mean matching less effective.
\begin{figure}
{\centering \includegraphics[width=.4\linewidth]{fig/mar_res6000-regms31}
\includegraphics[width=.4\linewidth]{fig/mar_res6000-regms32}
}
\caption{(Left) Coverage rate of pooled nominal 95\% CI for regression with three-way interaction \emph{without} age squared, including fpc. (Right) Average width
of 95\% CI.}
\label{fig:regms3}
\end{figure}
\subsubsection{Conditional Frequencies}
We also examine the quality of categorical imputations by estimating
cell frequencies of categorical variables. We restrict to cases where
$E(n_c)\times p_c\geq10$ and $E(n_c)\times (1-p_c)\geq10$, where $p_c$ is
the true proportion and $n_c$ is the cell size in a simple random sample, to make the normal approximation more plausible. Figure \ref{fig:prop_own_child} displays results from estimating the proportion of respondents with their own child under 18 in the home by sex, race and age. Recall that only race and the presence of the respondent's own child have missing values.
The HCMM-LD\ based imputations perform much better than MICE. Coverage rates are uniformly better under the HCMM-LD\ with CIs of comparable width.
Coverage rates for the HCMM-LD\
never drop below 84\% (versus 71\% for MICE in that case), and the difference is dramatic for a few estimands. Figure \ref{fig:prop_own_child_cov_by_n} shows that MICE
has very good or very poor coverage in large cells, consistent with
the lack of coverage arising from misspecification bias. The HCMM-LD\
tends to have slightly lower coverage in these larger cells than in the smaller cells, but not nearly to the extent of MICE.
\begin{figure}
{\centering \includegraphics[width=.4\linewidth]{./fig/mar_res6000-prop_own_kid_sex_race_age_cover_N1}
\includegraphics[width=.4\linewidth]{./fig/mar_res6000-prop_own_kid_sex_race_age_cover_N2}
}
\caption{(Left) Coverage rate of nominal 95\% CIs for proportion with
own child $<18$ in the household by age, race and sex. (Right) Average width
of 95\% CI.}
\label{fig:prop_own_child}
\end{figure}
\begin{figure}
{\centering \includegraphics[width=.55\linewidth]{fig/mar_res6000-prop_own_kid_sex_race_age}
}
\caption{Coverage by expected cell size for proportion with own child $<18$ in the household by age, race and sex. Lines connect the coverage rates that correspond to the same estimand. Blue dot-dashed lines indicate that the HCMM-LD\ coverage is closer to 95\% than MICE, with red dashed lines indicating the reverse.}
\label{fig:prop_own_child_cov_by_n}
\end{figure}
\section{Evaluating the SIPP Redesign}\label{sec:msipp}
We now evaluate the agreement between the data from the field test and the data from the constructed sample of the
2008 production SIPP panel. For brevity, we use SIPP to refer to the subset of the production panel and SIPP-EHC to refer to
the data from the field test. We focus on a subset of the data, namely household heads in 2010 from the SIPP-EHC and a
contemporaneous wave of the SIPP subsample. The sample sizes are 2,588 for the SIPP-EHC and 3,665 for SIPP.
Previously, Census Bureau researchers compared complete-case estimates from SIPP-EHC to those from production SIPP,
and also to administrative records where available \citep{USCensus2013}. However, most
variables have missing values, and for some variables the missingness is substantial as evident in Table \ref{tab:msippvar}.
\begin{table}[tdp]
\begin{center}
\caption{Variables used from field test and production panel, and their fractions of missing data. An * indicates that the missing data percentage is computed as a fraction of the units known to be in-universe; for example, the percentage reported for monthly earnings is the fraction of respondents who indicated employment during the reference period but did not report the earnings amount. These also correspond to the continuous variables; the remainder are discrete.}
\label{tab:msippvar}
\begin{tabular}{l|l|l|l}
Variable Type & Variable & SIPP-EHC & SIPP \\
\hline
Household characteristics & Proxy interview & 0.00\% & 0.00\% \\
& State & 0.00\% & 0.00\% \\
& Household composition & 0.00\% & 0.00\% \\
& No. persons in family & 0.00\% & 0.00\% \\
& No. children under 18 & 0.00\% & 0.00\% \\
\hline
Householder characteristics & Sex & 0.00\% & 0.00\% \\
& Race/Ethnicity & 0.08\% & 4.12\% \\
& Born in U.S. & 0.12\% & 0.05\% \\
& Nativity/Citizenship status & 0.12\% & 0.05\% \\
& Marital status & 0.70\% & 3.30\% \\
& Disabled & 5.18\% & 4.09\% \\
\hline
Work/Education & Educational attainment & 1.55\% & 0.00\% \\
& Enrolled in school & 0.00\% & 0.00\% \\
& Employment status & 15.96\% & 16.75\% \\
& Monthly earnings* & 22.70\% & 17.56\% \\
\hline
Program participation & Health insurance (any) & 2.43\% & 2.62\% \\
& OASDI & 4.17\% & 1.77\% \\
& SNAP & 1.85\% & 1.64\% \\
& SNAP benefit amount* & 58.87\% & 65.00\% \\
& TANF & 0.62\% & 0.22\% \\
& SSI & 3.28\% & 3.98\% \\
& Unemployment insurance & 4.13\% & 0.71\% \\
\hline
Assets & Own interest bearing account & 6.34\% & 2.21\% \\
& Own stocks/mutual funds & 5.80\% & 2.29\% \\
& Own retirement account & 6.38\% & 1.88\% \\
& Tenure in residence & 0.85\% & 0.03\% \\
& Own home value* & 28.37\% & 38.57\% \\
\end{tabular}
\end{center}
\end{table}%
Missing data rates are generally similar in SIPP and SIPP-EHC, but missing data patterns vary substantially
between the two surveys. For example,
Figure \ref{fig:agemis} shows density estimates of respondents' ages by whether their employment status is missing.
In the SIPP sample, respondents with missing employment status are more likely to be younger, whereas in the SIPP-EHC
they tend to be older. In each sample about 16\% of respondents are missing their employment status, so it is
unlikely that these differences are due to sampling variability. In neither case is employment status plausibly missing completely at random, so comparisons based solely on complete cases or pairwise deletion
are unreliable.
\begin{figure}
\begin{center}
\includegraphics[trim={0 0.9cm 0 0}, width=0.7\textwidth]{fig/age-mis}
\end{center}
\caption{Distributions of age in SIPP/SIPP-EHC by missing employment status indicator}
\label{fig:agemis}
\end{figure}
To account for missing data, we generated a set of $M=8$ completed datasets for SIPP and SIPP-EHC.
We used the HCMM-LD with the variables in Table \ref{tab:msippvar}, restricting the design vectors
to main effects only.
We imputed the SIPP-EHC and SIPP
separately to avoid biasing the comparisons.
The continuous variables actually have a spike at zero corresponding to the
unemployed, non-homeowners, or non-participants in SNAP. %
We decompose each of these into a binary indicator of a non-zero value and a continuous variable
that is treated as missing anywhere the indicator is zero. Imputations for the original variable are constructed as the product
of the indicator and the continuous variable (as in \cite{Heeringa2002}).
We run the MCMC for 130,000 iterations, discarding the first 50,000 and saving a completed dataset
every 10,000$^{th}$ iteration thereafter. Standard MCMC diagnostics again suggest this is conservative.
We compare the SIPP-EHC and SIPP on estimates of employment status, earnings, home value and SNAP benefit amounts,
since these variables have significant fractions of missing data. We
focus primarily on the effect that accounting for missing data has on comparing estimates from SIPP-EHC to the SIPP.
%
%
\subsection{Results: The impacts of accounting for missing data}\label{sec:SIPP-EHC:MI}
Figure \ref{fig:empstat} displays estimates of the proportion of respondents who were
employed at some point during the previous month, computed using the
HCMM-LD\ MI procedure and also using only the
cases with observed values. Compared to the complete case estimates, the MI
estimate is about 1\% higher for the SIPP and 1.5\% lower for
SIPP-EHC; hence, accounting for the missing data attenuates the
apparent differences in the estimates.
This attenuation is concordant with the differential age distributions
of the cases with missing employment displayed in Figure \ref{fig:agemis}. Further, in SIPP-EHC individuals who
report participating in SNAP are twice as likely to have missing income data as those who do not receive SNAP, whereas
in SIPP the rate of missingness for employment status is about the same regardless of SNAP participation. Therefore, we
expect more of those missing employment status in SIPP-EHC actually to be unemployed, and they are evidently imputed as such.
\begin{figure}
\begin{center}
\includegraphics[width=0.55\textwidth]{fig/msipp-empstat}
\end{center}
\vspace{-20pt}
\caption{Percent reporting employment during the
reference period, with 95\% confidence intervals, for SIPP-EHC and
the production SIPP subsample.}
\label{fig:empstat}
\end{figure}
Figure \ref{fig:medagesex} displays similar quantities for the median
earnings by age and sex strata. For most strata, accounting for
missing data appears not to have substantial impact on the
comparisons (most estimates and 95\% CIs are similar pre- and
post-imputation).
A notable exception occurs for 35-44 men: the MI estimate for SIPP-EHC is
substantially lower than the complete case estimate and appears more
in line with adjacent strata estimates.
{In this particular cell, there are 122 respondents with observed values, whereas on average there are 172 men in the imputed datasets (these observations were missing either employment status or earnings amounts -- age and sex are completely observed). }%
Evidently, the HCMM-LD\ uses information from other covariates to generate imputations that are more in line with what one would expect, {pulling down the median in this cell to a value more consistent with neighboring estimates}.
\begin{figure}
\begin{center}
\includegraphics[trim={0 0.5cm 0 0},width=0.95\textwidth]{fig/msipp-medagesex}
\end{center}
\vspace{-20pt}
\caption{Median earnings by age and sex in SIPP/SIPP-EHC computed in complete cases and multiply imputed datasets}
\label{fig:medagesex}
\end{figure}
Finally, Figure \ref{fig:snap-home} shows imputed and complete case
estimates of mean home values and SNAP benefits. We focus on means
because these distributions are less skewed than the earnings
distributions. Imputed mean home values are very similar to complete
case estimates, with the differences much smaller than relevant
standard errors.
The mean SNAP amounts are similar in the
imputed and corresponding complete case estimates, but for SIPP-EHC
the MI standard error is much higher than the complete case standard
error. This arises because of the relatively small sample
size; on average across imputations,
there are about 560 respondents in universe for SNAP benefit amounts but only
225 have observed values, so about 60\% of the values are
missing. Moreover, about 25\% of SNAP
recipients are missing earnings data in SIPP-EHC, and earnings are one
of the most important determinants of SNAP benefit amounts (the other
is household size, which is completely observed). Although the rate of
missing amounts in the production SIPP subsample is about the same, SIPP has a lower
rate of missing income data among SNAP recipients (16\% versus
25\%) and a larger sample size (ranging from 893-901 across imputed
datasets). These factors combine to allow for more precise
imputed estimates of SNAP benefits in the production SIPP subsample.
\begin{figure}
{\centering \includegraphics[trim={0 0.5cm 0 0},width=.45\linewidth]{fig/msipp-homeamt}
\includegraphics[trim={0 0.5cm 0 0},width=.45\linewidth]{fig/msipp-snapamt}
}
\caption{Estimates of mean home value and mean SNAP benefit in SIPP/SIPP-EHC computed in complete cases and multiply imputed datasets.}
\label{fig:snap-home}
\end{figure}
\subsection{Comparison of SIPP-EHC and production SIPP}
In their complete-case analyses, the Census Bureau researchers reported several
notable differences between the SIPP-EHC and production SIPP data \citep{USCensus2013}.
For variables with fairly low rates of item nonresponse (5\% or
less), of course, accounting for item nonresponse with MAR models is
not likely to alter these conclusions.
Our analyses of the variables with significant
missingness suggest nuanced conclusions about comparability. In
particular, even after adjusting for item nonresponse, the SIPP-EHC
respondents are more likely to report being employed during the
previous month and also to report lower mean
home values. Interestingly, Census Bureau researchers \citep{USCensus2013} linked complete cases
to administrative data and found that employment information was generally
more accurate in SIPP-EHC than in SIPP (these records
were not available to us). Across the two samples, differences in
earnings and mean SNAP benefits tend to be small relative to MI
standard errors, particularly for SNAP benefits where relying on
complete cases appears to underestimate standard errors.
Despite being drawn from the same frame and weighted to the same population, the two data sources do exhibit some
differences among completely observed variables. For example, in
SIPP-EHC, 56.7\% of the householders are female (SD 0.9\%), compared to
60.3\% (SD 0.8\%) in the production SIPP subsample. In SIPP-EHC, the mean householder age is
47.8 (SD 0.34) compared to 50.4 (SD 0.29) in the production SIPP sample, and the quartiles
show a similar difference of about 2 years. A number of factors may
contribute to these differences; one candidate is differential
attrition or unit nonresponse between SIPP and SIPP-EHC, which seems
plausible given the substantial differences in their designs. The Census
Bureau is continuing to examine possible sources for this discrepancy
(personal communication).
\section{Concluding Remarks}\label{sec:conclusion}
The repeated sampling simulation in Section \ref{sec:sipp-sim} demonstrates that the HCMM-LD\ can serve
as a reliable default multiple imputation engine. In fact, in the simulation
it often outperformed a default implementation of MICE, which is representative of the most widely used imputation routines.
Of course,
both MICE and the HCMM-LD\ could be modified to incorporate
dataset-specific prior knowledge -- and this is good practice -- but
in our experience many data users rely on default procedures.
We examined many other potential estimands in the simulation study.
For many estimands the difference between default MICE and the HCMM-LD\
are modest, but for others the improvement under
the HCMM-LD\ is substantial. We suspect that the performance gap to increase as the
sample size grows, because the differences appear to be driven mostly
by misspecification bias. Unlike default MICE, as a nonparametric
Bayesian joint model the HCMM-LD\ has the potential to increase in
complexity and capture additional features of the data.
We have not performed a systematic evaluation of the
properties of the HCMM-LD\ in large sample, high-dimensional settings;
this is an area for future research. We are optimistic about its
potential. Computationally, fitting the
HCMM-LD\ reduces to fitting a series of mixture and regression
models. Computational complexity scales roughly linearly with sample size, dominated by computing likelihoods when resampling
cluster indices.%
These steps could be optimized further in our existing implementation.
Increasing the dimension of the
categorical variables is clearly feasible; in an MPMN model \cite{Si2013}
considered simulations with 50 categorical
variables. Increasing the dimension of the continuous variables is more of a strain,
as the computation required grows quickly in
the dimension of the covariance matrices (for example, sampling the $H_i^{(\mathcal{Y})}$ for $1\leq i\leq n$ is $O(n{k^{\mathcal{(Y)}}} q^2)$).
Fitting large
mixtures of multivariate normals is a well-studied problem, however, and specialized, efficient algorithms exist to leverage
parallel computing architectures (e.g. \citet{suchard2010understanding}). These would be
straightforward to adapt to the HCMM-LD. {So too are alternative parameterizations of the component-specific covariance matrices,
such as factor-analytic forms in which $\Sigma_{\hy}= \Lambda_{\hy}\Lambda_{\hy}' + \Omega_\hy$ with $\Omega_\hy$ a diagonal $q\times q$ matrix and $\Lambda_{\hy}$ a $q\times b$ matrix of factor loadings (with $b<< q$). Such models reduce the number of free parameters and regularize the local covariances. They also render $Y_j$ and $Y_{j'}$ conditionally independent given some additional latent variables, simplifying imputation for $Y$.}
There are a number of interesting directions to extend the HCMM-LD. In
contexts with fully observed data, it can be advantageous to condition
on fully observed variables, such as design variables, so as not to
spend model fitting capacities on modeling these variables. Since the
HCMM-LD\ is modular in nature, it is conceptually possible to incorporate
such adaptations.
Similarly, the modular nature suggests that it should be possible to
adapt the model to incorporate other types of variables, such as counts or durations.
Finally, the current model does not account for structural zeros in contingency tables (from impossible
combinations or skip patterns), and linear restrictions
among continuous variables. We expect that it should be possible to
adapt the methods of \cite{Manrique-Vallier2012,
Manrique-Vallier2012a} to handle structural zeros and of
\cite{Kim2013} to handle linear constraints. Adapting these approaches to mixed data
is an active area of research.
|
\section{Introduction}
Any 2d quantum spin model with intrinsic topological order has a finite ground state degeneracy on a torus \cite{topo}.
This nontrivial ground state space features a distinguished basis which is given by the eigenvectors of the modular $T$-transformation ({Dehn} twist).
It is also this basis in which the topological $S$-matrix is expressed.
From the perspective of topological quantum field theory (\textsc{TQFT}) these distinguished basis states are in one-to-one correspondence with the (irreducible) charges that label the boundaries of the 2d surfaces in (2+1)d \textsc{TQFT}s.
In the corresponding lattice model, the charges label quasiparticle excitations which can be moved by string-like operators.
These operators have a surprisingly rich algebraic structure which encodes the mutual and self-statistics of the quasiparticle excitations, for example.
When wrapped around non-contractible loops, the string-like operators are also known as {Wilson} loop operators.
Furthermore, intrinsic topological order has been linked to different patterns of long-range entanglement under local unitary (LU) equivalence \cite{Chen:2010gb,Bachmann:2011kw}. For example, this manifests itself in the topological entanglement entropy which is a nonlocal quantity tied to bipartitions of ground states \cite{Hamma:2005p63,Kitaev:2006,Levin:2006ij}. Other entanglement measures show also topological contributions \cite{topoOther, topoGE1, topoGE2}. However, we do not know of a single measure which can even begin to capture all the intricate patterns of entanglement in a quantum many-body state\cite{Osterloh:2012gg}. In other words, our understanding of multipartite entanglement in quantum states is relatively poor at present.
Suppose we have access to a complete set of linearly independent ground states for a given 2d lattice model.
Tasked with identifying its universality class we will typically face several difficulties.
Firstly, the string-like operators encoding the properties of quasiparticle excitations are not known a priori.
Secondly, it may be impossible to actually perform {Dehn} surgery on many lattices.
However, it has been argued convincingly\cite{Dong:2008ce,Zhang:2012jc,Cincio:2013ku,Zhu:2013wt,Zhu:2014jf} that the distinguished basis states~$\set{\ket{\Xi_i}}$ are singled out by their entanglement properties. Namely, they minimise the entanglement entropy of certain non-contractible regions on the torus, which is why they have been called \emph{Minimally Entangled States} (MES). Thus we may neither need to know the string-like operators nor perform explicitly {Dehn} surgery to identify a topological universality class, provided the MES can be found.
The entanglement entropy is a measure of entanglement between
a region $A$ and its complement, denoted by $A^\perp$. It is defined by
the von Neumann entropy of the subsystem $A$, i.e., $S(\rho_A) = -{\rm tr} (\rho_A \log_2 \rho_A)$. Here the reduced density matrix $\rho_A={\rm
Tr}_{A^\perp}(|\Psi\rangle\langle\Psi|)$ is obtained by tracing out
the degrees of freedom in $A^\perp$. The entanglement entropy is bipartite in nature,
but each region $A$ or $A^\perp$ can contain many lattice sites. Rather than a bipartite measure of entanglement, here we
consider a multipartite measure, the
geometric entanglement (GE), and investigate how it can be used to
characterise a topological phase. Considering such a multipartite measure brings a number of advantages over the ``more usual'' bipartite entanglement, especially regarding its numerical evaluation. In particular, it was shown in Ref.\cite{topoGE2} that this quantity can be computed more efficiently \emph{and} more accurately than bipartite entanglement measures by means of a tensor network algorithm.
Motivated by the above, here we show that the GE evaluated on topologically
nontrivial partitions can be used to identify the topologically
distinguished basis of MES. This result has important implications both practical and fundamental. As said before, the
\textsc{GE} provides an alternative figure of merit which may prove
more convenient to evaluate numerically than other entanglement measures \cite{topoGE2}. Moreover, and from a fundamental perspective, the fact that the distinguished basis minimises the \textsc{GE} is a highly nontrivial statement about the structure of many-body entanglement in topologically ordered states.
The organisation of this paper is as follows: in Sec.II we review some preliminary concepts on MES, the GE, and the models that we will study. In Sec.III we provide our results for small-size systems, focusing on the doubled semion and doubled Fibonacci models. Sec. IV deals with large sizes for the toric code model using a tensor network approach. Finally, in Sec.V we provide the conclusions of our work and some perspectives for the future.
\section{Preliminary concepts}
Here we provide a brief introduction to the two main concepts to be discussed in this paper, namely: MES for topological 2d systems, and the geometric measure of entanglement for multipartite states.
\subsection{Minimally Entangled States}
Let us consider a 2d topological system on a torus. As is well-known, the system will display a topological degeneracy in the ground state subspace. Let us now consider the entanglement properties of the different states in the ground subspace by, say, considering the entanglement entropy $S(\rho_A)$ of a bipartition. As is widely believed, for smooth bipartitions with a contractible boundary of perimeter $L$, the entanglement entropy behaves like
\begin{equation}
S(\rho_A) = S_0 - S_{\gamma} + O(L^{-\nu}),
\end{equation}
with $S_0 \propto L$ (the so-called \emph{area-law} behaviour) and $S_{\gamma}$ the topological entanglement entropy, which for topological systems is a universal contribution and determines the presence of topological order by itself.
For bipartitions with a non-contractible boundary, the situation is quite different \cite{Zhang:2012jc}. In such a case, the topological contribution $S_{\gamma}$ actually depends on the specific choice of ground state within the ground subspace. Those states having the maximum topological component, or equivalently the minimum overall entanglement entropy, are called \emph{Minimal Entropy States}. For the sake of this paper, since we will be dealing with other entanglement measures rather than the entropy, we shall call them \emph{Minimally Entangled States}, or MES.
MES are interesting since, as is well-known by now \cite{Zhang:2012jc}, one can extract all the topological information about the system just from their mutual overlaps, e.g., $S$ and $T$ matrices. These states, which we call here $\set{\ket{\Xi_i}}$, are also from the distinguished topological basis discussed in the introduction.
\subsection{Geometric Entanglement}
Consider an $m$-partite normalised pure state $\ket{\Psi} \in \mathcal{H} =
\bigotimes_{i=1}^{m} \mathcal{H}^{[i]}$, where $\mathcal{H}^{[i]}$
is the Hilbert space of party $i$. For instance, in a system of $n$
spins each party could be a single spin, so that $m = n$, but could
also be a set of spins, either contiguous (a \emph{block}
\cite{geometric2}) or not. We wish now to determine how well the state
$\ket{\Psi}$ can be approximated by an unentangled (normalised)
state of the parties,
$\ket{\Phi}\equiv\mathop{\otimes}_{i=1}^{m}|\phi^{[i]}\rangle$. The
proximity of $\ket{\Psi}$ to $\ket{\Phi}$ is captured by their
overlap. The entanglement of $\ket{\Psi}$ is thus revealed by the
maximal overlap~\cite{ge},
$\Lambda_{\max}({\Psi})\equiv\max_{\Phi}|\ipr{\Phi}{\Psi}|$. The
larger $\Lambda_{\max}$ is, the less entangled is $\ket{\Psi}$. We
quantify the entanglement of $\ket{\Psi}$ via the quantity:
\begin{equation}
E_G({\Psi})\equiv-\log_2\Lambda^2_{\max}(\Psi), \label{eq:Entrelate}
\end{equation}
where we have taken the base-2 logarithm, and which gives zero for
unentangled states. $E_G(\Psi)$ is called \emph{geometric
entanglement} (GE). This quantity has been studied in a variety of
contexts, including critical systems and quantum phase transitions
\cite{geometric2, geometric3}, quantification of entanglement as a
resource for quantum computation \cite{resource}, local state
discrimination \cite{discrim}, and has been recently measured in NMR
experiments \cite{exper}. Also, one can choose the case of just two
sets of spins. In this case the GE $E_G(\Psi)$ coincides with the
so-called single-copy entanglement between the two sets,
\begin{equation}
\label{eq:single-copy}
E_1(\Psi)
=-\log_2
\nu_1(\rho),
\end{equation}
with $\nu_1(\rho)$ the largest eigenvalue of the reduced density
matrix $\rho$ of either set \cite{sc}.
The GE offers a lot of flexibility to study multipartite quantum
correlations in spin systems. For instance, one can choose each
party to be a single spin, but one can also choose blocks of
increasing boundary length $L$ \cite{geometric2}. Studying how the GE changes with $L$ provides information about how
close the system is to a product state under coarse-graining
transformations. What is more, one can
choose each block to consist of spins in non-contractible regions,
which is what we shall mainly use here in the investigation of
MES.
There have been two recent notable findings regarding the GE of quantum many-body states. The first of these results is that for renormalization group (RG) fixed points such as the toric code and other topological exactly-solvable models, the GE of blocks
exactly obeys $E_G = E_0 - E_{\gamma}$, with $E_{\gamma}$ a topological
contribution (the topological GE) and $E_0$ some non-universal term
\cite{topoGE1}. Moreover, it was observed that $E_{\gamma}$ coincided with the topological contribution to the entanglement entropy for
the considered models. As for $E_0$ it was found that $E_0 \propto
n_b L$, with $n_b$ the number of blocks with a boundary
of size $L$.
The second notable result has been the development of an efficient tensor network \cite{tn} algorithm based on Projected Entangled Pair States \cite{PEPS} to compute the GE for a torus partitioned into cylinders \cite{topoGE2}. This method was used to find sharp evidence of topological phase transitions in 2d systems with a string-tension perturbation. In fact, when compared to tensor network methods for R\'enyi entropies, this approach turned out to produce almost perfect accuracies close to criticality and, on top, was orders of magnitude faster than more ``standard'' R\'enyi entropy calculations.
In what follows we will show that the GE of topological ground states, when computed for a torus partitioned into cylinders, can be used to determine the distinguished topologically-preferred basis of MES. For systems with a small size we are able to do this almost exactly, i.e., without the need of any tensor network implementation. For larger systems, however, we make use of the algorithm introduced in Ref.\cite{topoGE2}. When combined with such a tensor network approach, the overall algorithm turns out to be a very efficient and precise way of identifying the MES for a topological 2d system.
{
Before we move on to describe the topological models, let us briefly discuss how one computes the GE in general. Here we describe an iterative
method to compute the maximal overlap, which has been described previously in the Supplemental Material of Ref.~\cite{exper} and was also discussed in Ref.~\cite{Bruss}. This
method can not only be implemented numerically, but can also be
carried out experimentally~\cite{exper}. To compute the maximal overlap for the
state $|\Psi\rangle$ with respect to product states
$|\Phi \rangle\equiv\bigotimes_{i=1}^{m}|\phi^{[i]}\rangle$, we use
the Lagrange multiplier $\lambda$ to enforce the constraint
$\langle \Phi|\Phi\rangle=1$,
\begin{equation}
f(\Phi)\equiv \langle \Phi|\Psi\rangle\langle \Psi |\Phi\rangle -\lambda
\langle \Phi|\Phi\rangle.
\end{equation}
Maximizing $f$ with respect to the local product state
$|\phi^{[i])}\rangle$, we obtain the extremal condition, originally derived in Ref.~\cite{ge},
\begin{equation}
{\cal H}_{\rm eff}^{[i]}|\phi^{[i]}\rangle = \lambda N^{[i]}
|\phi^{[i]}\rangle.
\end{equation}
Here the effective single-site Hamiltonian ${\cal H}_{\rm eff}^{[i]}\equiv (\bigotimes_{j\ne
i}^{m}\langle\phi^{[j])}|)|\Psi\rangle\langle \Psi|(\bigotimes_{j\ne
i}^{m}|\phi^{[j]}\rangle)$ is proportional to a local projector
$|\omega^{[i]}\rangle\langle\omega^{[i]}|$ at site $i$,
and the normalization $N^{[i]}\equiv\bigotimes_{j\ne
i}^{m}\langle\phi^{[j]}|\phi^{[j]}\rangle$
is unity if all the local states are properly normalized, as will be done in practice. From the
viewpoint of, e.g., a variational Matrix Product State (MPS), one fixes all local states
$|\phi^{[j]}\rangle$ but $|\phi^{[i]}\rangle$ and solves for the
corresponding optimal $|\phi^{[i]}\rangle$ and repeats the same
procedure for $i+1$, $i+2$, etc. until the $m$-th site and sweeps
the procedure back and forth until the eigenvalue $\lambda$
converges. The converged value $|\lambda|^2$ is the square of the
maximal overlap $\Lambda_{\max}^2$. To avoid getting trapped in possible local maxima, it is useful to repeat the procedure a few times with different initial random product states and use the largest overlap obtained.
}
\subsection{Topological Models}
Let us now briefly revisit some of the basic properties of the models to be studied in this paper, namely the toric code, doubled semion, and doubled Fibonacci models.
\subsubsection{Toric Code model}
The toric code \cite{Kitaev:2003} is the simplest example of a topologically non-trivial 2d system. It is the RG fixed point of the topological phase of a
$\mathbb{Z}_2$ gauge theory, and is equivalent under local transformations to a Levin-Wen model on a honeycomb lattice
\cite{Levin:2004p117}. The Hamiltonian of the toric code is given by
\begin{equation}
H_{{\rm TC}} = -\sum_s A_s -\sum_p B_p,
\end{equation}
where star operators $A_s$ and plaquette operators $B_p$ are defined as
\begin{equation}
A_s \equiv \prod_{j \in s} \sigma_x^{[j]} \ \ \ \ \ \ \ \ \ \ \ \ \ B_p \equiv \prod_{j \in p} \sigma_z^{[j]},
\end{equation}
with $\sigma_{\alpha}^{[j]}$ the $\alpha$-th Pauli matrix at link $j$ of the lattice. The properties of this model are well-known, including its MES \cite{Zhang:2012jc}, and its robustness to perturbations \cite{robust}.
\subsubsection{Doubled Semion Model}
The doubled semion model \cite{doublesemion} is given by the spin model on the
honeycomb lattice
\begin{equation}
H_{\rm DS}=-\sum_s A_s -\sum_p B'_p,
\end{equation}
where $A_s$ and $B'_p$ are mutually commuting and given by
\begin{equation}
A_s \equiv \prod_{j \in s} \sigma_x^{[j]} \ \ \ \ \ \ \ \ B'_p \equiv - \prod_{k \in {\rm legs \ of}\ p} i^{(1-\sigma_x^{[k]})/2} \prod_{j \in p} \sigma_z^{[j]}.
\end{equation}
As in the toric code, star and plaquette operators satisfy the non-local constraint
\begin{equation}
\prod_s A_s = \prod_p B'_p = \mathbb{I}.
\end{equation}
This model is known to be topologically ordered, corresponding to the universality class of a $U(1) \times U(1)$ Chern-Simons theory. The model is exactly solvable, and MES are also well known \cite{Levin:2004p117}.
\subsubsection{Doubled Fibonacci Model}
The doubled Fibonacci model \cite{Levin:2004p117} is defined on a honeycomb lattice with anyonic degrees of freedom on its edges. These degrees of freedom have two different states (say, $\ket{0}$ and $\ket{1}$), i.e., like a spin-1/2 model, but the overall Hilbert space is restricted to configurations that satisfy the Fibonacci branching rules at every vertex:
\begin{eqnarray}
0 \times q = q \times 0 &=& q ~~{\rm for}~~q \in \{ 0, 1 \} , \\ \nonumber
1 \times 1 &=& 0+1.
\end{eqnarray}
The Hamiltonian of the model is given by
\begin{equation}
H_{{\rm DF}} = -\sum_p \delta_{\Phi(p),0},
\end{equation}
where $\delta_{\Phi(p),0}$ is a projector on the states with zero flux $\Phi(p)$ through plaquette $p$. Again, the model is exactly solvable and many of its properties are well-known.
\section{MES from GE: small sizes}
In this section we consider the doubled semion and doubled Fibonacci models \cite{Levin:2004p117}. Both models have a topological ground state degeneracy of $4$ when placed on the torus, hence $\setbuilder{\ket{\Xi_i}} {0\leq i\leq 3}$. Small-size calculations were also performed for the toric code model \cite{Kitaev:2003}, but these produced equivalent results to those of the doubled semion model and are therefore not reproduced here. We shall, however, come back to the toric code model when considering large lattice sizes with tensor network methods in Sec.IV. For the geometric entanglement we partition the torus into $n_\mathrm{h}$ ($n_\mathrm{v}$) cylinders of equal width in horizontal (vertical) direction and define product states with respect to those cylinders. Clearly, the results for $n_{\mathrm{h},\mathrm{v}}\geq 3$ cylinders are the most interesting ones since these are beyond bipartite entanglement (recall that for the case of $2$ cylinders, the GE corresponds exactly to the infinite-R\'enyi entropy, or single-copy entanglement). Considering the three spins of a vertex as unit cell, a $3 \times 3$ honeycomb lattice on a torus would look like the one in Fig.~\ref{FigLatTopo}, i.e., with 27 spins on the whole. The lattice in this figure is also partitioned in three cylinders along the horizontal direction, for the sake of clarity.
\begin{figure}
\includegraphics[width=0.25\textwidth]{Fig0Top.pdf}
\caption{(Color online) $3 \times 3$ honeycomb lattice on a torus, where the unit cell is composed of the three spins around a vertex, and projected on the 2d plane. On the whole, there are 27 spins. The lattice is partitioned in 3 vertical cylinders. Spins on each cylinder have a different color on the link (blue, black, golden).}
\label{FigLatTopo}
\end{figure}
\subsection{Random sampling}
{
Our first procedure for these small-size systems works as follows. Given a complete set of
four linearly independent states $\{|G_i\rangle\}$ in the ground space ($i=0,1,2,3$), we shall sample a large set of random quantum states from this subspace
\begin{equation}
\ket{\psi} := \sum_{j =0}^3 c_ {j} \ket{G_j},
\end{equation}
and investigate their entanglement distribution. Let us remark that the coefficients $c_j$ are obtained from a column of random unitary matrices in $U(4)$ sampled according to the Haar measure. If $\ket{G_j}$'s are not orthonormal, we will first orthonormalize them and then superpose them with the random coefficients $c_j$'s.
We shall then attempt to identify the set of MES from those ground state samples with smallest GE. For the toric code and the doubled semion models, the set $\{\ket{\Xi_i}\}$ is fairly straightforward to write down, so we will actually take $\ket{G_j}=\ket{\Xi_i}$ for convenience and thus we have the benefit of being able to directly check the accuracy of our numerical results.
}
\subsection{Systematic minimisation}
{
Our second procedure consists of systematically minimising the GE in order to find the distinguished topological basis for this model. Here we do not assume the four linearly independent ground states $\ket{G_i}$ to be orthonormal, since orthonormality can be taken into account by considering
the overlap matrix $C_{ij}\equiv \langle G_j|G_i\rangle$ and diagonalizing it as $C_{ij} = (U^\dagger
\lambda U)_{ij}$, where $\lambda$ is a (positive) diagonal matrix
and $U$ is the unitary matrix that diagonalizes $C^T$.
We would like to find the set of MES~\cite{Zhang:2012jc,Zhu:2013wt,Zhu:2014jf} characterized not by the entanglement entropy but by the GE.
First we need to find a state with minimum GE within the full four-dimensional ground space, with an arbitrary state in it prametrized by $\sum_i
a_i |G_i\rangle$. Finding the minimum is generally a nonlinear optimization problem and for our purpose here we employ the Nelder-Mead simplex method \cite{NelderMead}.
Suppose that we found a global minimum GE state $|\Psi_0\rangle=\sum_i a_i^{[0]} |G_i\rangle$, how would we proceed for the remaining MES?
We need to impose orthogonality in the subsequent minimisation and it can be done as follows. Suppose we have two states
$|\Psi\rangle=\sum_i a_i |G_i\rangle$ and $|\Psi'\rangle=\sum_i
b_i|G_i\rangle$. Their overlap is $\langle \Psi'| \Psi \rangle= \sum_{ij}
b_j^* a_i \langle G_j|G_i\rangle=\vec{b}^{*}\cdot C \cdot \vec{a}=
\vec{b}^* U^\dagger \sqrt{\lambda}\sqrt{\lambda} U\vec{a}$. It is
convenient to define a matrix $T\equiv \sqrt{\lambda} U$ so that
$\langle \Psi'| \Psi \rangle=\vec{b}^* T^\dagger T\vec{a}$. Orthogonality is then imposed by restricting the parameters of $\vec{b}$ (used in the optimization program)
to those satisfying $\vec{b}^*T^\dagger T\vec{a}=0$.
We denote by $|\Psi_1\rangle=\sum_i a_i^{[1]} |G_i\rangle$ the state with minimum entanglement in
the subspace that is orthogonal to $|\Psi_0\rangle$, using the procedure described above. We continue to
find the next state $|\Psi_2\rangle$ with minimum entanglement restricted to be
orthogonal to the subspace spanned by $|\Psi_0\rangle$ and
$|\Psi_1\rangle$. This can be done via restricting
the parameters $\vec{c}$ used in the optimization to those satisfying both $\vec{c}^*T^\dagger T\vec{a}^{[0]}=0$ and $\vec{c}^*T^\dagger T\vec{a}^{[1]}=0$.
}
Having found the first three orthogonal states, the
fourth state $|\Psi_3\rangle$ is automatically determined. These four states would then constitute our candidates for the basis of MES.
\subsection{Results}
\subsubsection{Doubled semion model}
{
Calculations have been performed for the \emph{same} uniformly random sample~$\mathfrak{S}$ of 196608 states $\sum_{j=0}^3 c_{ij} \ket{\Xi_j}$ from the same set of random unit vectors in $\mathbb{C}^4$ but for different system sizes. These states are called logical states because they are effectively encoded two-qubit states on a torus. For 3 or more cylinders we obtain an approximation in terms of numerical upper \emph{bounds} on the GE via 10 lower bounds on the overlap~\footnote{The reason we refer to these as lower bounds is because numerically we cannot certify that the overlap from our procedure is absolutely maximum, hence only a lower bound on GE.}.
}
In order to obtain each lower bound, we draw a random product state and update the overlap one cylinder (party) at a time until convergence is reached. This algorithm yields a non-decreasing sequence of overlaps, i.e. a local maximum of the overlap, and is in fact the usual approach used also in tensor network methods to compute approximations to the GE~\cite{topoGE2}.
\begin{figure}
\includegraphics[width=0.5\textwidth]{Fig1Top}
\caption{(Color online) The overlaps $\abs{\braket{\psi_i}{\psi_j}}$ (for $i\geq j$) of the 64 logical states with smallest \textsc{GE} for the doubled semion model, for (a) horizontal and (b) vertical cylinders on $2\times 2$, $3\times 3$ and $3\times 4$ honeycomb lattices (27 and 36 spins, respectively). The logical states with lower GE do not depend on the size of the lattice, hence we get the same two plots for the three sizes.}
\label{fig:overlaps}
\end{figure}
\bigskip
\underline{\emph{(a) Small GE states.-}}
\bigskip
For each encoded state of the sample~$\mathfrak{S}$, we compute its total GE for a partition into cylinders for some given size of the system, and then sort the states by increasing GE value. Importantly, we have seen that the permutation to perform this sorting, for a given direction of the cylinders, is independent of the system size. Once the states are sorted, we pick those states with smallest GE, and study their properties.
First we analyze how orthogonal the small \textsc{GE} states are. Since the $\ket{\Xi_j}$ are orthonormal the inner product evaluates to $\braket{\psi_i}{\psi_j} = \sum_{k=0}^3 c_{ik}^* c_{jk}$. The overlaps are shown in Fig.~\ref{fig:overlaps}. Clearly these are either very close to one, or very close to zero. This picture is consistent with low-entanglement states: whenever the overlap of two states is close to one we interpret that these states are close to the same MES $\ket{\Xi_i}$, whereas whenever the overlap is close to zero we interpret that these are close to two different and orthogonal MES $\ket{\Xi_i}$ and $\ket{\Xi_j}$.
Given this we can extract a quasi-orthogonal basis for the ground state subspace from the smallest \textsc{GE} states. We do this by separating the states into sets of $k$ states which are almost identical, i.e., the overlap of any two states within a set is $ \ge 0.9$, and choosing the smallest $k$ that yields four sets. Then, from each set we pick the state with the smallest \textsc{GE}. In this way we build our four candidates for MES. We find that these always have almost maximum overlap with exactly one of the four states $\ket{\Xi_i}$, and almost zero with the remaining ones. In this way, we certify that we indeed found a very good approximation to the correct basis of MES by looking at states with minimal \textsc{GE}.
{
In fact, the basis of MES found in this way agrees, with good accuracy, with the one found using the systematic minimisation procedure described previously (and will be discussed in more detail for the case of the doubled Fibonacci model). Using systematic minimisation, we report here that we have successfully identified the correct four MESs, all with $E_G=6$ on the $3\times3$ honeycomb (same for the toric code), partitioned into three cylinders. From this set of states we can extract the modular matrices $S$ and $T$, if needed. It is important to stress that the four MES are the only four with the lowest GE values; any other states than the four have higher values of GE.
}
\begin{figure}
\includegraphics[width=0.45\textwidth]{Fig2Top}
\caption{(Color online) The \textsc{GE} (blue) versus single-copy entanglement (red) histograms for horizontal and vertical cylinders on honeycomb lattices of different sizes, for the doubled semion model: (a) 27 spins and 3 vertical cylinders, (b) 36 spins and 3 vertical cylinders, (c) 27 spins and 3 horizontal cylinders, and (d) 36 spins (as in (b)) and 4 horizontal cylinders.}
\label{fig:histograms_minentropy}
\end{figure}
We conclude, thus, that the \textsc{GE} correctly identifies the quantum states of the topologically distinguished basis. Let us stress that similar results were also found for the toric code model (not shown).
\bigskip
\underline{\emph{(b) Comparison to single-copy entanglement.-}}
\bigskip
\label{sec:ge_min_entropy}
One could be tempted to say that the \textsc{GE} of a state with respect to a partition into, say, $n$ cylinders on a torus, should be approximately given by $E_G \sim 2(n-1)S^{(\infty)}$ where $S^{(\infty)} = - \log \nu_1$ is the infinite-R\'enyi entropy or single-copy entanglement \cite{sc}, and $\nu_1$ is the largest eigenvalue of the reduced density matrix of a bipartition into two cylinders. This is inspired by a picture based on matrix product states (\textsc{MPS}) \cite{tn}, where the two boundaries of each cylinder should effectively decouple \emph{up to a topological contribution}, once the cylinder is wider than the correlation length. Clearly, for $n=2$ cylinders $S^{(\infty)}$ is the exact GE value, so we should regard the GE as a natural multipartite generalisation of the bipartite infinite-R\'enyi entropy. Still, the relation between them in the multipartite scenario \emph{and} for topological models is not so clear.
Here we try to throw and cast a bit of light on this question, by comparing the values of $E_G$ and $S^{(\infty)}$ for the states in our sample, for small lattice sizes of the doubled semion model. In Fig.~\ref{fig:histograms_minentropy} we can see a comparison of the histograms of states of our sample, i.e., the ratio of states with a given value of either the GE or the single-copy entanglement, for different partitions and lattice sizes, and cylinders of one lattice-site width. Our results seem to be in agreement with a global offset between the GE and the single-copy entanglement for all the states in the sample (which correspond to different ground states), far away from the conjectured (quasi)-decoupled regime for wide cylinders. In fact, our data for this model is consistent with the behaviour
\begin{equation}
E_G = S^{(\infty)}+ (n- 2) (L-1) ~~~{\rm (conjecture)}
\end{equation}
for cylinders of one unit cell width, where $n$ is the number of cylinders and $L$ their circumference (in unit cells). Such a dependence might very well be different once wider cylinders are considered, once a perturbation is added to the topological model, and once the spectrum of eigenvalues of the reduced density matrix stops being flat. It would also be of interest to know if such a relation, or similar, holds for models with more complex types of topological order.
{
Finally, let us remark that we do not necessarily need to restrict ourselves to consider the torus to be obtained from a square, but also $n$ cylinders with each having length $L$ and width $w$.
}
\subsubsection{Doubled Fibonacci model}
In order to analyze the properties of this model we used similar methods as in the previous section. However, and as we shall see, the systematic minimisation method becomes quite important here since some of the states in the topologically distinguished basis do not correspond to a global minimum in entanglement, given the non-Abelian nature of the system.
{This is in contrast to what we have found earlier in the toric code and doubled semion models, where all the states in the topologically distinguished basis, i.e., the MES, have the same global minimum value of GE.
}
\begin{figure}
\includegraphics[width=0.45\textwidth]{Fig3Top.pdf}
\caption{(Color online) GE in the linear superposition of $\cos\theta|\Psi_1\rangle+\sin\theta\,e^{i\phi}|\Psi_2\rangle$ for 3 cylinders.
At $\theta=0,\pi/2$, the corresponding states are $\ket{\Psi_1}$ and $\ket{\Psi_2}$, respectively (regardless of $\phi$). The figure shows that these two states are the only two minimum entangled states in this subspace, and any other superposition will necessarily have higher entanglement. Curiously, the entanglement is independent of the angle $\phi$.}
\label{fig:states23H2x2}
\end{figure}
\bigskip
\underline{\emph{(a) Small GE states and non-Abelian character.-}}
\bigskip
{
Earlier, we have applied the first method of random sampling to the doubled Fibonacci model and found that, unlike in the toric code and doubled semion models, there is only \emph{one} global minimum \textsc{GE} state. Upon closer inspection this fact is less surprising, since the corresponding entanglement \emph{entropy} depends on the quantum dimension of each distinguished basis state $\ket{\Xi_i}$. It is known that for this model the ground-state space $\mathcal{L}$ decomposes as $\mathcal{L}_1\oplus\mathcal{L}_\phi\oplus\mathcal{L}_{\phi^2}$ (where $\phi$ is the golden section), where the dimension of $\mathcal{L}_1$ and $\mathcal{L}_{\phi^2}$ is one and the dimension of $\mathcal{L}_\phi$ is two. From such a structure we may expect to find three distinct GE values for the MES, i.e., with the two topologically distinguished states in $\mathcal{L}_\phi$ having the same value of GE. If our observation is true, then this implies that the \textsc{GE} can actually be used to extract the number of quasiparticle excitations in each sector from the number of quasi-orthogonal states (i.e. subset of MES) and their \textsc{GE} values.
Motivated by the previous observation, we thus carry out the systematic minimisation procedure in order to find the distinguished topological basis for this model. Specifically, we considered lattices of 27 ``spins'' partitioned into 3 cylinders. Proceeding in this way, the four MES obtained via GE have
very similar (within less than 1\% error) coefficients to those
obtained via the entanglement entropy in a $2\times2$ system.
The specific values of GE that we find for these four states are $E_G(\Psi_0) = 4.8443$, $E_G(\Psi_1) = 6.5698$, $E_G(\Psi_2) = 6.5698$, and $E_G(\Psi_3) = 7.7303$. The fact that $|\Psi_1\rangle$ and $|\Psi_2\rangle$ possess almost the same GE confirms our expectation that there are indeed two MES with the same GE value, implying $|\Psi_1\rangle$ and $|\Psi_2\rangle$ are in
$\mathcal{L}_\phi$. As an illustration we plot in Fig.~\ref{fig:states23H2x2} the numerical landscape of entanglement in the $\mathcal{L}_\phi$ subspace, showing interesting features.
Moreover,
it appears that in our basis $|\Psi_2\rangle=
(|\Psi_1\rangle)^*$. Moreover, $|\Psi_0\rangle
\in\mathcal{L}_1$ with the lowest entanglement, and $|\Psi_3\rangle \in \mathcal{L}_{\phi^2}$ having the largest entanglement among the MES.
Furthermore we find the product of modular matrices $TS$ to be accurate up to an error of $< 1\%$, once global phases of the MES have been fixed:
\begin{widetext}
\begin{equation}
TS \approx
\begin{pmatrix}
+0.276 + 0.003 i &+ 0.448 & +0.445 + 0.001 i & + 0.724 - 0.001 i \\
-0.362 - 0.264 i & +0.224 + 0.160 i &-0.585 - 0.427 i & +0.361 + 0.262 i\\
-0.363 + 0.261 i & -0.586 + 0.424 i & +0.224 - 0.163 i & + 0.362 - 0.263 i \\
+0.723 - 0.002 i & -0.448 + 0.003 i & -0.447 + 0.001 i & +0.276
\end{pmatrix}.
\end{equation}
\end{widetext}
}
\section{MES from GE: large sizes}
To extend the study of MES to larger system sizes, we employ
a tensor network construction of the ground state. In our case we use PEPS,
which have proved to represent faithfully ground states of
topological models \cite{topoPEPS}. In some situations, an analytic derivation of
the tensors can be obtained directly from the Hamiltonian. In general, though, the tensors
are obtained after a numerical optimisation of the energy. The numerical contraction of
a PEPS allows a precise and efficient calculation of the
geometric entanglement even in the case of blocks, as explained in Ref.~\cite{topoGE2}.
\subsection{Reminder of the tensor network method}
We set up a PEPS on a square lattice with periodic boundary
conditions. The tensors represent the state $|\Psi(n,L)\rangle$,
where $n$ and $L$ are the sizes of the torus defined by the boundary conditions.
The GE is obtained from the set $|\Phi\rangle$ of $n$ states
of length $L$ that cover the entire torus across cylinders. These states are one-dimensional
and have periodic boundary conditions, therefore we approximate them by MPS with periodic boundary conditions, see Fig.~\ref{fig:peps_diag}(a).
Our goal is the optimisation of the MPS states so that the overlap
$$
\Lambda_{max} = \frac{|\langle\Phi|\Psi\rangle|}{\sqrt{\langle\Psi|\Psi\rangle\langle\Phi|\Phi\rangle}}
$$
is maximised. The quantities $\langle\Phi|\Psi\rangle$ and $\langle\Psi|\Psi\rangle$ can be efficiently estimated
using standard PEPS methods. For large systems (say, $L>20$) this calculation can be achieved very efficiently using the MPS description for each of the states $|\phi^{[i]}\rangle$ \cite{topoGE2}.
We remind here how the optimisation of the product state $|\Phi\rangle$
can be performed for a torus of size $n\times L$ (see Fig.~\ref{fig:peps_diag}(b) as a reference).
The key observation is that the optimisation is performed on each $|\phi^{[i]}\rangle$ iteratively,
optimising a single state in each step and sweeping along the torus until convergence is reached.
Following Ref.~\cite{topoGE2}, the procedure can be summarised as follows:
\bigskip
\emph{1.-} We start with a random choice for each $|\phi^{[i]}\rangle$ as the initial state of the optimisation.
\bigskip
\emph{2.-} Choose a position $k$, and by fixing all the remaining states optimise a new $|\phi^{[k]}\rangle$.
This state is obtained after contracting the entire PEPS, keeping the tensor structure to form the new $|\phi^{[k]}\rangle$.
For small systems one can perform a single contraction of the tensors for each position $i \neq k$ and continue the optimisation
as a purely 1d problem.
\bigskip
\emph{3.-} Move to position $k+1$. Sweep along the torus iterating these steps until convergence in the overlap is obtained.
In our simulations convergence is obtained after a few sweeps along the torus.
\bigskip
\subsection{Results: toric code model}
As an archetypical example of topological order we analysed here the toric code model with our tensor network method. Since this is the simplest 2d model with topological order, we have a very good control over its representation in terms of PEPS. We focus our attention on the toric code model
on a square lattice with periodic boundary conditions.
\begin{figure}
\includegraphics[width=0.5\textwidth]{Fig4Top.pdf}
\caption{(Color online) Diagrammatic representation of the optimisation performed to obtain the GE:
(a) we first set up a random set of 1d MPS $|\phi^{[i]}\rangle$ of bond dimension $\chi$ for all cylinders $i$ forming $|\Phi\rangle$. Each
of these states covers the torus along one direction. (b) At each step we pick a position $k$
where we obtain a new $|\phi^{[k]}\rangle$ after contracting the rest of the torus tensors. This
procedure is repeated sweeping back and forth along the torus until convergence is obtained. In the figure, $d$ is the physical dimension of the sites.}
\label{fig:peps_diag}
\end{figure}
\bigskip
\underline{\emph{(a) Small GE states from tensor networks.-}}
\bigskip
The contraction of PEPS allows a precise calculation of
the GE to be used in order to identify the
MESs. Using the tensor network construction, the states $|00\rangle$
and $|10\rangle$ (we follow the notation introduced in Ref.~\cite{topoGE1})
can be created easily in the PEPS picture with a small bond dimension $D=2$.
The MES $|\Xi_0\rangle =2^{-1/2}(|00\rangle+|10\rangle)$ can easily be constructed using a
bond dimension $D=4$. These conditions allow an optimal description of the problem as a PEPS,
and result in an efficient iterative search of the maximal overlap.
In order to show how to the GE can be used to clearly identify a MES with the tensor network approach for large system sizes, we
study the parametrization
$$
|\theta,\phi\rangle = \cos(\theta)|00\rangle + \sin(\theta)e^{i\phi}|10\rangle.
$$
The GE for this state is represented in Fig.~\ref{fig:ge_peps}(a) as a function of both $\theta$ and $\phi$.
The minimum of the GE appears at $\theta=\pi/4$, corresponding to
the superposition $|\Xi_0\rangle =2^{-1/2}(|00\rangle+|10\rangle)$.
A similar exploration can be performed for the superposition of $|00\rangle$ and $|01\rangle$, which is shown in
Fig.~\ref{fig:ge_peps}(b). In this case however, no MES is found and the GE increases for any value of
$\theta$ and $\phi$, as expected from entropy calculations \cite{Zhang:2012jc}. The two plots in Fig.~\ref{fig:ge_peps} are obtained for a system of $4\times 4$ sites.
Identical results are obtained for larger systems up to some overall displacement in the amount of GE due only to the size of the system. In practice, we checked this for systems up to $\sim 20 \times 20$, with no change in the conclusions.
\begin{figure}
\includegraphics[width=0.3\textwidth]{Fig5Top.pdf}
\caption{(Color online) (a) GE for the state $|\theta,\phi\rangle = \cos\theta|00\rangle + \sin\theta\,e^{i\phi}|10\rangle$.
We find a minimum at $\theta=\pi/4$ corresponding to $|\Xi_0\rangle =2^{-1/2}(|00\rangle+|10\rangle)$. (b)
The superposition of states $|00\rangle$ and $|01\rangle$ yields a larger value of the GE for any value of \
$\theta$ and $\phi$, with the maximum located at precisely $\theta=\pi/4$.}
\label{fig:ge_peps}
\end{figure}
\begin{figure}
\includegraphics[width=0.33\textwidth]{Fig6Top.pdf}
\caption{(Color online)
{
Evolution of the absolute error in the gradient method while searching for the state with minimum GE
in the space defined by $\theta$ and $\phi$ in $|\theta,\phi\rangle = \cos\theta|00\rangle + \sin\theta\,e^{i\phi}|10\rangle$, with (a) $\epsilon_\theta \equiv |\theta-\pi/4|$ (fixing $\phi = 0$) and (b) $\epsilon_{(\theta, \phi)} \equiv | (\theta, \phi) - (\pi/4,0)|$. The size of the system is $4\times 20$.}}
\label{fig:ge_peps2}
\end{figure}
\bigskip
\underline{\emph{(b) Systematic minimisation with tensor networks.-}}
\bigskip
Since we have a way to determine the GE for large lattices using PEPS, we can actually identify the MESs via GE using some numerical minimisation algorithm. We checked this by optimising over the 2-parameter space spanned by $\theta$ and $\phi$ shown in Fig.~\ref{fig:ge_peps}. This parameter space
can be explored by, e.g., a gradient method, in order to identify the minimum GE corresponding to MES.
We show in Fig.~\ref{fig:ge_peps2} the result of such an optimisation: starting from a random state, the value of GE evolves along the optimisation for the
superposition of $|00\rangle$ and $|10\rangle$ in a system of size $4\times 20$. The optimal value $\theta_{opt}=\pi/4$ is obtained
with an error $<10^{-8}$ after only a few iterations, being of the order of the precision imposed on the minimiser. Even though we only show here the combination
of two states of the ground state space, our results clearly suggest that this process can be extended to the full basis of ground states and to larger parameter spaces in order to perform a full search if required.
\section{Conclusions and outlook}
In this paper we have shown that the GE can be used as a powerful and useful tool to identify the distinguished basis of MES on a 2d system with topological order. We have seen this with calculations for the toric code, doubled semion and doubled Fibonacci models. Large-scale calculations for the toric code model have been done using a recently proposed tensor network approach \cite{topoGE2}. Our results for the doubled Fibonacci model also show that, indeed, it is possible to read off the number of Abelian quasiparticle excitations in the topological model directly from the optimisation procedure. Moreover, the results of this paper provide a very straightforward and efficient way of determining the topological properties of a strongly correlated system, especially when combined with the tensor network numerical approach. It would be very interesting to apply the ideas and methods discussed in this paper to other interesting 2d topological models, especially those having non-Abelian quasiparticle excitations. This will be the subject of future investigations.
\bigskip
{\bf Acknowledgements:} Discussions with B. Bauer, F. Pollmann and G. Vidal are acknowledged. T.-C.W. and A.G.-S. acknowledge the support by the National Science Foundation under Grants No. PHY 1314748 and No. PHY 1333903. This research was supported in part by Perimeter Institute for Theoretical Physics. Research at Perimeter Institute is supported by the Government of Canada through Industry Canada and by the Province of Ontario through the Ministry of Research and Innovation.
|
\section{Introduction}
Let $P$ be a set of $n$ points in the plane. For any two points $p,q\in P$, let $\CD{p}{q}$ denote the closed disk which has the line segment $\overline{pq}$ as diameter. Let $|pq|$ be the Euclidean distance between $p$ and $q$.
The {\em Gabriel graph} on $P$, denoted by $GG(P)$, is defined to have an edge between two points $p$ and $q$ if $\CD{p}{q}$ is empty of points in $P\setminus\{p,q\}$. Let $C(p,q)$ denote the circle which has $\overline{pq}$ as diameter. Note that if there is a point of $P\setminus\{p,q\}$ on $C(p,q)$, then $(p,q)\notin GG(P)$. That is, $(p,q)$ is an edge of $GG(P)$ if and only if $$|pq|^2<|pr|^2+|rq|^2\quad\quad \forall r\in P,\quad\quad r\neq p,q.$$
Gabriel graphs were introduced by Gabriel and Sokal \cite{Gabriel1969} and can be computed in $O(n\log n)$ time \cite{Matula1980}. Every Gabriel graph has at most $3n-8$ edges, for $n\ge 5$, and this bound is tight \cite{Matula1980}.
A {\em matching} in a graph $G$ is a set of edges without common vertices. A {\em perfect matching} is a matching which matches all the vertices of $G$.
In the case that $G$ is an edge-weighted graph, a {\em bottleneck matching} is defined to be a perfect matching in $G$ in which the weight of the maximum-weight edge is minimized. For a perfect matching $M$, we denote the {\em bottleneck} of $M$, i.e., the length of the longest edge in $M$, by $\lambda(M)$. For a point set $P$, a {\em Euclidean bottleneck matching} is a perfect matching which minimizes the length of the longest edge.
In this paper we consider perfect matching and bottleneck matching admissibility of higher order Gabriel Graphs. The {\em order-$k$ Gabriel graph} on $P$, denoted by \kGG{k}{}, is the geometric graph which has an edge between two points $p$ and $q$ iff $\CD{p}{q}$ contains at most $k$ points of $P\setminus\{p,q\}$. The standard Gabriel graph, $GG(P)$, corresponds to \kGG{0}{}. It is obvious that \kGG{0}{} is plane, but \kGG{k}{} may not be plane for $k\ge 1$. Su and Chang \cite{Su1990} showed that \kGG{k}{} can be constructed in $O(k^2n\log n)$ time and contains $O(k(n-k))$ edges. In \cite{Bose2013}, the authors proved that \kGG{k}{} is $(k+1)$-connected.
\subsection{Previous Work}
For any two points $p$ and $q$ in $P$, the {\em lune} of $p$ and $q$, denoted by $L(p,q)$, is defined as the intersection of the open disks of radius $|pq|$ centred at $p$ and $q$.
The {\em order-$k$ Relative Neighborhood Graph} on $P$, denoted by \kRNG{k}{}, is the geometric graph which has an edge $(p,q)$ iff $L(p,q)$ contains at most $k$ points of $P$.
The {\em order-$k$ Delaunay Graph} on $P$, denoted by \kDG{k}{}, is the geometric graph which has an edge $(p,q)$ iff there exists a circle through $p$ and $q$ which contains at most $k$ points of $P$ in its interior.
It is obvious that $$\text{\kRNG{k}{}}\subseteq\text{\kGG{k}{}}\subseteq\text{\kDG{k}{}}.$$
The problem of determining whether a geometric graph has a (bottleneck) perfect matching is quite of interest. Dillencourt showed that the Delaunay triangulation (\kDG{0}{}) admits a perfect matching \cite{Dillencourt1990}. Chang et al. \cite{Chang1992} proved that a Euclidean bottleneck perfect matching of $P$ is contained in \kRNG{16}{}.\footnote{They defined \kRNG{k}{} in such a way that $L(p,q)$ contains at most $k-1$ points of $P$.} This implies that \kGG{16}{} and \kDG{16}{} contain a (bottleneck) perfect matching of $P$. In \cite{Abellanas2009} the authors showed that \kGG{15}{} is Hamiltonian which implies that \kGG{15}{} has a perfect matching.
Given a geometric graph $G(P)$ on a set $P$ of $n$ points, we say that a set $K$ of points {\em blocks} $G(P)$ if in $G(P\cup K)$ there is no edge connecting two points in $P$, in other words, $P$ is an independent set in $G(P\cup K)$.
Aichholzer et al.~\cite{Aichholzer2013} considered the problem of blocking the Delaunay triangulation (i.e. \kDG{0}{}) for $P$ in general position. They show that $\frac{3n}{2}$ points are sufficient to block DT($P$) and at least $n-1$ points are necessary. To block a Gabriel graph, $n-1$ points are sufficient, and $\frac{3}{4}n-o(n)$ points are sometimes necessary \cite{Aronov2013}.
In a companion paper \cite{Biniaz2014}, we considered the matching and blocking problems in triangular-distance Delaunay (TD-Delaunay) graphs. The {\em order-$k$ TD-Delaunay graph}, denoted by \kTD{k}{}, on a point set $P$ is the graph whose convex distance function is
defined by a fixed-oriented equilateral triangle. Then, $(p,q)$ is an edge in \kTD{k}{} if there exists an equilateral triangle which has $p$ and $q$ on its boundary and contains at most $k$ points of $P\setminus\{p,q\}$. We showed that \kTD{6}{} contains a bottleneck perfect matching and \kTD{5}{} may not have any. As for maximum matching, we proved that \kTD{1}{} has a matching of size at least $\frac{2(n-1)}{5}$ and \kTD{2}{} has a perfect matching (when $n$ is even). We also showed that $\lceil\frac{n-1}{2}\rceil$ points are necessary and $n-1$ points are sufficient to block \kTD{0}{}. In \cite{Babu2013} it is shown that \kTD{0}{} has a matching of size $\lceil\frac{n-1}{3}\rceil$.
\subsection{Our Results}
In this paper we consider the following three problems: (a) for which values of $k$ does every \kGG{k}{} have a Euclidean bottleneck matching of $P$? (b) for a given value $k$, what is the size of a maximum matching in \kGG{k}{}? (c) how many points are sufficient/necessary to block a \kGG{k}{}? In Section~\ref{preliminaries} we review and prove some graph-theoretic notions. In Section~\ref{bottleneck-section} we consider the problem (a) and prove that a Euclidean bottleneck matching of $P$ is contained in \kGG{10}{}. In addition, we show that for some point sets, \kGG{8}{} does not have any Euclidean bottleneck matching. In Section~\ref{max-matching-section} we consider the problem (b) and give some lower bounds on the size of a maximum matching in \kGG{k}{}. We prove that \kGG{0}{} has a matching of size at least $\frac{n-1}{4}$, and this bound is tight. In addition we prove that \kGG{1}{} has a matching of size at least $\frac{2(n-1)}{5}$ and \kGG{2}{} has a perfect matching. In Section~\ref{blocking-section} we consider the problem (c). We show that at least $\lceil\frac{n-1}{3}\rceil$ points are necessary to block a Gabriel graph and this bound is tight. We also show that at least $\lceil\frac{(k+1)(n-1)}{3}\rceil$ points are necessary and $(k+1)(n-1)$ points are sufficient to block a \kGG{k}{}. The open problems and concluding remarks are presented in Section~\ref{conclusion}.
\section{Preliminaries}
\label{preliminaries}
Let $G$ be an edge-weighted graph with vertex set $V$ and weight function $w:E\rightarrow\mathbb{R^+}$. Let $T$ be a minimum spanning tree of $G$, and let $w(T)$ be the total weight of $T$.
\begin{lemma}
\label{not-mst-edge}
Let $\delta(e)$ be a cycle in $G$ which contains an edge $e\in T$. Let $\delta'$ be the set of edges in $\delta(e)$ which do not belong to $T$ and let $e'_{max}$ be the largest edge in $\delta'$. Then, $w(e)\le w(e'_{max})$.
\end{lemma}
\begin{proof}
Let $e=(u,v)$ and let $T_u$ and $T_v$ be the two trees obtained by removing $e$ from $T$. Let $e'=(x,y)$ be an edge in $\delta'$ such that one of $x$ and $y$ belongs to $T_u$ and the other one belongs to $T_v$. By definition of $e'_{max}$, we have $w(e')\le w(e'_{max})$. Let $T'=T_u\cup T_v \cup\{(x,y)\}$. Clearly, $T'$ is a spanning tree of $G$. If $w(e')<w(e)$ then $w(T')<w(T)$; contradicting the minimality of $T$. Thus, $w(e)\le w(e')$, which completes the proof of the lemma.
\end{proof}
For a graph $G=(V,E)$ and $S\subseteq V$, let $G-S$ be the subgraph obtained from $G$ by removing all vertices in $S$, and let $o(G-S)$ be the number of odd components in $G-S$, i.e., connected components with an odd number of vertices. The following theorem by Tutte~\cite{Tutte1947} gives a characterization of the graphs which have perfect matching:
\begin{theorem}[Tutte~\cite{Tutte1947}]
\label{Tutte}
$G$ has a perfect matching if and only if $o(G-S)\le |S|$ for all $S\subseteq V$.
\end{theorem}
Berge~\cite{Berge1958} extended Tutte’s theorem to a formula (known as the Tutte-Berge formula) for the maximum size of a matching in a graph. In a graph $G$, the {\em deficiency}, $\text{def}_G(S)$, is $o(G-S)-|S|$. Let $\text{def}(G)=\max_{S\subseteq V}{\text{def}_G(S)}$.
\begin{theorem}[Tutte-Berge formula; Berge~\cite{Berge1958}]
\label{Berge}
The size of a maximum matching in $G$ is $$\frac{1}{2}(n-\mathrm{def}(G)).$$
\end{theorem}
For an edge-weighted graph $G$ we define the {\em weight sequence} of $G$, \WS{G}, as the sequence containing the weights of the edges of $G$ in non-increasing order. A graph $G_1$ is said to be less than a graph $G_2$ if \WS{G_1} is lexicographically smaller than \WS{G_2}.
\section{Euclidean Bottleneck Matching}
\label{bottleneck-section}
Given a point set $P$, in this section we prove that \kGG{10}{} contains a Euclidean bottleneck matching of $P$. We also present a configuration of a point set $P$ such that \kGG{8}{} does not contain any Euclidean bottleneck matching of $P$. We use a similar argument as in \cite{Abellanas2009, Chang1991}. First consider the following lemma of \cite{Abellanas2009}:
\begin{lemma}[Abellanas et al.~\cite{Abellanas2009}]
\label{cone-lemma}
Let $0 < \theta \le \pi/5$. Let $C(A, \theta, L, R)$ be a cone with apex $A$, bounding rays $L$
and $R$ emanating from $A$ and angle $\theta$ computed clockwise from $L$ to $R$. Given two points $x, y \in C(A, \theta, L, R)$ and a constant $r > 0$. If $|xA| > 2r$ and $|yA|>2r$, then $|xy| < 2r$ or $|xy| < \max\{|xA|− r, |yA|− r\}$.
\end{lemma}
\begin{figure}[htb]
\centering
\includegraphics[width=.6\columnwidth]{fig/cone0.pdf}
\caption{Illustration for Theorem~\ref{10-GG-thr}.}
\label{10-GG-fig}
\end{figure}
\begin{theorem}
\label{10-GG-thr}
For every point set $P$, \kGG{10}{} contains a Euclidean bottleneck matching of $P$.
\end{theorem}
\begin{proof}
Let $\mathcal{M}$ be the set of all perfect matchings through the points of $P$. Define a total order on the elements of $\mathcal{M}$ by their weight sequence. If two elements have exactly the same weight sequence, break ties arbitrarily to get a total order.
Let $M^* = \{(a_1, b_1),\dots, (a_{\frac{n}{2}}, b_{\frac{n}{2}})\}$ be a perfect matching in $\mathcal{M}$ with minimal weight sequence. It is obvious that $M^*$ is a Euclidean bottleneck matching for $P$. We will show that all edges of $M^*$ are in \kGG{10}{}. Consider any edge $e = (a_i, b_i)$ in $M^*$ and its corresponding disk $\CD{a_i}{b_i}$. Suppose that $\CD{a_i}{b_i}$ contains $w$ points of $P\setminus\{a_i,b_i\}$. Let $U = \{u_1, u_2,\dots, u_w\}$ represent the points inside $\CD{a_i}{b_i}$, and $U'=\{r_1, r_2,\dots, r_w\}$ represent the points where $(r_i,u_i)\in M^*$. We will show that $w\le 10$. Let $r=|a_ib_i|/2$ be the radius of $\CD{a_i}{b_i}$.
{\em Claim 1}: For each $r_j\in U'$, $\min\{|r_ja_i|, |r_jb_i|\} \ge 2r$. To prove this, assume that $|r_ja_i|< 2r$ and let $M$ be the perfect matching obtained from $M^*$ by deleting $\{(a_i,b_i), (r_j,u_j)\}$, and adding $\{(a_i, r_j), (b_i, u_j)\}$. The two new edges are smaller than the old ones. Thus, $\WS{M}<\WS{M^*}$ which contradicts the minimality of $M^*$.
Let $D_1$ and $D_2$ respectively be the open disks with radius $2r$ centered at $a_i$ and $b_i$. By Claim 1, we may assume that no point of $U'$ lies inside $D_1\cup D_2$. In other words all points of $U'$ are contained in $\overline{D_1\cup D_2}$.
{\em Claim 2}: For each pair $r_j$ and $r_k$ of points in $U'$, $|r_jr_k|\ge \allowbreak\max\allowbreak\{|a_ib_i|, \allowbreak |r_ju_j|, \allowbreak |r_ku_k|\}$. To prove this, assume that $|r_jr_k|< \max\{|a_ib_i|, |u_jr_j|,\allowbreak |u_kr_k|\}$. Let $M$ be the perfect matching obtained from $M^*$ by deleting $\{(u_j,r_j),(u_k,r_k),(a_i, b_i)\}$ and adding $\{(a_i, u_j), (b_i, u_k),(r_j, r_k)\}$. Since $\max \{|a_iu_j|,\allowbreak |b_iu_k|,\allowbreak |r_jr_k|\}<\max\{|u_jr_j|,\allowbreak |u_kr_k|,\allowbreak |a_ib_i|\}$, $\WS{M} \allowbreak<\allowbreak \WS{M^*}$ which contradicts the minimality of $M^*$.
Let $c_i$ be the center of $\CD{a_i}{b_i}$. Consider a decomposition of the plane into 10 cones $C_1, \dots, C_{10}$ of angle $\pi/5$ with apex at $c_i$. See Figure~\ref{10-GG-fig}. By contradiction, we will show that each cone $C_i, 1\le i\le 10$, contains at most one point of $U'$. Suppose that a cone $C_i$ where $1\le i\le 10$ contains two points $r_j, r_k\in U'$. It is obvious that
\begin{equation}\label{equ1}
|r_ju_j|\ge |c_ir_j|-r \quad\quad\text{and} \quad\quad |r_ku_k|\ge |c_ir_k|-r.
\end{equation}
{\em Claim 3}: Each cone $C_i$ where $1\le i\le 10$ and $i\neq 3,8$ contains at most one point of $U'$. Suppose that $C_i$ contain two points $r_j,r_k \in U'$. By Claim 1, all points of $U'$ are contained in $\overline{D_1\cup D_2}$. Consider the disk $D_3$ with radius $2r$ centred at $c_i$, as shown in Figure~\ref{10-GG-fig}. Since $D_3\cap \allowbreak(\overline{D_1\cup D_2})=\allowbreak\emptyset$, $r_j$ and $r_k$ are outside $D_3$, i.e., $|r_jc_i| > 2r$ and $|r_kc_i|>2r$. By Lemma~\ref{cone-lemma}, $|r_jr_k| < 2r$ or $|r_jr_k| \allowbreak<\allowbreak \max\{|r_jc_i| − r, \allowbreak |r_kc_i|− r\}$. By inequality~(\ref{equ1}), $|r_jr_k| < \max\{|a_ib_i|,|r_ju_j|,|r_ku_k|\}$ which contradicts Claim 2.
\begin{figure}[htb]
\centering
\setlength{\tabcolsep}{0in}
$\begin{tabular}{cc}
\multicolumn{1}{m{.5\columnwidth}}{\centering\includegraphics[width=.27\columnwidth]{fig/cone1.pdf}}
&\multicolumn{1}{m{.5\columnwidth}}{\centering\includegraphics[width=.3\columnwidth]{fig/cone2.pdf}}\\
(a) & (b)
\end{tabular}$
\caption{(a) The angle $\angle bac$ is smaller than the angle $\angle abc$, and hence (b) $\angle r_ka'r_j <\angle a'r_kr_j$.}
\label{cone-fig}
\end{figure}
{\em Claim 4}: Each of $C_3$ and $C_8$ contains at most one point of $U'$. Let $\{S_1,S_2\}$ be the partition of $D_3\cap (\overline{D_1\cup D_2})$ which lies inside $C_3$ and $C_8$ as shown in Figure~\ref{10-GG-fig}. Because of symmetry, we only prove the claim for $C_3$. Suppose that $C_3$ contains two points $r_j, r_k \in U'$.
For the rest of the proof, refer to Figure~\ref{cone-fig}.
W.l.o.g. assume that $r_j$ is further from $c_i$ than $r_k$ and $r_k$ is to the left of $r_j$ (i.e., $r_k$ is to the left of the line through
$c_i$ and $r_j$ oriented from $c_i$ to $r_j$). If $r_k\notin S_1$ then $|r_kc_i|>2r$ and $|r_jc_i|>2r$. Then, by Lemma~\ref{cone-lemma} and Claim 2 we have a contradiction. Therefore, assume that $r_k\in S_1$. Let $L$ and $R$ be the two rays defining $C_3$. Let $a$ be the intersection of $R$ and $\CIRC{a_i}{b_i}$. Let $b$ be the intersection of the boundaries of $D_1$ and $D_3$ which is inside $C_3$. Define the point $c$ on $R$ such that $|bc|=2r$ and $c\neq c_i$. See Figure~\ref{cone-fig}(a). The triangle $\bigtriangleup cbc_i$ is isosceles, and hence $\angle bcc_i = \angle bc_ic <\frac{\pi}{5}$. This implies that $\angle cbc_i>\frac{3\pi}{5}$. On the other hand, in triangle $\bigtriangleup abc_i$, $|ab|>|ac_i|$, which implies that $\angle abc_i <\angle ac_ib<\frac{\pi}{5}$. Thus $\angle abc >\frac{2\pi}{5}$. In addition $\angle bac_i>\frac{3\pi}{5}$ and hence $\angle bac < \frac{2\pi}{5}$. Therefore in the triangle $\bigtriangleup abc$ we have $$\angle abc>\frac{2\pi}{5}>\angle bac.$$ Let $C(b,c)$ be the circle with radius $2r$ having $\overline{bc}$ as diameter, and let $A$ be the ray emanating from $b$ which goes through $c$ as shown in Figure~\ref{cone-fig}(b). The intersection of $C_3$ with $\overline{D_1\cup D_2}$ which lies to the right of $A$ is completely inside $C(b,c)$. Thus, if $r_j$ is to the right of $A$, $|r_jr_k|<\allowbreak 2r\allowbreak =\allowbreak |a_ib_i|$, which contradicts Claim 2. Therefore $r_j$ lies to the left of $A$. If $r_j$ is in the interior of $C_3$, rotate $C_3$ counter-clockwise around $c_i$ until $r_j$ lies on $R$. Since $r_k$ is to the left of $r_j$, the point $r_k$ is still in the interior of $C_3$. Let $a'$ be the intersection of the new $R$ with $\CIRC{a_i}{b_i}$. Note that $S_1$ and hence $r_k$ is contained in $\bigtriangleup abc$. In addition $r_j$ and $a'$ are outside $\bigtriangleup abc$ and to the left of the line through $a$ and $c$. Therefore, $\angle a'r_kr_j\ge \angle abc >\angle bac \ge r_ka'r_j$ and hence $$|r_jr_k|<|r_ja'|=|r_jc_i|-r\le |r_ju_j|,$$ which contradicts Claim 2.
By Claim 3 and Claim 4 each cone $C_i$ where $1\le i \le 10$ contains at most one point of $U'$. Thus, $w\le 10$, and $e=(a_i,b_i)$ is an edge of \kGG{10}{}.
\end{proof}
\begin{figure}[htb]
\centering
\includegraphics[width=.8\columnwidth]{fig/9-GG.pdf}
\caption{A set of 20 points such that \kGG{8}{} does not contain any Euclidean bottleneck matching.}
\label{8-GG-fig}
\end{figure}
Now, we will show that for some point sets, \kGG{8}{} does not contain any Euclidean bottleneck matching.
Consider Figure~\ref{8-GG-fig} which shows a configuration of a set $P$ of 20 points. The closed disk $\CD{a}{b}$ is centred at $c$ and has diameter one, i.e., $|ab|=1$. $\CD{a}{b}$ contains 9 points $U=\{u_1, \dots, u_9\}$ which lie on a circle with radius $\frac{1}{2}-\epsilon$ which is centred at $c$. Nine points in $U'=\{r_1,\dots,r_9\}$ are placed on a circle with radius 1.5 which is centred at $c$ in such a way that $|r_ju_j|= 1+\epsilon$, $|r_ja|>1+\epsilon$, $|r_jb|>1+\epsilon$, and $|r_jr_k|>1+\epsilon$ for $1\le j, k\le 9$ and $j\neq k$. Consider a perfect matching $M=\{(a,b)\}\cup \{(r_i, u_i): i=1,\dots, 9\}$ where each point $r_i\in U'$ is matched to its closest point $u_i$. It is obvious that $\lambda(M)=1+\epsilon$, and hence the bottleneck of any bottleneck perfect matching is at most $1+\epsilon$. We will show that any Euclidean bottleneck matching of $P$ contains $(a,b)$. By contradiction, let $M^*$ be a Euclidean bottleneck matching which does not contain $(a,b)$. In $M^*$, $a$ is matched to a point $x\in U\cup U'$. If $x \in U'$, then $|ax|>1+\epsilon$. If $x\in U$, w.l.o.g. assume that $x = u_1$. Thus, in $M^*$ the point $r_1$ is matched to a point $y$ where $y\neq u_1$. Since $u_1$ is the closest point to $r_1$ and $|r_1u_1|=1+\epsilon$, $|r_1y|>1+\epsilon$. In both cases $\lambda(M^*)> 1+\epsilon$, which is a contradiction. Therefore, $M^*$ contains $(a,b)$. Since $\CD{a}{b}$ contains 9 points of $P\setminus\{a,b\}$, $(a,b)\notin\text{\kGG{8}{}}$. Therefore \kGG{8}{} does not contain any Euclidean bottleneck matching of $P$.
\section{Maximum Matching}
\label{max-matching-section}
Let $P$ be a set of $n$ points in the plane. In this section we will prove that \kGG{0}{} has a matching of size at least $\frac{n-1}{4}$; this bound is tight. We also prove that \kGG{1}{} has a matching of size at least $\frac{2(n-1)}{5}$ and \kGG{2}{} has a perfect matching (when $n$ is even).
First we give a lower bound on the number of components that result after removing a set $S$ of vertices from \kGG{k}{}. Then we use Theorem~\ref{Tutte} and Theorem~\ref{Berge}, respectively presented by Tutte~\cite{Tutte1947} and Berge~\cite{Berge1958}, to prove a lower bound on the size of a maximum matching in \kGG{k}{}.
\begin{figure}[htb]
\centering
\setlength{\tabcolsep}{0in}
$\begin{tabular}{cc}
\multicolumn{1}{m{.5\columnwidth}}{\centering\includegraphics[width=.38\columnwidth]{fig/partition-graph.pdf}}
&\multicolumn{1}{m{.5\columnwidth}}{\centering\includegraphics[width=.4\columnwidth]{fig/partition-tree.pdf}}\\
(a) & (b)
\end{tabular}$
\caption{The point set $P$ of 16 points is partitioned into open/closed disks, open/closed squares, and crosses. (a) The graph $G(\mathcal{P})$, (b) The set $\mathcal{T}$ of straight-line edges corresponding to $MST(G(\mathcal{P}))$ is in bold, and the set $\mathcal{D}$ of their corresponding disks.}
\label{partition-fig}
\end{figure}
Let $\mathcal{P}=\{P_1, P_2,\dots\}$ be a partition of the points in $P$. For two sets $P_i$ and $P_j$ in $\mathcal{P}$ define the distance $d(P_i,P_j)$ as the smallest Euclidean distance between a point in $P_i$ and a point in $P_j$, i.e., $d(P_i,P_j)=\min\{|ab|:a\in P_i, b\in P_j\}$.
Let $G(\mathcal{P})$ be the complete edge-weighted graph with vertex set $\mathcal{P}$. For each edge $e=(P_i,P_j)$ in $G(\mathcal{P})$, let $w(e)=d(P_i,P_j)$. This edge $e$ is defined by two points $a$ and $b$, where $a\in P_i$ and $b\in P_j$. Therefore, an edge $e\in G(\mathcal{P})$ corresponds to a straight line edge $(a,b)$ in $P$; see Figure~\ref{partition-fig}(a). Let $MST(G(\mathcal{P}))$ be a minimum spanning tree of $G(\mathcal{P})$. It is obvious that each edge $e$ in $MST(G(\mathcal{P}))$ corresponds to a straight line edge $(a,b)$ in $P$. Let $\mathcal{T}$ be the set of all these straight line edges. Let $\mathcal{D}$ be the set of disks which have the edges of $\mathcal{T}$ as diameter, i.e., $\mathcal{D}=\{D[a,b]: (a,b)\in \mathcal{T}\}$. See Figure~\ref{partition-fig}(b).
\begin{observation}
\label{T-plane}
$\mathcal{T}$ is a subgraph of a minimum spanning tree of $P$, and hence $\mathcal{T}$ is plane.
\end{observation}
\begin{lemma}
\label{D-empty}
A disk $\CD{a}{b}\in\mathcal{D}$ does not contain any point of $P\setminus\{a,b\}$.
\end{lemma}
\begin{proof}
Let $e=(P_i,P_j)$ be the edge in $MST(G(\mathcal{P}))$ corresponding to $\CD{a}{b}$. Note that $w(e)=|ab|$. By contradiction, suppose that $\CD{a}{b}$ contains a point $c\in P\setminus\{a,b\}$. Three cases arise: (i) $c\in P_i$, (ii) $c\in P_j$, (iii) $c\in P_l$ where $l\neq i$ and $l\neq j$. In case (i) the edge $(c,b)$ between $c\in P_i$ and $b\in P_j$ is smaller than $(a,b)$; contradicting that $w(e)=|ab|$ in $G(\mathcal{P})$. In case (ii) the edge $(a,c)$ between $a\in P_i$ and $c\in P_j$ is smaller than $(a,b)$; contradicting that $w(e)=|ab|$ in $G(\mathcal{P})$. In case (iii) the edge $(a,c)$ (resp. $(c,b)$) between $P_i$ and $P_l$ (resp. $P_l$ and $P_j$) is smaller than $(a,b)$; contradicting that $e$ is an edge in $MST(G(\mathcal{P}))$.
\end{proof}
\begin{lemma}
\label{center-in-lemma}
For each pair $D_i$ and $D_j$ of disks in $\mathcal{D}$, $D_i$ (resp. $D_j$) does not contain the center of $D_j$ (resp $D_i$).
\end{lemma}
\begin{proof}
Let $(a_i,b_i)$ and $(a_j,b_j)$ respectively be the edges of $\mathcal{T}$ which correspond to $D_i$ and $D_j$. Let $C_i$ and $C_j$ be the circles representing the boundary of $D_i$ and $D_j$. W.l.o.g. assume that $C_j$ is the bigger circle, i.e., $|a_ib_i|<|a_jb_j|$. By contradiction, suppose that $C_j$ contains the center $c_i$ of $C_i$. Let $x$ and $y$ denote the intersections of $C_i$ and $C_j$. Let $x_i$ (resp. $x_j$) be the intersection of $C_i$ (resp. $C_j$) with the line through $y$ and $c_i$ (resp. $c_j$). Similarly, let $y_i$ (resp. $y_j$) be the intersection of $C_i$ (resp. $C_j$) with the line through $x$ and $c_i$ (resp. $c_j$).
\begin{figure}[htb]
\centering
\includegraphics[width=.6\columnwidth]{fig/center-in.pdf}
\caption{Illustration of Lemma~\ref{center-in-lemma}: $C_i$ and $C_j$ intersect, and $C_j$ contains the center of $C_i$.}
\label{center-in-fig}
\end{figure}
As illustrated in Figure~\ref{center-in-fig}, the arcs $\widehat{x_ix}$, $\widehat{y_iy}$, $\widehat{x_jx}$, and $\widehat{y_jy}$ are the potential positions for the points $a_i$, $b_i$, $a_j$, and $b_j$, respectively. First we will show that the line segment $x_ix_j$ passes through $x$ and $|a_ia_j|\leq|x_ix_j|$. The angles $\angle x_ixy$ and $\angle x_jx_y$ are right angles, thus the line segment $x_ix_j$ goes through $x$. Since $\widehat{x_ix}<\pi$ (resp. $\widehat{x_jx}<\pi$), for any point $a_i\in \widehat{x_ix}, |a_ix|\leq|x_ix|$ (resp. $a_j\in \widehat{x_jx}, |a_jx|\leq|x_jx|$). Therefore, $$|a_ia_j|\leq|a_ix|+|xa_j|\leq|x_ix|+|xx_j|=|x_ix_j|.$$
Consider triangle $\bigtriangleup x_ix_jy$ which is partitioned by segment $c_ix_j$ into $t_1=\bigtriangleup x_ix_jc_i$ and $t_2=\bigtriangleup c_ix_jy$. It is easy to see that $|x_ic_i|$ in $t_1$ is equal to $|c_iy|$ in $t_2$, and the segment $c_ix_j$ is shared by $t_1$ and $t_2$. Since $c_i$ is inside $C_j$ and $\widehat{yx_j}=\pi$, the angle $\angle yc_ix_j>\frac{\pi}{2}$. Thus, $\angle x_ic_ix_j$ in $t_1$ is smaller than $\frac{\pi}{2}$ (and hence smaller than $\angle yc_ix_j$ in $t_2$). That is, $|x_ix_j|$ in $t_1$ is smaller than $|x_jy|$ in $t_2$. Therefore,
$$|a_ia_j|\leq|x_ix_j|<|x_jy|=|a_jb_j|.$$
By symmetry $|b_ib_j|<|a_jb_j|$. Therefore $\max\{|a_ia_j|,|b_ib_j|\}<\max\{|a_ib_i|,|a_jb_j|\}$. In addition $\delta=(a_i,a_j,b_j,b_i,a_i)$ is a cycle and at least one of $(a_i,a_j)$ and $(b_i,b_j)$ does not belong to $\mathcal{T}$. This contradicts Lemma~\ref{not-mst-edge} (Note that by Observation~\ref{T-plane}, $\mathcal{T}$ is a subgraph of a minimum spanning tree of $P$).
\end{proof}
Now we show that four disks in $\mathcal{D}$ cannot intersect mutually. In other words, every point in the plane cannot lie in more than three disks in $\mathcal{D}$. In Section~\ref{proof-section} we prove the following theorem, and in Section~\ref{lower-bounds-section} we present the lower bounds on the size of a maximum matching in \kGG{k}{}.
\begin{theorem}
\label{four-circle-theorem}
For every four disks $D_1,D_2,D_3,D_4\in\mathcal{D}$, $D_1\cap D_2\cap D_3\cap D_4=\emptyset$.
\end{theorem}
\subsection{Proof of Theorem~\ref{four-circle-theorem}}
\label{proof-section}
Let $\mathcal{X}=D_1\cap D_2\cap D_3\cap D_4$ and let $x$ be a point in $\mathcal{X}$. Let $(a_i,b_i)$ be the edge in $\mathcal{T}$ which corresponds to $D_i$, let $c_i$ be the center of $D_i$, and let $C_i$ denote the boundary of $D_i$, where $1\le i\le 4$. Denote the angle $\angle a_ixb_i$ by $\alpha_i$, where $1\le i\le 4$. Since $(a_i,b_i)$ is a diameter of $D_i$ and $x$ lies in $D_i$, $\alpha_i \ge\frac{\pi}{2}$. First we prove the following observation.
\begin{observation}
\label{inclusion-exclusion}
For $1\le i,j\le 4$ where $i\neq j$, the angles $\alpha_i$ and $\alpha_j$ are disjoint or one is completely contained in the other.
\end{observation}
\begin{proof}
The proof is by contradiction. Suppose that $\alpha_i$ and $\alpha_j$ share some part and w.l.o.g. assume that $b_i$ is in the cone which is defined by $\alpha_j$ and $b_j$ is in the cone which is defined by $\alpha_i$. Three cases arise:
\begin{itemize}
\item $b_i\in\bigtriangleup xa_jb_j$. In this case $b_i$ is inside $D_j$ which contradicts Lemma~\ref{D-empty}.
\item $b_j\in\bigtriangleup xa_ib_i$. In this case $b_j$ is inside $D_i$ which contradicts Lemma~\ref{D-empty}.
\item $b_i\notin\bigtriangleup xa_jb_j$ and $b_j\notin\bigtriangleup xa_ib_i$. In this case $(a_i,b_i)$ intersects $(a_j,b_j)$ which contradicts Observation~\ref{T-plane}.
\end{itemize}
\end{proof}
We call $\alpha_i$ a {\em blocked angle} if $\alpha_i$ is contained in an angle $\alpha_j$ where $j\neq i$, otherwise we call $\alpha_i$ a {\em free angle}.
\begin{lemma}
\label{not-all-free-angles}
At least one $\alpha_i$, where $1\le i \le 4$, is blocked.
\end{lemma}
\begin{proof}
Suppose that all angles $\alpha_i$, where $1\le i \le 4$, are free. This implies that the $\alpha_i$s are pairwise disjoint and $\alpha=\sum_{i=1}^{4}{\alpha_i} \ge 2\pi$. If $\alpha > 2\pi$, we obtain a contradiction to the fact that the sum of the disjoint angles around $x$ is at most $2\pi$. If $\alpha = 2\pi$, then the four edges $(a_i,b_i)$ where $1,\le i\le 4$, form a cycle which contradicts the fact that $\mathcal{T}$ is a subgraph of a minimum spanning tree of $P$.
\end{proof}
\begin{figure}[htb]
\centering
\setlength{\tabcolsep}{0in}
$\begin{tabular}{cc}
\multicolumn{1}{m{.7\columnwidth}}{\centering\includegraphics[width=.40\columnwidth]{fig/trap.pdf}}
&\multicolumn{1}{m{.3\columnwidth}}{\centering\includegraphics[width=.06\columnwidth]{fig/trap3.pdf}}\\
(a) & (b)
\end{tabular}$
\caption{(a) The point $x$ should be inside the arc $\widehat{c_ib_i}$. (b) The $trap(a_i,b_i)$ which consists of two almond-shaped regions known as $trap(a_i)$ and $trap(b_i)$.}
\label{trap-fig}
\end{figure}
By Lemma~\ref{not-all-free-angles} at least one of the angles is blocked. Hereafter, assume that $\alpha_j$ is blocked by $\alpha_i$ where $1\le i,j\le 4$ and $i\neq j$. W.l.o.g. assume that $a_ib_i$ is a vertical line segment and the point $x$ (which belongs to $\mathcal{X}$) is to the left of $a_ib_i$. Thus, $a_jb_j$ and $c_j$ are to the right of $a_ib_i$. This implies that $a_ib_i\cap D_j\neq \emptyset$. See Figure~\ref{trap-fig}(a). By Lemma~\ref{center-in-lemma}, $c_i$ cannot be inside $D_j$, thus either $a_ic_i\cap D_j\neq \emptyset$ or $c_ib_i\cap D_j \neq\emptyset$, but not both. W.l.o.g. assume that $c_ib_i\cap D_j\neq \emptyset$. Let $C'$ be the circle with radius $|c_ib_i|$ which is centred at $b_i$. Let $d$ denote the intersection of $C'$ with $C_i$ which is to the right of $c_ib_i$. Consider the circle $C''$ with radius $|xb_i|$ centred at $d$. Let $\widehat{c_ib_i}$ be the closed arc of $C''$ to the left of $c_ib_i$ as shown in Figure~\ref{trap-fig}(a).
We show that $x$ cannot be outside $\widehat{c_ib_i}$. By contradiction suppose that $x$ is outside $\widehat{c_ib_i}$ (and to the left of $c_ib_i$). Let $l_1$ and $l_2$ respectively be the perpendicular bisectors of $xb_i$ and $xc_i$. Let $b'_i$ and $c'_i$ respectively be the intersection of $l_1$ and $l_2$ with $c_ib_i$ and let $d'$ be the intersection point of $l_1$ and $l_2$. Since $x$ is outside $\widehat{c_ib_i}$, the intersection point $d'$ is to the left of (the vertical line through) $d$ and inside triangle $\bigtriangleup b_ic_id$. If $c_j$ is below $l_1$ then $|c_jb_i|<|c_jx|$ and $D_j$ contains $b_i$ which contradicts Lemma~\ref{center-in-lemma}. If $c_j$ is above $l_2$ then $|c_jb_i|<|c_jx|$ and $D_j$ contains $c_i$ which contradicts Lemma~\ref{center-in-lemma}. Thus, $c_j$ is above $l_1$ and below $l_2$, and (by the initial assumption) to the right of $c_ib_i$. That is, $c_j$ is in triangle $\bigtriangleup b'_ic'_id'$. Since $\bigtriangleup b'_ic'_id'\subseteq \bigtriangleup b_ic_id\subseteq D_i$, $c_j$ lies inside $D_i$ which contradicts Lemma~\ref{center-in-lemma}. Therefore, $x$ is contained in $\widehat{c_ib_i}$.
By symmetry $D_j$ can intersect $a_ic_i$ and/or $c_j$ can be to the left of $a_ib_i$ as well. Therefore, if $\alpha_i$ blocks $\alpha_j$, the point $x$ can be in $\widehat{c_ib_i}$ or any of the symmetric arcs. For an edge $a_ib_i$ we denote the union of these arcs by $trap(a_i,b_i)$ which is shown in Figure~\ref{trap-fig}(b). For each disk $D_i$, let $trap(D_i)= trap(a_i,b_i)$ where $(a_i, b_i)$ is the edge in $\mathcal{T}$ corresponding to $D_i$. Therefore $x$ is contained in $trap(D_i)$ which implies that $$\mathcal{X}\subseteq trap(D_i).$$ Note that $trap(D_i)$ consists of two almond-shaped symmetric regions; for simplicity we call them $trap(a_i)$ and $trap(b_i)$, i.e., $trap(D_i)=trap(a_i)\cup trap(b_i)$.
\begin{lemma}
\label{angle-in-trap}
For any point $x\in trap(a_i,b_i)$, $\angle a_ixb_i \ge 150^\circ$.
\end{lemma}
\begin{proof}
See Figure~\ref{trap-fig}(a). The angle $\angle b_idc_i = 60^\circ$, which implies that $\widehat{c_ib_i}= 60^\circ$. Thus, for any point $x'$ on the arc $\widehat{c_ib_i}$, $\angle x'c_ib_i + \angle x'b_ic_i = 30^\circ$, and hence for any point $x$ in $\widehat{c_ib_i}$, $\angle xc_ib_i + \angle xb_ic_i \le 30^\circ$. This implies that in $\bigtriangleup xb_ic_i$, $\angle b_ixc_i \ge 150^\circ$. On the other hand $\angle b_ixc_i \le \angle b_ixa_i$, which proves the lemma.
\end{proof}
\begin{figure}[htb]
\centering
\setlength{\tabcolsep}{0in}
$\begin{tabular}{ccc}
\multicolumn{1}{m{.33\columnwidth}}{\centering\includegraphics[width=.15\columnwidth]{fig/trap-intersection1.pdf}}
&\multicolumn{1}{m{.33\columnwidth}}{\centering\includegraphics[width=.15\columnwidth]{fig/trap-intersection2.pdf}}
&\multicolumn{1}{m{.33\columnwidth}}{\centering\includegraphics[width=.14\columnwidth]{fig/trap-intersection3.pdf}}\\
(a) & (b)& (c)
\end{tabular}$
\caption{Illustration of Lemma~\ref{intersecting-trap}.}
\label{trap-intersection-fig}
\end{figure}
\begin{lemma}
\label{intersecting-trap}
For any two disks $D_i$ and $D_j$ in $\mathcal{D}$, $trap(D_i)\cap trap(D_j)=\emptyset$.
\end{lemma}
\begin{proof}
We prove this lemma by contradiction. Suppose $x\in trap(D_i)\cap trap(D_j)$ and w.l.o.g. assume that $x\in trap(a_i)\cap trap(a_j)$ as shown in Figure~\ref{trap-intersection-fig}. Connect $x$ to $a_i$, $c_i$, $a_j$, and $c_j$ ($a_i$ may be identified with $a_j$). As shown in the proof of Lemma~\ref{angle-in-trap}, $\min\{\angle a_ixc_i, \angle a_jxc_j\} > 150^\circ$. Two configurations may arise:
\begin{itemize}
\item $\angle c_ixc_j \le 60^{\circ}$. In this case $|c_ic_j|\le \max\{|xc_i|,|xc_j|\}$. W.l.o.g. assume that $|xc_i|\le|xc_j|$ which implies that $|c_ic_j|\le |xc_j|$; see Figure ~\ref{trap-intersection-fig}(a). Clearly $|xc_j|<|c_ja_j|$, and hence $|c_ic_j|<|c_ja_j|$. Thus, $D_j$ contains $c_i$ which contradicts Lemma~\ref{center-in-lemma}.
\item $\angle c_ixc_j > 60^{\circ}$. In this case $\angle a_ixc_j \le 60^\circ$ and $\angle a_jxc_i \le 60^\circ$, hence $|a_ic_j|\le \max\{|a_ix|,|c_jx|\}$ and $|a_jc_i|\le \max\{|a_jx|,|c_ix|\}$. Three configurations arise:
\begin{itemize}
\item $|a_ix|<|c_jx|$, in this case $|a_ic_j|< |c_jx|<|c_ja_j|$ and hence $D_j$ contains $a_i$. See Figure~\ref{trap-intersection-fig}(b).
\item $|a_jx|<|c_ix|$, in this case $|a_jc_i|< |c_ix|< |c_ia_i|$ and hence $D_i$ contains $a_j$.
\item $|a_ix|\ge|c_jx|$ and $|a_jx|\ge|c_ix|$, in this case w.l.o.g. assume that $|a_ix|\le|a_jx|$. Thus $|a_ic_j|\le |a_ix|\le |a_jx|<|a_jc_j|$ which implies that $D_j$ contains $a_i$. See Figure~\ref{trap-intersection-fig}(b).
\end{itemize}
All cases contradict Lemma~\ref{D-empty}.
\end{itemize}
\end{proof}
Recall that each blocking angle is representing a trap. Thus, by Lemma~\ref{not-all-free-angles} and Lemma~\ref{intersecting-trap}, we have the following corollary:
\begin{corollary}
\label{one-blocked-angle}
Exactly one $\alpha_i$, where $1\le i\le 4$, is blocked.
\end{corollary}
Recall that $\alpha_j$ is blocked by $\alpha_i$, $a_ib_i$ is vertical line segment, $c_j$ is to the right of $a_ib_i$, and $x\in\widehat{c_ib_i}$. As a direct consequence of Corollary~\ref{one-blocked-angle}, $\alpha_i$, $\alpha_k$, and $\alpha_l$ are free angles, where $1\le\allowbreak i,j,\allowbreak k,\allowbreak l\le\allowbreak 4$ and $i\neq j\neq k\neq l$. In addition, $c_k$ and $c_l$ are to the left of $a_ib_i$. It is obvious that $$\mathcal{X}\subseteq trap(D_i)\cap D_k \cap D_l.$$
\begin{figure}[htb]
\centering
\includegraphics[width=.5\columnwidth]{fig/trap2.pdf}
\caption{Illustration of Lemma~\ref{intersecting-trap2}.}
\label{trap2-fig}
\end{figure}
\begin{lemma}
\label{intersecting-trap2}
For a blocking angle $\alpha_i$ and free angles $\alpha_k$ and $\alpha_l$, $trap(D_i)\allowbreak\cap D_k \allowbreak\cap D_l=\emptyset$.
\end{lemma}
\begin{proof}
Since $\alpha_i$ is a blocking angle and $\alpha_k$, $\alpha_l$ are free angles, $c_k$ and $c_l$ are on the same side of $a_ib_i$.
By contradiction, suppose that $x\in trap(D_i)\cap D_j \cap D_k$. See Figure~\ref{trap2-fig}. It is obvious that $\max\{|xa_i|, |xb_i|\}< |a_ib_i|$, $\max\{|xa_k|, |xb_k|\}< |a_kb_k|$, and $\max\{|xa_l|, |xb_l|\}<|a_lb_l|$. By Lemma~\ref{angle-in-trap}, $\alpha_i \ge 150^\circ$. In addition $\alpha_k, \alpha_l \allowbreak\ge\allowbreak 90^\circ$. Thus, $\max\{\angle a_ixb_k, \angle a_kxb_l, \angle a_lxb_i\}\allowbreak\le\allowbreak 30^\circ$. Hence, $|a_ib_k|\allowbreak<\allowbreak\max\{|xa_i|,|xb_k|\}$, $|a_kb_l|<\max\{|xa_k|,|xb_l|\}$, and $|a_lb_i|<\max\{|xa_l|,\allowbreak|xb_i|\}$. Therefore, $\max\{|a_ib_k|,|a_kb_l|, |a_lb_i|\}\allowbreak<\max\allowbreak\{|a_ib_i|,\allowbreak|a_kb_k|,|a_lb_l|\}$. In addition $\delta=(a_i,\allowbreak b_i,a_l,\allowbreak b_l,a_k,\allowbreak b_k,a_i)$ is a cycle and at least one of $(a_i,b_k)$, $(a_k,b_l)$ and $(a_l,b_i)$ does not belong to $\mathcal{T}$. This contradicts Lemma~\ref{not-mst-edge}.
\end{proof}
Thus, $\mathcal{X}=\emptyset$; which complete the proof of Theorem~\ref{four-circle-theorem}.
\subsection{Lower Bounds}
\label{lower-bounds-section}
In this section we present some lower bounds on the size of a maximum matching in \kGG{2}{}, \kGG{1}{}, and \kGG{0}{}.
\begin{theorem}
\label{matching-2GG}
For a set $P$ of an even number of points, \kGG{2}{} has a perfect matching.
\end{theorem}
\begin{proof}
First we show that by removing a set $S$ of $s$ points from \kGG{2}{}, at most $s+1$ components are generated. Then we show that at least one of these components must be even. Using Theorem~\ref{Tutte}, we conclude that \kGG{2}{} has a perfect matching.
Let $S$ be a set of $s$ vertices removed from \kGG{2}{}, and let $\mathcal{C}=\{C_1, \dots, C_{m(s)}\}$ be the resulting $m(s)$ components, where $m$ is a function depending on $s$. Actually $\mathcal{C}=\text{\kGG{2}{}}-S$ and $\mathcal{P}=\{V(C_1),\dots, V(C_{m(s)})\}$ is a partition of the vertices in $P\setminus S$.
{\bf\em Claim 1.} $m(s)\le s+1$. Let $G(\mathcal{P})$ be the complete graph with vertex set $\mathcal{P}$ which is constructed as described above. Let $\mathcal{T}$ be the set of all edges in $P$ corresponding to the edges of $MST(G(\mathcal{P}))$ and let $\mathcal{D}$ be the set of disks corresponding to the edges of $\mathcal{T}$. It is obvious that $\mathcal{T}$ contains $m(s)-1$ edges and hence $|\mathcal{D}|=m(s)-1$. Let $F=\{(p,D):p\in S, D\in \mathcal{D}, p\in D\}$ be the set of all (point, disk) pairs where $p\in S$, $D\in \mathcal{D}$, and $p$ is inside $D$. By Theorem~\ref{four-circle-theorem} each point in $S$ can be inside at most three disks in $\mathcal{D}$. Thus, $|F|\le 3\cdot|S|$.
Now we show that each disk in $\mathcal{D}$ contains at least three points of $S$ in its interior.
Consider any disk $D\in \mathcal{D}$ and let $e=(a,b)$ be the edge of $\mathcal{T}$ corresponding to $D$. By Lemma~\ref{D-empty}, $D$ does not contain any point of $P\setminus S$. Therefore, $D$ contains at least three points of $S$, because otherwise $(a,b)$ is an edge in \kGG{2}{} which contradicts the fact that $a$ and $b$ belong to different components in $\mathcal{C}$. Thus, each disk in $\mathcal{D}$ has at least three points of $S$. That is, $3\cdot|\mathcal{D}|\le|F|$. Therefore, $3(m(s)-1)\le |F|\le 3s$, and hence $m(s)\le s+1$.
{\bf \em Claim 2}: $o(\mathcal{C})\le s$. By Claim 1, $|\mathcal{C}|=m(s)\le s+1$. If $|\mathcal{C}|\le s$, then $o(\mathcal{C})\le s$. Assume that $|\mathcal{C}|=s+1$. Since $P=S\cup \{\bigcup^{s+1}_{i=1}{V(C_i)}\}$, the total number of vertices of $P$ is equal to $n=s+\sum_{i=1}^{s+1}{|V(C_i)|}$. Consider two cases where (i) $s$ is odd, (ii) $s$ is even. In both cases if all the components in $\mathcal{C}$ are odd, then $n$ is odd; contradicting our assumption that $P$ has an even number of vertices. Thus, $\mathcal{C}$ contains at least one even component, which implies that $o(\mathcal{C})\le s$.
Finally, by Claim 2 and Theorem~\ref{Tutte}, we conclude that \kGG{2}{} has a perfect matching.
\end{proof}
\begin{theorem}
\label{matching-1GG}
For every set $P$ of $n$ points, \kGG{1}{} has a matching of size at least $\frac{2(n-1)}{5}$.
\end{theorem}
\begin{proof}
Let $S$ be a set of $s$ vertices removed from \kGG{1}{}, and let $\mathcal{C}=\{C_1, \dots, C_{m(s)}\}$ be the resulting $m(s)$ components. Actually $\mathcal{C}=\text{\kGG{1}{}}-S$ and $\mathcal{P}=\{V(C_1),\dots, V(C_{m(s)})\}$ is a partition of the vertices in $P\setminus S$. Note that $o(\mathcal{C})\le m(s)$.
Let $M^*$ be a maximum matching in \kGG{1}{}. By Theorem~\ref{Berge},
\begin{align}
\label{align0}
|M^*|&= \frac{1}{2}(n-\text{def}(\text{\kGG{1}{}})),
\end{align}
where
\begin{align}
\label{align1}
\text{def}(\text{\kGG{1}{}})&= \max\limits_{S\subseteq P}(o(\mathcal{C})-|S|)\nonumber\\
& \le \max\limits_{S\subseteq P}(|\mathcal{C}|-|S|)\nonumber\\
& = \max\limits_{0\le s\le n}(m(s)-s).
\end{align}
Define $G(\mathcal{P})$, $\mathcal{T}$, $\mathcal{D}$, and $F$ as in the proof of Theorem~\ref{matching-2GG}. By Theorem~\ref{four-circle-theorem}, $|F|\le 3\cdot|S|$.
By the same reasoning as in the proof of Theorem~\ref{matching-2GG}, each disk in $\mathcal{D}$ has at least two points of $S$ in its interior. Thus, $2\cdot|\mathcal{D}|\le|F|$. Therefore, $2(m(s)-1)\le |F| \le 3s$, and hence
\begin{equation}
\label{ineq1}
m(s)\le\frac{3s}{2}+1.
\end{equation}
In addition, $s+m(s)=|S|+|\mathcal{C}|\le |P|=n$, and hence
\begin{equation}
\label{ineq2}
m(s)\le n-s.
\end{equation}
By Inequalities~(\ref{ineq1}) and ~(\ref{ineq2}),
\begin{equation}
\label{ineq3}
m(s)\le \min\{\frac{3s}{2}+1, n-s\}.
\end{equation}
Thus, by (\ref{align1}) and (\ref{ineq3})
\begin{align}
\label{align2}
\text{def}(\text{\kGG{1}{}})&\le \max\limits_{0\le s\le n}(m(s)-s)\nonumber\\
&\le \max\limits_{0\le s\le n}\{\min\{\frac{3s}{2}+1, n-s\}-s\}\nonumber\\
&= \max\limits_{0\le s\le n}\{\min\{\frac{s}{2}+1, n-2s\}\}\nonumber\\
&= \frac{n+4}{5},
\end{align}
where the last equation is achieved by setting $\frac{s}{2}+1$ equal to $n-2s$, which implies $s=\frac{2(n-1)}{5}$. Finally by substituting (\ref{align2}) in Equation (\ref{align0}) we have
$$
|M^*|\ge \frac{2(n-1)}{5}.
$$
\end{proof}
By similar reasoning as in the proof of Theorem~\ref{matching-1GG} we have the following Theorem.
\begin{figure}[htb]
\centering
\includegraphics[width=.6\columnwidth]{fig/tight-0GG.pdf}
\caption{A \kGG{0}{} of $n = 17$ points with a maximum matching of size $\frac{n-1}{4}=4$ (bold edges). The dashed edges do not belong to the graph because any of their corresponding closed disks has a point on its boundary.}
\label{tight-0GG}
\end{figure}
\begin{theorem}
\label{matching-0GG}
For every set $P$ of $n$ points, \kGG{0}{} has a matching of size at least $\frac{n-1}{4}$.
\end{theorem}
The bound in Theorem~\ref{matching-0GG} is tight, as can be seen from the graph in Figure~\ref{tight-0GG}, for which the maximum matching has size $\frac{n-1}{4}$. Actually this is a Gabriel graph of maximum degree four which is a tree. The dashed edges do not belong to \kGG{0}{} because any closed disk which has one of these edges as diameter has a point on its boundary. Observe that each edge in any matching is adjacent to one of the vertices of degree four.
\begin{paragraph}{Note:}For a point set $P$, let $\nu_k(P)$ and $\alpha_k(P)$ respectively denote the size of a maximum matching and a maximum independent set in \kGG{k}{}. For every edge in the maximum matching, at most one of its endpoints can be in the maximum independent set. Thus,$$\alpha_k(P)\le |P| - \nu_k(P).$$
By combining this formula with the results of Theorems ~\ref{matching-0GG}, \ref{matching-1GG}, \ref{matching-2GG}, respectively, we have $\alpha_0(P)\le \frac{3n+1}{4}$, $\alpha_1(P)\le \frac{3n+2}{5}$, and $\alpha_2(P)\le \lceil\frac{n}{2}\rceil$. The \kGG{0}{} graph in Figure~\ref{tight-0GG} has an independent set of size $\frac{3n+1}{4}=13$, which shows that this bound is tight for \kGG{0}{}. On the other hand, \kGG{0}{} is planar and every planar graph is 4-colorable; which implies that $\alpha_0(P)\ge \lceil\frac{n}{4}\rceil$. There are some examples of \kGG{0}{} in \cite{Matula1980} such that $\alpha_0(P)= \lceil\frac{n}{4}\rceil$, which means that this bound is tight as well.
\end{paragraph}
\section{Blocking Higher-Order Gabriel Graphs}
\label{blocking-section}
In this section we consider the problem of blocking higher-order Gabriel graphs. Recall that a point set $K$ blocks \kGG{k}{(P)} if in \kGG{k}{(P\cup K)} there is no edge connecting two points in $P$.
\begin{theorem}
\label{blocking-thr1}
For every set $P$ of $n$ points, at least $\lceil\frac{n-1}{3}\rceil$ points are necessary to block \kGG{0}{(P)}.
\end{theorem}
\begin{proof}
Let $K$ be a set of $m$ points which blocks \kGG{0}{(P)}. Let $G(\mathcal{P})$ be the complete graph with vertex set $\mathcal{P}=P$. Let $\mathcal{T}$ be a minimum spanning tree of $G(\mathcal{P})$ and let $\mathcal{D}$ be the set of closed disks corresponding to the edges of $\mathcal{T}$. It is obvious that $|\mathcal{D}|=n-1$. By Lemma~\ref{D-empty} each disk $\CD{a}{b}\in\mathcal{D}$ does not contain any point of $P\setminus\{a,b\}$, thus, $\mathcal{T}\subseteq\text{\kGG{0}{(P)}}$. To block each edge of $\mathcal{T}$, corresponding to a disk in $\mathcal{D}$, at least one point is necessary. By Theorem~\ref{four-circle-theorem} each point in $K$ can lie in at most three disks of $\mathcal{D}$. Therefore, $m\ge\lceil\frac{n-1}{3}\rceil$, which implies that at least $\lceil\frac{n-1}{3}\rceil$ points are necessary to block all the edges of $\mathcal{T}$ and hence \kGG{0}{(P)}.
\end{proof}
\begin{figure}[htb]
\centering
\setlength{\tabcolsep}{0in}
$\begin{tabular}{cc}
\multicolumn{1}{m{.75\columnwidth}}{\centering\includegraphics[width=.65\columnwidth]{fig/tight-blocking.pdf}}
&\multicolumn{1}{m{.25\columnwidth}}{\centering\includegraphics[width=.23\columnwidth]{fig/dashed-blocking.pdf}}\\
(a)&(b)
\end{tabular}$
\caption{(a) \kGG{0}{} graph of $n=13$ points (in bold edges) which is blocked by $\lceil\frac{n-1}{3}\rceil=4$ white points, (b) dashed edges do not belomg to \kGG{0}{}.}
\label{blocking-fig}
\end{figure}
Figure~\ref{blocking-fig}(a) shows a \kGG{0}{} with $n=13$ (black) points which is blocked by $\lceil\frac{n-1}{3}\rceil=4$ (white) points. Note that all the disks, corresponding to the edges of every cycle, intersect at the same point in the plane (where we have placed the white points). As shown in Figure~\ref{blocking-fig}(b), the dashed edges do not belong to \kGG{0}{}. Thus, the lower bound provided by Theorem~\ref{blocking-thr1} is tight. It is easy to generalize the result of Theorem~\ref{blocking-thr1} to higher-order Gabriel graphs. Since in a \kGG{k}{} we need at least $k+1$ points to block an edge of $\mathcal{T}$ and each point can be inside at most three disks in $\mathcal{D}$, we have the following corollary:
\begin{corollary}
For every set $P$ of $n$ points, at least $\lceil\frac{(k+1)(n-1)}{3}\rceil$ points are necessary to block \kGG{k}{(P)}.
\end{corollary}
In \cite{Aronov2013} the authors showed that every Gabriel graph can be blocked by a set $K$ of $n-1$ points by putting a point slightly to the right of each point of $P$, except for the rightmost one. Every disk with diameter determined by two points of $P$ will contain a point of $K$. Using a similar argument one can block a \kGG{k}{} by putting $k+1$ points slightly to the right of each point of $P$, except for the rightmost one. Thus,
\begin{corollary}
For every set $P$ of $n$ points, there exists a set of $(k+1)(n-1)$ points that blocks \kGG{k}{(P)}.
\end{corollary}
Note that this upper bound is tight, because if the points of $P$ are on a line, the disks representing the minimum spanning tree are disjoint and each disk needs $k+1$ points to block the corresponding edge.
\section{Conclusion}
\label{conclusion}
In this paper, we considered the bottleneck and perfect matching admissibility of higher-order Gabriel graphs. We proved that
\begin{itemize}
\item \kGG{10}{} contains a Euclidean bottleneck matching of $P$ and \kGG{8}{} may not have any.
\item \kGG{0}{} has a matching of size at least $\frac{n-1}{4}$ and this bound is tight.
\item \kGG{1}{} has a matching of size at least $\frac{2(n-1)}{5}$.
\item \kGG{2}{} has a perfect matching.
\item $\lceil\frac{n-1}{3}\rceil$ points are necessary to block \kGG{0}{} and this bound is tight.
\item $\lceil\frac{(k+1)(n-1)}{3}\rceil$ points are necessary and $(k+1)(n-1)$ points are sufficient to block \kGG{k}{}.
\end{itemize}
We leave a number of open problems:
\begin{itemize}
\item Does \kGG{9}{} contain a Euclidean bottleneck matching of $P$?
\item What is a tight lower bound on the size of a maximum matching in \kGG{1}{}?
\end{itemize}
\bibliographystyle{abbrv}
|
\section{Motivation}
The diameter-constrained reliability measure was introduced in 2001 by H\'ector Cancela and Louis Petingi, inspired in delay sensitive applications~\cite{PR01}.
In telecommunications, there are several problems where the diameter (or the number of hops) in the communication is a major cause of concern.
In flooding-based systems, the number of hops should be controlled in order to avoid network congestion. Peer-to-peer networks
originally support file discovery protocols by means of flooding~\cite{Adar00freeriding}.
Internet Protocol version 6 (IPv6) has a ``Hop limit'' field, reserved for these cases~\cite{ietf:ipv6}.
Another hot topic in network design is fiber optics deployment. There, light-paths should be short in order to save bandwidth resources~\cite{Liu201344}.
The performance of degraded systems is dramatically deteriorated with distance~\cite{Cover:2006:EIT:1146355}.
A practical example is electrical networks, which suffer from Joule effect, causing power losses.
We invite the reader to find a rich discussion on diameter-constrained reliability and its applications in~\cite{CEKP2012}.\\
This paper is organized in the following manner. Section~\ref{dcr} formally presents the problem under study, and its computational complexity.
Section~\ref{exact} shows three exact methods to find network reliability, focused on factorization method.
The main contributions are included in Section~\ref{factor}. There, the determination of irrelevant links is fully characterized.
Additionally, we present invariants for the diameter-constrained reliability in a source-terminal scenario,
which should be considered to find the DCR exactly. Section~\ref{conclusions} presents concluding remarks and trends for future work.
\section{Diameter-Constrained Reliability}\label{dcr}
We will follow the terminology of Michael Ball~\cite{Ball1986}.
A \emph{stochastic binary system} (SBS) is a tern $(S,\phi,p)$, being $S=\{a_1,\ldots,a_m\}$ a ground-set with $m$ on-off elements, called \emph{components},
$\phi:\{0,1\}^m \to \{0,1\}$ a \emph{structure function} that assigns either up ($1$) or down ($0$) to each system state, and $p=(p_1,\ldots,p_m) \in [0,1]^m$
a vector that contains \emph{elementary probabilities of operation} for each component. Consider a random vector $X=(X_1,\ldots,X_m)$,
being $\{X_i\}_{i=1,\ldots,m}$ a set of independent Bernoulli random variables such that $P(X_i=1)=p_i$. The \emph{reliability of an SBS} is the
number $r$:
\begin{equation}
r = P(\phi(X)=1) = E(\phi(X))
\end{equation}
A pathset is a state $x \in \{0,1\}^m$ such that $\phi(x)=1$. A cutset is a state $x \in \{0,1\}^m$ such that $\phi(x)=0$.
Minimal pathsets (cutsets) are called \emph{minpaths} (\emph{mincuts}). Let us denote $\overline{x}_{i}$ to binary word $x$
with the complementary value for bit $i$. A component $i \in S$ is \emph{irrelevant} if $\phi(x)=\phi (\overline{x}_{i})$ for
all possible states $x \in \{0,1\}^m$. In words, a component is irrelevant when its elementary state does not affect the
global system state.\\
The classical network reliability problem considers a simple graph $G=(V,E)$ and a set of distinguished nodes $K\subseteq V$,
called \emph{terminal set}. The corresponding SBS is defined by the following tern:
\begin{itemize}
\item[1] The ground set is the ordered-set of links: $S=E=(e_1,\ldots,e_m)$.
\item[2] A corresponding probability vector: $p=(p_1,\ldots,p_m) \in [0,1]^m$.
\item[3] For each $E^{\prime} \subseteq E$, a binary word $w_{E^{\prime}} = (w_1,\ldots,w_m)$ such that $w_i=1$ if and only if
$e_i \in E^{\prime}$, and $\phi(w)=1$ if and only if all pair terminals in $G^{\prime}=(V,E^{\prime})$ are connected by some path.
\end{itemize}
The classical network reliability $r$ is historically termed connectedness probability as well~\cite{provan83}.
In the diameter-constrained scenario, the structure function is modified, and $\phi(w)=1$ additionally requires paths
with $d$ hops or less between each pair of terminals, being $d$ a positive integer called \emph{diameter}.
We will denote $R_{K,G}^{d}$ the diameter-constrained reliability.\\
The exact reliability computation is at least as hard as minimum cardinality cutset recognition~\cite{Ball1986}.
Arnon Rosenthal observed that minimum cardinality recognition in the classical reliability problem is precisely Steiner Tree Problem
in graphs~\cite{Rosenthal}.
Since this problem is included in Karp's list~\cite{Karp1972},
classical reliability computation belongs to the class of $\mathcal{NP}$-Hard problems.
The reader can observe that the exact diameter-constrained reliability computation is an extension of classical reliability.
Therefore, it also belongs to the class of $\mathcal{NP}$-Hard problems. H\'ector Cancela and Louis Petingi showed
the the problem is $\mathcal{NP}$-Hard even in the source terminal case $|K|=2$ when $d\geq 3$; see~\cite{CP2004} for a complete proof.
A full complexity analysis of different subproblems as a function of $k=|K|$ and $d$ is available in prior works~\cite{CanaleRomero,CP2004,CP2001}.
\section{Exact Methods}\label{exact}
In order to find the reliability of an SBS, Surech Rai et. al. suggest
the following classification of exact method~\cite{NET:NET3230250308}:
\begin{itemize}
\item Inclusion-Exclusion.
\item Sum of Disjoint Products.
\item Factorization.
\end{itemize}
The first one, also called Poincare's formula, is based on a full enumeration of minpaths (or mincuts).
Assume that $M_1, \ldots,M_l$ is the whole list of minpaths of a certain SBS $(S,\phi,p)$.
Inclusion-Exclusion method returns the reliability using the following expression:
\begin{equation}\label{ie}
r = P(\bigcup_{i=1}^{l}M_i) = \sum_{i=1}^{l-1}(-1)^{i-1}\sum_{I \subseteq \{1,2,\ldots,l\}; |I|=i} P(\bigcap_{j\in |I|}M_j)
\end{equation}
The number of minpaths can be exponential with the cardinal of the ground-set $m$. Moreover, the number of terms from Expression~\eqref{ie}
is even larger. Unless a cancellation of terms or special property is found, a full enumeration of minpaths or mincuts is avoided.\\
Observe that the events $\{M_i\}_{i=1,\ldots,l}$ are non-necessarily disjoint.
An alternative is to re-write Expression~\eqref{ie}, finding a mutually-exhaustive union of disjoint events.
Since components fail independently, events are then written as a product of the elementary reliabilities.
This is the key idea of Sum of Disjoint Products method.\\
Let us have a closer look to the third family of exact methods, called Factorization. The basic idea is to consider
conditional measure on the operation of some component $i$:
\begin{equation}
r = p_{i} r(\phi|i=1) + (1-p_i)r(\phi|i=0)(1-p_i),
\end{equation}
where $\phi|i=1$ (resp. $\phi|i=0$) is structure $\phi$ conditioned to the event ``component $i$ is in operation'' (resp. failure).
A shortcoming of this recursive method in its basic form is that it is strongly exponential.
If the system has irrelevant components they should be discarded, and the process can be largely accelerated. The determination
of irrelevant components depends on the specific structure under study.
The first work in Factorization in the field of network reliability is authored by Fred Moskowitz, inspired in electrical networks~\cite{6372698}:
\begin{equation}\label{fact1}
R_{K,G} = p_e R_{K^{\prime},G*e} + (1-p_e)R_{K,G-e},
\end{equation}
being $e \in E$ a certain link of graph $G=(V,E)$ with elementary probability $p_e$, $G-e = (V,E-e)$ the deletion
graph, $G*e$ the contraction graph (contraction of link $e$) and $K^{\prime}$ is the new terminal-set after link contraction.\\
Since contraction operation is not a diameter invariant, Expression~\eqref{fact1} (sometimes called deletion-contraction formula)
does not hold for the diameter-constrained reliability. However, a similar expression holds:
\begin{equation}\label{fact2}
R_{K,G}^{d} = p_e R_{K,G^{\prime}} + (1-p_e)R_{K,G-e},
\end{equation}
being now $G^{\prime}=G$ but with elementary reliability $p_e=1$. Expression~\eqref{fact2} suggests a recursive solution,
where links are either perfect or deleted in turns, until a halting condition is met
(either the network has perfect pathset, or there is no feasible pathset).\\
Recall that a recursive application of Expression~\eqref{fact2} is strongly exponential, and reductions/simplifications to successive
graphs should be performed. We term \emph{Factorization methods} including these aspects as well.
In classical network reliability, Factorization theory is mature~\cite{Satyanarayana,doi:10.1137/0214057}.
However, its extension to the diameter-constrained case deserves further research.
\section{Full Characterization of Irrelevant links in DCR} \label{factor}
H\'ector Cancela et. al. propose a sufficient condition for a link to be irrelevant.
They state the determination of irrelevant links in a source-terminal context is still an open problem~\cite{CEKP2012}.
Later effort has been carried-out by Louis Petingi, with a stronger sufficient condition~\cite{petingi2013diameter}.
A recent analysis shows a third sufficient condition, but it leaves the determination of irrelevant links as an open problem~\cite{CanaleRomeroRubino}.\\
Here, a full characterization of irrelevant links is introduced for a source-terminal scenario first, and later in a $K$-terminal context
(i.e., for an arbitrary terminal-set $K \subseteq V$).
First, we will show the three sufficient conditions for the source-terminal case,
available from the literature presented in a chronological order,
and the reasons that they fail to recognize irrelevant links in some graphs. Consider an arbitrary graph $G=(V,E)$,
a two-terminal set $K=\{s,t\}$, a diameter $d$ and a specific link under study $e=\{x,y\}$.
By an elementary analysis, the following conditions are sufficient for link $e$ to be irrelevant:
\begin{itemize}
\item[1)] $d_G(s,x)+d_G(y,t)\geq d$ and $d_{G}(s,y)+d_G(x,t)\geq d$;
\item[2)] $d_{G-e}(s,x)+d_{G-e}(y,t)\geq d$ and $d_{G-e}(s,y)+d_{G-e}(x,t)\geq d$;
\item[3)] $d_{G-y-t}(s,x) + d_{G-s-x}(y,t)\geq d$ and $d_{G-x-t}(s,y) + d_{G-s-y}(x,t)\geq d$.
\end{itemize}
Let us consider the graph $G$ sketched in Figure~\ref{fig:red}. Observe that link $e=\{1,2\}$ is irrelevant for diameter $d=5$, and even
for $d=6$ as well. However, $d_G(s,1)+d_G(2,t)=1+2=3<5$, so Condition 1 does not detect that $e$ is irrelevant for $d=5$ nor $d=6$.
Observe that $d_{G-e}(s,1)+d_{G-e}(2,t)=1+4=5$ and $d_{G-e}(s,2)+d_{G-e}(1,t)=4+1=5$, so Condition $2$ detects that $e$ is irrelevant
when $d=5$, but it is not the case for $d=6$. Finally, $d_{G-2-t}(s,1)+d_{G-1,s}(2,t)=1+5=6$ and
$d_{G-1-t}(2,s)+d_{G-s-2}(1,t)=\infty$, so Condition $3$ detects that $e$ is irrelevant in both cases.
Nevertheless, the reader can check that link $e^{\prime}=\{2,3\}$ is irrelevant when $d=6$, but no sufficient condition detects that
$e^{\prime}$ is irrelevant.
\begin{figure}[h!]\centering{
\begin{tikzpicture}
[scale=1.2,>=stealth,shorten >=0.1pt, auto, thick,
nodot/.style={circle, draw = black , very thin, minimum size=12pt, inner sep=0pt, font=\scriptsize },
]
\begin{scope}[xshift=0cm,scale=1]
\node [nodot] (s) at (0,0) {$s$};
\node [nodot] (1) at (1,0) {1};
\node [nodot] (2) at (1,1) {2};
\node [nodot] (3) at (2,2) {3};
\node [nodot] (4) at (3,2) {4};
\node [nodot] (5) at (4,2) {5};
\node [nodot] (6) at (5,1) {6};
\node [nodot] (t) at (5,0) {t};
\draw (s)--(1)--(2)--(3)--(4)--(5)--(6)--(t);
\draw (1)--(4);
\draw (1)--(t);
\end{scope}
\end{tikzpicture}} \caption{Sample graph $G$ with an irrelevant link $e=\{1,2\}$ when $d=6$.}
\label{fig:red}
\end{figure}
A basic result from its definition is that if a certain link $e$ is irrelevant for a diameter $d$,
then it will also be irrelevant for any diameter $d^{\prime}\leq d$.\\
Now, we will fully characterize irrelevant links. Let $G=(V,E)$ be a simple graph, $K=\{s,t\}$ the terminal set,
$d$ a positive integer (diameter) and $e=\{x,y\} \in E$ a certain link. Under these conditions, $e$ is relevant
if and only if there is some $s-t$ path $P$ composed by at most $d$ links, such that $e \in P$.\\
Equivalently, we will find two node-disjoint paths $P_1$ and $P_2$ from nodes $x,y$ to nodes $s,t$ with the minimum length-sum.
Then, $e$ is irrelevant if and only if path $P= P_1 \cup \{x,y\} \cup P_2$ has more than $d$ links.
In order to find the desired paths $P_1$ and $P_2$, let us extend the original network.
Consider two artificial nodes $u$ and $z$ with degree $2$.
Specifically, dode $u$ is connected with terminals $s$ and $t$, while node $z$ is connected with $x$ and $y$.
We should find two node-disjoint paths $P^{\prime}_1$
and $P^{\prime}_2$ between $u$ and $z$ with minimum length-sum. Suurballe's algorithm~\cite{NET:NET3230040204}
(or Bhandari's algorithm~\cite{DBLP:conf/iscc/Bhandari97}) provides precisely those paths.
After the deletion of artificial nodes, we obtain the desired paths $P_1$ and $P_2$. We proved the following
\begin{theorem}
Link $e$ is relevant if and only if $l(P_1 \cup P_2) \leq d-1$\\
(where the disjoint paths $P_1$ and $P_2$ are found using Suurballe's algorithm).
\end{theorem}
As a consequence, the deletion of irrelevant links is a DCR invariant. In the most general $K$-terminal context,
the determination of irrelevant links is performed analogously: just check all pair of terminals whether link $e$ is part of some path
between two terminals or not. The following elementary operations are DCR invariants as well,
for a source-terminal configuration:
\begin{itemize}
\item \emph{Pending-Node}: If the source $s$ (idem terminal $t$) is pending on a link $e=\{s,x\}$ with reliability $p_e$, then we contract link $e$, and
replace $G$ for its contraction $G*e$. The invariant is $R_{\{s,t\},G}^{d} = R_{\{s^{\prime},t\},G*e}^{d-1}$, being $e^{\prime}$ the
new source. All non-terminal pending nodes are deleted.
\item \emph{Perfect-Path}: If a path $P=\{v_1,\ldots,v_n\}$ is an induced subgraph for $G$ and links have elementary reliabilities
$p_{v_i,v_i+1}$, then we re-assign the link reliabilities $p_{v_i,v_{i+1}}=1$ for all $i=1,\ldots,n-2$ but
$p_{v_{n-1},v_n}= \prod_{i=1}^{n-1}p_{v_i,v_{i+1}}$.
\item \emph{Perfect-Neighbors}: if the source $s$ (idem terminal $t$) has all perfect links to its neighbors $N(s)$, then $s \cup N(s)$ is a
new vertex in $G^{\prime}$ and $R_{\{s,t\},G}^{d} = R_{\{s,t\},G^{\prime}}^{d-1}$.
\item \emph{Perfect-Cut-Node}: if $v$ is a cut-node (i.e., $G-v$ has more than one component), first delete
components with all non-terminal nodes (observe that we cannot finish with more than two components).
Second, apply Perfect-Neighbors to $v$ on both sides.
\item \emph{Parallel-Links}: If we find two links $e_1$ and $e_2$ from the same nodes with elementary reliabilities $p_{e_1}$ and $p_{e_2}$,
they are replaced by a new link $e$ with reliability $p_e = p_{e_1}+p_{e_2}-p_{e_1}p_{e_2}$.
\end{itemize}
These invariants are building blocks of a Factorization method, combined with the deletion of irrelevant links and
reduction of selected links.
\begin{algorithm}[H]
\caption{$R = Factor(G,s,t,d)$} \label{BLocal}
\begin{algorithmic}[1]
\IF{$HasPefectPath(G,d)$}
\RETURN $R=1$
\ENDIF
\IF{$Distance(s,t)<d$}
\RETURN $R=0$
\ENDIF
\STATE $G \leftarrow Delete(G,Suurballe)$
\STATE $(G,s,t,d) \leftarrow Invariants(G,s,t,d)$
\STATE $e \leftarrow NonPerfectRightMost(E(G))$
\RETURN $(1-p_e) \times Factor(G-e,s,t,d)+p_e \times Factor(G*e,s,t,d)$
\end{algorithmic}
\end{algorithm}
We put all together in $Factor$ Algorithm. It receives the graph $G$, two terminals $s,t$ and a diameter $d$,
and returns $R=R_{\{s,t\},G}^{d}$. The block of Lines $1$-$6$ test the termination (i.e., either a perfect pathset or no
feasible pathset). In Line $7$, Suurballe's algorithm is called in order to determine, for each link, whether it is relevant or not.
Observe that the order of this test does not matter, since the deletion of an irrelevant link does not remove any minpath.
In Line 8, a list of invariants help to further reduce and simplify the graph. So far, the list has five elementary operations,
as previously detailed. In Line 9, a certain link is selected in order to perform Factor decomposition. Here, we recommend
to choose the non-perfect link that is closest to the terminal $t$ (or one of them chosen uniformly at random in case of several
links). In this way, we improve the activity of Perfect-Neighbors operation (i.e., contracting all nodes close to neighbors from $t$,
and reducing the diameter in one unit).
\section{Conclusions and Trends for Future Work}\label{conclusions}
In this paper we discussed exact approaches for the exact diameter-constrained reliability (DCR),
focused in a source-terminal context. The exact DCR computation belongs to the class of $\mathcal{NP}$-Hard problems,
since it subsumes the classical network reliability problem.
Factorization techniques are available for the classical problem. However, the determination of irrelevant links has been
a shortcoming of previous works in the diameter-constrained version.\\
Here, an efficient method for the determination of irrelevant links is provided for the DCR. Additionally,
some DCR invariants are included, and a factorization algorithm has been introduced. It greedily selects the closest links
to one of the terminals, in order to increase the degree of perfect links from the terminals (or reduce the degree after link deletion).\\
Currently, we are implementing this algorithm and similar ones available from the literature in order to perform a faithful comparison
for both sparse and dense graphs. Additionally, the determination of new DCR invariants is both a challenging and useful task
in order to develop better exact factorization algorithms for the source-terminal DCR computation.
|
\section{Simplicity of restricted Lazarsfeld-Mukai bundles}
\label{sec:simple}
We fix a $K3$ surface $S$, a smooth curve $C\subset S$ of genus $g$ and a globally generated linear series $A\in W^r_d(C)$, with $h^0(C, A)=r+1$.
Using the evaluation sequence (\ref{eqn: F}), we form the vector bundle $F=F_{C, A}$; by dualizing, we obtain an exact sequence for the dual bundle $E=E_{C, A}:=F_{C, A}^{\vee}$:
\begin{equation}
\label{eqn: E}
0\longrightarrow H^0(C, A)^{\vee}\otimes \mathcal{O}_S\longrightarrow E_{C, A}\longrightarrow K_C\otimes A^{\vee}\longrightarrow 0.
\end{equation}
It is well-known \cite{M89}, \cite{L1} that $c_1(E)=[C]$ and $c_2(E)=d$; moreover
$h^0(S, F)=0$ and $h^1(S, E)=h^1(S, F)=0$. Finally, one also has that
$$\chi(S, E\otimes F)=2-2\rho(g, r, d);$$ in particular, if $E$ is a simple bundle, then $\rho(g, r, d)\geq 0$.
Assuming furthermore that $\mbox{Pic}(S)=\mathbb Z\cdot C$, it is also well-known that both $E$ and $F$ are $C$-stable bundles on $S$.
\subsection{The rank $2$ case}
We begin by showing that in rank $2$, irrespective of the structure of $\mbox{Pic}(S)$, a splitting of the restriction $E|_C$ can only be induced by an elliptic pencil on the $K3$ surface.
\begin{thm}
\label{thm: simple LM}
Let $C\subset S$ be as above and $A\in W^1_d(C)$ a base point free pencil of degree $2<d<g-1$ with $K_C\otimes A^{\vee}$
globally generated. The following conditions are equivalent:
\begin{enumerate}
\item[(i)] $E|_C\cong A\oplus (K_C\otimes A^{\vee})$;
\item[(ii)] There exists an elliptic pencil $N\in\mathrm{Pic}(S)$ such that
$N|_C= A$.
\item[(iii)] $h^0(S,E\otimes F)<h^0(C,E\otimes F|_C)$.
\end{enumerate}
\end{thm}
\begin{cor}
With notation as above, if $g\leq 2d-2$ and $A$ is not induced by an elliptic pencil on $S$, then
$E|_C$ is simple if and only if $E$ is simple.
\end{cor}
Note that it is easy to see that if $E|_C$ is simple, then $E$ is also simple. It is also known that if $E$ is simple, then automatically $g\leq 2d-2$.
\proof (of Theorem \ref{thm: simple LM})
\noindent {\em (ii)$\Rightarrow$(i).}
Let $N$ be an elliptic pencil with $N|_C=A$. Consider the exact sequence
\[
0\longrightarrow N^{\vee}\longrightarrow F\longrightarrow N(-C)\longrightarrow 0.
\]
Its restriction to $C$ gives a splitting of the dual of the sequence (\ref{eqn: E|_C})
characterizing $E|_C$. Observe that since $d<g-1$, there is no morphism from $A^{\vee}$ to $K_C^{\vee}\otimes A$.
\medskip
\noindent {\em (i)$\Rightarrow$(ii).}
Conversely, suppose that $E|_C= A\oplus (K_C\otimes A^{\vee})$.
Applying $\mathrm{Hom}(K_C\otimes A^{\vee},\ -\ )$ to the sequence
(\ref{eqn: F}), we obtain an exact sequence
\[
0\longrightarrow \mathrm{Ext}^1(K_C\otimes A^{\vee}, F)\longrightarrow
\mathrm{Ext}^1(K_C\otimes A^{\vee} ,H^0(C, A)\otimes \mathcal{O}_S)\longrightarrow
\mathrm{Ext}^1(K_C\otimes A^{\vee}, A).
\]
Since the extension class $[E]\in \mathrm{Ext}^1(K_C\otimes A^{\vee}, H^0(C, A)\otimes \mathcal{O}_S)$
maps to the trivial extension in $\mathrm{Ext}^1(K_C\otimes A^{\vee}, A)$, it
follows that there exists a rank $2$ bundle $G$ on $S$ which
fits into a commutative diagram:
\begin{equation}
\label{eqn: G}
\xymatrix{ &0\ar[d]&0\ar[d]&
\\
0\ar[r]& F\ar[r]\ar[d] & H^0(A)\otimes \mathcal{O}_S\ar[d]\ar[r]&A\ar@{=}[d]\ar[r]&0
\\
0\ar[r] & G \ar[d]\ar[r] &E \ar[d]\ar[r] & A\ar[r] &0
\\
& K_C\otimes A^{\vee} \ar@{=}[r]\ar[d] & K_C\otimes A^{\vee} \ar[d] & &
\\
& 0 &0 &}
\end{equation}
Using that $H^0(S, F)=H^1(S, F)=0$, we obtain
$H^0(S, G)\cong H^0(C, K_C\otimes A^{\vee})$. Since $h^0(S, E)=h^0(C, A)+h^1(C, A)=h^0(C, A)+h^0(S, G)$,
and $h^1(S, E)=0$, it follows that $H^1(S, G)=0$. From the second row of (\ref{eqn: G}), we find that
$H^0(S, G(-C))=0$.
Furthermore, we compute
$c_1(G)=0$ and $c_2(G)=2d-2g+2$. So
$c_2(G)<0=c_1^2(G)$, that is, $G$ violates Bogomolov's
inequality, and then it sits in an extension
\begin{equation}
\label{eqn: G ext}
0\longrightarrow M\longrightarrow G\longrightarrow M^{\vee}\otimes \mathcal I_{\Gamma/S}\longrightarrow 0,
\end{equation}
where $\Gamma$ is a zero-dimensional subscheme of $S$,
and $M\in \mbox{Pic}(S)$ is such that $M^2>0$ and $M\cdot H>0$ for any ample
line bundle $H$ on $S$. In particular, $H^0(S, M^{\vee})=0$,
and hence $H^0(S, M)\cong H^0(S, G)\cong H^0(C, K_C\otimes A^{\vee})\ne 0$. Moreover, since $$h^0(S,M^\vee\otimes \mathcal I_{\Gamma/S})=h^1(S,G)=0,$$
it also follows that $H^1(S,M)=0$.
On the other hand $H^0(S, F)=0$, which implies that the composed map
$$M\longrightarrow G\longrightarrow K_C\otimes A^{\vee}$$ is non-zero; in fact, we claim that it is surjective,
that is, $M|_C= K_C\otimes A^{\vee}$.
Suppose that $M|_C=K_C\otimes A^{\vee}(-D^\prime)$, with $D^\prime\ne 0$
an effective divisor on $C$.
Since $h^0(S, G(-C))=0$, we have $h^0(S, M(-C))=0$, which
implies $h^0(S, M)\le h^0(C, M|_C)$.
Since we assumed $K_C\otimes A^{\vee}$ to be globally generated, we have that
$$h^0(S, M)\le h^0(C, K_C\otimes A^{\vee}(-D^\prime))<h^0(C, K_C\otimes A^{\vee})=h^0(S, M),$$
a contradiction.
Setting $N:=M^{\vee}(C)$, we have shown that $N|_C=A$ and there is an exact sequence
\[
0\longrightarrow M^{\vee}\longrightarrow N\longrightarrow A\longrightarrow 0.
\]
Since $h^0(S, M^{\vee})=h^1(S, M^{\vee})=0$, it follows that $h^0(S, N)=h^0(C, A)=2$ and hence
$N$ defines an elliptic pencil.
\medskip
\noindent {\em (iii)$\Rightarrow$(i).}
From the sequence (\ref{eqn: F}) twisted by $E(-C)\cong F$, we obtain that
\[
H^0(S, E\otimes F(-C))\subset H^0(C, A)\otimes H^0(S, E(-C)),
\]
and, since $F$ has no sections, it follows that $H^0(S, E\otimes F(-C))=0$.
We have an exact sequence
\[
0\longrightarrow H^0(S, E\otimes F)\longrightarrow H^0(S, E\otimes F|_C)\longrightarrow H^1(S, E\otimes F(-C)).
\]
The hypothesis implies that $H^1(S, E\otimes F(-C))\ne 0$. From (\ref{eqn: F}) twisted by $E(-C)\cong F$, we obtain
the exact sequence in cohomology
\[
0\longrightarrow H^0(C, E|_C\otimes K_C^{\vee}\otimes A)\longrightarrow H^1(S, E\otimes F(-C))\longrightarrow
H^0(C, A)\otimes H^1(S, E(-C))=0,
\]
therefore $h^0(C, E|_C\otimes K_C^{\vee}\otimes A)\ne 0$. The sequence (\ref{eqn: E|_C})
yields to an exact sequence
\[
0=H^0(C, K_C^{\vee}\otimes A^{\otimes 2})\longrightarrow H^0(C, E|_C\otimes K_C^{\vee}\otimes A)\longrightarrow
H^0(C, \mathcal{O}_C)\to H^1(C, K_C^{\vee}\otimes A^{\otimes 2}).
\]
Then $H^0(C, E|_C\otimes K_C^{\vee}\otimes A)\to
H^0(C, \mathcal{O}_C)$ is an isomorphism and, under the coboundary map $$H^0(C, \mathcal{O}_C)\ni 1\mapsto 0
\in H^1(C, K_C^{\vee}\otimes A^{\otimes 2}),$$ that is, the sequence (\ref{eqn: E|_C}) is split.
\medskip
Note that we also have $h^1(S, E\otimes F(-C))=1$ and
$h^0(C, E\otimes F|_C)=h^0(S, E\otimes F)+1$.
\medskip
\noindent {\em (i)$\Rightarrow$(iii).}
From the hypothesis and from the sequence
(\ref{eqn: E|_C}), we find
$$h^0(C, E|_C\otimes A^{\vee})=h^0(C, K_C\otimes A^{\otimes (-2)})+1.$$
Furthermore, $h^0(S, E\otimes F)=h^0(C, E|_C\otimes A^{\vee})$;
twist (\ref{eqn: E}) by $F$ and use the vanishing
of $h^0(F)$ and that of $h^1(F)$.
On the other hand, since $E|_C\cong A\oplus K_C\otimes A^{\vee}$,
we have
$$h^0(C, E\otimes F|_C)=2+h^0(C, K_C\otimes A^{\otimes (-2)}),$$
hence $h^0(C, E\otimes F|_C)=h^0(S, E\otimes F)+1$.
\endproof
\vskip 4pt
\subsection{Lazarsfeld-Mukai bundles of higher rank}
We study when the restriction $E|_C$ is a simple vector bundle. Our main tool is a variant of the Bogomolov instability theorem.
\begin{thm}\label{simple}
Let $S$ be a $K3$ surface and $C\subset S$ a smooth curve of genus $g\geq 4$ such that $\mathrm{Pic}(S)=\mathbb{Z}\cdot C$.
We fix positive integers $r$ and $d$ such that
$$\rho(g,r,d)\ge 0, \ g\geq 2r+4 \mbox{ and } d\leq \frac{3r(g-1)}{2r+2}.$$
Then for any linear series $A\in W^r_d(C)$ such that $h^0(C, A)=r+1$ and $K_C\otimes A^{\vee}$ is globally generated,
the restricted Lazarsfeld-Mukai bundle $E|_C$ is simple.
\end{thm}
Note that in the special case $\rho(g, r, d)=0$, the constraints
from the previous statement give rise to the bound $g>2r+5$.
\proof
{\em Step 1.} We first establish that the natural extension (\ref{eqn: E|_C}), that is,
$$
0\longrightarrow Q_A \longrightarrow E|_C\longrightarrow K_C\otimes A^{\vee}\longrightarrow 0
$$
is non-trivial. Assuming that (\ref{eqn: E|_C}) is trivial. Then there is an
injective morphism from $K_C\otimes A^{\vee}$ to $E|_C$ and hence
a surjective map
$F(C)\to A$. Then
$$G:=\mbox{Ker}\{F(C)\to A\}$$ is a vector
bundle of rank $r+1$ with Chern classes
$c_1(G)=(r-1)[C]$ and
$$
c_2(G)=c_2(F(C))-c_1(F(C))\cdot C+\deg(A)=2d+r(r-3)(g-1).
$$
We compute the discriminant of $G$
$$
\Delta(G)=2{\rm rk}(G)c_2(G)-({\rm rk}(G)-1)c_1^2(G)=4d(r+1)-8r(g-1)<0,
$$
hence $G$ is unstable. Applying \cite{HL02} Theorem 7.3.4, there exists
a subsheaf $M\subset G$ with
$$
\xi_{M,G}^2\ge-\frac{\Delta(G)}{r(r+1)^2},
$$
where
$
\xi_{M,G}=c_1(M)/{\rm rk}(M)-c_1(G)/{\rm rk}(G).
$ Setting $c_1(M)=k\cdot [C]$ and $s:={\rm rk}(M)$, the previous inequality becomes
$$
\left(\frac{k}{s}-\frac{r-1}{r+1}\right)^2(2g-2)\ge \frac{8r(g-1)-4d(r+1)}{r(r+1)^2}.
$$
Note that $M$ destabilizes $G$, which coupled with the stability of $F(C)$ yields
$$
\frac{r-1}{r+1}\leq \frac{k}{s}< \frac{r}{r+1},
$$
implying after manipulations $2d(r+1)>3(g-1)r$, thus contradicting the hypothesis.
\medskip
{\em Step 2.} Assuming that $E|_C$ is non-simple, we
deduce that the extension (\ref{eqn: E|_C}) splits. We consider the exact sequence
$$
H^0(S, E\otimes F)\longrightarrow H^0(C, E\otimes F|_C)\longrightarrow H^1(S, E\otimes F(-C)).
$$
and it suffices to show that $H^1(S, E\otimes F(-C))=0$. Assuming this not to be the case, twisting (\ref{eqn: F}) by $E(-C)$ induces
the exact sequence
$$
H^0(C, A\otimes E|_C\otimes K_C^{\vee})\longrightarrow H^1(S, E\otimes F(-C))\longrightarrow H^0(C, A)\otimes H^1(S, E(-C)).
$$
Since $H^1(S, E(-C))=0$, we obtain that $H^0(C, A\otimes E|_C\otimes K_C^{\vee})\ne 0$.
Furthermore, $Q_A$ is a stable bundle and since $\mu(Q_A\otimes A\otimes K_C^{\vee})<0$, we find that
$$H^0(C, Q_A\otimes A\otimes K_C^{\vee})=0,$$ hence we also have the sequence
induced from (\ref{eqn: E|_C}) after twisting with $A\otimes K_C^{\vee}$
$$
0\longrightarrow H^0(C, E|_C\otimes K_C^{\vee}\otimes A)\longrightarrow
H^0(C, \mathcal O_C)\longrightarrow H^1(C, K_C^{\vee}\otimes A\otimes Q_A).
$$
We conclude that the coboundary map $H^0(C, \mathcal O_C)\to H^1(C, Q_A\otimes A\otimes K_C^{\vee})$
is trivial, that is, $E|_C\cong Q_A\oplus (K_C\otimes A^{\vee})
$, which completes the proof.
\endproof
\section{Stability of restricted Lazarsfeld-Mukai bundles}
\label{sec:stable}
\subsection{The rank $2$ case} If $C\subset S$ is an ample curve, then with one exception ($g=10$ and $C$ a smooth plane sextic), $\mbox{Cliff}(C)$ is computed by a pencil, see \cite{CP95} Proposition 3.3. We show that in rank $2$ the semistability of the LM bundle is preserved under restriction.
\begin{thm}
\label{thm: stable LM}
Let $S$ be a $K3$ surface, $C\subset S$ an ample curve of genus $g\geq 4$ and $A\in W^1_d(C)$ a pencil computing $\mathrm{Cliff}(C)$.
If $E_{C,A}$ is $C$-semistable on $S$, then $E|_C$ is also semistable on $C$. Moreover, if $E_{C,A}$ is $C$-stable on $S$, then $E|_C$ is stable on $C$.
\end{thm}
\proof
The proof of the stability is similar, and hence we discuss the semistability part only.
We write $A=\mathcal{O}_C(D)$, where $D$ is an effective divisor on $C$. Suppose $E|_C$ is unstable and consider an exact sequence
\[
0\longrightarrow L_1\longrightarrow E|_C\longrightarrow K_C\otimes L_1^{\vee}\longrightarrow 0,
\]
with
$\mathrm{deg}(L_1)\ge g$. Since $L_1\nsubseteq A$, the composed map
$L_1\to E|_C\to K_C\otimes A^{\vee}$
must be non-zero, that is, $L_1=K_C(-D-D_1)$,
where $D_1$ is an effective divisor on $C$. Set $d_1:=\mathrm{deg}(D_1)$. Consider the elementary modification
\begin{equation}
\label{eqn: V}
0\longrightarrow V\longrightarrow E\longrightarrow A(D_1)\longrightarrow 0
\end{equation}
induced by the composition
$E\to E|_C\to A(D_1)$. Then
$$c_1(V)=0\ \mbox{ and } \ c_2(V)=2d+d_1-2g+2<0,$$ hence $V$ is unstable with respect to any polarization and fits in an exact sequence
\begin{equation}
\label{eqn: V ext}
0\longrightarrow M\longrightarrow V\longrightarrow M^{\vee}\otimes \mathcal I_{\Gamma/S} \longrightarrow 0,
\end{equation}
where $\Gamma\subset S$ is a $0$-dimensional subscheme and $M$ is a divisor class that intersects positively any ample class on $S$ and with $M^2>0$.
From (\ref{eqn: V}) and (\ref{eqn: V ext}) we find that $H^0(S, M)\cong H^0(S, V)$ and $H^0(S, M(-C))=0$.
Dualizing (\ref{eqn: V}), we obtain the sequence
\[
0\longrightarrow F\longrightarrow V^{\vee}\longrightarrow K_C(-D-D_1)\longrightarrow 0,
\]
from which, using that $V\cong V^{\vee}$, we obtain $H^0(S, V)=H^0(C, K_C(-D-D_1))$.
\vskip 3pt
We claim that $\mathrm{Cliff}(A(D_1))=\mathrm{Cliff}(C)$.
Recall that $h^0(S, E)=h^0(C, A)+h^1(C, A)$, and, from the sequence
(\ref{eqn: V}) we write
$$h^0(S, E)\le h^0(C, A(D_1))+h^1(C, A(D_1)).$$
By assumption, the pencil $A$ computes $\mathrm{Cliff}(C)$, which implies
\[
\mathrm{Cliff}(C)=g+1-h^0(A)-h^1(A)\ge g+1-h^0(A(D_1))-h^1(A(D_1))
=\mathrm{Cliff}(A(D_1)).
\]
It follows that $\mathrm{Cliff}(A(D_1))=\mathrm{Cliff}(C)$, in particular $K_C(-D-D_1)$ is globally
generated.
\vskip 3pt
Clearly, $M\nsubseteq F$, hence the composition $\varphi: M\to V\to K_C(-D-D_1)$ is non-zero and one writes $\mbox{Im}(\varphi)=K_C(-D-D_1-D_2)$, where $D_2$
is an effective divisor on $C$.
If $D_2\ne 0$, then one has the sequence of inequalities
\[
h^0(S, M)\le h^0(C, K_C(-D-D_1-D_2))<h^0(C, K_C(-D-D_1))=h^0(S, M),
\]
a contradiction. Therefore $M|_C=K_C(-D-D_1)$. Viewing $M$ as a subsheaf of $E$, we find $\mu(M)=M\cdot C=\mathrm{deg}(L_1)>\mu(E)$, thus bringing the proof to an end.
\endproof
\begin{rmk}
If $E_{C, A}$ is stable, then it is simple and hence $d=\lfloor \frac{g+3}{2}\rfloor $, see \cite{L1}. Conversely, if $C'\subset S$ is an ample curve of genus $g$ and gonality $\lfloor \frac{g+3}{2}\rfloor$, then it was shown in \cite{LC} that the LM bundle $E_{C,A}$ corresponding to a general curve $C\in |\mathcal{O}_S(C')|$ and a pencil $A\in W^1_{\lfloor \frac{g+3}{2}\rfloor}(C)$ is $C$-semistable (even stable when $g$ is odd).
\end{rmk}
\subsection{Stability of Lazarsfeld-Mukai bundles of rank four}
We show that restrictions of LM bundles of rank $4$ on very general $K3$ surfaces of genus $g\geq 20$ are stable. Similar results were established in \cite{V} and \cite{FO12} for rank $2$ and $3$ respectively. We fix integers $i\geq 6$ and $\rho\geq 0$ and write
$$g:=4i-4+\rho \ \ \mbox{ and } \ \ d:=3i+\rho,$$
so that $\rho(g, 3, d)=\rho$. Let $S$ be a $K3$ surface and $C\subset S$ a curve of genus $g$ such that $\mbox{Pic}(S)=\mathbb Z\cdot C$, and pick a globally generated linear series $A\in W^3_d(C)$ with $h^0(C, A)=4$.
\vskip 5pt
\noindent \emph{Proof of Theorem \ref{thm: rank four}.} Our previous results show that $E|_C$ is simple, hence indecomposable. Suppose $E|_C$ is not stable and fix a maximal destabilizing sequence
\[
0\longrightarrow M\longrightarrow E|_C\longrightarrow N\longrightarrow 0.
\]
Put $d_N:=\mathrm{deg}(N)$ and $d_M:=\mathrm{deg}(M)=2g-2-d_N$. Since $M$ is destabilizing,
\begin{equation}
\label{eqn: M destab}
\frac{d_M}{\mathrm{rk}(M)}\ge \frac{g-1}{2},\ \ \ \frac{d_N}{\mathrm{rk}(N)}\le \frac{g-1}{2}.
\end{equation}
The bundle $N$, being a quotient of $E$, is globally generated. Since $H^0(C, E|_C^{\vee})=0$, clearly $N\neq \mathcal{O}_C$, therefore $h^0(C, N)\geq 2$. From the inequalities (\ref{eqn: M destab}) it follows that $\mbox{rk}(N)>1$, because $C$ has maximal gonality.
\medskip
{\em Step 1.} We prove that $M$ is a line bundle. Assume that, on the contrary,
$$\mbox{rk}(M)=\mbox{rk}(N)=2$$ and
consider the elementary modification $G:=\mathrm{Ker}\{E\to N\}$.
Its Chern classes are given as follows:
$$
c_1(G)=-[C],\ \ \
c_2(G)=d+d_N-2(g-1),
$$
and its discriminant equals
$\Delta(G)=-64 i + 110 + 8 d_N-14\rho<0$,
because of (\ref{eqn: M destab}). In particular, there exists a saturated subsheaf $F\subset G$
which verifies the inequalities
\begin{equation}
\label{eqn: destab}
\mu(G)\le \mu(F)<\mu(E), \ \ \mbox{ and }
\end{equation}
\begin{equation}
\label{eqn: HL}
\xi^2_{F,G}\ge -\frac{\Delta(G)}{48}.
\end{equation}
Write $c_1(F)=\alpha \cdot [C]$ and $\mathrm{rk}(F)=\beta\le 3$. The above
inequality (\ref{eqn: HL}) becomes
\[
\left(\frac{\alpha}{\beta}+\frac{1}{4}\right)^2(2g-2)
\ge -\frac{\Delta(G)}{48}.
\]
We apply (\ref{eqn: destab}) for $\mu(F)=\alpha(2g-2)/\beta$ and obtain
$$-\frac{1}{4}\leq \frac{\alpha}{\beta}<\frac{1}{4},$$
hence $\alpha =0$, and the inequality (\ref{eqn: HL}) reads in this case
$
d_N\ge 5i-10+\rho.
$ Recalling that $d_N\le g-1=4i-5+\rho$, we obtain a contradiction whenever $i\ge 6$.
\medskip
{\em Step 2.} We construct an elementary modification, in order to reach a contradiction.
\medskip
From (\ref{eqn: M destab}), we have $d_M\ge \frac{g-1}{2}$.
The composite map $M\rightarrow E|_C\rightarrow K_C\otimes A^{\vee}$ is not zero, for else $M\hookrightarrow Q_A$ and since $\mu(Q_A\otimes M^{\vee})<0$, one contradicts the semistability of $Q_A$. We set $A_1:=K_C\otimes A^{\vee}\otimes M^{\vee}$ and obtain a surjection $F(C){|_C}\to A\otimes A_1$
inducing, as before, an elementary modification
$$V:=\mbox{Ker}\{F(C)\rightarrow A\otimes A_1\}.$$
By direct computation we show that $\Delta(V)<0$.
Indeed, we compute
$$c_1(V)=2\cdot [C],\ \ \ c_2(V)=d+2g-2-d_M, \ \ \mbox{ hence }
$$
$$
\Delta(V)=8c_2(V)-3c_1^2(V)=8(d-d_M-g+1)=
8(5-d_M-i)<0.
$$
We obtain a destabilizing sheaf $P\subset V$,
with ${\rm rk}(P)=b\le 3$ and $c_1(P):=a\cdot [C]$, such that the following inequalities are both satisfied
\begin{equation}
\label{eqn: HL 2}
\left(\frac{a}{b}-\frac{1}{2}\right)^2(2g-2)\ge
-\frac{\Delta(V)}{48} \ \ \ \mbox{ and } \ \ \ \mu(V)\le\mu(P)<\mu(F(C)).
\end{equation}
The second inequality gives $\frac{1}{2}\le \frac{a}{b}< \frac{3}{4}$, which leaves two possibilities: either $a=1$ and $b=2$, when via (\ref{eqn: HL 2}) one finds
that $\Delta(V)\geq 0$, a contradiction, or else $a=2$ and $b=3$,
when inequalities (\ref{eqn: HL 2}) and (\ref{eqn: M destab}) clash.
\hfill
$\Box$
\section{Normal bundle of canonical curves of genus $7$}
The aim of this section is to prove Theorem \ref{g7} and we begin by recalling Mukai's results \cite{M3} on canonical curves of genus $7$. We choose a vector space $U:=\mathbb C^{10}$ and a non-degenerate quadratic form $q:U\rightarrow \mathbb C$, defining a smooth $8$-dimensional quadric
$Q\subset {\textbf P}(U)={\textbf P}^9$.
The algebraic group $\mbox{{\bf Spin}}(U)$ corresponding to the Dynkin diagram $D_5$ admits two $16$-dimensional half-spin representations $\mathcal{S}^+$ and $\mathcal{S}^-$, which correspond to maximal weights $\alpha^+$ and $\alpha^-$ respectively. The homogeneous spaces $V^{\pm}:=\mbox{{\bf Spin}}(U)/P(\alpha^{\pm})$ are both $10$-dimensional and can be realized as the two irreducible components of the Grassmannian $G_q(5, U)$ of projective $4$-planes inside ${\textbf P}(U)$ which are isotropic with respect to the quadratic form $q$. From now on, we set
$$V:=V^+\subset {\textbf P}(\mathcal{S}^+)={\textbf P}^{15}.$$
Note that $\mbox{Aut}(V)=SO(10)$. If $\mathcal{E}$ is the restriction to $V$ of the tautological bundle on $G(5, 10)$, one has an exact sequence of vector bundles on $V$:
\begin{equation}\label{spin}
0\longrightarrow \mathcal{E}^{\vee}\longrightarrow U\otimes \mathcal{O}_V\longrightarrow \mathcal{E}\longrightarrow 0.
\end{equation}
By the adjunction formula, smooth curvilinear sections of $V$ are canonical curves of genus $7$ and Mukai \cite{M3} showed that \emph{each} curve $[C]\in \mathcal{M}_7$ with $\mbox{Cliff}(C)=3$ appears in this way. Precisely, there is a birational map
$$\alpha:G(7, 16){/\!/} SO(10)\dashrightarrow \overline{\mathcal{M}}_7, \ \ \alpha(\Lambda):=[\Lambda\cap V],$$
where $\Lambda\cong {\textbf P}^6$. Given a curve $[C]\in \mathcal{M}_7$, the inverse $\alpha^{-1}([C])$ is constructed precisely via the twist of the conormal bundle on $C$ mentioned in the introduction.
\vskip 3pt
Let $C\subset {\textbf P}^{6}$ be a smooth canonical curve with $\mbox{Cliff}(C)=3$, and set $E:=N_{C{\textbf P}^6}^{\vee}(2)$. One has an identification $H^0(C, E)=I_2(K_C)$ and $E$ is a globally generated
bundle. The tautological map
$$\phi_E:C\rightarrow G(5, H^0(C,E))$$ is easily shown to be injective and its image lies on $V$. In particular, the vector bundle $E$ is the restricted spinorial bundle, that is, $E=\mathcal{E}_{|C}$ and one has an exact sequence:
\begin{equation}\label{spin2}
0\longrightarrow E^{\vee}\longrightarrow H^0(C, E)\otimes C\longrightarrow E\longrightarrow 0.
\end{equation}
Note that $W^1_4(C)=\emptyset$, while $W^1_5(C)$ is a curve. We are going to make essential use of the following fact:
\begin{lem}
Let $C$ as above and $A\in W^1_5(C)$. Then there are no surjections $E\twoheadrightarrow A$.
\end{lem}\label{surj}
\begin{proof} We proceed by contradiction. Assume that there is such a pencil $A\in W^1_5(C)$, then use the base point free pencil trick to write the following diagram:
\begin{equation}
\xymatrix{
0\ar[r]& E^{\vee}\ar[r]\ar[d] & H^0(C,E)\otimes \mathcal{O}_C\ar[d]\ar[r]&E\ar[d]\ar[r]&0
\\
0\ar[r] & A^{\vee} \ar[r] &H^0(C,A)\otimes \mathcal{O}_C \ar[r] & A\ar[d]\ar[r] &0
\\
& & & 0&
}
\end{equation}
In particular, $H^0(C, E\otimes A^{\vee})\neq 0$. Via the identification $H^0(C,E)=I_2(K_C)$, this implies that if $L:=K_C\otimes A^{\vee}\in W^2_7(C)$, then the multiplication map
$$\mbox{Sym}^2 H^0(C,L)\rightarrow H^0(C, L^{\otimes 2})$$ is not injective. This is possible only if $L$ is not birationally very ample, in particular, $C$ must be trigonal, which is not the case.
\end{proof}
\vskip 3pt
We are now in a position to prove that the twist $E$ of the conormal bundle of a canonical curve of genus $7$ is stable.
\vskip 5pt
\noindent \emph{Proof of Theorem \ref{g7}.} Suppose that $0\rightarrow F\rightarrow E\rightarrow M\rightarrow 0$ is a destabilizing sequence for the vector bundle $E$, that is, with
$\mu(F)\geq \mu(E)=\frac{24}{5}$. Since $E$ is globally generated, so is any of its quotient, in particular $M$ too. We distinguish several possibilities, depending on the ranks that appear:
\vskip 4pt
\noindent {\bf (i)} $\mbox{rk}(F)=4$ and $M$ is line bundle. Then $\mbox{deg}(F)\geq 20$, hence $\mbox{deg}(M)\leq 4$. Since $C$ is not tetragonal, $h^0(C,M)\leq 1$. Note that $M\neq \mathcal{O}_C$, for $H^0(C, E^{\vee})=0$. It follows that $M$ is not globally generated, a contradiction.
\vskip 3pt
\noindent {\bf (ii)} $\mbox{rk}(F)=1$ and we may assume that $\mbox{deg}(F)=5$. Suppose first that $h^0(C, F)=0$, therefore $h^0(C, K_C\otimes F^{\vee})=1$, and hence $K_C\otimes F^{\vee}$ is not globally generated. Since one has a surjection $E^{\vee}(1)\twoheadrightarrow K_C\otimes F^{\vee}$, we reach a contradiction by observing that $E^{\vee}(1)$ is globally generated. Indeed, via Serre duality, this last statement is equivalent to the equality $h^0(C, E(p))=h^0(C,E)=10$, for every point $p\in C$. From the exact sequence
$$0\longrightarrow E(p)\longrightarrow M_{K_C}\otimes K_C(p)\longrightarrow K_C^{\otimes 3}(p)\longrightarrow 0,$$
we obtain that $H^0(C, E(p))=\mbox{Ker}\Bigl\{H^0(C, M_{K_C}\otimes K_C(p))\rightarrow H^0(C, K_C^{\otimes 3}(p))\Bigr\}$.
The conclusion follows, since $H^0(C, M_{K_C}\otimes K_C)=H^0(C, M_{K_C}\otimes K_C(p))$.
\vskip 5pt
Suppose now that $h^0(C, F)\geq 1$. The case $h^0(C, F)\geq 2$ having been discarded in the course of proving Lemma \ref{surj}, we assume that $h^0(C, F)=1$, hence $h^0(C,K_C\otimes F^{\vee})=2$.
We obtain that the multiplication map
$$\mbox{Sym}^2 H^0(C, K_C\otimes F^{\vee})\rightarrow H^0(C, K_C^{\otimes 2}\otimes F^{\otimes (-2)})$$ is not injective, which contradicts the base point free pencil trick.
\vskip 5pt
\noindent {\bf (iii)} $\mbox{rk}(F)=3$, and then $\mbox{deg}(F)\geq 15$, hence $\mbox{deg}(M)\leq 9$. This time we may assume that $F$ is stable. If $M$ is not stable, we choose a line subbundle $A\subset M$ of maximal degree, which we pull-back under the surjection $E\twoheadrightarrow M$, to obtain the exact sequence
$$0\longrightarrow G\longrightarrow E\longrightarrow M/A\longrightarrow 0.$$
We obtain that $\mbox{deg}(M/A)\leq \mbox{deg}(M)/2\leq 9/2$, that is, $\mbox{deg}(M/A)\leq 4$. In particular, $M/A$ is not globally generated, which is again a contradiction, so we can assume that both $F$ and $M$ are stable vector bundles. Since $h^0(C,M)+h^0(C, F)\geq h^0(C,E)=10$, the strategy is to use the fact that the Mercat statements $(M_2)$ and $(M_3)$ have been established for curves $C$ of genus $7$ with maximal Clifford index, that is,
$$\mbox{Cliff}_2(C)=\mbox{Cliff}_3(C)=3,$$
see \cite{LN3} Theorem 4.5. In particular, if both $F$ and $M$ contribute to their respective Clifford indices, that is, $h^0(C,F)\geq 6$ and $h^0(C,M)\geq 4$ respectively, then we write
$$\frac{9}{2}+3\leq \frac{3}{2} \gamma(F)+\gamma(M)=\frac{1}{2}\Bigl(\mbox{deg}(F)+\mbox{deg}(M)\Bigr)-h^0(C, F)-h^0(C, M)+5,$$
that is, $h^0(C, F)+h^0(C,M)\leq \frac{19}{2}$, a contradiction.
\vskip 3pt
Assume now that one of the bundles $F$ or $M$ does not contribute to its Clifford index. Since $M$ is globally generated, $h^0(C,M)\geq 2$. We can have $h^0(C,M)=2$, only when $M=\mathcal{O}_C^{\oplus 2}$, which is impossible, for $\mathcal{O}_C^{\oplus 2}$ is not a direct summand of $E$. If $h^0(C,M)=3$, then $\mbox{deg}(M)\geq 7$, and one has equality if and only if $M=Q_L$, where $L\in W^2_7(C)$. Assuming this to be the case, we choose two points $p, q\in C$ that correspond to a node in the plane model $\phi_L:C\rightarrow {\textbf P}^2$, that is, $A:=L(-p-q)\in W^1_5(C)$. Then there is a surjection $Q_L\twoheadrightarrow A$, which by composition gives rise to a surjective morphism $E\twoheadrightarrow A$. This contradicts Lemma \ref{surj}.
Thus we may assume that $\mbox{deg}(M)\geq 8$, and accordingly, $\mbox{deg}(F)\leq 16$. Then we compute
$$\gamma(F)=\mu(F)-\frac{2}{3} h^0(C,F)+2\leq \frac{16}{3}-\frac{14}{3}+2<\mbox{Cliff}(C),$$
which again contradicts the equality $\mbox{Cliff}_3(C)=3$.
\vskip 4pt
\noindent {\bf (iv)} $\mbox{rk}(F)=2$, and then $\mbox{deg}(F)\geq 10$ and $\mbox{deg}(M)\leq 14$. We may assume this time that $M$ is stable. If $F$ is not stable, then it has a line subbundle $A\hookrightarrow F$ with $\mbox{deg}(A)\geq 5$, and we are back to case (ii). Thus both $M$ and $F$ are stable bundles, and we proceed precisely like in case (iii).
\hfill $\Box$
\vskip 5pt
It is instructive to remark that the normal bundle of a canonical curve of genus $g<7$ is never stable. More generally we have the following:
\begin{prop}
The normal bundle of a tetragonal canonical curve of genus $g$ is unstable.
\end{prop}
\begin{proof} More generally, we begin with a $k:1$ covering $f:C\rightarrow {\textbf P}^1$, and consider the rank $(k-1)$-vector bundle $\mathcal{F}^{\vee}:=f_*\mathcal{O}_C/\mathcal{O}_{{\textbf P}^1}$ on the projective
line. Then $\pi:X={\textbf P}(\mathcal{F})\rightarrow {\textbf P}^1$ is a scroll of dimension $k-1$, which contains the canonical curve $C$ and which can be embedded by the tautological bundle
$\mathcal{O}_{X}(1)$ in ${\textbf P}^{g-1}$ as a variety of degree $g-k+1$. Denoting by $H, R\in \mbox{Pic}(X)$ the class of the hyperplane section and that of the ruling respectively,
we have $$K_X\equiv -(k-1)H+(g-k-1)R,$$
whereas obviously $C\cdot H=2g-2$ and $C\cdot R=k$. We compute the degree of the normal bundle $N_{C/X}$ and find:
$$\mbox{deg}(N_{C/X})=\mbox{deg}(T_{X|C})+\mbox{deg}(K_C)=k(g+k-1).$$
We write the usual exact sequence relating normal bundles
$$0\longrightarrow N_{C/X}\longrightarrow N_{C/{\textbf P}^{g-1}}\longrightarrow N_{X/{\textbf P}^{g-1}}\otimes \mathcal{O}_C \longrightarrow 0,$$
and compare the slopes
$$\mu(N_{C/X})=\frac{k(g+k-1)}{k-2} \ \ \mbox{ and } \ \ \mu(N_{C/{\textbf P}^{g-1}})=\frac{2(g-1)(g+1)}{g-2}.$$
We conclude that for $k=4$ and $g\geq 6$, the normal bundle $N_{C/X}$ is a destabilizing subbundle of $N_{C/{\textbf P}^{g-1}}$. For $g$ at most $5$, every canonical curve of genus $g$ is a complete intersection which obviously produces a destabilizing line subbundle.
\end{proof}
|
\section{Introduction}
\label{sec:intro}
Lateral motion of membrane components is required for proper physiological functioning in
cellular biology \cite{berg77,gennisbook,ram09,hell09,lingwood} and serves as fertile grounds
for biophysical studies involving lipid bilayer model systems \cite{ott,saxton97,greenbook,keller_review,simonsvazrafts2004,brown_hydro_review,Ash2004158}.
A key aspect of dynamics taking place at the membrane surface is hydrodynamic flow, both within the membrane
itself and in the fluids surrounding the membrane. The membrane environment is thus distinct from traditional three dimensional (3D) or
two dimensional (2D) hydrodynamic systems, incorporating aspects of both 2D and 3D flow; the membrane
components move in a two dimensional space defined by the membrane surface, but are coupled to fluid motions in a full 3D space.
For this reason, it has been suggested that membrane hydrodynamics might best be described as ``quasi-2D" \cite{Haim09}.
Saffman and Delbr\"uck \cite{saf} (SD) introduced the standard hydrodynamic model for lipid bilayer membranes. In the SD model,
the membrane is treated as a flat thin incompressible fluid sheet with surface viscosity $\eta_m$, surrounded by an infinite incompressible aqueous bulk of viscosity $\eta$.
This model can be solved in the creeping flow limit to yield analytical predictions for the diffusion coefficient of cylindrical
bodies embedded within the membrane via no-slip boundary conditions \cite{saf,white81}, which are in good agreement
with experiments recording the diffusion of membrane proteins and lipid domains \cite{peters82,vaz82,cicuta07,petrov08,nguyen,nguyen2,krieg,ram10}. (However, some experimental studies claim deficiencies in
the SD model for membrane protein diffusion \cite{gambin06,gambin10}.) The
SD model also correctly predicts the dynamics of lipid domain boundary fluctuations \cite{stonemcconnell95,camleyesposito2010}
and the dynamics of phase separation in ternary model membrane systems \cite{camleybrown2010,camleybrownscaling2011,stanich2013coarsening,honerkamp2012experimental,haataja2009critical,inaura2008concentration}.
Perhaps the most concise way to express the physics contained within the SD model is through the Green's function
formulation of the hydrodynamic problem. The membrane's velocity response to in-plane forcing
is given by
\begin{equation}
\label{eq:v-field}
\mathbf{v}(\mathbf{r})= \int \mathrm{d}\mathbf{r}'\, \mathbf{T}(\mathbf{r}-\mathbf{r}') \mathbf{f}(\mathbf{r}')\,.
\end{equation}
Here, $\mathbf{v}(\mathbf{r})$ is the $xy$ velocity of the membrane at position $\mathbf{r}=(x,y)$. (We assume the membrane
plane is coincident with the $xy$ plane of our coordinate system.) $\mathbf{f}(\mathbf{r})$ is an in-plane force/area applied to the
membrane and the Green's function tensor $\mathbf{T}(\mathbf{r})$ captures the hydrodynamic response to forcing in the SD model.
Explicitly \cite{Haim09,levine02},
\begin{eqnarray}
\label{eq:mem_green}
T_{ij}(\mathbf{r}) &=& \frac{1}{4 \eta_m}\left \{ \left [ H_0(\tilde{r}) - \frac{H_1 (\tilde{r})}{\tilde{r}} -\frac{Y_0(\tilde{r})}{2} + \frac{Y_2(\tilde{r})}{2} + \frac{2}{\pi \tilde{r}^2} \right ] \right .\delta_{ij} \nonumber \\
& & \left . - \left [ H_0 (\tilde{r}) - \frac{2 H_1 (\tilde{r})}{\tilde{r}} + Y_2(\tilde{r}) + \frac{4}{\pi \tilde{r}^2} \right ] \frac{\tilde{r}_i \tilde{r}_j}{\tilde{r}^2} \right \}
\end{eqnarray}
where $Y_n$ and $H_n$ are Bessel functions of the second kind and Struve functions, respectively.
The dimensionless distance $\tilde{r} = |\mathbf{r}|/\ell_0$, where $\ell_0 = \eta_m / 2 \eta$ is
the Saffman-Delbruck length. $\ell_0$ defines a crossover between 2D and 3D-like
hydrodynamics in the quasi-2D SD model; for separations well below $\ell_0$ the Green's function is
that of a 2D fluid with surface viscosity $\eta_m$, whereas at separations larger than $\ell_0$ the Green's function is
similar to that of a 3D system of viscosity $\eta$ \cite{lubensky96,Haim09}. For typical lipid bilayer systems
$\ell_0$ is on the order of a micron \cite{saf}.
In principle, Eq. \ref{eq:v-field} is restricted to fluid regions and would not
immediately seem to be of use in solving problems involving particles suspended within the fluid (e.g. prediction of diffusion coefficients).
However, within creeping flow, there is no physical or mathematical distinction between a solid particle embedded within the fluid via no-slip
boundary conditions and a ``fluid" region that occupies the same physical space as the embedded particle with the supplemental restriction
to undergo only rigid body motions. Computational strategies that supplement Eq. \ref{eq:v-field} (typically the 3D version of it)
with the constraint of rigid-body motion over particle associated regions within the fluid have the capability to predict particulate
dynamics. Well known schemes that take advantage of this strategy to calculate diffusion coefficients and related properties associated
with particulate flows include the Kirkwood approximation \cite{bloomfield67,doi}, shell method
\cite{deutchbio76,swanson78} and related techniques \cite{peskin,torre,wang2013assessing}. (See Ref.~\citenum{torre} and the references within for
a review of these approaches.) Though the majority of work in this direction has been aimed at studying various solid bodies
immersed in homogeneous 3D hydrodynamic environments, the quasi-2D membrane case has previously been considered
with a method similar to the Kirkwood approximation by Levine Liverpool and MacKintosh \cite{levine04,levine04lett}. (It should
also be mentioned that the Green's function approach naturally leads to diffusion coefficients for circular fluid domains in the
membrane geometry \cite{dekoker,komura}.)
Though the Kirkwood approximation is simple to implement and often gives rise to qualitatively correct predictions,
the method is known to be inexact and, in extreme cases, can yield unphysical results \cite{zwanzig_weiss}.
In 3D geometries, the shell method \cite{deutchbio76,swanson78} resolves these shortcomings, however
this method is not easily applied to the SD problem as it requires use of the Rotne-Prager tensor \cite{rotne_prager};
a quasi-2D analog to the Rotne-Prager tensor has not been derived. An appealing alternative to the shell method is
the method of Regularized Stokeslets (RS) \cite{cortez01,cortez05}. (Section \ref{sec:rs-method} provides an introduction
of the RS method.)
This technique is known to reproduce the
correct velocity fields and particle diffusion coefficients for model problems in 3D \cite{cortez05} and is immediately extended
to the SD hydrodynamic geometry without difficulty. In recent work ~\cite{camley13}, two of us presented this
extension and verified that RS calculations quantitatively reproduce the known analytical results
for single-particle mobility problems (i.e. prediction of diffusion coefficients) in the quasi-2D environment.
While single-particle mobility/diffusion problems present a convenient test of computational
methodologies, the methods discussed above (RS included) are applicable to a broader
class of problems involving the motions of multiple embedded bodies and the hydrodynamic
interactions between them. The study of many body dynamics in fluid environments has
received considerable theoretical/numerical attention \cite{montgomery77,montgomeryDeutch77,batchelor76,murphy73,
kimbook,schmitz82,felder77,felder78,felder83,schmitz82b,felder88,mccammon,stokesian}, but
primarily in the context of 3D systems. The few prior studies in the membrane geometry
are restricted to limiting regimes and circular particles \cite{bussel92,Haim09,henle2009effective}, due to the focus
of these works on analytical calculations. The complicated form of Eq. \ref{eq:mem_green}
hinders analytical progress without invoking some manner of approximation. A versatile numerical
tool to quantitatively study the dynamics of multiple membrane-embedded bodies has yet
to be developed, despite the growing experimental interest in the dynamics of solid body
suspensions in membrane systems and related quasi-2D environments
\cite{weeks06,petrov08, petrov12, knight10}.
In this work, we extend the membrane RS method to consider many-body systems.
A general framework is presented, which is applied in detail to the the study of
two circular disks within the membrane. We also briefly revisit our earlier calculations
\cite{camley13} involving the diffusion of diamond shaped lipid domains and tethered protein
assemblies, as motivated by recent experiments \cite{petrov12,knight10}.
This work is structured as follows: In Sec.~\ref{sec:dif-tensor} we discuss and
present the diffusion matrix for a many-body system. In Sec.~\ref{sec:rs-method} we
generalize the quasi-2D RS method to calculations involving multiple bodies embedded in the membrane. This
allows for the calculation of the full many-body diffusion matrix.
Though both sections \ref{sec:dif-tensor} and \ref{sec:rs-method} focus primarily on the two-body case
for concreteness, the many-body generalization is straightforward and is discussed as well.
In Sec.~\ref{sec:res} we present detailed results for the case of two identical disks embedded within
the membrane. In Sections~\ref{sec:dimers} and
\ref{sec:diamonds} we consider the case of tethered two-disk dimers and correlations between diamond shaped domains, respectively.
Finally, Sec.~\ref{sec:summary} discusses our results and concludes.
\section{The Diffusion Matrix and Resistance Matrix}
\label{sec:dif-tensor}
The velocities and angular velocities of rigid bodies
embedded within a fluid are related to the forces and
torques applied to these bodies through the diffusion matrix $\mathbf{D}$
\cite{montgomery77,montgomeryDeutch77, felder88}:
\begin{equation}
\label{eq:dif-tensor}
\left(
\begin{array}{cc}
\mathbf{V} \\ \boldsymbol{\Omega}
\end{array}
\right) = \frac{\mathbf{D}}{k_B T} \left(
\begin{array}{cc}
\mathbf{F} \\ \boldsymbol{\tau}
\end{array}
\right)
=
\frac{1}{k_BT}
\left(
\begin{array}{cc}
\mathbf{D}_\mathrm{T} & \mathbf{D}_\mathrm{TR} \\
\mathbf{D}_\mathrm{RT} & \mathbf{D}_\mathrm{R}
\end{array}
\right)
\left(
\begin{array}{cc}
\mathbf{F} \\ \boldsymbol{\tau}
\end{array}
\right)\,.
\end{equation}
Here, $\mathbf{F}$, $\boldsymbol{\tau}$, $\mathbf{V}$ and
$\boldsymbol{\Omega}$ contain the vector components of force, torque, velocity and angular velocity for all the bodies
present in the fluid. $k_B$ is the Boltzmann constant ant $T$ is temperature.
This formulation, which describes an instantaneous transmission of force through an otherwise quiescent fluid, assumes an overdamped
``creeping flow" hydrodynamic regime neglecting inertial effects. (The mobility matrix $\mathbf{M} = \mathbf{D}/k_B T$
notation is preferred by some authors; the relation between mobility and diffusion is assured by
the Einstein relation \cite{montgomery77,condiff66}.)
The blocks $\mathbf{D}_\mathrm{T}$ and $\mathbf{D}_\mathrm{R}$ in Eq.~\eqref{eq:dif-tensor} are associated with
translational and rotational motion, respectively. The
blocks $\mathbf{D}_\mathrm{TR}$ and $\mathbf{D}_\mathrm{RT}$ couple
translation and rotation. Subblocks of $\mathbf{D}_\mathrm{T}$, $\mathbf{D}_\mathrm{R}$,
$\mathbf{D}_\mathrm{TR}$ and $\mathbf{D}_\mathrm{RT}$ associated with specific pairs of particles (e.g. $\mathbf{D}_\mathrm{T}^{1,1}$
or $\mathbf{D}_\mathrm{TR}^{3,5}$) are tensors.
Note that $\mathbf{D}_\mathrm{TR}=\mathbf{D}_\mathrm{RT}^{\intercal}$ where ``$\intercal$" denotes the transpose \cite{happel, condiff66}.
The inverse of $\mathbf{D}$ defines the resistance matrix.
\begin{equation}
\label{eq:diff-fric-tensor}
{\boldsymbol{\zeta} = k_BT \mathbf{D}^{-1}} .
\end{equation}
The natural physical interpretation of the resistance matrix relates the {\em hydrodynamic forces/torques}
(i.e. frictions or drags) imparted to the bodies as they move through the fluid with prescribed motions
\begin{equation}
\left(
\begin{array}{cc}
\mathbf{F_{hy}} \\ \boldsymbol{\tau}_{hy}
\end{array}
\right) = -\boldsymbol{\zeta}
\left (
\begin{array}{cc}
\mathbf{V} \\ \boldsymbol{\Omega}
\end{array}
\right ).
\end{equation}
However, the hydrodynamic drags exactly compensate the applied forces(torques) in creeping flow and
it is mathematically convenient to express this relationship in terms of the applied forces and
torques. This is the formulation obtained by inverting Eq.~\eqref{eq:dif-tensor}.
\begin{equation}
\left(
\begin{array}{cc}
\mathbf{F} \\ \boldsymbol{\tau}
\end{array}
\right) =
\label{eq:fric-tensor}
\boldsymbol{\zeta}
\left (
\begin{array}{cc}
\mathbf{V} \\ \boldsymbol{\Omega}
\end{array}
\right )
=
\left(
\begin{array}{cc}
\boldsymbol{\zeta}_\mathrm{T} & \boldsymbol{\zeta}_\mathrm{TR} \\
\boldsymbol{\zeta}_\mathrm{RT} & \boldsymbol{\zeta}_\mathrm{R}
\end{array}
\right)
\left (
\begin{array}{cc}
\mathbf{V} \\ \boldsymbol{\Omega}
\end{array}
\right )
\end{equation}
with
$\boldsymbol{\zeta}_\mathrm{TR}=\boldsymbol{\zeta}_\mathrm{RT}^\intercal$ \cite{condiff66,happel}.
Throughout this work we will be concerned with pairs of objects residing in flat membranes spanning the $xy$ plane.
Consequently, the two vectors appearing in Eqs.~\eqref{eq:dif-tensor} and ~\eqref{eq:fric-tensor} may be ordered as
$(\mathbf{F}, \boldsymbol{\tau}) \equiv (f_{1x},f_{1y},f_{2x},f_{2y},\tau_1,\tau_2)$
and $(\mathbf{V},\boldsymbol{\Omega}) \equiv (v_{1x},v_{1y},v_{2x},v_{2y},\Omega_1,\Omega_2)$ with the ${1,2}$ indices
specifying particle identity.
We will refer to these two vectors as the F-vector and V-vector, respectively.
Explicitly writing the full $6\times 6$ resistance matrix yields
(the diffusion matrix follows similarly)
\begin{equation}
\label{eq:2-obj-fric-tensor-general}
\left(
\begin{array}{c}
f_{1x} \\ f_{1y} \\ f_{2x} \\ f_{2y} \\ \tau_1 \\ \tau_2
\end{array}
\right)
=
\left(
\begin{array}{cccc|cc}
\zeta^\mathrm{11}_\mathrm{xx} & \zeta^\mathrm{11}_\mathrm{xy} & \zeta^\mathrm{12}_\mathrm{xx} & \zeta^\mathrm{12}_\mathrm{xy} & \zeta^\mathrm{11}_\mathrm{xR}&\zeta^\mathrm{12}_\mathrm{xR} \\
\zeta^\mathrm{11}_\mathrm{yx} & \zeta^\mathrm{11}_\mathrm{yy} & \zeta^\mathrm{12}_\mathrm{yx} & \zeta^\mathrm{12}_\mathrm{yy} & \zeta^\mathrm{11}_\mathrm{yR}&\zeta^\mathrm{12}_\mathrm{yR} \\
\zeta^\mathrm{21}_\mathrm{xx} & \zeta^\mathrm{21}_\mathrm{xy} & \zeta^\mathrm{22}_\mathrm{xx} & \zeta^\mathrm{22}_\mathrm{xy} & \zeta^\mathrm{21}_\mathrm{xR}&\zeta^\mathrm{22}_\mathrm{xR} \\
\zeta^\mathrm{21}_\mathrm{yx} & \zeta^\mathrm{21}_\mathrm{yy} & \zeta^\mathrm{22}_\mathrm{yx} & \zeta^\mathrm{22}_\mathrm{yy} & \zeta^\mathrm{21}_\mathrm{yR}&\zeta^\mathrm{22}_\mathrm{yR} \\
\hline
\zeta^\mathrm{11}_\mathrm{xR} & \zeta^\mathrm{11}_\mathrm{yR} & \zeta^\mathrm{12}_\mathrm{xR} & \zeta^\mathrm{12}_\mathrm{yR} & \zeta^\mathrm{11}_\mathrm{R}&\zeta^\mathrm{12}_\mathrm{R} \\
\zeta^\mathrm{21}_\mathrm{xR} & \zeta^\mathrm{21}_\mathrm{yR} & \zeta^\mathrm{22}_\mathrm{xR} & \zeta^\mathrm{22}_\mathrm{yR} & \zeta^\mathrm{21}_\mathrm{R}&\zeta^\mathrm{22}_\mathrm{R}
\end{array}
\right)
\left(
\begin{array}{c}
v_{1x} \\ v_{1y} \\ v_{2x} \\ v_{2y} \\ \Omega_1 \\ \Omega_2
\end{array}
\right)
\,,
\end{equation}
where the subscript $\mathrm{R}$ denotes elements associated with the
rotational motion.
The lines in
Eq.~\eqref{eq:2-obj-fric-tensor-general} divides the friction tensor into
$\boldsymbol{\zeta}_\mathrm{T}$, $\boldsymbol{\zeta}_\mathrm{R}$, $\boldsymbol{\zeta}_\mathrm{TR}$ and $\boldsymbol{\zeta}_\mathrm{RT}$
blocks,
see Eq.~\eqref{eq:fric-tensor}.
\begin{center}
\begin{figure}
\includegraphics[scale=0.6]{topside}
\caption{\label{fig:topside}
Schematic of the two disk system.
(a) Side view: The two disks reside in a Saffman-Delbr\"uck membrane, which treats the bilayer as a thin structureless sheet with surface viscosity $\eta_\mathrm{m}$
surrounded by a bulk fluid with viscosity $\eta_\mathrm{f}$.
(b) Top view: The disks have radius $r_0$ and center-to-center
distance $d$; equivalently, the closest approach separation is $h=d-2r_0$. }
\end{figure}
\end{center}
Most of the calculations in this work involve two identical disks of radius $r_0$,
separated by a center-to-center distance $d$ (see Fig.~\ref{fig:topside}).
This
situation introduces significant symmetry relative to the general case and the diffusion and
resistance matrices simplify considerably \cite{bussel92}. Without loss of generality, we assume the
disk centers lie on the $x$ axis, which yields
\begin{equation}
\label{eq:circ-r-mat}
\boldsymbol{\zeta}=\left(
\begin{array}{cccc|cc}
\zeta^\mathrm{s}_\mathrm{L} & 0 & \zeta^\mathrm{c}_\mathrm{L} & 0 & 0 & 0 \\
0 & \zeta^\mathrm{s}_\mathrm{T} & 0 & \zeta^\mathrm{c}_\mathrm{T} &\zeta^\mathrm{s}_\mathrm{RT} & \zeta^\mathrm{c}_\mathrm{RT} \\
\zeta^\mathrm{c}_\mathrm{L} & 0 & \zeta^\mathrm{s}_\mathrm{L} & 0 & 0 & 0 \\
0 & \zeta^\mathrm{c}_\mathrm{T} & 0 & \zeta^\mathrm{s}_\mathrm{T} &-\zeta^\mathrm{c}_\mathrm{RT} & -\zeta^\mathrm{s}_\mathrm{RT} \\
\hline
0 & \zeta^\mathrm{s}_\mathrm{RT} & 0 & -\zeta^\mathrm{c}_\mathrm{RT} &\zeta^\mathrm{s}_\mathrm{R} & \zeta^\mathrm{c}_\mathrm{R} \\
0 & \zeta^\mathrm{c}_\mathrm{RT} & 0 & -\zeta^\mathrm{s}_\mathrm{RT} &\zeta^\mathrm{c}_\mathrm{R} & \zeta^\mathrm{s}_\mathrm{R}\\
\end{array}
\right)\,,
\end{equation}
and
\begin{equation}
\label{eq:circ-dif-mat}
\mathbf{D}=\left(
\begin{array}{cccc|cc}
D^\mathrm{s}_\mathrm{L} & 0 & D^\mathrm{c}_\mathrm{L} & 0 & 0 & 0 \\
0 & D^\mathrm{s}_\mathrm{T} & 0 & D^\mathrm{c}_\mathrm{T} &D^\mathrm{s}_\mathrm{RT} & D^\mathrm{c}_\mathrm{RT} \\
D^\mathrm{c}_\mathrm{L} & 0 &D^\mathrm{s}_\mathrm{L} & 0 & 0 & 0 \\
0 & D^\mathrm{c}_\mathrm{T} & 0 & D^\mathrm{s}_\mathrm{T} &-D^\mathrm{c}_\mathrm{RT} & -D^\mathrm{s}_\mathrm{RT} \\
\hline
0 & D^\mathrm{s}_\mathrm{RT} & 0 & -D^\mathrm{c}_\mathrm{RT} &D^\mathrm{s}_\mathrm{R} & D^\mathrm{c}_\mathrm{R} \\
0 & D^\mathrm{c}_\mathrm{RT} & 0 & -D^\mathrm{s}_\mathrm{RT} &D^\mathrm{c}_\mathrm{R} & D^\mathrm{s}_\mathrm{R}\\
\end{array}
\right)\,.
\end{equation}
Here, we have introduced the subscripts $L$ and $T$ to denote motions longitudinal or transverse to the
disk separation axis and the superscripts $s$ and $c$ to denote self and coupling components to the matrices.
In our numerical calculations for the friction and diffusion matrices (see below), we do not
impose any of the above symmetries and calculate all 36 elements independently.
However, our results display the above symmetries to very high
numerical precision (e.g. the ``zero" elements in Eqs. \ref{eq:circ-r-mat} and \ref{eq:circ-dif-mat} are calculated to be $10^{-10}$ to $10^{-12}$ times smaller than the
``nonzero" elements in these matrices.), which provides a welcome check of our numerical algorithms.
The elements of the diffusion matrix predict the thermal fluctuations in position and orientation of the disks during a short time interval $\Delta t$
in the absence of any applied forcing. Generically, we express this as $2D_{\alpha \beta} \Delta t = \langle\Delta q_\alpha \Delta q_\beta \rangle$ with $q_{\alpha (\beta)}$
representing any of the coordinates or angular coordinates in our system and it is understood that $\Delta t$ must be short enough to ensure that
the observed $q_{\alpha(\beta)}$ fluctuations are too small to lead to significant changes in the elements of $\mathbf{D}$. More specifically: ($i=1,2$)
\begin{eqnarray}
2D^\mathrm{s}_\mathrm{L}\Delta t &=& \langle (\Delta x_i)^2\rangle \nonumber \\
2D^\mathrm{c}_\mathrm{L} \Delta t &=& \langle\Delta x_1\Delta x_2 \rangle \nonumber \\
2D^\mathrm{s}_\mathrm{T}\Delta t &=& \langle (\Delta y_i)^2\rangle \nonumber \\
2D^\mathrm{c}_\mathrm{T}\Delta t &=& \langle\Delta y_1\Delta y_2 \rangle\nonumber \\
2D^\mathrm{s}_\mathrm{RT}\Delta t &=& \langle \Delta y_1 \Delta \theta_1\rangle = -\langle \Delta y_2 \Delta \theta_2\rangle\nonumber \\
2D^\mathrm{c}_\mathrm{RT}\Delta t &=& \langle\Delta y_1\Delta \theta_2 \rangle = -\langle\Delta y_2\Delta \theta_1 \rangle\nonumber \\
2D^\mathrm{s}_\mathrm{R}\Delta t &=& \langle (\Delta \theta_i)^2 \rangle\nonumber \\
2D^\mathrm{c}_\mathrm{R}\Delta t &=& \langle\Delta \theta_1\Delta \theta_2 \rangle. \label{eq:fluctuations}
\end{eqnarray}
This provides a possible experimental route toward measuring the elements of $\mathbf{D}$ that does
not involve directly perturbing the system \cite{Haim09}.
\section{Regularized Stokeslets Method for a Many-Body System}
\label{sec:rs-method}
In previous work \cite{camley13}, two of us introduced a Regularized Stokeslets \cite{cortez01,cortez05} (RS)
numerical scheme for calculating the diffusion and resistance matrices for a single
solid body embedded within a membrane. The present paper extends this method
to calculations involving multiple solid bodies. This section briefly reviews the
membrane RS method and elaborates on those aspects of the calculation necessary
to consider multiple bodies. For details on the numerical implementation, the reader
is referred to our original treatment \cite{camley13}.
The starting point of the RS
method is a discretized version of Eq. \ref{eq:v-field}
\begin{equation}
\label{eq:rs}
\mathbf{v}[\mathbf{R}_{m}] =\sum_{n=1}^N \mathbf{T}(\mathbf{R}_{m}-\mathbf{R}_{n};\epsilon)
\,\mathbf{g}[\mathbf{R}_{n}]\,.
\end{equation}
This equation results from inserting
\begin{equation}
\mathbf{f}(\mathbf{r}') = \sum_{n=1}^{N} \mathbf{g}[\mathbf{R}_{n}] \phi_{\epsilon}(\mathbf{r}' - \mathbf{R}_{n})
\end{equation}
into Eq. \ref{eq:v-field} and defining the ``Regularized Stokeslet" as
\begin{equation}
\label{1}
\mathbf{T}(\mathbf{r};\epsilon) = \int \mathrm{d} \mathbf{r}'\,
\mathbf{T} (\mathbf{r}-\mathbf{r}') \,\phi_\epsilon(\mathbf{r}')\,.
\end{equation}
The ``blob function" $\phi_\epsilon({\mathbf{r}})$ represents a localized envelope over which
forces $\mathbf{g}[\mathbf{R}_{n}]$, at discrete spatial positions, $\mathbf{R}_{n}$, are transmitted to the fluid.
The blob function must be centered at zero, integrate to 1 and, as $\epsilon \to 0$,
must limit to the Dirac delta function $\delta(\mathbf{r})$. We employ
a Gaussian blob function \cite{camley13},
$\phi_\epsilon(\mathbf{r})=\frac{1}{2\pi\epsilon^2} e^{-r^2/(2\epsilon^2)}$ in all of our calculations, but other
choices are certainly possible. The numerical evaluation of $\mathbf{T}(\mathbf{r};\epsilon)$
for our quasi-2D Saffman-Delbr\"uck systems is described in detail in Ref.
\onlinecite{camley13}.
It should be clear that Eq. \ref{eq:rs} is a simple matrix equation. If we restrict ourselves to only
calculating velocities over the same set of positions where we impose forces, i.e. $\mathbf{R}_{m} \in \{ \mathbf{R}_{n} \}$,
the relevant matrix is square. We choose the blob positions, $\{\mathbf{R}_{n}\}$, to tile all particles embedded in
the fluid (see Fig. \ref{fig:circle-blob}) and solve Eq. \ref{eq:rs} to determine the forces, $\{\mathbf{g}[\mathbf{R}_{n}]\}$,
necessary to effect a specified set of velocities, $\{\mathbf{v}[\mathbf{R}_{m}] \}$. {\em The velocities are always chosen
to describe rigid body motions within each particle} so that $\mathbf{v}[\mathbf{R}_{m_i}]=\mathbf{V}_i+\mathbf{\Omega}_i\times ( \mathbf{R}_{m_i} - \mathbf{R}_{c_i})$ for
each particle $i$. Here, the subscript $m_i$ is used to index the subset of blobs associated with particle $i$ and $\mathbf{R}_{c_i}$ is the hydrodynamic center of particle $i$.
We note that blobs are chosen to cover the entire disk,
rather than just the perimeter. This is because, as discussed in Ref.\cite{camley13}
mandating rigid body motion of the perimeter does not ensure rigid
body motion in the disk interior. In practice, the generalized minimal residual method algorithm (GMRES) \cite{gmres}
is used to solve Eq. \ref{eq:rs} for the blob forces, $\{\mathbf{g}[\mathbf{R}_{n}]\}$.
The blob forces combine to yield the total forces and torques on the individual
particles,
\begin{eqnarray}
\label{eq:force-torque-rs}
\mathbf{F}_i &=& \sum_{m_i} \mathbf{g}[\mathbf{R}_{m_i}]\, \nonumber \\
\mathbf{\tau}_i &=& \sum_{m_i} (\mathbf{R}_{m_i}- \mathbf{R}_{c_i})\times \mathbf{g}[\mathbf{R}_{m_i}]\,.
\end{eqnarray}
Therefore, the RS method provides a route to determining the F-vector corresponding to a given V-vector, using
the language of the preceding section. The numerical calculations involve a large number of blobs ($N$) to discretize the particles,
but the F-vector and V-vector reflect forces/torques and velocities/angular velocities on the particles and are modestly sized (six elements each for wo
particle calculations). As a practical matter, it is most convenient to calculate the resistance
matrix by choosing the V-vector input to the RS calculation to be composed entirely of zeros, except for a single element.
The F-vector corresponding to this input then corresponds to a complete column of the resistance matrix. This process
is repeated for each element of the V-vector to generate the full resistance matrix, column by column. The diffusion matrix
is then obtained by inversion of the resistance matrix.
For concreteness and to emphasize numerical details we provide a step-by-step procedure for
calculating the diffusion matrix corresponding to the case of two identical disks of radius $r_0$ (see Fig. \ref{fig:topside})
embedded within the membrane.
\begin{figure}
\begin{center}
\includegraphics[scale=0.69]{circblob}
\caption{\label{fig:circle-blob}
Cartoon of the blob distribution on a pair of disks. In the actual calculations, thousands of blobs
are distributed on the disks. }
\end{center}
\end{figure}
\begin{enumerate}
\item Discretize each disk with $N/2$ blobs, see Fig.~\ref{fig:circle-blob}.
Similar to Ref.~\citenum{camley13}, the inter-blob spacing, $s$, is chosen to be between
$0.03-0.09 r_0$. $\epsilon$ is set to be half of the spacing.
\item All the elements of the V-vector are zeroed except for
$v_{1x}\ne0$, {\it i.e.} the disk 1 moves along the x-axis
without rotation and the disk 2 is stationary.
This V-vector is used as input for the RS routine to obtain the forces and torques on the disks (F-vector).
\item The obtained F-vector is divided by $v_{1x}$
to obtain the first column of the resistance matrix.
\item Repeat steps 2 and 3, keeping a single non-zero element of the V-vector each time
to obtain the remaining columns of the resistance matrix. For example, in the fifth run, we set all
the elements of the V-vector to zero except $\Omega_{1}\ne0$, {\it i.e.} disk 1
only rotates and the disk 2 is stationary. The resulting F-vector determines the fifth
column of the friction tensor.
\item Calculate
the diffusion tensor by numerically inverting the friction tensor, see Eq.~\eqref{eq:diff-fric-tensor}.
\item Repeat steps 1-5 for different values of the spacing
$s$. We find that the elements of
the diffusion matrix change linearly in $s$ and this
allows us to extrapolate to the infinite resolution limit $s=0$.
The reported results are obtained by this extrapolation.
\end{enumerate}
Step six of the above procedure provides a natural criterion for
establishing the credibility of our numerics.
The mean squared error of the linear fit in step 6 is readily determined for
a given physical situation (i.e. specification of
$d/r_0$ and $\ell_0/r_0$) by considering $n$ different resolutions: $s_1, s_2, \dots, s_n$ .
If $y_i$ is the numerical value of an element of the diffusion matrix for the resolution $i$ and $y_{fit}(x)$ is the best-fit line
to the points $y_1,y_2,\dots,y_n$, the root mean square error reads
\begin{equation}
\label{eq:error}
\mathrm{RMSE}=\sqrt{(1/n)\sum_{i=1}^{n} \big[y_i-y_{fit}(s_i)\big]^2}\,.
\end{equation}
The RMSE for all the ``nonzero" elements of the diffusion matrices which are
reported in this work are two to three orders of magnitude
smaller than the extrapolated elements themselves.
However, we point out that the data reported in this paper are restricted to separations $h/r_0>0.1$. At smaller separations the RMSE
starts to become large and the linear fit over the data taken at various resolutions is clearly poor;
slight adjustments to the resolution within the range indicated above lead to wild vacillations in the predictions. This is a clear
indication that the RS scheme at computationally practical resolutions is breaking down. So, although the RS scheme is
not able to directly probe the small separation limit, we have a clear numerical marker that this is the case. The numbers reported
in this work are meaningful and we have avoided reporting on any regimes that are problematic. We do note that the problematic
regime is readily captured via lubrication theory results and the RS numerics and lubrication
results share a window of common validity (see App.~\ref{app:lubrication}).
\section{Results for two identical disks}
\label{sec:res}
The preceding sections have detailed the mechanics for calculating the resistance and diffusion
matrices for multiple solid bodies embedded in a membrane. Here, we apply these methods
to the calculation of these matrices for the specific case of two identical disks embedded
within the bilayer. As mentioned previously, this case has the advantage of maximal symmetry
and allows us to explore some consequences of the hydrodynamic calculations while keeping
the set of physical and geometric parameters to a minimum. As we will see, even this simplest of systems
is rather complex to describe.
\subsection{Translational Coupled Diffusion}
\label{subsec:comp-w-haim}
The translational coupling elements of the diffusion matrix quantify correlations
in the thermal Brownian motions of the two disks (see Eq. \ref{eq:fluctuations}).
In the limit of point-like particles, $D_\mathrm{L}^\mathrm{c}$ and $D_\mathrm{T}^\mathrm{c}$, follow immediately
from the Green's function of Eq. \ref{eq:mem_green}. Specifically, if the disks
are separated by distance $d$ along the $x$ axis as diagrammed in fig. \ref{fig:topside},
the limiting forms are provided by
\begin{equation}
\label{eq:mu-c-L}
D_\mathrm{L}^\mathrm{c} \approx k_B T T_{xx}(d \hat{\mathbf{x}}) = \frac{k_BT}{4\eta_\mathrm{m}}\frac{\ell_0}{d} \left[
H_1(d/\ell_0) - Y_1(d/\ell_0) - \frac{2\ell_0}{\pi d}
\right]\,,
\end{equation}
and
\begin{multline}
\label{eq:mu-c-T}
D_\mathrm{T}^\mathrm{c} \approx k_B T T_{yy}(d \hat{\mathbf{x}}) = \frac{k_BT}{4\eta_\mathrm{m} } \left[
H_0(d/\ell_0)-\frac{H_1(d/\ell_0)}{d/\ell_0} \right.\\\left.- \frac{1}{2} \left( Y_0(d/\ell_0) - Y_2(d/\ell_0) \right)
+\frac{2\ell_0^2}{\pi d^2}
\right] \,.
\end{multline}
Direct use of these two expressions within the diffusion matrix, would be the membrane analog
to the Kirkwood approximation for these elements.
Recently \cite{Haim09}, Oppenheimer and Diamant have derived a finite size correction to
eqs. \ref{eq:mu-c-L} and \ref{eq:mu-c-T} in the regime where particle separation is well below
the Saffman Delbruck length (i.e. $d \ll \ell_0$). For disks of radius $r_0 \ll d \ll \ell_0$ they find
\begin{equation}
\label{eq:mu-size-corr}
D^\mathrm{c}_\mathrm{L,T} \approx\frac{k_BT}{4\pi\eta_\mathrm{m}}
\left[ \ln(2\ell_0/d) -\gamma_\mathrm{e} \pm \frac{1}{2} \mp \frac{r_0^2}{d^2}\right]\,,
\end{equation}
to leading order in $r_0/d$, where the upper (lower) signs correspond to the longitudinal (transverse) diffusion and
$\gamma_\mathrm{e}\approx0.5772$ is Euler's constant. (We mention that the
corresponding formula in Ref.\cite{Haim09} contains a sign error. A brief derivation
of Eq. \ref{eq:mu-size-corr} can be found in App.~\ref{app:corr} to justify the result
provided here.) In the limit that $r_0 / d \rightarrow 0$, the last term of Eq. 18 vanishes and the remaining
result is the small $d/\ell_0$ limit of eqs. \ref{eq:mu-c-L} and \ref{eq:mu-c-T}. It is important to emphasize
the limitation of this formula to the $d \ll \ell_0$ regime. We are unaware of any comparable analytical
formula(e) that extend the finite size results beyond this regime.
\begin{figure}
\begin{center}
\includegraphics[scale=0.69]{lsd1000}
\vspace{6mm}
\includegraphics[scale=0.7]{lsdp1}
\caption{\label{fig:HaimShort}
Comparison of the RS data for the coupled-diffusions $D_\mathrm{L}^\mathrm{c}$ and
$D_\mathrm{T}^\mathrm{c}$ with the related analytical expressions, for
the Saffman-Delbr\"uck length {\bf (a)} $\ell_0/r_0=10^3$ and
{\bf (b)} $\ell_0/r_0=0.1$:
The transverse coupled-diffusion $D_\mathrm{T}^\mathrm{c}$
scaled by $4\eta_\mathrm{m}/k_BT$ versus the ratio of the
surface-to-surface distance $h$ and the radius of the disks $r_0$.
The circles are the
RS data for the transverse coupled-diffusion. The blue and dashed lines
represent Eqs.~\eqref{eq:mu-c-T} and \eqref{eq:mu-size-corr}, respectively. As expected, the
dashed line, which includes corrections to account for finite disk size, improves upon the
point particle approximation (blue line).
Inset: the longitudinal coupled-diffusion $D_\mathrm{L}^\mathrm{c}$
scaled by $4\eta_\mathrm{m}/k_BT$
versus the ratio of the surface-to-surface distance $h$ and the disks' radius $r_0$.
The squares are the
RS data for the longitudinal coupled-diffusion. The red lines
represent Eq.~\eqref{eq:mu-c-L}.
}
\end{center}
\end{figure}
Results from RS numerical calculations for $D_\mathrm{L}^\mathrm{c}$ and $D_\mathrm{T}^\mathrm{c}$
are displayed in Fig. ~\ref{fig:HaimShort}. We display two different cases corresponding to the behavior
of small bodies (lipids or proteins, $\ell_0/r_0 =1000$) and large bodies (ten micron scale
lipid domains, $\ell_0/r_0 = 0.1$) in the bilayer. In both cases it is observed that the point-like-particles approximation
performs well until the disks come within several radii of one another. For separations of
$h=8 r_0$ and larger, eqs. ~\eqref{eq:mu-c-L} and \eqref{eq:mu-c-T} are indistinguishable from
the full numerical calculations.
There are clear deviations from the point-particle results for small inter-particle separations. In
the $\ell_0/r_0 \gg 1$ regime these differences are nicely explained by the theory of Oppenheimer
and Diamant (Eq. \ref{eq:mu-size-corr}), though the theory is not quantitatively accurate at very
small separations (as expected). In the regime of particles large compared to $\ell_0$ we lack
an analytical prediction for finite size corrections to eqs. ~\eqref{eq:mu-c-L} and \eqref{eq:mu-c-T}.
Though finite-particle-size effects make negligible contributions to $D_\mathrm{L}^\mathrm{c}$ for the $\ell_0/r_0 = 0.1$,
case, there are large effects seen in $D_\mathrm{T}^\mathrm{c}$. In fact, for
two identical circular domains with $r_0=10$ microns on a membrane with $\ell_0=1$ micron, at $h=1$ micron separation
the prediction of Eq.~\eqref{eq:mu-c-T} for the transverse mobility is off by more than 800\% from the numerical results.
Although it is perhaps unsurprising that finite particle size should play a major role at such separations, it is interesting
to see just how big the effect is, especially in contrast to the related (lack of) effect on the longitudinal motion.
\subsection{Self-Diffusion}
\label{subsec:change-in-mus}
The self-diffusion matrix elements,
$D_\mathrm{L}^\mathrm{s}$, $D_\mathrm{T}^\mathrm{s}$,
are affected by the presence of other proximal solid particles
~\cite{montgomery77,montgomeryDeutch77}, though this
influence requires a finite particle size and is a shorter
ranged hydrodynamic effect as compared to that observed
in the coupling diffusion elements. Indeed, many common theoretical/simulation
tools including the Kirkwood approximation \cite{doi}, Brownian dynamics with
hydrodynamic interactions \cite{mccammon} and related
methodologies neglect all influence of neighboring particles on self-diffusivity.
In this subsection, we present RS predictions for how
disk self diffusion is impacted by the presence of a second nearby disk. To this end, we compare the self-diffusion of a disk in the two disk system
with the self-diffusion of an isolated single disk control.
We consider
two different scenarios:
(1) both tracked disk and perturbing disk are freely moving, and (2)
the tracked disk is free, but the perturber is fixed in place.
We report
the relative change in the self-diffusions compared to the
single disk case as
\begin{equation}
\label{eq:rel-err}
\Delta D^\mathrm{s}_\mathrm{x} = \frac{D^0_\mathrm{x}-D^\mathrm{s}_\mathrm{x}}{D^0_\mathrm{x}}\,,
\end{equation}
where $D^0_\mathrm{x}$ is the (control) self-diffusion of an isolated single disk,
and $D^\mathrm{s}_\mathrm{x}$ is the self-diffusion of a disk
in the two disk system. x= L, T and R, for the
longitudinal, transverse and rotational mobilities.
\subsubsection{Both disks freely moving}
\label{sec:both_moving_self}
The relative change in longitudinal self-diffusion as a function of
$r_0/\ell_0$ is depicted in Fig.~\ref{fig:self-both-moving}a. The
plots display results for three different disk geometries corresponding to
$h/r_0 = 0.1, 0.5, 1.0$ swept over a wide range of Saffman-Delbr\"uck lengths.
For a given choice of $h/r_0$, any particular choice of $r_0$ (or $h$) yields an identical two-disk geometry
upon proper scaling and it is most useful to regard changes along horizontal axis ($r_0/\ell_0$) in this figure
as resulting from altering the membrane $\ell_0$ while maintaining a given two-disk geometry.
The first point to note about Fig. \ref{fig:self-both-moving}a is that the cases plotted cover a significantly
abbreviated range of values for $h/r_0$ relative to the cases considered in the context of coupled diffusion. The passive
effect of the second disk on the motion of the first is a considerably shorter ranged phenomenon than the
direct transmission of force from disk one to disk two at play in the coupling diffusion coefficients. At edge-to-edge
separations equal to one disk radius, the strongest observed influence on self diffusion is only on the order of 5\% and
shrinks still further for larger separations. Only for very closely placed disks does the effect become appreciable. We
also note that the extreme values of $\ell_0$ occurring at the edges of the plot place the system in either a limiting 3D hydrodynamic regime independent of $\ell_0$
(at small $\ell_0$) or the regularized 2D limit of the Saffman-Delbr\"uck model (at large $\ell_0$) with weak (logarithmic \cite{saf}, see Eq. \ref{eq:mu-size-corr})
dependence on $\ell_0$.
The relative change in the transverse self-diffusion $\Delta D_\mathrm{T}^\mathrm{s}$
remains very small at all separations, although it
displays the same trends as in Fig.~\ref{fig:self-both-moving}a.
The maximum relative change in the transverse self-diffusion
for the separation ratio $h/r_0=0.1$ is less than 5\% and we have
not explicitly included $\Delta D_\mathrm{T}^\mathrm{s}$ plots.
The most striking feature of Fig. \ref{fig:self-both-moving} is the
non-monotonic behavior for the longitudinal translation case, with
a maximum for disks of radius comparable to the
Saffman-Delbr\"uck length, {\it i.e.} $r_0 \sim \ell_0$. Although
this feature draws the eye, we do not believe it to be as
interesting as might first be believed. We recall that the
quantities plotted in Fig. \ref{fig:self-both-moving} are
{\em relative} differences in self-diffusion, with the difference itself
plotted in ratio to the corresponding self-diffusion of an isolated
single-disk control. Both the numerator and denominator of the ratio
display strong, but clearly monotonic dependence on $r_0 / \ell_0$ in the vicinity
of $r_0 / \ell_0$ (see Fig. \ref{fig:why-pick}) . The observed maximum simply results
from the product of two functions, one sharply increasing and the other sharply decreasing.
The appearance of a maximum can be qualitatively reproduced using less elaborate
numerical models, as is explicitly demonstrated in App.~\ref{app:dim-mon} using a Kirkwood approximation
for a simplified system.
\begin{figure}
\begin{center}
\includegraphics[scale=0.74]{DLmov}
\vspace{6mm}
\includegraphics[scale=0.73]{DRmov}
\caption{\label{fig:self-both-moving}
Two disks free to move and rotate: the relative change in the
(a) longitudinal and (b) rotational self-diffusion versus
the ratio of the radius of the disks $r_0$ and the Saffman-Delbr\"uck length $\ell_0$.
In both figures (a) and (b) the squares, triangles and circles represent
the separations $h/r_0=0.1,~0.5$ and 1, respectively. The percentage changes
to $\Delta D_\mathrm{T}^\mathrm{s}$ are all less than 5\% for the cases considered here
and are not explicitly plotted.
}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=0.75]{difinv}
\caption{\label{fig:why-pick}
Two disks free to move and rotate: Diamonds show the dimensionless difference between the
longitudinal self-diffusions of a disk in the presence of another one $(h/r_0=0.1)$ and a single disk
$(D^0_\mathrm{L}-D^\mathrm{s}_\mathrm{L})/D^0_\mathrm{L}(1)$ versus the ratio of the
radius of the disks and the Saffman-Delbr\"uck length $r_0/\ell_0$.
Down triangles represent the dimensionless inverse of the longitudinal self-diffusion of the single disk $D^0_\mathrm{L}(1)/D^0_\mathrm{L}$
versus $r_0/\ell_0$. $D^0_\mathrm{L}(1)$ is the longitudinal self-diffusion coefficient of
the single disk for $\ell_0/r_0=1$.
Inset: Short and intermediate distance close-up of $(D^0_\mathrm{L}-D^\mathrm{s}_\mathrm{L})/D^0_\mathrm{L}(1)$ on the linear scale.
}
\end{center}
\end{figure}
The relative changes in rotational self-diffusion are presented in Fig.~\ref{fig:self-both-moving}b.
The trends of $\Delta D_\mathrm{R}^\mathrm{s}$ are completely different than the $\Delta D_\mathrm{L}^\mathrm{s}$ case,
as might have been anticipated due to the shorter ranged hydrodynamic flows associated with rotational motion
relative to translational motion. Here, the larger effects are seen in the 2D hydrodynamic regime relative to the
3D regime and the crossover is completely monotonic. Unlike translational diffusion, rotational diffusion does not
suffer from a Stokes paradox in 2D and requires no regularization by a surrounding subphase to obtain finite results \cite{saf}; the large
$\ell_0$ limit for rotational diffusion saturates to a true 2D limit without any residual logarithmic corrections.
In reconciling the
somewhat backward trends between panes (a) and (b) of Fig. \ref{fig:self-both-moving} it is helpful to recognize that the effects
described in this figure result from the perturbation to single disk motion caused by constraining a nearby region of the membrane to undergo
only rigid-body motion. We expect that the largest changes to self diffusion will occur when the unperturbed flows around a moving isolated disk 1 deviate
most strongly from rigid body flows at the location of disk 2. In the limit of large $\ell_0$ in Fig. \ref{fig:self-both-moving}a the velocity field proximal to a single translating disk is very nearly constant.
The constraint of rigid body motion over the envelope of disk 2 is hardly a constraint at all and has minimal effects on the translational self diffusion. Rotational motions,
on the other hand, generate considerably shorter ranged flows which vary appreciably proximal to the rotating disk and a strong effect is seen
in the large $\ell_0$ limit of Fig. \ref{fig:self-both-moving}b. In the small $\ell_0$ (3D) regime the translational flows are not as flat as in the opposite regime,
which explains the relatively larger changes on the right of Fig. \ref{fig:self-both-moving} a. The rotational flows are also more rapidly varying in the 3D regime
relative to the 2D regime, and actually die off so rapidly away from the rotating disk that they are quite small in the vicinity of the second disk, explaining the small effects on the right.
of Fig. \ref{fig:self-both-moving} b.
\subsubsection{One Disk Stationary, the Other Free to Move and Rotate}
\label{sec:fixed_disk}
In biological contexts, it is common for proteins and larger assemblies to be immobilized
on the membrane surface due to interactions with cytoskeletal filaments and related cellular
structures \cite{lodish}. It is thus natural to consider the influence of immobilized perturbers
on particle self diffusion in addition to the case considered in the previous section.
The self-diffusion of a disk adjacent to a completely immobile (i.e. no translation and
no rotation) neighboring disk is a distinct hydrodynamic problem from that considered
in the previous section. Indeed, constraining $\mathbf{v}=0$ in a region proximal
to the diffuser is clearly a significantly more severe restriction than only requiring rigid-body
motion over the same region. Within the RS framework, immobility of disk 2 is
enforced by setting $v_{2x}=v_{2y}=\Omega_2=0$ in Eq. \ref{eq:2-obj-fric-tensor-general}.
Further, we only require the $f_{1x}$, $f_{1y}$ and $\tau_1$ components of the F-vector
to determine the mobilities of disk 1 as the forces on the immobilized disk are irrelevant to us.
Eq.\ref{eq:2-obj-fric-tensor-general} is thus reduced to an effective single particle resistance
problem
\begin{eqnarray}
\label{eq:1-obj-fric-tensor-general}
\left(
\begin{array}{c}
f_{1x} \\ f_{1y} \\ \tau_1
\end{array}
\right)
&=&
\left(
\begin{array}{ccc}
\zeta^\mathrm{11}_\mathrm{xx} & \zeta^\mathrm{11}_\mathrm{xy} & \zeta^\mathrm{11}_\mathrm{xR} \\
\zeta^\mathrm{11}_\mathrm{yx} & \zeta^\mathrm{11}_\mathrm{yy} & \zeta^\mathrm{11}_\mathrm{yR} \\
\zeta^\mathrm{11}_\mathrm{xR} & \zeta^\mathrm{11}_\mathrm{yR} & \zeta^\mathrm{11}_\mathrm{R}
\end{array}
\right)
\left(
\begin{array}{c}
v_{1x} \\ v_{1y} \\ \Omega_1 \end{array}
\right) \nonumber \\
& = & \left(
\begin{array}{ccc}
\zeta^s_L & 0 & 0 \\
0 & \zeta^s_T & \zeta^s_{RT} \\
0 & \zeta^s_{RT} & \zeta^s_{R}
\end{array}
\right)
\left(
\begin{array}{c}
v_{1x} \\ v_{1y} \\ \Omega_1 \end{array}
\right)
\, .
\end{eqnarray}
The first of these equations corresponds to the case of two general objects with object 2 immobilized;
the second equation is specific to the symmetries present in the two disk case with the second disk immobilized. We stress that the individual elements of
the above matrices are identical to those appearing in Eqs. \ref{eq:2-obj-fric-tensor-general} and \ref{eq:circ-r-mat}
and need to be calculated via the full 2 body RS scheme described in section \ref{sec:rs-method}. No new
RS calculations are needed here, beyond those made to calculate the elements of the general two-body resistance
matrix. The effect of second body immobilization on the self-diffusion of body one is captured through the different
matrix inversion required in this case relative to the case considered in the previous section and as indicated in Eq. \ref{eq:diff-fric-tensor}.
The self-diffusivities of particle 1 in the present case are obtained by inverting the truncated $3\times 3$ resistance matrix
of Eq. \ref{eq:1-obj-fric-tensor-general} and not the full $6 \times 6$ resistance matrix.
As expected, compared to a mobile adjacent disk, a stationary disk exerts a considerably stronger
influence on translational self-diffusions
,$\Delta D^\mathrm{s}_\mathrm{L,T}$ ( see Fig.~\ref{fig:self-one-stat}a).
The effect is particularly strong in the 2D regime of large $\ell_0$. Although
enforcing rigid body motion over the perturbing disk has little effect due to the
nearly homogeneous velocity field around the diffuser at large $\ell_0$, forcing complete immobility
of the perturbing disk has a huge effect and can slow diffusion by nearly a factor
of two. This is true for both longitudinal and transverse motions. In the 3D regime,
flows die off more rapidly away from the diffuser and the perturbation by disk 2 is smaller than
in the 2D regime. However, the effect of an immobile perturber is always stronger than a mobile
one; forcing immobility on the fluid flow in a region in space is a stronger perturbation than just
requiring rigid body motion over the same region.
The relative change in the rotational self-diffusion, $\Delta D_\mathrm{R}^\mathrm{s}$,
is presented in Fig.~\ref{fig:self-one-stat}b.
Comparing Figs.~\ref{fig:self-both-moving}b and ~\ref{fig:self-one-stat}b,
the rotational self-diffusion is affected by the immobile disk
between 5 \% to 10\% more strongly than the free disks case, but the trends
are similar.
\begin{figure}
\begin{center}
\includegraphics[scale=0.63]{DLTst}
\vspace{5mm}
\includegraphics[scale=0.64]{DRst}
\caption{\label{fig:self-one-stat}
An immobile and a free disk (the free disk is free to move and rotate):
The relative change in the (a) transverse and (b) rotational self-mobilities
versus the ratio of the radius of the disks $r_0$ and the Saffman-Delbr\"uck length $\ell_0$.
In both figures (a) and (b) the squares, triangles and circles represent
the separations $h/r_0=0.1,~0.5$ and 1, respectively.
Inset: The relative change in the longitudinal self-diffusion.
}
\end{center}
\end{figure}
\section{Linked Dimers: Rotating vs Non-Rotating Monomers}
\label{sec:dimers}
Inspired by the recent experiments of Knight and coworkers
\cite{knight10}, we consider the motion of two-disk dimers (Fig. \ref{fig:dimerss}) on the
membrane surface.
The experimental systems involve constructs of two lipid associated proteins bound
together by molecular (peptide) linkers of various lengths. (Trimers were also studied
in the experiments, but are not considered here.) In a previous study \cite{camley13},
we treated these dimers as a completely rigid body and modeled the translational diffusion
for direct comparison to experiment. We did not consider rotational diffusion in that work because rotation
was not tracked experimentally and we had not developed the theoretical methodology to treat any
motions beyond rigid body translations and rotations. Given the flexible nature of the peptide linkers
used in experiment, the assumption of a completely rigid dimer is questionable and might naively
be expected to be particularly problematic for rotational motions. Requiring the individual
monomers of the dimer to rotate with the same angular velocity as the whole assembly is a significant
constraint. The purpose of this section is to evaluate how relaxing this constraint affects overall dimer rotation.
As we have seen in Sec. \ref{sec:fixed_disk}, the calculation of particle mobilities/diffusivities
under the constraint that certain motions are disallowed is straightforward.
Once the elements of the resistance matrix have been calculated for the general case absent constraints,
the hard work has been done. Imposing the constraint(s) simply means choosing input V-vectors consistent
with the constraint(s) and only tracking those component(s) of the F-vector orthogonal to constrained motions.
The reduced resistance matrix obtained in this manner may then be inverted to obtain the diffusion matrix for
the unconstrained motions. This procedure was particularly simple for the example of a completely immobile
monomer studied in Sec. \ref{sec:fixed_disk}, since our starting basis involved single disk motions and the
reduced basis was obtained simply by zeroing out half of the V-vector corresponding to motions of disk 2.
Similarly, the forces/torque associated with disk 2 were ignored, which was also completely natural within
the single disk basis.
The present case involving dimer motions is slightly more complicated because our resistance matrix elements
have been calculated in the single disk basis, but we need to impose constraints that involve both disks. These constraints
are handled much more naturally in a basis of center-of-mass and relative motions of the two disks. To this end,
we consider a V-vector
\begin{equation}
\label{eq:rigid-dimer-Uv}
\left(
\begin{array}{c}
v_{\mathrm{cm},x} \\
v_{\mathrm{cm},y} \\
\Omega_{\mathrm{cm}} \\
\Delta\Omega_1 \\
\Delta\Omega_2 \\
v_{\mathrm{rel}}
\end{array}
\right)
=
\mathbf{T}\,\cdot
\left(
\begin{array}{c}
v_{1x} \\
v_{1y} \\
v_{2x} \\
v_{2y} \\
\Omega_1 \\
\Omega_2
\end{array}
\right)
\end{equation}
defined by the transformation matrix
\begin{equation}
\label{eq:Uv-rigid}
\mathbf{T}=\left(
\begin{array}{cccccc}
\frac{1}{2} & 0 & \frac{1}{2} & 0 & 0 & 0 \\
0 & \frac{1}{2} & 0 & \frac{1}{2} & 0 & 0 \\
0 & -\frac{1}{d} & 0 & \frac{1}{d} & 0 & 0 \\
0 & \frac{1}{d} & 0 & -\frac{1}{d} & 1 & 0 \\
0 & \frac{1}{d} & 0 & -\frac{1}{d} & 0 & 1 \\
-1 & 0 & 1 & 0 & 0 & 0\\
\end{array}
\right).
\end{equation}
\begin{figure}
\begin{center}
\includegraphics[scale=0.45]{dimers}
\caption{\label{fig:dimerss}
Two disk dimers are derived from the general two disk systems of the previous sections under the constraint of constant separation
between the disks. ``Rigid dimers" are further constrained to treat the dimer assembly as a single rigid body. ``Dimers with rotating monomers''
allow the monomers to freely spin with angular velocities that are not slaved to the overall rotation of the dimer about its center of mass. }
\end{center}
\end{figure}
$v_{\mathrm{cm},x}$, $v_{\mathrm{cm},y}$ and $\Omega_{\mathrm{cm}}$ are the translational and angular velocities of the
dimer relative to its center of mass. $v_{\mathrm{rel}}$ is the relative velocity of the monomers along the axis joining them and
$\Delta \Omega_{1,2}$ are the angular velocities of the individual disks measured relative to $\Omega_{\mathrm{cm}}$. The
dimers we consider will always have a fixed separation and $v_{\mathrm{rel}}=0$. Additionally, ``rigid dimers" are defined
by $\Delta \Omega_{1,2}=0$, whereas dimers with freely rotating monomers impose no such constraint; see Fig. \ref{fig:dimerss}.
If we desire to preserve a similar expression of the fluctuation-dissipation relationship in our new basis as seen in the original
two-particle basis (i.e. Eq. \ref{eq:fluctuations}), it is essential to define the elements of our transformed F-vector as
\begin{equation}
\label{eq:rigid-dimer-Uf}
\left(
\begin{array}{c}
f_{\mathrm{cm},x} \\
f_{\mathrm{cm},y} \\
\tau_\mathrm{cm}\\
\Delta\tau_1 \\
\Delta\tau_2 \\
f_{rel} \\
\end{array}
\right)
=(\mathbf{T}^\intercal)^{-1}\, \cdot
\left(
\begin{array}{c}
f_{1x} \\
f_{1y} \\
f_{2x} \\
f_{2y} \\
\tau_1 \\
\tau_2
\end{array}
\right)
\end{equation}
with the inverse of the transpose of $\mathbf{T}$ given explicitly by
\begin{equation}
\label{eq:Uf-rigid}
(\mathbf{T}^{-1})^{\intercal}=(\mathbf{T}^\intercal)^{-1}=\left(
\begin{array}{cccccc}
1 & 0 & 1 & 0 & 0 & 0 \\
0 & 1 & 0 & 1 & 0 & 0 \\
0 & -\frac{d}{2} & 0 & \frac{d}{2} & 1 & 1 \\
0 & 0& 0 & 0 & 1 & 0 \\
0 &0 & 0 & 0 & 0 & 1 \\
-\frac{1}{2} & 0 &\frac{1}{2} & 0 & 0 & 0 \\
\end{array}
\right)\,.
\end{equation}
It then follows that the resistance and mobility matrices in our new basis are related
to their counterparts in the two-particle basis as
\begin{eqnarray}
\boldsymbol{\zeta}' & = &(\mathbf{T}^\intercal)^{-1} \boldsymbol{\zeta} \mathbf{T}^{-1} \nonumber \\
\mathbf{D}' &=& \mathbf{T} \mathbf{D} \mathbf{T}^\intercal. \label{eq:transformed_matrices}
\end{eqnarray}
Further, the matrices obey the expected inverse relation to one another, $\mathbf{D}' = k_B T \boldsymbol{\zeta}'^{-1}$
and the elements of $\mathbf{D}'$ are simply related to the transformed coordinate displacements in a short time interval in analogy to Eq.
\ref{eq:fluctuations}: $2D'_{\alpha \beta} \Delta t = \langle\Delta q'_\alpha \Delta q'_\beta \rangle$.
We calculate the diffusion coefficients for the rigid dimer by considering only the diagonal upper left $3 \times 3$ corner of the $\boldsymbol{\zeta}'$
matrix defined by Eq. \ref{eq:transformed_matrices} and inverting this reduced resistance matrix to obtain a diagonal $3 \times 3$ diffusion matrix. The
resulting diffusion coefficients $D_L$ (longitudinal translational motion along the dimer axis), $D_T$ (transverse translational motion perpendicular to the
dimer axis) and $D_{rot}$ (rotational motion of the dimer) have the explicit forms
\begin{eqnarray}
\label{eq:rigid-dimer-diffusion}
D_\mathrm{L}/(k_BT) &=& \frac{1}{2(\zeta^\mathrm{s}_\mathrm{xx}+\zeta^\mathrm{c}_\mathrm{xx})} \,\,\,\, \mbox{(rigid dimer results)} \nonumber\\
D_\mathrm{T}/(k_BT) &=& \frac{1}{2(\zeta^\mathrm{s}_\mathrm{yy}+\zeta^\mathrm{c}_\mathrm{yy})} \\
D_\mathrm{R}/(k_BT)&=&\frac{1}
{2(\zeta^\mathrm{s}_\mathrm{r}+\zeta^\mathrm{c}_\mathrm{r})
-2 d (\zeta^\mathrm{s}_\mathrm{rt}+\zeta^\mathrm{c}_\mathrm{rt})
+ \frac{d^2}{2} (\zeta^\mathrm{s}_\mathrm{yy}-\zeta^\mathrm{c}_\mathrm{yy})} \nonumber \,.
\end{eqnarray}
It was verified that results obtained in this manner are numerically identical to direct RS calculations treating
the entire dimer as a single rigid body (as in Ref. \cite{camley13}). The corresponding quantities for the dimer with rotating monomers
are calculated by considering the upper left $5 \times 5$ corner of $\boldsymbol{\zeta}'$
and performing the $5 \times 5$ inversion to obtain the diffusion matrix. This matrix is not diagonal, however the first three elements
along the diagonal still correspond to mean square displacements and rotation of the dimer about the center of mass. The results are
\begin{eqnarray}
\label{eq:mobil-non-rigid-mobil}
D_\mathrm{L}/(k_BT) &=& \frac{1}{2(\zeta_\mathrm{xx}^\mathrm{c}+\zeta_\mathrm{xx}^\mathrm{s})} \,\,\,\, \mbox{(rotating monomers results)} \nonumber \\
D_\mathrm{T}/(k_BT)&=& \frac{\zeta_\mathrm{r}^\mathrm{c}-\zeta_\mathrm{r}^\mathrm{s}}
{2\big((\zeta_\mathrm{rt}^\mathrm{c}-\zeta_\mathrm{rt}^\mathrm{s})^2+
(\zeta_\mathrm{r}^\mathrm{c}-\zeta_\mathrm{r}^\mathrm{s})
(\zeta_\mathrm{yy}^\mathrm{c}+\zeta_\mathrm{yy}^\mathrm{s})
\big)} \,, \\
D_\mathrm{R}/(k_BT)&=& \frac{2(\zeta_\mathrm{r}^\mathrm{c}+\zeta_\mathrm{r}^\mathrm{s})}
{d^2\big(
(\zeta_\mathrm{r}^\mathrm{c}+\zeta_\mathrm{r}^\mathrm{s})
(\zeta_\mathrm{yy}^\mathrm{s}-\zeta_\mathrm{yy}^\mathrm{c})
-(\zeta_\mathrm{rt}^\mathrm{c}+\zeta_\mathrm{rt}^\mathrm{s})^2
\big)} \,.\nonumber
\end{eqnarray}
Our interest is in the comparison between rotational diffusion for the two different dimer assemblies.
To this end, we compare the rotational diffusion of a rigid dimer with fixed monomers $\widetilde{D}_\mathrm{R}^\mathrm{fixed}$ and
a dimer with rotating monomers $\widetilde{D}_\mathrm{R}^\mathrm{rot}$ via the relative difference in the rotational self-diffusions as
\begin{equation}
\label{eq:rel-dif-drot}
\Delta \widetilde{D}_\mathrm{R}=\frac{\widetilde{D}^\mathrm{rot}_\mathrm{R}-\widetilde{D}^\mathrm{fixed}_\mathrm{R}}{\widetilde{D}^\mathrm{rot}_\mathrm{R}}\,.
\end{equation}
\begin{figure}
\begin{center}
\includegraphics[scale=0.9]{dimer}
\caption{\label{fig:dimer}
Relative difference between rotational self-diffusions for dimers with and without freely rotating monomers
(calculated by Eq.~\eqref{eq:rel-dif-drot}) versus the ratio of the
distance of the closest approach between the monomers $h$ and the radius of the monomers $r_0$. The upper and lower plots are
for the Saffman-Delbr\"uck lengths $\ell_0/r_0 = 10^3$ and 0.1, respectively.
The triangles correspond to the full calculation including all
hydrodynamic interactions between the monomers. The dashed curves represent the
results obtained by neglecting monomer-monomer hydrodynamic interactions (see text).}
\end{center}
\end{figure}
Figure~\ref{fig:dimer} plots $\Delta \widetilde{D}_\mathrm{R}$
as a function of the dimer separation for two Saffman-Delbr\"uck lengths:
$\ell_0/r_0= 10^3$ (protein-sized monomers) and 0.1 (domain-sized monomers).
The results for the two regimes are rather similar, with only modest differences between
the two models. However, it is interesting to note that this similarity between the two regimes
as well as the small magnitudes seen in the plots both result from particle-particle hydrodynamic
interactions damping otherwise more significant effects. The dashed lines in fig. \ref{fig:dimer}
result from approximating the full calculation by turning off all hydrodynamic interactions between
the monomers (i.e. by zeroing all off-diagonal contributions to the resistance matrix and calculating
all diagonal elements in the absence of the second disk in Eq. \ref{eq:circ-r-mat} prior to transforming
to $\boldsymbol{\zeta}'$). In this limit, Eq. \ref{eq:rel-dif-drot} takes the simple form
$\Delta \widetilde{D}_\mathrm{R} = 1/(1 + \frac{d^2 \zeta_{xx}^{s0}}{4 \zeta_{r}^{s0}})$, where the ``$s0$"
superscript indicates the self resistances are calculated for an isolated single particle.
It is expected that $\Delta \widetilde{D}_\mathrm{R}$ is a positive quantity. Constraining the monomers
to rotate with the overall rotation of the dimer can only lead to additional dissipation beyond what the
unconstrained system would experience under a prescribed angular velocity. As the dashed lines in fig. \ref{fig:dimer}
indicate, this effect is potentially large (especially for protein sized monomers) when calculated in
the absence of direct monomer-monomer hydrodynamic coupling. It is expensive to force the monomers
to rotate. However, this cost is substantially reduced when the full calculation is performed. With hydrodynamic
coupling present, the monomers naturally spin in the same direction as the dimer rotations, so forcing the
completely in-phase motion between dimer and monomers is less costly. The effect is especially prominent
for dimers in the 2D hydrodynamic regime, where the hydrodynamic interactions are strong.
\section{An example of noncircular bodies: diamond shaped domains}
\label{sec:diamonds}
Though the preceding examples have been limited to circular disks
to maximize symmetry and simplify analysis, a strength of the RS methodology is the ability to consider bodies of arbitrary shape.
As a specific example, we consider the diamond-shaped
lipid domains studied in Ref. \cite{petrov12}. Previously, we used the RS technique
to study single diamond-shaped domains \cite{camley13}. This section considers the diffusion
of diamond domains in proximity to other diamond shaped domains, which
corresponds to the actual experimental conditions of Ref. \cite{petrov12}.
In these experiments, micron-scale diamond-shaped (with aspect ratio of 1.42, see Fig. \ref{fig:diamonds-blob}) solid lipid domains in a
background fluid lipid phase (with Saffman-Delbr\"uck length $\ell_0 \sim 1 \mbox{$\mu$m}$) were tracked in time to measure the translational and rotational diffusion coefficients
of the domains. Though precise area coverages of the diamonds are not reported for the experiments, the provided images clearly display diamonds in proximity to one another
and it is natural to ask how the diffusive characteristics of a tracked domain are expected to be altered by a neighbors. The results of Sec. \ref{sec:both_moving_self}
for disks suggest that that since the diamond size is close to $\ell_0$ in these experiments, the effect should be nearly as strong as is possible.
To determine the impact of hydrodynamic interactions on the
self-diffusion of the diamonds,
RS calculations were performed for diamond pairs on a
membrane with the Saffman-Delbr\"uck length of 1 micron (see
the discussion in Ref.~\cite{camley13} regarding the estimation of $\ell_0$ in the diamonds experiment).
\begin{figure}
\begin{center}
\includegraphics[scale=0.65]{blobdia}
\caption{\label{fig:diamonds-blob}
Cartoon of the blob distribution for the
(a) side-to-side and (b) tip-to-tip diamonds. In the calculations,
thousands of blobs tile each diamond.
}
\end{center}
\end{figure}
In accordance with typical experimental numbers \cite{petrov12}, the short axis of the
diamonds is chosen to be 2.00\,$\mu m$ and the long axis 2.84\,$\mu m$,
corresponding to the aspect ratio of 1.42 mentioned above. This coincides with
an effective radius $r_0\approx 0.95\mu m$, for direct comparison to the data
in Ref. \cite{petrov12}. (The effective radius is the radius of a hypothetical disk
that shares the same area as the diamond.)
\begin{figure}
\begin{center}
\includegraphics[scale=0.64]{diarel}
\caption{\label{fig:diamond}
Two identical diamonds free to move and rotate: the relative change in the
longitudinal and rotational (inset) self-diffusions versus the separation between the
diamonds, for $\ell_0 = 1\mu m$\,. The aspect ratio of the diamonds
is 1.42 and the shorter diagonal is 2\,$\mu m$
(equivalent to the effective radius of $r_0\approx 0.95\mu m$\,). $d$ is the
separation between diamond centers.
The black diamonds and blue squares respectively represent the data for the
side-to-side and tip-to-tip geometries,
see Fig.~\ref{fig:diamonds-blob}. The circles show the data for the disks with the radius $r_0=0.95~\mu m$
on the same membrane. Note that the shortest separations are excluded for the tip-to-tip geometry as they
correspond to overlapping configurations.
}
\end{center}
\end{figure}
Figure~\ref{fig:diamond} plots the relative change (as defined in Eq. \ref{eq:rel-err}) in the longitudinal and rotational self-diffusions
for side-to-side and tip-to-tip diamond geometries in addition to the corresponding quantities obtained by approximating the diamonds
as effective circles with areas identical to the diamond areas.
The relative change depends both on the separation and orientation of the diamonds. For
separations in the range $2\mu m \lesssim d \lesssim3\mu m$
the relative change is between 5\% to 25\% and dies off rapidly with increasing separation
(especially for rotational motion). Although closely spaced neighboring domains can quantitatively
affect the measured diffusion, the effect is not large and is not strikingly different from what could
be estimated by approximating the domains as circles. In particular, domain-domain
hydrodynamic interactions probably can not explain the apparent experimental inconsistencies
between rotational and translational domain motion discussed at some length in Ref. \cite{camley13}.
\section{Conclusion}
\label{sec:summary}
The RS approach \cite{cortez01,cortez05} is a versatile and powerful tool for studying
particulate flows in traditional 3D hydrodynamic geometries and, more recently, in
the quasi-2D geometries associated with lipid bilayer membranes \cite{camley13}.
This work has further validated the membrane RS methodology of Ref. \cite{camley13} and
demonstrated its utility in application to problems involving multiple membrane-embedded
objects. The applications considered herein involve comparisons to analytical theory and
experimental measurements, but one of the most promising future applications of this
work will be its use in parameterizing and validating approximate simulation schemes for the
dynamics of protein laden membranes and related inhomogeneous biomembrane systems.
Standard techniques for simulating particulate flows in 3D \cite{mccammon,stokesian} start with
a pairwise approximation to the diffusion matrix for the full many body system. Typically, the
Rotne-Prager approximation \cite{rotne_prager} is employed. In the case of membrane systems,
there is no Rotne-Prager analog available. One possibility for developing membrane simulations
analogous to ``Brownian Dynamics with Hydrodynamic Interactions" \cite{mccammon} is to
numerically tabulate the coupling elements of the diffusion matrix as a function of particle
separation to build up a pairwise approximation to the diffusion matrix that can be implemented
in simulations, as has been suggested for 3D systems \cite{wang2013assessing}. The RS calculations presented in this work are a natural method to use for such
a tabulation and could, perhaps, be supplemented with the lubrication results of Ref. \cite{bussel92}
to capture near-field effects, as in Stokesian dynamics \cite{stokesian} simulations in 3D.
An alternate strategy for studying protein laden membranes based on an immersed boundary (IB)
scheme \cite{peskin} has recently been introduced \cite{camley14}. This method does not
require input of a diffusion matrix, but rather tracks fluid flow explicitly and derives hydrodynamic
interactions from the fluid velocity field. However, the treatment of particles in this scheme
is approximate (as in any IB simulation) and has only been verified to agree
with the far-field predictions related to those of Oppenheimer and Diamant \cite{Haim09}. Comparison
between RS results and the membrane IB simulations will be an important future test of
this simulation methodology and may suggest future improvements to this (or other)
biomembrane simulation scheme incorporating hydrodynamic interactions. Work along these
lines is currently being pursued.
\begin{acknowledgments}
F.B. thanks Haim Diamant and Gilad Haran for helpful conversations.
This research was supported in part by a grant from the BSF (Grant No. 2012084 to FB) and
the NSF (Grant No. CHE-1153096 to FB).
\end{acknowledgments}
|
\section{Introduction}
Subterranean termites are tropical social insects having an economic impact in the billions of dollars all over the world each year \cite{su1990economically}. For this reason, their patterns of movement have fomented both practical and theoretical studies in the past. Most of these studies, however, have concentrated in displacements at colony level and inside tunnels leaving the study at the individual level largely overlooked. Also, while animal movement in the context of resource acquisition has received a fair deal of attention \cite{sims2008scaling, viswanathan2008levy, bartumeus2008fractal, reynolds2009levy}, its links to individual interaction and --ultimately-- to sociality still need better focus. Providing that interindividual interactions lie at the very heart of sociality \cite{Miramontes.DeSouza.2008} and that interactivity depends on movement and space \cite{DeSouza.Miramontes.2004}, effective space exploration must play prominent role in social behaviour in general, and in termites in particular \cite{miramontes2014social}.
A common misconception about termite spatial orientation is that physical restrictions imposed by tunnel walls would guide effective space sweeping. While this could be true for termites foraging within tunnels underground, once inside their nest the scenario changes completely. There, galleries merge and split in unpredictable ways usually forming an entangled set of paths and chambers whose layout does not seem to help orientation in most cases. Termites could, then, be guided by specific chemical cues and vibratory signals to effectively find nestmates. But would that be entirely dependent on external clues? Or, as sustained by a long forgotten hypothesis \cite{Jander.Daumer.1974}, would termites orient themselves concatenating external stimuli with an optimized search strategy, as with other animals \cite{boyer2006scale}?
Here we present evidence in favour of such a hypothesis, describing exploratory spatial behaviour in isolated termite workers kept in large containers, free from the constrained movements they experience within tunnels. In this way we were able to assess individual free exploratory behaviour in clueless environments and away from social interactions. We conclude that their searching patterns are compatible with scale-free strategies based on a fractal exploration of space and that these are key to the efficient flow of information between nestmates, thereby providing expressive hints on how self-organization underlies social cohesion \cite{Sumpter.2006}
Our study includes the analysis of anomalous diffusion where the mean squared displacement (MSD) is estimated as a base to identify superdiffusive aspects of termites movements. Within the framework of anomalous diffusion theory \cite{klafter1987stochastic,ramos2004levy }, the scaling exponent of the MSD is related mathematically to a L\'evy probability distribution $P(l)\propto l^{-\mu}$ of the steps lengths $l$, with $1<\mu\leq3$. We evaluate these relationships and compare them versus the results that can be extracted from a Kolmogorov structure functions analysis. We progress into exhibiting that a termite walking is a $1/f$ noise process with power-laws in the correlation function indicating long-memory that is compatible with the fractal structures revealed by an Iterated Function System algorithm. Finally, we use maximum likelyhood estimation (MLE) to show that waiting times with power-law scaling exponent values are also present in the termite exploratory behaviour with scaling exponents values compatible with those predicted by the physics of transport phenomena.
\begin{figure}[!ht]
\centering
\includegraphics[width=100mm]{Figure01.pdf}
\caption{{ \small{Picture of the observation set.} (A) One \emph{Cornitermes cumulans} worker with a painted abdomen was allocated into a circular glass arena (205mm inner dia.). When detected by the video recording system, the individual appears as an image measuring 5$\times$5 pixels representing 4.7 mm$^2$ aprox. In (B) the termite worker is the small black dot at the top of the circular area. A single total trajectory is drawn in (C) showing the typical entangled pattern of individual steps. This particular example contained 35,000 points sampled at 0.5 seconds intervals. Notice that most of the trajectory occurs near the arena border, however inner exploratory excursions are also frequent.}}
\label{fig:arena}
\end{figure}
\section*{Materials and Methods}
\subsection*{Experimental set-up}\hspace{2mm}\textit{Cornitermes cumulans} (Kollar) (Blattaria: Isoptera: Termitidae: Nasutitermitinae) workers were collected from a wild colony at the campus of the Federal University of Vi\c cosa, Minas Gerais, Brazil (20$^\circ$ 45' S, 42$^\circ$ 52' W), on 14 Dec 2010. \emph{Cornitermes} spp. are Neotropical termites occurring in several habitats, including forests, ``cerrados'' (Brazilian savannas), and man-modified habitats. No specific permissions were required for the locations or activities reported in this manuscript. The field studies did not involve endangered or protected species. The study did not involved human participants, specimens or tissue samples, or vertebrate animals, embryos or tissues.
\subsection*{Movement recordings}\hspace{2mm}Experimental arena consisted of a glass Petri dish (\O{} = 205 mm) upside down over a sand-blasted flat glass. To allow evaluation of free-walking behaviour, no obstacle, nestmates or food was present in the arenas. Within the arena, a single termite worker was inserted and its movement was tracked from above, continuously for 5--6 hours, with a closed-circuit video camera (Panasonic WV-BP334 B\&W CCD) equipped with a macro lens (Fujinon YV5$\times$2.7 R4B-SA2L). To allow for full contrast between termites and the background, thereby permitting appropriate video recording, termites were painted with non-toxic water soluble dye \cite{Brunow.etal.2005}. We also performed experiments in containers of diameters 140 mm, 90 mm and 55 mm and found that the size did not significantly altered the results here reported. In total, half a million worth of termite displacements were recorded. We concentrate in reporting the data from the largest containers and choose four representative examples for our analysis and discussion.
\begin{figure}[!h]
\centering
\includegraphics[width=70mm]{Figure02.pdf}
\caption{{ \small{A termite walking trajectory segment as reconstructed in 3D, with a time axis added.} Three walking behaviours are visible. (C) is the individual walking termite following the circular geometry of the container border, (B) is a straight long nearly-ballistic displacement from one container side to the opposite, and (W) is a waiting time.}}
\label{fig:tra}
\end{figure}
Termite trajectories were captured and digitized at a sample rate of one point every 0.5 s with an automatic video-tracking software (EthovisionXT version 8.5.614, Noldus Information Technology). The paths were converted as a series of Cartesian coordinates coupled with respective time records, which allowed posterior calculations of spatial-temporal displacements of termites and their associated parameters, as shown in Fig. \ref{fig:arena}. Recordings included all three displacing behaviours presented by termites in the arena: (i) walking along the border of the arena, (ii) free moving, and (iii) static waiting-times (see Fig.\ref{fig:tra}).
\begin{figure}[!h]
\centering
\includegraphics[width=90mm]{Figure03.pdf}
\caption{{ \small{Examples of time-series containing traveled distances by four different \emph{Cornitermes cumulans} individuals.} The time-series totaled 35,000 data points in (A) and 43,000 in (B, C and D), but a window of 18,000 points is shown for each. Sample rate was one point at every 0.5 seconds.}
\label{fig:distances}}
\end{figure}
\begin{figure}[!h]
\centering
\includegraphics[width=50mm]{Figure04.pdf}
\caption{{\small{Sampling rates.} Different sampling rates give very different results when the length of the steps are plotted as a histogram of their frequency (log-binned). In the plot, three different sampling rates (sr) were exemplified as the time series as captured by the video recording device at each 0.5, 2.0, and 5.0 sec. Note that a region resembling a power-law scaling is only obvious at 2-sec sampling rate.}
\label{fig:histograms}}
\end{figure}
\subsection*{Path definitions} Paths were defined as the sequence of line segments walked by the termite individual every 0.5 seconds inside the Petri dish (Fig. \ref{fig:distances}). Waiting times are the time elapsed when the distance recorded was zero.
Both quantities, distances and waiting times, are heavily dependent on the sampling rate. Under-sampling occurs if the time interval between two samples misses turning points and oversampling happens when two video shots are taken from the same line segment. Avoiding both, under- and over-sampling, is a matter of prime importance. However, it is not always clear how to do it. From Fig. \ref{fig:histograms} it is obvious that different sampling rates would result in different conclusions if an histogram is the only choice made in order to measure potential scaling exponents, coming from hypothetical power-laws. Current trend of model fitting may even led toward selecting a model that excels the statistical criteria but has no phenomenological meaning. Here, we circumvent this, completing a series of alternate data analysis to evince key aspects of the termite walking behaviour while, at the same time, its dynamical richness is exhibited.
\section*{Results}
\subsection*{MSD and anomalous difussion
\begin{figure}[!h]
\centering
\includegraphics[width=80mm]{Figure05.pdf}
\caption{{\small{ Anomalous diffusion.}\emph{Cornitermes cumulans} termites exhibit anomalous diffusion ($\alpha > 1$) in their walking patterns because the mean squared displacement grows faster than it does in the normal diffusion of a Brownian particle (black), where $\alpha=1$. MSD superdiffusive scaling exponent values of four termite workers are $\alpha=1.90$ (purple), $\alpha=1.86$ (green), $\alpha=1.66$ (blue) and $\alpha=1.75$ (red). Notice that the termite MSD scaling separate away from a power-law at values of $\tau > 10$ and beyond, this is common and correspond to the typical diffusive behaviour of truncated motion in confined environments. $D_i$ is the diffusion coefficient of each individual termite}}.
\label{fig:msd}
\end{figure}
A widely-used method to evaluate the diffusive properties of mobile objects is the mean-square displacement (MSD) $\langle\Delta \vec{r}_{\tau}(t)^ 2 \rangle$, where $ \Delta \vec{r}(t)= [\vec{r}(t)-\vec{r}(t+\tau)]$ is the object displacement at time $t$ with a time delay $\tau$. Normal diffusion, as in the case of Brownian motion, occurs when the MSD increases linearly so that $\langle\Delta \vec{r}_{\tau}(t)^ 2 \rangle=2Dt$, where D is the diffusion coefficient of the object. Diffusive regimes where the MSD is not linear in time but proportional to $t^\alpha$ with $\alpha \neq 1$ are known as anomalous diffusion. Furthermore, when $\alpha$ is the diffusion exponent and $\alpha < 1$, the process is called subdiffusive, and if $\alpha > 1$, the process is called superdiffusive.
\begin{figure}[!h]
\centering
\includegraphics[width=90mm]{Figure06.pdf}
\caption{{ \small{Kolmogorov $q$-functions.} Plots shown at the left column depict four examples of Kolmogorov $q$-functions and their power-law scaling (red lines). First eight values of the exponent $q$ were calculated but only four are shown for the sake of clarity. The column at the right depicts the linear scaling of $\zeta(q)$ (red lines), resulting in four $\delta$ slope values: 0.93, 0.87, 0.85 and 0.8. These correspond to L\'evy exponents of values $\mu=2.0$, $\mu=2.14$, $\mu=2.17$ and $\mu=2.25$, respectively. }}
\label{fig:q-functions}
\end{figure}
\emph{Cornitermes cumulans} diffusion in the arena can be estimated and characterized when the value of the MSD is calculated and when, for the sake of clarity, it is compared against a simulated Brownian particle of a known $\alpha=1$. It is easy to see, in Fig. \ref{fig:msd}, that while the Brownian particle has the expected scaling exponent of $\alpha = 1$, all termite's scaling exponents are $\alpha > 1$ and so their movements are properly superdiffusive. Termite step lengths and waiting times statistics (see Figs.\ref{fig:q-functions} and \ref{fig:waiting-times}) have broad distributions corresponding to superdiffusive (L\'evy) process where the mean squared displacement grows faster than it does in the normal diffusion of a Brownian particle \cite{klafler2005anomalous}. Phenomenologically, termite workers exhibit anomalous diffusion in their walking patterns because these individuals remain in motion without changing direction for a time that follows a scale-free distribution. These less-frequent long distance displacements are clearly visible in our experimental arenas (Fig. \ref{fig:tra}).
Physical systems where superdiffusion and L\'evy fligths were first identified together and theoretically related to each other, have been studied in great detail up to now. These systems include the transport of passive particles in non-linear fluids where fluid motion may be retarded at the border of regions having different velocity vectors. Such regions cause ``sticking" that result in long-tailed distributions, both in the distances traveled and the waiting-times \cite{solomon1993observation}. For all practical purposes both superdiffusion and L\'evy fligths are two sides of the same coin, in such a way that a measure of one is related to a measure on the other. Superdiffusion and L\'evy fligths scaling exponents ($\alpha$ and $\mu$) are know to be theoretically related by a simple expression \cite{klafter1987stochastic,solomon1993observation}:
\begin{equation}
\label{relation1}
\mu = 4 - \alpha,
\end{equation}
\noindent that can be used to estimate L\'evy fligths exponents from a MSD measure, as detailed below.
\subsection*{Structure functions and scaling}
A well know method to characterize potential free-scale behaviour is due to Andrey Kolmogorov in the context of turbulent flows \cite{frisch1991global, yu2003structure, Arenas.Chorin.2006}. A generalization of the Kolmogorov method is known as structure functions of order $q$, being expressed as $\langle|\Delta l_{\tau}|^q \rangle\approx \tau^{\zeta(q)}$, \noindent where $l$ is the traveled distance between two ($x$,$y$) points in the space, $\Delta l$ is the distance increments, $\tau$ is a time lag that does the role of spatially coarse-graining the data set at different scales, so that scale-invariant observables could be spotted through the scaling of the exponent function $\zeta(q)$ and, finally, the angled brackets denote an average operator. Since the 1941 Kolmogorov seminal work on turbulence, the $q$-order structure functions have been used extensively in physics \cite{Benzi.etal.1993,Chechkin.Gonchar.2000,padoan2002structure,padoan2003structure,padoan2004structure,chapman2005scaling}, its use in biology for studying scale-free movement patterns in animals is relatively new and mostly applied to the analysis of invertebrates such as copepods \cite{Schmitt.Seuront.2001} and fruit flies \cite{Reynolds.Frye.2007}.
In order to study the free-scale patterns of termite movements under the Kolmogorov framework, we considered the trajectories yield by the walking of representative solitary individuals confined in a Petri dish arena where the values of the re-scaling lag $\tau$ was estimated from
\begin{equation}
\label{funciones2}
|\Delta l_{\tau}|=((x(t+\tau)-x(t))^{2}+((y(t+\tau)-y(t))^2)^{1/2}.
\end{equation}
Then the averages over all increments were calculated and elevated to the proper $q$ power, so that the $q$-structure function is calculated: $\langle|\Delta l_{\tau}|^q \rangle \approx \tau^{\zeta(q)}$
in such a way that $log(\langle|\Delta l_{\tau}|^q \rangle)\approx log(\tau^{\zeta(q)})\approx \zeta(q)log(\tau)$. Notice that the last expression is a linear equation with slope $\zeta(q)$ that can be easily evaluated. The last step needed under this procedure is to realize that $\zeta(q)=\delta q$ and that the value of the slope $\delta$ is related to the {L\'evy} exponent $\mu$ as $\delta=(\mu-1)^{-1}$. It is also important to remark that if $\zeta(q)=\frac{1}{2}q$, the process is normal-diffusive but if $\zeta(q)\neq \frac{1}{2}q$ then the process exhibits anomalous diffusion, with $\zeta(q) > \frac{1}{2}q$ being the superdiffusive regime (a condition fulfilled by all termites examined here).
Equation \ref{funciones2} is used to estimate the structure functions $\langle|\Delta l_{\tau}|^q \rangle$ of our time series for $q$-orders $=\{1,2..,8\}$. Results are plotted in Fig. \ref{fig:q-functions} (left column) where linear scaling is seen on the log-log plot for values of $\tau < 10$, therefore a power-law scaling is evident. The values of these slops are, in turn, plotted in Fig. \ref{fig:q-functions} (right column) These plots evince the linear scaling behaviour of $\zeta(q)$ and so the value of the slopes $\delta=(\mu-1)^{-1}$, where $\mu$ is the L\'evy exponent.
L\'evy flight exponents in the termite superdiffusive walking estimated by the two previous methods are compared in Table \ref{table:levy}. Notice that the mean values for the $\mu$
exponents gives $\langle\mu_1\rangle$$\pm$sd = 2.20$\pm$0.11 (from MSD) and $\langle\mu_2\rangle$$\pm$sd = 2.16$\pm$0.04 (from $q$-functions) that are remarkable similar.
\begin{table}[!ht]
\centering
\caption{\bf {Scaling exponents of four representative examples of termite walking.}
}
\begin{tabular}{|c|c|c|c|c|}
\hline
Series & $\alpha$ (MSD) & $\delta$ & $\mu_1=4-\alpha$ & $\mu_2=(1/\delta) +1 $ \\
\hline
Termite A & 1.90 & 0.93 & 2.10 & 2.08\\
Termite B & 1.88 & 0.87 & 2.12 & 2.15 \\
Termite C & 1.75 & 0.85 & 2.25 & 2.18 \\
Termite D & 1.66 & 0.80 & 2.34 & 2.25 \\
\hline
\end{tabular}
\begin {center}\small{The values of exponents $\alpha$ and $\mu_1$ are related by theoretical results \cite{solomon1993observation}, \\
and are also compatible with the results of a L\'evy exponent $\mu_2$ via \\
a Kolmogorov structure functions analysis.}
\end{center}
\label{table:levy}
\end{table}
\subsection*{Distance temporal fluctuations
The self-similar features of the distances traveled by a termite walking can be visualized by means of a Fast Fourier Transform that converts time fluctuations into a frequency fluctuations domain. This is useful for detecting temporal features such as periodicity or broad frequency spectra. The analysis is also very illustrative because it helps detecting self-similarity when the power spectrum scales as a power-law that, in turn, may hint the presence of colored noise, self-organization or critical dynamics. In addition, and for comparison purposes, it is possible to artificially generate a power-law time-series using well know algorithms in order to exhibit the resemblance of both power spectra. This procedure adds further support in characterizing the termite walking pattern as a fractal process in the traveled distance distributions.
\begin{figure}[!h]
\centering
\includegraphics[width=75mm]{Figure07.pdf}
\caption{\small{Power spectrum of a \emph{Cornitermes cumulans} termite walking time-series (A) and an artificially generated one (B). Both time series contained 4096 points and were transformed with a Fast Fourier Transform (FFT) algorithm.}}
\label{fig:FFT}
\end{figure}
For generating the artificial time-series, a transformation method between random probability distributions is preferred \cite{newman2005power}. The method can generate a random power-law-distributed real number
$x = x_{min}(1 - r)^{(-1/(\mu-1))}$ in the range $x_{min} \leq x < \infty$ with exponent $\mu$, when feed with a random real number $r$ uniformly distributed in the range $0 \leq r < 1$. After the Fourier transform was applied to a 4096 long segment of the termite walking, its power-law scaling $P(f)\propto f^{-\beta}$ was evident (Fig. \ref{fig:FFT}A), scaling with an exponent $\beta=1.3$. This information was used to generate a power-law with the same scaling exponent (Fig. \ref{fig:FFT}B). The resemblance between both spectra is remarkable, adding support for considering the termite walking pattern as a power-law relaxation process.
\begin{figure}[!h]
\centering
\includegraphics[width=110mm]{Figure08.pdf}
\caption{\small{{IFS algorithm.} (A) A \emph{Cornitermes cumulans} termite walking time-series as seen with a IFS algorithm. Notice the subtle details of a self-similar structure. (B) An artificially correlated time-series generated with a relaxation return map $x_{t+1}=\eta x_t + (1-\eta)\epsilon_t$ where $\epsilon_t$ is a normally distributed random variable with zero mean and unit variance and $\eta$ is real valued parameter whose value determines the colour of the resulting time-series scaling. Colour in this context means the classification of a noise mode $1/f^{\beta}$, where $f$ is the frequency in a Fourier transformed space. $\beta$ is the scaling exponent and when $\beta=0$, the process is uncorrelated white noise (C), $\beta=1$ is correlated pink noise (D) and $\beta=2$ is a correlated brown noise (E). Termite walking (A) lies in between a pink (D) and brown noise (E) scaling, being compatible with the fact that the termite scaling exponent in the FFT is $\beta=1.3$. For details on how the IFS algorithm operates, see \cite{miramontes1998intrinsically}.}}
\label{fig:ifs}
\end{figure}
\subsection*{Fractal signatures and IFS
L\'evy flights have spatial self-similarity due to their power-law nature. Because of this, time-series containing L\'evy-distributed random variables can be visualized to evince their subtle structural details, a pattern shared with time-series containing correlated coloured noise that can be generated artificially. An example of this is given in Fig. \ref{fig:ifs} where a fractal recurrence method known as Iterated Function System (IFS), has been used to reveal the detailed self-similarity found in a termite walking.
IFS fractal reconstruction is a method widely used to visualize biological data such as in DNA sequences \cite{jeffrey1990chaos}, insect population dynamics \cite{miramontes1998intrinsically} or the temporal activity of social insects \cite{miramontes2001neural, Miramontes.DeSouza.2008}. Here the method is used for the first time to visualize the self-similar features of an animal walking while performing exploratory behaviour. Fig. \ref{fig:ifs} depicts one of the termite time-series. It reveals in full detail the self-similarity of it, on a set of fractal triangular structures. A similar picture arises when the IFS algorithm is used to visualize a time series of an artificial power-law distribution of a known FFT scaling exponent $\beta=1.3$. The foregoing result is sound evidence in favor of considering the termite walking pattern as a self-similar process with power-law distribution of step lengths.
\subsection*{Step correlations
Successive steps in a \emph{Cornitermes cumulans} termite motion are correlated to a certain extent, as evident from the fact that these resemble correlated pink noise (Fig. \ref{fig:ifs}). In a termite walking, successive steps are no independent from each other with temporal dependencies that decrease along the trajectory. One method to identify and characterize such correlating distances is the correlation function (CF):
\begin{equation}
\centering
|C(\tau)|= \frac{1}{(N-\tau)\sigma^2} \sum_{t=1}^{N-\tau}[l(t)-\bar{\theta}][l(t+\tau)-\bar{\theta}],
\end{equation}
\noindent where $N$ is the time-series length, $\tau$ is a time increment, $\sigma^2$ is the variance and $\bar{\theta}$ is the mean of the steps $l$. Correlation function is useful for finding temporal patterns in the time series such as periodicity or independence. In the case of white noise $|C(\tau)|=0$, so no correlation is present and the successive values are independent. We learn from Fig. \ref{fig:correlation} that this is not the case of termites. On the other hand, it is common that CF decays exponentially with rapid rates as $|C(\tau)|\propto e^{-\tau/\tau_s}$, in which case the time-series have correlations only among short distances $\tau_s$. Non-trivial correlation over long distances usually decay asymptotically as power-laws $|C(\tau)|\propto \tau^{-\gamma}$, with $0< \gamma <1$.
\begin{figure}[!h]
\centering
\includegraphics[width=70mm]{Figure09.pdf}
\caption{\small{Long-range correlations. Termite walking exhibit power-law decaying long-range correlations as measured by a correlation function along the walk time-series (blue). An artificial correlated time-series, as explained in Fig. 8 was used also to compare a correlated decaying process (red). The black line is a power-law with a scaling exponent $\gamma=-0.3$.}}
\label{fig:correlation}
\end{figure}
\emph{Cornitermes cumulans} walking patterns have been found here to have long-range correlations decaying as power-laws. This means that a step length chosen by the termite at a particular time, will influence the length of future steps in a process akin to long-memory. For years physicists have been studying systems that have power-law decaying CFs and have linked them to either systems which are far from equilibrium or very close to a critical point (e.g., a phase transition). We do not know the details of the locomotive physiological mechanism at play that generates a correlated termite walk but we can compare it to an artificially generated time series in order to gain more insights. Consider Fig. \ref{fig:correlation} where the same auto-correlated time-series explored before (Fig. \ref{fig:ifs}) is again analyzed. We now know that this artificial time-series has a power-law decaying CF but has originated by a multiplicative growth process. There are differences worth mentioning. Notice that the termite CF varies softly in contiguous values of $\tau$ while the artificial signal has a CF varying wildly in amplitude along of all its decay, evidencing its stochastic origin.
\subsection*{Waiting times
Foraging patterns of mobile individuals often exhibit resting periods known as waiting times. In an ecological context, it has been speculated that these may result from the encounter of places of interest that need time for being explored and exploited so that observed statistical properties of waiting times will reveal the distribution of resource patch sizes \cite{boyer2006scale}. In a behavioural approach waiting times may be originated in the fact that a mobile organism actually need periods of resting between activities. A number of foraging studies on the statistical properties of waiting times report that these behave as random-like burstiness that appear to be scale-free and so well-approximated by power-laws, meaning that short waiting periods are frequent while long are rare \cite{ramos2004levy}. It has also been recently speculated that the power-law scaling of waiting times is a behavioural ``rule of thumb'' evolved to optimize move-wait decisions in unpredictable environments \cite{wearmouth2014scaling}. Remarkably these waiting times resemble the scaling laws found in the study of superdiffusive passive trackers moving in non-linear flows and where these waiting times are a well-known indivisible aspect of anomalous diffusion \cite{solomon1993observation}. In fact, let's return to equation \ref{relation1} where a relationship between the values of the MSD and L\'evy scaling exponents was established. As a matter of fact, equation \ref{relation1} holds best if $\lambda>2$ \cite{zumofen1995laminar, ramos2004levy}. Since the equation actually holds in our case because the estimated values of $\mu_1$ and $\mu_2$ are consistent (Table 1), we expect that the termite waiting times will be described as power-laws $w(t) \propto t^{-\lambda}$ with $\lambda>2.0.$ We show in Fig. \ref{fig:waiting-times} that \emph{Cornitermes cumulans} worker termites indeed exhibit power-law waiting times when performing exploratory behaviour in navigational clueless environment, where no food or nestmates are present. Waiting times were calculated from the termite walking time series when the isolated termite workers did not changed position. Before walking again a termite may stop and wait for as long as 0.5 seconds to as much as 200 seconds. The log-log plot of the waiting time data, for the example shown, resulted in a negative power-law function with an exponent $\lambda=2.64$; other measured values were 2.28, 2.52 and 2.96, with $\langle\lambda\rangle$$\pm$sd = 2.60$\pm$0.28.
\begin{figure}[!h]
\centering
\includegraphics[width=100mm]{Figure10.pdf}
\caption{\small{Waiting times. Waiting times are an ubiquitous pattern of animal movement behaviour and they may follow power-law scaling, as is the case of \emph{Cornitermes cumulans} workers when performing exploratory behaviour. The graph at the left (A) depicts a typical example of termite spatial distribution of accumulated waiting times over the circular arena (squared root axis for enhancing visualization). The plot in (B) is the waiting-time bouts histogram showing a power-law with a scaling exponent value of $\lambda=2.64$ (straight line slope, calculated with a MLE procedure \cite{Clauset.etal.2009}.)}}
\label{fig:waiting-times}
\end{figure}
\subsection*{Turning angle distribution
Inside the entangled network of tunnels, termites follow well defined routes for traveling between different places of interest, from the foraging areas, to the royal chamber, etc. Here we are interested in exploring to what extent there is persistence in the direction of consecutive steps in termites that would reveal, for example, if termites move exploring space in preferential angles or not. This is because it has been speculated that they may be doing movements with angles of about 40--60 degrees as an optimal searching strategy that may be related to the branching angle of termite tunnels \cite{Bardunias.Su.2009}. When exploring the distribution of turning angles between successive steps, it is found here that \emph{Cornitermes cumulans} workers in open spaces do not show evidence of pre-wired preferential angles. There is a directional trend to move forwards but the resulting probability distribution is a uniform bell shaped curve centered in 0 degrees, as seen in Fig. \ref{fig:angles}. The absence of preferential angles of around 40--60 degrees may imply that the geometry of the tunnels may be either the outcome of social interactions or a response of termites to terrain clues, or a combination of both.
\begin{figure}[!h]
\centering
\includegraphics[width=90mm]{Figure11.pdf}
\caption{\small{ Turning angle distribution. Turning angle distribution in termite walking. Four examples are depicted exhibiting a bell shaped distribution centered at 0 degrees. No preferential angles were identified apart from the persistence of moving forwards.}}
\label{fig:angles}
\end{figure}
\section*{Discussion}
The study of animal movements is of prime importance for understanding ecological and behavioural traits of individual displacements, needed to the efficient use of space. These movements in a social context are extremely important because these regulate the rate of interactions that are the basis for building self-organization global patterns of behaviour \cite{Miramontes.DeSouza.2008} which are, ultimately, essential to cooperation phenomena and social stasis. Social individuals exchange information related to both the status of individuals and the social group and this information flows between individuals who move and encounter each other.
Our results show that free termite displacement, in space and time, present self-similar patterns highly consistent with anomalous diffusion. In social animals interacting nestmates are valuable resources to look for, and that is achieved by moving about in an efficient manner. Isolated termite displacement patterns revealed in this study, add support for the hypothesis that animals adopt L\'evy flights because these confer an advantage in terms of higher fitness resulting from greater efficiency in finding nestmates to whom socially interact. The current trend that highlights statistical model fitting has perhaps put to much attention in fitting methodologies, which are important of course, but other qualitative and quantitative aspects of the phenomena should not remain overlooked, specially the biological phenomenology. Superdiffusion, L\'evy flights and waiting times are quantitatively interconnected in ways that must be explored further. It is intriguing that passive trackers in fluid dynamics do behave qualitatively and quantitatively the same way that termites do. Superdiffusion is a movement pattern observed in a wide range of organisms thought to be related to optimal search strategies. Here we show that termites perform superdiffusive displacements, suggesting that this movement pattern is important for developing social interactions upon which self-organization at the colony level is built. Our results suggest that scale-free superdiffusion may have played an important role in the evolution of societies and cooperation because it may confer an advantage in terms of fitness to individuals that profit from efficient social encounter rates, that translates into efficient information transfer.
What is remarkable in our study, is the fact that assayed termites exhibited superdiffusion and L\'evy flights even in the absence of external stimuli or clues: the arenas contained no food nor other nestmates. It seems, therefore, that such a movement pattern is indeed hardwired in termite ``instincts''. Or recalling a long forgotten idea in termitology, termites orient themselves concatenating external stimuli with some form of internal momentum \cite{Jander.Daumer.1974}. The important connotation arising from such a result is that termite diffusion can proceed independently of a reactive phasis, as expected for a walker with limited (if any) cognitive abilities \cite{Viswanathan.etal.2011}. More specifically, blind termites seem to be equipped with strategies which secure them to find their targets (food or nestmates) even in environments where these are scarce or criptically located.
\section*{Acknowledgments}
OM thanks DGAPA-PAPIIT Grant IN101712 and the Brazilian Ci\^encia Sem Fronteiras program (CSF-CAPES) 0148/2012. ODS is supported by CNPq-Brasil, fellowship 305736/2013-2. We very much appreciate valuable comments, suggestions and support from D. Boyer, D. Lemus, A. Chopps and C. Corona.
\bibliographystyle{IEEEtran-esp}
|
\section{Introduction}
Large-area hybrid silicon pixel detectors are widely used for imaging experiments at synchrotron radiation sources and Free-Electron Lasers (FELs). Examples are Large Area Medipix-Based Detector Array (LAMBDA) \cite{LAMBDA} at PETRA III, Cornell-SLAC Pixel Array Detector (CSPAD) \cite{CSPAD} at LCLS and Adaptive Gain Integrating Pixel Detector (AGIPD) \cite{AGIPD} at the European XFEL. The large detection area of these detectors is obtained by tiling single hybrid modules together. However, the drawback is the formation of a dead region between individual modules, typically of a few mm. This results in X-ray information missing in this region and may cause further problems during image reconstruction.
The main causes of the dead region from single hybrid detector modules are the guard-ring structure of the sensor, the space occupied by the wire-bonding connection between the ASIC chips and the circuit board. Figure \ref{EdgelessDetector} (left) shows the cross section of a conventional hybrid detector module and its dead-space region. The dead space can be reduced by utilizing edgeless detectors with edgeless sensor replacing conventional silicon sensor, with Through-Silicon-Vias (TSVs) and a Re-Distribution Layer (RDL) implemented into the ASIC chips, and with an integration from ASIC chips to circuit board through a Ball-Grid-Array (BGA). Figure \ref{EdgelessDetector} (right) shows the cross section of an edgeless detector module. At DESY, the edgeless silicon pixel detector based on current LAMBDA detector system is under development.
\begin{figure}[htbp]
\small
\centering
\includegraphics[scale=0.27]{Edgeless.eps}
\caption{From conventional to edgeless hybrid detector module.}
\label{EdgelessDetector}
\end{figure}
Recent progress in active-edge techonology of silicon sensors enables the development of edgeless detectors. Such technology has been proven in successful productions of ATLAS and Medipix-based silicon pixel sensors by a few foundries \cite{MBomben, JKCharge}. The key process involves a Deep Reactive Ion Etching (DRIE) to form a trench surrounding the active pixels and a side implantation with the same dopants as backside implant to shield the leakage current from defects located at the edge produced by DRIE \cite{XWu}. The aim of this work is to (1) investigate the radiation hardness of edgeless sensors with active edges using TCAD simulation, (2) understand the charge-collection behavior of edgeless sensors through model calculation, and (3) optimize the edge design for better charge collection by edge pixels and for higher breakdown voltage after high-dose X-ray irradiation.
\section{Investigated structures}
The investigated edgeless sensor is a commercial available product fabricated by VTT (which was later on transferred to its spin-off company Advacam \cite{Advacam}) in a Multi-Project Wafer (MPW) run.
\begin{figure}[htbp]
\small
\centering
\includegraphics[scale=0.2]{Simulated_structure.eps}
\caption{Investigated edgeless sensor produced by VTT/Advacam. Left: Cross section of the edgeless sensor along the cut-line through the center of pixels; Right: Top view of the edgeless sensor (picture reproduced from \cite{JKalliopuska}).}
\label{CrossSection}
\end{figure}
Figure \ref{CrossSection} shows the top view of the investigated edgeless sensor and its cross section. The pitch of pixels is 55 $\mu$m, the width/diameter of pixel implant 30 $\mu$m, which leaves a gap of \mbox{25 $\mu$m} between neighbouring pixel implants. The metal overhang, which is the overlap between aluminum plate and implant window of pixel, is 5 $\mu$m. The distance from the implant boundary of the last pixel to the active edge is 50 $\mu$m, also referred to "edge space" or "last pixel-to-edge distance". The thickness of the SiO$_{2}$ layer, extracted from the capacitance of a MOS capacitor produced on the same silicon wafer biased into accumulation, is 700 nm for n-type and 680 nm for p-type silicon. The junction depth of pixel electrode is $\sim$1.2 $\mu$m for both boron and phosphorus dopants, and the depth of p-spray of 0.9 $\mu$m, which were obtained either from process simulations or from a direct measurement with spreading resistance method. The thickness of SiO$_{2}$ and profiles of implants at pixel side are critical parameters for breakdown simulation.
The edgeless sensors with the following polarities have been investigated: p$^{+}$n, p$^{+}$p, n$^{+}$n and n$^{+}$p with either p-spray or p-stop. For sensors with p-stop, the p-stop implant is 5 $\mu$m wide and located in the center of gaps between pixel implants and in the "edge space" region.
\section{Simulation of electrical properties of edgeless sensors}
The active volume and breakdown voltage as function of radiation dose have been simulated with Synopsys TCAD \cite{TCAD}. In the simulation, the following physics models have been implemented: (1) drift and diffusion of carriers, (2) band-gap narrowing, (3) doping dependence and high-field satuaration of carrier mobility, (4) carrier-carrier scattering and mobility degradation at the interface, (5) doping, temperature dependence and electric field enhancement of Shockley-Read-Hall (SRH) recombination, (6) band-to-band tunneling with Hurkx model, (7) avalance process using vanOverstraetenMan model with the gradient of quasi-Fermi potential as driving force, and (8) fixed charges and generation-recombination at the Si-SiO$_{2}$ interface. In addition, Neumann boundary condition was used on top of SiO$_{2}$, which represents a sensor operation in vacuum or dry atmosphere without any influence due to humidity effect.
The carrier lifetime for both electrons and holes used in the simulation is 1.35 ms, which was extracted from the leakage current of a diode biased into full depletion. The doping concentrations, obtained from the Capacitance-Voltage (CV) measurements on diodes, are $7 \times 10^{11}$ cm$^{-3}$ and $1.1 \times 10^{12}$ cm$^{-3}$ for n-type and p-type silicon, respectively. Before irradiation, the oxide-charge densities are $1.0 \times 10^{10}$ cm$^{-2}$ (n-type), and $3.0 \times 10^{10}$ cm$^{-2}$ (p-type); the surface-recombindation velocities are 1.35 cm/s (n-type) and 3.54 cm/s (p-type).
The simulations were performed at 20$^{\circ}$C in 2D covering the region from the active edge to pixel-10 (10$^{\textrm{th}}$ pixel counted from the active edge). A large simulation region was chosen so that the simulated boundary cutting through pixel-10, where a Neumann boundary applies, will not affect the electric field distribution close to the sensor edge. Sensors with different thicknesses have been simulated, but results for 300 $\mu$m thick Si will be shown.
\subsection{Active volume}
The sensor active volume is obtained by simulating the potential distribution of edgeless sensors biased at different voltages.
Figure \ref{ActiveVolume} (left) shows the results for p$^{+}$n and n$^{+}$p (p-stop/p-spray) sensors biased at 100 V, \mbox{150 V} and 200 V (note the sign of the bias voltage for different polarities). The white line in the figure refers to the depletion boundary of the sensor. For these sensors, the depletion starts from the pixel side and extends to the active edge and backside of the sensor, which leaves an "inactive" (non-depleted) region close to the corner between edge implant and backside implant. The volume of the "inactive" region is influenced by bias voltage as seen in the figure, doping concentration and "edge space". The volume shrinks with increasing bias voltage for a given layout and doping concentration. This can change the electric field distribution close to the edge region and result in a bias voltage dependence of charge collection by pixels close to the active edge. To minimize the volume of the "inactive" region and thus eliminate the dependence of charge collection on bias voltage, the bias should be kept much higher than the depletion voltage of the sensor. Typically, a bias voltage of $\sim$150 V above the depletion voltage is expected to be able to remove such influence for a 300 $\mu$m thick silicon sensor.
\begin{figure}[htbp]
\small
\centering
\includegraphics[scale=0.5]{Active_volume.eps}
\caption{The potential distributions at different bias voltages for 300 $\mu$m thick sensors without irradiation. Left: Results for p$^{+}$n and n$^{+}$p (p-stop/p-spray) sensors. Right: Results for n$^{+}$n (p-stop/p-strap) and p$^{+}$p sensors. Note the sign of the bias voltage for different polarities.}
\label{ActiveVolume}
\end{figure}
In figure \ref{ActiveVolume} (right), the potential distributions for n$^{+}$n (p-stop/p-spray) and p$^{+}$p sensors are shown. As the depletion starts from the active edge and the backside electrode, the entire sensor volume will be fully depleted once the depletion reaches the pixel electrodes. This will not leave "inactive" region inside the sensor. Thus, the charge collection by pixels close to the active edge does not strongly depend on bias voltage once the sensor is biased above its depletion voltage.
\subsection{Sensor breakdown and radiation hardness}
One of the main aims of this work is to investigate the breakdown voltage of edgeless sensors as function of X-ray dose. For sensors after X-ray irradiation, oxide charges, interface traps and border traps will be introduced either in the SiO$_{2}$ insulating layer or at the Si-SiO$_{2}$ interface \cite{JZMOS}. In the simulation, oxide charges and charged interface traps were represented as fixed charges at the Si-SiO$_{2}$ interface. In addition, the current generated by the states of interface traps close to the silicon mid-gap was described by surface-recommbination velocity and implemented in TCAD simulation. The details of the input parameters related to X-ray radiation damage can be found in \cite{JZMOS, JS}. The effect of border traps were not simulated due to the lack of quantative characterization.
\begin{figure}[htbp]
\small
\centering
\includegraphics[scale=0.55]{Field_pn_pp.eps}
\caption{Electric field close to the sensor corner at pixel side and its distribution along the cut-line at \mbox{10 nm} below the Si-SiO$_{2}$ interface as function of X-ray irradiation dose represented by fixed charges (\mbox{$N_{ox}=3 \times 10^{12}$ cm$^{-2}$} corresponds to $\sim$10 MGy). Left: p$^{+}$n sensor; Right: p$^{+}$p sensor.}
\label{ElectricFieldPN}
\end{figure}
Figure \ref{ElectricFieldPN} shows the electric field distributions close to the sensor corner at the pixel side (\mbox{50 $\mu$m} depth) for p$^{+}$n and p$^{+}$p sensors. For both sensors, the highest electric field appears at the Si-SiO$_{2}$ interface below the aluminium of the first pixel before irradiation; after high-dose irradiation, the positive charges in the SiO$_{2}$ and charged interface traps induce a layer of electrons accumulating below the Si-SiO$_{2}$ interface, which conducts the high voltage on the backside of the sensor through the active edge to the first pixel and thus results in a high electric field at the junction of the first pixel. In this case, the first pixel behaves like a Current-Collection Ring (CCR) of a conventional sensor and thus breakdown first. It should be noted that, for both sensors, the breakdown votlage is $\sim$500 V before irradiation, and it sharply decreases with X-ray irradiation dose and finally reaches a value of $\sim$20 V at a high dose equivalent to $\sim$10 MGy. The results are consistent with those published in \cite{JSDesign} for a sensor with 700 nm thick SiO$_{2}$ and without any guard rings.
\begin{figure}[htbp]
\small
\centering
\includegraphics[scale=0.55]{Field_nn_nn.eps}
\caption{Electric field close to the sensor corner at pixel side and its distribution along the cut-line at \mbox{10 nm} below the Si-SiO$_{2}$ interface as function of X-ray irradiation dose represented by fixed charges (\mbox{$N_{ox}=3 \times 10^{12}$ cm$^{-2}$} corresponds to $\sim$10 MGy). Left: n$^{+}$n (and similarly n$^{+}$p) sensor with p-stop; Left: n$^{+}$n (and similarly n$^{+}$p) sensor with p-spray.}
\label{ElectricFieldPP}
\end{figure}
Electric field distributions for n$^{+}$n (n$^{+}$p) sensor with either p-stop or p-spray are shown in figure \ref{ElectricFieldPP}. For n$^{+}$n (n$^{+}$p) sensor with p-stop, the highest electric field is located at the junction of the p-stop between the active edge and the first pixel before irradiation; after irradiation, the high density of electrons from the accumulation layer creates an additional high electric field region at the interface between the edge implant and the accumulation layer and causes a breakdown at the corner of the sensor. For n$^{+}$n (n$^{+}$p) sensor with p-spray, the highest electric fields are located below the aluminum of the first pixel and at the interface between the p-spray layer and the implant of the first pixel before irradiation; after irradiation, a corner breakdown is also observed similar to n$^{+}$n (n$^{+}$p) sensor with p-stop. The breakdown voltage of n$^{+}$n (n$^{+}$p) sensor with p-spray increases with irradiation till the density of fixed charges, $N_{ox}$, larger than the dose of p-spray layer (\mbox{$\sim$1.5$\times$10$^{12}$ cm$^{-2}$}) and then the breakdown voltage starts to decrease. For both n$^{+}$n (n$^{+}$p) sensors with p-stop and p-spray, their breakdown voltage after irradiation is as low as $\sim$25 V at $\sim$10 MGy.
It should be noted that the simulation with drift-diffusion model and fixed charges implemented in gives pessimistic estimation for the breakdown.
\subsection{Discussion: Towards radiation-hard design}
Potentially, there are a few ways to enhance the radiation hardness of edgeless sensors with active edges. The first is to add a guard-ring structure (CCR + floating guard rings) to the region between the active edge and the first pixel. However, this will leave an "inactive" region, because charges generated in some parts of the sensor will be collected by the CCR. The second is to optimize the technological parameters to achieve a high breakdown voltage for a sensor without any guard-ring structure. Here we propose the following methods according to different choices for sensor polarity:
(1) For p$^{+}$n and p$^{+}$p sensors, the radiation hardness can be achieved by optimizing the oxide thickness below the field plate, junction depth and the doping concentration. In previous work done for the AGIPD sensor, in the case where no floating guard rings are present, the breakdown voltage has been simulated as function of oxide thickness for different junction depth \cite{JSDesign}. Considering the minimal requirement for edgeless sensor operation is that the operating voltage, $V_{op}$, should be larger than the depletion voltage, $V_{dep}$, but smaller than the breakdown voltage, $V_{bd}$, the maximal doping concentration $N_{d}$ can be calculated for silicon sensors with different thicknesses according to $V_{dep} = \frac{q_{0}T_{si}^{2}N_{d}}{2\varepsilon_{0}\varepsilon_{si}} - V_{bi} \leq V_{op} \leq V_{bd}$. Figure \ref{Breakdown} shows the calculated results on the maximal doping concentration required by 300 $\mu$m and 500 $\mu$m thick sensors as function of technological parameters in order to achieve a breakdown voltage larger than the depletion voltage. It can be seen that, for example, by keeping the doping concentration less than $1.2 \times 10^{12}$ cm$^{-3}$ for 300 $\mu$m Si and $4.0 \times 10^{11}$ cm$^{-3}$ for 500 $\mu$m Si, the breakdown voltage can be higher than the depletion voltage for a sensor with 250 nm thick SiO$_{2}$ and junction depth of 2.4 $\mu$m. Therefore, careful selection of technological parameters makes the radiation hardness of p$^{+}$n and p$^{+}$p edgeless sensors possible.
\begin{figure}[htbp]
\small
\centering
\includegraphics[scale=0.5]{Doping_vs_Tox.eps}
\caption{Maximal doping in edgeless sensors for radiation hardness purpose. Results re-calculated based on the results from \cite{JSDesign}.}
\label{Breakdown}
\end{figure}
(2) For the n$^{+}$n and n$^{+}$p sensors with p-spray layer, it has been seen that the breakdown voltage increases as function of irradiation at low doses. This indicates that an optimized/increased dose of p-spray layer can improve the radiation hardness of edgeless sensor to a higher irradiation dose. However, the drawback by increasing the p-spray dose is the reduction of breakdown voltage before irradiation as the increased boron concentration increases the electric field at the interface between the p-spray layer and the pixel implant at a fixed bias voltage. To balance this, the p-spray dose can be chosen as half of the saturation value of the density of fixed charges. Thus, the density of fixed charges has to be characterized as function of dose from the test structures produced by the same technology from the vendor. Another option is to use a p-spray dose same as the saturation value of the density of fixed charges and perform a pre-irradiation to the sensors to certain dose level in order to reduce the effect from the high-dose p-spray. The second needs more investigations before the sensor works.
\section{Modeling charge-collection behavior of edgeless sensors}
The charge collection of edgeless sensors has been widely investigated \cite{JKCharge, RBates}. It has been commonly observed that the charge collection by edge pixels are highly non-uniform. To understand the charge-collection behavior of edgeless sensors, we have developed a model. The model is used to understand the measurement results from thin sensors and predict the performance for thick silicon sensors.
\subsection{Model development}
The model is based a Finite Volume Method (FVM), which segments the sensor into finite volumes. The model considers the following main processes: (1) Photon absorption in silicon sensor and generation of electron-hole pairs, (2) drift of free carriers along electric field and lateral diffusion, and (3) carrier collection by electrodes and signal processing in readout electronics.
In the first process, X-ray photons are absorbed in silicon sensor through photoelectric effect, which is the dominant effect for photon energy less than 20 keV. The probability of photon absorption in each finite volume is given by $1/ \lambda (E_{xray}) \cdot \textrm{exp} [- y/ \lambda (E_{xray})]$, where $\lambda (E_{xray})$ is the attenuation length of X-ray photons with energy $E_{xray}$ and $y$ the distance from the point where the photon is absorbed to the silicon surface where X-rays enter. Once the photon is absorbed, a charge cloud consisting of electrons and holes are created. The number of electron-hole pairs inside the charge cloud is given by the X-ray energy divided by 3.6 eV, which is the mean energy needed to generate one pair of electron and hole in silicon.
In the second process, the electrons and holes drift to either pixel or backside/edge electrode, depending on the sensor polarity. The drifting path of the center of mass of electrons and holes follows the eletric field lines (vector) obtained from TCAD simulation. Then, the drifting time of carriers to pixels is calculated by $t = \int _{x,y} \frac{\textrm{d}y}{\mu \cdot E(x,y)}$, with $\mu$ the carrier mobilty and $E(x,y)$ the electric field. The size of charge cloud when carriers reach pixels is given by the lateral diffusion of carrers: $\sigma = \sqrt{2D\cdot t}$, with $D$ the diffusion coefficient.
Finally, the carriers are collected by different pixels, depending upon their locations when they reach the pixel side. Results from the model calculation can be output either as number of photons or charges depending on whether the ASIC chip is operated in photon-counting or charge-integrating mode. In addtion, a threshold $E_{thr}$ can be set in simulation so that results can be compared to measurements with a certain threshold. The final spectrum can be deconvoluted with the profile of the beam spot for a beam scan experiment.
\subsection{Comparison to measurements}
Results from model calculation have been compared to measurements in a beam scan experiment done at the Diamond Light Source. The X-ray beam with energy of 15 keV and a spot of FWHM=\mbox{11 $\mu$m} was shot into the backside (non-pixel side) of 150 $\mu$m thick edgeless sensor coupled to a Medipix chip \cite{RBates}. The counts by edge pixels were measured with a threshold setting at 5 keV in Medipix chip as function of distance from the X-ray beam to the edge of the sensor. On top of figure \ref{Comparison}, the measurement scheme is shown.
\begin{figure}[htbp]
\small
\centering
\includegraphics[scale=0.15]{Measurement.eps}
\includegraphics[scale=0.39]{Comparison_model_meas.eps}
\caption{Comparison between model calculation and measurements. Top: Measurement scheme and electric field lines inside the sensor; Bottom left: Counts as function of beam position; Bottom right: Normalised counts for edge pixels.}
\label{Comparison}
\end{figure}
The comparison between model calculation and measurements is shown on the bottom of figure \ref{Comparison}. A good agreement was obtained. The counts for different edge pixels as function of distance from X-ray beam to the sensor edge is shown in the left figure: Non-uniform charge collection by pixels close to the sensor edge is observed. The non-uniformity of charge collection is explained by the bending of electric field close to the sensor edge: Figure \ref{Comparison} (top) shows the distributions of electric field lines ending at the middle of pixel gaps (10 $\mu$m below the Si-SiO$_{2}$ interface). By integrating the individual distributions for each pixel, the total counts in the scan for each pixels can be obtained, as shown in figure \ref{Comparison} (right). The counts for the edge pixels have been normalised to the count of a central pixel. The result corresponds to the counting behavior of edge pixels in a flat-field image. For 150 $\mu$m thick sensor, the last two pixels count differently compared to the other pixels.
\subsection{Prediction}
In photon science application with hard X-rays, thick silicon sensors are preferred and commonly used in order to obtain a good quantum efficiency. With this developed model, it is possible to predict the charge-collection behavior for thick edgeless silicon sensors.
\begin{figure}[htbp]
\small
\centering
\includegraphics[scale=0.60]{Charge_collection_thick_Si.eps}
\caption{Prediction of charge-collection behavior for 300 $\mu$m and 500 $\mu$m thick edgeless sensors with 50 $\mu$m "edge space" from model calculation. Top row: Electric field lines; Middle row: Counts as function of beam position in a beam scan experiment; Bottom row: Normalised counts for edge pixels.}
\label{Prediction}
\end{figure}
Figure \ref{Prediction} shows the electric field lines inside 300 $\mu$m and 500 $\mu$m thick edgeless silicon sensors with 50 $\mu$m "edge space", counts as function of beam position in a beam scan experiment and their normalised counts for edge pixels. It is found that the results depend on the thickness of the sensor and the energy of X-rays. approximately 7 pixels for 300 $\mu$m Si and 10 pixels for 500 $\mu$m Si close to the edge are influenced due to the bending of electric field caused by edge implantation. In addition, for low-energy X-rays ($E_{xray} < 10$ keV), from two to four pixels cannot effectively respond to photons for 300 $\mu$m and 500 $\mu$m thick sensors, respectively.
Therefore, in order to maintain the possibilty of edge pixels to see low-energy photons, we propose the distance from the last pixel to the active edge should be kept at a distance at least 50\% of the sensor thickness.
\section{Summary and outlook}
In this work , radiation hardness of edgeless sensor with active edges has been investigated through TCAD simulation. Results indicate poor radiation hardness for current designs. To improve the radiation hardness, a few methods have been discussed for different sensor polarity choices. In addition, a model has been developed, which makes it possible to reproduce the measurement results at a beam scan experiment with X-rays and predict the charge-collection behavior for different sensor layouts. Finally, the distance from the last pixel to active edge is proposed based on the model calculation in order to obtain a better sensitivity to low-energy photons for edgeless sensors.
\acknowledgments
The work was performed within the LAMBDA project. J. Zhang would like to thank J. Kalliopuska of Advacam co. for providing process information for TCAD simulation, and D. Maneuski of Glasgow University sharing the measurement data taking at the Diamond Light Source for comparison with the model calculation.
|
\section{Introduction}
The task of recovering the latent topics underlying a given corpus of $D$ documents has been in the forefront of active research in statistical
machine learning for more than a decade, and continues to receive the dedicated contributions from many researchers from around the world.
Since the introduction of Latent Dirichlet Allocation (LDA) \cite{Blei03latentdirichlet} and then the extension to correlated topic models (CTM) \cite{Blei06correlatedtopic}, a series of excellent contributions have been made to this exciting field, ranging from slight extension
in the modelling structure to the development of scalable topic modeling algorithms capable of handling extremely large
collections of documents, as well as selecting an optimal model among a collection of competing models or using the output of topic modelling as entry points (inputs)
to other machine learning or data mining tasks such as image analysis and sentiment extraction, just to name a few. As far as correlated topic models are concerned,
virtually all the contributors to the field have so far concentrated solely on the use of the logistic normal topic model. The seminal paper on correlated topic model\cite{Blei06correlatedtopic} adopts a variational approximation approach to model fitting while subsequent authors like \cite{Mimno_gibbssampling} propose a Gibbs sampling scheme with data augmentation of uniform random variables. More recently, \cite{Tsinghua:2013:1} presented an exact and scalable Gibbs sampling algorithm with Polya-Gamma distributed auxiliary variables which is a recent development of efficient sampling of logistic model. Despite the inseparable relationship between logistic and probit model in statistical modelling, the probit model has not yet been proposed, probably due to its computational inefficiency for multiclass classification problem and high posterior dependence between auxiliary variables and parameters. As for practical application where topic models are commonly employed, having multiple topics is extremely prevalent. In some cases, more than 1000 topics will be fitted to large datasets such as Wikipedia and Pubmed data. Therefore, using MCMC probit model in topic modeling application will be impractical and inconceivable due to its computational inefficiency. Nonetheless, a recent work on diagonal orthant probit model \cite{Johndrow:2013:1} substantially improved the sampling efficiency while maintaining the predictive performance, which motivated us to build an alternative correlated topic modeling with probit normal topic distribution. On the other hand,
probit models inherently capture a better dependency structure between topics and co-occurrence of words within a topic as it doesn't assume the IIA (independence of irrelevant alternatives) restriction of logistic models.\\
\\
The rest of this paper is organized as follows: in section 2, we present a conventional formulation of topic modelling along with our general notation and the correlated topic models
extension.
Section 3 introduces our adaptation of the diagonal orthant probit model to topic discovery in the presence correlations among topics, along with the corresponding auxiliary variable sampling scheme for updating the probit model parameters and the remainder of all the posterior distributions of the parameters of the model. Unlike with the logistic normal formulation
where the non-conjugacy leads to the need for sophisticated sampling scheme, in this section we clearly reveal the simplicity of our proposed method resulting from the natural conjugacy inherent in the auxiliary formulation of the updating of the parameters. We also show compelling computational demonstrations of the efficiency of the diagonal orthant approach compared to the traditional multinomial probit for on both the auxiliary variable sampling and the estimation of the topic distribution.
Section 4 presents the performance of our proposed approach on the Associated Press data set, featuring the intuitively appealing topics discovered, along with
the correlation structure among topics and the loglikelihood as a function of topical space dimension. Section 5 deals with our conclusion, discussion
and elements of our future work.
\section{General aspects of topic models}
In a given corpus, one could imagine that each document deals with one or more topics. For instance, one of the collection considered
in this paper is provided by the Associated Press and covers topics as varied as {\it aviation, education, weather, broadcasting, air force, navy,
national security, international treaties, investing, international trade, war, courts, entertainment industry, politics}, and etc.
From a statistical perspective, a topic is often modeled as a {\it probability distribution over words}, and as a result a given document
is treated as a {\it mixture of probabilistic topics} \cite{Blei03latentdirichlet}. We consider a setting where we have a total of $V$ unique words in the reference vocabulary and $K$ topics underlying the $D$ documents provided. Let $w_{dn}$ denote the $n$-th word in the $d$-th document, and let $z_{dn}$ refer to the label of the topic assigned to the $n$-th word of that $d$-th document. Then the probability of $w_{dn}$ is given by
\begin{eqnarray}
\Pr(w_{dn}) = \sum_{k=1}^K{\Pr(w_{dn}| z_{dn}=k)\Pr(z_{dn}=k)},
\label{eq:tm:1}
\end{eqnarray}
where $\Pr(z_{dn}=k)$ is the probability that the $n$th word in the $d$th document is assigned to topic $k$.
This quantity plays an important role in the analysis of correlated topic models. In the seminal article on correlated
topic models \cite{Blei06correlatedtopic}, $\Pr(z_{dn}=k)$ is modeled for each document $d$ as a function of a $K$-dimensional vector
$\nmathbf \eta_d$ of parameters. Specifically, the logistic-normal defines $\nmathbf \eta_d = (\eta_{d}^1, \eta_{d}^2, \cdots, \eta_{d}^K)$ where the last element $\eta_{d}^K$ is typically set to zero for identifiability and assumes
with $\nmathbf \eta_d \sim {\tt MVN}(\nmathbf \mu, \nmathbf \Sigma)$ with
$$
\theta_{d}^k = \Pr[z_{dn}^k = 1 | \nmathbf \eta_d] =f(\nmathbf \eta_d) = \frac{e^{\eta_{d}^k}}{\sum_{j=1}^K{e^{\eta_{d}^j}}},\quad k=1,2,\cdots,K-1 \quad \textit{and} \quad \theta_{d}^K = \frac{1}{\sum_{j=1}^K{e^{\eta_{d}^j}}},
$$
Also, $\forall n \in \{1,2, \cdots, N_d\}\,\,\,$ and $z_{dn} \sim {\tt Mult}(\nmathbf \theta_{d})$, and
$w_{dn} \sim {\tt Mult}(\nmathbf \beta)$. With all these model components defined, the estimation task in correlated topic modelling from a Bayesian perspective
can be summarized in the following posterior
\begin{eqnarray}
p(\nmathbf \eta_d,\nmathbf Z |\nmathbf W, \nmathbf \mu, \nmathbf \Sigma) &\propto& p(\nmathbf W|\nmathbf Z)\prod_{d=1}^D\left\{\prod_{n=1}^{N_d} p(z_{dn}) p(\nmathbf \eta_d|\nmathbf \mu, \nmathbf \Sigma)\right\}\\ \nonumber
&=&\prod_{k=1}^{K}{\frac{\delta(C_{k}+ \nmathbf \beta)}{\delta(\nmathbf \beta)}}\prod_{d=1}^D\left\{\left(\prod_{n=1}^{N_d} \theta_d^{z_{dn}}\right) \mathcal{N}(\nmathbf \eta_d|\nmathbf \mu, \nmathbf \Sigma)\right\},
\label{eq:post:1}
\end{eqnarray}
where $\delta(\cdot)$ is defined using the Gamma function $Gamma(\cdot)$ so for a $K$-dimension vector $\nmathbf u$,
$$
\delta(\nmathbf u) = \displaystyle \frac{\displaystyle \prod_{k=1}^K{\Gamma(u_k)}} {\Gamma\left(\displaystyle \sum_{k=1}^K{u_k}\right)}.
$$
\eqref{eq:post:1} provides the ingredients for estimating the parameter vectors $\nmathbf \eta_d$ that help capture the correlations among topics, and the matrix $\nmathbf Z$
that contains the topical assignments. Under the logistic normal model, sampling from the full posterior of $\nmathbf \eta_d$ derived from the joint posterior in \eqref{eq:post:1} requires the use of sophisticated sampling schemes like the one used in \cite{Tsinghua:2013:1}. Although these authors managed to achieve great performances on large corpuses of documents, we thought it useful to contribute to correlated topic modelling by way of the multinomial probit. Clearly, as indicated earlier, most authors concentrated on logistic-normal even despite non-conjugacy, and the lack of probit topic modeling can be easily attributed to the inefficiency of the corresponding sampling scheme. In the most raw formulation of the multinomial probit that intends to capture the full extend of all the correlations among the topics, the topic assignment probability is defined by \eqref{eq:probit:0}.
\begin{eqnarray}
\Pr(z_{dn}=k)=\theta_{d}^k = \int\int\int\cdots\int{\phi_{K}(u; \nmathbf \eta_d, R) d u}
\label{eq:probit:0}
\end{eqnarray}
The practical evaluation of \eqref{eq:probit:0} involves a complicated high dimensional integral which is typically computationally intractable when the number of categories is greater than $4$. A relaxed version of \eqref{eq:probit:0}, one that still captures more correlation than the logit and that is also very commonly used in practice, defines $\theta_d^k$ as
\begin{eqnarray}
\theta_{d}^k = \int_{-\infty}^{+\infty}{\left\{\prod_{j=1, j\neq k}^{K}{\Phi(v + \eta_{d}^k - \eta_{d}^{j})}\right\}\phi(v) dv}
= \mathbb{E}_{\phi(v)}\left\{\prod_{j=1, j\neq k}^{K}{\Phi(V + \eta_{d}^k - \eta_{d}^{j})}\right\},
\label{eq:probit:prob:1}
\end{eqnarray}
where $\phi(v)=\frac{1}{\sqrt{2\pi}}e^{-\frac{1}{2}v^2}$ is the standard normal density, and $\Phi(v)=\int_{-\infty}^{v}{\phi(u)du}$ is the standard normal distribution function.
Despite this relaxation, the multinomial probit in this formulation still has major drawbacks namely: (a) Even when one is given the vector $\nmathbf \eta_d$, the calculation of
$\theta_d^k$ remains computationally prohibitive even for moderate values of $K$. In practice, one may consider using a monte carlo approximation to that integral in \eqref{eq:probit:prob:1}. However, such an approach in the context of a large corpus with many underlying latent topics renders the probit formulation almost unusable.
(b) As far as the estimation of $\nmathbf \eta_d$ is concerned, a natural approach to sampling from the posterior of $\nmathbf \eta_d$ in this context would be to use the Metropolis-Hastings
updating scheme, since the full posterior in this case is not available. Unfortunately, the Metropolis in this case is excruciatingly slow with poor mixing rates and high sensitivity to the proposal distribution. It turns out that an apparently appealing solution in this case could come from the auxiliary variable formulation as described in \cite{AlbertChib93}. Unfortunately, even this promising formulation fails catastrophically for moderate values $K$ as we will demonstrate in the subsequent section, due to the high dependency structure between auxiliary variables and parameters. Essentially, the need for Metropolis is avoided by defining an auxiliary vector $Y_d$ of dimension $K$. For $n=1,\cdots,N_d$, we consider the vector $z_{dn}$ containing the current topic allocation and we repeatedly sample $Y_{dn}$ from a $K$-dimensional multivariate Gaussian
until the component of $Y_{dn}$ that corresponds to the non-zero index in $z_{dn}$ is the largest of all the components of $Y_{dn}$, ie.
\begin{eqnarray}
Y_{dn}^{z_{dn}} = \underset{k=1,\cdots,K}{\max}\{Y_{dn}^k\}.
\label{eq:max:aux:1}
\end{eqnarray}
The condition in \eqref{eq:max:aux:1} typically fails to be fulfilled even when $K$ is moderately large. In fact, we demonstrate later that in some cases, it becomes impossible to
find a vector $Y_{dn}$ satisfying that condition. Besides, the dependency of $Y_{dn}$ on the current value of $\nmathbf \eta_{d}$ further complicates the sampling scheme especially in the case of large topical space. In the next section,
we remedy these inefficiencies by proposing and developing our adaptation of the diagonal orthant multinomial probit.
\section{Diagonal Orthant Probit for Correlated Topic Models}
In a recent work, \cite{Johndrow:2013:1} developed the diagonal orthant probit approach to multicategorical classification. Their approach circumvents the bottlenecks mentioned earlier and substantially improves the sampling efficiency while maintaining the predictive performance. Essentially, the diagonal orthant probit approach successfully makes the most of the benefits of binary classification, thereby substantially reducing the high dependency that made the condition \eqref{eq:max:aux:1} computationally unattainable. Indeed, with the diagonal orthant multinomial model, we achieved three main benefits
\begin{itemize}
\item A more tractable and easily computatble definition of topic distribution $\theta_d^k=\Pr(z_{dn}=k|\nmathbf \eta_d)$
\item A clear and very straightforward and adaptable auxiliary variable sampling scheme
\item The capacity to handle a very large number of topics due to the efficiency and low dependency.
\end{itemize}
Under the diagonal orthant probit model, we have
\begin{eqnarray}
\theta_{d}^k = \frac{(1-\Phi(-\eta_d^k))\prod_{j\neq k}\Phi(-\eta_d^{j})}{\displaystyle \sum_{\ell=1}^{K}{(1-\Phi(-\eta_d^\ell))\prod_{j\neq \ell}\Phi(-\eta_d^{j})}}.
\label{eq:orthant:prob:1}
\end{eqnarray}
The generative process of our probit normal topic models is essentially identical to logistic topic models except that the topic distribution for each document now is obtained by a probit transformation of a multivariate Gaussian variable \eqref{eq:orthant:prob:1}. As such, the generating process of a document of length $\textit\mbox{N}_d $ is as follows:
\begin{enumerate}
\item Draw $\eta \sim {\tt MVN}(\mu, \Sigma)$ and transform $\eta_{d}$ into topic distribution $\theta_{d}$ where each element of $\theta$ is computed as follows:
\begin{eqnarray}
\theta_{d}^k = \frac{(1-\Phi(-\eta_d^k))\prod_{j\neq k}\Phi(-\eta_d^{j})}{\displaystyle \sum_{\ell=1}^{K}{(1-\Phi(-\eta_d))\prod_{j\neq \ell}\Phi(-\eta_d^{j})}}.
\end{eqnarray}
\item For each word position $\textit{n}$ $\in (1,\cdots, \textit\mbox{N}_d )$
\begin{enumerate}
\item Draw a topic assignment $Z_n \sim {\tt Mult} (\theta_d)$
\item Draw a word $W_n \sim {\tt Mult} (\varphi^{z_n})$
\end{enumerate}
\end{enumerate}
Where $\Phi(\cdot)$ represents the cumulative distribution of the standard normal.
We specify a Gaussian prior for $\nmathbf \eta_d$, namely $(\nmathbf \eta_d | \cdots) \sim {\cal N}_{K}(\nmathbf \mu, \nmathbf \Sigma)$.
Throughout this paper, we'll use $\phi_{K}(\cdot)$ to denote the $K$-dimensional multivariate Gaussian density function,
$$
\phi_{K}(\nmathbf \eta_d; \nmathbf \mu, \nmathbf \Sigma) = \frac{1}{\sqrt{(2\pi)^K |\Sigma|}}\exp\left\{-\frac{1}{2}(\nmathbf \eta_d-\nmathbf \mu)^\top\nmathbf \Sigma^{-1} (\nmathbf \eta_d-\nmathbf \mu)\right\}.
$$
To complete the Bayesian analysis of our probit normal topic model, we need to sample from the joint posterior
\begin{eqnarray}
p(\nmathbf \eta_d, \nmathbf Z_d | \nmathbf W, \nmathbf \mu, \nmathbf \Sigma) \propto p(\nmathbf \eta_d | \nmathbf \mu, \nmathbf \Sigma)p(\nmathbf Z_d | \nmathbf \eta_d)p(\nmathbf W|\nmathbf Z_d).
\label{eq:post:joint:2}
\end{eqnarray}
As noted earlier, the second benefit of the diagonal orthant probit model lies in its clear, simple, straightforward yet powerful auxiliary variable sampling scheme.
We take advantage of that diagonal orthant property when dealing with the full posterior for $\nmathbf \eta_d$ given by
\begin{eqnarray}
p(\nmathbf \eta_d | \nmathbf W, \nmathbf Z_d, \nmathbf \mu, \nmathbf \Sigma) \propto p(\nmathbf \eta_d | \nmathbf \mu, \nmathbf \Sigma)p(\nmathbf Z_d | \nmathbf \eta_d).
\label{eq:post:full:eta:1}
\end{eqnarray}
While sampling directly from \eqref{eq:post:full:eta:1} is impractical, defining a collection of auxiliary variables $\nmathbf Y_d$
allows a scheme that samples from the joint posterior $p(\nmathbf \eta_d, \nmathbf Z_d, \nmathbf Y_d|\nmathbf W, \nmathbf \mu, \nmathbf \Sigma)$ using the following:
For each document $d$, the matrix $\nmathbf Y_d \in \mathbb{R}^{N_d \times K}$ contains all the values of the auxiliary variables,
$$
\nmathbf Y_d = \left[\begin{array}{cccccc}
Y_{d1}^1 & Y_{d1}^2 & \cdots & Y_{d1}^k & \cdots & Y_{d1}^K\\
Y_{d2}^1 & Y_{d2}^2 & \cdots & Y_{d2}^k & \cdots & Y_{d2}^K\\
\vdots & \vdots & \cdots & \ddots & \cdots & \vdots\\
Y_{d,{N_d-1}}^1 & Y_{d,{N_d-1}}^2 & \cdots & Y_{d,{N_d-1}}^k & \cdots & Y_{d,{N_d-1}}^K\\
Y_{d,{N_d}}^1 & Y_{d,{N_d}}^2 & \cdots & Y_{d,{N_d}}^k & \cdots & Y_{d,{N_d}}^K
\end{array}\right]
$$
Each row $Y_{dn} = (Y_{dn}^1, \cdots, Y_{dn}^k, \cdots, Y_{dn}^K)^\top$ of $\nmathbf Y_d$ has $K$ components, and the diagonal orthant
updates them readily using the following straightforward sampling scheme: Let $k$ be the current topic allocation for the nth word.
\begin{itemize}
\item For the component of $Y_{dn}$ whose index corresponds to the label of current topic assignment of word $n$
sample from a truncated normal distribution with variance $1$ restricted to positive outcomes
$$
(Y_{dn}^{k} | \eta_d^{k}) \sim \mathcal{N}_{+}(\eta_d^{k}, 1) \quad z_{dn}^k = 1
$$
\item For all components of $Y_{dn}$ whose indices do correspond to the label of current topic assignment of word $n$
sample from a truncated normal distribution with variance $1$ restricted to negative outcomes
$$
(Y_{dn}^{j} | \eta_d^{j}) \sim \mathcal{N}_{-}(\eta_d^{j}, 1) \quad \quad z_{dn}^{j}\neq 1
$$
\end{itemize}
Once the matrix $\nmathbf Y_d$ is obtained, the sampling scheme updates the parameter vector $\nmathbf \eta_d$ by conveniently drawing
$$
(\nmathbf \eta_d | \nmathbf Y_d, \nmathbf A, \nmathbf \mu, \nmathbf \Sigma) \sim MVN(\nmathbf \mu_{\nmathbf \eta_d}, \nmathbf \Sigma_{\nmathbf \eta_d}),
$$
where
$$
\nmathbf \mu_{\nmathbf \eta_d} = \nmathbf \Sigma_{\nmathbf \eta_d}(\nmathbf \Sigma^{-1} \nmathbf \mu + \nmathbf X_d^\top \nmathbf A^{-1}{\rm vec}(\nmathbf Y_d))
\quad
\text{and}
\quad
\nmathbf \Sigma_{\nmathbf \eta_d} = (\nmathbf \Sigma^{-1} + \nmathbf X_d^\top \nmathbf A^{-1} \nmathbf X_d)^{-1}.
$$
with $\nmathbf X_d = {\nmathbf 1}_{N_d} \otimes \nmathbf I_K$ and ${\rm vec}(\nmathbf Y_d)$ representing the row-wise vectorization of
the matrix $\nmathbf Y_d$.
Adopting the fully Bayesian treatment of our probit normal correlated topic model, we add an extra layer to the hierarchy in order to capture
the variation in the mean vector and the variance-covariance matrix of the parameter vector $\nmathbf \eta_d$. Taking advantage of
conjugacy, we specify a normal-Inverse-Wishart prior for $(\nmathbf \mu, \nmathbf \Sigma)$, namely,
$$
p(\nmathbf \mu, \nmathbf \Sigma) = NIW(\nmathbf \mu_0, \kappa_0, \Psi_0, \nu_0),
$$
meaning that $\nmathbf \Sigma|\nu_0, \Psi_0 \sim IW(\Psi_0, \nu_0)$ and $(\nmathbf \mu | \nmathbf \mu_0, \nmathbf \Sigma, \kappa_0) \sim MVN(\nmathbf \mu_0, \Sigma/\kappa_0)$.
The corresponding posterior is normal-inverse-Wishart, so that we can write
$$
p(\nmathbf \mu, \nmathbf \Sigma| \nmathbf W, \nmathbf Z, \nmathbf \eta) = NIW(\nmathbf \mu^\prime, \kappa^\prime, \Psi^\prime, \nu^\prime),
$$
where $\kappa^\prime=\kappa_0+D$, $\nu^\prime = \nu_0 + D$, $\nmathbf \mu^\prime = \frac{D}{D+\kappa_0}\bar{\nmathbf \eta} + \frac{\kappa_0}{D+\kappa_0}\nmathbf \mu_0$,
and
$$
\Psi^\prime = \Psi_0 + Q + \frac{\kappa_0}{\kappa_0+D}(\bar{\nmathbf \eta}-\nmathbf \mu_0)(\bar{\nmathbf \eta}-\nmathbf \mu_0)^\top,
$$
where
$$
Q = \sum_{d=1}^{D}{(\nmathbf \eta_d-\bar{\nmathbf \eta})(\nmathbf \eta_d-\bar{\nmathbf \eta})^\top}.
$$
As far as sampling from the full posterior distribution of $Z_{dn}$ is concerned, we use the expression
$$
\Pr[z_{dn}^k = 1 | \nmathbf Z_{\neg n}, w_{dn}, \nmathbf W_{\neg dn}] \propto p(w_{dn}|z_{dn}^k = 1,\nmathbf W_{\neg dn},\nmathbf Z_{\neg n}) \theta_{d}^k \propto \frac{C_{k,\neg n}^{w_{dn}}+\beta_{w_{dn}}}{\sum_{j=1}^V{C_{k,\neg n}^{j}} + \sum_{j=1}^V{\beta_j}} \theta_{d}^k.
$$
where the use of $C_{\cdot,\neg n}$ is used to indicate that the $n$th is not included in the topic or document under consideration.
\section{Computational results on the Associated Press data}
In this section, we used a famous Associated Press data set from \cite{Grün:Hornik:2011:JSSOBK:v40i13} in R to uncover the word topic distribution, the correlation structure between various topics as well as selecting optimal models. The Associated Press corpus consists of 2244 documents and 10473 words. After preprocessing the corpus by picking frequent and common terms, we reduced the size of the words from 10473 to 2643 for efficient sampling. \\
In our first experimentation, we built a correlated topic modelling structure based on the traditional multinomial
probit and then tested the computational speed for key sampling tasks. The high posterior dependency structure between auxiliary variables and parameters make
multinormal probit essentially unscalable for situations where it is impossible for the sampler to yield a random variate of the auxiliary variable corresponding the current topic allocation label that is also the maximum \eqref{eq:max:aux:1}. For a random initialization of topic assignment, the sampling of auxiliary variable cannot even complete one single iteration. In the case of good initialization of topical prior $\nmathbf \eta_d$ which leads to smooth sampling of auxiliary variables, the computational efficiency is still undesirable and we observed that for larger topical space such as K=40, the auxiliary variable stumbled again after some amount of iterations, indicating even good initialization will not ease the troublesome dependency relationship between the auxiliary variables and parameters in larger topical space. Unlike with the traditional probit model for which the computation of $\theta_d^k$ is virtually impractical for large $K$,the diagonal orthant approach makes this computation substantially faster ever for large $K$. The comparison of the computational speed of two essential sampling tasks between the multinomial probit model and digonal orthant probit model are shown as below in table 1 \eqref{tab:comp:efficiency:1}. \\
\begin{table}[!htbp]
\centering
\begin{tabular}{lrr}
\hline
Sampling Task (K=10) & {MNP} & {DO Probit} \\
Topic Distribution $\theta$ & $18.3$ & $0.06$ \\
Auxiliary variable $Y_d$ & ($108$ to NA) & $3.09$ \\
\hline
\end{tabular}
\begin{tabular}{lrr}
\hline
Sampling Task (K=20) & {MNP} & {DO Probit} \\
Topic Distribution $\theta$ & $63$ & $0.13$ \\
Auxiliary variable $Y_d$ & ($334$ to NA) & $3.39$ \\
\hline
\end{tabular}
\begin{tabular}{lrr}
\hline
Sampling Task (K=30) & {MNP} & {DO Probit} \\
Topic Distribution $\theta$ & $123$ & $0.21$ \\
Auxiliary variable $Y_d$ & ($528$ to NA) & $3.49$ \\
\hline
\end{tabular}
\begin{tabular}{lrr}
\hline
Sampling Task (K=40) & {MNP} & {DO Probit} \\
Topic Distribution $\theta$ & $211.49$ & $0.33$ \\
Auxiliary variable $Y_d$ & ($1785$ to NA) & $3.79$ \\
\hline
\end{tabular}
\caption{All the numbers in this table represent the processing time (in seconds), and are computed in R on PC using a parallel algorithm acting on 4 CPU cores.
NA here represents situations where it is impossible for the sampler to yield a random variate of the auxiliary variable corresponding the current topic allocation label that is also the maximum}
\label{tab:comp:efficiency:1}
\end{table}
In addition to the drastic improvement of the overall sampling efficiency, we noticed that the computational complexity for sampling the auxiliary variable and topic distribution is close to O(1) and O(K) respectively, suggesting that probit normal topic model now becomes an attainable and feasible tool of the traditional correlated topic model.\\
Central to topic modelling is the need to determine for a given corpus the optimal number of latent
topics. As it is the case for most latent variable models, this task can be formidable at times, and there
is no consensus among machine learning researchers as to which of the existing methods is the best.
Figure \eqref{fig:loglikelihood:full:1} shows the loglikelihood as a function of the number of topics
discovered in the model. Apart from the loglikelihood, many other techniques are commonly used such as perplexity,
harmonic mean method and so on.\\
\begin{figure}[!htbp]
\centering
\epsfig{figure=ctm-probit.eps, width=10cm, height=10cm}
\caption{Loglikelihood as a function of the number of topics}
\label{fig:loglikelihood:full:1}
\end{figure}
As we see, the optimal number of topics in this case is 30.
In table \eqref{tab:topics:1c}, we show a subset of the 30 topics uncovered where each topic is represented by the 10 most frequent words. It can be seen that our probit normal topic model is able to capture the co-occurrence of words within topics successfully. In figure 2, we also show the correlation structure between various topics which is the essential purpose of employing the correlated topic model. Evidently, the correlation captured intuitively reflect the natural relationship between similar topics.\\
\begin{table}[!htbp]
\centering
\begin{tabular}{llllllll}
\toprule
& {\bf Topic 25} & {\bf Topic 18} & {\bf Topic 23} & {\bf Topic 11} & {\bf Topic 1} & {\bf Topic 24} & {\bf Topic 27} \\
\hline
{\bf Word1} & court & company & bush & students & tax & fire & air \\
{\bf Word2} & trial & billion & senate & school & budget & water & plane \\
{\bf Word3} & judge & inc & vote & meese & billion & rain & flight \\
{\bf Word4} & prison & corp & dukakis & student & bill & northern & airlines \\
{\bf Word5} & convicted & percent & percent & schools & percent & southern & pilots \\
{\bf Word6} & jury & stock & bill & teachers & senate & inches & aircraft \\
{\bf Word7} & drug & workers & kennedy & board & income & fair & planes \\
{\bf Word8} & guilty & contract & sales & education & legislation & degrees & airline \\
{\bf Word9} & fbi & companies & bentsen & teacher & taxes & snow & eastern \\
{\bf Word10} & sentence & offer & ticket & tax & bush & temperatures & airport \\
\hline
\end{tabular}
\end{table}
\begin{table}[!htbp]
\centering
\begin{tabular}{llllllll}
\toprule
& {\bf Topic 6} & {\bf Topic 12} & {\bf Topic 20} & {\bf Topic 2} & {\bf Topic 22} & {\bf Topic 16} & {\bf Topic 15} \\
\hline
{\bf Word1} & percent & space & military & soviet & aid & police & dollar \\
{\bf Word2} & stock & shuttle & china & gorbachev & rebels & arrested & yen \\
{\bf Word3} & index & soviet & chinese & bush & contras & shot & rates \\
{\bf Word4} & billion & nasa & soldiers & reagan & nicaragua & shooting & bid \\
{\bf Word5} & prices & launch & troops & moscow & contra & injured & prices \\
{\bf Word6} & rose & mission & saudi & summit & sandinista & car & price \\
{\bf Word7} & stocks & earth & trade & soviets & military & officers & london \\
{\bf Word8} & average & north & rebels & treaty & ortega & bus & gold \\
{\bf Word9} & points & korean & hong & europe & sandinistas & killing & percent \\
{\bf Word10} & shares & south & army & germany & rebel & arrest & trading \\
\hline
\end{tabular}
\end{table}
\begin{table}[!htbp]
\centering
\begin{tabular}{llllllll}
\toprule
& {\bf Topic19} & {\bf Topic 14} & {\bf Topic 7} & {\bf Topic 4} & {\bf Topic 30} & {\bf Topic 8} & {\bf Topic 17} \\
\hline
{\bf Word1} & iraq & trade & israel & navy & percent & south & film \\
{\bf Word2} & kuwait & percent & israeli & ship & oil & africa & movie \\
{\bf Word3} & iraqi & farmers & jewish & coast & prices & african & music \\
{\bf Word4} & german & farm & palestinian & island & price & black & theater \\
{\bf Word5} & gulf & billion & arab & boat & cents & church & actor \\
{\bf Word6} & germany & japan & palestinians & ships & gasoline & pope & actress \\
{\bf Word7} & saudi & agriculture & army & earthquake & average & mandela & award \\
{\bf Word8} & iran & japanese & occupied & sea & offers & blacks & band \\
{\bf Word9} & bush & tons & students & scale & gold & apartheid & book \\
{\bf Word10} & military & drought & gaza & guard & crude & catholic & films \\
\hline
\end{tabular}
\caption{Representation of topics discovered by our method}
\label{tab:topics:1c}
\end{table}
\begin{figure}[!htbp]
\centering
\epsfig{figure=correlation.eps, width=10cm, height=8cm}
\caption{Graphical representation of the correlation among topics}
\label{fig:correlation:topics:1}
\end{figure}
\newpage
\section{Conclusion and Discussion}
In the context of topic modelling where many other researchers seem to have avoided it. By adapting the diagonal orthant probit model, we proposed a probit alternative to the logit approach to the topic modeling. Compared to the multinomial probit model we constructed, our topic discovery scheme using diagonal orthant probit model enjoyed several desirable properties; First,we gained the efficiency in computing the topic distribution $\nmathbf \theta_d^{k}$ ; Second, we achieved a clear and very straightforward and adaptable auxiliary variable sampling scheme that substantially reduced the strength of the dependence structure between auxiliary variables and model parameters, responsible for absorbing state in the Markov chain; Thirdly, as a consequence of good mixing, our approach made the probit model a viable and competitive alternatives to its logistic counterpart. In addition to all these benefits, our proposed method offers a straightforward and inherent conjugacy, which helps avoid those complicated sampling schemes employed in the logistics normal probit model.\\
In the Associated Press example explored in the previous section, not only does our method produce a better likelihood than the logistic normal topic model with variational EM, but also discovers meaningful topics along with underlying correlation structure between topics. Overall, the method we developed in this paper offers another feasible alternatives in the context of correlated topic model that we hope will be further explored and extended by many other researchers\\
Based on the promising results we have seen in this paper, the probit normal topic model opens the door for various future works. For instance, \cite{DBLP:conf:sdm:SalomatinYL09} proposed a multi-field correlated topic model by relaxing the assumption of using common set of topics globally among all documents, which can also be applied to the probit model to enrich the comprehensiveness of structural relationships between topics . Another potential direction would be to enhance the scalability of the model. Currently we used a simple distributed algorithm proposed by \cite{yao2009efficient} and \cite{Newman:2009:DAT:1577069.1755845} for efficient Gibbs sampling. The architecture for topic models presented by \cite{smola2010architecture} can be further utilized to reduce the computational complexity substantially while delivering comparable performance. Furthermore, a novel sampling method involving the Gibbs Max-Margin Topic \cite{journals/corr/ZhuCPZ13} will further improve the computational efficiency.\\
\bibliographystyle{chicago}
|
\section{Introduction}
The exponential family is a large class of probability distributions widely used in statistics. Distributions in this class possess many useful properties that make them useful for inferential and algorithmic purposes, especially with reference to maximum likelihood estimation. Despite this, when closed form solutions are not available, computational problems can arise in the fitting of a particular distribution in this family to real data. Starting values far from the optimum and/or bounded parameter spaces may cause common optimization algorithms like Newton-Raphson to fail. Convergence along a canyon and the phenomenon of \textit{parameter evaporation} (i.e. the algorithm is lost in a plateau and pushes the parameters to infinity) can be problematic too \citep[see][]{Transtum2012}. In such cases algorithms with more reliable convergence must be employed to find the maximum likelihood estimates.
The Levenberg-Marquardt algorithm was developed for the minimization of functions expressed as the sum of squared functions, usually squared errors, where its convergence properties have been theoretically demonstrated. Its performance is usually very good for functions that are mildly non-linear, and therefore many authors have attempted to adapt it to the maximization of the likelihood. In literature such adaptations have been based upon good heuristics and have been shown to provide reliable results, but to the best of our knowledge formal arguments for their convergence are still lacking. \citet{Smyth2002} gave an algorithm for REML (restricted or residual maximum likelihood) scoring with a Levenberg-Marquardt restricted step. It was applied to find estimates in heteroscedastic linear models with data normally distributed, and, taking advantage of the strong global convergence properties of the Levenberg-Marquardt algorithm, the author concluded that the result \textit{\ldots can therefore be expected to be globally convergent to a solution of the REML equations subject to fairly standard regularity conditions \ldots}. \citet{statmodSmyth} considered another adaptation of the Levenberg-Marquardt algorithm to fit a generalized linear model with secure convergence (for gamma generalized linear model with identity links or negative binomial generalized linear model with log-links, see the R package \textit{statmod}). \citet{aitchison2003statistical} used a Levenberg-Marquardt step for the maximization of the likelihood of two distributions on the simplex that belong to the exponential family. \citet{Stinis2005} used the standard Levenberg-Marquardt algorithm to minimize an error function related to a moment-matching problem arising in maximum likelihood estimation of distributions in the exponential family.
In this paper we evaluate a class of adaptations of the Levenberg-Marquardt algorithm to find the maximum likelihood estimates in the exponential family and we give formal proof of convergence for such an algorithm in the general case of generalized linear models with natural link. We examine its performance on real and simulated data.
\section{Exponential Family}
We consider the problem of maximum likelihood estimation in the exponential family. With $y$ we denote a $K$-dimensional observation. The densities of the distributions in the exponential family can be written in their natural parametrization as
\begin{equation}\label{EF}
f(y|\theta) = \exp \left\{ y' \theta - b(\theta) + c(y) \right\}
\end{equation}
where the $K$-dimensional vector $\theta$ is the natural parameter, $b(\theta)$ is the log-partition function and $\exp\{c(y)\}$ is the base measure. The natural parameter space is the set of natural parameters for which the log-partition function is finite. The covariance matrix of a random vector with distribution in the exponential family is usually supposed to be positive definite in the interior of the natural parameter space.
The Newton-Raphson algorithm and the Fisher scoring algorithm are two algorithms that are commonly used to find maximum likelihood estimates. They are equivalent in the multivariate exponential family with natural parametrization because for a sample of $n$ i.i.d. observations the Hessian matrix of the loglikelihood is $n$ times the Hessian of the log-partition function and is independent of the data. Further, for densities in the exponential family the Hessian is related to the covariance matrix and it is invertible, insuring good convergence properties for the algorithms. In what follows we denote with $l$ the loglikelihood function, with $s$ the corresponding score function (the gradient of the loglikelihood function with respect to the parameters) and with $H_l$ the Hessian of the loglikelihood. With this notation a Newton-Raphson iteration can be expressed as
\begin{equation}\label{NRiteration}
\theta^{(t+1)} = \theta^{(t)} - [{H_l(\theta^{(t)})}]^{-1} s(\theta^{(t)}).
\end{equation}
The algorithm can also be written in the equivalent form of IRLS, Iteratively Reweighted Least Squares (see appendix A). In appendix A we provide all the necessary notation and results to extend the Newton-Raphson algorithm to the case of generalized linear model with natural link. They will be the basis for the proof in appendix B.
\section{A class of algorithms for maximum likelihood estimation}\label{sec:LM}
The Levenberg-Marquardt algorithm \citep{Levenberg1944,Marquardt1963} is an adaptive algorithm that is used to solve least squares problems. It is based on a modification of the Gauss-Newton method. Specifically, the Levenberg-Marquardt algorithm uses in its iterations a penalized version of $J'J$ where $J$ is the Jacobian matrix of the target function, which guarantees that the matrix can be inverted. The penalization is opportunely tuned through a damping parameter. In practice the Hessian of the function to be minimized is replaced by
\begin{equation}\label{dampLM}
J'J + \gamma \mathrm{diag}(J'J)
\end{equation}
where $\gamma$ is the damping parameter. Here we show how similar ideas can be applied directly to the maximization of the loglikelihood in exponential families. In appendix B we give a formal proof of a local convergence property for the algorithm outlined below when this is applied to generalized linear model with natural link. This case includes the case of distributions in the exponential family as can easily be seen considering design matrices equal to identity matrices (see appendix A).
The algorithm for maximum likelihood estimation in the exponential family that we evaluate in this work can be thought as a penalized version of the Newton-Raphson algorithm, with a penalization similar to the one used in the Levenberg-Marquardt algorithm, \textit{i.e.}, Equation (\ref{dampLM}). The iterations are given by
\begin{equation}\label{LMiteration}
\theta^{(t+1)} = \theta^{(t)} - [{H_l(\theta^{(t)})} + \gamma^{(t)}{P(\theta^{(t)})}]^{-1} s(\theta^{(t)})
\end{equation}
where $\gamma^{(t)}$ is a positive damping parameter and $P(\theta^{(t)})$ is a symmetric negative definite matrix. In practice $P(\theta^{(t)})$ is always a diagonal matrix to be used as penalization. The proposal of \citet{Levenberg1944} for the least squares problem was to use as penalization the identity matrix $I$. For maximum likelihood estimation the corresponding proposal is $P(\theta^{(t)})=-I$. The proposal of \citet{Marquardt1963} instead corresponds to $P(\theta^{(t)})=\mathrm{diag} H_l{(\theta^{(t)})}$. In this paper we use the latter form which has the advantage of better following the curvature of the function being maximized. The damping parameter will play a central role to make the algorithm adaptive. This can decrease the penalty, and thus speed up the convergence. Moreover, for a bounded parameters space, a careful tuning of the damping parameter is required to avoid large steps that could bring the parameters outside the allowed region. The penalization $\gamma^{(t)}{P(\theta^{(t)})}$ can be used to ensure that $[{H_l(\theta^{(t)})} + \gamma^{(t)}{P(\theta^{(t)})}]$ is invertible. In regular exponential families the inversion of $H_l$ is usually possible, but such a matrix can still be poorly conditioned and therefore algorithms for matrix inversion can fail or perform poorly. The penalization can avoid this problem. Other features related to the implementation of the algorithm will be discussed in the next section.
The iterations in Equation (\ref{LMiteration}) are a sound basis for a stable algorithm. In appendix B we give a formal proof for the convergence of the algorithm.
\subsection{Damping Parameter and Stopping Criteria}\label{secDampPar}
The damping parameter $\gamma$ in Equation (\ref{LMiteration}) influences the step size of the iterations. As $\gamma$ reaches zero, the algorithm reduces to the Newton-Raphson algorithm. Since the Newton-Raphson algorithm has a quadratic rate of convergence in the neighbourhood of the maximum we want a small $\gamma$ in such a situation. However, if the current iteration is far from the maximum, small steps can avoid some of the problems discussed in the previous section. These steps are provided by a large value of $\gamma$. To achieve the desired changes in the damping parameter we adopt the strategy proposed by \cite{Nielsen1998}. We use the following gain function:
\begin{displaymath}
\varrho^{(t+1)} = \frac{l(\theta^{(t+1)})-l(\theta^{(t)})}{ -0.5(\theta^{(t+1)}-\theta^{(t)})'H_l(\theta^{(t)})(\theta^{(t+1)}-\theta^{(t)})}.
\end{displaymath}
With this function we are comparing the actual increase or decrease of the loglikelihood function with the second order term of its Taylor approximation. The denominator is always positive (see appendix A) and therefore a positive value of the gain function indicates that we are moving in the right direction. With a large positive value of the gain function we can reduce the parameter $\gamma^{(t+1)}$. In this way we approximate the Newton-Raphson algorithm. A small value of $\varrho^{(t+1)}$ indicates instead that the Taylor's approximation is not working very well and in this case it is better to penalize the steps by increasing $\gamma^{(t+1)}$.
The Hessian matrix in the denominator of the gain function can be replaced by its actual approximation in the squared brackets of Equation (\ref{LMiteration}). We do this when small values of the Hessian lead to values close to zero in the denominator of the gain function.
The damping parameter can be updated as follows:
\begin{equation}\label{DampPar}
\nonumber \gamma^{(t+1)} = \gamma^{(t)} \max \left\{\frac{1}{3},1 - (2 \varrho^{(t+1)} - 1)^3 \right\}I_{\varrho^{(t+1)} > 0} + \gamma^{(t)} 2 I_{\varrho^{(t+1)} \leq 0}
\end{equation}
where $\gamma^{(0)}=1$. This value for $\gamma^{(0)}$ proposes therefore an initial penalized step. Other choices of course are possible. The above function is similar to the one suggested by \citet{IMM2004}, it is positive and is continuous in $\varrho^{(t+1)}$.
We stop the algorithm as soon as one of three criteria is reached: if the norm of the score is very close to zero: $ \parallel s(\theta^{(t)}) \parallel < \epsilon_1$; if the changes in the parameters in successive iterations are very small: $ \parallel \theta^{(t+1)} - \theta^{(t)} \parallel < \epsilon_2 ( \parallel \theta^{(t)} \parallel + \epsilon_2) $; if the number of iterations is greater than a pre-established threshold, \textrm{maxit}. In this work we consider $\epsilon_1=\epsilon_2=10^{-8}$ and \textrm{maxit}=1000. The algorithm is considered to have reached convergence only if the final estimates are inside the natural parameter space.
\section{Examples}
In this section we apply the algorithm described above to two distributions belonging to the exponential family. Specifically, we focus on two distributions for the analysis of compositional data, \textsl{i.e.}, positive data that sum up to one.
Within numerical accuracy, all the algorithms we compared reach the same optimum upon convergence (the distributions in the natural exponential family are convex), therefore we focus our comparisons on the computational efficiency and stability.
\subsection{First case: Dirichlet distribution}
If we denote with $\alpha = (\alpha_1,\ldots,\alpha_K)$ the vector of parameters of the Dirichlet distribution then its loglikelihood for $n$ i.i.d. observations can be written as
\begin{equation}\label{lDirichlet}
l(\alpha)=n \ln \Gamma(\sum_{k=1}^K \alpha_k) - n\sum_{k=1}^K \ln \Gamma(\alpha_k) + \sum_{k=1}^K (\alpha_k - 1)\sum_{r=1}^n \ln y_{rk}.
\end{equation}
With this notation the parameters must be greater than zero. The transformation $\theta_k = \alpha_k - 1$ gives the natural parameters. The Dirichlet distribution can arise as a transformation of independently distributed gamma variables and as a consequence has independence properties that can bound its use \citep[see][p. 59-60]{aitchison2003statistical}. For high dimensional data however it is reasonable to assume that many components are almost independent and therefore we study its performance in this situation.
We compare the algorithm in (\ref{LMiteration}) (henceforth, simply LM) with the Newton-Raphson (NR) algorithm and a fixed point iteration (FPI) algorithm. The former has quadratic convergence in a neighbourhood of the maximum but can often fail, whereas the latter is very stable but can be slow. We also consider an implementation of (\ref{LMiteration}) with a fixed damping parameter \citep[see][]{giordan2014} to evaluate the differences with the suggested adapting damping parameter. As starting values for the algorithms we employ four different strategies: the method of moments and the proposals based upon the works of \citet{Dishon1980}, \citet{Ronning1989} and \citet{Wicker2008}. The method of moments has the advantage of simplicity but works only on the marginal distributions and can give estimates outside the natural parameter space. The method of \citet{Dishon1980} is an improvement that considers information from all the data for the estimation of each single parameter. The proposal of \citet{Ronning1989} always gives initial parameters inside the parameter space, whereas \citet{Wicker2008} developed an approximation of the likelihood that is useful for high-dimensional data.
\subsubsection{Simulations}
In this simulation study we generate data from a Dirichlet distribution with dimension 1000. The number of simulated samples is equal to 20. To avoid the rounding of many randomly generated values to zero due to machine precision we simulate data from distributions with large parameters. Specifically we consider an hypothetic sum of the parameters, $\sum \alpha_k$, ranging from 10000 to 50000 with step size of 2000 and we draw each parameter from a uniform distribution between $\sum \alpha_k/K -2$ and $\sum \alpha_k/K +2$ with $K=1000$.
In \figurename ~\ref{fig:Sim} the convergence rate for each combination of starting values and methods (upper panel) and the mean number of iterations required for convergence (lower panel) are shown. NR has convergence problems when the starting values are given by the methods of moments. The other two methods instead are very stable. They provide convergence with all the starting values strategies. The increasing number of iterations needed for FPI (lower panel) suggest that by raising \textrm{maxit} we can always have convergence for this method. NR, as expected, shows a very fast convergence. FPI requires a large number of iterations for convergence. The LM implementation with the fixed damping parameter has a performance very close to that of FPI. The adapting damping parameter instead brings the required number of iterations for convergence very close to the number of iterations that NR is using, while having greater stability than the FPI algorithm.
\begin{figure*}[!p]
\centering
\includegraphics[scale=0.50,angle=270]{Stability1Add.eps}
\includegraphics[scale=0.50,angle=270]{Efficiency1Add.eps}
\caption{simulations from the Dirichlet distribution. We compare three algorithms (LM, FPI and NR) and four different strategies for the starting values (Wicker, Dishon, Ronning and moments). With plus we report the results of LM using a fixed damping parameter.}\label{fig:Sim}
\end{figure*}
\subsubsection{Apple data set}
We now want to compare the performance of the algorithms on real data. We do this only for the two best algorithms from the previous simulation study: LM with the adapting damping parameter and NR. The apple data set \citep{Franceschi2012} contains mass spectrometric measurements and it is publicly available in the R package BioMark \citep{Wehrens2012}. We consider a subset of twenty samples (10 controls and 10 spiked-in samples) of positive ionization data. After an appropriate normalization to get compositional data the final data set has 1602 variables. The convergence results are reported in \tablename ~\ref{tab:Apple}.
The LM algorithm always reaches convergence for all the starting values strategies whereas NR fails to converge for two out of four initialization strategies. The number iterations required for reaching convergence is very similar to the number required by the NR algorithm when it does reach convergence. Therefore the LM variant is both stable and fast.
\begin{table*}
\centering
\caption{results for the apple data set. LM converges for all initailization methods, whereas NR fails to converge in two out of four cases.}\label{tab:Apple}
\begin{tabular}{lcccc}
\toprule
& convergence NR & convergence LM & n iterations NR & n iterations LM \\
\midrule
Wicker & no & yes & - & 11 \\
Dishon & yes & yes & 21 & 22 \\
Ronning & yes & yes & 30 & 31 \\
moments & no & yes & - & 55 \\
\bottomrule
\end{tabular}
\end{table*}
\subsection{Second Case: Aitchison distribution}
For low dimensional data the independence structure of the Dirichlet distribution can be too strong to allow a good fitting. \citet[][p. 310-315]{aitchison2003statistical} introduced a more flexible alternative that arises as a generalization of the Dirichlet distribution and the additive logistic normal distribution. We refer to it as the \textit{Aitchison distribution} in this paper. Its parametric form is again in the exponential family and the corresponding loglikelihood can be expressed as:
\begin{equation}\label{lAitchison}
l= -n \log c(\alpha,\beta) + \sum_{i=1}^{K}(\alpha_i - 1)U_i + \sum_{i=1}^{K-1}\sum_{j=i+1}^{K} \beta_{ij}V_{ij}
\end{equation}
where
$$
U_i=\sum_{r=1}^n \log y_{ri}
$$
and
$$
V_{ij}= -\frac{1}{2}\sum_{r=1}^n (\log y_{ri} - \log y_{rj})^2
$$
with $i=1,\ldots,K-1; j=i+1,\ldots,K$. We remark that the Aitchison distribution has a more complex structure than the additive logistic normal distribution. Contrary to the additive logistic normal and the normal on simplex where a change in coordinate allows to work with standard techniques for the multivariate normal distribution \citep[see, for example,][]{Figueras2008}, the Aitchison distribution requires computational tools to find the maximum likelihood estimates.
It is evident from equations (\ref{lDirichlet}) and (\ref{lAitchison}) that the price to pay for the generalization is the increased number of parameters. This can be quite high, even for a data set of modest dimension. For a compositional data set with $K$ variables the number of parameters to be estimated is:
$$
K + \frac{1}{2}\left( K(K-1) \right)= \frac{1}{2}\left( K(K+1) \right).
$$
Moreover, the normalizing factor $c(\alpha,\beta)$ has no closed form and therefore it and its derivatives must be evaluated numerically. The computation of the loglikelihood is therefore particularly demanding even for a low dimensional data set. For example, for a compositional data set with 5 variables the Dirichlet distribution requires only 5 parameters while the Aitchison distribution requires 15. Further, the lack of a closed form for the Hessian matrix implies the numerical evaluation of 120 integrals in each iteration.
For the Dirichlet distribution the natural parameter space is the set of vectors with positive elements $\{ \alpha =(\alpha_1,\ldots,\alpha_K) \ | \ \alpha_i > 0 \ \mathrm{for} \ i=1,\ldots,K \}$. For the Aitchison distribution a similar description of the natural parameter space is not available. In \citet[][p. 311-312]{aitchison2003statistical} two different restrictions are proposed to obtain proper density functions. However, these conditions are sufficient but not necessary. In practice the normalizing constant of a proper density is finite and for a current vector of parameters this must be numerically evaluated in the algorithms. Therefore if the algorithms converge to finite values these must be in the natural parameter space because the corresponding normalizing constant (log-partition function) must be finite. To accurately calculate the normalizing factor $c(\alpha,\beta)$ we use Gauss-Hermite integration following the suggestions in \citet[][p 314-315]{aitchison2003statistical}.
The closed-form solutions for the maximum likelihood estimates of the additive logistic normal distribution will be used as starting values for the algorithms. These estimates are simply the sample mean and variance of the additive log-ratio transformation of the original compositional data \citep[see][p. 113, 313-314]{aitchison2003statistical}.
\subsubsection{Simulations}
For the Aitchinson distribution, the calculation of the gradient requires the calculation of several complex numerical integrals, and we therefore only examined a case with a low number of variables.
To generate samples from the Aitchison distribution we used the R package \textit{compositions} version 1.30-1 \citep{van2013analyzing}. To ensure good starting values we used 2000 simulations where the sample covariance matrix from the log-ratio transformation was positive definite. The number of samples in each simulated data set was 20. The dimensions of the compositional data sets were 3 and 5, corresponding to parameter vectors ($\alpha,\beta$) of length 6 and 15 respectively. The parametrization used in the above package is slightly different from the one used in the paper; for the simulations we considered parameters according to the example of \citet[][p. 66]{van2013analyzing}.
A summary of the simulations is given in \tablename ~\ref{AitSim}. The number of succesful convergences dramatically increases when we use LM instead of NR, and the increase in the number of iterations required (in the subset of cases when NR does converge) is very modest, making it an obvious improvement for maximum likelihood estimation.
\begin{table*}
\centering
\caption{simulations from the Aitchison distribution. The first column indicates the number of parameters to be estimated. In columns two and three the number of convergences over 2000 simulations are given for the NR algorithm and the LM algorithm, respectively. Similarly, the last two columns show the mean numbers of iterations in case of convergence.}\label{AitSim}
\begin{tabular}{ccccc}
\toprule
n parameters & n convergences NR & n convergences LM & n iterations NR & n iterations LM \\
\midrule
6 & 22 & 879 & 7.95 & 13.32 \\
15 & 12 & 745 & 8.75 & 12.90 \\
\bottomrule
\end{tabular}
\end{table*}
\subsubsection{Applications to known data sets}
We now apply the NR and LM algorithms to four data sets and we compare their performance. The data sets are publicly available in the R package \textit{robCompositions} \citep{Templ2011} and/or the package \textit{compositions}. They are briefly described below (more information is available inside the actual packages):
\begin{itemize}
\item[] Data set 1, skye lavas. It is a data set with 3 variables: magnesium, sodium-potassium and iron. We used the variable \textit{iron} for the log-ratio transform.
\item[] Data set 2, arctic lake. It is a data set with 3 variables: clay, silt and sand. We used the variable \textit{sand} for the log-ratio transform.
\item[] Data set 3, machine operators. It is a data set with 4 variables: high quality production, low quality production, setting and repair. We used the variable \textit{repair} for the log-ratio transform.
\item[] Data set 4, expenditures. It is a data set with 5 variables: housing, food stuffs, alcohol, services and other. We used the variable \textit{other} for the log-ratio transform.
\end{itemize}
In \tablename ~\ref{AitDataSets} we summarize the convergence results for the Aitchison distribution. Only LM is able to give convergence for the arctic lake data set. Both algorithms failed to converge for the expenditures data set. For the remaining data sets both algorithms converge to the same parameters although the number of required iterations by LM is slightly greater than those required by NR.
\begin{table*}
\centering
\caption{analysis of 4 data sets with the Aitchison distribution. The first column indicates the data sets. In the columns two and three we see if the NR algorithm and the LM algorithm have converged. In the last two columns we give the number of iterations used for the convergence.}\label{AitDataSets}
\begin{tabular}{lcccc}
\toprule
& convergence NR & convergence LM & n iterations NR & n iterations LM \\
\midrule
skye lavas & yes & yes & 5 & 14 \\
arctic lake & no & yes & - & 12 \\
machine operators & yes & yes & 5 & 14 \\
expenditures & no & no & - & - \\
\bottomrule
\end{tabular}
\end{table*}
\section{Discussion and conclusions}
In this paper we have investigated the use of the Levenberg-Marquardt algorithm to find the maximum likelihood estimates of distributions in the exponential families. We have given formal proof of convergence to the optimum for a class of possible adaptations and we have shown through real and simulated data that the LM-variant outperforms other algorithms in many settings.
The penalization used in the paper is related to the curvature of the loglikelihood and it ensures a well-conditioned negative definite matrix in each iteration of the algorithm. This provide a stable algorithm at the price of an increased number of iterations for the convergence (using as reference the Newton-Raphson algorithm). However, the damping parameter used in the penalization is adaptive and therefore it can speed up the convergence of the algorithm. We have shown in a simulation study that the difference with a fixed damping parameter can be substantial. We have compared the efficiency of the algorithms taking into account the number of iterations rather than the time. Since Equation (\ref{LMiteration}) is essentially Equation (\ref{NRiteration}) with the added computation for the penalty and the Fixed Point Iteration algorithm requires a much greater computational effort than both other algorithms, we can guarantee that the order of the efficiency comparison is in any case preserved.
In this work we have focused our attention on distributions for the analysis of compositional data. The computational performance improvements, however, are expected to hold also for other distributions in the exponential family because the algorithm is not related to compositional data. In particular we have used the Dirichlet distribution and the Aitchison distribution to analyze high and low dimensional data respectively. For both distributions the adaptation of the Levenberg-Marquardt algorithm has shown substantial stability advantages over other algorithms. Despite this, the number of required iterations is very low thanks to the adaptive damping parameter. The algorithm studied in the paper is therefore a powerful computational tool for maximum likelihood estimation in the exponential family.
|
\section{Introduction}
\label{sec:introduction}
We are in the golden age of observational cosmology, in which General Relativity (GR) is being put to the test at the largest observable distances~\cite{Jain:2010ka,Joyce:2014kja}. Consequently, it has become an important task to develop consistent competitor theories which modify $\Lambda$CDM predictions on cosmological scales. Moreover, there is still no fundamental understanding of the dark sector, which constitutes the main part of the energy budget in the $\Lambda$CDM model. The most pressing issue from a theory standpoint is the cosmological constant problem (see \cite{Weinberg:1988cp} for a seminal work and \cite{Burgess:2011va} for a more recent discussion). This provides a strong motivation to look for consistent infrared modifications of gravity.
A prominent candidate is the model of \textit{brane induced gravity} (BIG) \cite{Dvali:2000hr, Dvali:2000xg} according to which our four dimensional universe (the brane) and all its matter content is localized in a $d$-dimensional infinite space-time (the bulk). Despite the fact that the extra dimensions are infinite in extent, 4D gravity is nevertheless recovered at short enough distances on the brane, thanks to an intrinsic Einstein-Hilbert term (or brane induced gravity term) on the brane. This results in a modification of gravity characterized by a single length scale $r_c$ which discriminates between two gravitational regimes: a conventional 4D regime on scales $\ell\ll r_c$, for which the Newtonian potential is proportional to $1/r$ up to small corrections; and a $d$-dimensional regime on scales $\ell\gg r_c$, for which gravity on the brane is effectively weakened and the scaling becomes $1/r^{d-3}$. In order to be in accordance with gravitational measurements on solar system scales, the \textit{cross-over} scale $r_c$ has to be large enough, {\it e.g.}, for $d=5$ lunar laser ranging experiments demand $r_c^{(5)} \gtrsim 0.04 H_0^{-1}$~\cite{Afshordi:2008rd}. Thus, cosmology represents the ideal playground for testing these theories.
Brane induced gravity models are interesting also for other reasons. At the linear level, the effective 4D graviton is a resonance, {\it i.e.}, an infinite superposition of massive graviton states. Historically it turned out to be notoriously difficult to give a mass to the 4D graviton on a nonlinear level without introducing Boulware-Deser ghost instabilities (for recent reviews, see~\cite{Hinterbichler:2011tt,deRham:2014zqa}). This has been achieved recently with dRGT gravity~\cite{deRham:2010kj}. Extra dimensional constructions, such as BIG, offer promising arenas to devise ghost-free examples. Another motivation comes from the {\it degravitation} approach to the cosmological constant problem~\cite{Dvali:2002pe, Dvali:2002fz, ArkaniHamed:2002fu, Dvali:2007kt, deRham:2007rw}. The massive/resonant graviton leads to a weakening of the gravitational force law at large distances, which makes gravity effectively insensitive to a large cosmological constant. There are linear~\cite{deRham:2007rw} and nonlinear~\cite{Charmousis:2001hg} indications for that claim.
The best-known and most extensively studied example is the Dvali-Gabadadze-Porrati (DGP) model~\cite{Dvali:2000hr}, corresponding to $d=5$. The cross-over scale in this case is given by $r_c^{(5)} = \frac{M_{\rm Pl}^2}{2M_5^3}$, where
$M_5$ is the bulk Planck scale. For cosmology, the DGP setup gives rise to a modified Friedmann equation~\cite{Deffayet:2000uy}, $H^2 \pm \frac{H}{r_c^{(5)}} = \frac{\rho}{3M_{\rm Pl}^2}$, featuring an additional term controlled by $r_c$. Accordingly, the modification can be neglected for early times and large curvature ($H \gg 1/r_c^{(5)}$), whereas it becomes significant at late times and small curvature ($H \lesssim 1/r_c^{(5)}$). The plus and minus sign correspond to two different branches of solutions, the ``normal'' and the ``self-accelerating'' branch, respectively. The former is characterized by a weakening of gravity since the energy density gets effectively reduced, while the latter describes a gravitational enhancement. The self-accelerated branch is widely believed to suffer from perturbative ghost instabilities~\cite{Luty:2003vm,Nicolis:2004qq,Koyama:2005tx,Charmousis:2006pn,Gregory:2007xy,Gorbunov:2005zk}. The normal branch is perturbatively stable. Confronting DGP with cosmological observations yields a rather stringent bound on the cross-over scale: $r_c^{(5)} \gtrsim 3 H_0^{-1}$~\cite{Lombriser:2009}.
A natural generalization of the DGP model are higher-codimension scenarios ($d>5$) \cite{Dvali:2000xg}. Several difficulties have impeded their development:
\begin{itemize}
\item According to claims in the literature, the model propagates a linear ghost on a Minkowski background~\cite{Dubovsky:2002jm,Hassan:2010ys}, which questions the quantum consistency of the whole theory.
\item Bulk fields are generically divergent at the position of a higher-codimension brane and require a regularization prescription.
\item A non-trivial cosmology on the brane implies the existence of gravitational waves which are emitted into the bulk. (In $d=5$, the symmetries of the geometry imply a static bulk, because there is a generalization of Birkhoff's theorem to planar symmetry~\cite{Taub:1951}. However, no such theorem exists for cylindrical symmetry, and Einstein-Rosen waves \cite{EinsteinRosen1937} are in fact a counter-example.) Including these waves in the dynamical description makes it much more difficult to solve the full system.
\end{itemize}
The first point, which clearly would be the most severe, was recently proven to be wrong~\cite{Berkhahn:2012wg}. Through a detailed constraint analysis, it was shown rigorously in~\cite{Berkhahn:2012wg} that the would-be ghost mode is {\it not} dynamical and is instead subject to a constraint. This is analogous to the conformal mode of standard 4D GR. For $d=6$ the positive definiteness of the Hamiltonian was explicitly shown in~\cite{Berkhahn:2012wg}. Consequently, in a weakly coupling regime on a Minkowski background the model is healthy. This result offered a new window of opportunity for investigating consistently modified cosmologies at the largest observable scales.
In the present paper we explore cosmological solutions in the simplest case: brane induced gravity in $ d=6 $ dimensions. Those solutions are obviously interesting for observational purposes, but they also test the non-perturbative stability of the model.
To overcome the second issue listed above, we introduce in Sec.~\ref{sec:model} a regularization which replaces the infinitely thin brane by a hollow cylinder of finite size $R$. We stabilize this size by introducing an appropriate azimuthal pressure. The microscopical origin of this pressure component is not specified, but we check {\it a posteriori} whether the required source is physically reasonable ({\it i.e.}, whether it satisfies the standard energy conditions).
We first check the consistency of our framework by deriving known solutions for a static cosmic string in 6D in Sec.~\ref{sec:static_sol}. Based on these solutions the geometry of the setup is illustrated and a distinction between sub- and super-critical branes is motivated.
According to the third issue listed above, which is discussed in more detail in Sec.~\ref{sec:Brane_Bulk}, a key feature of the higher codimensional models is the existence of bulk gravitational waves which are emitted by the brane and affect its dynamics. For $d=6$ they correspond to a higher-dimensional generalization of Einstein-Rosen waves. Consequently, we must resort to numerics, introduced in Sec.~\ref{sec:num_impl}, to find the most general solutions.
We then solve Einstein's field equations in the bulk in the presence of FRW matter (and brane induced gravity terms) on the brane and present the results in Sec.~\ref{sec:num_sol}. We stress that these solutions have been derived from the full system of nonlinear Einstein equations without making any approximations or additional assumptions other than having FRW symmetries on the brane and a source-free bulk. This result makes it possible for the first time to discuss the phenomenological viability of the six dimensional BIG model with respect to cosmological observations.
Depending on the model parameters, we find two qualitatively different classes of solutions:
\begin{itemize}
\item { \it Degravitating} solutions for which the system approaches the static cosmic string solution, {\it i.e.}, the 4D Hubble parameter becomes zero despite the presence of a non-vanishing on-brane source.
\item { \it Super-accelerating} solutions for which Hubble grows unbounded for late times.
\end{itemize}
The solution of the first type constitutes the first example of a dynamically realized degravitation mechanism. Accordingly, the brane tension is shielded from a 4D observer by exclusively contributing to extrinsic curvature.
We dismiss the second type due to its pathological run-away behavior. In addition, the effective energy density that sources 6D gravity turns negative for these solutions. This bears strong resemblance with the self-accelerating branch in the DGP model and thus questions their perturbative quantum stability.
It is shown that the degravitating and super-accelerating solutions are separated by a physical singularity. Thus, it is not possible to dynamically evolve from one regime to the other. We derive an {\it analytic} expression for the separating surface in parameter space. This in turn allows us to derive a necessary condition to be in the degravitating regime:
\begin{equation}\label{eq:stab_bound}
\left( H r_c\right)^2 < \frac{3}{2} \left|H\right| R\,,
\end{equation}
with $2 \pi R$ the circumference of the cylinder and $r_c$ the crossover scale\footnote{Here and henceforth, $ r_c $ refers to the 6D crossover scale, defined below in \eqref{eq:defCrossover}.}. However, a phenomenologically viable solution has to fulfill two requirements: First, $H r_c \gg 1$ for early times which ensures that the deviation from standard Friedmann cosmology is small. Second, $H R \ll 1$ in order to be insensitive to unknown UV physics that led to the formation of the brane. Obviously, these two conditions are incompatible with the bound~\eqref{eq:stab_bound}. {\it As a consequence of these considerations, the degravitating solutions are ruled out phenomenologically.}
We conclude in Sec.~\ref{sec:conclusion} with some remarks on super-critical energy densities.
A number of technical results have been relegated to a series of appendices. In particular, we repeat the analysis with a different regularization scheme in Appendix~\ref{ap:dynReg} to check the insensitivity of our results to the regularization details.
We adopt the following notational conventions: capital Latin indices $ A, B, \dots $ denote six-dimensional, small Latin indices $ a, b, \ldots $ five-dimensional, and Greek indices $ \alpha, \beta, \ldots $ four-dimensional space-time indices. Small Latin indices $ i, j, \ldots $ run over the three large spatial on-brane dimensions and corresponding vectors are written in boldface.
The space-time dimensionality $ d $ of some quantity $ Q $ is sometimes made explicit by writing $ Q^{(d)} $.
Our sign conventions are ``$ +++ $'' as defined (and adopted) in~\cite{Misner}.
We work in units in which $ c = \hbar = 1 $.
\section{The model}
\label{sec:model}
The action of the BIG model in $D=4+n$ dimensions is the sum of three terms:
\begin{align}
\label{eq:ActionBIG}
\mathcal{S}=
\mathcal{S}_{\rm EH}+
\mathcal{S}_{\rm BIG}+
\mathcal{S}_{\rm m}[h]\;.
\end{align}
The first term,
\begin{align}
\mathcal{S}_{\rm EH}=
M_{D}^{D-2}\int {\rm d}^D X\;\sqrt{-g}\; \mathcal{R}^{(D)}\,,
\end{align}
describes Einstein-Hilbert gravity in $D$ infinite space-time dimensions. The bulk Planck scale is denoted by $M_D$.
The bulk is assumed to be source-free; in particular, the bulk cosmological constant is set to zero for simplicity.
The second term is the induced gravity term on a codimension-$n$ brane:
\begin{align}
\label{eq:S_BIG_delta_0}
\mathcal{S}_{\rm BIG}=M_{\rm Pl}^2\int {\rm d}^4 x\,\sqrt{- h}\; \mathcal{R}^{(4)} \,.
\end{align}
This describes intrinsic gravity on the brane, with $h_{\mu\nu}$ denoting the induced metric. To match standard GR in the 4D regime, $M_{\rm Pl}$ is identified as the usual 4D Planck scale.
From the effective field theory point of view, the BIG term can be thought to arise from integrating out heavy matter fields on the brane. The last term in~\eqref{eq:ActionBIG}, $\mathcal{S}_{\rm m}[h]$,
is the action for matter fields localized on the brane, which by definition couple to $h_{\mu\nu}$.
Henceforth we will focus on $D=6$, corresponding to the codimension $n=2$ case.
\subsection{Regularization schemes}
\label{sec:regul_schemes}
In general, a localized codimension-two source leads to a singular geometry, {\it i.e.}, the bulk metric diverges logarithmically at the position of the brane. This is well known for static solutions, reviewed in Sec.~\ref{sec:static_sol}. For the pure tension case, the space-time develops a conical singularity---the bulk geometry stays flat arbitrarily close to the brane but diverges exactly at the brane. For more general static and non-static solutions we have to deal with curvature singularities other than the purely conical one. These singularities can be properly dealt with by introducing a certain brane width.
In this work, we adopt a regularization which consists of blowing up the brane to a circle of circumference $2 \pi R$~\cite{Kaloper:2007ap, Burgess:2008yx}. In other words, the brane is now a codimension-one object, with topology ${\cal M}_4\times \mathcal{S}_1$. The matter fields are smeared out on the $\mathcal{S}_1$. This amounts to the substitution
\begin{equation}
\label{reg_sub}
\mathcal{S}_{\rm BIG} ~\longrightarrow~ M_5^3\int_{\mathcal{M}_4 \times \mathcal{S}_1}\!\!\!\!\! {\rm d}^5 x\,\sqrt{- h^{(5)}}\; \mathcal{R}^{(5)} \;,
\end{equation}
where $M_5^3=\frac{M_{\rm Pl}^2}{2 \pi R}$, and $h^{(5)}_{ab}$ is the five dimensional induced metric.
Furthermore, in the main body of the paper, we follow a \textit{static regularization} scheme, which makes the evolution completely insensitive to the geometry inside the regularized brane.
This scheme can be viewed from two, equivalent perspectives:
\begin{itemize}
\item The brane is a boundary of space-time, and there is no interior geometry to speak of. This is the {\it hollow cylinder} perspective. In this case, the equations of motion consist of Einstein's field equations in the exterior, supplemented by Israel's junction conditions~\cite{Israel:1966, Israel:1967} at the brane,
\begin{eqnarray}
\nonumber
T^{(5) a}_{\hphantom{(5)a} b}-M_5^3 G^{(5)a}_{\hphantom{[h]a} b} &=& M_6^4 \big(K^{~c}_{{\rm out}\,c} \delta^{a}_{\hphantom{a}b} - K_{{\rm out}\,b}^{~a}\big) \\
&-& \frac{1}{R} \left( \delta^{a}_{\hphantom{a}b} - \delta^a_{~\phi}\delta^{\phi}_{\hphantom{\phi}b}\right)\,,
\label{israel}
\end{eqnarray}
where $K_{{\rm out}\,ab}$ is the extrinsic curvature tensor. In the second line, we have extracted from $T^{(5) a}_{\hphantom{(5)a} b}$ a cosmological constant
along $\mathcal{M}_4$. This is necessary to ensure that the deficit angle vanishes when $T^{(5) a}_{\hphantom{(5)a} b} \rightarrow 0$.
\item The brane has an interior geometry, such that the junction condition now becomes
\begin{equation}
T^{(5) a}_{\hphantom{(5)a} b}-M_5^3 G^{(5)a}_{\hphantom{[h]a} b} = M_6^4 \big([K^c_{~c}] \delta^{a}_{\hphantom{a}b} - [K_{~b}^{a}]\big)\,,
\label{israel2}
\end{equation}
where $[K_{ab}] \equiv K_{{\rm out}\, ab} - K_{{\rm in}\, ab}$. However, to ensure that the interior region does not introduce any dynamics on the brane,
we demand that $K_{ab}^{\rm in}$ is equal to a constant value corresponding to a static cylinder:
\begin{align}
K^{~\phi}_{{\rm in}\,\phi}=\frac{1}{R}\,; \qquad K^{~0}_{{\rm in}\,0} = K^{~i}_{{\rm in}\,j}= 0 \,.
\label{eq:RegII}
\end{align}
With this choice, the junction condition~\eqref{israel2} agrees with~\eqref{israel}, and the two descriptions give identical brane geometry and
exterior space-time. We will not be concerned with the brane interior.
\end{itemize}
{\it A priori} one naturally expects that the solutions thus obtained should not depend sensitively on the details of the regularization,
as long as the characteristic time scale ($H^{-1}$, in the case of interest) is much longer than the radius of the circle, {\it i.e.},
\begin{equation}\label{UV_IR_Limit}
H^{-1}\gg R\;.
\end{equation}
We explicitly check this expectation in Appendix~\ref{ap:dynReg}, by studying a different regularization scheme called \textit{dynamical regularization}.
In this scheme, the gravitational dynamics are fully resolved inside the cylinder. We find that the time-averaged Hubble evolution
on the brane agrees with the static regularization result in the limit~\eqref{UV_IR_Limit}.
Let us stress that only by performing this fully self-consistent GR analysis, which in particular implements regularity at the symmetry axis, was it possible to quantify the effect of having some interior dynamics and thus to show that our results are regularization independent. Moreover, this analysis revealed that the static regularization corresponds to the favorable case where the effects of the interior dynamics are minimized and perfectly smoothed out. The presentation in the main part of the paper therefore uses the simpler static regularization.
The interested reader is referred to the Appendix~\ref{ap:dynReg} for more details.
\subsection{Bulk geometry}
The assumed symmetries are homogeneity, isotropy and (for simplicity) spatial flatness along the three spatial brane dimensions, as well as axial symmetry about the brane.
As shown in Appendix~\ref{ERcoords}, given these symmetries and the fact that the space-time is empty away from the brane, the bulk metric can be brought to the form:
\begin{align}
\label{eq:met_cyl_symm_2}
\mathrm{d} s^2_6 &= \mathrm{e}^{2(\eta - 3\alpha)} \left( -\mathrm{d} t^2 \!+ \mathrm{d} r^2 \right) + \mathrm{e}^{2\alpha} \mathrm{d} \vec{x}^2 + \mathrm{e}^{-6\alpha} r^2 \mathrm{d}\phi^2 \; .
\end{align}
Note that by formally replacing $3\alpha \rightarrow \alpha$ in the first and last term and $\vec{x} \rightarrow z$, we recover the ansatz that was used by Einstein and Rosen to derive the existence of cylindrically symmetric waves in GR \cite{EinsteinRosen1937} (see also, \textit{e.g.},~\cite{Marder1958}). The additional factor 3 in the generalized case simply counts the dimensionality of the symmetry axis. In the remainder of the paper we will refer to \eqref{eq:met_cyl_symm_2} as the \textit{Einstein-Rosen coordinates}.
The Einstein field equations in the exterior (vacuum) region become
\begin{subequations}
\begin{empheq}[box=\widefbox]{align}
\partial_t^2 \alpha &= \partial_r^2 \alpha + \frac{1}{r}\partial_r\alpha \label{eq:2D_wave}\\
\partial_r\eta &= 6r\Big((\partial_r\alpha)^2 + (\partial_t\alpha)^2\Big) \label{eq:etaPrime_vac}\\
\partial_t\eta &= 12 r \, \partial_r \alpha \, \partial_t\alpha \;. \label{eq:etaDot_vac}
\end{empheq}
\label{eq:einstein_vacuum}
\end{subequations}
The fact that $\alpha$ obeys the linear\footnote{Despite the linearity of this equation, the complete brane-bulk system is still highly nonlinear due to the junction conditions, discussed below.} 2D wave equation \eqref{eq:2D_wave} makes the coordinate choice \eqref{eq:met_cyl_symm_2} unique and especially convenient for numerical implementation.
\subsection{Brane geometry}
The induced cosmological metric on the brane is
\begin{equation}\label{eq:induced_g_3}
\mathrm{d} s^2_5 =-\mathrm{d} \tau^2 + \mathrm{e}^{2\alpha_0} \mathrm{d} \vec{x}^2 + R^2 \mathrm{d}\phi^2\;,
\end{equation}
where the subscript ``0'' denotes evaluation at the brane position. The scale factor is recognized
as $a(\tau) \equiv \mathrm{e}^{\alpha_0}$, with Hubble parameter $H \equiv \mathrm{d} \alpha_0/\mathrm{d} \tau$.
The proper time $\tau$ is related to the ``bulk'' time via
\begin{equation}\label{eq:dTauDT}
\mathrm{d} \tau = \frac{\mathrm{e}^{-3\alpha_0}}{\gamma}\mathrm{d} t \,,
\end{equation}
where
\begin{equation}\label{eq:defGamma}
\gamma \equiv \frac{\mathrm{e}^{-\eta_0}}{\sqrt{1 - \left(\frac{{\rm d}r_0}{{\rm d}t}\right)^2}} = \sqrt{\mathrm{e}^{-2\eta_0} +\dot{r}_0^2\mathrm{e}^{-6\alpha_0} } \;,
\end{equation}
with $r_0(t)$ describing the position of the brane in the extra-dimensional space, and $\dot{r}_0 \equiv \frac{{\rm d}r_0}{{\rm d}\tau}$.
Here and henceforth, dots refer to ${\rm d}/{\rm d}\tau$.
To recover 4D gravity in the appropriate regime, we assume that the proper circumference (divided by $ 2\pi $) is stabilized:
\begin{equation}
\label{eq:defR}
R \equiv r_0 \mathrm{e}^{-3\alpha_0} = \text{const.}
\end{equation}
The justification is clear: A realistic defect would have some underlying bulk forces to keep its core stable. Technically, this is imposed
by introducing a suitable azimuthal pressure component $P_\phi $. We must of course check {\it a posteriori} whether the pressure thus inferred
satisfies physically reasonable energy conditions, such as the Null Energy Condition.
As an immediate consequence of the stabilization condition, the 4D Planck mass,
\begin{equation}
M_{\rm Pl}^2 = 2\pi R M_5^3\,,
\label{M4}
\end{equation}
is constant. Moreover,~\eqref{eq:defR} implies $\dot{r}_0 = 3Hr_0$, which allows us to rewrite~\eqref{eq:defGamma} as
\begin{equation}
\gamma = \sqrt{\mathrm{e}^{-2\eta_0} + 9H^2R^2} \, .
\label{gammaredef}
\end{equation}
The symmetries of our system allow for a fluid ansatz of the localized 5D energy-momentum tensor
\begin{align}
\label{eq:EMT}
T^{(5) a}_{\hphantom{(5)a} b}=\frac{1}{2\pi R} {\rm diag}(-\rho, P,P,P,P_{\phi})\;,
\end{align}
where the overall factor is such that $T_{ab} = 2\pi R T^{(5)}_{ab}$ defines a 4D energy-momentum tensor. Fixing $R$ also implies that
the energy density and pressure satisfy the standard 4D conservation equation
\begin{empheq}[box=\widefbox]{equation}
\dot \rho + 3H \left( \rho + P \right) = 0\,.
\label{eq:enCons4D}
\end{empheq}
\subsection{Junction conditions} \label{sec:junction_cond}
In the next step, we explicitly evaluate the junction conditions~\eqref{israel}. The outward-pointing unit normal vector is given by $n^A = \mathrm{e}^{3\alpha_0} \left(3HR, \gamma, 0,0,0,0 \right)$.
It is straightforward to show that $K_{{\rm out}\,ab}$ has components
\begin{subequations}
\label{eq:Kcompsout}
\begin{align}
K^{~0}_{{\rm out}\,0} &= \frac{3R}{\gamma}\left(\dot H +H \dot \eta_0 \right) +n^A\partial_A\left(\eta - 3\alpha\right)\vert_0 \,, \\
K^{~i}_{{\rm out}\,j} &= \delta^i_{~j} n^A\partial_A \alpha\vert_0 \, \,,\\
K^{~\phi}_{{\rm out}\,\phi} &= \frac{\gamma}{R} - 3 n^A\partial_A \alpha\vert_0 \,.
\end{align}
\end{subequations}
Using~\eqref{eq:defR},~\eqref{M4} and~\eqref{eq:EMT}, the $(0,0)$ component of the junction conditions gives a modified Friedmann equation
\begin{empheq}[box=\widefbox]{equation}
H^2 =\frac{\rho}{3 M_{\rm Pl}^2} + \frac{1}{r_c^2} \left( \gamma - 1 \right)\,,
\label{eq:rhoJunctCond}
\end{empheq}
where $\gamma$ is given by~\eqref{gammaredef}, and $r_c$ denotes the cross-over scale
\begin{equation}
\label{eq:defCrossover}
r_c^2 \equiv \frac{3 M_{\rm Pl}^2}{2\pi M_6^4}\,.
\end{equation}
The modification to the standard Friedmann equation is controlled by this cross-over scale. Assuming $|\gamma-1| \sim 1$, one can already tell that in the regime where $ H \gg r_c^{-1} $ the modification is negligible and the model reproduces the standard 4D evolution. When $ H $ becomes of order $ r_c^{-1} $, however, the modification becomes important and we expect a transition to a 6D regime. This is of course the way the model was engineered to work in the first place. It is also very similar to the 5D (DGP) case, where the modification term is simply $ \pm H / r_c^{(5)} $, with the appropriate 5D crossover scale $ r_c^{\rm DGP} = \frac{M_{\rm Pl}^2}{2M_5^3}$. But the crucial difference is that in the 6D case, the modification term cannot be directly expressed in terms of on-brane quantities like $ H $. It knows something about the bulk geometry through its dependence on $ \eta_0 $, and in order to make quantitative predictions one has to solve the bulk Einstein equations \eqref{eq:einstein_vacuum} as well.
The $(i,j)$ component of the junction conditions, combined with the vacuum Einstein equations \eqref{eq:etaPrime_vac} and \eqref{eq:etaDot_vac} in the limit $r \rightarrow r_0^+$, can be expressed as
\begin{equation}
\label{eq:pJunctCondRConst}
\boxed{\dot H = -\frac{3}{2f(\tau)}\left[ \frac{P}{3 M_{\rm Pl}^2} + H^2 - \frac{1}{r_c^2} \Big( \gamma\, g(\xi, \chi)-1\Big )\right]\,,}
\end{equation}
where
\begin{equation}
f(\tau) \equiv 1- \frac{9R^2}{2r_c^2\gamma} \, ,
\label{eq:def_f}
\end{equation}
and
\begin{subequations}
\begin{align}
g(\xi, \chi) &\equiv 1 + 2 \left (9\chi - 1 \right ) \bigl [ 3\chi + \xi \left ( 3\xi - 2 \right ) \left ( 9\chi - 1 \right ) \bigr ] \, , \label{eq:def_g}\\
\xi & \equiv r \partial_r \alpha\vert_0 \, , \qquad
\chi \equiv \frac{H^2R^2}{\gamma^2} \,.
\end{align}
\end{subequations}
In our analysis, we will see that the sign of $f(\tau)$
allows to discriminate between a stable and an unstable class of solutions.
The closed set of equations describing the bulk-brane system comprises the bulk equations of motion~\eqref{eq:einstein_vacuum}, the energy conservation equation~\eqref{eq:enCons4D}
and the Friedmann equation~\eqref{eq:rhoJunctCond}. The $\dot H$ equation~\eqref{eq:pJunctCondRConst} follows from these, as usual. For the purpose of numerical implementation,
however, we will integrate the $\dot H$ equation. The Friedmann equation will only be implemented at the initial time and later on will serve as a numerical consistency check.
Finally, the $(\phi,\phi)$ component of the junction conditions can be used to determine the azimuthal pressure:
\begin{eqnarray}
\nonumber
\frac{P_\phi}{3M_{\rm Pl}^2} &=& -\dot{H} \left( 1 - \frac{3R^2}{r_c^2\gamma} \right) - 2H^2 \\
& & + \frac{6\gamma}{r_c^2} \left\{\chi + \bigl[ 3\chi - \xi ( 9\chi - 1 ) \bigr]^2 \right\}
\label{eq:pPhiJunctCond}
\end{eqnarray}
In our analysis, we will compute $P_\phi$ explicitly to check, for instance, whether the equation of state along the azimuthal direction satisfies the Null Energy Condition.
Before investigating the dynamical solutions, let us pause to recover the well-known static solutions from our setup.
\section{Static Solutions}
\label{sec:static_sol}
The static case constitutes an important check of the above equations and will provide a first physical insight into the geometry of the system\footnote{Note that in this case the static regularization (used in the main text) and the dynamical one (discussed in Appendix~\ref{ap:dynReg}) coincide by construction. Indeed, the only non-singular static geometry inside the cylinder is Minkowski space, hence the extrinsic curvature at the inner boundary is exactly the one given by~\eqref{eq:RegII}.}.
For a purely static solution $\dot r_0=0$ and all metric functions solely depend on $r$.
The exterior field equations~\eqref{eq:einstein_vacuum} yield the solution
\begin{equation}
\alpha=c \log{\frac{r}{r_0}}+\alpha_0
\quad\text{and}\quad
\eta=6\, c^2 \log{\frac{r}{r_0}}+\eta_0\;.
\label{staticprofile}
\end{equation}
By rescaling coordinates tangential to the brane, we can set $\alpha_0=0$ without loss of generality.
The remaining constants $c$ and $\eta_0$ are determined by the junction conditions~\eqref{eq:rhoJunctCond} and~\eqref{eq:pJunctCondRConst}:
\begin{subequations}
\begin{align}
\eta_0&=-\log{\left(1-\frac{\rho}{\rho_{\rm crit}} \right)}\;,\label{eq:eta0}\\
c& = \frac{1}{3} \left( 1 - \sqrt{\frac{2 \rho_{\rm crit} + (1 + 3w) \rho }{2(\rho_{\rm crit} - \rho)}} \right )\;, \label{eq:c}
\end{align}
\end{subequations}
where $w = P/\rho$ is the equation of state. Here we have introduced the critical density $\rho_{\rm crit} \equiv 2\pi M_6^4$. The third junction condition~\eqref{eq:pPhiJunctCond} then becomes
\begin{equation}
P_{\phi}=6 c^2 \left(\rho_{\rm crit}-\rho \right)\;.
\end{equation}
Note that~\eqref{eq:eta0} is ill-defined for $\rho>\rho_{\rm crit}$; we will come back to this point shortly. The line element for the exterior reads
\begin{eqnarray}
\nonumber
\mathrm{d} s^2 &=& \mathrm{e}^{2\eta_0} \left(\frac{r}{r_0}\right)^{12c^2-6c}\!\!\!\!\left( -\mathrm{d} t^2 \!+ \mathrm{d} r^2 \right)\\
&+& \left(\frac{r}{r_0}\right)^{2c} \mathrm{d} \vec{x}^2 + \left(\frac{r}{r_0}\right)^{-6c} r^2 \mathrm{d}\phi^2 \; .
\end{eqnarray}
Since the brane induced terms vanish identically for static configurations, this solution is the direct generalization of the exterior metric of a static cylinder in 4D, first derived by Levi-Civita~\cite{Levi:1919} and later reviewed for example in \cite{Thorne:1965}.
Consider the case of pure 4D tension on the brane:
\begin{subequations}
\begin{gather}
\rho = - P \equiv \lambda \\
\Rightarrow\, c = 0 = P_\phi \,.
\label{puretension}
\end{gather}
\end{subequations}
The coordinate rescaling $( \bar t,\bar r)=( \mathrm{e}^{\eta_0} t, \mathrm{e}^{\eta_0} (r-r_0)+r_0)$ yields the famous wedge geometry in Gaussian normal coordinates, characterized by the deficit angle $\delta \equiv \lambda / M_6^4$:
\begin{equation}\label{eq:ds_wedge}
\mathrm{d} s^2 = -\mathrm{d} \bar t^2 \!+ \mathrm{d} \bar r^2 + \mathrm{d} \vec{x}^2 + W(\bar r)^2 \mathrm{d}\phi^2 \; ,
\end{equation}
where
\begin{equation}
W(\bar r)=
\begin{cases}
\bar r & \mbox{for } \bar r\leq r_0\\
\frac{\delta}{2\pi}r_0+\left(1-\frac{\delta}{2\pi}\right)\bar r & \mbox{for } \bar r>r_0\;.
\end{cases}
\end{equation}
Note that this solution corresponds to the generalization of the cosmic string geometry \cite{Vilenkin:1981zs, Hiscock:1985uc} to 6D. The coordinates cover again the whole space-time including the interior. A well-known fact about this solution is that the intrinsic brane geometry is flat and the energy on the brane only affects the extrinsic curvature, thereby creating a deficit angle. This property makes the higher codimensional models in particular interesting with respect to the cosmological constant problem because $\lambda$ is effectively ``filtered out'' from the perspective of a brane observer; see~\cite{Burgess:2011va} and~\cite{Dvali:2002pe} in the case of large or infinite extra dimensions, respectively.
For sub-critical tensions $\delta<2\pi$ we find for the ratio of physical radius and circumference: $ \bar r/W(\bar r)=1$ for $\bar r\leq r_0$ and $\bar r/W(\bar r)>1$ for $\bar r>r_0$. In an embedding picture this corresponds to a capped cone, as shown in Fig.~\ref{fig:embedding_geometry}. In the critical limit $\delta\rightarrow2\pi$, the embedding geometry becomes ``cylindrical''.
In the super-critical case, $\delta>2\pi$, the circumference $2\pi W(\bar r)$ decreases for $\bar r>r_0$ and vanishes for a certain radius $\bar r_1$, implying the existence of a second axis. However, in general the geometry is not elementary flat at that position, {\it i.e.}, $W^{\prime}(\bar r_1) \neq 1$, which indicates the existence of a naked singularity. It has been argued that this (conical) singularity is an artifact of the static approximation and is resolved once the full dynamics are taken into account \cite{Cho:1998xy}.
\begin{figure*}
\subfloat[$ \delta<\delta_{\rm crit}$]{
\includegraphics[width=0.3\textwidth]{figures/capped_cone-1.pdf}
}
\hfil
\subfloat[$ \delta=\delta_{\rm crit} $]{
\includegraphics[width=0.3\textwidth]{figures/capped_cone-2.pdf}
}
\hfil
\subfloat[$ \delta>\delta_{\rm crit} $]{
\includegraphics[width=0.3\textwidth]{figures/capped_cone-3.pdf}
}
\caption{Embedding diagrams of the regularized static geometry in the case of a pure tension brane. The circle at $\bar r=r_0$ describes the brane. As the tension approaches the critical value, the deficit angle approaches $2\pi$, and the bulk geometry becomes cylindrical (b). For super-critical tensions, a naked singularity develops in the bulk a finite distance away from the brane.} \label{fig:embedding_geometry}
\end{figure*}
The derivation of the junction conditions in the Einstein-Rosen language is not compatible with the super-critical scenario. This is clear in the static case, as already mentioned, since~\eqref{eq:eta0} does not allow a real solution for $\eta_0$ in the super-critical regime. See Appendix~\ref{ERcoords} for a more detailed discussion of this point in the context of dynamical solutions, and~\cite{Niedermann:2014yka} for a detailed investigation of super-critical cosmic strings. We henceforth exclude the super-critical regime from our analysis.
\section{Interlude: Bulk-brane dynamics}
\label{sec:Brane_Bulk}
The analysis of cosmological solutions on the brane is greatly complicated by the fact that the assumed symmetries allow
for axially symmetric gravitational waves propagating in the bulk. This is unlike the much-studied codimension-one case,
where the assumption of planar symmetry enforces a version of Birkhoff's theorem~\cite{Taub:1951}: The only vacuum 5D solutions
are Minkowski or Schwarzschild. The Schwarzschild mass parameter enters the brane Friedmann equation as the coefficient
of a ``dark radiation'' term. In particular, the brane Friedmann equation is completely local.
The codimension-two case of interest is qualitatively very different. The bulk field equations \eqref{eq:einstein_vacuum} explicitly show that in this case gravitational waves are in fact compatible with all the symmetries. As a consequence, it would be possible to prepare a wave packet in the bulk that reaches the brane at some arbitrary time. Since the amplitude of the wave is given by the metric function $ \alpha(t, r) $, while the 4D scale factor is determined by $ \alpha_0(t) \equiv \alpha(t, r_0) $, the 4D cosmological evolution will inevitably be influenced by such a wave packet. As a result, it cannot be possible to derive a closed local on-brane evolution equation for $ \alpha_0 $, without imposing additional restrictions on the bulk geometry.
What could these restrictions be? As a first guess, one could try to assume a flat bulk geometry, just as could be done in the DGP case. After all, this is also what happens in the static pure tension solution. However, it turns out that this is no longer possible after one demands $ \alpha_0 $ to have non-trivial dynamics. To show this, let us try to set the $(t,x^1,t,x^1)$- and $(t,r,t,r)$-components of the Riemann tensor to zero, which is a necessary condition for flatness. This in turn demands
\begin{equation}
\left(\partial_t \alpha\right)^2- \left(\partial_r \alpha\right)^2=0 \qquad \text{and} \qquad r\,\partial_r \alpha=0\,.
\end{equation}
The only solution to these equations is indeed the trivial configuration $\alpha = \mathrm{constant}$.
So a dynamical codimension-two brane inevitably curves the extra-dimensional space-time, and since the on brane geometry will be time-dependent, so will be the bulk geometry. In other words, gravitational waves are not only possible for a non-trivial cosmology in this setup, but in fact necessary.
One could still try to arrive at a closed on-brane system by implementing an ``outgoing wave condition'' at the outer boundary of the brane to exclude incoming bulk waves. Physically, this is clearly a necessary condition because we assume a source-free, infinite bulk. However, it is well known that such a condition is necessarily non-local (in time) in the case of cylindrically symmetric waves (see \cite{Givoli:1991} for a review, and \cite{Hofmann:2013zea} for a discussion in the context of GR).
Moreover, because the coordinate position of the brane $ r_0(t) $ will in general be time dependent, the resulting on-brane system would be non-local both in space \emph{and} time. It is clear that solving such a non-local system would not be any easier than solving the full bulk system from the start. In other words, if one tried to accommodate for all allowed bulk configurations in the on-brane system, one would end up with not only one, but infinitely many ``constants of integration''. This is what makes the codimension-two problem much harder to solve.
Therefore, there seems to be no way around solving the full bulk geometry in order to see what 4D cosmology emerges in the codimension-two BIG model. This can in general only be done numerically, and we will do so in the next sections.
\section{Numerical implementation and Initial Data}
\label{sec:num_impl}
We now turn to the numerical implementation of the full brane-bulk system~\eqref{eq:einstein_vacuum},~\eqref{eq:enCons4D} and~\eqref{eq:rhoJunctCond}.
Solutions were obtained by specifying initial data, as explained below, and numerically integrating this initial value problem forward in time. Since the dynamical bulk equation~\eqref{eq:2D_wave} is nothing but the standard (flat space) cylindrically symmetric scalar wave equation, it is straightforward to find a stable integration scheme for the PDE part of the problem. There is only a slight complication stemming from the matching procedure. Even though the physical brane size $ R $ is fixed, its coordinate position $ r_0 $ is generally time-dependent. Therefore, if one chooses a fixed spatial grid size in the bulk (as we do), one has to allow $ r_0 $ to lie in between those grid points. We deal with this problem by using some suitable interpolation scheme. The details of the numerical implementation can be found in Appendix~\ref{ap:numImpl}.
The numerical integration starts at some initial time $ t = t_i $, $ \tau = \tau_i $. Let us denote all functions evaluated at this time with a subscript $ i $.
Through a global rescaling of coordinates, we can always set $\alpha = 0$ on the brane initially, {\it i.e.},
\begin{equation}
\label{eq:initCondAlpha0}
\left(\alpha_{0}\right)_{i} = 0\,.
\end{equation}
Consequently, the initial brane position is
\begin{equation}
\left(r_{0}\right)_{i} = R \,.
\end{equation}
In the bulk we must specify the initial radial profile $\alpha_i( r)$ and its time derivative $\partial_t \alpha_i( r)$.
To be definite, as initial profile we choose the static profile given by~\eqref{staticprofile}, namely
\begin{equation}
\label{eq:initProfileAlpha}
\alpha_{i}(r) = c \ln\left( \frac{r}{R} \right) \, ,
\end{equation}
where the constant $ c $ is given by~\eqref{eq:c} with $\rho \rightarrow \rho_i$. In particular, for a cosmological constant ($ w=-1 $), we get $ c = 0 $, and hence $ \alpha_i(r ) = 0 $. Note that by choosing the static profile we are not putting any potential energy into the bulk gravitational field initially.
At the brane position, the velocity profile is related to the initial Hubble parameter $ H_i $ via
\begin{eqnarray}
\nonumber
\partial_{ t} \alpha_{0i} &=& \frac{\mathrm{d} \alpha_{0i} }{\mathrm{d} t}- \frac{\mathrm{d} r_{0i} }{\mathrm{d} t}\,\partial_{ r} \alpha_{0i} \\
\nonumber
&=& \frac{\mathrm{d} \alpha_{0i} }{\mathrm{d} t}\left( 1 - 3 \partial_{ r} \alpha_{0i} \right) \\
&=& \frac{H_i}{\gamma_i}(1-3c)\;.
\label{eq:alpha0TildeDot_init}
\end{eqnarray}
Extending this to the bulk, we write
\begin{equation}
\partial_t \alpha_i(r) = \frac{H_i}{\gamma_i} (1 - 3c) F(r) \,,
\end{equation}
where $ F( r) $ is some profile function satisfying the boundary condition $F(R) = 1$.
To minimize the amount of kinetic energy put into the gravitational field initially, which could impact the brane cosmology for long times,
we will choose profile functions which are sharply localized around the brane. For definiteness, we will focus on a Gaussian profile of width $\sigma$,
\begin{equation}
F(r) = \exp\left[ - \frac{(r-R)^2}{\sigma^2} \right]\,,
\label{GaussianF}
\end{equation}
With these choices, we expect the on-brane evolution to rapidly become insensitive to the initial conditions.
This completes the specification of initial data. Indeed, the remaining variable, $ \eta_{0i} $, is fixed by the constraint~\eqref{eq:rhoJunctCond}, together with the relation~\eqref{gammaredef}\footnote{The full radial profile $ \eta(r) $ can be calculated from \eqref{eq:etaPrime_vac}, but is actually not needed for the evolution of $ \alpha $. Only $ \eta_0 $ enters through the junction conditions, and it can be calculated at later times from its initial value using \eqref{eq:etaPrime_vac} and \eqref{eq:etaDot_vac} only locally at the brane position.}. Specifically,
\begin{equation}
\frac{\rho_i}{\rho_{\rm crit}} = r_c^2 H_i^2 + 1 - \sqrt{ \mathrm{e}^{-2\eta_{0i}} + 9 H_i^2 R^2 } \, .
\end{equation}
Note that this equation does not always have a (real) solution for $ \eta_{0i} $. The existence of a real solution places an upper bound on the energy density:
\begin{equation}
\label{eq:criticalityBound}
\frac{\rho}{\rho_{\rm crit}} < r_c^2 H^2 + 1 - 3 \left|H\right| R \, .
\end{equation}
Since the constraint has to hold for all times, we were able to drop the subscript $i$. We will refer to this as the \textit{criticality bound}, separating the sub- and super-critical regimes. As soon as~\eqref{eq:criticalityBound} is violated, the initial constraint cannot be fulfilled.
The reason is that in this parameter regime the Einstein-Rosen coordinates as used in our derivation are no longer valid. The interested reader is referred to Appendix~\ref{ERcoords}
for more details. Since this super-critical regime is not compatible with our coordinate choice, it will not be considered in this paper.
As a check on~ \eqref{eq:criticalityBound}, note that it correctly reproduces the static criticality bound $ \rho < \rho_{\text{crit}} = 2\pi M_6^4 $ in the static limit $ H \to 0 $. In the dynamical case, however, the bound is more general. In particular, for $ r_c = 0 $, {\it i.e.}, without the induced gravity terms, the bound becomes stronger---the critical point is reached for a smaller value of $ \rho $ than in the static case.
Physically, the reason is that for $ H \neq 0 $, there is additional kinetic energy in the system. For $ r_c \neq 0 $, on the other hand, the induced gravity terms can absorb (or ``shield'') part of the energy density from the bulk, thereby allowing much larger values for $ \rho $ than in the static case.
The final ingredient is the choice of grid spacing for the numerical calculation. We use a scheme in which the temporal and radial grid spacing is the same and constant:
\begin{align}
\Delta t & = \Delta r \equiv \epsilon \,.
\label{stepsize}
\end{align}
The system can then be evolved forward in time using \eqref{eq:einstein_vacuum} and \eqref{eq:pJunctCondRConst} for any given $ H_i $, $ \sigma $, $ R $, $ r_c $, $ \rho_i $ and equation of state parameter $ w $. (In fact, the quantities $ H_i $, $ R $ and $ r_c $ enter the equations only in the combinations $ H_i r_c $ and $ H_i R $, so only two of them need to be specified while the third one is degenerate.)
The constraint equation \eqref{eq:rhoJunctCond} can be used as an important consistency check for the numerical solver.
Further details of the numerical implementation are given in Appendix~\ref{ap:numImpl}. In what follows we will present the results.
\section{Numerical Solutions}
\label{sec:num_sol}
\begin{figure*}[htb]
\subfloat[The Hubble parameter on the brane exhibits degravitation. It starts out positive and asymptotically tends to zero.]{
\includegraphics[width=0.45\textwidth]{figures/hubble_cc_degrav_statreg.pdf}
\label{fig:hubble_cc_degrav}
}
\hfill
\subfloat[The radial profile for $ \alpha $ at different values of $ \tau $. The dots indicate the brane position as a function of time.]{
\includegraphics[width=0.45\textwidth]{figures/alpha_cc_degrav_statreg.pdf}
\label{fig:alpha_cc_degrav}
}
\\
\subfloat[Equation of state of $ P_\phi $ that is needed to keep the brane circumference fixed for $ w=-1 $. It never falls below the value $-1$ corresponding to unphysical matter.]{
\includegraphics[width=0.45\textwidth]{figures/pPhiEOS_cc_degrav_statreg.pdf}
\label{fig:pPhi_cc_degrav}
}
\hfill
\subfloat[The effective energy density, $\hat\rho\equiv \rho - 3M_4^2 H^2$, as ``seen'' by 6D GR. This approaches a positive value consistent with the static solution.]{
\includegraphics[width=0.45\textwidth]{figures/rhoHat_cc_degrav_statreg.pdf}
\label{fig:rhoHat_cc_degrav}
}
\caption{Example of a degravitating solution.}
\label{fig:cc_degrav}
\end{figure*}
\begin{figure*}[htb]
\subfloat[The Hubble parameter on the brane grows in time, indicating super-acceleration.]{ %
\includegraphics[width=0.45\textwidth]{figures/hubble_cc_pathol_statreg.pdf} %
\label{fig:hubble_cc_pathol} %
} %
\hfill %
\subfloat[The radial profile of the function $ \alpha $ at different values of $ \tau $. At fixed $r$, $\alpha$ grows in time.]{ %
\includegraphics[width=0.45\textwidth]{figures/alpha_cc_pathol_statreg.pdf} %
\label{fig:alpha_cc_pathol} %
} %
\\
\subfloat[Equation of state of $ P_\phi $ that is needed to keep the brane circumference fixed. It is negative and falls rapidly below $-1$.]{ %
\includegraphics[width=0.45\textwidth]{figures/pPhiEOS_cc_pathol_statreg.pdf} %
\label{fig:pPhi_cc_pathol} %
} %
\hfill %
\subfloat[The effective energy density, as ``seen'' by 6D GR, becomes negative. This is interpreted as the source of the physical instability.]{ %
\includegraphics[width=0.45\textwidth]{figures/rhoHat_cc_pathol_statreg.pdf} %
\label{fig:rhoHat_cc_pathol} %
} %
\caption{Example of a super-accelerating solution.}\label{fig:cc_pathol} %
\end{figure*}
\begin{figure*}
\subfloat[Behavior of solutions for different choices of $r_c$ and $\rho_i$. The green region (Region (1)) shows stable solutions; the red region (Region (2)) shows unstable solutions. The solid line in between corresponds to $f = 0$. The gray region corresponds to super-critical solutions, which are not covered in our analysis.]{
\includegraphics[width=0.42\textwidth]{figures/contourPlot_R_0_05_statreg.pdf}
\label{fig:contour_plot_full}
}
\hfill
\raisebox{0.1\height}{\includegraphics[width=0.11\textwidth]{figures/contourPlot_legend.pdf}}
\hfill
\subfloat[Zoom into the small blue rectangle depicted in Fig.~\ref{fig:contour_plot_full}. The yellow/orange regions show solutions which hit the singularity at $f = 0$ in a finite time. The dashed lines have been inferred from the numerical results.]{
\includegraphics[width=0.43\textwidth]{figures/ContourPlot_zoomed_IV_w_-1.pdf}
\label{fig:contour_plot_zoom}
}\\
\raisebox{0.75\height}{\includegraphics[width=0.15\textwidth]{figures/fPlotLegend_statreg.pdf}}
\subfloat[The evolution of $f(\tau) = 1 - \frac{9R^2}{2r_c^2\gamma}$ for $H_i r_c=0.15$ and different values of $\rho_i$. The color/numerical labels of the curves match those of Fig.~\ref{fig:contour_plot_zoom}. The yellow (4) and the orange (5) lines hit the singularity at $f=0$ in finite time, while the green (1) and red (2) curves avoid the singularity.]{
\includegraphics[width=0.5\textwidth]{figures/fPlotZoom_statreg.pdf}
\label{fig:f_plot_zoom}
}
\caption{(color online) Results of the numerical stability analysis of the model.}
\label{fig:contour_plot}
\end{figure*}
We have found two, qualitatively different classes of solutions, depending on the initial conditions.
The first class, called {\it degravitating} solutions, features a geometry which at late times approaches
the static profile. In particular, $H\rightarrow 0$ on the brane. The second class, called {\it super-accelerating}
solutions, features a run-away behavior for the Hubble parameter on the brane. The source for this apparent instability
is an effective energy density on the brane which violates the Null Energy Condition.
After describing a fiducial degravitating (Sec.~\ref{sec:degrav_sol}) and super-accelerating (Sec.~\ref{sec:pathol_sol}) solution,
we will discuss the regions of parameter space spanned by each class in Sec.~\ref{sec:contourPlot}.
\subsection{A degravitating solution}
\label{sec:degrav_sol}
As a first example, let us consider a 4D cosmological constant source ($ w=-1 $) with parameters\footnote{For completeness, the width of the initial Gaussian profile~\eqref{GaussianF} is set to $\sigma = R/50$, and the step size for integration~\eqref{stepsize} is $\epsilon = 2 \times 10^{-4} R$.}
\begin{equation}
H_i r_c = \frac{1}{10}\, ; ~~H_i R = \frac{1}{20} \, ; ~~\rho = \frac{4}{5} \rho_{\rm crit} \, .
\label{eq:parameteres_dgrav}
\end{equation}
For this choice, the energy density lies in the sub-critical regime. Meanwhile, the cross-over scale $r_c$ is smaller than the initial Hubble radius, hence we expect a large modification to standard 4D gravity.
This can be seen directly from the Friedmann equation~\eqref{eq:rhoJunctCond}: The modification term $(\gamma-1)$ is controlled by $r_c$.
The results of the numerics are depicted in Fig.~\ref{fig:cc_degrav}. Fig.~\ref{fig:hubble_cc_degrav} shows the Hubble parameter on the brane as a function of time.
(The numerical error estimates for $ H $, discussed in Appendix~\ref{ap:numErrors}, are smaller than the line thickness.) We see that $H$ initially decreases to negative values, turns around and approaches zero at late times. This confirms that the static solutions of Sec.~\ref{sec:static_sol} have a finite basin of attraction. This is one of the central results of this work: it is the first example of dynamical {\it degravitation}, and demonstrates how the brane tension can be absorbed into extrinsic curvature while the intrinsic brane geometry tends to flat, Minkowski space. The evolution of the bulk geometry, characterized by $\alpha$, is shown in Fig.~\ref{fig:alpha_cc_degrav}. The initial configuration, as discussed in the last section, leads after a few time steps to a rather narrow Gaussian profile. As time evolves, we see that $\alpha$ describes a two dimensional gravitational wave that moves outwards, gets more and more diluted and asymptotically settles to a constant.
It remains to check the physicality of the azimuthal pressure component $P_\phi$ required for stabilization. The equation of state corresponding to this pressure component is shown in Fig.~\ref{fig:pPhi_cc_degrav}. The equation of state satisfies the Null Energy Condition ($w_\phi \geq -1$), and is therefore physically reasonable. At late times, $P_\phi\rightarrow 0$, which is consistent
with the static solution for a 4D cosmological constant---see~\eqref{puretension}. Figure~\ref{fig:rhoHat_cc_degrav} shows the effective energy density (including the brane induced terms) that sources the 6D bulk gravity theory, $\hat\rho\equiv \rho - 3M_4^2 H^2$. This quantity remains positive at all times, which indicates a healthy source from the bulk perspective. At late times, $H\rightarrow 0$,
and $\hat\rho$ approaches $\frac{4}{5}\rho_{\rm crit}$, which is consistent with a static solution with brane density given by~\eqref{eq:parameteres_dgrav}.
We have repeated the analysis with a dust $(w=0)$ or radiation $(w=1/3)$ component on the brane and found similar behavior. The system approaches the corresponding static, deficit-angle solutions at late times. The azimuthal pressure $P_\phi$ and effective density $\hat\rho$ are healthy at all times.
\subsection{A super-accelerating solution}
\label{sec:pathol_sol}
Consider once again a 4D cosmological constant source ($ w=-1 $), with the same parameters as before except for a somewhat larger value of $r_c$:
\begin{equation}
H_i r_c = \frac{1}{4} \, .
\label{eq:param_pathol}
\end{equation}
In this case we find completely different behavior. The Hubble parameter on the brane, shown in Fig.~\ref{fig:hubble_cc_pathol}, grows monotonically in time,
which indicates an effective violation of the Null Energy Condition. This growth propagates into the bulk, as can be seen from Fig.~\ref{fig:alpha_cc_pathol}: the wave
function $\alpha(\tau,r)$ grows in time at any $r$.
This pathological behavior is reflected in the azimuthal pressure $P_\phi$, whose equation of state (Fig.~\ref{fig:pPhi_cc_pathol}) becomes less than $-1$ and tends to $-\infty$.
Such an equation of state violates the Null Energy Condition and is rather unphysical. This suggests that no consistent stabilization mechanism exists for a super-accelerating solution. One might wonder
whether this apparent instability is solely due to this strange azimuthal component required to fix the brane circumference. We found that this is not the case. In Appendix~\ref{sec:vanishingPPhi},
we show that fixing $ P_\phi = 0 $ by hand, and therefore allowing the circumference to evolve in time, still results in super-acceleration.
The instability can be clearly seen by looking at the effective energy density $\hat\rho\equiv \rho - 3M_4^2 H^2$ that sources 6D gravity. As shown in Fig.~\ref{fig:rhoHat_cc_pathol},
$\hat{\rho}$ starts out positive but eventually turns around and reaches negative values. This behavior bears resemblance to the DGP model, where the self-accelerating branch leads to a negative
effective energy density~\cite{Gregory:2007xy}. The self-accelerating branch is widely believed to contain a ghost in the spectrum~\cite{Luty:2003vm,Nicolis:2004qq,Koyama:2005tx,Charmousis:2006pn,Gregory:2007xy,Gorbunov:2005zk}. Although the study of perturbations is beyond the scope of this paper, we also expect that the super-accelerating solutions in 6D are likely to have ghosts. (The instability is even more severe in our case, since $\hat{\rho}$ decreases monotonically at late times whereas it is bounded below in DGP.) Note that this instability uncovered here is a nonlinear result which can only be inferred from the full Einstein equations. On a Minkowski background the linear 6D model is stable~\cite{Berkhahn:2012wg}.
\subsection{Contour plot}
\label{sec:contourPlot}
As the above examples show emphatically, our 6D model yields qualitatively very different solutions, depending on the choice of parameters. To study this more systematically,
we now perform a scan over $\rho_i$ and $r_c$, keeping $ H_i R = 0.05 $ fixed. This will allow us, in particular, to understand the border delineating degravitating
and super-accelerating solutions.
The results are shown in Fig.~\ref{fig:contour_plot_full}, where each dot corresponds to one set of parameters for which we ran the numerics.
The green region (also labeled (1)) corresponds to degravitating solutions. As in the example of Sec.~\ref{sec:degrav_sol}, the brane Hubble parameter $H$ tends to zero at late times,
and the effective energy density $ \hat\rho $ is always positive. The red region (also labeled (2)) indicates super-accelerating solutions. As in Sec.~\ref{sec:pathol_sol}, $H$ grows unbounded,
while $ \hat\rho $ eventually becomes negative, indicating a classical instability. Finally, the gray region (labeled (3)) corresponds to parameter choices for which the criticality
bound~\eqref{eq:criticalityBound} is violated. As explained earlier, our coordinate system is ill-defined in this case, and hence we cannot make any statements about solutions in this region.
It turns out that the border between the stable and unstable regions matches perfectly the location in parameter space where
\begin{equation}
f(\tau) \equiv 1- \frac{9R^2}{2r_c^2\gamma} \, ,
\label{eq:def_f_bis}
\end{equation}
first introduced in~\eqref{eq:def_f}, vanishes. This is drawn as a solid line in Fig.~\ref{fig:contour_plot_full}. In the degravitating regime, $ f $ is negative, and in the super-accelerating regime it is positive.
Since $ f $ appears in the denominator on the right-hand side of the $\dot{H}$ equation~\eqref{eq:pJunctCondRConst}, the evolution of $ H $ becomes ill-defined when $f$ vanishes. The system hits a (physical) singularity, where the numerics of course break down.
To better understand the boundary between the stable and unstable regions, Fig.~\ref{fig:contour_plot_zoom} zooms in on the boxed region of Fig.~\ref{fig:contour_plot_full}.
For parameters sufficiently close to the $f = 0$ line, $ f (\tau)$ dynamically approaches zero after a short time, and the system hits a singularity.
The basin of attraction for the singularity corresponds to the yellow region (labeled (4)), in which case one starts in the ``healthy'' region,
and the orange region (labeled (5)), in which case one starts in the ``unstable'' region. This is shown in more detail in Fig.~\ref{fig:f_plot_zoom}.
This yellow-orange attractor region of the singularity, which is hardly visible in Fig.~\ref{fig:contour_plot_full}, can be broadened by injecting
more energy into the bulk initially. This can be achieved by widening the initial Gaussian velocity profile.
We checked that these results are largely unchanged if one uses dust ($ w=0 $) or radiation ($ w=1/3 $) on the brane. Furthermore, we repeated the entire analysis for a different value of the circumference, namely \ $ R = 0.025 H_i^{-1} $, and found similar agreement. In particular, the border between the stable and unstable regimes again coincides with the $ f=0 $ line in parameter space.
\subsection{Interpretation}
\label{sec:interpretation}
The main lesson from the above analysis can be summarized as follows: For sub-critical energy densities, the model is stable if and only if the function $ f(\tau) < 0$.
Using the constraint~\eqref{eq:rhoJunctCond} to eliminate $ \gamma $, this stability condition can be cast into the form
\begin{equation}
\label{eq:stabilityBound}
\frac{\rho}{\rho_{\mathrm{crit}}} > r_c^2 H^2 + 1 - \frac{9R^2}{2 r_c^2} \, .
\end{equation}
If this bound is violated, the model is unstable. The stable and unstable regions are separated by a physical singularity, so it is not possible to evolve dynamically from one region to the other.
It is instructive to compare this result with the analogous situation in the DGP model. In that case, the modified Friedmann equation reads~\cite{Deffayet:2000uy}
\begin{equation}
H^2 = \frac{\rho}{3M_\mathrm{Pl}^2} \pm \frac{\left|H\right|}{r_c^{(5)}}\, ,
\end{equation}
where $r_c^{(5)} \equiv \frac{M_\mathrm{Pl}^2}{2M_5^3}$. The $ - $ sign corresponds to the ``normal'' branch and the $ + $ sign to the ``self-accelerated'' branch. At initial time, this can be rewritten as
\begin{equation}
\frac{\rho_i}{6M_5^3 H_i} = H_i r_c^{(5)} \mp 1
\end{equation}
The ratio $\frac{\rho_i}{6M_5^3 H_i}$, which is the 5D analogue of $ \frac{\rho}{2\pi M_6^4}$, is fixed (up to the choice of branch) for a given crossover scale $r_c^{(5)}$.
Therefore, the DGP parameter space is only one-dimensional. This difference is due to the fact that in 6D there additional freedom in choosing the initial deficit angle.
The resulting DGP ``contour'' plot, shown in Fig.~\ref{fig:DGP_plot}, is remarkably similar to the 6D setup. The green line corresponds to the normal branch of DGP; this branch is stable, and
the effective density $\hat{\rho}$ is positive. The red line is the self-accelerated branch. On this branch, $H$ is always larger than $ H_{\text{self}} \equiv 1/r_c^{(5)} $, and $ \hat\rho $ is always negative.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figures/DGPPlot.pdf}
\caption{The ``contour'' plot for the DGP model consists of two disjoint lines. The green line is the normal branch, which is stable. The red line is the self-accelerated branch, which is unstable.}
\label{fig:DGP_plot}
\end{figure}
Our results generalize this peculiarity of the DGP model to codimension-two. The main differences are: (i) the stable/unstable solutions lie on disconnected branches in the DGP model, whereas they are
separated by a physical singularity in 6D; (ii) there is no criticality bound on $\rho$ in DGP, hence no gray region.
\section{Phenomenology}
\label{sec:phenomen}
The stable/degravitating (green) region of Fig.~\ref{fig:contour_plot_full} is bounded from above by the critical bound~\eqref{eq:criticalityBound}, and from below by the stability bound~\eqref{eq:stabilityBound}.
Since we have analytic expressions for both borders, we can discuss how this stable region depends on model parameters. Of particular interest is whether phenomenologically viable
points can lie inside this region.
\begin{figure*}[htb]
\subfloat[$ H R=0.1 $]{
\includegraphics[width=0.3\textwidth]{figures/contourPlotAnalytic_R_0_1.pdf}
}%
\hfill
\subfloat[$ H R=0.05 $]{
\includegraphics[width=0.3\textwidth]{figures/contourPlotAnalytic_R_0_05.pdf}
}%
\hfill
\subfloat[$ H R=0.01 $]{
\includegraphics[width=0.3\textwidth]{figures/contourPlotAnalytic_R_0_01.pdf}
}%
\caption{Contour plots for different values of $ H R $. The dotted lines correspond to the dynamical regularization discussed in Appendix \ref{ap:dynReg}.
The color scheme is the same as in Fig.~\ref{fig:contour_plot}.}
\label{fig:contour_plots_analytic}
\end{figure*}
Fig.~\ref{fig:contour_plots_analytic} shows three contour plots for different values of $H R $. In the limit $H R \to 0 $, the degravitating region gets squeezed towards the $H r_c=0$ axis, while approaching $ \rho = \rho_{\text{crit}} $ from below. The dotted lines are the corresponding boundaries for the dynamical regularization discussed in Appendix~\ref{ap:dynReg}. As $H R$ decreases, the dotted and solid lines approach each other, implying that the two regularization schemes agree in this limit, as expected.
The bounds~\eqref{eq:criticalityBound} and~\eqref{eq:stabilityBound} imply that sub-critical, stable solutions exists if and only if
\begin{equation}
\label{eq:maxCrossover}
\left( H r_c\right)^2 < \frac{3}{2} \left|H\right| R \,.
\end{equation}
This bound can also be derived in the dynamical regularization, in which case it is only a necessary condition.
For phenomenological reasons, we need $ H r_c \gg 1$ to reproduce standard 4D cosmological evolution on the brane, at least at early times.
Indeed, if instead $ H r_c \lesssim 1 $, then the system will exhibit a 6D behavior. On the other hand, we must have $H R\ll 1$, as mentioned in~\eqref{UV_IR_Limit},
in order for brane physics to admit an effective 4D description. Clearly, these two requirements---$ H r_c \gg 1$ and $H R\ll 1$---are mutually incompatible,
given~\eqref{eq:maxCrossover}. In other words, {\it the model admits no (sub-critical) solutions that are both stable and phenomenologically viable}.
In the super-accelerating (red) region of Fig.~\ref{fig:contour_plot_full}, on the other hand, there is no problem with achieving arbitrarily large values of $Hr_c$.
Fig.~\ref{fig:hubble_4D_regime} shows the Hubble evolution for different values of $r_c$ (black curves), compared to the standard 4D evolution (blue curve).
The matter consists of dust and cosmological constant, with
\begin{equation}
\rho_i^{\text{cc}} = \rho_i^{\text{dust}} = \frac{1}{2} \left( H_i^2 r_c^2 + 0.8 \right) \rho_{\text{crit}} \,.
\end{equation}
As expected, the larger the $r_c$ value, the longer the standard evolution is traced. Once the modification kicks in, however, the evolution becomes unstable and super-accelerating.
This instability, accompanied by a negative effective energy density, should be regarded as strong indications against the physical relevance of those solutions. We expect fluctuations
around such backgrounds to exhibit ghost instabilities, analogous to the DGP model. It would of course be worthwhile to verify this expectation through explicit calculation.
While it would be desirable to further verify this last claim, we think that our current results already suggests that the super-accelerating solutions should not be regarded as consistent alternative cosmologies.
\begin{figure}
\includegraphics[width=0.45\textwidth]{figures/hubblePlot4D_statreg.pdf}
\caption{The Hubble evolution for different values of the cross-over scale $r_c$ (black curves), compared to the standard 4D evolution (blue curve). Since $ H_i r_c > 1 $, these curves all lie deep inside the super-accelerating/red region. As the value of $r_c$ is increased, the solution traces the 4D evolution for longer.}
\label{fig:hubble_4D_regime}
\end{figure}
\section{Conclusion}
\label{sec:conclusion}
In this work, the cosmology of the brane induced gravity model in $ 6 $ dimensions has been investigated. The existence of bulk gravitational waves, and the fact that a (nontrivial) FRW codimension-two brane cannot be embedded in a Minkowski bulk, makes it impossible to derive a local on-brane Friedmann equation as in the DGP case. Therefore, we solved the full (nonlinear) system of bulk-brane equations numerically.
We found that the model can show two qualitatively different behaviors: either the solutions degravitate, \textit{i.e.}, they dynamically approach the static deficit angle solution, or they super-accelerate, \textit{i.e.}, the Hubble parameter grows unbounded. This instability originates from the effective energy density $ \hat\rho $, which sources six-dimensional GR, becoming negative in those cases. It is very likely---though we have not shown this in the present work---that perturbations around those solutions would allow for ghosts, on top of the classical instability of the background itself. It would certainly be desirable to verify this claim; one strong indication for it is that this is exactly what happens in the DGP case: ghosts are present in fluctuations around the self-accelerated branch, which also has $ \hat\rho < 0 $. But in 6D the instability already shows up in the background solution, which is why we already consider them physically irrelevant.
Whether a solution degravitates or super-accelerates depends on the three independent (dimensionless) parameters $ H R $, $ H r_c $ and $ \rho / \rho_{\mathrm{crit}} $. We were able to derive an analytic expression that determines the border between the two regimes and showed that it corresponds to a physical singularity. Thus, a solution can never dynamically evolve from one regime to the other.
Unfortunately, it turned out that the stable, degravitating solutions are not phenomenologically viable because they never lead to an almost 4D behavior, and thus could never match the past history of our universe which is very well described by the standard FRW evolution. On the other hand, phenomenologically interesting parameters $ H r_c \gg 1 $, $ H R \ll 1 $ which are indeed able to mimic a 4D evolution, always lead to an instable behavior once the modification sets in. Unless there is some way to make sense of those instable solutions---which seems very unlikely---we conclude that the BIG model in $ d=6 $ is ruled out (for sub-critical energy densities).
It should be noted that we have not investigated super-critical energy densities. An effective field theory (EFT) analysis in Appendix~\ref{ap:EFT} shows that for large enough values of the regularization scale ($R> M_6^{-1}$) this constitutes the remaining window in parameter space which could allow for a phenomenologically interesting solution. %
Finally, we have not considered a cosmological constant in the bulk. It might be interesting to check how relaxing this assumption would change the size of the healthy region in parameter space.
\begin{acknowledgments}
We thank Felix Berkhahn, Gia Dvali and Michael Kopp for helpful discussions.
FN and RS would like to thank the Department of Physics and Astronomy at the University of Pennsylvania for its hospitality in the course of this work. The work of SH was supported by the DFG cluster of excellence `Origin and Structure of the Universe' and by TRR 33 `The Dark Universe'. The work of FN and RS was supported by the DFG cluster of excellence `Origin and Structure of the Universe'. JK is supported in part by NSF CAREER Award PHY-1145525 and NASA ATP grant NNX11AI95G.
\end{acknowledgments}
|
\section{Introduction}
In~\cite{kopparty2014roots} Kopparty and Wang considered the zero-nonzero
pattern of a univariate polynomial $P(X)$ over ${\mathbb{F}}_q$ and its
relation to the number of roots in ${\mathbb{F}}_q^\ast$. Their main
theorem~\cite[Th.\ 1]{kopparty2014roots} states that a polynomial with
many zeros cannot have long sequences of consecutive coefficients all
being equal to zero. Then in~\cite[Th.\ 2]{kopparty2014roots} they
gave necessary and sufficient conditions for a product of pairwise
different linear factors to have sequences of zero
coefficients of maximal possible length for any polynomial with prescribed number of roots. In
this note we generalize the abovementioned results to polynomials in
more variables.\\
In Section~\ref{sec2} we start by recalling
the results by Kopparty
and Wang. In Section~\ref{sec3} we then present and prove the generalizations.
\section{Univariate polynomials}\label{sec2}
The main theorem in~\cite{kopparty2014roots} is their Theorem 1 which
we present in a slightly stronger version.
\begin{theorem}\label{the1}
Let $P(X) \in {\mathbb{F}}_q[X]$ be a nonzero polynomial of degree at
most $q-2$, say $P(X)=\sum_{i=0}^{q-2}b_iX^i$. Let $m$ be the number
of $x\in {\mathbb{F}}_q^\ast$ with $P(x) \neq 0$. Then there does not
exist any $k \in \{0, \ldots , q-2\}$ where all the $m$ coefficients
$b_k$, $b_{k+1 {\mbox{ mod }} (q-1)}, \ldots, b_{k+m-1 {\mbox{ mod }}
(q-1)}$ are zero.
\end{theorem}
The modification made in Theorem~\ref{the1} is that we consider $k\in
\{0, \ldots , q-2\}$ rather than just $k \in \{0, \ldots ,
q-1-m\}$. The proof in~\cite{kopparty2014roots} is easily modified to
cover this more general situation. Alternatively, one can deduce it by
writing $P(X)=X^sQ(X)$ with $s$ maximal and then
applying~\cite[Th.\ 1]{kopparty2014roots} to $Q(X)$.\\
Obviously, if we consider a product of $q-1-m$ pairwise different
linear factors $X-x$ with $x\neq 0$, this polynomial has exactly $m$
non-roots in ${\mathbb{F}}_q^\ast$ and we have $b_{q-m}=\cdots =b_{q-2}=0$ which is a
sequence of $m-1$ consecutive zero coefficients modulo $q-1$. The
below theorem, corresponding to~\cite[Th.\ 2]{kopparty2014roots},
gives sufficient and necessary conditions for
a sub-sequence of $m-1$ consecutive zeros among $b_0, \ldots ,
b_{q-m-2}$ to exist.
\begin{theorem}\label{the2}
Let $S$ be a subset of ${\mathbb{F}}_q^\ast$ of size $q-1-m$, where $m
\geq 2$ and consider
\begin{equation}
P(X)=\prod_{a \in S}(X-a)=\sum_{i=0}^{q-1-m}b_iX^i. \label{eqtrekant}
\end{equation}
There exists a $k\in \{1, \ldots , q-2m\}$ such that $b_k= \cdots
=b_{k+m-2}=0$ if and only if ${\mathbb{F}}_q^\ast \backslash S$ is
contained in $\gamma H$ for some $\gamma \in {\mathbb{F}}_q^\ast$ and
for some proper multiplicative subgroup $H$ of ${\mathbb{F}}_q^\ast$.
\end{theorem}
Inspecting the proof in~\cite{kopparty2014roots} one sees that for
polynomials of the form~(\ref{eqtrekant}) the
existence of one sub-sequence of $m-1$ consecutive zero coefficients in
$b_0, \ldots , b_{|S|-1}$
is equivalent to the existence of $(q-1)/|H|$ such disjoint sequences.
\begin{proposition}\label{pro1}
Let $P(X)$ be a polynomial as in~(\ref{eqtrekant}) satisfying the
condition of Theorem~\ref{the2}. That is, there exists a $k \in \{1,
\ldots , q-2m\}$ such that $b_k=\cdots
=b_{k+m-2}=0$ where $m=| {\mathbb{F}}_q^\ast \backslash S|$. Write
$d=|H|$ where $H$ is the subgroup corresponding to $P$. The
coefficients $b_{jd}, b_{jd+(d-m)}$, $j=0, \ldots , \frac{q-1}{d}-1$ are nonzero and the only other
possible nonzero coefficients of $P(X)$ are $b_{jd+1},
b_{jd+1},\ldots , b_{jd+(d-m)-1}$, $j=0, \ldots ,
\frac{q-1}{d}-1$.
\end{proposition}
\noindent {\bf{Proof:}} According to~\cite[Proof of Th.\ 2]{kopparty2014roots}, if $P(X)$
satisfies the conditions in Theorem~\ref{the2} then it can be
written
$$\bigg( \sum_{j=1}^{(q-1)/d}b_jX^{(q-1)-jd}\bigg) \cdot U(X)$$
where $U$ is a product of $d-m$ pairwise different expressions $X-x$
with $x \in {\mathbb{F}}_q^\ast$.
\begin{example}\label{ex1}
Let $\alpha$ be a primitive element of ${\mathbb{F}}_{16}$. We first
consider
$$T=\{\beta \mid \beta^3=\alpha^3\}=\{\alpha,\alpha^6,\alpha^{11}\}.$$
The support of $P(X)$ becomes $\{1, X^3, X^6,X^9,X^{12}\}$. If we
choose $T$ to be a subset of $\{ \alpha, \alpha^6, \alpha^{11}\}$ of
size $2$ then the support of $P(X)$ becomes $\{1,
X,X^3,X^4,X^6,X^7,X^9,X^{10},X^{12},X^{13}\}$. Consider next
$$T=\{ \beta \mid \beta^5=\alpha^{10} \}=\{\alpha^2,
\alpha^5,\alpha^8,\alpha^{11}, \alpha^{14}\}.$$
The support of $P(X)$ becomes $\{1, X^5,X^{10}\}$. Finally, if we
choose $T$ to be a subset of $\{\alpha^2,
\alpha^5,\alpha^8,\alpha^{11}, \alpha^{14}\}$ of size 3 then we can
conclude:
\begin{eqnarray}
\{1, X^2,X^5, X^7,X^{10},X^{12}\} \subseteq \mbox{Supp} P \subseteq \{1, X,X^2,X^5, X^6,X^7,X^{10},X^{11}, X^{12}\}.\nonumber
\end{eqnarray}
\end{example}
\section{Multivariate polynomials}\label{sec3}
The crucial observation used in the proof of Theorem~\ref{the1} is
that a univariate polynomial $F(X)$ can at most have $\deg F$
roots. For multivariate polynomials over general fields there does not
exist a similar result as typically such polynomials have infinitely
many roots when the field under consideration is infinite. For
multivariate polynomials over finite fields, however, we do have a
counterpart to the bound used in the proof of Theorem~\ref{the1}. We
describe this bound in terms of roots from $({\mathbb{F}}_q^\ast)^n$ in
Proposition~\ref{pro1} below. To motivate the bound we need a few
results from Gr\"{o}bner basis theory.\\
Let ${\mathbb{F}}$ be a field and $I \subseteq {\mathbb{F}}[X_1, \ldots , X_n]$ an
ideal. Throughout this section assume that an arbitrary fixed monomial
ordering $\prec$ has been chosen. Following~\cite{onorin} we define the
footprint of $I$ by
\begin{eqnarray}
\Delta_\prec (I) &=&\{ X_1^{i_1} \cdots X_n^{i_n} \mid X_1^{i_1}
\cdots X_n^{i_n} {\mbox{ is not }} \nonumber \\
&&\, \, \, \, \, \, \, \, {\mbox{ a leading monomial of any polynomial
in }}I\}.\nonumber
\end{eqnarray}
From~\cite[Prop.\ 4, page 229]{clo} we know that $\{M+I \mid M \in
\Delta_{\prec}(I)\}$ constitutes a basis for ${\mathbb{F}}[X_1, \ldots , X_n]/I$
as a vector space over ${\mathbb{F}}$. Assume $I$ is finite dimensional (which
simply means that $\Delta_\prec(I)$ is a finite set). Consider $\ell$
pairwise different points $P_1, \ldots , P_\ell$ in the zero-set of
$I$ (over ${\mathbb{F}}$). The map ${\mbox{ev}}: {\mathbb{F}}[X_1, \ldots , X_n]/I
\rightarrow {\mathbb{F}}^\ell$ given by ${\mbox{ev}}(F+I)=(F(P_1), \ldots ,
F(P_\ell))$ is a surjective vector space homomorphism (surjectivity
follows by Lagrange interpolation). Therefore
\begin{equation}
\ell \leq
|\Delta_\prec(I)|\label{eqabove}
\end{equation}
(this result is often called the footprint bound
\cite{onorin}). In particular we derive:
\begin{proposition}\label{pro2}
Consider $P(\vec{X}) \in {\mathbb{F}}_q[X_1, \ldots , X_n]$ with
leading monomial equal to $X_1^{i_1}\cdots X_n^{i_n}$ such that $i_s
< q-1$ for $s=1, \ldots ,n$. Let $m$ be the number of elements in
$({\mathbb{F}}_q^\ast)^n$ which are not roots of $P$. Then $m \geq
\prod_{s=1}^n (q-1-i_s)$.
\end{proposition}
\noindent {\bf{Proof:}} The proor follows by applying (\ref{eqabove}) to the ideal $I=\langle
P(\vec{X}), X_1^{q-1}-1, \ldots , X_n^{q-1}-1 \rangle$. The footprint
of this ideal
is a subset of
\begin{eqnarray}
\{X_1^{j_1} \cdots X_n^{j_n} \mid 0 \leq j_s < q-1,
s=1, \ldots , n, {\mbox{ not all }} j_s {\mbox{ satisfy }} i_s \leq j_s\}.\nonumber
\end{eqnarray}
Therefore,
the number of non-roots is at least $| \{ (j_1, \ldots , j_n) \mid i_s
\leq j_s < q-1, s=1, \ldots , n\}|$.
Observe that for $n=1$ the statement in Proposition~\ref{pro2} is
but the well-known fact that a multivariate polynomial $P$ has
at least $q-1-\deg P$ non-roots in ${\mathbb{F}}_q^\ast$.\\
Before giving the generalization of Theorem~\ref{the1} we introduce
the set $U(q,m,n)$. This set shall play the role as did the set of consecutive monomials
$\{X^{q-1-m}, \ldots , X^{q-2}\}$ in connection with Theorem~\ref{the1}.
\begin{definition}
Given positive integers $m$ and $n$ let
$${\mathcal{M}}(q,n)=\{ X_1^{i_1}\cdots X_n^{i_n} \mid 0 \leq i_1,
\ldots , i_n < q-1\},$$
$$U(q,m,n)=\{X_1^{i_1} \cdots X_n^{i_n} \in {\mathcal{M}}(q,n) \mid \prod_{s=1}^m(q-1-i_s)\leq m\}.$$
\end{definition}
\begin{theorem}\label{the3}
Given a positive integer $n$ write $\vec{X}=(X_1, \ldots , X_n)$ and
consider a nonzero polynomial $P(\vec{X}) \in
{\mathbb{F}}_q[\vec{X}]$ with $\deg_{X_i}P<q-1$, $i=1, \ldots ,
n$. Let $m$ be the number of $\vec{x} \in ({\mathbb{F}}_q^\ast)^n$ with
$P(\vec{x})\neq 0$. Then there does not exist any $(k_1, \ldots
,k_n) \in \{0, 1, \ldots, q-2\}^n$ such that
\begin{eqnarray}
\mbox{Supp} (X_1^{k_1} \cdots X_n^{k_n} P(\vec{X}) {\mbox{ mod }}
\{X_1^{q-1}-1, \ldots , X_n^{q-1}-1\}) \cap U(q,m,n)=\emptyset. \nonumber
\end{eqnarray}
\end{theorem}
Observe that for $n=1$ we have $U(q,m,n)=\{X^{q-1-m}, \ldots ,
X^{q-1-1}\}$ which is a list of $m$ consecutive monomials. Hence,
Theorem~\ref{the3} is a natural generalization of Theorem~\ref{the1}
to polynomials in more variables.
\noindent {\bf{Proof:}}
Let $P(\vec{X})$ and $m$ be as in the theorem. Aiming for a
contradiction assume that an $X_1^{k_1} \cdots X_n^{k_n}$ exists such that
\begin{eqnarray}
\mbox{Supp} \big( X_1^{k_1} \cdots X_n^{k_n} P(\vec{X}) {\mbox{ mod }}
\{X_1^{q-1}-1, \ldots , X_n^{q-1}-1\} \big) \cap U(q,m,n) =\emptyset . \nonumber
\end{eqnarray}
According to Proposition~\ref{pro2}
$$X_1^{k_1} \cdots X_n^{k_n} P(\vec{X}) {\mbox{ mod }}
\{X_1^{q-1}-1, \ldots , X_n^{q-1}-1\}$$
has at least $m+1$
non-roots in ${\mathbb{F}}_q^\ast$; and so has $P(\vec{X})$.\\
The generalization of Theorem~\ref{the2} is as follows:
\begin{theorem}\label{the4}
Consider sets $S_i\subseteq {\mathbb{F}}_q^\ast$, $i=1, \ldots ,
n$. Write $s_i=|S_i|$ and assume $0<s_i <q-1$, $i=1, \ldots , n$, not
all $s_i$ being equal to $q-2$. Define
$T_i={\mathbb{F}}_q^\ast\backslash S_i$ and let $t_i=|T_i|$ and
$m=\prod_{i=1}^nt_i$ (by the above assumption on $s_i$ we have $m \geq
2$). Consider
\begin{equation}
P(\vec{X})=\prod_{i=1}^n\prod_{x \in
S_i}(X_i-x).\label{eqtrekantm}
\end{equation}
There exists an $X_1^{k_1}\cdots X_n^{k_n}$ with $0 <
k_1, \ldots , k_n < q-1$ such that
\begin{eqnarray}
\mbox{Supp} (X_1^{k_1} \cdots
X_n^{k_n}P(\vec{X}) {\mbox{ mod }} \{X_1^{q-1}-1, \ldots ,
X_n^{q-1}-1\}) \cap U(q,m-1,n)=\emptyset \label{eqsnabel}
\end{eqnarray}
if and only if for $i=1, \ldots , n$ it holds that $T_i$ is contained
in $\gamma_iH_i$ for some $\gamma_i\in {\mathbb{F}}_q^\ast$ and for some
proper multiplicative subgroup $H_i$ of ${\mathbb{F}}_q^\ast$.
\end{theorem}
We note that the role of the assumption $m \geq 2$ is to make $U(q,m-1,n)$
non-empty.\\
As already observed, for $n=1$ we have $U(q,m-1,n)=\{X_1^{q-m},
\ldots , X_1^{q-2}\}$. Therefore for $n=1$ the assumption~(\ref{eqsnabel}) is equivalent
to saying that
$${\mathcal{M}}(q,n) \backslash \big( \mbox{Supp} P \cup
U(q,m-1,n) \big)$$
contains a set $D$ such that
$$U(q,m-1,n) \subseteq
X_1^{k_1}D {\mbox{ mod }} \{X_1^{q-1}-1\}$$
(a similar remark does not
hold for $n > 1$.) In other words, for $n=1$, ${\mathcal{M}}(q,n) \backslash
\mbox{Supp} P$ contains besides $U(q,m-1,n)$ also a translated copy of
$U(q,m-1,n=1)$ which is disjoint from $U(q,m-1,n)$. We have argued that
Theorem~\ref{the4} reduces to Theorem~\ref{the2} in the case that
$n=1$.\\
Turning to the general case of $n \geq 1$ one sees by inspection that $U(q,m-1,n) \subseteq {\mathcal{M}}(q,n)
\backslash \mbox{Supp} P$. The condition $0<k_i<q-1$, $i=1,\ldots
, n$ means that the sets assumed to exist or proved to exist,
respectively, in Theorem~\ref{the4} are different from $U(q,m-1,n)$
itself; but they may have an overlap with this set.\\
Before giving the proof we illustrate the theorem with an
example.
\begin{example}\label{ex2}
This is a continuation of Example~\ref{ex1} where we considered
polynomials $P(X) \in {\mathbb{F}}_{16}[X]$ of the form~(\ref{eqtrekant}) satisfying the
conditions in Theorem~\ref{the2}. In this example we consider a
polynomial $P(X_1,X_2) \in {\mathbb{F}}_{16}[X_1,X_2]$ of the
form~(\ref{eqtrekantm}) satisfying the condition in
Theorem~\ref{the4}. Choosing $T_1=\{\alpha^2,\alpha^5,
\alpha^8,\alpha^{11},\alpha^{14}\}$ and
$T_2=\{\alpha,\alpha^6,\alpha^{11}\}$ we get that the support of
$P(\vec{X})$ is
\begin{eqnarray}
\{1, X_1^5,X_1^{10}, X_2^3, X_1^5X_2^3,X_1^{10}X_2^3,
X_2^6, X_1^5X_2^6,X_1^{10}X_2^6, \, \, \, \, \, \, \nonumber \\
X_2^9, X_1^5X_2^9,X_1^{10}X_2^9,
X_2^{12}, X_1^5X_2^{12},X_1^{10}X_2^{12}\}. \nonumber
\end{eqnarray}
Clearly, $m=5 \cdot 3=15$.
In Figure~\ref{figo1} the support is illustrated with diamonds. A
set $D$ is illustrated with filled circles. This set satisfies that $D
\subseteq {\mathcal{M}}(q,n) \backslash \mbox{Supp} P$ and that
$$X_1^5X_2^3D {\mbox{ mod }} \{X_1^{q-1}-1, \ldots X_m^{q-1}-1\}=U(q,m-1,2).$$
Hence,
$$\mbox{Supp} (X_1^5X_2^3 P {\mbox{ mod }} \{X_1^{15}-1,X_2^{15}-1\}) \cap
U(q,m-1,2) = \emptyset.$$
\begin{figure}
\begin{center}
\includegraphics[width=50mm]{figurver2.eps}
\end{center}
\caption{The situation in Example~\ref{ex2}.}
\label{figo1}
\end{figure}
\end{example}
It is possible to give a proof of Theorem~\ref{the4} which as a main
tool uses Theorem~\ref{the2} and Proposition~\ref{pro1} in combination with a study of the shape
of $U(q,m-1,n)$. Using this approach the proof of the ``if'' part
becomes straight forward whereas the proof of the ``only if''
part becomes technical and
requires more care. Instead of stating the technical proof of the
``only if'' part we shall present a self contained
proof of the ``only if'' part based
on the technique from~\cite{kopparty2014roots}. Our proof calls for
the following lemma which has some interest in itself. We state the
lemma in a slightly more general version than shall be needed (we will employ
the lemma with ${\mathbb{F}}={\mathbb{F}}_q$ and $A_1=\cdots = A_n= {\mathbb{F}}_q^\ast$).
\begin{lemma}\label{lem1}
Given a field ${\mathbb{F}}$ let $A_1, \ldots , A_n \subseteq
{\mathbb{F}}$ be finite sets. Consider proper subsets $B_1 \subsetneq
A_1, \ldots , B_n \subsetneq A_n$ and write
$$P(\vec{X})=\prod_{i=1}^n \prod_{x \in B_i}(X_i-x).$$
Assume that $G(\vec{X}) \in {\mathbb{F}}[\vec{X}]$ is a polynomial
with $\deg_{X_i} G<| A_i|$, $i=1, \ldots , n$ such that
$$\{ \vec{x} \mid \vec{x} \in A_1 \times \cdots \times A_n,
F(\vec{x})=0\} \subseteq \{ \vec{x} \mid \vec{x} \in A_1 \times \cdots \times A_n,
G(\vec{x})=0\}.$$
Then $F(\vec{X})$ divides $G(\vec{X})$.
\end{lemma}
\noindent {\bf{Proof:}} It is enough to prove that $(X_s-x)$ divides $G(\vec{X})$ for
arbitrary $s \in \{1, \ldots , n\}$ and $x \in B_s$. We can write
$G(\vec{X})=Q(\vec{X})(X_s-x)+R(\vec{X})$ where $R(\vec{X})$ is a
polynomial in ${\mathbb{F}}[X_1, \ldots , X_{s-1},X_{s+1}, \ldots ,
X_n]$ and where $\deg_{X_i} R < |A_i|$ for $i \in \{1, \ldots , s-1,
s+1, \ldots , n\}$. We observe that $(\alpha_1, \ldots ,
\alpha_{s-1},x,\alpha_{s+1}, \ldots , \alpha_n)$ is a root of $P$ and
thereby also of $G$, for
all $(\alpha_1, \ldots , \alpha_{s-1},\alpha_{s+1}, \ldots ,
\alpha_n)$ in $A_1 \times \cdots \times A_{s-1} \times A_{s+1} \times
\cdots \times A_n$. But then $(\alpha_1, \ldots , \alpha_{s-1},\alpha_{s+1}, \ldots ,
\alpha_n)$ is a root of $R$ and from the Chinese remainder theorem it
follows that $R(\vec{X})=0$.\\
We are now ready to prove Theorem~\ref{the4}.\\
\noindent {\bf{Proof:}} Assume that there
exists an $X_1^{k_1}\cdots X_n^{k_n}$ with $0 <
k_1, \ldots , k_n < q-1$ such that~(\ref{eqsnabel}) holds true.
Write
$$G(\vec{X})=X_1^{k_1} \cdots X_n^{k_n}P(\vec{X}) {\mbox{ mod }}
\{X_1^{q-1}-1, \ldots , X_n^{q-1}-1\}.$$
Clearly the roots of $P(\vec{X})$ in $({\mathbb{F}}_q^\ast)^n$ are
also roots of $G(\vec{X})$. Hence, by Lemma~\ref{lem1},
$G(\vec{X})=Q(\vec{X})P(\vec{X})$ for some $Q(\vec{X}) \in
{\mathbb{F}}_q[\vec{X}]$. Recall that $X_1^{s_1}\cdots X_n^{s_n}$ is
the leading monomial of $P(\vec{X})$. If we consider a monomial $N$ such that
$NX_1^{s_1} \cdots X_n^{s_n} \in {\mathcal{M}}(q,n)$ then either $N=1$
or $NX_1^{s_1} \cdots X_n^{s_n} \in U(q,m-1,n)$. From
assumption~(\ref{eqsnabel}) it therefore follows that
$G(\vec{X})=\alpha P(\vec{X})$ for some $\alpha \in
{\mathbb{F}}_q^\ast$. This implies that
$$P(\vec{X}) (X_1^{k_1} \cdots X_n^{k_n}-\alpha)=0 {\mbox{ mod }}
\{X_1^{q-1}-1, \ldots , X_n^{q-1}-1\}.$$
However, then all non-roots of $P(\vec{X})$ in $({\mathbb{F}}_q^\ast)^n$ --
that is the elements of $T_1 \times \cdots \times T_n$ -- must be
roots of $X_1^{k_1}\cdots X_n^{k_n}-\alpha$. In other words, for
$\vec{x} \in T_1\times \cdots \times T_n$ we have
$x_i^{k_i}=\alpha_i$, $i=1, \ldots , n$ where
$\prod_{i=1}^n \alpha_i =\alpha$. Consider an $i$ such that $t_i
>1$. Let $y$ and $z$ be two different elements in $T_i$. For fixed
$x_j \in T_j$, $j \in \{1, \ldots , i-1, i+1, \ldots , n\}$ both
$(x_1, \ldots ,x_{i-1},y,x_{i+1}, \ldots , x_n)$ and $(x_1, \ldots
,x_{i-1},z,x_{i+1}, \ldots , x_n)$ satisfy that
they produce the value $\alpha$
when plugged into
$X_1^{k_1}\cdots X_n^{k_n}$.
Hence,
$\alpha_1, \ldots , \alpha_n$ are unique. Let $H_i=\{\beta \in
{\mathbb{F}}_q^\ast \mid \beta^{k_i}=1\}$ (which is a proper subgroup
of ${\mathbb{F}}_q^\ast$ as $0 < k_i < q-1$), and $\gamma_i \in
T_i$. Then $T_i \subseteq \gamma_i H_i$.\\
We next prove the ``if'' part of the theorem. Assume that $T_i
\subseteq \gamma_i H_i$, $i=1, \ldots , n$ and write $d_i = |
H_i|$. Define
$$W_i=\big\{ X_i^v \mid v\in \{jd_i, \ldots , jd_i+(d_i-t_i)\mid j=0,
\ldots , \frac{q-1}{d_i}-1\} \big\}$$
and
$$W=\{X_1^{v_1} \cdots X_n^{v_n} \mid X_i^{v_i} \in W_i, i=1,
\ldots , n\}.$$
Proposition~\ref{pro1} tells us that $\mbox{Supp} P \subseteq W$ and by
inspection we find that $U(q-1,m-1,n)$ is contained in
${\mathcal{M}}(q,n)\backslash W$. By symmetry we have
$$X_1^{k_1} \cdots X_n^{k_n} W {\mbox{ mod }} \{X_1^{q-1}-1, \ldots ,
X_n^{q-1}-1\}=W$$
for all $(k_1, \ldots , k_n)$ where for $i=1, \ldots , n$,
$k_i=\ell_id_i$ for some $\ell_i$. The theorem follows.
\section*{Acknowledgements}
This work was supported by the Danish Council for Independent
Research, grant no.\ DFF-4002-00367.
\bibliographystyle{plain}
|
\section*{\Large Supplementary Materials}
\end{center}
\paragraph{This PDF file includes the following:}
\newenvironment{myitemize}
{ \begin{itemize}
\setlength{\itemsep}{0pt}
\setlength{\parskip}{0pt}
\setlength{\parsep}{0pt} }
{ \end{itemize} }
\begin{myitemize}
\itemsep0em
\setlength{\itemindent}{4.5pt}
\item[] Material and Methods
\item[] Supplementary Text
\item[] Figures S1 to S7
\item[] Tables S1 to S4
\item[] References 37--64
\end{myitemize}
\paragraph*{\large Materials and Methods}
\subparagraph{Selection of Light Curves.}
To include a SN in our sample, we require a photometric measurement before 5 days after {\it B}-band maximum,
at least one photometric measurement after {\it B}-band maximum,
and at least three separate epochs having a signal-to-noise ratio (S/N) $>5$ measurement. Only photometry through 50 days
after {\it B}-band maximum enters the fitting with MLCS2k2.
A fraction of the SNe in our sample were followed by multiple teams and published in more than one
of the light-curve collections listed in Table~\ref{tab:lcs}.
We selected the light curve according to the order in Table~\ref{tab:lcs}, with light-curve collections at the beginning of the list having precedence.
Rearranging the order does not significantly change the Hubble-residual statistics that we measure.
\subparagraph{MLCS2k2 Light-Curve Fitting.}
Except for CSP SN~Ia photometry, for which we use magnitudes in the natural system, all fitting is performed to
fluxes in the Landolt standard system \cite{landolt92}. Heliocentric SN~Ia redshifts are transformed to the cosmic microwave background (CMB) rest frame using the dipole velocity \cite{fixencheng96}. We correct fluxes for Milky Way extinction \cite{schlaflyfinkbeinerSFD11} using the O'Donnell 1994 reddening law \cite{odonnell94}. Since MLCS2k2 performs light-curve fitting to rest-frame fluxes, the observed SN fluxes are transformed, prior to fitting, from the observer frame into the rest frame using a K-correction \cite{nug02}. When performing MLCS2k2 fitting using SNANA, we apply an exponential prior on $A_V$ with $\tau=0.3$\,mag (PRIOR\_AVEXP $= 0.3$) convolved with a Gaussian kernel having $\sigma=0.02$\,mag (PRIOR\_AVRES $=$ 0.02). We obtain similar Hubble-residual scatter among SNe in UV-bright environments when priors instead with $\tau=0.1$\,mag and $\tau=0.5$\,mag are used while fitting.
To calculate the Hubble residuals of individual SNe, we first find the redshift-distance relation that minimizes the $\chi^2$ value of the fit to the MLCS2k2 distance moduli. For the purpose of fitting for the redshift-distance relation, we add an uncertainty of 0.07\,mag in quadrature to the statistical uncertainty on each distance computed by MLCS2k2. After minimizing the $\chi^2$ value, we repeat the fit with only the SNe having Hubble residuals smaller than 0.3\,mag. The best-fitting relation is not significantly affected by the specific values of the additional uncertainty, or of the criterion used to identify outlying SNe.
Figures~\ref{fig:LCcuts} and \ref{fig:DeltaVSUV} show the light-curve parameter cuts that we apply to the sample of SN~Ia light curves.
A principal purpose of light-curve cuts used in cosmological analyses is to remove SNe that may be systematically offset from the light-curve width/color/luminosity relation, that show extremely large dispersion around the redshift-distance relation, or that have light curves that indicate high reddening. As Figs. \ref{fig:LCcuts}\ and \ref{fig:DeltaVSUV} show, we selected light-curve cuts that achieved these aims.
As Figure~\ref{fig:DeltaVSUV} shows, fast-declining ($\Delta > 0.4$) SNe~Ia in our sample are absent from UV-bright host-galaxy environments.
Since these fast-declining SNe have Hubble residuals that are systematically low (Fig.~\ref{fig:LCcuts}),
including these SNe in our analysis sample would increase the Hubble-residual scatter of the full SN~Ia sample, but not affect the UV-bright subset.
Therefore, the addition of the fast-declining SNe would increase the statistical significance of the
comparatively small Hubble-residual scatter we observe among SNe in UV-bright regions.
Nonetheless, we exclude the fast-declining SNe to show that SNe~Ia having typical light curves exhibit a smaller Hubble-residual scatter, when found in regions having high UV brightness.
\subparagraph{Calculating Statistical Significance.}
Through Monte Carlo simulations, we have assessed the probability that the small Hubble-residual scatter we measure for SNe~Ia in UV-bright
host-galaxy apertures is only a random effect.
An additional possibility is that, from random effects, the SNe erupting instead, for example, in
UV-faint regions could have exhibited small apparent Hubble-residual
scatter. If we expect that we would have reported such an alternative
pattern as a discovery, then the $p$-values we compute should be
adjusted to account for multiple searches.
For $n$ separate searches, the probability of finding at least one random pattern having significance $p$ is
$p' = 1 - (1 - p)^n$.
We have calculated, assuming a single search, a $p$-value of 0.2\%\ for the measured Hubble-residual scatter of 0.073$\pm$0.012\,mag for SNe~Ia in regions with high NUV surface brightness (see Table~\ref{tab:stddev}).
Assuming instead $n = 4$ separate searches, for example, would yield an adjusted $p$-value of 0.8\%.
However, a principal focus of our analysis from its inception was to identify SNe~Ia in
UV-bright, star-forming environments using {\it GALEX} imaging, motivated by
a recent analysis of H$\alpha$ flux near explosion sites
\cite{rigaultcopin13}.
Indeed, while high UV surface brightness within the $r=$ 5\,kpc\
aperture provides strong evidence for a young stellar population,
interpretation of low average UV surface brightness is less
straightforward. As shown in Figure~\ref{fig:mosaic}, the apertures
with lower NUV surface brightness consist of a heterogeneous
mix of SNe in low-mass galaxies having small physical sizes, SNe with
large host offsets, and SNe erupting in old stellar populations.
Given the mixed set of environments, it probably would have been unlikely that we would
feel we had found a robust pattern. If SNe with very small scatter
exploded from the oldest stellar populations, we also note that such
an association probably would already have been identified using the
host-galaxy morphology, spectroscopy, or optical and infrared photometry \cite{hi09b,guptadandrea11,dandrea11,childressaldering13,johanssonthomas13,haydengupta13,pansullivan14}.
\subparagraph{KPNO Image Reduction and Calibration.}
We acquired {\it ugriz} imaging of SN~Ia host galaxies with the T2KA and T2KB cameras mounted on the Kitt Peak National Observatory (KPNO) 2.1\,m telescope in June 2009, March 2010, May 2010, October 2010, and December 2010. The raw images were processed using standard IRAF\footnote{http://iraf.noao.edu/} reduction routines. A master bias frame was created from the median stack of the bias exposures taken each night, and subtracted from the flat field and from images of the host galaxies. For each run, we constructed a median dome-flat exposure that we used to correct the host-galaxy images. To improve the flat-field correction, a median stack of all host-galaxy frames was computed after removing all objects detected using SExtractor \cite{bert96}.
Even after dividing images by the flat field and by stacked, object-subtracted science images, the edge to the south of the T2KB CCD array showed a $>5$\% spatially dependent background variation. To remove this poorly corrected region of the detector, we trimmed the 400 pixel columns closest to south edge of the T2KB pixel array, after overscan subtraction. We constructed a fringe model for {\it z}-band images from object-subtracted science images, and scaled the model to best remove fringing from the affected images.
\subparagraph{Lick Observatory Image Reduction and Calibration.} We used the bias and flat-field corrected {\it BVRI} ``template'' images of the host galaxies of SNe~Ia acquired after the SNe had faded, as part of the LOSS follow-up program \cite{fili01}. The imaging was obtained using the 0.76\,m Katzman Automatic Imaging Telescope (KAIT) and the 1\,m Anna Nickel telescope at Lick Observatory \cite{ganeshalingam10}. During the period when the observations were taken, the CCD detector that was mounted on KAIT was replaced twice, while the {\it BVRI} filter set was replaced once.
\subparagraph{Astrometric and Photometric Calibration of KPNO and Lick Observatory Imaging.}
We used the \textit{Astrometry.net} routine \cite{lang10}, which matches sources in each input image against positions in the USNO-B catalog, to generate a World Coordinate System (WCS) for each KPNO\,2.1\,m, KAIT, and Nickel image. After updating each image with the WCS computed by \textit{Astrometry.net}, we performed a simultaneous astrometric fit that included every exposure of each host galaxy across all passbands using the {\tt SCAMP} package \cite{bertin06}, and the 2MASS \cite{skru06} point-source catalog as the reference.
The resulting WCS solution was passed to {\tt SWarp} \cite{bert02} and used to resample all images to a common pixel grid. To extract stellar magnitudes, we next used {\tt SExtractor} \cite{bert96} to measure magnitudes inside circular apertures with $4''$ radius. {\tt PSFEx} \cite{bertin11} was used to fit a Moffat point-spread function (PSF) model and calculate an aperture correction for each image.
We computed {\it ugriz} or {\it BVRI} zeropoints from the stellar locus using the publicly available {\tt big macs}\footnote{https://code.google.com/p/big-macs-calibrate/} package developed by the authors \cite{kellyvonderlinden14}. The routine synthesizes the expected stellar locus for an instrument and detector combination using the wavelength-dependent transmission and a spectroscopic model of the SDSS stellar locus.
\subparagraph{Host-Galaxy Photometry.}
To measure UV emission from host galaxies, we used both FUV and NUV imaging from the All-Sky Imaging Survey and Medium Imaging Survey performed by the {\it GALEX} \cite{martinfanson05} satellite. The {\it GALEX} FUV passband has a central wavelength of 1528\,\AA\ and a PSF full width at half-maximum intensity (FWHM) of $\sim6''$, while the central wavelength of the NUV passband is 2271\,\AA\ and the PSF FWHM is $\sim4.5''$. Photometry at optical wavelengths was measured from {\it ugriz} images taken with the SDSS or the KPNO 2.1\,m telescope, or from {\it BVRI} images taken with KAIT or the 1\,m Nickel telescope at Lick Observatory. All optical images were convolved to have a PSF of $\sim4.5''$ to match that of the {\it GALEX} NUV images before extracting host-galaxy photometry.
When assembling mosaics of host-galaxy images, we record the MJD of each image. We exclude exposures taken from between two weeks prior to through 210 days past the date when the SN reached {\it B}-band maximum.
For the SN host galaxies that have images without contaminating flux, we resample all exposures to a common grid centered on the SN explosion coordinates using the {\tt SWarp} software program. We measure the host-galaxy flux within a circular $r=$ 5\,kpc\ aperture, and compute K-corrections using {\tt kcorrect} \cite{bl07}.
If the uncertainty of a magnitude in a specific bandpass exceeds 0.3\,mag, then we use the magnitude synthesized from the {\tt kcorrect} model that best fits the full set of measured magnitudes, instead of adding a K-correction to the measured magnitude.
We adjust surface-brightness estimates calculated from K-corrected magnitudes for $(1 + z)^{-4}$ cosmological dimming.
When both UV and optical host photometry are available, we estimate the star-formation surface density using the PEGASE2 \cite{fi99} stellar population synthesis models. In Table~\ref{tab:data}, we list measured light-curve parameters, host-galaxy measurements, and Hubble residuals.
\paragraph*{\large Supplementary Text}
\subparagraph{Additional Evidence for a Low-Scatter Population from SNe with Multiple Light Curves.}
A fraction of SNe have separate light-curve measurements published in the LOSS collection as well as in a second collection of light curves (CfA2, CfA3, CfA4, or CSP).
After completing an advanced draft of the paper, we compared the pairs of MLCS2k2 $R_V=1.8$ distance moduli to these SNe.
Since these distances are not affected by the peculiar motion of the host galaxies (unlike $\mu_z$), we can additionally study $z<0.02$
SNe which are not part of the Hubble-residual sample.
Figure~\ref{fig:distComp} shows that there is greater agreement between the distance moduli
to the SNe~Ia that explode in regions with high NUV surface brightness.
This provides additional evidence, discovered after identifying the low-scatter population, that SNe~Ia in UV-bright regions yield more precise distances.
\subparagraph{$R_V$ and the Hubble Residuals of SNe in Regions with High NUV Surface Brightness.}
Figure~\ref{fig:RV} shows Hubble residuals against $A_V$ for MLCS2k2 distance moduli computed using $R_V=1.8$ and $R_V=3.1$.
\subparagraph{FUV Surface Brightness.}
Figure \ref{fig:FUV} shows that the FUV surface brightness yields comparably small Hubble-residual scatter as NUV surface brightness.
Among the Hubble residuals of the $N=$ 17\ SNe~Ia in environments brighter than 24.9} %~mag~arcsec$^{-2}$\,mag\,arcsec$^{-2}$, the root-mean-square scatter is 0.074$\pm$0.012\,mag. When we examine only the $N=$ 10\ SNe~Ia with distance modulus statistical uncertainty $\sigma_{\mu_{\rm SN}} < 0.075$\,mag, the root-mean-square scatter is 0.067$\pm$0.014\,mag.
Additional statistics are listed in Table~\ref{tab:stddev}.
\subparagraph{Color-Composite Images of Explosion Environments.}
Figure~\ref{fig:mosaic} displays panels of SDSS color-composite images of the host-galaxy environments for
SNe~Ia whose circular $r=$ 5\,kpc\ apertures have high or low average NUV surface brightness.
These images show that SNe that occur within the host-galaxy outskirts, as well as in small low-mass galaxies, have
low average NUV surface brightness within the aperture.
\subparagraph{Si~II Velocity.}
In Figure~\ref{fig:NUVsiiv}, we show the velocities of SNe~Ia that have \ionpat{Si}{ii} measurements within seven days of
{\it B}-band maximum brightness against NUV surface brightness within the circular $r=$ 5\,kpc\ aperture. The highly standardizable population
found in high surface brightness regions shows no strong differences in the distributions of \ionpat{Si}{ii} velocities
near maximum brightness.
\begin{table*}[htp!]
\centering
\begin{tabular}{ccc}
Light-Curve Dataset&Filters&References \\
\hline
Lick Observatory Supernova Search (LOSS) & {\it BVRI} & \cite{ganeshalingam10,leaman11,li11a} \\
CfA2 & {\it (U)BVRI} & \cite{jhakirshner06} \\
CfA3 & {\it (U)BVRIr'i'} & \cite{hi09a} \\
CfA4 & {\it (U)BVr'i'} & \cite{hickenchallis12} \\
Carnegie Supernova Project (CSP) & {\it (u)BVgri} & \cite{contrerashamuy10,stritzingerphillips11} \\
\hline
\end{tabular}
\caption{SN~Ia light-curve samples used for the analysis. When fitting SN light curves, we did not include measured {\it U}- or {\it u}-band magnitudes. }
\label{tab:lcs}
\end{table*}
\label{sec:sample}
\begin{table*}
\centering
\begin{tabular}{lc}
\hline
Criterion & SNe \\
\hline
$N_{\rm DOF} > 5$ & 307\\
$\chi^2_{\nu} < 2$ & 292\\
$z>0.02$ & 165\\
$A_V < 0.4$\,mag & 143\\
$-0.35 < \Delta < 0.4$ & 107\\
MLCS2k2 & 107 \\
Not SNF20080522-000 & 106 \\
MW $A_V<0.5$\,mag & 103 \\
\hline
Uncontaminated FUV; NUV & 83 \\
$R_V=1.8$ Hubble residual $<0.3$\,mag & 77 \\
\hline
Uncontaminated FUV; NUV; {\it ugriz} or {\it BVRI} & 67 \\
$R_V=1.8$ Hubble residual $<0.3$\,mag & 61 \\
\hline
\end{tabular}
\caption{Construction of SN sample. $\chi^2_{\nu}$ is the reduced chi-squared statistic calculated for the MLCS2k2 ($R_V=1.8$) light-curve fits. $A_V$ is the best-fitting extinction value computed by MLCS2k2. Milky Way extinction is computed from foreground dust maps \cite{schlaflyfinkbeinerSFD11}. We exclude host-galaxy images taken from 14 days prior to through 210 days after $B$-band maximum light. }
\label{tab:sample}
\end{table*}
\begin{figure*
\centering
\includegraphics[angle=0,width=5in]{LOSS_nuvSB_dldiff.pdf}
\caption{Difference between MLCS2k2 distance moduli $\mu$ measured using LOSS and a second published light curve against mean NUV surface brightness within 5\,kpc\ of SN.
SNe that erupt in regions with a high near-UV surface brightness may show greater agreement between distance measurements from independent light curves.
Diamonds mark SNe having $z>0.02$ (included in the Hubble-residual sample), while squares mark SNe having $z<0.02$.
The parallel vertical lines show the 24.45} %~mag~arcsec$^{-2}$\,mag\,arcsec$^{-2}$ (dashed) and 24.7} %~mag~arcsec$^{-2}$\,mag\,arcsec$^{-2}$~(solid) NUV surface brightness upper limits used to construct samples of SNe~Ia having low Hubble-residual scatter.
}
\label{fig:distComp}
\end{figure*}
\begin{figure*
\centering
\subfigure{\includegraphics[angle=0,width=3.1in]{AV_18_HR.pdf}}
\subfigure{\includegraphics[angle=0,width=3.1in]{DELTA_18_HR.pdf}}
\caption{Hubble residuals plotted against MLCS2k2 $R_V=1.8$ light-curve parameters. When constructing our sample of SNe~Ia light curves, we apply the $A_V<$ 0.4\,mag upper limit shown in the left panel by the vertical line. We include only SNe with decline-rate parameter $-0.35$\ $< \Delta <$ 0.4\ marked by the
pair of vertical lines in the right panel.
}
\label{fig:LCcuts}
\end{figure*}
\begin{figure*
\centering
\subfigure{\includegraphics[angle=0,width=3.1in]{NUV_Delta.pdf}}
\caption{MLCS2k2 light-curve decline parameter $\Delta$ against mean NUV surface brightness within $r=$ 5\,kpc\ aperture around SN.
More rapidly declining SNe~Ia have greater values of $\Delta$ and are, on average, less luminous explosions.
The parallel horizontal lines are the $-0.35$~$< \Delta <$~0.4~limits that we apply to construct our
light-curve sample, while the parallel vertical lines show the 24.45} %~mag~arcsec$^{-2}$\,mag\,arcsec$^{-2}$ (dashed) and 24.7} %~mag~arcsec$^{-2}$\,mag\,arcsec$^{-2}$ (solid) NUV surface brightness upper limits.
}
\label{fig:DeltaVSUV}
\end{figure*}
\begin{figure*
\centering
\subfigure{\includegraphics[angle=0,width=4.5in]{FUV_MLCS2k2_18_HR.pdf}}
\caption{SNe~Ia in regions with high NUV (upper) and FUV (lower) surface brightness ($<$24.7} %~mag~arcsec$^{-2}$\,mag\,arcsec$^{-2}$ and 25.15} %~mag~arcsec$^{-2}$\,mag\,arcsec$^{-2}$, respectively) inside a circular aperture with 5\,kpc\ radius exhibit smaller Hubble-residual dispersion.
}
\label{fig:FUV}
\end{figure*}
\begin{figure*
\centering
\subfigure{\includegraphics[angle=0,width=3.1in]{AVHISTAN_18_HR.pdf}}
\subfigure{\includegraphics[angle=0,width=3.1in]{AVHISTAN_31_HR.pdf}}
\caption{Hubble residuals against MLCS2k2 $A_V$ parameter for $R_V=1.8$ and $R_V=3.1$ model dust extinction laws for SNe with average NUV surface brightness $<$24.7} %~mag~arcsec$^{-2}$\,mag\,arcsec$^{-2}$ within $r=$ 5\,kpc\ aperture around SN. As shown in Table~\ref{tab:stddev}, distance moduli computed using MLCS2k2 with an $R_V=1.8$ dust extinction law exhibit a smaller dispersion among their Hubble residuals, than when computed using an $R_V=3.1$ model dust extinction law.
}
\label{fig:RV}
\end{figure*}
\begin{figure*}[t]
\centering
\subfigure{\includegraphics[angle=0,width=6.5in]{mosaic_below.pdf}}
\subfigure{\includegraphics[angle=0,width=6.5in]{mosaic_above.pdf}}
\caption{Representative SNe~Ia explosion sites with high and low average NUV surface brightness within $r=$ 5\,kpc\ aperture around SN. White circles represent the apertures used for photometry. The upper mosaic shows SDSS color composite images of the host galaxies where the aperture NUV surface brightnesses is brighter than 24.7} %~mag~arcsec$^{-2}$\,mag\,arcsec$^{-2}$, while the lower panels contain images of SNe with environments that have lower NUV surface brightness. Text next to each SN name gives the NUV surface brightness (in units of mag\,arcsec$^{-2}$), while the crosshairs are centered on the location of the explosion. }
\label{fig:mosaic}
\end{figure*}
\begin{figure*
\centering
\subfigure{\includegraphics[angle=0,width=4.5in]{NUV_siiv.pdf}}
\caption{NUV surface brightness against \ionpat{Si}{ii} photospheric velocity of SNe~Ia. Horizontal line shows the 11,800\,km\,s$^{-1}$ line used by Wang et al. (2009) to separate low- and high-velocity SNe~Ia \cite{xwang09}. The population erupting in apertures with high NUV surface brightness does not differ strongly from that found in lower surface brightness regions.
}
\label{fig:NUVsiiv}
\end{figure*}
\clearpage
\begin{table*}[htp!]
\centering
\scriptsize
\caption{Host-Galaxy Environment}\begin{tabular}{lcccccc}
\hline
Name & $z_{\rm helio}$ & {\it FUV} SB & {\it NUV} SB & SFR Density & HR & HR\\
& & mag arcsec$^{-2}$ & mag arcsec$^{-2}$ & dex & $R_V$ = 1.8 & $R_V$ = 3.1\\
\hline
SN 1997dg & 0.034 & 25.72$\pm$0.19 & 25.82$\pm$0.10 & -2.48$\pm$0.29 & 0.29$\pm$0.09 & 0.04$\pm$0.11\\
SN 1999cc & 0.031 & 25.33$\pm$0.17 & 24.48$\pm$0.08 & -2.14$\pm$0.33 & -0.12$\pm$0.12 & -0.10$\pm$0.12\\
SN 1999dg & 0.022 & 27.89 (model) & 26.10$\pm$0.11 & -2.77$\pm$0.25 & -0.07$\pm$0.09 & -0.03$\pm$0.09\\
SN 2000dn & 0.032 & 27.61 (model) & 26.95$\pm$0.25 & -2.84$\pm$0.29 & 0.19$\pm$0.06 & 0.22$\pm$0.06\\
SN 2000fa & 0.021 & 25.02$\pm$0.08 & 24.00$\pm$0.06 & -2.15$\pm$0.21 & -0.10$\pm$0.05 & -0.13$\pm$0.06\\
SN 2001br & 0.021 & 27.17 (model) & 26.51$\pm$0.13 & -2.80$\pm$0.23 & 0.22$\pm$0.06 & 0.21$\pm$0.07\\
SN 2001cj & 0.024 & 28.63$\pm$0.26 & 27.84$\pm$0.07 & -3.46$\pm$0.30 & 0.09$\pm$0.04 & 0.13$\pm$0.04\\
SN 2001ck & 0.035 & 24.11$\pm$0.13 & 23.71$\pm$0.08 & -1.72$\pm$0.26 & 0.01$\pm$0.06 & 0.05$\pm$0.06\\
SN 2001cp & 0.022 & 28.13 (model) & 28.03$\pm$0.28 & -3.45$\pm$0.35 & -0.01$\pm$0.05 & 0.00$\pm$0.06\\
SN 2001ie & 0.031 & 29.23 (model) & 28.23$\pm$0.30 & -3.30$\pm$0.25 & -0.14$\pm$0.13 & -0.20$\pm$0.15\\
SN 2002aw & 0.026 & 27.53 (model) & 26.89$\pm$0.10 & .\,.\,. & -0.16$\pm$0.07 & -0.24$\pm$0.09\\
SN 2002de & 0.028 & 24.18$\pm$0.06 & 23.67$\pm$0.06 & .\,.\,. & 0.07$\pm$0.04 & 0.04$\pm$0.05\\
SN 2002eb & 0.028 & 28.17 (model) & 26.50$\pm$0.07 & -3.35$\pm$0.28 & 0.01$\pm$0.03 & -0.00$\pm$0.04\\
SN 2002el & 0.030 & 28.81 (model) & 27.51$\pm$0.29 & -3.16$\pm$0.38 & -0.30$\pm$0.05 & -0.27$\pm$0.05\\
SN 2002hu & 0.037 & 26.90 (model) & 26.45$\pm$0.20 & -3.14$\pm$0.25 & -0.15$\pm$0.06 & -0.12$\pm$0.07\\
SN 2003U & 0.028 & 25.10$\pm$0.05 & 24.61$\pm$0.06 & -2.26$\pm$0.30 & -0.14$\pm$0.10 & -0.12$\pm$0.10\\
SN 2003W & 0.020 & 24.60$\pm$0.08 & 23.98$\pm$0.06 & -1.87$\pm$0.29 & -0.07$\pm$0.04 & -0.11$\pm$0.05\\
SN 2003ch & 0.025 & 29.32 (model) & 28.15 (model) & -3.32$\pm$0.20 & 0.18$\pm$0.09 & 0.21$\pm$0.09\\
SN 2003cq & 0.033 & 25.61$\pm$0.28 & 24.84$\pm$0.11 & -2.19$\pm$0.27 & -0.15$\pm$0.08 & -0.20$\pm$0.09\\
SN 2003gn & 0.035 & 26.52 (model) & 25.52$\pm$0.12 & -2.52$\pm$0.36 & 0.19$\pm$0.06 & 0.14$\pm$0.06\\
SN 2003kc & 0.033 & 24.15$\pm$0.06 & 23.74$\pm$0.06 & .\,.\,. & -0.06$\pm$0.13 & -0.11$\pm$0.20\\
SN 2004as & 0.031 & 25.09$\pm$0.16 & 24.65$\pm$0.09 & -2.25$\pm$0.21 & 0.11$\pm$0.05 & 0.05$\pm$0.05\\
SN 2004at & 0.022 & 26.48$\pm$0.05 & 25.89$\pm$0.06 & -2.61$\pm$0.35 & -0.08$\pm$0.05 & -0.04$\pm$0.05\\
SN 2004bg & 0.021 & 25.77$\pm$0.04 & 25.08$\pm$0.06 & -2.38$\pm$0.34 & -0.04$\pm$0.05 & -0.00$\pm$0.05\\
SN 2004br & 0.023 & 28.04 (model) & 26.64$\pm$0.15 & -2.90$\pm$0.24 & -0.27$\pm$0.04 & -0.23$\pm$0.05\\
SN 2004bw & 0.021 & 25.09$\pm$0.06 & 24.48$\pm$0.06 & -2.01$\pm$0.31 & -0.05$\pm$0.06 & -0.02$\pm$0.06\\
SN 2004ef & 0.031 & 26.40$\pm$0.24 & 25.41$\pm$0.10 & .\,.\,. & -0.12$\pm$0.05 & -0.16$\pm$0.05\\
SN 2004gu & 0.045 & 27.77 (model) & 27.04$\pm$0.10 & -3.21$\pm$0.14 & -0.07$\pm$0.05 & -0.15$\pm$0.06\\
SN 2005ag & 0.079 & 28.44 (model) & 26.81 (model) & -2.93$\pm$0.25 & -0.14$\pm$0.04 & -0.18$\pm$0.05\\
SN 2005bg & 0.023 & 23.96$\pm$0.06 & 23.41$\pm$0.06 & .\,.\,. & 0.02$\pm$0.05 & 0.04$\pm$0.06\\
SN 2005eq & 0.030 & 25.88$\pm$0.09 & 25.66$\pm$0.06 & -2.35$\pm$0.33 & -0.00$\pm$0.05 & -0.03$\pm$0.06\\
SN 2005eu & 0.035 & 25.40$\pm$0.08 & 25.32$\pm$0.06 & -2.37$\pm$0.24 & -0.04$\pm$0.06 & 0.00$\pm$0.06\\
SN 2005hc & 0.046 & 25.57$\pm$0.07 & 25.50$\pm$0.06 & -2.53$\pm$0.27 & 0.24$\pm$0.05 & 0.28$\pm$0.06\\
SN 2005iq & 0.034 & 25.93$\pm$0.28 & 25.85$\pm$0.14 & .\,.\,. & 0.22$\pm$0.05 & 0.25$\pm$0.05\\
SN 2005ku & 0.050 & 24.29$\pm$0.04 & 23.78$\pm$0.06 & -1.86$\pm$0.34 & 0.17$\pm$0.10 & 0.14$\pm$0.13\\
SN 2005ms & 0.025 & 29.04 (model) & 28.00$\pm$0.09 & -3.53$\pm$0.33 & 0.22$\pm$0.06 & 0.25$\pm$0.07\\
SN 2006S & 0.032 & 25.11$\pm$0.18 & 25.04$\pm$0.11 & -2.26$\pm$0.25 & 0.19$\pm$0.05 & 0.16$\pm$0.07\\
SN 2006ac & 0.023 & 24.40$\pm$0.07 & 23.92$\pm$0.06 & -1.81$\pm$0.28 & 0.04$\pm$0.10 & 0.08$\pm$0.11\\
SN 2006an & 0.064 & 27.13 (model) & 27.00 (model) & -3.43$\pm$0.26 & 0.07$\pm$0.07 & 0.11$\pm$0.07\\
SN 2006az & 0.031 & 26.58$\pm$0.30 & 25.26$\pm$0.09 & -2.54$\pm$0.20 & -0.22$\pm$0.05 & -0.18$\pm$0.05\\
SN 2006cf & 0.042 & 24.84$\pm$0.14 & 24.20$\pm$0.08 & -2.08$\pm$0.18 & 0.07$\pm$0.09 & 0.11$\pm$0.09\\
SN 2006cj & 0.067 & 25.85$\pm$0.13 & 25.62$\pm$0.07 & -2.69$\pm$0.12 & 0.14$\pm$0.07 & 0.18$\pm$0.07\\
SN 2006cp & 0.022 & 26.24$\pm$0.25 & 25.83$\pm$0.11 & -2.56$\pm$0.29 & 0.02$\pm$0.06 & -0.01$\pm$0.07\\
SN 2006cq & 0.048 & 26.41$\pm$0.15 & 25.69$\pm$0.07 & -2.85$\pm$0.17 & 0.18$\pm$0.09 & 0.17$\pm$0.11\\
SN 2006en & 0.032 & 24.43$\pm$0.10 & 23.47$\pm$0.06 & .\,.\,. & 0.02$\pm$0.07 & -0.06$\pm$0.09\\
SN 2006et & 0.022 & 27.30 (model) & 26.10$\pm$0.11 & .\,.\,. & 0.01$\pm$0.10 & -0.06$\pm$0.12\\
SN 2006lu & 0.053 & 24.33$\pm$0.22 & 24.16$\pm$0.12 & .\,.\,. & -0.09$\pm$0.07 & -0.09$\pm$0.08\\
SN 2006mp & 0.023 & 24.93$\pm$0.11 & 24.77$\pm$0.07 & .\,.\,. & 0.15$\pm$0.08 & 0.12$\pm$0.10\\
SN 2006oa & 0.060 & 25.94$\pm$0.11 & 25.73$\pm$0.07 & -2.73$\pm$0.23 & 0.13$\pm$0.07 & 0.14$\pm$0.07\\
SN 2006on & 0.070 & 28.30 (model) & 26.51$\pm$0.14 & -3.04$\pm$0.18 & -0.13$\pm$0.11 & -0.21$\pm$0.15\\
SN 2006py & 0.060 & 28.18 (model) & 26.84$\pm$0.12 & -3.14$\pm$0.16 & -0.11$\pm$0.10 & -0.22$\pm$0.11\\
SN 2006sr & 0.024 & 24.87$\pm$0.05 & 24.32$\pm$0.06 & -2.03$\pm$0.30 & 0.13$\pm$0.07 & 0.16$\pm$0.07\\
SN 2007F & 0.024 & 24.45$\pm$0.07 & 24.05$\pm$0.06 & -2.00$\pm$0.28 & 0.06$\pm$0.05 & 0.09$\pm$0.06\\
SN 2007R & 0.031 & 24.63$\pm$0.11 & 23.80$\pm$0.06 & -2.02$\pm$0.18 & 0.05$\pm$0.07 & 0.09$\pm$0.07\\
SN 2007ae & 0.064 & 25.85 (model) & 25.59$\pm$0.14 & .\,.\,. & -0.20$\pm$0.08 & -0.16$\pm$0.08\\
\hline
\end{tabular}
\end{table*}\begin{table*}[htp!]
\centering
\scriptsize\begin{tabular}{lcccccc}
\hline
Name & $z_{\rm helio}$ & {\it FUV} SB & {\it NUV} SB & SFR Density & HR & HR\\
& & mag arcsec$^{-2}$ & mag arcsec$^{-2}$ & dex & $R_V$ = 1.8 & $R_V$ = 3.1\\
\hline
SN 2007bd & 0.031 & 25.18$\pm$0.13 & 24.91$\pm$0.08 & -2.25$\pm$0.19 & -0.15$\pm$0.07 & -0.11$\pm$0.07\\
SN 2007bz & 0.022 & 23.43$\pm$0.03 & 23.17$\pm$0.05 & -1.66$\pm$0.27 & 0.59$\pm$0.07 & 0.52$\pm$0.10\\
SN 2007cp & 0.037 & 24.89$\pm$0.20 & 23.93$\pm$0.07 & -2.15$\pm$0.18 & 0.00$\pm$0.17 & 0.03$\pm$0.18\\
SN 2007cq & 0.026 & 24.89$\pm$0.08 & 24.59$\pm$0.06 & -2.13$\pm$0.29 & -0.11$\pm$0.06 & -0.08$\pm$0.06\\
SN 2007is & 0.030 & 24.81$\pm$0.14 & 24.28$\pm$0.08 & -2.01$\pm$0.22 & -0.06$\pm$0.10 & -0.04$\pm$0.13\\
SN 2007jg & 0.040 & 26.48 (model) & 26.02$\pm$0.25 & -2.90$\pm$0.28 & 0.19$\pm$0.08 & 0.22$\pm$0.09\\
SN 2007qe & 0.024 & 27.03 (model) & 27.03$\pm$0.14 & .\,.\,. & 0.05$\pm$0.06 & 0.02$\pm$0.07\\
SN 2008Q & 0.008 & 31.12 (model) & 29.93$\pm$0.19 & -3.89$\pm$0.08 & -4.75$\pm$0.10 & -4.74$\pm$0.08\\
SN 2008Z & 0.021 & 29.02 (model) & 28.06$\pm$0.19 & -3.26$\pm$0.24 & 0.32$\pm$0.09 & 0.26$\pm$0.11\\
SN 2008ar & 0.026 & 25.56$\pm$0.19 & 24.77$\pm$0.09 & -2.40$\pm$0.30 & 0.30$\pm$0.06 & 0.32$\pm$0.07\\
SN 2008bf & 0.024 & 29.64 (model) & 28.42$\pm$0.29 & -3.41$\pm$0.23 & -0.02$\pm$0.05 & 0.02$\pm$0.05\\
SN 2008bz & 0.060 & 27.22 (model) & 26.04$\pm$0.23 & -2.68$\pm$0.30 & 0.06$\pm$0.08 & 0.09$\pm$0.08\\
SN 2008cf & 0.046 & 25.00$\pm$0.28 & 25.00$\pm$0.15 & .\,.\,. & -0.08$\pm$0.07 & -0.05$\pm$0.07\\
SN 2008dr & 0.041 & 26.60 (model) & 25.69$\pm$0.07 & .\,.\,. & -0.00$\pm$0.06 & 0.01$\pm$0.07\\
SN 2008fr & 0.039 & 27.12$\pm$0.21 & 26.33$\pm$0.08 & -3.10$\pm$0.25 & -0.46$\pm$0.06 & -0.42$\pm$0.06\\
SN 2008gl & 0.034 & 30.01 (model) & 29.98 (model) & .\,.\,. & -0.03$\pm$0.08 & 0.00$\pm$0.08\\
SN 2008hj & 0.038 & 25.81$\pm$0.09 & 25.20$\pm$0.06 & -2.73$\pm$0.20 & 0.02$\pm$0.09 & 0.05$\pm$0.09\\
SN 2009D & 0.025 & 28.86 (model) & 28.29$\pm$0.27 & -3.79$\pm$0.31 & -0.12$\pm$0.06 & -0.08$\pm$0.07\\
SN 2009ad & 0.028 & 25.82$\pm$0.28 & 24.90$\pm$0.10 & -2.49$\pm$0.32 & -0.02$\pm$0.06 & 0.01$\pm$0.06\\
SN 2009al & 0.022 & 31.67 (model) & 29.74 (model) & -3.86$\pm$0.34 & 0.20$\pm$0.06 & 0.10$\pm$0.07\\
SN 2009do & 0.040 & 28.00 (model) & 26.07$\pm$0.07 & -2.93$\pm$0.24 & -0.10$\pm$0.07 & -0.09$\pm$0.08\\
SN 2009lf & 0.046 & 30.79 (model) & 28.66 (model) & -3.66$\pm$0.21 & -0.71$\pm$0.07 & -0.67$\pm$0.07\\
SN 2009na & 0.021 & 24.22$\pm$0.04 & 23.83$\pm$0.05 & -1.79$\pm$0.27 & -0.04$\pm$0.07 & -0.02$\pm$0.08\\
SN 2010ag & 0.034 & 25.86$\pm$0.24 & 25.50$\pm$0.11 & .\,.\,. & -0.18$\pm$0.08 & -0.23$\pm$0.10\\
SN 2010dt & 0.053 & 26.30$\pm$0.16 & 25.94$\pm$0.08 & -2.63$\pm$0.21 & 0.13$\pm$0.21 & 0.16$\pm$0.22\\
SNF20080514-002 & 0.022 & 28.75 (model) & 26.90$\pm$0.10 & -2.94$\pm$0.22 & 0.06$\pm$0.06 & 0.10$\pm$0.06\\
SNF20080522-011 & 0.038 & 25.29$\pm$0.06 & 25.18$\pm$0.06 & -2.55$\pm$0.23 & 0.01$\pm$0.06 & 0.04$\pm$0.06\\
SNF20080909-030 & 0.031 & 27.26 (model) & 26.48$\pm$0.08 & -2.57$\pm$0.26 & -0.11$\pm$0.12 & -0.15$\pm$0.13\\
\hline
\end{tabular}
\caption{The NUV and FUV surface brightnesses are measured within a circular $r$\ $=$\ 5\,kpc~aperture centered at the SN position. When the uncertainty of the measured flux exceeds 0.3\,mag, we use the magnitude synthesized from the {\tt kcorrect} \cite{bl07} model that best fits the full set of measured magnitudes, instead of the K-corrected measured magnitude. The ``HR'' columns show the Hubble residuals of each SN Ia from the best-fitting redshift-distance relation for $R_V=1.8$ and $R_V=3.1$ MLCS2k2 light-curve fits.}
\label{tab:data}
\end{table*}
\end{document}
For AB magnitudes, surface brightness dimming varies as $(1+z)^3$.
|
\section{Introduction}
Chiral effective field theory (\ensuremath{\chi\mathrm{EFT}}{}) for nuclear physics provides
a theoretical framework for a common description of various nuclear
processes through the systematic generation of two- three- and many-
body interactions and currents. Undoubtedly, it has allowed advances
in ab initio structure calculations of light~\cite{Maris2013} and
medium mass~\cite{Roth2012,Hergert2013,Hagen2012} atomic nuclei. The
quantitative predictions of \ensuremath{\chi\mathrm{EFT}}{} depend on the numerical values
of a set of low-energy constants (LECs), which have become a limiting
factor in medium mass nuclei where current sets of LECs fail to
simultaneously predict binding, spectra, and radii. It is therefore
relevant to constrain these LECs such that all predictions within the
realm of applicability of \ensuremath{\chi\mathrm{EFT}}{} can be quantified together with
statistical uncertainties. This is important for advancing modern
many-body calculations into regions of the nuclear chart and the
physical processes where experimental data for verification is
limited, such as neutrino-less double beta decay or
structure and reactions near the neutron drip line. We present
results that constrain the \ensuremath{\pi{}N}{}-sector of \ensuremath{\chi\mathrm{EFT}}{}, with
accompanying confidence intervals (CI), up to fourth order in the
expansion of the effective Lagrangian. At this order, the sub-leading
long range three-nucleon interaction enters, and will thus be fully
constrained by the results presented here. Further, many operator
currents, such as the axial-vector current, depend on only \ensuremath{\pi{}N}{} LECs
and three nucleon contact LECs. Therefore careful uncertainty
quantification of the \ensuremath{\pi{}N}{} LECs is also critical for comparing
processes driven by these currents to experimental cross sections and
decay measurements.
The \ensuremath{\chi\mathrm{EFT}}{} interaction Lagrangian $\mathcal{L}_{\rm eff}$ for
atomic nuclei can be separated into two terms $\mathcal{L}_{\rm
eff}=\mathcal{L}_{\ensuremath{NN}{}}+\mathcal{L}_{\ensuremath{\pi{}N}{}}$, and the two different
contributions each depend explicitly on a distinctive set of LECs. The
first term parametrizes the short-ranged contact-interactions and the
second term describes the long-ranged and pion-mediated part of the
two- and three-nucleon interaction. While the \ensuremath{NN}{}-contact sector
must be constrained using nucleon-nucleon data, the LECs in
$\mathcal{L}_{\ensuremath{\pi{}N}{}}$ can be determined from experimental \ensuremath{\pi{}N}{}
scattering-data, completely separately from the \ensuremath{NN}{} terms. This is
one example of how \ensuremath{\chi\mathrm{EFT}}{} can link separate physical processes that
are relevant for the description of atomic nuclei. Previous
constraints for the \ensuremath{\pi{}N}{} LECs have been determined from peripheral
\ensuremath{NN}{} scattering phase shifts~\cite{Entem2002,Ekstrom2013} or \ensuremath{\pi{}N}{}
elastic scattering phase
shifts~\cite{Buttiker2000,Fettes2000a,Krebs2012}. Indeed, these
efforts have produced various sets of LECs that closely reproduce the
respective phase shift analyses (either
\ensuremath{NN}{}~\cite{Stoks1993,Arndt2007} or
\ensuremath{\pi{}N}{}~\cite{Koch1985,Koch1986,Workman2012,SAID:WI08});
however, the lack of reliable uncertainties on the input phase shifts
prevents meaningful uncertainty quantification of the \ensuremath{\pi{}N}{} LECs. It
should be noted that available scattering phase shifts are not
experientially measured data, but the result of a partial-wave
analysis of measured data. In contrast with previous determinations of
the \ensuremath{\pi{}N}{} LECs, the analysis presented here is grounded in
experimental scattering data. This allows us to estimate meaningful
statistical uncertainties. For the first time we can therefore explore
the consistency of \ensuremath{\chi\mathrm{EFT}}{} by predicting CI's for the peripheral
\ensuremath{NN}{} phase shifts determined by \ensuremath{\pi{}N}{}-data.
\begin{table}
\caption{\label{tab:lecs}Numerical values of the \ensuremath{\pi{}N}{} LECs that
result from the optimization with respect to explrimental
observables. The resulting values are grouped from left to right in
the order they appear in the Lagrangian.}
\begin{ruledtabular}
\begin{tabular}{r D{,}{\,\pm\,}{-1} | r D{,}{\,\pm\,}{-1} | r D{,}{\,\pm\,}{-1}}
\multicolumn{2}{c|}{$\mathcal{O}(Q^1)$ LECs } & \multicolumn{2}{c|}{$\mathcal{O}(Q^2)$ LECs} & \multicolumn{2}{c}{$\mathcal{O}(Q^3)$ LECs}\\
\multicolumn{2}{c|}{[GeV$^{-1}$]} & \multicolumn{2}{c|}{[GeV$^{-2}$]} & \multicolumn{2}{c}{[GeV$^{-3}$]} \\
\hline
$c_{1}$ & -1.40,0.12 & $\bar{d}_1+\bar{d}_2$ & +5.80,0.14 &$\bar{e}_{14}$ & +1.53,0.31\\
$c_{2}$ & +1.71,0.33 & $\bar{d}_3$ & -5.66,0.08 &$\bar{e}_{15}$ & -11.91,0.87\\
$c_{3}$ & -4.56,0.11 & $\bar{d}_5$ & +0.03,0.06 &$\bar{e}_{16}$ & +11.43,1.23\\
$c_{4}$ & +3.72,0.27 & $\bar{d}_{14}-\bar{d}_{15}$ & -11.50,0.12 &$\bar{e}_{17}$ & +0.73,0.51\\
& & & &$\bar{e}_{18}$ & +0.57,1.36\\
\end{tabular}
\end{ruledtabular}
\end{table}
\section{Optimization}
We seek a set of \ensuremath{\pi{}N}{} LECs $\bm{c}_{\star}$ that minimize the least-squares objective
function
\footnote{See Refs.~\cite{Arndt2007} for details on their process for
a slightly older solution (SP06)}:
\begin{align}
\label{eqn:Chi2:Obs}
\ensuremath{\chi^2_{\mathrm{red}}}(\bm{c},\bm{N}) &= \frac{1}{n_{df}}\left(\sum_{i}R_{i}(\bm{c},\bm{N})^2+\sum_j r_i(\bm{N})^2\right)
\\
R_i(\bm{c},\bm{N}) &= \frac{N_{j_i}O_{i}^{\ensuremath{\chi\mathrm{EFT}}{}}(\bm{c})-O_{i}^{\mathrm{Exp.}}}{\Delta_{i}^{Exp.}}
\\
r_j(\bm{N}) &= \frac{N_j-1}{\Delta_j}
\end{align}
where $O_{i}^{\ensuremath{\chi\mathrm{EFT}}{}}(\bm{c})$ denotes the value of the scattering
observable computed from \ensuremath{\chi\mathrm{EFT}}{}, while $O_{ji}^{\mathrm{Exp.}}$ and
$\Delta_{ji}^{Exp.}$ denotes the experimentally measured value and
uncertainty, respectively, for the corresponding observable. $\bm{c}$ is
a vector of LECs spanning all included $c_i$, $d_i$, and $e_i$ LECs.
$\bm{N}$ is a vector of normalization coefficients $N_j$, where all
points from a single experimental angular distribution share the same
$N_j$; $\Delta_j$ encodes uncertainty of the experimental systematics
for $N_j$. The number of degrees of freedom is given by $n_{\rm df} =
(n_d+n_{N}-n_{NF})-(n_{\rm lecs}+n_{N})$, where $n_d$ is the number of
data included in the fit, $n_{\rm lecs}$ is the number of LECs being
fit, and $n_N$ is the number of unknown normalization coefficients and
$n_{NF}$ is the number of floated coefficients (contribute no residual
term in $r_j$ as $\delta_j=\infty$). For a given experiment $j$, the
included data points run over a series of scattering angles at a fixed
lab frame momentum $Q_{\rm Lab}$.
For our fitting dataset, we adopt the database from the most recent
\ensuremath{\pi{}N}{} partial wave analysis~\cite{Workman2012}, referred to as
WI08{}. By construction, \ensuremath{\chi\mathrm{EFT}}{} is a low-energy theory, therefore
we exclude data with lab-frame momentum $Q_{\mathrm{Lab}}>160\, \mbox{MeV}$.
This leaves us with an experimental database consisting of
differential scattering cross-sections and polarization cross sections
from $\pi^{\pm}+p\rightarrow\pi^{\pm}+p$ and
$\pi^{-}+p\rightarrow\pi^{0}+n$ processes. In total, there are
$n_d=1246$ data points, consisting of $1194$ differential unpolarized
cross-sections and $52$ differential singly-polarized cross-sections.
There are $n_N=110$ normalization coefficients, with $n_{NF}=9$
floated coefficients. There are other measurable observables,
such as the spin rotation parameters, but experimental data only
exists for momentum well beyond the range of validity of the EFT. The
cutoff in lab frame momentum ($160\, \mbox{MeV}$) was chosen such that
increasing or decreasing the cutoff would lead to a larger minimum
value of $\ensuremath{\chi^2_{\mathrm{red}}}(\bm{c},\bm{N})$, maximizing amount of included data
while avoiding fitting past the radius of convergence of the EFT.
For the calculated observables, we use the strong amplitudes presented
in Ref.~\cite{Krebs2012}
(Refs.~\cite{Fettes1998,Fettes2000,Fettes2000a,Fettes2001} give a more
complete presentation of the \ensuremath{\pi{}N}{} scattering amplitudes, but use a
different power counting scheme for relativistic corrections.) For
the strong amplitude, we adopt their conventions for fixing
$\bar{d}_{18}$, absorbing $\bar{e}_{19,20,21,22,35,36,37,38}$,
$\bar{l}_{3}$ into $c_{1,2,3,4}$. We also work with exact isospin
symmetry and use an averaged pion mass ($m_{\pi^{\vphantom{\pm}}} =
(m_{\pi^0}+2m_{\pi^\pm})/3$). We adopt the electromagnetic treatment
that is used in the WI08{} partial wave analysis, which is described in
detail inRefs.~\cite{Tromborg1978,Tromborg1974,Tromborg1977,Bugg1973}. For the
electromagnetic corrections, we explicitly break isospin symmetry and
use the physical pion masses within the coulomb amplitudes. Actual
fits were performed using the TAO package~\cite{TAO}.
\section{Results}
The central values and $1\sigma$ uncertainties of the \ensuremath{\pi{}N}{} LECs up
to fourth order in \ensuremath{\chi\mathrm{EFT}}{} from fitting against scattering data are
presented in Table~\ref{tab:lecs}. For this fit, we find that
$\ensuremath{\chi^2_{\mathrm{red}}}=2.29$. As a comparison, the LECs from the fit against WI08{}
phase shifts of Ref.~\cite{Krebs2012} generate $\ensuremath{\chi^2_{\mathrm{red}}}=3.63$ with
respect to our objective function. Not surprisingly, our fit will
reproduce experimental data better than the fits with respect to phase
shifts. While our LECc are consistent with the spread of previous
analyses, we find that no single analysis is entirely consistent with
our fit at the $95\%$ confidence level, though the WI08 \ensuremath{\pi{}N}{} phase
shift fit from Ref.~\cite{Krebs2012} lies just outside our interval.
\begin{table*}
\caption{\label{tab:covcor}Covariance (lower triangle) and
correlation (upper triangle) matrices for our fit to experimental
data}
\begin{ruledtabular}
\begin{tabular}{c| D{.}{.}{-1} D{.}{.}{-1} D{.}{.}{-1} D{.}{.}{-1} D{.}{.}{-1} D{.}{.}{-1} D{.}{.}{-1} D{.}{.}{-1} D{.}{.}{-1} D{.}{.}{-1} D{.}{.}{-1} D{.}{.}{-1} D{.}{.}{-1} }
LEC & \multicolumn{1}{c}{$c_{1}$} & \multicolumn{1}{c}{$c_{2}$} & \multicolumn{1}{c}{$c_{3}$} & \multicolumn{1}{c}{$c_{4}$} & \multicolumn{1}{c}{$\bar{d}_1+\bar{d}_2$} & \multicolumn{1}{c}{$\bar{d}_3$} & \multicolumn{1}{c}{$\bar{d}_5$} & \multicolumn{1}{c}{$\bar{d}_{14}-\bar{d}_{15}$} & \multicolumn{1}{c}{$\bar{e}_{14}$} & \multicolumn{1}{c}{$\bar{e}_{15}$} & \multicolumn{1}{c}{$\bar{e}_{16}$} & \multicolumn{1}{c}{$\bar{e}_{17}$} & \multicolumn{1}{c}{$\bar{e}_{18}$}\\
\cline{1-14}
$c_{1}$ & +0.03 & \multicolumn{1}{|D{.}{.}{-1}}{+0.95} & +0.24 & +0.56 & +0.51 & -0.33 & -0.48 & -0.45 & -0.31 & +0.37 & -0.85 & +0.15 & -0.54\\
\cline{3-3}
$c_{2}$ & +0.07 & +0.21 & \multicolumn{1}{|D{.}{.}{-1}}{-0.08} & +0.55 & +0.54 & -0.37 & -0.48 & -0.56 & -0.32 & +0.56 & -0.96 & +0.18 & -0.54\\
\cline{4-4}
$c_{3}$ & +0.01 & -0.01 & +0.02 & \multicolumn{1}{|D{.}{.}{-1}}{+0.13} & -0.02 & +0.12 & -0.06 & +0.26 & +0.08 & -0.58 & +0.26 & -0.12 & -0.08\\
\cline{5-5}
$c_{4}$ & +0.03 & +0.09 & +0.01 & +0.14 & \multicolumn{1}{|D{.}{.}{-1}}{+0.97} & -0.75 & -0.77 & -0.74 & -0.39 & +0.36 & -0.55 & +0.18 & -0.94\\
\cline{6-6}
$\bar{d}_1+\bar{d}_2$ & +0.01 & +0.05 & -0.00 & +0.07 & +0.03 & \multicolumn{1}{|D{.}{.}{-1}}{-0.81} & -0.76 & -0.75 & -0.43 & +0.46 & -0.57 & +0.18 & -0.91\\
\cline{7-7}
$\bar{d}_3$ & -0.01 & -0.02 & +0.00 & -0.03 & -0.02 & +0.01 & \multicolumn{1}{|D{.}{.}{-1}}{+0.24} & +0.67 & +0.43 & -0.45 & +0.44 & -0.24 & +0.74\\
\cline{8-8}
$\bar{d}_5$ & -0.01 & -0.02 & -0.00 & -0.02 & -0.01 & +0.00 & +0.00 & \multicolumn{1}{|D{.}{.}{-1}}{+0.52} & +0.25 & -0.28 & +0.48 & -0.04 & +0.69\\
\cline{9-9}
$\bar{d}_{14}-\bar{d}_{15}$ & -0.01 & -0.04 & +0.01 & -0.04 & -0.02 & +0.01 & +0.01 & +0.02 & \multicolumn{1}{|D{.}{.}{-1}}{+0.34} & -0.52 & +0.62 & -0.47 & +0.81\\
\cline{10-10}
$\bar{e}_{14}$ & -0.02 & -0.05 & +0.00 & -0.05 & -0.03 & +0.02 & +0.01 & +0.02 & +0.13 & \multicolumn{1}{|D{.}{.}{-1}}{-0.80} & +0.50 & -0.28 & +0.41\\
\cline{11-11}
$\bar{e}_{15}$ & +0.07 & +0.29 & -0.10 & +0.15 & +0.10 & -0.05 & -0.02 & -0.09 & -0.33 & +1.30 & \multicolumn{1}{|D{.}{.}{-1}}{-0.76} & +0.32 & -0.40\\
\cline{12-12}
$\bar{e}_{16}$ & -0.22 & -0.71 & +0.06 & -0.33 & -0.17 & +0.07 & +0.05 & +0.15 & +0.30 & -1.40 & +2.59 & \multicolumn{1}{|D{.}{.}{-1}}{-0.25} & +0.56\\
\cline{13-13}
$\bar{e}_{17}$ & +0.01 & +0.05 & -0.01 & +0.04 & +0.02 & -0.02 & -0.00 & -0.04 & -0.06 & +0.22 & -0.24 & +0.36 & \multicolumn{1}{|D{.}{.}{-1}}{-0.50}\\
\cline{14-14}
$\bar{e}_{18}$ & -0.14 & -0.41 & -0.02 & -0.58 & -0.27 & +0.13 & +0.08 & +0.21 & +0.24 & -0.75 & +1.48 & -0.49 & +2.70
\end{tabular}
\end{ruledtabular}
\end{table*}
For uncertainty analysis, we apply a standard gradient expansion of
$\chi^2(\bm{c})$ at the optimum $\bm{c}_{\star}$(see
Ref.~\cite{Dobaczewski2014} and references therein for further detail):
\begin{equation}
\chi^2(\bm{c}) = \chi^2(\bm{c}_{\star}) +
\frac{1}{2}\sum_{a,b} \left(\bm{c}-\bm{c}_{\star}\right)_a \bm{H}_{a,b}
\left(\bm{c}-\bm{c}_{\star}\right)_b +... \,,
\end{equation}
\begin{equation}
\bm{H}_{a,b} = \frac{\partial^2\chi^2}{\partial c_a \partial c_a} \bigg|_{\bm{c}=\bm{c}_{\star}}\,,
\end{equation}
where $\chi^2(\bm{c})$ is the full (not reduced by $n_{df}$) objective
function. If the residual vectors ($R_i$ and $r_j$) are normally
distributed, then the covariance matrix of our fit is given by
\begin{align}
\label{eqn:cov}
\bm{C}_{a,b} = \operatorname{cov}(c_ac_b) &\approx \ensuremath{\chi^2_{\mathrm{red}}} \operatorname{inv}(\bm{H})_{a,b}\,,
\\
\operatorname{corr}(c_ac_b) &= \bm{C}_{a,b}/\sqrt{\bm{C}_{a,a}\bm{C}_{b,b}}\,.
\end{align}.
The covariance matrix ($\bm{C}_{a,b}$) and correlation matrix
$\operatorname{corr}(c_ac_b)$ are presented in
table~\ref{tab:covcor}.
\begin{figure*}
\centering
\includegraphics{Obs_DSG.pdf}
\includegraphics{Obs_P.pdf}
\\
\includegraphics{Obs_A.pdf}
\includegraphics{Obs_R.pdf}
\caption{\ensuremath{\pi{}N}{} Observables computed at $\theta_{\rm
c.m.}=45^\circ$. The blue (green) lines with diamond (square)
markers show results for the
$\pi{}^{+(-)}+p\rightarrow\pi{}^{+(-)}+p$ processes while the red
lines with circular markers show the charge exchange process. The
dashed lines show observables computed using our fit to experimental
data. The dotted lines shows results reconstructed from LECs of
Ref.~\cite{Krebs2012}. P is the polarization, while A and R are the
spin rotation parameters, all presented as ratios with respect to
the differential cross-section. Definitions of these observables
can be found in Ref.~\cite{Hohler1983}
\label{fig:PiN:Obs}}
\end{figure*}
Figure~\ref{fig:PiN:Obs} shows calculations of \ensuremath{\pi{}N}{} scattering
observables using our LECs fit data as well as LECs from the phase
shift fit of Ref.~\cite{Krebs2012}. At smaller $Q_{\rm lab}$, the
difference is minimal, but it is clear at higher momentum that phase
shifts fits are inadequate. This is especially apparent in the
calculations of the polarization (P) and spin rotation parameters (A
and B), suggesting a phase shift fit may not adequately capture the
underlying tensor effects that are critical to nuclear observables.
\begin{figure*}
\centering
\includegraphics{PS_all}
\caption{\ensuremath{\pi{}N}{} phase shifts. The blue line is a prediction of
our fit with the band showing the $95\%$ interval. For
comparison, the red dotted lines are the phase shifts from
\cite{Krebs2012}. The markers show phase shifts from two
partial wave analyses~\cite{Koch1985,Workman2012}. The
partial waves are denoted as [L][2I][2J] where L is the
orbital angular momentum, I is the isospin, and J is the
total angular momentum.
\label{fig:PiN:PhaseShifts}}
\end{figure*}
As a validation of our analysis, we plot \ensuremath{\pi{}N}{} partial wave phase
shifts in Fig.~\ref{fig:PiN:PhaseShifts} and compare with two
different partial wave analyses(WI08{}~\cite{Workman2012} and
KA84~\cite{Koch1985}) The blue bands denotes the $95\%$ CI from our
fit. For S- and P-waves, we have good agreement with the WI08{}
partial wave analysis. In the D-waves, where the $\mathcal{O}(Q^3)$
LECs have significant contributions, we find poor agreement in
the $J=\tfrac{3}{2}$ channels (D13 and D33).
To better understand where this error in the $J=\tfrac{3}{2}$ D-waves
is coming from, we can use the eigenstate of the covariance matrix to
gain insight on the quality of the LEC constraints:
\begin{align}
\bm{C} \bm{v}_j &= \nu_j \bm{v}_j.
\end{align}
The vectors of LEC combinations, $\bm{v}_j$, form pseudo LECs that are
completely uncorrelated, and eigenvalues $\nu_j$ are effective
variances of these pseudo LECs. In table~\ref{tab:pca}, we examine
the LEC content of the least constrained eigenvectors (those with
largest eigenvalue). It is clear that the largest eigenvectors are
dominated by the $\bar{e}_i$ LECs, and in particular the pair
$\bar{e}_{16}$, $\bar{e}_{18}$ which dominate the two most
significant vectors. This indicates that $\bar{e}_{16}$ and
$\bar{e}_{18}$ are poorly constrained. Within the scattering
amplitudes, $\bar{e}_{16}$ and $\bar{e}_{18}$ span the non-spin-flip
and spin-flip amplitudes respectively~\cite{Krebs2012}, suggesting
that a lack of low momentum data for spin observables could be at
fault.
\begin{table}
\caption{\label{tab:pca} Eigenvalues ($\nu_j$) of
the covariance matrix, with contributions to eigenvector ($v_j$)
summed over LECs of same chiral order. These eigenvalue
correspond to the variances $\sigma^2_j$ of effective
uncorrelated pseudo LECs. For this analysis the covariance
matrix was reduced to a dimensionless form using appropriate
powers of the nucleon mass.}
\begin{ruledtabular}
\begin{tabular}{rrrrrrr}
\multicolumn{1}{c}{j} &
\multicolumn{1}{c}{$\nu_j$} &
\multicolumn{1}{c}{$\sum_{c_i}v^2_{j,c_i}$} &
\multicolumn{1}{c}{$\sum_{\bar{d}_i}v^2_{j,\bar{d}_i}$} &
\multicolumn{1}{c}{$\sum_{\bar{e}_i}v^2_{j,\bar{e}_i}$} &
\multicolumn{1}{c}{$v^2_{j,\bar{e}_{16}}$} &
\multicolumn{1}{c}{$v^2_{j,\bar{e}_{18}}$} \\
\hline
1 & 4.1147 & 0.0525 & 0.0094 & 0.9382 & 0.3829 & 0.4207 \\
2 & 1.4185 & 0.0362 & 0.0052 & 0.9587 & 0.2558 & 0.4798 \\
3 & 0.5114 & 0.1067 & 0.0013 & 0.8920 & 0.2038 & 0.0038 \\
4 & 0.2412 & 0.0748 & 0.0236 & 0.9016 & 0.0241 & 0.0157 \\
5 & 0.0516 & 0.2721 & 0.0459 & 0.6819 & 0.0004 & 0.0021 \\
6 & 0.0060 & 0.0076 & 0.9696 & 0.0227 & 0.0001 & 0.0044 \\
7 & 0.0047 & 0.0134 & 0.9647 & 0.0220 & 0.0006 & 0.0008 \\
8 & 0.0015 & 0.2194 & 0.6919 & 0.0887 & 0.0147 & 0.0368 \\
9 & 0.0007 & 0.6323 & 0.2427 & 0.1250 & 0.0790 & 0.0002 \\
10 & 0.0003 & 0.8666 & 0.0346 & 0.0987 & 0.0235 & 0.0348 \\
11 & 0.0001 & 0.7617 & 0.0125 & 0.2258 & 0.0007 & 0.0007 \\
12 & 0.0000 & 0.0080 & 0.9918 & 0.0002 & 0.0000 & 0.0000 \\
13 & 0.0000 & 0.9488 & 0.0068 & 0.0444 & 0.0144 & 0.0001 \\
\end{tabular}
\end{ruledtabular}
\end{table}
\section{Application to Uncertainty of \ensuremath{NN}{} Phase Shifts}
Accurate calculations accompanied with quantitative uncertainty
estimates is of-course one of the principal goals of low-energy
nuclear theory. This is critical to the study of nuclei near the
neutron drip-line or for $0\nu\beta\beta$ decay where this
insufficient data to validate many-body calculations.
As a first step towards quantifying the accuracy of the \ensuremath{\pi{}N}{} sector
of the \ensuremath{NN}{} interaction, we will predict the confidence intervals of
the peripheral ($J\ge4$) \ensuremath{NN}{} scattering phase shifts at \NxLO[3]{}. These
partial waves are fully determined by the long-ranged pion physics. It
should also be noted that this is the first \NxLO[3]{} calculation of these
phase shifts that is actually grounded in experimental scattering
data.
Figure~\ref{fig:UQnn} shows the predicted $95\%$ CIs of the proton-
neutron elastic scattering phase shifts for all total angular momentum
$J=4,5$ partial waves. For comparison, phase shifts from two
different partial wave analyses are presented, PWA93~\cite{Stoks1993}
and SP07~\cite{Arndt2007}. We also compare to phase shifts computed
using the high-precision Idaho-\NxLO[3]{} interaction~\cite{Entem2002},
which reproduces the SM99~\cite{Machleidt2001} \ensuremath{NN}{} database with
$\ensuremath{\chi^2_{\mathrm{red}}}\sim{}1$, and to
\NxLO[3]{} \ensuremath{NN}{} interactions computed using the LECs from
Ref.~\cite{Krebs2012}. Our LECs perform quite favorably
compared the \ensuremath{\pi{}N}{} phase shift fit, and surprisingly favorable in
many channels to the phase shifts computed from Idaho-\NxLO[3]{}. For many
channels, our error bands are remarkably small.
\begin{figure*}
\includegraphics{UQ_J4_PiN}
\includegraphics{UQ_J5_PiN}
\caption{Elastic proton-neutron $J=4,5$ phase shifts at \NxLO[3]{}
($\Lambda=500$ MeV) with $95\%$ CIs from \ensuremath{\pi{}N}{} data (red band).
(Dotted line) the phase shifts from the \NxLO[3]{} interactions
of Ref.~\cite{Entem2003}. (Dashed line) the phase shifts from the
Nijmegen partial wave analysis of Ref.~\cite{Stoks1993}.}
\label{fig:UQnn}
\end{figure*}
\section{Conclusions}
We constrained the \ensuremath{\pi{}N}{} LECs using experimental data with
$\chi^2/\textrm{datum} = 2.29$, generating a set of LECs with error
estimates that are compatible for use with modern \NxLO{3}
Hamiltonians and currents. We find that our lower order LECs are of
natural size, while the higher order LECs tend to be unnaturally large (or
small in the case of $\bar{d}_5$.) We find that even with fairly
large error bars for some LECs, the proton-neutron peripheral phase shifts
are very well constrained at lab scattering energies below $100\, \mbox{MeV}$.
The \ensuremath{\pi{}N}{} LECs not only fix the long range part of the \NxLO[3]{}
three body force, but our analysis provides a means to examine
uncertainty of the three body force in nuclear bound states as well as
uncertainty from currents in observables. Progressing forward,
simultaneous constraints of \ensuremath{\pi{}N}{} and \ensuremath{NN}{} LECs with quantitative
statical analysis will yield predictions of few- and many-body systems
with quantified uncertainty and is a topic for many future
investigations.
\section{Acknowledgments}
The authors would like to thank H.~Krebs, T.~Papenbrock, D.~Phillips,
and R.~Workman for their useful discussion. This work was supported
in part by the U.S. Department of Energy (DOE) under Grant
Nos.~DEFG02-96ER40963 (University of Tennessee), DE-SC0008499 (NUCLEI
SciDAC collaboration), Oak Ridge National Laboratory the Research
Council of Norway under contract ISP-Fysikk/216699, and by the
European Research Council under the European Community's Seventh
Framework Programme (FP7/2007-2013) ERC grant agreement no. 240603.
Oak Ridge National Laboratory is supported by the DOE Office of
Science under Contract No.~DE- AC05-00OR22725.
|
\section{Introduction}
We fix a prime number $p$.
Let $F$ be a field, which will always be assumed to contain a root of unity of order $p$.
Let $G_F$ be the absolute Galois group of $F$.
Recent works by Hopkins, Wickelgren, Min\'a\v c, and T\^an revealed a remarkable new property of the mod-$p$ Galois cohomology groups $H^i(G_F,\dbZ/p)$, $i=1,2$, related to triple Massey products.
This property, which they proved in several important cases, puts new restrictions on the possible group-theoretic structure of maximal pro-$p$ Galois groups of fields, and in particular, of absolute Galois groups.
In particular, in \cite{MinacTan13} Min\'a\v c and T\^an use this method to give new examples of pro-$2$ groups which cannot occur as absolute Galois groups of fields.
More specifically, for an arbitrary profinite group $G$ which acts trivially on $\dbZ/p$,
let $H^i(G)=H^i(G,\dbZ/p)$.
We recall that the triple Massey product is a multi-valued map $\langle\cdot,\cdot,\cdot\rangle\colon H^1(G)^3\to H^2(G)$ (see \S\ref{section on Massey products} for its precise definition).
We consider the following cohomological condition:
{\it If $\chi_1,\chi_2,\chi_3\in H^1(G)$ and $\langle\chi_1,\chi_2,\chi_3\rangle\subseteq H^2(G)$ is nonempty, then it contains zero.}
Following Min\'a\v c and T\^an, we call this condition the {\sl vanishing triple Massey product property} for $G$.
When $G=G_F$ for a field $F$ as above, this property can be rephrased in a more basic Galois-theoretic language,
in terms of the groups $\dbU_n(\dbF_p)$ of unipotent upper-triangular $n\times n$ matrices over $\dbF_p$.
Namely, $\chi_1,\chi_2,\chi_3$ are the Kummer characters corresponding the elements $a_1,a_2,a_3$ of $F^\times$, which we assume for simplicity to be $\dbF_p$-linearly independent in $F^\times/(F^\times)^p$.
Then (see Corollary \ref{Massey products and Un}):
\begin{itemize}
\item
$\langle\chi_1,\chi_2,\chi_3\rangle\neq\emptyset$ if and only if each of $F(a_1^{1/p},a_2^{1/p})$ and $F(a_2^{1/p},a_3^{1/p})$ embeds inside a Galois extension of $F$ with Galois group $\dbU_3(\dbF_p)$;
\item
$0\in\langle\chi_1,\chi_2,\chi_3\rangle$ if and only if $F(a_1^{1/p},a_2^{1/p},a_3^{1/p})$ embeds inside a Galois extension of $F$ with Galois group $\dbU_4(\dbF_p)$.
\end{itemize}
The vanishing triple Massey product condition was shown to hold for $G=G_F$ in the following situations:
\begin{enumerate}
\item[(i)]
$p=2$ and $F$ is a global field \cite{HopkinsWickelgren15};
\item[(ii)]
$p=2$ and $F$ is arbitrary \cite{MinacTan13};
\item[(iii)]
$p$ is arbitrary and $F$ is a global field \cite{MinacTan14}.
\end{enumerate}
Moreover, Min\'a\v c and T\^an conjecture that $G_F$ satisfies the vanishing triple Massey product property for every field $F$
containing a root of unity of order $p$ \cite{MinacTan15}.
If true, this would give new kinds of examples of pro-$p$ groups which are not realizable as absolute Galois groups for arbitrary primes $p$,
along the lines of \cite{MinacTan13} (where this is done for $p=2$).
In view of the interpretations of triple Massey products in terms of $\dbU_3(\dbF_p)$- and $\dbU_4(\dbF_p)$-Galois extension,
one also obtains an ``automatic realization" principle in the above cases, and conjecturally always.
In this note we relate these recent developments to classical results in the theory of central simple algebras and Brauer groups.
We investigate another cohomological property of $G$, which we call the \textsl{cup product--restriction property} for characters $\chi_1\nek\chi_r\in H^1(G)$.
This property for $r=2$ implies the vanishing triple Massey product property, but unlike the latter property, it does not involve external cohomological operations.
Now when $G=G_F$, we prove this cup product--restriction property for global fields and arbitrary $r$, using the Albert--Brauer--Hasse--Noether theorem and an injectivity theorem for $H^1$ due to Artin and Tate.
In the case where $p=2$ and $r=2$ the cup-product--restriction property for $G_F$ was proved by Tignol \cite{Tignol81a}*{Cor.\ 2.8}, and is an easy consequence of a refinement, also due to Tignol \cite{Tignol79}*{Th.\ 1}, of a result of Albert on the decomposition of central simple algebras as a tensor product of two quaternion algebras.
Specifically, let $G$ be an arbitrary profinite group, let $\chi_1\nek\chi_r\in H^1(G)=\Hom(G,\dbZ/p)$, and set $K=\bigcap_{i=1}^r\Ker(\chi_i)$.
We define a multi-linear map $\Lam_{(\chi_i)}\colon H^1(G)^r\to H^2(G)$ by $\Lam_{(\chi_i)}(\varphi_1\nek\varphi_r)=\sum_{i=1}^r\chi_i\cup\varphi_i$.
It lifts to a homomorphism $\Lam_{(\chi_i)}\colon H^1(G)^{\tensor r}\to H^2(G)$.
We say that the \textsl{cup product--restriction property} holds for $\chi_1\nek\chi_r$ if the sequence
\begin{equation}
\label{cup product-restriction}
H^1(G)^{\tensor r}\xrightarrow{\Lam_{(\chi_i)}}H^2(G)\xrightarrow{\res_K}H^2(K)
\end{equation}
is exact.
Note that (\ref{cup product-restriction}) is always a complex.
Further, its exactness depends only on $K$ (but not on the choice of $\chi_1\nek\chi_r$; see Proposition \ref{dependence on K}).
Now when $\chi_1,\chi_2,\chi_3\in H^1(G)$, the cup product--restriction property for $\chi_1,\chi_3\in H^1(G)$ implies the vanishing triple Massey product
property for $\langle\chi_1,\chi_2,\chi_3\rangle$ (Proposition \ref{cohomological lemma}).
\begin{main thm}
\label{thm on relative Brauer group}
Let $a_1\nek a_r\in F^\times$.
Suppose that one of the following conditions holds:
\begin{enumerate}
\item[(1)]
$F$ is a global field;
\item[(2)]
$p=2$ and $r=2$.
\end{enumerate}
Then the cup product--restriction property holds for the Kummer elements $(a_1)\nek (a_r)$ in $H^1(F)$.
\end{main thm}
In particular, by restricting to $r=2$, this gives alternative algebra-theoretic proofs of the results of Hopkins, Wickelgren, Min\'a\v c, and T\^an on the vanishing of triple Massey products, and relates these works to the above-mentioned classical results on Brauer groups.
As remarked above, the case (2) was earlier proved in \cite{Tignol81a}*{Cor.\ 2.8}, and is brought here in order to relate the respective results on Massey products to their algebra-theoretic counterparts (see \S5).
The cup-product--restriction property for $G=G_F$ is also closely related to the subgroup $\Dec(L/F)$ of ${}_p\Br(F)$, introduced in \cite{Tignol81a}, \cite{Tignol81b} where $L$ is the fixed field of $K$ in the separable closure of $F$.
In the case $p=2$ the cup-product--restriction property for absolute Galois groups of fields is essentially the property $P_2$ studied in \cite{Tignol81a} and \cite{ElmanLamTignolWadsworth83}.
In particular, in this situation, case (1) of the Main Theorem was earlier proved in \cite{ElmanLamTignolWadsworth83}*{Cor.\ 3.18}.
On the other hand, constructions of Tignol \cite{Tignol87} and McKinnie \cite{McKinnie11} show that, for $p$ odd, there exist fields $F$ for which the cup product--restriction property with $r=2$ does not hold (Example \ref{McKinnies example}).
This shows that some assumptions on $F$, as in the Main Theorem, are needed.
We thank the referees of this paper for very valuable comments and in particular for drawing our attention to several important references, which we included in the reference list.
\section{Preliminaries in Galois cohomology}
Let $F$ be again a field containing a root of unity of order $p$.
We abbreviate $H^i(F)=H^i(G_F,\dbZ/p)$.
We fix an isomorphism between the group $\mu_p$ of the $p$-th roots of unity and $\dbZ/p$.
This isomorphism induces the Kummer isomorphism $H^1(F)\isom F^\times/(F^\times)^p$.
Given $a\in F^\times$ let $(a)\in H^1(F)$ be the corresponding Kummer element.
Let $\Br(F)$ be the Brauer group of $F$ and let ${}_p\Br(F)$ be its subgroup consisting of all elements with exponent dividing $p$.
Given a field extension $L/F$ we write $\Br(L/F)$ for the kernel of the restriction map $\Br(F)\to\Br(L)$.
The isomorphism $\mu_p\isom\dbZ/p$ also induces in a standard way an isomorphism $H^2(F)\isom{}_p\Br(F)$.
For $a,b\in F^\times$, let $(a,b)_F$ be the corresponding symbol $F$-algebra of degree $p$.
The cup product $(a)\cup(b)$ in $H^2(F)$ then corresponds to the similarity class $[(a,b)_F]$ in ${}_p\Br(F)$.
\section{The cup product--restriction property}
\label{comments on cup res property}
We first show that the cup product--restriction property depends only on the subgroup $K=\bigcap_{i=1}^r\Ker(\chi_i)$, but not on the choice of $\chi_1\nek\chi_r$.
\begin{prop}
\label{dependence on K}
Let $G$ be a profinite group and let $K$ be an open subgroup of $G$.
Consider $\chi_1\nek \chi_r,\mu_1\nek\mu_s\in H^1(G)$ such that $K=\bigcap_{i=1}^r\Ker(\chi_i)=\bigcap_{j=1}^s\Ker(\mu_j)$.
Then the cup product--restriction property holds for $\chi_1\nek \chi_r$ if and only if it holds for $\mu_1\nek\mu_s$.
\end{prop}
\begin{proof}
There is a perfect pairing
\[
G/G^p[G,G]\times H^1(G)\to\dbZ/p, \quad (\bar g,\varphi)\mapsto \varphi(g)
\]
\cite{EfratMinac11}*{Cor.\ 2.2}.
It induces a perfect pairing $G/K\times\langle\chi_1\nek\chi_r\rangle\to\dbZ/p$, and similarly for the $\mu_j$.
Therefore $\langle\chi_1\nek \chi_r\rangle=\langle\mu_1\nek\mu_s\rangle$.
It follows that
\[
\chi_1\cup H^1(G)+\cdots+\chi_r\cup H^1(G)=\mu_1\cup H^1(G)+\cdots+\mu_s\cup H^1(G),
\]
i.e., the homomorphisms
\[
\Lam_{(\chi_i)}\colon H^1(G)^{\tensor r}\to H^2(G), \quad \Lam_{(\mu_j)}\colon H^1(G)^{\tensor s}\to H^2(G)
\]
have the same image, and the assertion follows.
\end{proof}
Consequently, for every open subgroup $K$ of $G$ such that $G/K$ is an elementary abelian $p$-group,
we may define the \textsl{cup product--restriction property for $K$} to be the cup product--restriction property
for $\chi_1\nek\chi_r$, where $\chi_1\nek\chi_r$ is any list of elements in $H^1(G)$ such that $K=\bigcap_{i=1}^r\Ker(\chi_i)$.
\begin{exam}
\label{example r=1}
\rm
Let $\chi=\chi_1\in H^1(G)$, so $K=\Ker(\chi)$.
The cup product--restriction property for $\chi$ means that the sequence
\begin{equation}
\label{exact sequence for r=1}
H^1(G)\xrightarrow{\chi\cup} H^2(G)\xrightarrow{\res} H^2(K),
\end{equation}
is exact.
This is trivial when $\chi=0$, so we assume that $\chi\neq0$ and therefore $(G:K)=p$.
When $p=2$ (\ref{exact sequence for r=1}) is always exact, and is in fact a segment of the infinite Arason exact sequence \cite{Arason75}*{Satz 4.5}:
\[
\cdots\xrightarrow{\res}H^i(K)\xrightarrow{\Cor}H^i(G)\xrightarrow{\chi\cup}H^{i+1}(G)
\xrightarrow{\res}H^{i+1}(K)\xrightarrow{\Cor}\cdots\ .
\]
When $p$ is an arbitrary prime and $F$ is a field (containing as always a fixed root of unity of order $p$), we may write $\chi=(a)$ for some $a\in F^\times$.
Then $L=F(a^{1/p})$ is the $\dbZ/p$-Galois extension corresponding to $K$.
There is an isomorphism
\[
F^\times/N_{L/F}L^\times\xrightarrow{\sim}\Br(L/F), \quad xN_{L/F}L^\times\mapsto [(a,x)_F]
\]
\cite{Draxl83}*{p.\ 73, Th.\ 1}.
It follows that (\ref{exact sequence for r=1}) is exact, i.e., the cup product--restriction property for $\chi$ holds.
This is again a part of a more general fact:
Based on results of Voevodsky \cite{Voevodsky03}*{\S5}, it was shown by Lemire, Min\'a\v c and Swallow \cite{LemireMinacSwallow07}*{Th.\ 6} that for every $i\geq1$ the following sequence is exact:
\[
H^i(L)\xrightarrow{\Cor}H^i(F)
\xrightarrow{\chi\cup}H^{i+1}(F)\xrightarrow{\res} H^{i+1}(L),
\]
\end{exam}
\begin{rem}
\label{Dec}
\rm
Suppose that $F$ is a field containing a root of unity of order $p$, and $L$ is a Galois extension of $F$ with $\Gal(L/F)$ an elementary abelian $p$-group.
As in \cite{Tignol81a}, \cite{Tignol81b} let $\Dec(L/F)$ be the subgroup of $\Br(L/F)$ generated by all subgroups $\Br(L'/F)$, where $L'$ ranges over all cyclic $p$-extensions of $F$ contained in $L$.
Then the cup product--restriction property holds for the subgroup $G_L$ of $G_F$ if and only if $\Dec(L/F)={}_p\Br(L/F)$.
\end{rem}
\begin{exam}
\label{McKinnies example}
\rm
For $p$ odd there are constructions due to Tignol \cite{Tignol87}*{Th.\ 1, Rem.\ 1.3(a)} and McKinnie \cite{McKinnie11} (see also \cite{Saltman79} and \cite{Rowen82} for related works), of division algebras $D$ over a field $F$ which contains a root of unity of order $p$, such that
\begin{enumerate}
\item[(a)]
$D$ splits in $L=F(a_1^{1/p},a_2^{1/p})$ for certain $a_1,a_2\in F^\times$;
\item[(b)]
$[D]\not\in\Dec(L/F)$.
\end{enumerate}
In view of Remark \ref{Dec}, this means that the cup product--restriction property does not hold for the subgroup $K=G_L$ of $G=G_F$.
\end{exam}
\section{Global fields}
Throughout this section we assume that $F$ is a global field containing a root of unity of order $p$.
For every place $v$ on $F$ we write $F_v$ for the completion of $F$ relative to $v$, and denote the canonical extension of $v$ to $F_v$ also by $v$.
There is a canonical monomorphism $\Inv_v\colon \Br(F_v)\to\dbQ/\dbZ$ which is an isomorphism for $v$ non-archimedean.
Restricting to ${}_p\Br(F)$, we obtain a monomorphism $\Inv_v\colon H^2(F_v)\to\frac1p\dbZ/\dbZ$.
It is an isomorphism unless $v$ is archimedean and $p\neq2$ (and in the latter case $H^2(F_v)=0$).
Given a finite extension $E$ of $F_v$, there is a commutative square
\[
\xymatrix{
\Br(E)\ar[r]^{\Inv_u} &\dbQ/\dbZ \\
\Br(F_v)\ar[r]^{\Inv_v}\ar[u]^{\res_E} &\dbQ/\dbZ\ar[u]_{[E:F_v]},
}
\]
where the map on the right means multiplication by the degree $[E:F_v]$ \cite{SerreLocalFields}*{Ch.\ XIII, \S3, Prop.\ 7}.
Consequently, if $p|[E:F_v]$, then $\res_E\colon H^2(F_v)\to H^2(E)$ is the zero map.
We recall that, by classical results of Albert, Brauer, Hasse and Noether, the following sequence is exact:
\[
0\to\Br(F)\xrightarrow{\res}\bigoplus_v\Br(F_v)\xrightarrow{\sum_v\Inv_v}\dbQ/\dbZ\to 0.
\]
It gives rise to an exact sequence
\begin{equation}
\label{LGP}
0\to H^2(F)\xrightarrow{\res}\bigoplus_v H^2(F_v)\xrightarrow{\sum_v\Inv_v}\tfrac1p\dbZ/\dbZ.
\end{equation}
\begin{lem}
\label{v0}
Let $S$ be a finite set of places on $F$ and let $a_1\nek a_r\in F^\times$.
Suppose that $a_1\not\in (F^\times)^p$.
Then there exists a place $v_0$ on $F$ such that $v_0\not\in S$ and
$a_1,a'_2\nek a'_r\not\in (F_{v_0}^\times)^p$, where for each $2\leq i\leq r$ either $a'_i=a_i$ or $a'_i=a_1a_i$.
\end{lem}
\begin{proof}
The restriction map $H^1(F)\to\prod_{v\not\in S}H^1(F_v)$ is injective \cite{ArtinTate}*{Ch.\ IX, Th.\ 1}.
Hence there is a place $v_0\not\in S$ with $(a_1)\neq0\in H^1(F_{v_0})$.
Then for every $i$ with $2\leq i\leq r$ we have $(a_i)\neq0\in H^1(F_{v_0})$ or $(a_1a_i)=(a_1)+(a_i)\neq0\in H^1(F_{v_0})$, and we can choose $a'_i$ accordingly.
\end{proof}
\medskip
\begin{proof}[Proof of Case (1) of the Main Theorem]
Let $a_1\nek a_r\in F^\times$.
If $a_1\nek a_r\in (F^\times)^p$, then the cup product--restriction property for $(a_1)\nek(a_r)\in H^1(F)$ is trivial.
We may therefore assume that $a_1\not\in (F^\times)^p$.
Let $L=F(a_1^{1/p}\nek a_r^{1/p})$,
and consider $\alp\in H^2(F)$ with $\res_L(\alp)=0$ in $H^2(L)$.
We have to show that $\alp\in (a_1)\cup H^1(F)+\cdots+ (a_r)\cup H^1(F)$.
To this end let $S$ be the set of all places $v$ on $F$ such that $\alp_{F_v}\neq0$, where $\alp_{F_v}$ denotes the restriction of $\alp$ to $H^2(F_v)$.
By (\ref{LGP}), $S$ is finite.
Let $v_0$ and $a'_2\nek a'_r$ be as in Lemma \ref{v0}.
In particular, $\alp_{F_{v_0}}=0$.
Now $L=F(a_1^{1/p},(a'_2)^{1/p}\nek (a'_r)^{1/p})$ and
\[
\begin{split}
&(a_1)\cup H^1(F)+(a_2)\cup H^1(F)+\cdots+(a_r)\cup H^1(F)\\
=&(a_1)\cup H^1(F)+(a'_2)\cup H^1(F)+\cdots+(a'_r)\cup H^1(F).
\end{split}
\]
We may therefore replace $a_i$ by $a'_i$, $i=2\nek r$, to assume without loss of generality that $a_1\nek a_r\not\in (F_{v_0}^\times)^p$.
Next for $1\leq i\leq r$ let
\[
S_i=\{v\ |\ a_i^{1/p}\not\in F_v\}.
\]
Thus $v_0\in (S_1\cap\cdots\cap S_r)\setminus S$.
If $v\in S$, then $L\not\subseteq F_v$, so $a_i^{1/p}\not\in F_v$ for some $1\leq i\leq r$.
This shows that $S\subseteq S_1\cup\cdots\cup S_r$.
Hence we may decompose $S=S'_1\discup\cdots\discup S'_r$ with $S'_i\subseteq S_i$, $i=1,2\nek r$.
Note that $v_0\not\in S'_i$ for every $i$.
For every $i$ let
\[
t_i:=\sum_{v\in S'_i} \Inv_v(\alp_{F_v}).
\]
Then (\ref{LGP}) gives rise to $\alp_i\in H^2(F)$ with local invariants
\[
\Inv_v((\alp_i)_{F_v})
=\begin{cases}
\Inv_v(\alp_{F_v}), & \hbox{if } v\in S'_i,\\
-t_i,& \hbox{if } v=v_0,\\
0,& \hbox{otherwise}.\\
\end{cases}
\]
{\bf Claim 1:} \ For every place $w$ on $F$,
\begin{equation}
\label{equality of local invariants}
\Inv_w(\alp_{F_w})=\sum_{i=1}^r\Inv_w((\alp_i)_{F_w}).
\end{equation}
Indeed, when $w\in S'_i$ for some $i$, this follows from the disjointness of $S'_1\nek S'_r$.
When $w=v_0$ we compute using (\ref{LGP}):
\[
\begin{split}
\Inv_w(\alp_{F_w})=0&=-\sum_v\Inv_v(\alp_{F_v})=-\sum_{v\in S}\Inv_v(\alp_{F_v})\\
&=-\sum_{i=1}^r\sum_{v\in S'_i} \Inv_v(\alp_{F_v})=-\sum_{i=1}^rt_i=\sum_{i=1}^r\Inv_w((\alp_i)_{F_w}).
\end{split}
\]
For all other places, the left-hand side of (\ref{equality of local invariants}) is zero, by the definition of $S$, and all summands on the right-hand side are zero by the choice of $\alp_i$.
This proves the claim.
We conclude from Claim 1 and from (\ref{LGP}) that $\alp=\sum_{i=1}^r\alp_i$ in $H^2(F)$.
\medskip
{\bf Claim 2:} \
For every $1\leq i\leq r$ and every place $u$ on $F(a_i^{1/p})$ one has
\[
(\alp_i)_{F(a_i^{1/p})_u}=0.
\]
To see this let $v$ be the place on $F$ which lies under $u$.
\medskip
Case 1: $v\in S_i$.
Then $p=[F_v(a_i^{1/p}):F_v]\bigm|[F(a_i^{1/p})_u:F_v]$, so as we have seen, $\res\colon H^2(F_v)\to H^2(F(a_i^{1/p})_u)$ is the zero map.
In particular, $(\alp_i)_{F(a_i^{1/p})_u}=0$.
Case 2: $v\not\in S_i$.
Then $a_i^{1/p}\in F_v$ and $v\neq v_0$.
Hence $F(a_i^{1/p})_u=F_v$.
The choice of $\alp_i$ implies that $\Inv_v((\alp_i)_{F_v})=0$, so again,
$(\alp_i)_{F(a_i^{1/p})_u}=(\alp_i)_{F_v}=0$.
\medskip
We conclude from Claim 2 and from (\ref{LGP}) that $(\alp_i)_{F(a_i^{1/p})}=0$, $i=1,2\nek r$.
Therefore $\alp_i\in (a_i)\cup H^1(F)$ (Example \ref{example r=1}).
It follows that
\[
\alp=\sum_{i=1}^r\alp_i\in (a_1)\cup H^1(F)+\cdots+(a_r)\cup H^1(F).
\qedhere
\]
\end{proof}
\section{The case $p=2$}
Let $p=2$ and let $F$ be a field of characteristic $\neq2$.
By a classical result of Albert \cite{Albert39}, every central simple $F$-algebra of degree $4$ and exponent $2$ is $F$-isomorphic to a tensor product of two $F$-quaternion algebras.
This was extended by Tignol \cite{Tignol79}*{Th.\ 1} as follows.
Recall that an involution on a central simple $F$-algebra is of the \textsl{first kind} if it is the identity on $F$.
\begin{thm}[Tignol]
\label{stronger Albert}
Let $A$ be a central simple $F$-algebra with involution of the first kind and which is split by
a Galois extension $M$ of $F$ with Galois group $(\dbZ/2\dbZ)^2$.
Let $L_1,L_2$ be quadratic extensions of $F$ with $M=L_1L_2$.
Then there are quaternion $F$-algebras $Q_1,Q_2$ such that $L_i\subset Q_i\subset A$, $i=1,2$, and $A\isom_F Q_1\tensor_F Q_2$.
\end{thm}
For a closely related result see \cite{Rowen84}*{Cor.\ 5}.
The case $p=2$, $r=2$ of the Main Theorem is an easy corollary of Theorem \ref{stronger Albert}.
In a different terminology it was obtained by Tignol in \cite{Tignol81a}*{Cor.\ 2.8}, however we provide a short proof showing its relation to Theorem \ref{stronger Albert}.
\begin{proof}[Proof of Case (2) of the Main Theorem]
Let $a_1,a_2\in F^\times$ and denote $M=F(\sqrt a_1,\sqrt a_2)$.
Thus in the terminology of sequence (\ref{cup product-restriction}), $K=G_M$.
Since the sequence (\ref{cup product-restriction}) is always a complex, we need to show that for every central simple $F$-algebra $A$ of exponent $2$ and which splits in $M$
the similarity class $[A]$ of $A$ in ${}_2\Br(F)$ is contained in $(a_1)\cup H^1(G)+(a_2)\cup H^1(G)$.
We may assume that $A$ does not split in $F$.
If $a_1,a_2$ have $\dbF_2$-linearly dependent cosets in $F^\times/(F^\times)^2$, then we are done by Example \ref{example r=1}.
So assume that $a_1,a_2$ have $\dbF_2$-linearly independent cosets in $F^\times/(F^\times)^2$.
Since $A$ splits in $M$, it is similar to a central simple $F$-algebra $A'$ of degree $[M:F]=4$ and which contains $M$ \cite{Draxl83}*{p.\ 64, Th.\ 7}.
The exponent of $A'$ is also $2$, and therefore it has an involution of the first kind \cite{Albert39}*{Ch.\ X, Th.\ 19}.
Let $L_i=F(\sqrt{a_i})$, $i=1,2$.
Theorem \ref{stronger Albert} yields quaternion $F$-subalgebras $Q_1,Q_2$ of $A'$ which contain $L_1,L_2$ respectively, and such that $A'\isom_F Q_1\tensor_F Q_2$.
Then $Q_1\isom_F(a_1,x)_F$ and $Q_2\isom_F(a_2,y)_F$ for some $x,y\in F^\times$ \cite{Draxl83}*{p.\ 104, Th.\ 4}.
Therefore $[A]=[A']=(a_1)\cup (x)+(a_2)\cup(y)$, as desired.
\end{proof}
\begin{rem}
\label{rems on the odd prime case}
\rm
There are no known direct generalizations of Theorem \ref{stronger Albert} and \cite{Rowen84}*{Cor.\ 5} for odd primes.
For instance, when $p=3$, it seems that the best result to date is that a central simple $F$-algebra which contains a maximal subfield $L$
which is Galois over $F$ with $\Gal(L/F)\isom\dbZ/3\times\dbZ/3$ (i.e., $A$ is a $\dbZ/3\times\dbZ/3$-crossed product over $F$) is similar to the tensor product of $\leq31$ symbol algebras of degree $3$ over $F$ \cite{Matzri14}.
\end{rem}
\section{Massey products}
\label{section on Massey products}
We recall the definition and basic properties of Massey products of degree $1$ cohomology elements.
For more information see e.g., Fenn \cite{Fenn83}, Kraines \cite{Kraines66}, Dwyer \cite{Dwyer75} (and in a more general setting, May \cite{May69}).
Note that the various sources use different sign conventions.
We first recall that a \textsl{differential graded algebra} over a ring $R$ ($R$-DGA) is a graded $R$-algebra
$C^\bullet=\bigoplus_{s=0}^\infty C^s$ equipped with $R$-module homomorphisms
$\partial^s\colon C^s\to C^{s+1}$ such that $(C^\bullet,\bigoplus_{s=0}^\infty\partial^s)$ is a complex satisfying the \textsl{Leibnitz rule}
$\partial^{r+s}(ab)=\partial^r(a)b+(-1)^ra\partial^s(b)$ for $a\in C^r$, $b\in C^s$.
Set $Z^r=\Ker(\partial^r)$, $B^r=\Img(\partial^{r-1})$, and $H^r=Z^r/B^r$, and let $[c]$ denote the class of $c\in Z^r$ in $H^r$.
We fix an integer $n\geq 2$.
Consider a system $c_{ij}\in C^1$, where $1\leq i\leq j\leq n$ and $(i,j)\neq(1,n)$.
For any $i,j$ satisfying $1\leq i\leq j\leq n$ (including $(i,j)=(1,n)$) we define
\[
\widetilde c_{ij}=-\sum_{r=i}^{j-1}c_{ir}c_{r+1,j}\in C^2.
\]
One says that $(c_{ij})$ is a \textsl{defining system of size $n$} in $C^\bullet$ if
$\partial c_{ij}=\widetilde c_{ij}$ for every $1\leq i\leq j\leq n$ with $(i,j)\neq(1,n)$.
We also say that the defining system $(c_{ij})$ is \textsl{on $c_{11}\nek c_{nn}$}.
Note that then $c_{ii}$ is a $1$-cocycle, $i=1,2\nek n$.
Further, $\widetilde c_{1n}$ is a $2$-cocycle (\cite{Kraines66}*{p.\ 432}, \cite{Fenn83}*{p.\ 233}).
Its cohomology class depends only on the cohomology classes $[c_{11}]\nek[c_{nn}]$ \cite{Kraines66}*{Th.\ 3}.
Given $c_1\nek c_n\in Z^1$, the \textsl{$n$-fold Massey product} of $\langle [c_1]\nek [c_n]\rangle$ is the subset of $H^2$ consisting of all cohomology classes $[\widetilde{c_{1n}}]$
obtained from defining systems $(c_{ij})$ of size $n$ on $c_1\nek c_n$ in $C^\bullet$.
This construction is functorial in the natural sense.
When this subset is nonempty one says that $\langle [c_1]\nek [c_n]\rangle$ is \textsl{defined}.
Note that $\langle [c_1]\nek [c_n]\rangle$ contains $0$ if and only if there is an array $(c_{ij})$, $1\leq i\leq j\leq n$, in $C^1$ such that $\partial c_{ij}=\widetilde c_{ij}$ for every $1\leq i\leq j\leq n$
(including $(i,j)=(1,n)$).
In this case one says that the Massey product $\langle [c_1]\nek [c_n]\rangle$ is \textsl{trivial}.
When $n=2$, $\langle [c_1],[c_2]\rangle$ is always defined and consists only of $-[c_1][c_2]$.
Next we record some well-known facts on the case $n=3$.
\begin{prop}
\label{structure of triple Massey products}
Let $c_1,c_2,c_3\in Z^1$.
\begin{enumerate}
\item[(a)]
$\langle [c_1],[c_2],[c_3]\rangle$ is defined if and only if $[c_1][c_2]=[c_2][c_3]=0$;
\item[(b)]
If $(c_{ij})$ is a defining system on $[c_1],[c_2],[c_3]$, then
$\langle [c_1],[c_2],[c_3]\rangle=[\widetilde {c_{13}}]+[c_1]H^1+[c_3]H^1$.
\end{enumerate}
\end{prop}
\begin{proof}
(a) \quad
Having a defining system on $c_1,c_2,c_3$ means that there exist $c_{12},c_{23}\in C^1$ with $\partial c_{12}=-c_1c_2$ and $\partial c_{23}=-c_2c_3$,
i.e., $c_1c_2,c_2c_3\in B^2$.
\medskip
(b) \quad
Suppose that $(c'_{ij})$ is another defining system on $c_1,c_2,c_3$.
For $d_{12}=c'_{12}-c_{12}$ and $d_{23}=c_{23}-c'_{23}$ we have $\partial d_{12}=\widetilde{c'_{12}}-\widetilde{c_{12}}=-c_1c_2+c_1c_2=0$.
Thus $d_{12}\in Z^1$, and similarly $d_{23}\in Z^1$.
By a direct calculation $[\widetilde{c'_{13}}]=[\widetilde{c_{13}}]+[c_1][d_{23}]+
[c_3][d_{12}]$.
Conversely, for every $d_{12},d_{23}\in Z^1$, the system $(c'_{ij})$ is also a defining system on $c_1,c_2,c_3$, where we take
$c'_{ii}=c_i$, $c'_{12}=c_{12}-d_{12}$ and $c'_{23}=c_{23}+d_{23}$.
One has $[\widetilde{c'_{13}}]=[\widetilde{c_{13}}]+[c_1][d_{23}]+[c_3][d_{12}]$.
\end{proof}
Now for the fixed prime number $p$, let $G$ be a profinite group acting trivially on $\dbZ/p$.
Let $C^\bullet=\bigoplus_{s=0}^\infty C^s(G,\dbZ/p)$ be the $\dbZ/p$-DGA of continuous cochains from $G$ to $\dbZ/p$, with the cup product.
Thus in our previous notation, $H^i=H^i(G)$.
\begin{prop}
\label{cohomological lemma}
Let $\chi_1,\chi_2,\chi_3\in H^1(G)$.
Suppose that the triple Massey product $\langle\chi_1,\chi_2,\chi_3\rangle$ is defined, and that the cup product--restriction property holds for $\chi_1,\chi_3$.
Then $0\in\langle\chi_1,\chi_2,\chi_3\rangle$.
\end{prop}
\begin{proof}
Take $\alp\in\langle\chi_1,\chi_2,\chi_3\rangle$.
Let $K=\Ker(\chi_1)\cap\Ker(\chi_3)$.
The functoriality of the Massey product implies that
\[
\res_K(\alp)\in \langle\res_K(\chi_1),\res_K(\chi_2),\res_K(\chi_3)\rangle=\langle0,\res_K(\chi_2),0\rangle=\{0\}.
\]
By the exactness of (\ref{cup product-restriction}), $\alp\in \Img(\Lam_{(\chi_1,\chi_3)})$, that is,
$\alp=\chi_1\cup\beta_1+\chi_3\cup\beta_3$ for some $\beta_1,\beta_3\in H^1(G)$.
Now Proposition \ref{structure of triple Massey products} implies that
\[
0=\alp-\chi_1\cup\beta_1-\chi_3\cup\beta_3\in \langle\chi_1,\chi_2,\chi_3\rangle.
\qedhere
\]
\end{proof}
Dwyer \cite{Dwyer75} relates $n$-fold Massey products in $C^\bullet(G,\dbZ/p)$
to unipotent upper-triangular $n+1$-dimensional representations of $G$ as follows
(\cite{Dwyer75} works in a discrete context and with more a general coefficients ring;
see \cite{Efrat14}*{\S8} for the profinite context, and \cite{Wickelgren12} for a generalization to the case of non-trivial actions).
For $n\geq2$ let $\dbU_{n+1}(\dbF_p)$ be as before the group of all unipotent upper-triangular $(n+1)\times(n+1)$-matrices over $\dbF_p$.
Its center consists of all matrices which are $0$ on all off-diagonal entries, except possibly for entry $(1,n+1)$.
Let $\bar\dbU_{n+1}(\dbF_p)$ be the quotient of $\dbU_{n+1}(\dbF_p)$ by this center.
Its elements may be viewed as unipotent upper-triangular $(n+1)\times(n+1)$-matrices with the $(1,n+1)$-entry deleted.
We notice that $\dbU_{n+1}(\dbF_p)$ is a $p$-group, and its Frattini subgroup is the kernel of the epimorphism
$\dbU_{n+1}(\dbF_p)\to(\dbZ/p)^n$, $(c_{ij})\mapsto(c_{12},c_{23}\nek c_{n,n+1})$, and similarly for $\bar \dbU_{n+1}(\dbF_p)$.
Given an array $(c_{ij})$, $1\leq i\leq j\leq n$, in $C^1(G,\dbZ/p)$, we define a continuous map $\gam\colon G\to\dbU_{n+1}(\dbF_p)$
by $\gam(\sig)_{ij}=(-1)^{j-i}c_{i,j-1}(\sig)$ for $\sig\in G$ and for $1\leq i<j\leq n+1$.
Then $\widetilde{c_{ij}}=\partial c_{ij}$ for every $i<j$ if and only if $\gam\colon G\to\dbU_{n+1}(\dbF_p)$ is a homomorphism.
Similarly, $\widetilde{c_{ij}}=\partial c_{ij}$ for every $i<j$ with $(i,j)\neq(1,n)$ if and only if the induced map $\bar\gam\colon G\to\bar\dbU_{n+1}(\dbF_p)$ is a homomorphism.
We write $\gam_{ij},\bar\gam_{ij}$ for the projections of $\gam,\bar\gam$, respectively, on the $(i,j)$-coordinate.
\begin{prop}
\label{Massey products and homorphisms}
Let $\chi_1\nek\chi_n\in H^1(G)$ be $\dbF_p$-linearly independent.
\begin{enumerate}
\item[(b)]
$\langle\chi_1\nek\chi_n\rangle$ is defined if and only if
there exists a continuous homomorphism $\bar\gam\colon G\to\bar\dbU_{n+1}(\dbF_p)$ such that $\bar\gam_{i,i+1}=\chi_i$, $i=1,2\nek n$,
\item[(B)]
$\langle\chi_1\nek\chi_n\rangle$ is trivial if and only if
there exists a continuous homomorphism $\gam\colon G\to\dbU_{n+1}(\dbF_p)$ such that $\gam_{i,i+1}=\chi_i$, $i=1,2\nek n$.
\end{enumerate}
Moreover, such homomorphisms $\gam,\bar\gam$ are necessarily surjective.
\end{prop}
\begin{proof}
(a) and (b) follow from the above discussion.
The surjectivity of $\gam$ and $\bar\gam$ follows by a Frattini argument.
\end{proof}
This and Proposition \ref{structure of triple Massey products}(a) imply the following facts, mentioned in the Introduction, which are also implicit in \cite{MinacTan14}*{Cor.\ 3.2}:
\begin{cor}
\label{Massey products and Un}
Let $\chi_1,\chi_2,\chi_3\in H^1(G)$ be $\dbF_p$-linearly independent.
Then:
\begin{enumerate}
\item[(a)]
$\langle\chi_1,\chi_2,\chi_3\rangle$ is defined if and only if there exist continuous epimorphisms $\gam',\gam''\colon G\to\dbU_3(\dbF_p)$
such that
$\gam'_{12}=\chi_1$, $\gam'_{23}=\chi_2$, $\gam''_{12}=\chi_2$, $\gam''_{23}=\chi_3$.
\item[(b)]
$\langle\chi_1,\chi_2,\chi_3\rangle$ is trivial if and only if there exists a continuous epimorphism $\gam\colon G\to\dbU_4(\dbF_p)$ such that
$\gam_{12}=\chi_1$, $\gam_{23}=\chi_2$ and $\gam_{34}=\chi_3$.
\end{enumerate}
\end{cor}
\begin{bibdiv}
\begin{biblist}
\bib{Albert39}{book}{
author={Albert, A. Adrian},
title={Structure of Algebras},
series={American Mathematical Society Colloquium Publications, Vol. XXIV},
publisher={American Mathematical Society, Providence, R.I.},
date={1939},
}
\bib{Arason75}{article}{
author={Arason, J{\'o}n Kr.},
title={Cohomologische Invarianten quadratischer Formen},
journal={J. Algebra},
volume={36},
date={1975},
pages={448--491},
}
\bib{ArtinTate}{book}{
author={Artin, Emil},
author={Tate, John},
title={Class Field Theory},
note={Reprinted with corrections from the 1967 original},
publisher={AMS Chelsea Publishing, Providence, RI},
date={2009},
pages={viii+194},
}
\bib{Draxl83}{book}{
author={Draxl, P.K.},
title={Skew Fields},
series={London Math. Soc.\ Lect.\ Notes Series},
volume={81},
publisher={Cambridge University Press},
place={Cambridge},
date={1983},
}
\bib{Dwyer75}{article}{
author={Dwyer, William G.},
title={Homology, Massey products and maps between groups},
journal={J. Pure Appl. Algebra},
volume={6},
date={1975},
pages={177\ndash190},
}
\bib{Efrat14}{article}{
author={Efrat, Ido},
title={The Zassenhaus filtration, Massey products, and representations of profinite groups},
journal={Adv.\ Math.},
volume={263},
date={2014},
pages={389\ndash411},
}
\bib{EfratMinac11}{article}{
author={Efrat, Ido},
author={Min\' a\v c, J\'an},
title={On the descending central sequence of absolute Galois groups},
journal={Amer.\ J.\ Math.},
volume={133},
date={2011},
pages={1503\ndash1532},
}
\bib{ElmanLamTignolWadsworth83}{article}{
author={Elman, Richard},
author={Lam, T. Y.},
author={Tignol, Jean-Pierre},
author={Wadsworth, A.R.},
title={Witt rings and Brauer groups under multiquadratic extensions, I},
journal={Amer. J. Math.},
volume={105},
date={1983},
pages={1119\ndash1170},
}
\bib{Fenn83}{book}{
author={Fenn, Roger A.},
title={Techniques of Geometric Topology},
Series={London Math.\ Soc.\ Lect. Notes Series},
volume={57},
publisher={Cambridge Univ. Press},
date={1983},
place={Cambridge}
}
\bib{HopkinsWickelgren15}{article}{
author={Hopkins, Michael J.},
author={Wickelgren, Kirsten G.},
title={Splitting varieties for triple Massey products},
journal={J. Pure Appl. Algebra},
volume={219},
date={2015},
pages={1304--1319},
}
\bib{Kraines66}{article}{
author={Kraines, David},
title={Massey higher products},
journal={Trans.\ Amer.\ Math.\ Soc.},
volume={124},
date={1966},
pages={431\ndash449},
}
\bib{LemireMinacSwallow07}{article}{
author={Lemire, Nicole},
author={Min\'a\v c, J\'an},
author={Swallow, John},
title={Galois module structure of Galois Cohomology and partial Euler-Poincare Characteristics},
journal={J.\ reine angew.\ Math.},
volume={613},
date={2007},
pages={147\ndash173},
}
\bib{Matzri14}{article}{
author={Matzri, Eliyahu},
title={$\dbZ_3\times\dbZ_3$-crossed products},
journal={J.\ Algebra},
volume={418},
date={2014},
pages={1\ndash7},
}
\bib{May69}{article}{
author={May, J. Peter},
title={Matric Massey products},
journal={J. Algebra},
volume={12},
pages={533\ndash 568},
date={1969}
}
\bib{McKinnie11}{article}{
author={McKinnie, Kelly},
title={Degeneracy and decomposability in abelian crossed products},
journal={J. Algebra},
volume={328},
date={2011},
pages={443\ndash460},
}
\bib{MinacTan13}{article}{
author={Min\'a\v c, J\'an},
author={T\^an, Nguyen Duy},
title={Triple Massey products and Galois theory},
journal={J.\ Eur.\ Math.\ Soc.},
status={to appear},
eprint={arXiv:1307.6624},
date={2013},
}
\bib{MinacTan14}{article}{
author={Min\'a\v c, J\'an},
author={T\^an, Nguyen Duy},
title={Triple Massey products over global fields},
eprint={arXiv:1407.4586},
date={2014},
}
\bib{MinacTan15}{article}{
author={Min\'a\v c, J\'an},
author={T\^an, Nguyn Duy},
title={The kernel unipotent conjecture and the vanishing of Massey products for odd rigid fields},
journal={Adv. Math.},
volume={273},
date={2015},
pages={242--270},
status={(with an appendix by I.\ Efrat, J.\ Min\'a\v c, and N.D. T\^an)},
}
\bib{Rowen82}{article}{
author={Rowen, Louis Halle},
title={Cyclic division algebras},
journal={Israel J. Math.},
volume={41},
date={1982},
pages={213--234},
note={Correction: Israel J.\ Math.\ {\bf43} (1982), 277\ndash 280},
}
\bib{Rowen84}{article}{
author={Rowen, Louis H.},
title={Division algebras of exponent $2$ and characteristic $2$},
journal={J. Algebra},
volume={90},
date={1984},
pages={71--83},
}
\bib{Rowen}{book}{
author={Rowen, Louis Halle},
title={Graduate Algebra: Noncommutative View},
series={Graduate Studies in Mathematics},
volume={91},
publisher={Amer.\ Math.\ Soc., Providence, RI},
date={2008},
pages={xxvi+648},
}
\bib{Saltman79}{article}{
author={Saltman, David J.},
title={Indecomposable division algebras},
journal={Comm. Algebra},
volume={7},
date={1979},
pages={791--817},
}
\bib{SerreLocalFields}{book}{
author={Serre, Jean-Pierre},
title={Local Fields},
series={Grad.\ Texts Math.},
volume={67},
publisher={Springer-Verlag, New York-Berlin},
date={1979},
pages={viii+241},
}
\bib{Tignol79}{article}{
author={Tignol, J.-P.},
title={Central simple algebras with involution},
conference={
title={Ring theory (Proc. Antwerp Conf.},
address={NATO Adv. Study Inst.), Univ. Antwerp, Antwerp},
date={1978},
},
book={
series={Lecture Notes in Pure and Appl. Math.},
volume={51},
publisher={Dekker, New York},
},
date={1979},
pages={279--285},
}
\bib{Tignol81a}{article}{
author={Tignol, Jean-Pierre},
title={Corps \`a involution neutralis\'es par une extension ab\'elienne \'el\'ementaire,},
conference={
title={Groupe de Brauer (Les Plans-sur-Bex 1980, M.\ Kervaire and M.\ Ojanguren, Eds.)}, },
book={series={Lecture Notes in Math.},
volume={844},
publisher={Springer, Berlin},
}
date={1981},
pages={1\ndash34},
}
\bib{Tignol81b}{article}{
author={Tignol, Jean-Pierre},
title={Produits crois\'es ab\'eliens},
journal={J. Algebra},
volume={70},
date={1981},
pages={420--436},
}
\bib{Tignol87}{article}{
author={Tignol, Jean-Pierre},
title={Alg\`ebres ind\'ecomposables d'exposant premier},
journal={Adv.\ Math.},
volume={65},
pages={205\ndash228},
date={1987},
}
\bib{Voevodsky03}{article}{
author={Voevodsky, Vladimir},
title={Motivic cohomology with $\dbZ/2$-coefficients},
journal={Publ. Math. Inst. Hautes \'Etudes Sci.},
volume={98},
pages={59\ndash104},
date={2003},
}
\bib{Wickelgren12}{article}{
author={Wickelgren, Kirsten},
title={$n$-nilpotent obstructions to $\pi_1$ sections of $\mathbb{P}^1-\{0,1,\infty\}$ and Massey products},
conference={
title={Galois-Teichm\"uller theory and arithmetic geometry},
},
book={
series={Adv. Stud. Pure Math.},
volume={63},
publisher={Math. Soc. Japan, Tokyo},
},
date={2012},
pages={579--600},
}
\end{biblist}
\end{bibdiv}
\end{document}
|
\section{Introduction}
The {\em generalized cluster algebras\/}
were introduced by Chekhov and Shapiro
\cite{Chekhov11}.
They naturally generalize the (ordinary) cluster algebras by
Fomin and Zelevinsky \cite{Fomin03a}.
The main feature of the generalized cluster algebras
is the appearance of {\em polynomials\/}
in the exchange relations of cluster variables and
coefficients, instead of {\em binomials\/} in the ordinary case.
Generalized cluster algebras naturally appear so far in
Poisson dynamics \cite{Gekhtman02}, Teichm\"uller theory
\cite{Chekhov11}, representation theory \cite{Gleitz14},
exact WKB analysis \cite{Iwaki14b}, etc.
It has been shown
in
\cite{Chekhov11,Nakanishi14a}
that essentially all important properties
of the ordinary cluster algebras
are naturally extended to the generalized ones.
In this note we demonstrate that the notion of {\em quantum cluster algebras\/}
is also extended to the generalized ones.
To be more precise,
there are two kinds of formulations of quantum cluster algebras,
the one quantizing the {\em cluster variables\/} by \cite{Berenstein05b} and the one quantizing the {\em coefficients\/} by \cite{Fock03,Fock07},
and it is known that they are closely related to each other.
Here, we concentrate on the latter one.
As shown by \cite{Fock03,Fock07},
in the ordinary case,
the quantization of the coefficients is tightly integrated with
the {\em quantum dilogarithm\/} \cite{Faddeev93,Faddeev94}.
Similarly,
in the generalized case,
it is tightly integrated with certain generalizations
of the quantum dilogarithm, which we call the {\em quantum dilogarithms of higher degrees}.
As an application,
we derive the identities of these generalized quantum dilogarithms
associated with any period of quantum $Y$-seeds,
which are also parallel to the ones in the ordinary case.
The main message of this note is
that
the fundamental (and perhaps all) features of the quantum cluster algebras
are also extended to the generalized ones.
\section{Quantum dilogarithms of higher degrees}
To begin with,
let us recall some basic facts about the dilogarithm, the $q$-dilogarithm, and the quantum dilogarithm.
The {\em dilogarithm\/} $\mathrm{Li}_2(x)$ is defined by
\begin{align}
\mathrm{Li}_2(x)=\sum_{n=1}^{\infty} \frac{x^n}{n^2}.
\end{align}
Let $q$ be a formal variable.
The {\em $q$-dilogarithm} is defined
as follows.
\begin{align}
\label{eq:L1}
\mathcal{L}_{2,q}(x)=
\sum_{n=1}^{\infty}
\frac{x^n}{n(q^n-q^{-n})}
=
\frac{1}{q-q^{-1}}\sum_{n=1}^{\infty}
\frac{x^n}{n[n]_q},
\end{align}
where $[n]_q=(q^n-q^{-n})/(q-q^{-1})$ is the standard $q$-number.
The power series \eqref{eq:L1} converges for $|x|<1$ and $|q|<1$,
and the following asymptotic behavior holds when $q\rightarrow 1^{-}$,
\begin{align}
\label{eq:asym1}
\mathcal{L}_{2,q}(x)
\sim
\frac{\mathrm{Li_2}(x)}{q^2-1}\sim
\frac{\mathrm{Li_2}(x)}{\log q^2}.
\end{align}
This is clear form the second expression of $\mathcal{L}_{2,q}(x)$ in \eqref{eq:L1}
and the property $\lim_{q\rightarrow 1} [n]_q=n$.
Following \cite{Faddeev93,Faddeev94} (up to some convention),
we introduce the {\em quantum dilogarithm\/} $\Psi_{q}(x)$,
which is a formal power series in $x$
with coefficients in $\mathbb{C}(q)$, as follows.
\begin{align}
\Psi_{q}(x)&=
\prod_{m=0}^{\infty}
\left(
1+q^{2m+1}x
\right)^{-1}.
\end{align}
In particular, the quantum dilogarithm $\Psi_{q}(x)$
should be distinguished from the $q$-dilogarithm $\mathcal{L}_{2,q}(x)$.
The formal power series $\Psi_{q}(x)$ is characterized by the
following recursion relation with initial condition,
\begin{align}
\label{eq:rec1}
\Psi_{q}(0)&=1,
\quad
\Psi_{q}(q^{\pm2}x)
=
\left(
1+qx
\right)^{\pm1}
\Psi_{q}(x),
\end{align}
where two relations in the latter equality are equivalent to each other.
A little confusingly, the quantum dilogarithm is actually the exponential of the $q$-dilogarithm;
namely,
\begin{align}
\label{eq:psi1}
\Psi_{q}(x) = \exp\left(
-\mathcal{L}_{2,q}(-x)
\right).
\end{align}
This is easily shown by using the recursion relation \eqref{eq:rec1}.
Alternatively, one may define
the {\em dilogarithm\/} $\mathrm{Li}_2(x)$ by the integral
\begin{align}
\mathrm{Li}_2(x)=-\int_0^x \log(1-y)
\frac{dy}{y}
=
-\int_0^{-x} \log(1+y)
\frac{dy}{y}.
\end{align}
Then we have
\begin{align}
\label{eq:asym2}
\begin{split}
\log \Psi_{q}(x)
&=-
\sum_{m=0}^{\infty}
\log(1+q^{2m+1}x)\\
&=\frac{-1}{1-q^2}
\sum_{m=0}^{\infty}
\log(1+q^{2m+1}x)
\frac{q^{2m+1}x - q^{2m+3}x}{q^{2m+1}x}\\
&\sim
\frac{-1}{1-q^2}
\int_0^{x} \log(1+y)
\frac{dy}{y}
=
\frac{1}{1-q^2}\mathrm{Li}_2(-x)
\quad
(q\rightarrow 1^-).
\end{split}
\end{align}
This completely agrees with
\eqref{eq:asym1} and \eqref{eq:psi1}.
Now let us generalize the quantum dilogarithm $\Psi_{q}(x)$
to the ones with higher degrees.
For any field $F$, let $F(q)$ be the field of the rational functions in
the variable $q$.
\begin{defn}
\label{defn:dilog1}
Let $F$ be a field,
let $d$ be a positive integer,
and let $\mathbf{z}=(z_1,\dots,z_{d-1})$ be a $d-1$-tuple of elements
in $F$.
When $d=1$, $\mathbf{z}$ is regarded as the empty sequence $()$.
We set $z_0=z_d=1$.
Then, we define a formal power series
$\Psi_{d,\mathbf{z},q}(x)$ in $x$ with coefficients in $F(q)$ as follows:
\begin{align}
\label{eq:psi2}
\Psi_{d,\mathbf{z},q}(x)&=
\prod_{m=0}^{\infty}
\left(
\sum_{s=0}^d
z_s q^{s(2m+1)} x^s
\right)^{-1}.
\end{align}
When $d=1$, it is the usual quantum dilogarithm $\Psi_{1,(), q}(x)=
\Psi_{q}(x)$.
We call $\Psi_{d,\mathbf{z},q}(x)$ the {\em quantum dilogarithm of degree $d$
with coefficients $\mathbf{z}$}.
\end{defn}
\begin{prop}
The formal power series $\Psi_{d,\mathbf{z},q}(x)$ is characterized by the
the following recursion relation with initial condition:
\begin{align}
\label{eq:rec2}
\Psi_{d,\mathbf{z},q}(0)&=1,
\\
\label{eq:rec3}
\Psi_{d,\mathbf{z},q}(q^{\pm2}x)
&=
\left(
\sum_{s=0}^{d} z_s q^{\pm s} x^s
\right)^{\pm1}
\Psi_{d,\mathbf{z},q}(x),
\end{align}
where
two relations in \eqref{eq:rec3} are equivalent to each other.
\end{prop}
\begin{proof}
For example, we have
\begin{align}
\begin{split}
\Psi_{d,\mathbf{z},q}(q^{2}x)
&=
\prod_{m=0}^{\infty}
\left(
\sum_{s=0}^d
z_s q^{s(2m+1)} q^{2s} x^s
\right)^{-1}\\
&=
\prod_{m=1}^{\infty}
\left(
\sum_{s=0}^d
z_s q^{s(2m+1)} x^s
\right)^{-1}
=
\left(
\sum_{s=0}^{d} z_s q^{s} x^s
\right)
\Psi_{d,\mathbf{z},q}(x).
\end{split}
\end{align}
The rest of the properties are easily shown.
\end{proof}
For any integer $a$, let us introduce the {\em sign function\/}
\begin{align}
\mathrm{sgn}(a)=
\begin{cases}
+ & a > 0\\
0 & a =0\\
- & a<0.
\end{cases}
\end{align}
Here and below, we identify the signs $\pm$ with numbers $\pm 1$.
The following formula will be useful later.
\begin{prop}
\label{prop:rec1}
For any integer $a$, the following equality holds.
\begin{align}
\Psi_{d,\mathbf{z},q}(q^{2a}x)
=
\Biggl(
\prod_{m=1}^{|a|}
\Biggl(
\sum_{s=0}^d
z_s
q^{\mathrm{sgn}(a) (2m-1)s}
x^s
\Biggr)^{\mathrm{sgn}(a)}
\Biggr)
\Psi_{d,\mathbf{z},q}(x).
\end{align}
\end{prop}
\begin{proof}
This is obtained from
\eqref{eq:rec3} by induction on $a$.
\end{proof}
In some cases
the quantum dilogarithms of higher degrees
are
factorized by the ordinary quantum dilogarithm.
\begin{prop}
\label{prop:fac1}
Factorization formula.
Suppose that the following factorization
\begin{align}
\label{eq:fac1}
\sum_{s=0}^d
z_s x^s
=
\prod_{s=1}^d (1-w_s x)
\end{align}
occurs for some $w_1,\dots, w_d \in F$.
Then, we have
\begin{align}
\label{eq:fac2}
\Psi_{d,\mathbf{z},q}(x)
=\prod_{s=1}^d \Psi_q(-w_sx).
\end{align}
\end{prop}
\begin{proof}
One can directly observe the factorization in \eqref{eq:fac2}
as
\begin{align}
\Psi_{d,\mathbf{z},q}(x)=
\prod_{m=0}^{\infty}
\left(
\sum_{s=0}^d
z_s q^{s(2m+1)} x^s
\right)^{-1}
=
\prod_{m=0}^{\infty}
\prod_{s=1}^d (1-w_s q^{2m+1} x)^{-1}.
\end{align}
Alternatively,
under the assumption \eqref{eq:fac1},
the right hand side of \eqref{eq:fac2} satisfies
\eqref{eq:rec2} and \eqref{eq:rec3}.
Thus, thanks to Proposition
\ref{prop:rec1}, we have \eqref{eq:fac2}.
\end{proof}
\begin{ex} Let us consider the special case
where $F=\mathbb{C}$ and
the coefficients $\mathbf{z}$ is trivial, i.e., $\mathbf{z}=\mathbf{1}:=(1,\dots,1)$.
In this case we have the factorization
\begin{align}
\label{eq:pro6}
\sum_{s=0}^d x^s
=\prod_{s=1}^d (1-\omega^s x),
\end{align}
where
\begin{align}
\omega = \exp(2\pi i / (d+1)).
\end{align}
Thus, by Proposition \ref{prop:fac1}, we have
\begin{align}
\label{eq:pro5}
\Psi_{d,\mathbf{1},q}(x)
=
\prod_{s=1}^d
\Psi_q
(-\omega^s x).
\end{align}
On the other hand,
there is another factorization formula,
\begin{align}
\label{eq:pro3}
\Psi_{d,\mathbf{1},q}(x)=
\Psi_{q^{d+1}}(
-x^{d+1})
\Psi_{q}(-x) ^{-1}.
\end{align}
This is due to the following alternative expression of $\Psi_{d,\mathbf{1},q}(x)$,
\begin{align}
\label{eq:psi3}
\Psi_{d,\mathbf{1},q}(x)
=
\prod_{m=0}^{\infty}
\frac{
1- q^{2m+1}x
}{
\displaystyle
1- (q^{2m+1}x)^{d+1}
}.
\end{align}
Therefore, by \eqref{eq:asym1} and \eqref{eq:psi1},
we have the following asymptotic behavior in the limit $q\rightarrow 1^-$:
\begin{align}
\log \Psi_{d,\mathbf{1},q}(x)
\sim
&\frac{1}{1-q^2} \sum_{s=1}^d \mathrm{Li}_2(\omega^s x)\\
\sim&\frac{1}{1- q^2} \left(\frac{1}{d+1} \mathrm{Li}_2( x^{d+1})
-\mathrm{Li}_2( x)
\right).
\end{align}
In fact, these two expressions coincide due to
the well-known identity for $\mathrm{Li}_2(x)$ called the
{\em factorization formula\/} \cite[Eq.~(1.14)]{Lewin81},
\begin{align}
\label{eq:Lid2}
\frac{1}{d+1}\mathrm{Li}_{2}(x^{d+1})
=
\sum_{s=0}^d
\mathrm{Li}_{2}(\omega^s x).
\end{align}
As a side remark, in view of
\eqref{eq:psi1},
the expressions
\eqref{eq:pro5} and \eqref{eq:pro3} imply
the equality
\begin{align}
\label{eq:Lid1}
\mathcal{L}_{2,q^{d+1}}(x^{d+1})
=
\sum_{s=0}^d
\mathcal{L}_{2,q}(\omega^s x),
\end{align}
which is regarded as
the $q$-analogue of \eqref{eq:Lid2}.
The equality \eqref{eq:Lid1} is also obtained directly from
\eqref{eq:L1} and the equality
\begin{align}
\sum_{s=0}^d \omega^{sn}
=
\begin{cases}
d+1 & n \equiv 0 \mod d+1\\
0 & n \not\equiv 0 \mod d+1.\\
\end{cases}
\end{align}
\end{ex}
\begin{ex}
Let us consider the case where $F=\mathbb{C}$
with arbitrary coefficients $\mathbf{z}$.
Let us introduce the {\em dilogarithm of degree $d$ with coefficients $\mathbf{z}$}
by the integral,
\begin{align}
\mathrm{Li}_{2;d,\mathbf{z}}(x)=
-\int_0^{-x} \log\Biggl(
\sum_{s=0}^d
z_s y^s
\Biggr)
\frac{dy}y.
\end{align}
Then,
by the same calculation as in \eqref{eq:asym2},
we have the following asymptotic behavior,
\begin{align}
\label{eq:asym3}
\begin{split}
\log \Psi_{d,\mathbf{z},q}(x)
&=\frac{-1}{1-q^2}
\sum_{m=0}^{\infty}
\log
\Biggl(
\sum_{s=0}^d
z_s (q^{2m+1}x)^s
\Biggr)
\frac{q^{2m+1}x - q^{2m+3}x}{q^{2m+1}x}\\
&\sim
\frac{1}{1-q^2}\mathrm{Li}_{2;d,\mathbf{z}}(-x)
\quad
(q\rightarrow 1^-).
\end{split}
\end{align}
\end{ex}
\section{Generalized mutations of quantum $Y$-seeds}
In this section, following the idea of
\cite{Fock03,Fock07},
we introduce the quantum version of the
generalized mutation of generalized cluster algebras.
Here, we use the formulation of generalized cluster algebras
by \cite{Nakanishi14a}.
Let $B=(b_{ij})_{i,j=1}^n$ be a skew-symmetrizable integer matrix.
Let $\mathbf{d}=(d_1,\dots,d_n)$ be an $n$-tuple of
positive integers.
For given $B$ and $\mathbf{d}$,
we arbitrarily choose an $n$-tuple of
positive integers
$\mathbf{r}=(r_1,\dots,r_n)$
such that
\begin{align}
\label{eq:skew1}
r_i d_i b_{ij} = - r_jd_j b_{ji}.
\end{align}
Such an $\mathbf{r}$ exists (not uniquely) due to the skew-symmetrizable property of the matrix $B$.
Let $q$ continue to be a formal variable, and
let $Y=(Y_i)_{i=1}^n$ be an $n$-tuple of
noncommutative formal variables with
commutation relation
\begin{align}
\label{eq:Ycom1}
Y_i Y_j = q^{2r_j d_j b_{ji}} Y_j Y_i.
\end{align}
The relation \eqref{eq:Ycom1} makes sense
due to the skew-symmetric property in \eqref{eq:skew1}.
We call such a pair $(B,Y)$ a {\em quantum $Y$-seed}.
We use the notation
\begin{align}
q_i:= q^{r_i d_i},
\quad
i=1,\dots,n.
\end{align}
Then, \eqref{eq:Ycom1} is also written as
\begin{align}
\label{eq:Ycom3}
Y_i Y_j = q_j^{2 b_{ji}} Y_j Y_i
=q_i^{-2 b_{ij}} Y_j Y_i.
\end{align}
Let $F$ be any field.
For the above $\mathbf{d}=(d_1,\dots,d_n)$ we arbitrarily choose
a collection of elements in $F$,
\begin{align}
\mathbf{z}=(z_{i,s})_{i=1,\dots,n; s=1,\dots, d_i-1}
\end{align}
satisfying the {\em reciprocity condition\/} in \cite{Nakanishi14a}
\begin{align}
z_{i,s} = z_{i,d_i-s}.
\end{align}
(The use of symbol $\mathbf{z}$ here slightly conflicts with the one in
Definition \ref{defn:dilog1}, but we find that it is convenient.)
Let us set $z_{i,0}=z_{i,d_i}=1$.
We also introduce the notation
\begin{align}
\mathbf{z}_i = (z_{i,s})_{s=1,\dots,d_i-1},
\quad
i=1,\dots,n.
\end{align}
Under these notations we have the associated
quantum dilogarithm
$\Psi_{d_i,\mathbf{z}_i,q_i}(x)$ of degree $d_i$
for each $i=1,\dots,n$.
Below we assume that any element in $F$ commutes with variables $Y_i$.
\begin{defn}
For a quantum $Y$-seed $(B,Y)$,
the {\em $(\mathbf{d},\mathbf{z})$-mutation (generalized mutation) $(B',Y')=\mu_{k}(B,Y)$ of
$(B,Y)$ at $k$} is defined by
\begin{align}
\label{eq:bmut1}
b'_{ij}&=
\begin{cases}
-b_{ij} & \text{$i=k$ or $j=k$}\\
b_{ij} + d_k([-b_{ik}]_+ b_{kj}
+b_{ik}[b_{kj}]_+)
& i,j \neq k,
\end{cases}
\\
\label{eq:Ymut1}
Y'_{i}
&=
\begin{cases}
Y_k^{-1} & i = k\\
\displaystyle
q_i^{b_{ik} d_k [\varepsilon b_{ki}]_+}
Y_i Y_k^{d_k[\varepsilon b_{ki}]_+}\\
\displaystyle
\qquad
\times \prod_{m=1}^{|b_{ki}|}
\left(
\sum_{s=0}^{d_k}
z_{k,s}
q_k^{-\varepsilon \mathrm{sgn}(b_{ki}) (2m-1) s}
Y_k^{\varepsilon s}
\right)^{-\mathrm{sgn}(b_{ki})}
& i \neq k,
\end{cases}
\end{align}
where $\varepsilon=\pm$, and
\begin{align}
[a]_+=
\begin{cases}
a & a > 0\\
0 & a \leq 0.
\end{cases}
\end{align}
Actually, the right hand side of \eqref{eq:Ymut1} does not depend on
the choice of the sign $\varepsilon$ (see Lemma \ref{lem:mut1} (i)).
We call $\mathbf{d}$ and $\mathbf{z}$ the {\em mutation degrees}
and the {\em frozen coefficients}, respectively, in accordance with
\cite{Nakanishi14a}.
\end{defn}
When we formally set $q=1$,
the relation \eqref{eq:Ymut1}
reduces to
\begin{align}
\label{eq:ymut1}
Y'_i&=
\begin{cases}
\displaystyle
Y_k^{-1}
& i=k\\
\displaystyle
Y_i
Y_k^{d_k[\varepsilon {b}_{ki}]_+}
\Biggl(
\sum_{s=0}^{d_k}
z_{k,s} Y_k^{\varepsilon s}
\Biggr)^{-{b}_{ki}}
& i\neq k,\\
\end{cases}
\end{align}
which is
the
generalized mutation of coefficients ($y$-variables)
in generalized cluster algebras formulated in
\cite{Nakanishi14a}.
On the other hand,
when we set $d_k=1$, it reduces to the ordinary mutation of
quantum
$Y$-seeds by Fock and Goncharov \cite{Fock03,Fock07}.
The following properties are easily checked.
\begin{lem}
\label{lem:mut1}
(i) The right hand side of \eqref{eq:Ymut1} does not depend on
the choice of the sign $\varepsilon$.
\par
(ii) For the matrix $B'$, the condition
\begin{align}
r_i d_i b'_{ij} = - r_jd_j b'_{ji}.
\end{align}
holds.
\par
(iii) The $(\mathbf{d},\mathbf{z})$-mutation is involutive,
i.e., $\mu_k(\mu_k(B,Y))=(B,Y)$.
\end{lem}
In the rest of the section,
we will justify the relation \eqref{eq:Ymut1}
as a ``good" quantization of
the classical one \eqref{eq:ymut1}
in the sense of \cite{Fock03,Fock07}.
To start, let us consider
\begin{align}
\mathrm{Ad}(\Psi_{d_k,\mathbf{z}_k,q_k}(Y_k^{\varepsilon}))^{\varepsilon}(Y_i)
:=&\
\Psi_{d_k,\mathbf{z}_k,q_k}(Y_k^{\varepsilon})^{\varepsilon} Y_i
\Psi_{d_k,\mathbf{z}_k,q_k}(Y_k^{\varepsilon})^{-\varepsilon},
\end{align}
which we call the {\em adjoint action\/} of
$\Psi_{d_k,\mathbf{z}_k,q_k}(Y_k^{\varepsilon})$
on quantum $Y$-variables.
The following is the key formula which connects
the generalized mutation of quantum $Y$-seeds
and the quantum dilogarithms of higher degrees.
\begin{lem}
\label{lem:ad1}
\begin{align}
\label{eq:ad1}
\begin{split}
\mathrm{Ad}(\Psi_{d_k,\mathbf{z}_k,q_k}(Y_k^{\varepsilon}))^{\varepsilon}(Y_i)
=Y_i
\prod_{m=1}^{|b_{ki}|}
\left(
\sum_{s=0}^{d_k}
z_{k,s}q_k^{-\varepsilon \mathrm{sgn}(b_{ki}) (2m-1) s}
Y_k^{\varepsilon s}
\right)^{-\mathrm{sgn}(b_{ki})}.
\end{split}
\end{align}
\end{lem}
\begin{proof}
For example, in the case $\varepsilon=+$,
\begin{align}
\begin{split}
\Psi_{d_k,\mathbf{z}_k,q_k}(Y_k) Y_i
\Psi_{d_k,\mathbf{z}_k,q_k}(Y_k)^{-1}
&=
Y_i
\Psi_{d_k,\mathbf{z}_k,q_k}(q_k^{-2 b_{ki}}Y_k)
\Psi_{d_k,\mathbf{z}_k,q_k}(Y_k)^{-1}\\
&=
Y_i
\prod_{m=1}^{|b_{ki}|}
\left(
\sum_{s=0}^{d_k}
z_{k,s}
q_k^{- \mathrm{sgn}(b_{ki}) (2m-1) s}
Y_k^{ s}
\right)^{-\mathrm{sgn}(b_{ki}) },
\end{split}
\end{align}
where we used
\eqref{eq:Ycom1} and
Proposition \ref{prop:rec1} in the first and second equalities,
respectively.
The case $\varepsilon=-$ can be shown in the same way.
\end{proof}
The right hand side of \eqref{eq:ad1}, excluding the factor $Y_i$,
is a part of
\eqref{eq:Ymut1},
and for $d_k=1$ it is called the ``automorphism part" of \eqref{eq:Ymut1}
in \cite{Fock03,Fock07}.
Next let us consider the ``monomial part" of
\eqref{eq:Ymut1}.
Let us set
\begin{align}
\label{eq:Ymut2}
Z^{(\varepsilon)}_{i}:=
\begin{cases}
Y_k^{-1} & i = k\\
\displaystyle
q_i^{b_{ik} d_k [\varepsilon b_{ki}]_+}
Y_i Y_k^{d_k[\varepsilon b_{ki}]_+}& i \neq k.
\end{cases}
\end{align}
By Lemma \ref{lem:ad1},
the $(\mathbf{d},\mathbf{z})$-mutation \eqref{eq:Ymut1} is
expressed as the composition
\begin{align}
\label{eq:YZ1}
Y'_i=\mathrm{Ad}(\Psi_{d_k,\mathbf{z}_k,q_k}(Y_k^{\varepsilon}))^{\varepsilon}(Z_i^{(\varepsilon)}).
\end{align}
\begin{lem}
\label{lem:Z1}
The following commutation relation holds.
\begin{align}
\label{eq:Zcom2}
Z^{(\varepsilon)}_i Z^{(\varepsilon)}_j = q^{2r_jd_j b'_{ji}} Z^{(\varepsilon)}_j Z^{(\varepsilon)}_i,
\end{align}
where $b'_{ij}$ is given by \eqref{eq:bmut1}.
\end{lem}
\begin{proof}
This is easily verified by the case check.
\end{proof}
\begin{prop}
\label{prop:com1}
Under the $(\mathbf{d},\mathbf{z})$-mutation in \eqref{eq:Ymut1},
the following commutation relation holds:
\begin{align}
\label{eq:Ycom2}
Y'_i Y'_j = q^{2r_jd_j b'_{ji}} Y'_j Y'_i.
\end{align}
\end{prop}
\begin{proof}
By Lemma \ref{lem:Z1} and
\eqref{eq:YZ1},
we have
\begin{align}
\begin{split}
Y'_i Y'_j &=
\mathrm{Ad}(\Psi_{d_k,\mathbf{z}_k,q_k}(Y_k^{\varepsilon}))^{\varepsilon}(Z_i^{(\varepsilon)}Z_j^{(\varepsilon)})\\
&=
\mathrm{Ad}(\Psi_{d_k,\mathbf{z}_k,q_k}(Y_k^{\varepsilon}))^{\varepsilon}
(
q^{2r_jd_j b'_{ji}} Z^{(\varepsilon)}_j Z^{(\varepsilon)}_i)\\
&=
q^{2r_jd_j b'_{ji}} Y'_j Y'_i.
\end{split}
\end{align}
\end{proof}
Lemmas \ref{lem:ad1}, \ref{lem:Z1},
and Proposition \ref{prop:com1}
naturally extend
the fundamental properties
of the mutation of quantum $Y$-seeds in \cite{Fock03,Fock07}.
\section{Quantum dilogarithm identities of higher degrees}
Let us give an application of generalized mutations of quantum $Y$-seeds
to quantum dilogarithm identities of higher degrees.
Since they are parallel to the one for ordinary quantum dilogarithm identities
studied in \cite{Keller11,Kashaev11},
we only give the minimal description here.
We ask the reader to consult \cite[Section 3]{Kashaev11} for more details.
Consider a sequence of $(\mathbf{d},\mathbf{z})$-mutations of quantum $Y$-seeds,
\begin{align}
\label{eq:seq1}
(B(1),Y(1))
{\buildrel \mu_{k_1} \over \leftrightarrow }
(B(2),Y(2))
{\buildrel \mu_{k_2} \over \leftrightarrow }
\cdots
{\buildrel \mu_{k_L} \over \leftrightarrow }
(B(L+1),Y(L+1)),
\end{align}
and suppose that it has the periodicity
\begin{align}
\label{eq:period1}
b_{\sigma(i)\sigma(j)}(L+1)
=
b_{ij}(1),
\quad
Y_{\sigma(i)}(L+1)=
Y_{i}(1)
\end{align}
for some permutation $\sigma$ of $1,\dots,n$.
Then, we have the associated sequence of $(\mathbf{d},\mathbf{z})$-mutations of
(nonquantum) $Y$-seeds
of a (nonquantum) generalized cluster algebra,
\begin{align}
\label{eq:seq2}
(B(1),y(1))
{\buildrel \mu_{k_1} \over \leftrightarrow }
(B(2),y(2))
{\buildrel \mu_{k_2} \over \leftrightarrow }
\cdots
{\buildrel \mu_{k_L} \over \leftrightarrow }
(B(L+1),y(L+1)),
\end{align}
and it has the same periodicity
\begin{align}
\label{eq:period2}
y_{\sigma(i)}(L+1)=
y_{i}(1).
\end{align}
Let us further assume the {\em sign-coherence property\/} of the sequence
\eqref{eq:seq2} (see, .e.g., \cite{Nakanishi14a}).
Let $\varepsilon_t$ and $c_t$ ($t=1,\dots,L$) be the {\em tropical sign\/}
and the {\em $c$-vector\/} of $y_{k_t}(t)$ defined in \cite{Nakanishi14a}.
Let us denote the initial seed $(B(1),Y(1))$ as $(B,Y)$.
Let $\mathbb{T}(B)$ be the {\em quantum torus\/}
generated by noncommutative variables $Y^{\alpha}$ ($\alpha\in \mathbb{Z}^n$)
with the relations
\begin{align}
q^{\langle \alpha, \beta\rangle} Y^{\alpha} Y^{\beta}
=Y^{\alpha+\beta},
\quad
\langle \alpha, \beta\rangle
= \sum_{i,j=1}^n
\alpha_i d_i b_{ij} \beta_j.
\end{align}
We identify $Y_i = Y^{e_i}$, where $e_i$ is the $i$th unit vector.
\begin{thm}
Under the assumption of the periodicity \eqref{eq:period1}
and the sign-coherence property
of the sequence \eqref{eq:seq2},
we have the following identities of
the quantum dilogarithms of higher degrees
associated to the sequence \eqref{eq:seq1}.
\par
(i) Quantum dilogarithm identities in tropical form
(cf.~\cite[Theorem 3.5]{Kashaev11}).
\begin{align}
\label{eq:id1}
\Psi_{d_{k_1},\mathbf{z}_{k_1},q_{k_1}}(Y^{\varepsilon_1c_1})^{\varepsilon_1}
\cdots
\Psi_{d_{k_L},\mathbf{z}_{k_L},q_{k_L}}(Y^{\varepsilon_1c_L})^{\varepsilon_L}
=1,
\end{align}
where $Y^{\varepsilon_tc_t}\in \mathbb{T}(B)$.
\par
(ii) Quantum dilogarithm identities in universal form
(cf.~\cite[Corollary 3.7]{Kashaev11}).
\begin{align}
\label{eq:id2}
\Psi_{d_{k_L},\mathbf{z}_{k_L},q_{k_L}}(Y_{k_L}(L))^{\varepsilon_L}
\cdots
\Psi_{d_{k_1},\mathbf{z}_{k_1},q_{k_1}}(Y_{k_1}(1))^{\varepsilon_1}
=1.
\end{align}
\end{thm}
We omit the proof,
since it is completely parallel to
the one for Theorem 3.5 and Corollary 3.7 of \cite{Kashaev11}.
\begin{ex}
Let us consider the simplest nontrivial example of a generalized cluster algebra
with
\begin{align}
B=\begin{pmatrix}
0 & -1\\
1 & 0\\
\end{pmatrix},
\quad
\mathbf{d}=(2,1),
\quad
\mathbf{z}=(z_{1,1}).
\end{align}
This example was studied in \cite[Section 2.3]{Nakanishi14a}
for the nonquantum case.
Now let us choose
\begin{align}
\mathbf{r}=(1,2).
\end{align}
Thus, we have $q_1=q_2=q^2$, and
the commutation relation for $Y=(Y_1,Y_2)$ is given by
\begin{align}
Y_1Y_2=q^4 Y_2Y_1.
\end{align}
Let us set $(B(1),Y(1)):=(B,Y)$ and consider the following sequence of mutations
\begin{align}
\label{eq:seq3}
\begin{split}
(B(1),Y(1))\
{\buildrel \mu_{1} \over \leftrightarrow }\
(B(2),Y(2))\
&{\buildrel \mu_{2} \over \leftrightarrow }\
(B(3),Y(3))\
{\buildrel \mu_{1} \over \leftrightarrow }
(B(4),Y(4))\\
{\buildrel \mu_{2} \over \leftrightarrow }\
(B(5),Y(5))\
&{\buildrel \mu_{1} \over \leftrightarrow }\
(B(6),Y(6))\
{\buildrel \mu_{2} \over \leftrightarrow }\
(B(7),Y(7)).
\end{split}
\end{align}
Then, we have
\begin{align}
B(t)=(-1)^{t+1} B,
\end{align}
and the quantum $Y$-variables mutate as follows,
where we set $z=z_{1,1}$ for simplicity.
(If we set $q=1$, we recover the result in
\cite[Table 1]{Nakanishi14a} for the nonquantum case.)
\begin{align}
&
\begin{cases}
Y_1(1)=Y_1\\
Y_2(1)=Y_2,\\
\end{cases}
\\
&
\begin{cases}
Y_1(2)=Y_1^{-1}\\
Y_2(2)=Y_2(1+zq^2 Y_1 + q^4 Y_1^2),\\
\end{cases}
\\
&
\begin{cases}
Y_1(3)=Y_1^{-1}(1+q^2Y_2 + zY_1Y_2 + q^{-2}Y_1^2 Y_2)\\
Y_2(3)=Y_2^{-1}(1+zq^{-2} Y_1 + q^{-4} Y_1^2)^{-1},\\
\end{cases}
\\
\allowbreak
&
\begin{cases}
Y_1(4)=Y_1(1+q^{-2}Y_2 + zq^{-4}Y_1Y_2 + q^{-6}Y_1^2 Y_2)^{-1}\\
Y_2(4)=q^{-4}Y_1^{-2}Y_2^{-1}
(1+q^2 Y_2 + q^6 Y_2 + q^8 Y_2^2\\
\qquad\qquad\qquad
+z Y_1Y_2 +zq^2 Y_1 Y_2^2
+q^{-4} Y_1^2 Y_2^2
),\\
\end{cases}
\\
\allowbreak
&
\begin{cases}
Y_1(5)=q^{-2}Y_1^{-1}Y_2^{-1}
(1+q^{2} Y_2)\\
Y_2(5)=q^{-4}Y_1^2 Y_2
(1+q^{-6} Y_2 + q^{-2} Y_2 + q^{-8} Y_2^2\\
\qquad\qquad\qquad
+z q^{-4}Y_1Y_2 +zq^{-10} Y_1 Y_2^2
+q^{-10} Y_1^2 Y_2^2
)^{-1},\\
\end{cases}
\\
\allowbreak
&
\begin{cases}
Y_1(6)=q^{-2}Y_1Y_2
(1+q^{-2} Y_2)\\
Y_2(6)=Y_2^{-1},\\
\end{cases}
\\
\allowbreak
&
\begin{cases}
Y_1(7)=Y_1\\
Y_2(7)=Y_2.\\
\end{cases}
\end{align}
Among them, the calculation of $Y_2(4)$ is the most tedious one.
Now we observe the periodicity of the sequence
\eqref{eq:seq3} with $\sigma=\mathrm{id}$.
We also have the following data of the tropical signs and
the $c$-vectors in \cite[Section 3.4]{Nakanishi14a}
\begin{align}
&\varepsilon_1=\varepsilon_2=+,
\quad
\varepsilon_3=\varepsilon_4=\varepsilon_5=\varepsilon_6=-,\\
\begin{split}
&c_1=(1,0),\
c_2=(0,1),\
c_3=(-1,0),\\
&c_4=(-2,-1),\
c_5=(-1,-1),\
c_6=(0,-1),
\end{split}
\end{align}
which can be also read off from the above result by setting $q=1$.
Now, by substituting these data, the identity \eqref{eq:id1} reads
\begin{align}
\label{eq:id3}
\begin{split}
&\Psi_{2,(z),q^2}(Y_1)
\Psi_{q^2}(Y_2)
\Psi_{2,(z),q^2}(Y_1)^{-1}\\
&\qquad
\times
\Psi_{q^2}(q^{-4}Y_1^2Y_2)^{-1}
\Psi_{2,(z),q^2}(q^{-2}Y_1Y_2)^{-1}
\Psi_{q^2}(Y_2)^{-1}
=1,
\end{split}
\end{align}
while the identity \eqref{eq:id2} reads
\begin{align}
\label{eq:id4}
\begin{split}
&
\Psi_{q^2}(Y_2)^{-1}
\Psi_{2,(z),q^2}((1+q^2Y_2)^{-1}q^2Y_2Y_1)^{-1}
\\
&
\times\Psi_{q^2}(
(1+q^2 Y_2 + q^6 Y_2 + q^8 Y_2^2
+z Y_1Y_2 +zq^2 Y_1 Y_2^2
+q^{-4} Y_1^2 Y_2^2
)^{-1}
q^{4} Y_2Y_1^2)^{-1}
\\
&
\times
\Psi_{2,(z),q^2}((1+q^2Y_2 +zY_1Y_2 + q^{-2}Y_1^2Y_2)^{-1}Y_1)^{-1}
\\
&
\times
\Psi_{q^2}(Y_2(1+zq^2Y_1+q^4Y_1^2))
\Psi_{2,(z),q^2}(Y_1)
=1.
\end{split}
\end{align}
\end{ex}
In general, we conjecture that if the underlying nonquantum sequence
\eqref{eq:seq2} has a periodicity \eqref{eq:period2},
then the the corresponding quantum sequence
\eqref{eq:seq1} also has the same periodicity \eqref{eq:period1}.
(The converse is trivial as already stated.)
This was proved for the ordinary cluster algebras in
\cite[Proposition 3.4]{Kashaev11} when $B=B(1)$ is
skew-symmetric.
|
\section{Introduction}
Brownian motion, as a theory for the dynamics of a particle immersed in a resting fluid, has played a substantial role in the development of the statistical mechanics of diverse fluctuation phenomena both in and out of equilibrium situations \cite{Zwanzig,HM}. Even though a Brownian particle rests on average, it explores with time ever larger parts of space if not hindered by confining walls or other spatial constrifctions. This spreading can be quantified in terms of the particle's mean square displacement $\langle (\delta X(t))^2 \rangle$ which, in a resting, three-dimensional fluid at thermal equilibrium, grows at large times $t$ proportionally to $t$. In this case of so-called {\it normal} diffusion, the diffusion constant $D$ defined as the proportionality factor in the spreading law $\langle (\delta X(t))^2 \rangle = D t $ is related to the friction coefficient $\gamma$ experienced by the Brownian particle in the fluid and the temperature of the fluid by the Einstein relation \cite{Einstein}
\begin{equation}
D=\frac{2 k_B T}{M\gamma}
\label{ERnd}
\end{equation}
This relation actually goes back to Sutherland \cite{Sutherland}.
Other forms of diffusion laws known as {\it anomalous} diffusion may occur in the presence of spatial constriction, or in non-equilibrium situations \cite{BG,MK}. Often, the spreading may still be described by a power-law in time with an exponent differing from one, $\langle (\delta X(t))^2 \rangle = D_\alpha t^\alpha$. For example, in so-called single file diffusion, the location of a hard sphere in a one-dimensional row of other equal impenetrable spheres spreads with the exponent $\alpha=1/2$ \cite{AP}. In extreme contrast to this dwindling dispersion of single file diffusion, the distance $R$ of a pair of particles which are advected in a turbulent fluid in three dimensions spreads according to Richardson's law, $\langle R^2(t) \rangle \propto t^3 $, \cite{R,B}, with the exponent $\alpha=3$.
In a two-dimensional fluid at equilibrium, mode-coupling theory predicts the diffusive motion of a Brownian particle as well as the self-diffusion of fluid particles to follow a logarithmic correction to the algebraic behavior of the form $\langle (\delta X(t))^2 \rangle \propto t \ln t$ \cite{AW}. A selfconsistent mode-coupling theory yields a slightly weaker increase given by $\langle (\delta X(t))^2 \rangle \propto t \sqrt{\ln} t$ \cite{K,WAG}.
For the motion of a Brownian particle, and also of a tagged fluid particle a formally exact description exists in terms of a generalized Langevin equation, which constitutes a linear integro-differential equation for the considered particle's momentum. Given the molecular interactions, the memory kernel and the fluctuating force which are the ingredients of the generalized Langevin equation, can formally be derived by using the projection operator techniques of Zwanzig \cite{Zwanzig2} and Mori \cite{Mori}. The nature of the resulting motion of the Brownian particle is determined by the behavior of the integral of the memory kernel.
The motion is diffusive at large times if the time-integral of the memory kernel converges in the upper limit to a finite, non-vanishing value, called the static friction. It becomes super-diffusive, i.e., roughly speaking, the mean square displacement grows with an exponent $\alpha >1$, if this integral vanishes, and the motion is sub-diffusive ($\alpha <1$) if the static friction diverges.
These properties follow from the mutual dependence of the memory kernel and the momentum autocorrelation function at large times as first found by Corngold \cite{Corngold}.
Here we demonstrate that, at large times, properly defined time-dependent friction and diffusion coefficients stay in a reciprocal relation to each other. Their product is determined by a generalized asymptotic Einstein relation which, for normal diffusion, agrees with its well-known form (\ref{ERnd}). Most notably, the generalized relation allows one to determine the scaling exponent $\alpha$ of the particle's mean square displacement. This is the main difference between the present, generalized asymptotic Einstein relation and previous forms of generalized Einstein relations \cite{BG,BF} which are not restricted to the asymptotic large-time regime but do not explicitly contain the scaling exponent.
Unfortunately, the exact molecular expressions for the memory kernel and the fluctuating force are extremely involved and most often cannot be analytically determined other than in limiting cases. Molecular dynamics (MD) simulations
provide an alternative, convenient means to study the Brownian motion as well as the motion of a tagged fluid particle. As an example, we investigate a fluid in two spatial dimensions modeled by $N$ soft spheres interacting pairwise via purely repulsive Lennard-Jones potentials.
The paper is organized as follows. In Section II, the microscopic model
and the parameters for the molecular dynamics simulation are specified. After a short review of the generalized Langevin equation, the generalized asymptotic Einstein relation is derived from the generalized Langevin equation by means of the Tauberian theorem in Section III. For the sake of completeness and also for comparison, we present the proof of the already mentioned generalization of the Einstein relation \cite{BG,BF} at the end of the same Section. In Section IV the scaling behavior of the time-dependent diffusion and friction coefficients are determined in the framework of the molecular dynamics simulations. The scaling exponents are found to be compatible with each other in the sense that they concordantly imply the same scaling exponent $\alpha$ for the mean square displacement. We also find good agreement with the value of $\alpha$ following from the generalized asymptotic Einstein relation. The slight difference most likely can be attributed to the presence of slowly varying functions which may modify the apparent algebraic scaling behavior. Because from the available data the scaling behavior of the diffusion and friction coefficients is visible only for relatively short times it is impossible to identify and separate the contribution of such a possibly existing, slowly varying function to the apparent scaling. In the generalized asymptotic Einstein relation the slowly varying contributions of the diffusion and friction coefficients compensate each other. This effect leads to a more reliable estimate of the exponent $\alpha$. The paper closes with concluding remarks and a discussion of the present findings to previously published results in Section V.
\section{Microscopic model and Molecular dynamics simulation method}\label{MD}
We considered a standard model of a Brownian particle suspended in a two-dimensional fluid \cite{AT}. It consists of a single probe particle of mass $M$ and diameter ${\sigma}_\textrm{BB}$ and $N$ solvent particles of mass $m$ and diameter ${\sigma}_{\textrm{SS}}$ enclosed in a two-dimensional domain of side-length $L_{x}$ and $L_{y}$ with periodic boundary conditions.
Assuming that the system is isolated, the motion of particles follows the Hamiltonian dynamics with the Hamiltonian given as
\begin{equation}\label{Hamiltonian}
{\mathcal H} = \sum_i \frac{1}{2m}\left|{\mathbf p}_i \right|^2 + \frac{1}{2M}\left|{\mathbf P}\right|^2 + U \left({\mathbf r}^N, {\mathbf R} \right)\:,
\end{equation}
where ${\mathbf r}_i$ and ${\mathbf p}_i$ are the position and the momentum of the $i$th solvent particle. Accordingly, ${\mathbf R}$ and ${\mathbf P}$ refer
to the probe particle, and $U({\mathbf r}^N,{\mathbf R})$ is the potential energy describing the mutual interaction of the solvent particles among each other and with the Brownian particle.
The total linear momentum of the system as well as its energy are conserved. The angular momentum is not conserved due to periodic boundary condition. This assumption applies to the standard MD model and generates the NVEp ensemble for which the number of particles, volume, total energy and total linear momentum are fixed.
The potential energy is determined by pairwise interactions and hence given by
\begin{equation}
U(\mathbf{r}^N,\mathbf{R}) = \sum_{i>j} u_{SS}(|\mathbf{r}_i -\mathbf{r}_j|) + \sum_i u_{BS}(|\mathbf{R}-\mathbf{r}_i|)\;.
\label{Uuu}
\end{equation}
For the pair-potentials $u_{SS}(r)$, and $u_{BS}(r)$, we used
purely repulsive Lennard-Jones potentials \cite{Weeks} of the form
\begin{equation}\label{WCA_ER2D}
u_{\alpha \beta}(r)=\left\{ \begin{array}{ll}
4\epsilon \left[\left({\sigma}_{\alpha \beta}/r\right)^{12} - \left({\sigma}_{\alpha \beta}/r\right)^{6} \right] + \epsilon & \mbox{for $r < 2^{1/6} {\sigma}_{\alpha \beta}$} \\
0 & \mbox{for $r \geq 2^{1/6} {\sigma}_{\alpha \beta}$}. \end{array}
\right.
\end{equation}
where $\alpha$ and $\beta$ stands for either $S$ or $B$ referring to the fluid and Brownian particles, respectively. The length-scale ${\sigma}_\textrm{SB}=\left({\sigma}_{\textrm{BB}}+{\sigma}_{\textrm{SS}}\right)/2$ determines the range of the repulsive fluid-Brownian-particle interaction.
In the MD simulations, dimensionless units were used. The lengths are measured in units of $\sigma_{SB} $.
For the number density $n$ of fluid particles we took $n \sigma_{SB}^2 =0.3,\; 0.4,\;0.5,\;0.8$.
With the choice of the fluid particle size, the diameter $\sigma_{BB}$ of the Brownian particle is also determined.
The total number $N$ of fluid particles was taken as $N = 32\:000$ in the majority of cases. In order to identify finite size effects, simulations with less particles were performed at constant number density by taking $N=160,\; 1\:000,\;4\:000,\;16\:000$. The configuration space consisted of a rectangle glued together at opposite sides. The side lengths were chosen as $L_x= \sqrt{n/N} (N/n + \pi \sigma^2_{BB}/4)$ and $L_y=\sqrt{N/n}$. Note that, for the considered numbers of particles, the deviation from a square is only minor.
All masses were measured in units of the fluid particle mass $m$, and energies in units of the strength $\epsilon$ of the pair interactions at the distance $r=\sigma_{\alpha \beta}$.
A consistent unit of time then follows as $\tau \equiv \sigma_{SB} \sqrt{m/\epsilon}$.
As initial conditions of the MD simulations the centers of the fluid particles were put on the points of a simple cubic (100) surface and the Brownian particle was positioned at the center of one of the squares formed by four neighboring fluid particles. For each solvent particle a vector $\sqrt{k_B T^{\text{set}}/m}\: \mathbf{e}$ was chosen, where $k_B$ denotes the Boltzmann constant and $T^{\text{set}}$ the target temperature. The two-dimensional vector $\mathbf{e}$ was taken randomly with independent, Gaussian distributed Cartesian components having vanishing averages and unit variances. The algebraic average over all these $N$ vectors was subtracted from the individual vectors. The resulting vectors were used as initial velocities of the fluid particles. The initial velocity of the Brownian particle was put to zero.
By construction, the total momentum of the system consisting of fluid and Brownian particles vanished.
The values of target temperature $T^{\text{set}}$ were taken as $0.2 \epsilon / k_B$, $0.5 \epsilon / k_B$
and $ 1.0 \epsilon / k_B$.
The equations of motion for each individual particle were solved using the velocity Verlet algorithm with a time step $\Delta t = 0.001\tau $.
At the beginning of the simulation, a velocity rescaling was performed every $10,000$ steps to adjust the energy of the system to the desired value (all together $10$ times).
We found that the thermal equilibrium was established within $2\times 10^{6}$ time steps.
We used the data resulting from the subsequent $1 \times 10^{7}$ time steps for the construction of the time correlation functions of the relevant dynamical variables and the GLE.
If necessary, $20$ simulations were performed for identical thermodynamic parameters to obtain ensemble-averaged quantities.
\section{Generalized Mori-Zwanzig Langevin equation for a probe particle in a 2D fluid}\label{GLE}
\subsection{Generalized Langevin equation}
According to the Mori-Zwanzig theory \cite{Zwanzig} the dynamics of the probe particle can be represented as a generalized Langevin equation for the momentum ${\mathbf P}(t)$ reading
\begin{equation}\label{GLE_tensor}
\frac{d}{dt}{\mathbf P}(t) = {\mathbf \Omega}\cdot {\mathbf P}(t)-\int^t_0{{\mathbf k}(t^\prime)\cdot {\mathbf P}\left(t-t^\prime\right) dt^\prime }+{\mathbf F}^\dagger(t)\:.
\end{equation}
The first term on the right hand side describes the reversible, instantaneous change of the two-dimensional momentum $\mathbf{P}(t)$ with Cartesian components $P_x(t)$ and $P_y(t)$. It is determined by the frequency matrix
$ \mathbf{\Omega}$, which is defined as
\begin{equation}\label{omega_GLE}
{\mathbf \Omega } \equiv \left({\mathbf P}, L{\mathbf P}\right)\cdot {\left({\mathbf P},{\mathbf P}\right)}^{-1}\:,
\end{equation}
where $\left (f,g \right ) =$ \\
$ \int d \Gamma f(\Gamma) g(\Gamma) e^{- \mathcal{H}(\Gamma)/( k T^{\text{set}})} /\int d \Gamma e^{- \mathcal{H}(\Gamma)/( k T^{\text{set}})}$ denotes the Mori scalar product of phase-space functions $f(\Gamma)$ and $g(\Gamma)$ with $\Gamma$ being a point in phase-space and $d\Gamma= d^N \mathbf{p}d^N \mathbf{r} d \mathbf{P} d \mathbf{R}$ the corresponding infinitesimal phase-space volume element. This scalar product yields the equilibrium correlation of its arguments.
The Liouville operator is defined as $ L f= \{H,f \} $, where $\{f,g\}$ denotes the Poisson bracket of the phase-space functions $f$ and $g$. Because the weight used in the Mori scalar product is stationary, the frequency matrix $\Omega$ vanishes.
A projection operator $\mathcal{P}$ is conveniently defined in terms of the Mori scalar product assigning to each phase-space function $f(\Gamma)$ the component parallel to $\mathbf{P}$ as
\begin{equation}
\mathcal{P} f = \mathbf{P} \cdot \left (\mathbf{P}, f \right )\:,
\label{PP}
\end{equation}
where the dot indicates the scalar product of two-dimensional vectors $\mathbf{U} \cdot \mathbf{V} = U_x V_x +U_y V_y$.
The fluctuating force $\mathbf{F}^\dagger(t)$ can then be expressed as
\begin{equation}\label{fluc_GLE}
{\mathbf F}^\dagger(t) \equiv \left({\mathbf 1}-{\mathcal P}\right) \exp{\left\lbrace \left({\mathbf 1}-{\mathcal P}\right) Lt \right\rbrace} L{\mathbf P}\:.
\end{equation}
Its mean value vanishes, $\langle \mathbf{F}^\dagger(t) \rangle =0$, where the brackets indicate a thermal equilibrium average.
The memory kernel $\mathbf{k}(t)$ is defined in terms of the fluctuating force-force correlation function as
\begin{equation}\label{kernel_GLE}
{\mathbf k}(t) \equiv \left({\mathbf F}^\dagger,{\mathbf F}^\dagger(t)\right)\cdot {\left({\mathbf P},{\mathbf P}\right)}^{-1}\:.
\end{equation}
Due to the isotropy of the present model \cite{aniso} , the memory kernel is proportional to the $2 \times 2$ unit matrix $\mathbbm{1}$, i.e. it becomes
\begin{equation}
\mathbf{k}(t) = k(t) \mathbbm{1}\:.
\label{kiso}
\end{equation}
Hence, the Cartesian components of the momentum do not couple to each other and obey the identical generalized Langevin equations of the form
\begin{equation}
\frac{d}{dt} P_\alpha(t) = -\int_0^t k(t') P_\alpha(t-t') dt' +F^\dagger_\alpha(t)\:,\quad \alpha=x,y\:.
\label{GLE_alpha}
\end{equation}
Performing the Mori scalar product with the momentum component $P_\alpha$ on both sides of this equation one obtains for the momentum autocorrelation function (MACF) $C(t) \equiv (P_x,P_x(t) ) =(P_y,P_y(t) ) $
a homogeneous integro-differential equation reading \cite{Zwanzig}
\begin{equation}\label{GLE_MACF}
\frac{d}{dt}C \left(t\right) = -\int^{t}_0{k \left(t^\prime\right) C \left(t-t^\prime \right)dt^\prime}\:.
\end{equation}
The inhomogeneity vanishes because of the orthogonality of the initial momentum and the fluctuating force, i.e., $( P_x, F^\dagger_x(t)) = ( P_y, F^\dagger_y(t)) =0$.
From the knowledge of the instantaneous
momenta of
the probe particle, we estimated the stationary momentum auto-correlation function $C(t)$. In principle, knowing $C(t)$ one can find the memory kernel by solving Eq.~(\ref{GLE_MACF}) for $k(t)$.
A straightforward formal solution is obtained by means of a Laplace transformation which yields
\begin{equation}
\hat{C}(z) = \frac{C(0)}{z+\hat{k}(z)}\:,
\label{Cz}
\end{equation}
where $\hat{C}(z)= \int_0^\infty e^{- z t} C(t)$ and $\hat{k}(z)= \int_0^\infty e^{- z t} k(t)$.
However, the inverse Laplace transform poses notorious numerical problems and therefore one has to resort to Eq.~(\ref{GLE_MACF}) in the time regime.
This amounts to the solution of a homogeneous Volterra equation of the first kind, which is also known for its numerical instability. We therefore pursued a more stable approach as described in Ref. \cite{Shin}. For this purpose Eq.~(\ref{GLE_MACF}) is differentiated with respect to time yielding
\begin{equation}\label{GLE_MACF_diff}
\ddot{C}(t) = -C(0)k(t)-\int^t_0{k(t^\prime)\dot{C} \left(t-t^\prime\right)dt^\prime}\:,
\end{equation}
which is a Volterra equation of the second kind.
The first derivative of the MACF is given by the stationary cross-correlation between the momentum and its first derivative with respect to time. Therefore it agrees with the cross-correlation of the momentum and the total force $F_\alpha(t) \equiv -\int_0^t k(t') P_\alpha(t-t') + F^\dagger_\alpha(t)$ acting on the Brownian particle yielding
\begin{equation}\label{MACF_diff_GLE}
\dot{C}(t) = \left(P_x, F_x(t) \right) =\left(P_y, F_y(t) \right)\:.
\end{equation}
Likewise, the
second derivative of the MACF is given by the negative autocorrelation function of the total force acting on the Brownian particle, such that
\begin{equation}\label{MACF_diff_diff_GLE}
\ddot{C}(t) = -\left( F_x, F_x(t) \right) = -\left( F_x, F_x(t) \right)\:.
\end{equation}
Because the total force exerted on the Brownian particle can be determined from the MD simulation, also the first and the second derivative of the MACF can be estimated from the data in terms of the correlation functions $(P_\alpha,F_\alpha(t))$ and $(F_\alpha,F_\alpha(t))$ without invoking numerical differentiation. Finally, the memory kernel can reliably be determined from Eq.~(\ref{GLE_MACF_diff}) in a numerically stable way \cite{Shin}.
\begin{figure}\begin{center}
\includegraphics[width=8.5cm]{VACF_N.pdf}
\caption{The MACF $C(t)$ of a probe particle of the same size and mass as the fluid particle is displayed as a function of time in a doubly logarithmic presentation for different system sizes, $N=160$ (brown), $N=1,000$ (magenta), $N=4,000$ (orange), $N=8,000$ (green), $N=16,000$ (blue), and $N=32,000$ (red) are plotted. Density and temperature of the fluid are $n=0.4$ and $T^{\text{set}}=1.0$, respectively. The approach to an algebraic behavior at large times becomes more pronounced with increasing system size.}
\label{VACF_N}
\end{center}\end{figure}
\begin{figure}\begin{center}
\includegraphics[width=8.5cm]{VACF_loglog_M___.pdf}
\caption{The MACF $C(t)$ of a probe particle of the same size as the other $N=16,000$ fluid particles is displayed as a function of time $t$ in a doubly logarithmic plot. The density and temperature are chosen as $n=0.4$ and $T^{\text{set}}=1.0$, respectively. Three cases with different solute masses, $M=m$ (black), $M=10 m$ (red), and $M=100m$ (green), are compared.
With increasing mass the MACF decays slower with a decay law that gradually changes from an algebraic to a more exponential-like behavior at large times.}
\label{VACF_loglog_M___}
\end{center}\end{figure}
\begin{figure}\begin{center}
\includegraphics[width=8.5cm]{kernel_N.pdf}
\caption{The memory function $k(t)$ is presented as a function of time for self-diffusion in a fluid at the density $n=0.4$ and temperature $T^{\text{set}}=0.6$ for a fluid consisting of $N=160$ (blue), $N=1\:000$ (red), $N=4\:000$ (yellow), $N=8000$ (magenta), $N=16\:000$ (green) and $N=32\:000$ (black) particles. As for the MACF the memory functions become less rugged the larger the system size becomes. Differences between the memory functions for $N=16\:000$ and $N=32\:000$ are hardly visible. The inset displays the short-time behavior of the memory function. For better visibility, the negative part is magnified by a factor of 50. A dependence of the memory kernel on the system size $N$ becomes only sensible for larger times when $k(t)$ becomes negative.}
\label{F_ktN}
\end{center}\end{figure}
\begin{figure}\begin{center}
\includegraphics[width=8.5cm]{kernel_M___.pdf}
\caption{The memory function $\langle P^2\rangle k(t)$ of the solute particle is presented for short times.
As in Fig.~\ref{VACF_loglog_M___}, three different masses with $M=m$ (black), $M=10m$ (red), and $M=100m$ (green) are displayed.
For the sake of better visibility, in the region where the memory function becomes negative its value is magnified by a factor of 10.
With increasing mass the decay gradually becomes slower in the region where the memory function is positive and, at the same time, the negative part becomes more pronounced.
This trend corresponds to the result displayed in the inset which displays the integrated memory kernel as a function of time. The larger the mass the more pronounced is the maximum of memory kernel.}
\label{kernel_M___}
\end{center}\end{figure}
Fig.~\ref{VACF_N} displays the dependence of the MACF on the system size for self-diffusion ($M=m$) in a fluid with density $n=0.4$ and $T^\textrm{set}=1.0$.
Here and in the following we use the dimensionless units as introduced in Section II. At large times for the largest system size $N=32\:000$ the MACF follows an algebraic decay approximately as $t^{-1}$ with some small bumpy deviations at the largest displayed times. These deviations increase in size and move to shorter times with decreasing $N$.
Fig.~\ref{VACF_loglog_M___} and Fig.~\ref{kernel_M___} exemplify the MACFs and corresponding memory functions of the Brownian particle at the same density and temperature as in Fig.~\ref{VACF_N}, but with different solute masses.
The memory functions were numerically determined from Eq.~(\ref{GLE_MACF_diff}) by means of the estimated momentum-total-force and total-force-total-force correlation functions (\ref{MACF_diff_GLE}) and (\ref{MACF_diff_diff_GLE}), respectively. The obtained memory functions are characterized by a rapidly decaying peak at short times, then cross the zero line and only slowly re-approach zero from the negative side.
Fig.~\ref{F_ktN} exemplifies this behavior for self-diffusion ($M=m$, $\sigma_{BB}=\sigma_{SS}$) at the density $n=0.4$, and temperature $T^{\text{set}}=1.0$ for different system sizes. It is interesting to note that the positive short-time part of the memory kernel is virtually unaffected by the system size. This positive part can be understood as resulting from sequences of independent binary collisions between the tagged particle and the fluid particles. Hence it does not contain finite size effects. At larger times when the memory kernel becomes negative correlations between collisions may build up in a way depending on the system size.
The larger the system size the smoother the memory kernel becomes.
Fig.~\ref{kernel_M___} demonstrates the behavior of the memory kernel for different masses at the same density and temperature for $N=16\:000$. The inset of this figure displays the integrated memory kernels, or friction kernels as we will denote them below. They continue to decrease until the largest times without reaching a plateau value possibly indicating sub-diffusive behavior.
\subsection{Generalized Einstein Relation}
The position of a Brownian particle in an isotropic medium typically spreads at large times according to an algebraic law imposing for the variance $\langle (\delta X(t))^2 \rangle$ a growing as
\begin{equation}
\langle (\delta X(t))^2 \rangle \sim D_\alpha t^\alpha \quad \text{for}\; t \to \infty\:,
\label{Dalpha}
\end{equation}
where $\delta X(t) = X(t) -X(0)$ is the fluctuation of a component of the Brownian particle position $\mathbf{R}(t)$, and $\alpha>0$ is a scaling exponent. Possibly existing ``slow'' corrections to the algebraic law will be considered later. In the simplest case described by Eq.~(\ref{Dalpha}) the diffusion is {\it normal} for $\alpha =1$, {\it sub-diffusive} for $\alpha < 1$ and {\it super-diffusive} for $\alpha>1$. By the $\sim$ symbol we indicate the asymptotic equality of two functions $f(t)$ and $g(t)$, i.e. $f(t) \sim g(t)$ for $t \to T$ if $\lim_{t \to T} f(t)/g(t) =1$.
Because of $\delta X(t) = M^{-1} \int_0^t ds P(s)$, where $P(t)$ is the same component of the momentum vector $\mathbf{P}(t)$ as $X(t)$ of the position vector $\mathbf{R}(t)$, the variance of the position is related to the momentum autocorrelation by $\langle (\delta{X}(t))^2 \rangle = M^{-2} \int_0^t ds_1 \int_0^t ds_2 C(s_1-s_2)$. Introducing the time-dependent {\it diffusion coefficient} $D(t)$ as the time-derivative of the position variance we obtain
\begin{equation}
\begin{split}
D(t) & \equiv \frac{d}{dt} \langle (\delta X(t))^2 \rangle \\
& = \frac{2}{M^2} \int_0^t ds\: C(s)\:.
\end{split}
\label{Dt}
\end{equation}
The Laplace transform of the diffusion coefficient, $\hat{D}(z) = \int_0^\infty D(t) e^{-zt}$, is consequently related to the Laplace transform of the momentum autocorrelation by
\begin{equation}
\hat{D}(z) = \frac{2}{M^2 z} \hat{C}(z)\:.
\label{Dz}
\end{equation}
Introducing the time-dependent {\it friction coefficient}
\begin{equation}
\gamma(t) \equiv \int_0^t ds \: k(s)\:,
\label{gt}
\end{equation}
and, accordingly, its Laplace transform
\begin{equation}
\begin{split}
\hat{\gamma}(z) &\equiv \int_0^\infty \gamma(t) e^{-z t}\\
&= \frac{1}{z} \hat{k}(z)\:,
\end{split}
\label{gz}
\end{equation}
we obtain from Eq.~(\ref{Cz})
\begin{equation}
z^2 \hat{D}(z)(1+\hat{\gamma}(z)) = 2 C(0)/M^2\:.
\label{zDg}
\end{equation}
In order to elucidate the physical meaning of the time-dependent friction coefficient we consider a weak external, time-dependent force $F_e(t)$ acting on the Brownian particle. Within the linear response regime, the dynamics of the particle is then governed by the generalized Langevin equation
\begin{equation}
\dot{P}(t) = -\int_0^t ds k(t-s) P(s) +F_e(t) + F^+(t)\:.
\label{PFe}
\end{equation}
For the sake of simplicity, the direction of the force is supposed to be constant in time and only the motion in the direction of the external force is considered. Averaging over the realizations of the fluctuating force we obtain a linear equation for the mean value of the momentum governed by the memory kernel and the applied external force reading
\begin{equation}
\langle \dot{P}(t) \rangle_e = - \int_0^t ds k(t-s) \langle P(s) \rangle_e + F_e(t)\:,
\label{mPFe}
\end{equation}
where $\langle \cdot \rangle_e$ denotes an non-equilibrium average in the presence of the driving force $F_e(t)$.
From the mean value equation (\ref{mPFe}) it follows that the Brownian particle moves with constant mean momentum through the fluid, provided that the external force is proportional to the integral of the memory function, i.e. proportional to the time-dependent friction coefficient:
\begin{equation}
F_e(t) \propto \gamma(t) \quad \Longrightarrow \langle P(t) \rangle_e = \text{constant}\:.
\label{Fg}
\end{equation}
This generalizes the Archimedean law of uniform motion at constant force in a medium that causes normal diffusion to achieving a uniform motion of a particle in an environment that causes anomalous diffusion.
To further elucidate the relation between the diffusion and friction coefficients at large times we will investigate Eq.~(\ref{zDg}) for small, positive values of the Laplace variable $z$.
For the asymptotic behavior of the time-dependent diffusion coefficient we consider the slightly more general form
\begin{equation}
D(t) \sim t^{\alpha -1} L(t)
\label{DaL}
\end{equation}
than the one that immediately follows from the purely algebraic dependence by differentiating Eq.~(\ref{Dt}) with respect to time. This relation agrees in the scaling exponent $\alpha -1$ with the time-derivative of (\ref{Dt}) but, additionally, contains a slowly varying function $L(t)$, i.e. a function satisfying the asymptotic relation $\lim_{t\to \infty} L(t x)/L(t) =1$ for all positive $x$ \cite{F}. That means that $L(t)$ may grow to infinity, or decrease to zero slower than any power of $t$; typical examples of slowly growing functions are $\ln t$, $\ln(\ln t)$, $\ln^2 t$; their respective reciprocals $1/\ln t$, etc., provide examples of slowly vanishing functions.
The presence of a slowly varying function may change the normal diffusion behavior at $\alpha =1$ to super-diffusion if $L(t)$ slowly increases and to sub-diffusion if it decreases. Only if $L(t)$ has a finite, non-zero limit normal diffusion is observed. This limit then is related to the value of the normal diffusion constant $D_1$ by $\lim_{t\to \infty}L(t) = D_1 M^2/(2 C(0))$.
Because any estimate of the exponent $\alpha$ -- be it on the basis of numerically or experimentally determined data -- will be always contaminated by errors, the identification of the slowly varying function presents a difficult problem; even more, its presence also may considerably impede the estimate of the exponent $\alpha$ as it leads to slightly curved graphs rather than to a straight line with a well-defined slope in a doubly logarithmic plot of the position variance versus time.
Using the Tauberian theorem \cite{F}, see also appendix \ref{T}, we find, as an immediate consequence of the asymptotic behavior of the diffusion coefficient at large times, the corresponding behavior of the Laplace transform $\hat{D}(z)$ at small, positive $z$ to be given by
\begin{equation}
\hat{D}(z) \sim z^{-\alpha} \Gamma(\alpha) L(1/z)\quad \text{for}\: z\to 0\:,
\label{Daz}
\end{equation}
where $\Gamma(z)$ denotes the Gamma-function.
Putting this result into Eq.~(\ref{zDg}) we obtain for the Laplace transform of the friction coefficient
\begin{equation}
\hat{\gamma}(z) = \frac{2 C(0)}{\Gamma(\alpha) M^2 z^{2-\alpha}} L^{-1}(1/z) -1\quad \text{for}\: z\to 0\\:.
\label{gza}
\end{equation}
For $\alpha \geq 2$, the constant term $-1$ dominates in the limit $z\to 0$ and, hence, the time-dependent friction coefficient $\gamma(t)$ quickly converges to zero. We shall not further consider this extreme case of super-ballistic motion. For $0 < \alpha <2$ the constant term in Eq.~(\ref{gza}) can be neglected. We may then again employ the Tauberian theorem to obtain
\begin{equation}
\gamma(t) \sim \frac{2 C(0)}{M^2 \Gamma(\alpha) \Gamma(2-\alpha)} t^{1-\alpha} L^{-1}(t)\:.
\label{gta}
\end{equation}
Combining the asymptotic results for the time-dependent diffusion and friction coefficients we find for the limiting behavior of the product of these two functions a generalized asymptotic Einstein relation of the form
\begin{equation}
\lim_{t\to \infty} D(t) \gamma(t) = \frac{2 C(0)}{M^2 \Gamma(\alpha) \Gamma(2-\alpha)}\quad \text{for} \;\;0<\alpha<2\:.
\label{Dga}
\end{equation}
We want to stress that the slowly varying functions modifying the algebraic behavior of the diffusion and the friction coefficients compensate each other in the product entering the limit on the right hand side. Therefore this relation may be used as the basis of a more precise estimate of the scaling exponent $\alpha$.
The standard form of the Einstein relation is recovered for $\alpha =1$ yielding $D \gamma$ as the limit on the left hand side and $2 C(0)/M^2 = 2 k_B T^{\text{set}} /M$ on the right hand side.
Finally we note that taking the Laplace transformation on both sides of Eq.~(\ref{mPFe}) one can obtain the following formal expression for the Laplace transform of the average momentum in the presence of an external force, $\langle \hat{P}(z) \rangle_e = \int_0^\infty dt e^{-z t} \langle P(t) \rangle_e$,
\begin{equation}
\begin{split}
\langle \hat{P}(z) \rangle_e & = \frac{\hat{F}_e(z)}{z+\hat{k}(z)}\\
&= \frac{\hat{C}(z) \hat{F}_e(z)}{C(0)}\:,
\end{split}
\label{PFKz}
\end{equation}
where we denoted $\hat{F}_e(z) = \int_0^\infty dt e^{-z t} F_e(t) $ and put $\langle P(0) \rangle_e =0$. Making use of $\langle \delta \hat{X}(z) \rangle_e =\int_0^\infty dt e^{-z t} X(t) \rangle = \langle \hat{P}(z) \rangle_e/z$ in combination with Eq.~(\ref{Dz}) we obtain upon an inverse Laplace transformation
\begin{equation}
\langle \delta X(t) \rangle_e = \frac{2 M}{C(0)} \int_0^\infty ds D(s) F_e(t-s)\:,
\label{XDFt}
\end{equation}
and for a constant force $F_e(t) = F$
we find with $\int_0^t ds D(s) = \langle (\delta X (t))^2 \rangle$
\begin{equation}
\langle \delta X(t) \rangle_e = \frac{2 M}{C(0)} \langle (\delta X (t))^2 \rangle F_e\:.
\label{XXF}
\end{equation}
This equality relating the average position dynamics in the presence of a weak constant external force to the equilibrium mean square deviation of the position is also known as generalized Einstein relation \cite{BG,BF}.
It is valid for all times but, in contrast to the generalized asymptotic Einstein relation (\ref{Dga}), does not explicitly contain the scaling exponent.
\section{Results}\label{Results}
We investigated the diffusion coefficient, the friction coefficient and the generalized asymptotic Einstein relation in the particular case of self-diffusion, i.e. for $\sigma_{BB} = \sigma_{SS}$ and $M=m$, at different densities $n= 0.3,0.4, 0.5$ and 0.8 of the considered fluid. The corresponding temperatures were chosen as $T^{\text{set}} = 0.5,$ and 1.0 in the cases of the three lowest densities and $T^{\text{set}}=0.2$ in the high-density case.
In each case, the diffusion coefficient is determined from a numerical integration of the MACF according to the second line of Eq.~(\ref{Dt}). The friction coefficient is obtained from the time-integral of the memory kernel. Both, the diffusion and the friction coefficients are displayed in doubly logarithmic plots in order to make a possible scaling behavior visible. In the case of the low densities, $n=0.3,\;0.4,\;0.5$, two scaling regimes can be distinguished for the diffusion coefficient, one for short times $t<1/e$ and the other one for large times $t>e$, and similarly for the friction coefficient reaching up to $t\approx e^{-3}$ for the short-time scaling regime and above $t\approx 1/e$ for the large-time scaling regime.
The maximal time $t\approx 100$ is sufficiently small that an influence from the finite sound velocity can be excluded for systems with $N=32 000$ particles.
The left and the middle panel of Fig.~\ref{scaling_ld} display the diffusion and friction coefficient for $n=0.3$ and $T^{\text{set}}=0.5$.
\begin{figure*}[t]
\includegraphics[width=0.32\textwidth]{D-t_density=03_T=05}\nolinebreak
\includegraphics[width=0.32\textwidth]{Gam-t_density=03_T=05}\nolinebreak
\includegraphics[width=0.32\textwidth]{D-t_Gam-t_density=03_T=05}
\caption{Diffusion (left), friction (middle) and Einstein (right panel) coefficients multiplied by $M^2/(2 C(0))$ for self-diffusion ($M=m$, $\sigma_BB=\sigma_SS$) in a fluid consisting of $N=32\:000$ particles at density $n=0.3$ and at temperature $T^{\text{set}}=0.5$ as functions of time. The red solid curves in the left and middle panel represent the diffusion and friction coefficients, respectively, determined from the molecular dynamics simulations. The scaling behavior of the diffusion coefficient is characterized at short times by $\alpha=2$ and corresponds to ballistic motion. The black broken straight lines in the left and middle panels are least square fits to the large time behavior of the diffusion and friction coefficients. From their respective slopes $s_D$ and $s_\gamma$ the estimates $\alpha_D=1+s_D$ and $\alpha_\gamma=1-s_\gamma$ of the scaling exponent can be obtained. For the results see Table~\ref{t1}. The red curve in the right panel again represents the scaled Einstein coefficient obtained from the molecular dynamics simulation. The green broken line displays the asymptotic value reached at large times. For comparison, the blue and the green broken lines, which almost fall on top of each other, indicate the values of the factor $1/(\Gamma(\alpha) \Gamma(2-\alpha))$ entering the generalized asymptotic Einstein relation (\ref{Dga}) for $\alpha=\alpha_\gamma$ and $\alpha=\alpha_D$, respectively. The results for the other two densities $n=0.4$ and $n=0.5$ as well as those for the higher temperature $T=1.0$ are qualitatively similar and therefore not shown.}
\label{scaling_ld}
\end{figure*}
The large time behavior of these coefficients can be described by (\ref{DaL}) and (\ref{gta}) with
scaling exponents which almost agree with each other, see Table~\ref{t1}.
\begin{table}
\begin{tabular}[t]{|c|c||c|c|c|}
\hline
$T$& $n$&$\alpha^D$&$\alpha^\gamma$&$\alpha^{\text{ER}}$\\
\hline
\hline
0.5&0.3&1.073&1.070&1.092\\\hline
0.5&0.4&1.090&1.092&1.114\\\hline
0.5&0.5&1.10&1.098&1.118\\\hline
1.0&0.3&1.069&1.068&1.092\\\hline
1.0&0.4&1.087&1.085&1.096\\\hline
1.0&0.5&1.099&1.101&1.158\\\hline
0.2&0.8&1.013&1.019&1.032\\
\hline
\end{tabular}
\caption{Estimates of the scaling exponent $\alpha$ obtained from the behavior of $D(t)$, $\gamma(t)$ and the Einstein coefficient $D(t) \gamma(t)$ on the large-time range $[40,100]$ are collected for different temperatures and densities. In most cases, the difference between the values of the scaling exponents $\alpha^D$ and $\alpha^\gamma$ differ from each other only insignificantly. This indicates that, within the considered time interval $[40,100]$, both coefficients behave asymptotically according to eqs. (\ref{DaL}) and (\ref{gta}). In all cases, the exponent $\alpha^{\text{ER}}$ following from the generalized asymptotic Einstein relation is larger than the other two estimates. This suggests the presence of slowly varying functions which influence $\alpha^D$ and $\alpha^\gamma$ in the same way but cancel in the Einstein coefficient. }
\label{t1}
\end{table}
\begin{figure*}[t]
\includegraphics[width=0.32\textwidth]{D-t_density=08_T=02}\nolinebreak
\includegraphics[width=0.32\textwidth]{Gam-t_density=08_T=02}\nolinebreak
\includegraphics[width=0.32\textwidth]{D-t_Gam-t_density=08_T=02}
\caption{Diffusion (left), friction (middle) and Einstein (right panel) coefficients for self-diffusion ($M=m$) in a fluid with density $n=0.8$ and at the temperature $T^{\text{set}}=0.2$. Lines and color codes agree are the same as in Fig.~\ref{scaling_ld}. In the right most panel all three broken lines fall on top of each other.}
\label{scaling_hd}
\end{figure*}
In the regime of low densities $n=0.3,\;0.4,\;0.5$ the scaling exponent $\alpha$ is close to $1.1$ for both temperatures $T^{\text{set}}=0.5,\;1.0$. This super-diffusive scaling behavior is different from the results of mode-coupling theory according to which the momentum autocorrelation decays as slowly as $\propto t^{-1}$ for a two-dimensional fluid \cite{AW} giving rise to a diffusion coefficient which grows $\propto \ln(t)$. The self-consistent mode-coupling theory predicts the momentum auto-correlation function to decay $\propto t^{-1} (\ln(t))^{-1/2}$ \cite{K,WAG} and the diffusion coefficient to increase $\propto (\ln(t))^{1/2}$. Hence, according to both theories the scaling exponent would be $\alpha=1$ in disagreement with the present numerical findings. The actually observed algebraic law with $\alpha \approx 1.1$ most likely is still modified by a slowly varying function. The precise nature of this slowly varying function is difficult to extract from the existing numerical data but its presence is strongly suggested by the fact that the apparent scaling exponents of the diffusion and the friction coefficients agree quite well with each other, but not as accurately with the value of $\alpha$ following from the generalized asymptotic Einstein relation.
A detailed understanding of the mechanism underlying this behavior is still missing.
When going to the larger density $n=0.8$ the scaling exponent becomes approximately normal with $\alpha \approx 1$, see Fig.~\ref{scaling_hd}. The
asymptotic large time regime presumably is not yet fully reached as can be inferred from the larger difference of $\alpha_D$ and $\alpha_\gamma$ and also from the relatively large fluctuations of the Einstein coefficient at for times larger than $20$.
An extension of the considered time-range would be problematic because of the sound velocity, which at this relatively high density is larger than in the low density cases. A compensation of this effect by considering larger systems at the same density would requires to consider even larger numbers of particles which makes the simulation very time-consuming.
At even higher densities, for numerical reasons, it becomes increasingly difficult to reach the large-time asymptotic regime. Preliminary results indicate a continuing tendency of the friction coefficient to increase with increasing time possibly giving rise to sub-diffusive behavior. The Figure~\ref{Normal_diffusion_curve_n-T} roughly indicates the border region between super-and sub-diffusive motion in a density-pressure phase diagram.
\begin{figure}\begin{center}
\includegraphics[width=8cm] {Normal_diffusion_curve_n-T.pdf}
\caption{The figure displays the density($n$)-temperature($T$) plane characterizing the self-diffusive motion in 2D soft-disk fluids. The yellow narrow region III separates the super-diffusive low-density fluid behavior I from a very slow, possibly sub-diffusive behavior II at large densities. Along this boundary, the self-diffusion is normal. Normal self-diffusive motion was confirmed by numerical simulations for density-temperature values corresponding to the
indicated points within the narrow diffusive strip.}
\label{Normal_diffusion_curve_n-T}
\end{center}\end{figure}
\section{Conclusions}
Based on Mori's generalized Langevin equation for the momentum of a Brownian particle in combination with the assumption that its mean square displacement is of regular variation as a function of time, we derived a generalized asymptotic Einstein relation. It connects the time-dependent diffusion and the friction coefficients at large times for anomalously diffusing particles in an analogous way as the Einstein relation does for normal diffusion in a thermal environment. The only, but most relevant modification consists in a factor that multiplies the second moment of the momentum. This factor depends only on the scaling exponent of the considered anomalous diffusion process and becomes unity for normal diffusion.
Here, the time-dependent diffusion and friction coefficients are defined as the derivative of the Brownian particle's mean square displacement with respect to time and the time-integral of the memory function entering the generalized Langevin equation, respectively.
The proof of the generalized asymptotic Einstein relation is based on the fact that,
under the given condition of a mean square displacement regularly varying at large times, the diffusion and friction coefficients are inversely proportional to each other, i.e. $D(t) \propto \gamma^{-1}(t) \propto t^{\alpha -1} L(t)$.
Here we used that any function varying regularly in time may be represented as $t^\alpha L(t)$ with $\alpha >0$ and $L(t)$ being slowly varying.
A corresponding relation between the velocity autocorrelation function and the memory kernel has been known in the case of purely algebraic scaling, i.e. with a trivial, constant slowly varying function \cite{Corngold,Morgado,MCO}.
The use of the generalized asymptotic Einstein relation allows one to estimate the scaling exponent in a way that is not influenced by slowly varying functions.
We confirmed the inverse scaling behavior of the diffusion and the friction coefficients by means of molecular dynamics simulations of self-diffusion in a two dimensional liquid. The apparent deviation of the scaling exponents found from the diffusion and friction coefficients on the one hand and the generalized asymptotic Einstein relation
can be explained by the influence of a slowly varying function.
We identified a large region in the temperature-density plane with super-diffusive self-diffusion and a cross-over region where the self-diffusion is normal.
A super-diffusive behavior was found by Isobe \cite{Isobe} for a hard-disk fluid at moderate densities. The scaling exponents found by Isobe are similar to the ones we found for a soft disk fluid.
The mechanism leading to the observed algebraic behavior is not known. In particular it is not explained by the existing mode-coupling theories \cite{AW,K,WAG}.
In the case of a two-dimensional Lorentz gas the existence of ``empty corridors'' is responsible for the occurrence of super-diffusion \cite{AHO}. Whether a similar mechanism may explain the more dynamic situation of a probe particle moving under the mutual influence of other particles presents an open question.
The existence of a region with normal diffusion was reported by Liu, Goree and Vaulina \cite{LGV} for a system of particles mutually interacting via Yukawa potentials in two dimensions in a density-temperature region that is comparable with to the one where we observe normal self-diffusion.
It is interesting to note that both the largest Lyapunov exponent and the Kolmogorov Sinai entropy have their maxima as functions of the density in the same region where one finds normal diffusion \cite{FP}. These quantities present measures of chaoticity of a system which is apparently largest for normal diffusion.
The derivation of the generalized asymptotic Einstein relation also holds for the motion of Brownian particles which may have mass and size that differ from those of the fluid particles. In contrast to other works, as in \cite{BMW}, in which the Einstein relation is extended to non-equilibrium systems that still display normal diffusive behavior, our extension refers to equilibrium systems exhibiting anomalous diffusion. The generalized asymptotic Einstein relation must also be distinguished from another form of generalized Einstein relation that expresses the mean position response on a small constant force in terms of the particle's mean square displacement in thermal equilibrium \cite{BG,BF}.
In the numerical confirmation of the generalized asymptotic Einstein relation we restricted ourselves to self-diffusion, i.e. to the motion of a test-particle with the same mass and size as all other fluid particles.
|
\section{Introduction and Formula} \label{s:intro}
Toric varieties are parameterized by monomials. They
form an important and rich class of examples in algebraic geometry and often provide a testing ground for
theorems \cite{CLS11}. Here we work with toric varieties that need not be normal.
We fix an algebraically closed field $K$, and we set
$K^{*}=K \backslash \{0\}$. Our first ingredient is an arbitrary projective toric
variety of codimension $2$ in $\mathbb{P}^{n+1}$. This is defined as follows.
Fix an $n\times (n+2)$ integer matrix $A$ whose columns span the lattice $\mathbb{Z}^n$:
\[
A\,\,=\,\,\begin{bmatrix}
\mb{a}_0 & \mb{a}_1 & \ldots & \mb{a}_{n+1} \\
\end{bmatrix}.
\]
The column vectors $\mb{a}_i$ correspond to Laurent monomials
$\bt^{\mb{a}_i}=t_{1}^{a_{1,i}}t_{2}^{a_{2,i}}\ldots t_{n}^{a_{n,i}}$
in the variables $\bt=(t_1,t_2,\ldots,t_{n})$.
These $n+2$ monomials specify a monomial map
\[
\Phi_{A}: (K^*)^n \rightarrow (K^*)^{n+2}, \,
\bt\to (\bt^{\mb{a}_0},\bt^{\mb{a}_1},\ldots,\bt^{\mb{a}_{n+1}}).
\]
Throughout this paper we assume that all columns of $A$ sum to the
same positive integer $d$. Under this hypothesis, we obtain an induced map
$\, \Phi_{A}: (K^*)^n \rightarrow \PP^{n+1} $. The toric variety $X_A$ is the
closure in $\PP^{n+1}$ of the image of $\Phi_A$.
The degree of $X_A$ is the normalized volume of the polytope ${\rm conv}(A)$, and
the equations defining $X_A$ form the {\em toric ideal},
a well-studied object in combinatorial commutative algebra
and its applications \cite{Stu96}.
A natural extension of toric theory is the study of
{\em complexity one T-varieties} \cite{IS11}.
There are varieties with an action of a torus whose
general orbits have codimension one. They can be
viewed as a family of (possibly reducible) toric varieties over a curve. Our aim here
is to explore such T-varieties from the point of view
of symbolic computation. For simplicity we assume
that T-varieties are rational, i.e. the underlying curve is rational, and we focus on
projective hypersurfaces.
Our second ingredient is a vector of univariate polynomials in a new variable $x$:
\begin{equation}\label{def:Y}
{\bf f} \,= \,
\bigl(f_{0}(x),f_{1}(x),\ldots,f_{n+1}(x)\bigr) \,\, \in \,K[x]^{n+2}.
\end{equation}
This vector specifies a parametric curve $Y_{\mathbf{f}} \subset \PP^{n+1}$,
namely the closure of the set of points
$(f_{0}(x):f_{1}(x):\ldots:f_{n+1}(x))$.
Let $Z_{A,{\bf f}}$ denote the
\emph{Hadamard product} in $\PP^{n+1}$ of the toric variety $X_A$ and the curve $Y_\mathbf{f}$.
By definition, this is the Zariski closure of the~set
\[
\bigl\{(\bt^{\mb{a}_0}f_{0}(x):\bt^{\mb{a}_1}f_{1}(x):\ldots:\bt^{\mb{a}_{n+1}}f_{n+1}(x))\in \PP^{n+1}
\,|\,\,\bt\in (K^{*})^{n},x\in K \bigr\} \,\, \subset \,\, \PP^{n+1}.
\]
Under some mild hypotheses (see Theorem \ref{thm:main}(a)),
the variety $Z_{A,{\bf f}}$ has codimension $1$,
and we call it the {\em almost-toric hypersurface}
associated with $(A,{\bf f})$. We shall present a fast
method for implicitizing $Z_{A,{\bf f}}$. The output of our algorithm
is the irreducible polynomial
in $K[u_0,u_1,\ldots,u_{n+1}]$ that vanishes on $Z_{A,{\bf f}}$.
The torus action given by $A$ ensures
that its Newton polytope ${\rm Newt}(Z_{A,{\bf f}}) $
lies in a plane in $ \RR^{n+2}$, so it is a polygon.
Our first main result is the following
combinatorial formula for this Newton polygon.
The ingredients in our formula are two matrices, which we now define.
The {\em Pl\"{u}cker matrix} associated with $A$ is the
$(n+2) \times (n+2)$-matrix $P_A = (p_{ij}) $ with entries
\[
p_{ij}=\begin{cases}
\frac{1}{\delta} (-1)^{i+j}\det(A_{[i,j]}), & i<j; \\
-p_{ji}, & i>j; \\
0, & i=j,
\end{cases}
\]
where $\delta$ is the greatest common divisor of all $\det(A_{[i,j]})$.
Here $A_{[i,j]}$ is the $n \times n$ submatrix of $A$
obtained by deleting the columns ${\bf a}_i$ and ${\bf a}_j$.
The Pl\"ucker matrix $P_A$ is skew-symmetric and has rank $2$,
and its row space is the kernel of $A$. The latter property implies that
all rows and all columns of $P_A$ sum to zero.
The {\em valuation matrix} associated with ${\bf f}$ is
an integer matrix $V_{\bf f}$ with $n+2$ rows that is defined as follows.
Since $K$ is algebraically closed, each of our polynomials
$f_i(x)$ factors into linear factors in $K[x]$.
Let $g_{1}(x),g_{2}(x),\ldots,g_{m}(x)$ be the list of all distinct linear factors of
$f_0(x) f_1(x) \ldots f_{n+1}(x)$.
We write ${\rm ord}_{g_j}f_i$ for the order of vanishing of
$f_i(x)$ at the unique root of $g_j(x)$.
We organize these numbers into the vectors
\[
\mb{u}_{j}=(\text{ord}_{g_{j}}f_{0},\text{ord}_{g_{j}}f_{1},\ldots,\text{ord}_{g_{j}}f_{n+1})\in \NN^{n+2}
\qquad \hbox{ for $1\le j\le m$.}
\]
We now aggregate these vectors according to the lines they span.
Let $S=\{\mb{u}_{1},\ldots,\mb{u}_{m}\}$. If two vectors in $S$ are linearly dependent, then we delete them
and add their sum to the set. We repeat this procedure. After finitely many steps, we end up with
a new set $\,S'=\{\mb{v}_{1},\mb{v}_{2},\ldots,\mb{v}_{\ell}\}\,$
whose vectors span distinct lines. The {\em valuation matrix} is
\[
V_{\bf f} \,\,= \,\,
\begin{bmatrix}\,
\mb{v}_{1}^{T} & \mb{v}_{2}^{T} & \ldots &\mb{v}_{l}^{T} & (-\sum_{j=1}^{l}{\mb{v}_{j}})^{T}
\end{bmatrix}.
\]
The last vector represents the valuation at $\infty$.
It ensures that the rows of $V_{\bf f}$ sum to zero.
The following theorem allows us to derive the Newton polygon
from $A$ and ${\bf f}$.
\begin{thm} \label{thm:main}
The edges of the Newton polygon of $Z_{A,{\bf f}}$ are the columns of the product of
the Pl\"{u}cker matrix $P_A$ and the valuation matrix $V_{\bf f}$.
More precisely, given $A$ and ${\bf f}$,
\begin{itemize}
\item[(a)] if $\,{\rm rank}(P_A \cdot V_{\bf f}) = 0\,$ then $Z_{A,{\bf f}}$ is not a hypersurface;
\item[(b)] if $\,{\rm rank}(P_A \cdot V_{\bf f}) = 1\,$ then $Z_{A,{\bf f}}$ is a toric hypersurface;
\item[(c)] if $\,{\rm rank}(P_A \cdot V_{\bf f}) = 2\,$ then $Z_{A,{\bf f}}$ is a hypersurface but not
toric.
The directed edges of the Newton polygon of $Z_{A,{\bf f}}$ are the nonzero column vectors of
$P_{A} \cdot V_{\bf f}$.
\end{itemize}
\end{thm}
The rest of this paper is organized as follows.
In Section \ref{s:prf} we present the proof
of Theorem \ref{thm:main}, and we illustrate
this result with several small examples.
Earlier work of Philippon and Sombra \cite[Proposition 4.1]{PS08} yields
an expression for the degree of the almost-toric hypersurface
$Z_{A,{\bf f}}$ as a certain sum of integrals over ${\rm conv}(A)$.
Section \ref{s:alg} is concerned with computational issues.
Our primary aim is to give a fast algorithm for computing the
implicit equation of the almost-toric hypersurface $Z_{A,{\bf f}}$.
We develope such an algorithm and implement it in {\tt Maple 17}. A case
study of hard implicitization problems demonstrate
that our method performs very well.
The mathematics under the hood of Theorem \ref{thm:main} is
{\em tropical algebraic geometry} \cite{MS14}.
The proof relies on a technique known as
{\em tropical implicitization} \cite{STY07, ST08, SY08}.
Thus this article offers a concrete demonstration
that tropical implicitization can serve as an efficient
and easy-to-use tool in computer algebra. It lays the
foundation for future work that will extend toric algebra \cite{Stu96}
and its numerous applications to the almost-toric setting
of complexity one T-varieties \cite{IS11}.
\vfill
\section{Proof, Examples, and the Philippon-Sombra Formula }\label{s:prf}
In this section we prove Theorem \ref{thm:main}. In our proof we use the technique of
\emph{tropicalization}. We then present examples to illustrate our main theorem. We also consider an easier
task: finding the degree of the almost-toric hypersurface. We compare our result to existing results,
including the Philippon-Sombra Formula in \cite[Proposition 1.2]{PS08} and another formula in \cite{SY08}.
First we prove some of the claims made in Section \ref{s:intro}.
\begin{lem}\label{lem:gon}
Let $Z_{A,{\bf f}}$ be an almost-toric hypersurface. Then ${\rm Newt}(Z_{A,{\bf f}}) $ is at most
$2$-dimensional in $\RR^{n+2}$.
\end{lem}
\begin{proof}
Substituting variables $u_{0},\ldots,u_{n+1}$ by the parameterization $u_{i}=\bt^{\mb{a}_i}f_{i}(x)$ in the
implicit equation
$p(u_{0},\ldots,u_{n+1})$ of $Z_{A,{\bf f}}$, we get another polynomial in variables
$t_{1},t_{2},\ldots,t_{n},x$. The latter polynomial is the zero polynomial. After this substitution, each term
in $p$ becomes the product of a monomial in variables $t_{1},t_{2},\ldots,t_{n}$ and a polynomial in $x$.
Since $p$ is the generator of the principal ideal corresponding to the almost-toric hypersurface, all such
monomials in variables $t_{1},t_{2},\ldots,t_{n}$ are the same; otherwise the almost-toric hypersurface would
vanish on a polynomial that contains some terms of $p$, a contradiction. Suppose after substitution the
monomial is $ \prod_{i=1}^{n}{t_{i}^{\alpha_{i}}}$. If ${\bf v}=(v_{0},\ldots,v_{n+1})$ is a vertex of ${\rm
Newt}(Z_{A,{\bf f}}) $, then $p$ contains a term $\prod_{i=0}^{n+1}{u_{i}^{v_{i}}}$. Therefore $ {\bf v}\cdot
A^{T}=(\alpha_{1},\ldots,\alpha_{n})$. So vertices of ${\rm Newt}(Z_{A,{\bf f}}) $ satisfy $n$ independent
linear equations and we conclude that ${\rm Newt}(Z_{A,{\bf f}}) $ is at most $2$-dimensional.
\end{proof}
\begin{rem}
Even if $Z_{A,{\bf f}}$ is a hypersurface, its Newton polygon could be a degenerate
polygon that has only two vertices.
\end{rem}
\begin{lem}
Let $P_A$ be a Pl\"{u}cker matrix. Then $P_{A}$ is skew-symmetric. The rank of $P_{A}$ is $2$ and the entries
in each row and column of $P_{A}$ sum to $0$.
\end{lem}
\begin{proof}
The first claim follows directly from the definition of $P_{A}$. For the second claim, note that the inner
product of the $i$-th row of $P_{A}$ and the $j$-th row of $A$ is the determinant of the following $(n+1)\times
(n+1)$ matrix up to sign: append the $j$-th row of $A$ to matrix $A$ and then delete the $i$-th column of the new
matrix. Therefore ${\rm row}(P_{A})\subseteq {\rm row}(A)^{\perp}$, which means the rank of $P_{A}$ is at most
$2$. Since $A$ has full rank, by the definition of $P_{A}$ there is a nonzero non-diagonal entry $p_{ij}$ of
$P_{A}$. By the first claim, $p_{ji}$ is nonzero too. So we get a nonzero $2\times 2$ minor in $P_{A}$. For
the last claim, since $P_{A}$ is
skew-symmetric it is enough to prove the claim for rows. It turns out that up to sign, the sum of entries in the
$i$-th row is the determinant of the matrix formed by $A_{[i]}$ and the vector ${\bf 1}=(1,1,\ldots,1)\in
K^{n+1}$, where $A_{[i]}$ is the matrix obtained from $A$ by deleting the $i$-th column. Since each column of
$A$ has sum $d$, the vector ${\bf 1}$ lies in the row space of $A_{[i]}$, so the matrix is singular and its
determinant
is zero.
\end{proof}
Next, we explore $Z_{A,{\bf f}}$. We consider its \emph{tropicalization} (we follow the definition from
\cite[Definition 3.2.1]{MS14}). Let $X_{A},Y_{\bf f},Z_{A,{\bf f}}$ be defined as in Section \ref{s:intro}.
Note that $Z_{A,{\bf f}}$ is the Hadamard product $\overline{X_{A}*Y_{\bf f}}$ of the varieties $X_{A}$ and
$Y_{\bf f}$. We have the following result relating the tropicalizations of $X_{A},Y_{\bf
f},Z_{A,{\bf f}}$:
\begin{prop}\cite[Corollary 3.3.6]{Cue10}\label{prop:ha}
\[
{\rm trop}(Z_{A,{\bf f}})={\rm trop}(X_{A})+{\rm trop}(Y_{\bf f}),
\]
where the sum is Minkowski sum.
\end{prop}
So in order to find $\text{trop}(Z_{A,{\bf f}})$, it is enough to find
$\text{trop}(X_{A})$ and $\text{trop}(Y_{\bf f})$.
\begin{lem}\label{lem:X}
If ${\rm row}(A)$ is the row space of $A$ with real coefficients, then
\[
{\rm trop}(X_{A})={\rm row}(A).
\]
\end{lem}
\begin{proof}
By Proposition \ref{prop:ha}, it is enough to prove the case when $n=1$ and then use induction. For $n=1$ we
need to show that if $a_{0},a_{1},\ldots,a_{n+1}$ are integers and
\[
X_{A}=\text{cl}(\{(t^{a_{0}}:t^{a_{1}}:\ldots:t^{a_{n+1}})\in (\PP^{*})^{n+1}|t\in K^{*}\}),
\]
then
\[
\text{trop}(X_{A})=\{r\cdot{\bf a}|r\in \RR\},
\]
where ${\bf a}=(a_{0},a_{1},\ldots,a_{n+1})$. In this case the ideal $I(X_{A})$ is generated by binomials as
follows (cf. \cite[Corollary 4.3]{Stu96}):
\[
I(X_{A})=\langle {\bf x}^{\bf u}-{\bf x}^{\bf v}|{\bf a\cdot u=a\cdot v}\rangle.
\]
Then all points in $\text{trop}(X_{A})$ are scalar multiples of ${\bf a}$.
\end{proof}
\begin{lem}\label{lem:Y}
Let $Y_{\bf f}$ be defined as in (\ref{def:Y}), and $S=\{g_{1},\ldots,g_{m},\infty\}$. Then
\begin{equation}\label{eq:Y}
{\rm trop}(Y_{\bf f})=\bigcup_{z\in
S}{\{\lambda(\text{ord}_{z}{f_{0}},\text{ord}_{z}{f_{1}},\ldots,\text{ord}_{z}{f_{n+1}})\in \mathbb{R}^{n+2}
|\lambda \ge 0\}}.
\end{equation}
In addition, ${\rm trop}(Y_{\bf f})$ is an $1$-dimensional balanced polyhedral fan in $\RR^{n+2}$ and the rays
are spanned by the vectors ${\bf v}_{z}$, where
${\bf v}_{z}=(\text{ord}_{z}{f_{0}},\text{ord}_{z}{f_{1}},\ldots,\text{ord}_{z}{f_{n+1}})$.
\end{lem}
\begin{proof}
Let $K'=K\{\!\{t\}\!\}$ be the field of \emph{Puiseux series}\cite[Example 2.1.3]{MS14} in variable $t$ with coefficients in $K$. Then $K'$ has a nontrivial valuation and is algebraically closed. Let $Y'_{\bf f}$ be a variety parameterized as in (\ref{def:Y}) but $x$ varies in $K'$ instead. Note that $Y_{\bf f}$ and $Y'_{\bf f}$ have the same ideal, so $\text{trop}(Y_{\bf f})=\text{trop}(Y'_{\bf f})$.
By \cite[Theorem 3.2.3]{MS14}, $\text{trop}(Y'_{\bf f})=\overline{\text{val}(Y'_{\bf f})}$, where
\[
\text{val}(Y'_{\bf f})=\{(\text{val}(f_0(x)),\ldots,\text{val}(f_{n+1}(x)))|x\in K'\}.
\]
Since $\QQ$ is dense in $\RR$, the right hand side of (\ref{eq:Y}) is the closure of
\[
B=\bigcup_{z\in S}{\{\lambda(\text{ord}_{z}{f_{0}},\text{ord}_{z}{f_{1}},\ldots,\text{ord}_{z}{f_{n+1}})\in
\mathbb{R}^{n+2} |\lambda\in \mathbb{Q^{+}}\}}.
\]
It then suffices to show $B=\text{val}(Y'_{\bf f})$. We first show that $B\subseteq \text{val}(Y'_{\bf f})$. Fixing $z\in S$ and $\lambda>0$, we get a vector
\[{\bf u}=\lambda(\text{ord}_{z}{f_{0}},\text{ord}_{z}{f_{1}},\ldots,\text{ord}_{z}{f_{n+1}})\in B.\]
\begin{itemize}
\item If $z\ne \infty$, then $z=g_{j}$ for some $1\le j\le m$. Since each $g_{j}$ is linear, we may assume $g_{j}(x)=x-r_j$ where $r_j\in K$. For each $i$, we have $f_{i}(x)=(x-r_{j})^{\text{ord}_{g_{j}}{f_{i}}}h_{i}(x)$, where $h_{i}(r_j)\ne 0$. Then $\text{val}(h_{i}(r_j))=0$. Now if we take $x=r_{j}+t^{\lambda}\in K'$, then $f_{i}(x)=t^{\lambda\text{ord}_{g_{j}}{f_{i}}}h_{i}(r_{j}+t^{\lambda})$. Notice that $h_{i}(x)$ is a
polynomial in $K[x]$. Then $t^{\lambda}$ divides
$h_{i}(r_{j}+t^{\lambda})-h_{i}(r_j)$ in $K'$, so $\text{val}(h_{i}(r_{j}+t^{\lambda})-h_{i}(r_j))>0$. Then
$\text{val}(h_{i}(r_{j}+t^{\lambda}))=0$. Thus $\text{val}(f_{i}(x))=\lambda\text{ord}_{g_{j}}{f_{i}}$, which means
${\bf u}\in \text{val}(Y'_{\bf f})$.
\item If $z=\infty$, then $\text{ord}_{z}{f_{i}}=-\deg(f_{i})$. We take $x=t^{-\lambda}\in K'$. Then among all
terms in $f_{i}(x)$, the term with smallest valuations is the leading term, because $\lambda>0$. Then
\[
\text{val}(f_{i}(x))=(-\lambda)\deg(f_{i})=\lambda\text{ord}_{\infty}{f_{i}}.
\]
So ${\bf u}\in \text{val}(Y'_{\bf f})$, too.
\end{itemize}
We next show that $\text{val}(Y'_{\bf f}) \subseteq B$. Suppose ${\bf u}\in \text{val}(Y'_{\bf f})$. Then there exists
$x_0\in K'$ such that ${\bf u}=(\text{val}(f_{0}(x_0)),\ldots,\text{val}(f_{n+1}(x_0)))$. We may assume that ${\bf u}\ne 0$. We must deal with two cases.
\begin{itemize}
\item All terms of $x_0$ are nonnegative powers of $t$. Since
$f_{i}(x)=\prod_{j=1}^{m}{g_{j}(x)^{\text{ord}_{g_{j}}{f_{i}}}}$, we have
\[
\text{val}(f_{i}(x_0))=\sum_{j=1}^{m}{\text{ord}_{g_{j}}{f_{i}}\cdot \text{val}(g_{j}(x_0))}.
\]
Note that all $g_j$ are linear functions, then for $1\le j<j'\le m$ we have $\text{val}(g_{j}(x_0)-g_{j'}(x_0))=0$. Hence for $1\le j\le m$ there is at most one nonzero $\text{val}(g_{j}(x_0))$, while they are not all zero because ${\bf u}\ne 0$. Suppose $\text{val}(g_{j}(x_0))=\lambda>0$. Then ${\bf u}=\lambda(\text{ord}_{z}{f_{0}},\ldots,\text{ord}_{z}{f_{n+1}})$ where $z=g_j\in S$ and ${\bf u}\in B$.
\item At least one term of $x_0$ is a negative power of $t$. Suppose in $x_0$ the term with least degree is
$ct^{-\frac{p}{q}}$, where $\frac{p}{q}\in \mathbb{Q}_{\ge 0}$. Then ${\bf u}=\frac{p}{q}(\text{ord}_{\infty}{f_{0}},\text{ord}_{\infty}{f_{1}},\ldots,\text{ord}_{\infty}{f_{n+1}})$
is in $B$, too.
\end{itemize}
Finally, by \cite[Proposition 3.4.13]{MS14}, $\text{trop}(Y'_{\bf f})$ is $1$-dimensional and balanced.
\end{proof}
By the definition of $V_{\bf f}$, the tropicalization $\text{trop}(Y_{\bf f})$ is exactly the union of rays
generated by column vectors in $V_{\bf f}$. This implies the following corollary:
\begin{cor}\label{cor:trop} The tropicalization of $Z_{A,{\bf f}}$ is
\[
\{\mb{u}+\lambda\cdot \mb{v}^{T}|\mb{u}\in {\rm row}(A),\mb{v}\text{ is a column vector
of }V_{\bf f},\lambda\ge 0\}.
\]
\end{cor}
We now find ${\rm Newt}(Z_{A,{\bf f}}) $.
\begin{prop}\label{prop:direc}
The edges of ${\rm Newt}(Z_{A,{\bf f}}) $ are parallel to the nonzero column vectors in $P_{A}\cdot V_{\bf
f}$.
\end{prop}
\begin{proof}
By \cite[Proposition 3.1.10]{MS14}, $\text{trop}(Z_{A,{\bf f}})$ is the support of an $(n+1)$-dimensional
polyhedral fan, which is the \emph{$(n+1)$-skeleton} of the normal fan of ${\rm Newt}(Z_{A,{\bf f}}) $. Since
${\rm Newt}(Z_{A,{\bf f}})$ is a polygon, every cone in the $(n+1)$-skeleton of its normal fan is a cone
spanned by ${\rm row}(A)$ (which is exactly the orthogonal complement of the plane that contains the polygon)
and a ray inside this plane that is orthogonal to the corresponding edge of the polygon. By Corollary
\ref{cor:trop}, every one of these directed edges belongs to $\ker(A)$ and is orthogonal to the corresponding
column vectors of $V_{\bf f}$. Let ${\bf v}$ be a column vector of $V_{\bf f}$. Since $P_{A}$ is
skew-symmetric, we have
\[
{\bf v}^T \cdot P_{A} \cdot {\bf v}=0.
\]
Hence $P_{A}\cdot {\bf v}$ is orthogonal to ${\bf v}^T$, which means that there exists a scalar $c$ such that
$c(P_{A}\cdot {\bf v})$ represents an edge of ${\rm Newt}(Z_{A,{\bf f}})$.
\end{proof}
It remains to show that the length of column vectors in $P_A \cdot V_{\bf f}$ coincide with the edges of
${\rm Newt}(Z_{A,{\bf f}}) $. To analyze these lengths we recall the notion of \emph{multiplicity} of a
polyhedron which is maximal in a weighted polyhedral complex. We adopt the definition in \cite[Definition
3.4.3]{MS14}.
We have the following result:
\begin{lem}\cite[Lemma 3.4.6]{MS14}\label{lem:length}
The lattice length of any edge (defined as the number of lattice points on the edge minus $1$) of
${\rm Newt}(Z_{A,{\bf f}}) $ is the multiplicity of the corresponding
$(n+1)$-dimensional polyhedron in the normal fan of ${\rm Newt}(Z_{A,{\bf f}}) $.
\end{lem}
Note that if an edge of ${\rm Newt}(Z_{A,{\bf f}}) $ is expressed by a vector, then its lattice length
is the \emph{content} of that vector. Therefore it is enough to find out $\text{mult}(\sigma)$ for each
$\sigma$ in the $(n+1)$-skeleton of the normal fan of ${\rm Newt}(Z_{A,{\bf f}}) $. We cannot achieve this
directly from the definition of multiplicity, because it involves the implicit polynomial of $Z_{A,{\bf f}}$,
which is unknown to us. However, we can find out those multiplicities with the help of another result in
\cite{MS14}. The following proposition will lead to our main theorem.
\begin{prop}\label{prop:mult}
For every maximal cell $\sigma$ in the $(n+1)$-skeleton of the normal fan of ${\rm Newt}(Z_{A,{\bf f}}) $,
$\text{mult}(\sigma)$ is the content of the corresponding column vector of $P_A \cdot V_{\bf f}$.
\end{prop}
\begin{proof}
Let $\Sigma'$ be $\text{trop}(Z_{A,{\bf f}})$, which is a pure weighted polyhedral complex in $\RR^{n+2}$.
Suppose there are $m$ roots $z_{1},\ldots,z_{m}\in K$ of $\prod_{i=0}^{n+1}{f_{i}(z)}$. We define another pure
weighted polyhedral complex $\Sigma \subseteq \RR^{n+m+2}$ as follows:
\[
\Sigma=\{({\bf u,v})\in\RR^{n+m+2}|{\bf u}\in \text{row}(A),{\bf v}=\lambda\cdot e_{i},
1\le i\le m+1,\lambda\ge 0\},
\]
where $e_{i}$ is the vector with $i$-th component $1$ and others $0$ for $1\le i\le m$, and $e_{m+1}=-{\bf
1}$. Note that
\[
\Sigma=\text{trop}(Z),
\]
where $Z$ is the variety parameterized by
\[
\{((\bt^{\mb{a}_0}:\bt^{\mb{a}_1}:\ldots:\bt^{\mb{a}_{n+1}}),(x-z_{1},x-z_{2},\ldots,x-z_{m}))\in
\PP^{n+1}\times \CC^{m}|\,\,\bt\in (K^{*})^{n},x\in K\}.
\]
Note that we can get $\Sigma'$ from $\Sigma$ via a projection $\phi$, where $\phi$ keeps $u\in \text{row}(A)$
fixed and sends each $e_{i}$ to the transpose of the $i$-th column vector ${\bf v}_{i}$ in $V_{\bf f}$. So
$\phi$
maps a maximal cell $\sigma\in\Sigma $ to a maximal cell $\sigma'\in \Sigma'$. This correspondence of maximal
cells is a bijection. Suppose $\sigma'$ in the $(n+1)$-skeleton of the normal fan of ${\rm Newt}(Z_{A,{\bf
f}})$ corresponds to the vector ${\bf v}_{i}$. Then by \cite[(3.6.2)]{MS14}
\[
\text{mult}(\sigma')=\text{mult}(\sigma)\cdot[N_{\sigma'}:\phi(N_{\sigma})],
\]
where $N_{\sigma'}\subseteq \RR^{n+2}$ is the lattice generated by all integer points in the span of row
vectors of
$A$ and ${\bf v}_{i}^{T}$, and $N_{\sigma}\subseteq \RR^{n+m+2}$ is the lattice generated by all integer
points in the space ${\rm row}(A)\oplus \RR{\bf e}_{i}$. Let $N\subseteq \RR^{n+2}$ be the lattice generated
by row vectors of $A$ and ${\bf v}_{i}^{T}$. Then $N\subseteq N_{\sigma'},N_{\sigma}$ and
$[N_{\sigma'}:\phi(N_{\sigma})]=\frac{[N_{\sigma'}:N]}{[N_{\sigma}:N]}$. The lattice index $[N_{\sigma'}:N]$
is the greatest common divisor of all maximal minors of the matrix formed by all generating vectors of
$N_{\sigma'}$. Here this matrix is obtained from $A$ by adding one row vector ${\bf v}_{i}^{T}$ and its
maximal minors are the product of $\delta$ and one entry in the $i$-th column of $P_A \cdot V_{\bf f}$, so
$[N_{\sigma'}:N]$ is the product of $\delta$ and the content of the the $i$-th column vector of $P_A \cdot
V_{\bf f}$. Similarly, $[N_{\sigma}:N]$ is the greatest common divisor of all maximal minors of $A$, which is
$\delta$.
Finally the initial ideal $in_{\sigma}(I(Z))$ with respect to $\sigma$ of the ideal of $Z$ is the direct sum
of two prime ideals: the first one is toric, the second one is generated by linear polynomials. Then the quotient ring of $K[u_0,\ldots,u_{n+1}]$ modulo the toric ideal is a domain. Note that the quotient ring of $K[u_0,\ldots,u_{n+1},v_1,\ldots,v_m]$ modulo $in_{\sigma}(I(Z))$ is the quotient ring of $K[u_0,\ldots,u_{n+1}]$ modulo the toric ideal adjoining some variable(s) in $\{v_1,\ldots,v_m\}$, which is also a domain. Hence $in_{\sigma}(I(Z))$ is also prime. Then by the definition of multiplicity, $\text{mult}(\sigma)=1$. This finishes the proof.
\end{proof}
\begin{proof}[Proof of Theorem (\ref{thm:main})]
If $\,{\rm rank}(P_A \cdot V_{\bf f}) = 0\,$ then all column vectors of $V_{\bf f}$ belong to ${\rm row}(A)$,
which means $Z_{A,{\bf f}}$ contains the toric variety $X_{A}$ and has codimension $2$, so it is not a
hypersurface. If $\,{\rm rank}(P_A \cdot V_{\bf f}) = 1\,$, then essentially $\bf f$ provides one parameter
not appearing in $X_{A}$, which means that $Z_{A,{\bf f}}$ has codimension $1$ and is a toric hypersurface. If
$\,{\rm rank}(P_A \cdot V_{\bf f}) = 2\,$, then ${\rm Newt}(Z_{A,{\bf f}}) $ is a nondegenerate polygon and
with Proposition \ref{prop:direc}, Lemma \ref{lem:length} and Proposition \ref{prop:mult} we have proved
Theorem \ref{thm:main}.
\end{proof}
We illustrate Theorem \ref{thm:main} with the following example.
\begin{exa}\label{exa:H}
Let $Z_{A,{\bf f}}$ admit the following parameterization over $\CC$:
\[
(t_{1}^{2}(x^{2}+1):t_{1}t_{2}x^{3}(x-1):t_{1}t_{3}x(x+1):t_{2}^2(x-2)(x^2+1):t_{3}^{2}(x-1)^{2}(x+1)).
\]
In this example $A=\begin{bmatrix}
2&1&1&0&0 \\
0&1&0&2&0 \\
0&0&1&0&2
\end{bmatrix},d=2$. Then $P_{A}=\begin{bmatrix}
0&-2&2&1&-1 \\
2&0&-4&0&2 \\
-2&4&0&-2&0 \\
-1&0&2&0&-1 \\
1&-2&0&1&0
\end{bmatrix}$ and $\delta=2$. The linear factors of the univariate polynomials are $x,x-1,x+i,x-i,x+1,x-2$.
But we can combine $x\pm i$ into $x^2+1$. So the vectors are
\[
(0,3,1,0,0),(0,1,0,0,2),(2,0,0,2,0),(0,0,1,0,1),(0,0,0,1,0),(-2,-4,-2,-3,-3).
\]
Then the valuation matrix of ${\rm Newt}(Z_{A,{\bf f}})$ is
\[
V_{\bf f}=\begin{bmatrix}
0&0&2&0&0&-2 \\
3&1&0&0&0&-4 \\
1&0&0&1&0&-2 \\
0&0&2&0&1&-3 \\
0&2&0&1&0&-3 \\
\end{bmatrix}.
\]
Using ideal elimination in {\tt Macaulay 2} we can compute the implicit polynomial of ${\rm Newt}(Z_{A,{\bf
f}})$ in variables
$u_{0},u_{1},u_{2},u_{3},u_{4}$:
\begin{equation*}
\begin{split}
&16u_{1}^{4}u_{2}^{16}u_{3}^{2}-40u_{0}u_{1}^{4}u_{2}^{14}u_{3}^{2}u_{4}+8u_{0}^{2}u_{1}^{2}u_{2}^{14}u_{3}^{3}u_{4}-16u_{0}u_{1}^{6}u_{2}^{12}u_{3}u_{4}^{2}+20u_{0}^{2}u_{1}^{4}u_{2}^{12}u_{3}^{2}u_{4}^{2}\\
+&159u_{0}^{3}u_{1}^{2}u_{2}^{12}u_{3}^{3}u_{4}^{2}+u_{0}^{4}u_{2}^{12}u_{3}^{4}u_{4}^{2}+54u_{0}^{2}u_{1}^{6}u_{2}^{10}u_{3}u_{4}^{3}-77u_{0}^{3}u_{1}^{4}u_{2}^{10}u_{3}^{2}u_{4}^{3}+379u_{0}^{4}u_{1}^{2}u_{2}^{10}u_{3}^{3}u_{4}^{3}\\
+&5u_{0}^{2}u_{1}^{8}u_{2}^{8}u_{4}^{4}-27u_{0}^{3}u_{1}^{6}u_{2}^{8}u_{3}u_{4}^{4}-29u_{0}^{4}u_{1}^{4}u_{2}^{8}u_{3}^{2}u_{4}^{4}+163u_{0}^{5}u_{1}^{2}u_{2}^{8}u_{3}^{3}u_{4}^{4}-12u_{0}^{3}u_{1}^{8}u_{2}^{6}u_{4}^{5}\\
-&35u_{0}^{4}u_{1}^{6}u_{2}^{6}u_{3}u_{4}^{5}-425u_{0}^{5}u_{1}^{4}u_{2}^{6}u_{3}^{2}u_{4}^{5}+4u_{0}^{6}u_{1}^{2}u_{2}^{6}u_{3}^{3}u_{4}^{5}+87u_{0}^{5}u_{1}^{6}u_{2}^{4}u_{3}u_{4}^{6}+717u_{0}^{6}u_{1}^{4}u_{2}^{4}u_{3}^{2}u_{4}^{6}\\
+&103u_{0}^{6}u_{1}^{6}u_{2}^{2}u_{3}u_{4}^{7}-115u_{0}^{7}u_{1}^{4}u_{2}^{2}u_{3}^{2}u_{4}^{7}+12u_{0}^{7}u_{1}^{6}u_{3}u_{4}^{8}+4u_{0}^{8}u_{1}^{4}u_{3}^{2}u_{4}^{8}.\\
\end{split}
\end{equation*}
The vertices of Newton polygon of this implicit polynomial are
\[
(0,4,16,2,0),(2,8,8,0,4),(3,8,0,6,5),(7,6,0,1,8),(8,4,0,2,8),(4,0,12,4,2).
\]
The directed edges are
\[
\begin{split}
&(2,4,-8,-2,4),(-4,4,4,-2,-2),(-4,-4,12,2,-6) \\
&(1,-2,0,1,0),(4,-2,-6,1,3),(1,0,-2,0,1),
\end{split}
\]
and the product $P_{A}\cdot V_{\bf f}$ is
\[
\begin{bmatrix}
-4&-4&2&1&1&4\\
-4&4&4&-2&0&-2\\
12&4&-8&0&-2&-6\\
2&-2&-2&1&0&1\\
-6&-2&4&0&1&3\\
\end{bmatrix}.
\]
The column vectors of this matrix are the directed edges of ${\rm Newt}(Z_{A,{\bf f}})$.
\end{exa}
Next we compare our result with existing work. The Philippon-Sombra formula \cite[Proposition 4.1]{PS08}
computes the degree of $Z_{A,{\bf f}}$ from $A$ and $\bf f$. Let $B$ be a $2\times (n+2)$ matrix such that its
entries are integers and its row vectors span the kernel of $A$, and ${\bf b}_{0},\ldots,{\bf b}_{n+1}$ be the
column vectors of $B$. Given any column vector ${\bf v}=(v_{0},\ldots,v_{n+1})^{T}\in \RR^{n+2}$, we define a
polytope as the convex hull of the following $n+2$ vertices: ${\bf c}_{i}$ is the vector formed by the $i$-th
column vector of $A$ and $v_{i-1}$ for $1\le i\le n+2$. Then the vector $B\cdot {\bf v}$ admits a
triangulation $T$ of this polytope: for all $0\le i<j\le n+1$, the $n$-dimensional simplex formed by vertices
excluding ${\bf c}_{i},{\bf c}_{j}$ belongs to this triangulation if and only if the vector $B\cdot {\bf v}$
belongs to the nonnegative span of ${\bf b}_{i}$ and ${\bf b}_{j}$. Now we define a sum
\[
\partial_{\bf v}(A)=\frac{1}{\delta}\mathop{\sum}_{\sigma \in T}{|A_{\sigma}|\sum_{i\in \sigma}{v_{i}}}.
\]
Here $A_{\sigma}$ is the $n\times n$ minor of $A$ that contains all columns corresponding to $\sigma$.
\begin{prop}\label{prop:ps}
The degree of an almost-toric hypersurface $Z_{A,{\bf f}}$ is
\[
\mathop{\sum}_{{\bf v} \text{ is a column vector of } V_{f}}{\partial_{\bf v}(A)}.
\]
\end{prop}
This proposition is a direct corollary of \cite[Proposition 4.1]{PS08}.
\begin{rem}
For some ${\bf v}$, $\partial_{\bf v}(A)$ is negative, so the computation of degree using the Philippon-Sombra
formula is not very simple. The author attempted to obtain an alternative interpretation of the formula such
that all
summands are positive, but failed. However in Section \ref{s:alg} we present an algorithm to compute the
implicit polynomial of ${\rm Newt}(Z_{A,{\bf f}}) $, and in step $(3)$ we can compute the degree of this
polynomial efficiently.
\end{rem}
\cite[Theorem 5.2]{STY07} also provides an alternative way to compute the degree of $Z_{A,{\bf f}}$, using
tropical geometry. Applying this theorem to our almost-toric hypersurface $Z_{A,{\bf f}}$ we have the
following corollary.
\begin{cor}\label{cor:deg}
For a generic column vector ${\bf w}\in \RR^{n+2}$, the $i$-th coordinate of the vertex $face_{\bf w}({\rm
Newt}(Z_{A,{\bf f}}))$ is the number of intersection points, each counted with its \emph{intersection
multiplicity}, of the tropical hypersurface $\text{trop}(Z_{A,{\bf f}})$ with the half line ${\bf w}+\RR_{\ge
0}{\bf e}_{i}$.
\end{cor}
Here intersection multiplicity is defined in the remark after \cite[Theorem 5.2]{STY07} and ${\bf e}_{i}\in
\RR^{n+2}$ is the column vector with $i$-th component $1$ and others $0$ for $1\le i\le n+2$. From Corollary
\ref{cor:deg} we get the following proposition.
\begin{prop}\label{prop:deg}
Let $p(u_0,\ldots,u_{n+1})$ be the implicit polynomial of $Z_{A,{\bf f}}$. Then for a generic vector ${\bf
w}\in \RR^{n+2}$, the initial monomial $in_{\bf w}p$ is
\[
\mathop{\prod}_{(i,j)\in S}{u_{i}^{|{\bf e}_{i}P_{A}{\bf v}_{j}|}},
\]
where
\[
\begin{split}
S=&\{(i,j)|1\le i\le n+2,{\bf v}_{j} \text{ is a column vector of }V_{\bf f},\\
&{\bf e}_{i}^{T}P_{A}{\bf w},{\bf e}_{i}^{T}P_{A}{\bf v}_{j},{\bf v}_{j}^{T}P_{A}{\bf w}\text{ have the same sign.}\}.
\end{split}
\]
Hence the degree of $Z_{A,{\bf f}}$ is
\[
\mathop{\sum}_{(i,j)\in S}{|{\bf e}_{i}P_{A}{\bf v}_{j}|}.
\]
\end{prop}
\begin{proof}
If $\bf w$ is generic, then $\text{trop}(Z_{A,{\bf f}})$ and ${\bf w}+\RR_{\ge 0}{\bf e_i}$ have at most one
intersection point. Let ${\bf v}_{j}$ be the $j$-th column vector of matrix $V_{\bf f}$ and ${\rm
col}(A^{T})=\{{\bf u}^{T}|{\bf u}\in {\rm row}(A)\}$. Suppose there is an intersection point of the maximal
cone spanned by ${\rm row}(A)$ and the half line $\RR_{\ge 0}{{\bf v}_{j}}$ and ${\bf w}+\RR_{\ge 0}{\bf
e}_i$. Since $\bf w$ is generic, the $2$-dimensional subspace spanned by ${\bf w},{\bf e}_i$ has a unique
common point with a translation of a codimension $2$ subspace: ${\bf w}-{\rm col}(A^{T})$. Then there exists a
unique pair of nonzero (because $\bf w$ is generic) real numbers $\lambda_1,\lambda_2$ and ${\bf u}\in {\rm
row}(A)$ such that ${\bf u}^{T}+\lambda_{1}{\bf v}_{j}={\bf w}+\lambda_{2}{\bf e}_{i}$. Then the intersection
point exists if and only if $\lambda_1,\lambda_2>0$. Note that $\lambda_{1}>0$ if and
only if ${\bf v}_{j}$ and ${\bf w}$ are on the same side of the hyperplane spanned by ${\rm col}(A^{T})=\{{\bf
u}^{T}|{\bf u}\in {\rm row}(A)\}$ and ${\bf e}_{i}$ (if they don't span a hyperplane, then ${\bf e}_{i}\in
{\rm col}(A^{T})$ and no intersection point exists since $\bf w$ is generic). So $\det(\begin{bmatrix}{\bf
v}_{j} & {\bf e}_{i} & A^{T} \end{bmatrix})$ and $\det(\begin{bmatrix}{\bf w} & {\bf e}_{i} & A^{T}
\end{bmatrix})$ have the same sign.
Note that for any column vectors ${\bf a,b}\in \RR^{n+2}$, $\det(\begin{bmatrix}{\bf b}^{T} \\ {\bf a}^{T} \\
A \end{bmatrix})=\det(\begin{bmatrix}{\bf b} & {\bf a} & A^{T} \end{bmatrix})=\delta{\bf a}^{T}P_{A}{\bf b}$.
Hence $\lambda_{1}>0$ if and only if ${\bf e}_{i}^{T}P_{A}{\bf w}$ and ${\bf e}_{i}^{T}P_{A}{\bf v}_{j}$
have the same sign. Similarly, $\lambda_{2}>0$ if and only if ${\bf w}$ and ${\bf e}_{i}$ are on the
opposite sides of the hyperplane spanned by ${\rm col}(A^{T})$ and ${\bf v}_{j}$, which is equivalent to ${\bf
e}_{i}^{T}P_{A}{\bf v}_{j}$ and ${\bf v}_{j}^{T}P_{A}{\bf w}$ having the same sign. So $(i,j)\in S$ if and only
if there is an intersection point of the maximal cone spanned by ${\rm row}(A)$ and the half line $\RR_{\ge
0}{{\bf v}_{j}}$ with the half line ${\bf w}+\RR_{\ge 0}{\bf e_i}$. Next it suffices to show that the
intersection multiplicity of this point is $|{\bf e}_{i}P_{A}{\bf v}_{j}|$. By the definition of intersection
multiplicity, for this point it is the lattice index of the lattice spanned by ${\bf e}_{i},{\bf v}_{j}$ and
the transpose of row vectors of $A$, so the intersection multiplicity is $|{\bf e}_{i}P_{A}{\bf v}_{j}|$.
\end{proof}
\begin{rem}
Proposition \ref{prop:deg} enables us to compute the degree of $Z_{A,{\bf f}}$ without knowing ${\rm
Newt}(Z_{A,{\bf f}})$.
\end{rem}
\section{Algorithm, Implementation and Case Study}
\label{s:alg}
\subsection*{Algorithm to compute the implicit polynomial}
For an almost-toric hypersurface, we would like to compute its implicit polynomial in $n+2$ variables
$u_{0},u_{1},\ldots,u_{n+1}$ from $A$ and ${\bf f}$. An existing approach uses ideal elimination with
Gr\"{o}bner
bases, which is inefficient when $n$ is large. Based on Theorem \ref{thm:main} we have the following
alternative
approach:
\begin{enumerate}
\item Compute $P_{A}$ from $A$, factorize $f_{0},f_{1},\ldots,f_{n+1}$ over $K$ into irreducible factors
to get $V_{\bf f}$.
\item Compute $P_{A}\cdot V_{\bf f}$ and verify it has rank $2$.
\item Find ${\rm Newt}(Z_{A,{\bf f}}) $ using Theorem \ref{thm:main}.
\item Determine all possible monomials in variables $u_{0},u_{1},\ldots,u_{n+1}$ that could appear in the
implicit polynomial.
\item Use linear algebra to compute the coefficients of these monomials.
\end{enumerate}
We now explain our implementation of this method using the software {\tt Maple 17}. Among the five steps, the
first and second are trivial to implement ({\tt Maple 17} has the command {\it factor} which factors a
polynomial into irreducible factors over a given field).
\subsubsection*{Step (3)}
Theorem \ref{thm:main} tells us that the set of directed edges are the column vectors of $P_{A}\cdot V_{\bf
f}$. Then we need to arrange them in the correct order. We could project these vectors to a $2$-dimensional
space, by choosing two of the coordinates $1\le c_{1}<c_{2}\le n+2$. There is still the problem of
orientation: these directed edges admit two different arrangements. The correct orientation is determined by
the sign of the $(P_{A})_{c_{1},c_{2}}$.
Now suppose all directed edges are arranged in correct order and are the column vectors of a
matrix
\[
\begin{bmatrix}
c_{i,j}
\end{bmatrix}_{1\le i\le n+2,1\le j\le m}.
\]
If the vertex that corresponds to the first and $m$-th edges has coordinates $r_{1},\ldots,r_{n+2}$, then the
other
vertices have coordinates
\[
(r_{1}+\mathop{\sum}_{j=1}^{k}{c_{1,j}},r_{2}+\mathop{\sum}_{j=1}^{k}{c_{2,j}},\ldots,r_{n+2}+\mathop{\sum}_{j=1}^{k}{c_{n+2,j}}),k=1,\ldots
,m-1.
\]
We notice that for $1\le i\le n+2$, the $i$-th coordinate of the vector of vertices corresponds to the
exponents of $t_{i}$. Since the implicit polynomial is irreducible, the minimum of these exponents must be
$0$:
\[
\min_{1\le k\le m}\{r_{i}+\mathop{\sum}_{j=1}^{k}{c_{i,j}}\}=0, 1\le i\le n+2.
\]
Hence
\[
r_{i}=-\min_{1\le k\le m}\{\mathop{\sum}_{j=1}^{k}{c_{i,j}}\}, 1\le i\le n+2.
\]
So ${\rm Newt}(Z_{A,{\bf f}}) $ is uniquely determined by $P_{A}\cdot V_{\bf f}$ and we can compute its
vertices using the formula above.
\subsubsection*{Step (4)}
Given the vertices of ${\rm Newt}(Z_{A,{\bf f}}) $, we need to find all lattice points of this polygon.
Since all of them lie in the translation of a $2$-dimensional subspace, projection onto $2$ coordinates would
work. We find all lattice points within a convex polygon (which may be degenerate), then recover the
corresponding lattice points in ${\rm Newt}(Z_{A,{\bf f}}) $. These lattice points correspond to all possible
monomials in the implicit polynomial of $Z$: components of each vector are the exponents of variables
$u_{0},u_{1},\ldots,u_{n+1}$.\\
\subsubsection*{Step (5)}
Given all monomials of the implicit polynomial $p(u_0,\ldots,u_{n+1})$, it is enough to find the coefficients
of them. Consider the coefficients as undetermined unknowns. After the substitution
$u_{i}=\bt^{\mb{a}_i}f_{i}(x)$ we get another polynomial $q(t_1,\ldots,t_n,x)$. Each term in $q$ is the
product of a coefficient, a monomial in the variables $t_1,\ldots,t_n$ and a polynomial in the variable $x$.
We claim that all these monomials in variables $t_1,\ldots,t_n$ are the same. Assume the opposite situation.
Then we can pick all terms in $q$ with a particular monomial in variables $t_1,\ldots,t_n$ and get the
corresponding monomials in $p$, which form another polynomial $p'(u_0,\ldots,u_{n+1})$ with less terms than
$p$. Since $p$ is the implicit polynomial, polynomial $q$ must be identically zero, then $p'$ vanishes
everywhere too, so $p'$ belongs to the principal ideal of our hypersurface, a contradiction! So after the
substitution we can cancel the unique monomial in variables $t_1,\ldots,t_n$ and get a univariate polynomial
in $x$ with undetermined coefficients. Since this polynomial is identically zero, the undetermined
coefficients satisfy a system of homogeneous linear equations.\\
Next we use interpolation. Suppose there are $k$ possible monomials in the implicit polynomial, then we
replace $x$ by integers ranging from $-r$ to $r$, where $r=\lfloor \frac{k}{2} \rfloor$. Each interpolation
gives a linear equation with $k$ coefficients. Then we use the \emph{solve} command in {\tt Maple 17} to solve
these coefficients. Since this is a homogeneous linear system, the solution space should be $1$-dimensional.
This leads us to add another equation, for example $a_{1}=1$ where $a_{1}$ is one of the coefficients, to
guarantee the uniqueness of solution. After getting the solution, if all coefficients are rational, we
normalize them so that their content is $1$.
\begin{exa}\label{exa:Z}
Let $Z_{A,{\bf f}}$ be the almost-toric surface in $\PP^{3}$ parameterized by
\[
(s^{3}x^{2}(x-1):s^{2}t(x^{2}+1):st^{2}x(x+1)^{2}:t^{3}(x-1)(x-2)).
\]
Then
$
A=\begin{bmatrix}
3&2&1&0 \\
0&1&2&3 \\
\end{bmatrix}
$
and
$V_{f}=
\begin{bmatrix}
2&1&0&0&0&-3 \\
0&0&2&0&0&-2 \\
1&0&0&2&0&-3 \\
0&1&0&0&1&-2
\end{bmatrix}$. The vertices of ${\rm Newt}(Z_{A,{\bf f}}) $ are
\[
(6,0,0,6),(4,0,6,2),(2,2,8,0),(1,4,7,0),(0,7,4,1),(2,6,0,4).
\]
We find all monomials in the implicit polynomial to be
\[
\begin{split}
&u_{0}^6u_{3}^6, u_{0}^5u_{1}u_{2}u_{3}^5, u_{0}^5u_{2}^3u_{3}^4, u_{0}^4u_{1}^3u_{3}^5,
u_{0}^4u_{1}^2u_{2}^2u_{3}^4, u_{0}^4u_{1}u_{2}^4u_{3}^3, u_{0}^4u_{2}^6u_{3}^2, u_{0}^3u_{1}^4u_{2}u_{3}^4,
u_{0}^3u_{1}^3u_{2}^3u_{3}^3,\\ &u_{0}^3u_{1}^2u_{2}^5u_{3}^2, u_{0}^3u_{1}u_{2}^7u_{3},
u_{0}^2u_{1}^6u_{3}^4, u_{0}^2u_{1}^5u_{2}^2u_{3}^3, u_{0}^2u_{1}^4u_{2}^4u_{3}^2, u_{0}^2u_{1}^3u_{2}^6u_{3},
u_{0}^2u_{1}^2u_{2}^8, u_{0}u_{1}^6u_{2}^3u_{3}^2,\\ &u_{0}u_{1}^5u_{2}^5u_{3}, u_{0}u_{1}^4u_{2}^7,
u_{1}^7u_{2}^4u_{3}.
\end{split}
\]
We use interpolation to solve for the coefficients of these monomials and obtain the implicit polynomial
\[
\begin{split}
&8u_{0}^6u_{3}^6+52u_{0}^5u_{1}u_{2}u_{3}^5-28u_{0}^5u_{2}^3u_{3}^4-48u_{0}^4u_{1}^3u_{3}^5+58u_{0}^4u_{1}^2u_{2}^2u_{3}^4-82u_{0}^4u_{1}u_{2}^4u_{3}^3\\
&+25u_{0}^4u_{2}^6u_{3}^2-1312u_{0}^3u_{1}^4u_{2}u_{3}^4-175u_{0}^3u_{1}^3u_{2}^3u_{3}^3+476u_{0}^3u_{1}^2u_{2}^5u_{3}^2-50u_{0}^3u_{1}u_{2}^7u_{3}\\
&+72u_{0}^2u_{1}^6u_{3}^4-5760u_{0}^2u_{1}^5u_{2}^2u_{3}^3+5056u_{0}^2u_{1}^4u_{2}^4u_{3}^2-1194u_{0}^2u_{1}^3u_{2}^6u_{3}+25u_{0}^2u_{1}^2u_{2}^8\\
&-4176u_{0}u_{1}^6u_{2}^3u_{3}^2-3256u_{0}u_{1}^5u_{2}^5u_{3}-72u_{0}u_{1}^4u_{2}^7-576u_{1}^7u_{2}^4u_{3}.
\end{split}
\]
\end{exa}
\subsection*{Efficiency Test}
To test the efficiency of our algorithm, we try some simple examples using our implementation in {\tt Maple
17} and using ideal elimination in {\tt Macaulay 2}. The result is in Table \ref{tab:sim}:
\begin{table}[h]
\begin{tabular}{|c|c|c|c|c|}
\hline
sample & degree & \# of terms & our time cost & ideal elimination's time cost \\
\hline
Example \ref{exa:H} & 22 & 24 & 0.094s & 1.875s\\
\hline
Example \ref{exa:Z} & 10 & 16 & 0.047s & 0.078s.\\
\hline
\end{tabular}
\caption{Simple Examples}\label{tab:sim}
\end{table}
We then try some examples that both {\tt Macaulay 2} and {\tt Sagemath} cannot solve in a reasonable time. We
generate some samples as the input using the following method: let $n$ be the dimension of the torus, $d$ the
degree of the homogeneous monomials and $k$ a positive integer. Then we choose $n+2$ degree $d$ monomials
randomly from all possible $\binom{n+d-1}{d}$ choices. For the univariate polynomials, we choose $n+2$
polynomials of the form $(x-2)^{*}(x-1)^{*}x^{*}(x+1)^{*}(x+2)^{*}$, where each $*$ is a random integer
between $0$ and $k$. Table \ref{tab:com} shows the time needed to find the implicit polynomial of almost-toric
hypersurfaces given by randomly generated inputs. It turns out that our implementation improves the efficiency
of finding the implicit polynomial of almost-toric hypersurfaces.
\begin{table}[hb]
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
sample & degree & \# of terms & time cost & sample & degree & \# of terms & time cost\\
\hline
1 & 213 & 109 & 12.484s & 6 & 179 & 97 & 8.110s \\
\hline
2 & 109 & 80 & 1.594s & 7 & 40 & 32 & 0.156s \\
\hline
3 & 172 & 129 & 10.421s & 8 & 27 & 14 & 0.140s \\
\hline
4 & 474 & 275 & 156.969s & 9 & 79 & 71 & 1.766s \\
\hline
5 & 291 & 137 & 20.375s & 10 & 281 & 148 & 20.719s \\
\hline
\end{tabular}
\caption{$n=4,d=4,k=5$}\label{tab:com}
\end{table}
|
\section{Introduction}
In this paper we address two basic questions:
\begin{enumerate}
\item Can Alice and Bob test whether their joint state is maximally entangled while exchanging only a constant number of qubits?
More precisely, Alice and Bob hold two halves of a quantum state
$\ket{\psi}$ on a $D^2$-dimensional space for large $D$, and would like to check whether $\ket{\psi}$ is the maximally entangled state
$\ket{\phi_D} = \frac{1}{\sqrt{D}} \sum_x \ket{x}\ket{x}$ or whether it is orthogonal to that state.
The first entanglement tester
is the {\em hashing protocol} of the influential 1996 paper by Bennett, DiVincenzo, Smolin and Wootters~\cite{BDSW96}; further results
are summarized in table \tabref{ent-test}. Entanglement testing has found various applications, including entanglement distillation and error correction~\cite{BDSW96}, state authentication~\cite{BCGST02} and bounding the communication capacities of
bipartite unitary operators~\cite{HL07}.
As can be seen from this table,
all known protocols for this task
require resources (communication, shared randomness or catalyst)
which grow with $D$ \cite{BDSW96, BCGST02, HL07}.
\item Is there a counterexample to the generalized area law? A sweeping conjecture in condensed matter physics, and one of the most
important open questions in quantum Hamiltonian complexity theory, is
the so called ``area law,'' which asserts that ground states of quantum
many body systems on a lattice
have limited entanglement. Specifically, assume the system is described by a gapped local Hamiltonian\footnote{Here and later, by gapped local Hamiltonian we mean a Hamiltonian whose difference between ground energy and next excited
energy is $\Omega(1)$.} $H = H_1 + \ldots + H_m$, where each $H_i$
describes a local interaction between two neighboring particles of a
lattice. The area law conjectures that for every subset $S$ of the
particles, the entanglement entropy between $S$ and $\bar{S}$ for the
ground state of $H$ is bounded by a constant times the size of the
boundary of $S$. The area law, which has been proven for 1-D
lattices~\cite{ref:Has07} and is conjectured for higher degree
lattices, is of central importance in condensed matter physics as it
provides the basic reason to hope that ground states of gapped local
Hamiltonians on lattices might have a (relatively) succinct classical
description. The {\em generalized} area law (a folklore conjecture)
transitions from this physically motivated phenomenon to a very clean and general graph theoretic formulation, where in place of edges of the lattice, the terms of the Hamiltonian correspond to edges of an arbitrary graph. The generalized area law then states that for any subset $S$ of vertices (particles), the entanglement entropy between $S$ and $\bar{S}$ for the ground state is bounded by some constant times the cut-set of $S$ (the number of edges leaving $S$).
\end{enumerate}
We affirmatively answer both questions, based on a common technique that may be thought of as applying quantum expanders in a distributed fashion. Indeed these two questions which at first sight seem completely unrelated turn out to be two sides of the same coin.
\begin{table}
\begin{tabular}{@{}llp{1.5in}p{3in}@{}}
\toprule
reference & form of $\ket\psi$ & communication required & other resources \\
\midrule
\cite{BDSW96} & $n$ EPR pairs & $O(n\log(1/\varepsilon))$ & consumes only $O(\log(1/\varepsilon))$ EPR pairs \\
\cite{BCGST02} & $n$ EPR pairs & $(2+o(1))\log(n/\varepsilon)$ & $O(n/\log(n/\varepsilon))$ bits of shared randomness \\
\cite{HL07} & $n$ EPR pairs & $O(\log(1/\varepsilon))$ & $n/\varepsilon$ EPR pairs \\
this paper & $n$ EPR pairs & $O(\log(1/\varepsilon))$ & \\
\bottomrule
\end{tabular}
\caption{Comparison of different entanglement-testing protocols. When
we say that communication $x$ is required, this means that we need
to consume either $x$ qubits or, alternatively (by
teleportation), $2x$ classical bits and $x$ EPR pairs. The exception is the first row, which
uses classical communication to verify entanglement, and hence the
entanglement cost is lower. \label{tab:ent-test}}
\end{table}
\subsection{Main idea and results}
The main ingredient in both proofs is the notion of quantum expanders, which we discuss further in \secref{expanders}.
A quantum expander can be thought of as a collection of $d$ unitaries
$U_1,\ldots,U_d$, (think of $d$ as a constant) each acting on a (possibly large)
dimension-$D$ Hilbert space. For any
matrix $X$ on the $D$-dimensional Hilbert space, the operator associated with the expander,
$\mathcal{E}(X)= \frac{1}{d}\sum_{i=1}^d U_i X U_i^\dagger$,
has the unique eigenvalue 1 for the eigenvector $X=\bbI$ and next highest singular value $\lambda < 1$. It thus shrinks any
matrix orthogonal to the identity by a constant factor.
The key to the results in the paper is an equivalent way to view quantum expanders, by considering their action on maximally entangled states. It is well known that for any $U$, $U\otimes U^*$ acting on the maximally entangled state $\ket{\phi_D} = \frac{1}{\sqrt D}\sum_{i=1}^D \ket i \otimes \ket i$ leaves it as is. Of course, this remains true even if $U$ is drawn uniformly at random from the set $U_1,\ldots,U_d$ of the expander. Remarkably, even though quantum expanders use only a constant number $d$ of unitaries, they leave intact {\it only} the maximally entangled state, and cause all other states to shrink by at least a constant amount.
For the entanglement-testing problem, we use the above intuition to
derive a communication protocol which uses only a constant number of
qubits, and detects a maximally entangled state of arbitrary
dimension. This is described in Section \ref{sec:EPR-test}.
The idea is to determine whether Alice and Bob share a
state that is invariant under $U_i \otimes U_i^*$ for $i=1,\ldots,d$, or
far from invariant; i.e. whether the shared state is $\ket{\phi_D}$ or
something orthogonal. To achieve this with $O(1)$ communication,
suppose Alice and Bob each had access to a joint register initialized
with $1/\sqrt{d} \sum_{i=1}^d \ket{i}$. Each could then apply controlled
operators from this register to their share of the state $\ket{\psi}$:
Alice would apply a controlled $U_i$ and Bob a controlled $U_i^*$.
This ``shared register'' model could more naturally be achieved by
having Alice create the state and perform her controlled $U_i$ before
sending the state to Bob, who then performs his controlled $U_i^*$.
Bob should then test that the control state remained intact, which happens iff the original
state of the $D$-dimensional registers was indeed the maximally
entangled state. With only a little more algebra, this shows that for
any $D$, $\epsilon > 0$, there exists a protocol which uses $O(\log
1/\epsilon)$ qubits of communication, after which Bob always accepts
if the shared state is $\ket{\phi_D}$. If the shared state is
orthogonal to $\ket{\phi_D}$, he accepts with probability at most
$\epsilon$. Moreover, if Alice and Bob do start with the maximally
entangled state $\ket{\phi_D}$, the protocol does not damage the
state.
\begin{figure}[h]
\begin{center}
\includegraphics[width=9cm]{introfigure}
\caption{\small a) A counterexample to the generalized area law, consisting of a chain of complete graphs separated by the middle edge. The entropy across the cut is shown in \secref{concreteH} to
grow as $\Omega\left(n^{c}\right)$, where $n$ is the total number of particles. b) A four-particle construction, analyzed in \secref{HLMR}. c) Short-chain framework for proving 1-D area law, from \cite{AradKLV12area}.}
\label{fig-intro}
\end{center}
\end{figure}
For a counterexample to the generalized area law, we use the above
intuition to exhibit a gapped local Hamiltonian acting on the graph
featured in Figure~\ref{fig-intro}\,a), for which the entanglement
entropy of the ground state across the middle cut is $\Omega (n^c)$
for some $0 < c < 1$ (rather than $O(1)$ as predicted by the
generalized area law). The core step in generating this example is the
construction of a simpler system consisting of four particles on a
line in Figure~\ref{fig-intro}\,b): two particles of dimension $d = 3$
(qutrits) in the middle, and two particles of dimension $D$ at the two
ends, with arbitrarily large $D$. The gapped Hamiltonian is of the
form $H = H_L + H_M + H_R$, where $H_L$ acts between the left particle
and the left qutrit, $H_M$ between the two qutrits, and $H_R$ between
the right qutrit and the right particle. Crucially, the entanglement
entropy of the ground state across the middle cut is $\Omega(\log D)$,
as shown in Section~\ref{sec:HLMR}. The idea here, like in the
communication protocol, is to use the middle particles to synchronize
the application of a quantum expander on the left and right sides.
This requires only a single term of the Hamiltonian, acting on two
$d$-dimensional particles.
Enforcing a large amount of entanglement (in the ground state) by the
single constraint $H_M$ acting on a constant-dimensional system is a
surprising quantum phenomenon. In the analogous probabilistic
situation, consider a graph whise vertices are each associated with
constant-dimensional variables, and whose edges are associated with
classical constraints. Each constraint forbids some subset of the
possible assignments to the variables at the two ends of the
edge. This describes a constraint satisfaction problem
(CSP)\footnote{This analogy between local Hamiltonians and constraint
satisfaction problems is commonly used in quantum Hamiltonian
complexity, see e.g., \cite{AAV13}.}. Now consider the uniform
distribution over the set of all possible solutions to this set of
constraints, namely all assignments that violate no constraint. It is
easy to see that the middle constraint in the graph in
Figure~\ref{fig-intro}\,a) can only enforce a convex combination of a
constant number of product distributions \footnote{This phenomena can
be viewed as the zero temprature case; it in fact extends also to
the Gibbs distribution at any temprature, where the two endpoints of
a chain are always conditionally independent given the values of the
spins in the middle.}.
This simple example of a four-particle system is already important within the context of proofs of the 1-D area law and prospects for extending those techniques to higher dimensions.
The best current 1-D area law \cite{AradKLV12area} works
within a model very similar to our four-body Hamiltonian,
except the middle link in \cite{AradKLV12area} is extended into a finite chain of $s = \Omega(\log^2 d/\epsilon)$ particles, each of dimension $d$ (see figure~\ref{fig-intro}\,c).
This yields an area law bound of $S_{1-D} = O(\log^3(d) / \epsilon)$ across the middle cut. It was observed in \cite{AradKLV12area} that
any slight improvement in the exponent of $\log d$ would imply a non-trivial sub-volume law for 2-D systems. The crucial parameter in improving the result is the length of the middle chain;
Our four-body Hamiltonian shows that in the extreme case of a length $1$ chain,
no area law holds.
Understanding the intermediate regime is thus an important open question.
Our four-particle example involves non-physical particles of arbitrarily large dimension.
In Section~\ref{sec:concreteH} this example is converted to a counterexample
to the generalized area law with
bounded dimensional particles (albeit with unbounded degree of interaction).
This is done by applying Kitaev's circuit-to-Hamiltonian construction to implement the $U_i$,
followed by an application of the strengthening gadgets of \cite{NagajCao}
(see Section~\ref{sec:concreteH} for details).
What is the connection between our two results?
The above described Hamiltonian constructions are based on quantum expanders, just like our entanglement testers.
In Section~\ref{sec:comm} we explore a deeper connection between very efficient communication protocols for EPR testing and violations of generalized area laws. We show that it is possible to derive a counterexample to the generalized area law, starting from a solution to the first problem (an EPR testing protocol with limited communication) and converting it using Kitaev's circuit-to-Hamiltonian construction into a Hamiltonian violating the generalized area law. This, we believe, points at a fundamental link
between the two seemingly unrelated topics. We discuss this and many other open questions and related work
in Section \ref{sec:discussion}.
{\em Notation:} For a matrix $X$, let $X^*$ be the entry-wise complex conjugate of $X$ and $X^\dag$ the transpose of $X^*$. Define the Frobenius norm $|X|:=\sqrt{\tr X^\dag X}$; the operator norm $\|X\|$ is the largest singular value of $X$.
\section{Quantum Expanders}\label{sec:expanders}
The key structural component to our results are quantum expanders.
We will only make use here of expanders based on applying one out of $d$
unitaries at random
(a more general definition using Kraus operators exists).
\begin{definition}
The operator $\mathcal{E}: L(\ensuremath{\mathbb{C}} ^D) \rightarrow L(\ensuremath{\mathbb{C}}^D)$ (here we use $L(\ensuremath{\mathbb{C}}^D)$ to denote the set of linear operators on $\ensuremath{\mathbb{C}}^D$) is termed a {\it $(D, d, \lambda)$ quantum expander} if
\bitem
\item There are $d$ unitaries, $U_1, U_2,...,U_d$,
such that $\mathcal{E}(X) = \frac{1}{d} \sum_{i=1}^dU_i X U_i^\dagger $.
\item Interpreted as a linear map,
$\mathcal{E}$ has second-largest singular
value $\leq \lambda$.
\eitem
\end{definition}
Just as classical expanders may be thought of as constant-degree
approximations to the complete graph, quantum expanders are
constant-degree approximations to the application of unitaries drawn
at random from the
Haar measure.
By definition, the identity map $X=\mathbb{I}$
is the unique fixed point of $\mathcal{E}$.
The second condition is equivalent to saying
that for any $X$ with $\tr(X)=0$
\begin{eqnarray} | \mathcal{E} (X)| \leq \lambda |X|
\label{eq:frob-exp}.\end{eqnarray}
This interpretation suggests an alternate formulation
where we think of each $X\in L(\ensuremath{\mathbb{C}}^D)$ as a vector in $\ensuremath{\mathbb{C}}^D \otimes \ensuremath{\mathbb{C}}^D$ and the operator $\mathcal{E}$ then gets mapped to the
operator $\hat{\mathcal{E}} = \frac{1}{d} \sum_{i=1}^dU_i \otimes U_i^* $. Then $\hat{\mathcal{E}}$
fixes the maximally entangled state $\ket{\phi_D} =\frac{1}{\sqrt{D}}\sum_{x\in[D]} \ket{x}\ket{x}$, and has second largest singular value $\lambda$.
Quantum expanders were introduced independently in
\cite{Hastings-expander1} and \cite{BST07-expander-orig} although many of the relevant ideas
were implicit in \cite{AS04}. In \cite{Hastings-expander2}, it was proved that
taking $U_i$ for $i\in \{1,...,d\}$
to be Haar uniform results in a ``Ramanujan'' expander
with high probability; that is, $\lambda \approx 1/\sqrt{d}$.
Since random unitaries cannot be constructed
efficiently, other work~\cite{BST08-expander, H-cayley,
QuantumMargulisExpanders, HH09}
gave efficient constructions, in which the unitaries can be applied
by polynomial-size quantum circuits.
Essentially all of these constructions achieve $\log(d) = O(\log
1/\lambda)$.
For our communication protocols, we will need $d$ to be a variable
(since the error depends on it);
whereas for the area law counter example, we will take $d$ to be a
small constant. In what follows we will assume for simplicity of exposition that $d=3$ is possible,
although the smallest $d$ that has been verified is $d=8$ using
\cite{QuantumMargulisExpanders}.
Why are expanders relevant to our results? To understand the gap condition better,
let us see why
$\ket{\phi_D}$, the maximally entangled state on $\bbC^D \otimes \bbC^D$,
is a $+1$ eigenvector.
Observe that for any $D\times D$ matrix $X$,
we have $(X\otimes I)
\ket{\phi_D} =(I\otimes X^T) \ket{\phi_D}$.
Thus $\frac{1}{d}\sum_{i=1}^d (U_i \otimes U_i^*)\ket{\phi_D} =
\frac{1}{d}\sum_{i=1}^d (U_iU_i^\dag \otimes I)\ket{\phi_D} = \ket{\phi_D}$.
Since the second-largest singular
value of $\hat{\mathcal{E}}$ is $\lambda$, then we have
\begin{eqnarray} \L\|\hat{\mathcal{E}} - \proj{\phi_D} \right\| = \lambda \label{eq:exp-gap}.\end{eqnarray}
Thus, $\hat{\mathcal{E}}$ gives an approximation of a projector onto $\ket{\phi_D}$
up to operator-norm error $\lambda$.
In the rest of our paper we will explore two settings in which this
allows us to use resources proportional to $d$ (which we should think
of as small) to force a state on $\bbC^D \otimes \bbC^D$ (with $D$ large)
to be close to $\ket{\phi_D}$.
\bitem
\item In \secref{EPR-test} we will show how $\log(d)$ qubits of communication
can perform the projective measurement $\{\proj{\phi_D}, I -
\proj{\phi_D}\}$ up to error $1/d^{\Omega(1)}$.
\item In \secref{HLMR} we will show how interactions between a pair of
constant-dimensional
and a pair of two $D$-dimensional particles can have a ground state
with maximal entanglement on the $D$-dimensional particles and a constant gap.
\eitem
\section{A communication protocol for certifying global entanglement}
\label{sec:EPR-test}
\subsection{The EPR testing problem}
As above, set $\ket{\phi_D}$ to be
the maximally entangled state on $\bbC^D \otimes \bbC^D$. The EPR testing
problem is to determine whether a given shared state
$\ket\psi\in\bbC^D \otimes \bbC^D$ is equal to or orthogonal to
$\ket{\phi_D}$. More precisely, two parties (Alice and Bob) would
like to simulate the joint two-outcome POVM $\{\proj{\phi_D}, \bbI - \proj{\phi_D}\}$.
An $(D,\varepsilon)$ EPR tester is a communication protocol for performing a two-outcome measurement $\{M,\bbI-M\}$ such that $\|M - \proj{\phi_D}\|\leq \varepsilon$.
In general EPR testers may differ in a variety of ways:
\bitem
\item If $M\geq \proj{\phi_D}$ then we say the EPR tester has {\em perfect completeness}.
\item The communication requirements and computational efficiency may vary.
\item The protocol may be performed with quantum or classical communication. If quantum communication is used, then it is reasonable to assume that upon input $\rho$ the post-measurement state is $M^{1/2}\rho M^{1/2} / \tr[M\rho]$ or
$(\bbI-M)^{1/2}\rho (\bbI-M)^{1/2} / \tr[(\bbI-M)\rho]$, depending on the
outcome. If classical communication is used, we need to also consume
some entanglement. We say that the test consumes $k$ EPR pairs if given
an input of $n$ EPR pairs, it outputs at least $n-k$ EPR pairs (up to
$\varepsilon$ error) when it
reports success. There are no guarantees for orthogonal input states.
\eitem
We are aware of three previous implementations of EPR testers
(previously described in \tabref{ent-test}). Ref.~\cite{BDSW96} gave a
$\left(2^n,2^{-k}\right)$ EPR tester with perfect completeness that used a
message of $O(nk)$ classical bits and consumed $k$ EPR pairs. This
was improved by \cite{BCGST02} to a $(2^n,\frac{2n}{k(2^k+1)})$ EPR
tester that sent $2k$ classical bits, consumed $k$ EPR pairs and used
$\approx n/k$ bits of shared randomness. The paper \cite{HL07}
provides a protocol which uses only $\log 1/\varepsilon$ communication qubits, but with the
assistance of an additional $n/\varepsilon$ trusted EPR pairs. Our
protocol achieves a protocol with this amount of communication,
without the need for any extra resources.
\subsection{EPR testing with constant quantum communication}
\label{sec:EPRprotocol}
Our main result in this
section is an EPR tester using only a constant amount of quantum communication that is {\em
independent} of the dimension $D$.
\begin{theorem}\label{thm:EPR-test}
For any $D$ and any $\varepsilon>0$ there exists a $(D,\varepsilon)$ EPR tester with
perfect completeness using one-way communication from Alice to Bob.
The protocol has several variants:
\bitem
\item Using $(2 + o(1))\log(1/\varepsilon)$ qubits sent from Alice to Bob, but $\mathrm{poly}(D)$ run-time.
\item Using $C\log(1/\varepsilon)$ qubits sent from Alice to Bob and
$\mathrm{poly}\log(D)$ run-time for some universal constant $C>0$.
\item Using either $(8 + o(1))\log(1/\varepsilon)$ or $\approx\!
4C\log(1/\varepsilon)$ classical bits sent from Alice to Bob (depending on
whether computational efficiency is needed) and consuming the same
number of EPR pairs.
\item Using 2 bits of shared randomness and 1 qubit of communication,
$\mathrm{poly}\log(D)$ run-time and achieving $\varepsilon=\frac{8 + \sqrt 5}{16} \leq 0.64$.
\eitem
\end{theorem}
We remark that replacing the state $\ket{\phi_D}$ in \thmref{EPR-test}
with a general entangled state can result in a much larger (and
$D$-dependent) communication cost~\cite{HL07}. Thus we refer to the
result as an EPR tester rather than a general entanglement tester.
One application of this result relates to the open question of whether entanglement helps
quantum communication complexity. Classically, shared randomness
does not significantly reduce communication
complexity because large random strings can be replaced by
pseudo-random strings that fool protocols\cite{Newman91}. This is
called a blackbox reduction because it replaces the random input but
does not change the protocol. Quantumly such blackbox
reductions are ruled out by efficient entanglement-testing
protocols, since they cannot be fooled by any low-entanglement state.
A similar result is in \cite{JRS08} but their construction does not
yield an EPR tester. See also \cite{SZ08} for a non-blackbox entanglement
reduction that increases the communication cost by an exponential amount.
\begin{proof}[Proof of \thmref{EPR-test}]
The main idea is to interpret the results of \secref{expanders} as a way to
test maximally entangled states. By \secref{expanders} it
suffices for Alice and Bob to
implement a two-outcome measurement $\{M,\bbI-M\}$ on their shared state
with $M = \hat\mathcal{E}$ for $\mathcal{E}$ a $(D,k,\sqrt\varepsilon)$ expander.
However, it is not immediately clear how to implement this
measurement. To do this, we will use a trick that has been
used in a variety of contexts (e.g. \cite{BBDEJM96}, \cite{HL07} and Section 2.2.2 of
\cite{money-ICS}) and can be thought of as a variant of phase estimation.
The protocol (depicted in Figure~\ref{fig:commgames}a) is as follows.
\benum
\item Alice and Bob initially share a state in registers $L,R$.
\item Alice prepares the $\log(d)$-qubit state $\frac{1}{\sqrt{d}}\sum_{i=1}^d
\ket i$ in register $a$.
\item She performs $W=\sum_{i=1}^{d} \proj i \otimes U_i$ on
$a,L$.
\item She sends system $a$ to Bob.
\item Bob performs $W^* = \sum_{i=1}^d \proj i \otimes U_i^*$ on
$a,R$.
\item Bob does a two-outcome measurement on $a$, with the “accept”
outcome corresponding to the state $\frac{1}{\sqrt{d}} \sum_{i=1}^d \ket i$
and the ``reject'' outcome corresponding to the orthogonal subspace.
\eenum
If Alice and Bob start with the shared state $\ket\psi$, then step 5,
their state is
\begin{eqnarray}
\frac{1}{\sqrt d}\sum_{i=1}^d \ket{i}^a \otimes (U_i \otimes U_i^*)
\ket{\psi}^{LR}.\end{eqnarray}
Step 6 then accepts with probability equal to the norm squared of $\frac{1}{d}\sum_{i=1}^d (U_i \otimes U_i^*)
\ket{\psi} =: M\ket\psi$ where we have defined
\begin{eqnarray} M = \frac{1}{d}\sum_{i=1}^d U_i \otimes U_i^*.\end{eqnarray}
This results in the two-outcome measurement $\{M^\dag M, I - M^\dag M\}$.
By Eq.~\eq{exp-gap},
$M^\dag M$ is $\varepsilon$ close to the desired measurement operator
$\proj{\phi_D}$.
The communication cost is $\log(d)$. If we do not care about
computational efficiency, we can obtain $\varepsilon \approx 1/\sqrt{d}$
using random unitaries~\cite{Hastings-expander2}. For a
$\mathrm{poly}\log(d,D)$ run-time, we can iterate an efficient expander;
e.g. applying the construction of \cite{QuantumMargulisExpanders} $k$
times yields $d=8^k$ and $\varepsilon \leq (2\sqrt 5/8)^k$. To use classical
bits instead, we first use the construction of \cite{BCGST02} which
uses $O(\log 1/\varepsilon)$ classical bits to verify $O(\log 1/\varepsilon)$ EPR
pairs. Those EPR pairs are then used to teleport the qubits in the
above protocol, which can therefore be applied to verify the rest of
the EPR pairs.
For the last construction that uses two rbits and one qubit, we start
with the quantum Margulis expander~\cite{QuantumMargulisExpanders}.
This consists of unitaries $U_1,\ldots,U_4, U_5 = U_1^\dag, \ldots,
U_8=U_4^\dag$. The modified protocol is as follows. Let $r \in
\{1,2,3,4\}$ be the value of the shared randomness. Run the above
protocol with the pair of unitaries $\{I, U_r\}$. The resulting
measurement operator, conditioned on $r$, is $M_r = \frac{I + U_r \otimes
U_r^*}{2}$. Averaging over $r$ we obtain
$$M^\dag M := \frac{1}{4}\sum_{r=1}^4 M_r^\dag M_r
= \frac{I + \frac{1}{8} \sum_{i=1}^8 U_i \otimes U_i^*}{2}.$$
From the expansion properties proved in
\cite{QuantumMargulisExpanders}, we have that $\|M^\dag M -
\proj{\phi_D}\| \leq \frac 1 2(1 + \frac{2\sqrt 5}{8}) = \frac{8 +
\sqrt 5}{16}$.
\end{proof}
To help prepare for the next sections, it is useful to view
this test also in matrix form, as follows.
If the initial state of the left/right registers was $\ket{x}_L\ket{y}_R$, after Alice's
operation, the state has to have the form
\begin{align}
\frac{1}{\sqrt{d}}\sum_{i=1}^d \ket{i} (U_i \ket{x}_L) \ket{y}_R.
\end{align}
After Bob gets the ancilla and performs his operation $W^*$, the state has to have the form
\begin{align}
\frac{1}{\sqrt{d}}\sum_{i=1}^d \ket{i} \left(U_i \ket{x}_L\right) \left(U_i^* \ket{y}_R\right).
\end{align}
Let us represent the initial state $\ket{\psi}_{AB}=\sum_{k,\ell}x_{k,\ell}\ket{k,\ell}$ by a matrix $X$, such that $X_{k,\ell}= x_{k,\ell}$.
We now rewrite the final state as a matrix with components $\beta_{(a,L), R}$:
\begin{align}
\beta = \frac{1}{\sqrt{d}} \left[\begin{array}{c}
U_1 X U_1^\dagger\\
U_2 X U_2^\dagger\\
\vdots
\end{array}\right].
\end{align}
Passing the final test now means that
\begin{align}
U_i X U_i^\dagger = U_j X U_j^\dagger, \quad \forall i,j,
\end{align}
which is possible (if we have a quantum expander) only for $X=\mathbb{I}$.
This means the initial state was $\ket{\psi}_{LR} = \ket{\phi_D}$,
and that the final state is
$\left(\frac{1}{\sqrt{d}}\sum_{i=1}^{d} \ket{i}_a \right) \otimes \ket{\phi_D}$.
\section{A counterexample to the generalized area law}
\label{sec:HLMR}
In this Section we present our second result: a Hamiltonian with a small “bridge” term connecting two large halves of a system. Strikingly, this single-link bridge of constant dimensions has a large influence on the entanglement entropy between the two parts of the system, in the ground state.
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=1.4]
\tikzstyle{LabelStyle}=[fill=white,sloped]
\Vertex[x=0,y=0, style={minimum size=1.2cm}, LabelOut=false,L=$\Sigma_{L}$]{1}
\Vertex[x=1.2,y=0, LabelOut=false,L=$\sigma_{1}$]{2}
\Vertex[x=2.4,y=0, LabelOut=false,L=$\sigma_{2}$]{3}
\Vertex[x=3.6,y=0, style={minimum size=1.2cm}, LabelOut=false,L=$\Sigma_{R}$]{4}
\Edges[label={L}](1,2)
\Edges[label={M}](2,3)
\Edges[label={R}](3,4)
\end{tikzpicture}
\end{center}
\caption{\small A single-link chain with side operators $L$ and $R$.}
\label{fig-label}
\end{figure}
\medskip
\subsection{Background about local Hamiltonians}
We consider Hamiltonians on
finite-size spin systems, where each term in the Hamiltonian is a
bounded-strength interaction between a bounded number of spins; in
fact our constructions will involve only pairwise interactions.
Unlike many physical systems, we do not require spatial locality but
allow interactions between any pair of spins. We will also consider
systems in which individual spin dimension $d_i$ can be large.
A quantum state on $n$ particles, of dimensions $d_1,..,d_n$ respectively,
is a unit vector in $\mathcal{H} := \bbC^{d_1} \otimes \dots \otimes \bbC^{d_n}$. A
Hamiltonian $H$ is a Hermitian matrix acting on $\mathcal{H}$. A $k$-local
Hamiltonian can be written as $H = \sum_{i=1}^m H_i$ where each $H_i$
acts nontrivially on at most $k$ particles.
Conventionally,
each $\|H_i\| \leq 1$. If the $H_i$'s are all diagonal, $H$ is
equivalent to a classical constraint satisfaction problem; in
general the $H_i$ may not always be diagonal or even commute. The
eigenvector of $H$ with the smallest eigenvalue is called the ground state.
We say
that a Hamiltonian is {\em frustration free} if the ground state of $H$ is
also the eigenvector with lowest eigenvalue for each $H_i$; otherwise
we call $H$ frustrated.
If the eigenvalues of $\mathcal{H}$ are $E_0 \leq E_1
\leq \cdots \leq E_{D-1}$
then the {\em gap} of
$H$ is defined to be $E_1 - E_0$. When $H$ belongs to a family of
Hamiltonians indexed by $n$, we say this family is {\em gapped} if the gap
of each $H$ is lower-bounded by a constant independent of $n$.
(Otherwise the family is said to be {\em gapless}.) Often we identify
$H$ with the family of Hamiltonians, and simply say that $H$ itself is gapped
or gapless.
\subsection{Construction of the Hamiltonian}
\label{sec:Hconstruct}
Let the system $W$ consist of two qutrits ($\sigma_{1}$ and $\sigma_{2}$) and two high dimensional systems
($\Sigma_{L}$ and $\Sigma_{R}$):
\[
W = \Sigma_{L}\otimes \sigma_{1} \otimes \sigma_{2} \otimes \Sigma_{R} = C^{D}\otimes C^{3} \otimes C^{3} \otimes C^{D}.
\]
We design a gapped Hamiltonian $H = H_L+ H_M+ H_R$, where $H_L$ (left) $H_M$ (middle) and $H_R$ (right)
are projectors acting on $\Sigma_{L}\otimes \sigma_{1} $, $\sigma_{1} \otimes \sigma_{2}$ and $\sigma_{2} \otimes \Sigma_{R}$,
respectively, that defies the area law
through the cut $\Sigma_{L}\otimes \sigma_{1} \, \mid\, \sigma_{2} \otimes \Sigma_{R}$. For convenience we write all elements
of $W$ in the form
\[
\sum_{i,j\in[3]} |\psi_{i,j}\rangle|i\rangle |j\rangle,
\]
where $\psi_{i,j} \in \Sigma_{L} \otimes \Sigma_{R}$, and $\sum_{i,j\in[3]} |\psi_{i,j}|^2 = 1$. If we fix a basis in $\Sigma_{L}$ and $\Sigma_{R}$, respectively,
we can think of $\psi_{i,j}$ for every $i,j\in [3]$ as $D\times D$ matrices. Our construction will rely on
quantum expanders using $D\times D$ unitary matrices with $U_1=I$, and $U_{2}$ and
$U_{3}$ such that for any $D\times D$ matrix $X$ with
$|X|^2 = \tr(XX^\dag) = 1$, $\tr(X)=0$ we have:
\begin{eqnarray}\label{eq:qexp}
|\mathcal{E}(X)| =
\frac 1 3 |X + U_{2}XU_{2}^\dag + U_{3}XU_{3}^\dag| \le (1-c),
\end{eqnarray}
where $c:=1-\lambda>0$ is a fixed constant, independent of $D$.
Equation (\ref{eq:qexp}) and the triangle inequality imply that
for any $D\times D$ matrix $X$ with
$|X|= 1$, $\tr(X)=0$:
\begin{eqnarray}\label{eq:qexp2}
|U_{2}XU_{2}^\dag -X| + |U_{3}XU_{3}^\dag -X| \ge 3c
\end{eqnarray}
We now define projectors $H_L$, $H_R$ and $H_M$ via their zero subspaces
${\cal L}$, ${\cal R}$ and ${\cal M}$. We describe these subspaces by writing states of $W$ in the block matrix form
\[
\left(\begin{array}{lll}
\psi_{1,1} & \psi_{1,2} & \psi_{1,3} \\
\psi_{2,1} & \psi_{2,2} & \psi_{2,3} \\
\psi_{3,1} & \psi_{3,2} & \psi_{3,3} \\
\end{array}\right).
\]
Note that our way to present a (pure) state of $W$ is unlike the density matrix presentation, and it is only meaningful, because $W$ is a tensor product
of four components. The above matrix form (of a vector) is
simply a convenient way of rendering the $(3 D)^2$ coordinates of a state in $W$. In this presentation ${\cal L}$, ${\cal R}$ and ${\cal M}$
have convenient expressions.
${\cal L}$ is the set of states of the form
\[
\left(\begin{array}{rrr}
\psi_{1,1} & \psi_{1,2} & \psi_{1,3} \\
U_{2}\psi_{1,1} & U_{2}\psi_{1,2} & U_{2}\psi_{1,3} \\
U_{3}\psi_{1,1} & U_{3}\psi_{1,2} & U_{3}\psi_{1,3} \\
\end{array}\right),
\]
where $\psi_{1,1}$, $\psi_{1,2}$ and $\psi_{1,3}$ are arbitrary.
${\cal R}$ is the set of states of the form
\[
\left(\begin{array}{rrr}
\psi_{1,1} & \psi_{1,1}U_{2} & \psi_{1,1}U_{3} \\
\psi_{2,1} & \psi_{2,1}U_{2} & \psi_{2,1}U_{3} \\
\psi_{3,1} & \psi_{3,1}U_{2} & \psi_{3,1}U_{3} \\
\end{array}\right),
\]
where $\psi_{1,1}$, $\psi_{2,1}$ and $\psi_{3,1}$ are arbitrary.
${\cal M}$ is the set of states of the form
\[
\left(\begin{array}{lll}
\psi_{1,1} & X & Y \\
X & \psi_{2,2} & \psi_{2,3} \\
Y & \psi_{3,2} & \psi_{3,3} \\
\end{array}\right), \hspace{0.5in}
\]
where $X$, $Y$ and the remaining $\psi_{i,j}$'s are arbitrary.
It is easy to check that $H_L, H_M, H_R$ are indeed local.
For instance
${\cal M}$ is a tensor product of $\Sigma_{L} \otimes \Sigma_{R}$ with the
subspace $S$ of $\sigma_{1}\otimes \sigma_{2}$ that equates the coefficients of
$|1\rangle |2\rangle$ and $|2\rangle |1\rangle$ and also the
coefficients of $|1\rangle |3\rangle$ and $|3\rangle |1\rangle$.
Explicitly
\begsub{H-terms}
H_L &:= \bbI - \frac 1 3 \sum_{i,i' = 1}^3 U_i U_{i'}^\dag \otimes \ket i \bra i', \\
H_R &:= \bbI - \frac 1 3 \sum_{j,j' = 1}^3
\ket j \bra j' \otimes \left(U_j U_{j'}^\dag\right)^T, \\
H_M &:= \frac{(\ket{12}-\ket{21})(\bra{12}-\bra{21})}{2} +
\frac{(\ket{13}-\ket{31})(\bra{13}-\bra{31})}{2}.\label{HM}
\endsub
\subsection{The ground state is highly entangled}
\label{sec:entangledG}
\begin{lemma}\label{lem:ground1}
The unique
normalized ground state $\ket{G}$ of $H = H_L + H_R + H_M = (\bbI -
\Pi_{\cal L}) + (\bbI - \Pi_{\cal R}) + (\bbI - \Pi_{\cal M})$ written out
as a matrix is
\[
G = \frac{1}{3\sqrt{D}} \left(\begin{array}{lll}
I & U_{2} & U_{3} \\
U_{2} & U_{2}^{2} & U_{2}U_{3} \\
U_{3} & U_{3}U_{2} & U_{3}^{2} \\
\end{array}\right).
\]
It satisfies $H\ket{G} = H_L\ket{G} = H_M\ket{G} = H_R\ket{G} =0$.
\end{lemma}
\begin{proof}
Equation (\ref{eq:qexp2}) guarantees that $I$ is the only $D\times D$ matrix that commutes with
both $U_{2}$ and $U_{3}$. From this together with the above forms of
${\cal L}, {\cal R}$ and ${\cal M}$,
it follows that $\ket{G}$
is the only normalized state vector in ${\cal L} \cap {\cal R} \cap {\cal M}$.
\end{proof}
\begin{lemma}
The entanglement entropy of $\ket{G}$
along the $\Sigma_{L}\otimes \sigma_{1} \, \mid\, \sigma_{2} \otimes \Sigma_{R}$ cut
is
$\log_{2} D$.
\label{lem:entangledstate}
\end{lemma}
\begin{proof}
Let $\ket{Z}$ be any state on our-four particle system, and let $Z$ be its matrix notation.
By a direct calculation, the reduced density matrix of $\ket{Z}$ on the $\Sigma_L \otimes \sigma_1$ systems is exactly
$ZZ^\dag$. Since $G = \frac{1}{3\sqrt{D}} \sum_{i,j=1}^3 U_iU_j \otimes
\ket i \bra{j}$ (letting $U_1 := \bbI$), we have that for the ground state $\ket{G}$ the reduced density matrix
\begin{eqnarray}
GG^\dag = \frac{1}{3D} \sum_{i,i'=1}^3 U_iU_{i'}^\dag \otimes \ket i\bra{i'} .
\end{eqnarray}
To diagonalize $GG^\dag$, let $W := \sum_{i=1}^3 U_i \otimes \proj i$.
Then
\begin{eqnarray}
W^\dag (GG^\dag) W = \frac{\mathbb{I}}{D} \otimes \frac 1 3 \sum_{i,i'=1}^3 \ket i \bra{i'},
\end{eqnarray}
which has $D$ eigenvalues equal to $1/D$.
\end{proof}
\subsection{The Hamiltonian is gapped}
\label{sec:Hgapped}
\begin{lemma}
\label{lem:gap1}
Denote the energy gap above the ground space for the Hamiltonian
$H=H_L+H_M+H_R$ by $\Delta$. Then $\Delta\geq c/4$ with $c$ defined in \eq{qexp}.
\end{lemma}
First we state a Lemma about the spectrum of the sum of two projectors.
\begin{lemma}\label{lem:two-proj}
Let $P_1,P_2$ be projectors onto subspaces $V_1, V_2$. Let
\begin{eqnarray} \mu :=
\min\{\bra{\psi}P_2\ket{\psi} : \ket{\psi}\in V_1,
\braket{\psi}{\psi}=1\}
= \lambda_{\min}(P_1 P_2 P_1).\end{eqnarray}
Then the minimum eigenvalue of $\bbI-P_1 + P_2$ is $1-\sqrt{1-\mu}\geq
\mu/2$.
\end{lemma}
\begin{proof}
By Jordan's Lemma~\cite{Jordan75}, it suffices to consider the case
when
\begin{eqnarray} P_1 = \bpm 1 & 0 \\ 0 & 0 \epm
\qquad\text{and}\qquad
P_2 = \bpm \mu & \sqrt{\mu(1-\mu)} \\
\sqrt{\mu(1-\mu)} & 1-\mu\epm\end{eqnarray}
In this case $P_1 + P_2 = I - (1-\mu)\sigma_z + \sqrt{\mu(1-\mu)}\sigma_x$
which has eigenvalues $1 \pm \sqrt{1-\mu}$.
\end{proof}
\begin{proof}[Proof of \lemref{gap1}]
Let $V_{LR}$ be the ground space of $H_L + H_R$ and let $\tilde V_{LR}$ be the
subspace of $V_{LR}$ that is orthogonal to $\ket{G}$. Let
$P_{LR}, \tilde P_{LR}$ be the corresponding projectors and observe
that
\begin{eqnarray} H_L + H_R \geq \bbI - P_{LR}
\qand
H_L + H_R + \proj G \geq \bbI - \tilde P_{LR}.
\end{eqnarray}
Let $\ket\psi \in
\tilde V_{LR}$ be a unit vector. Since $\ket\psi\in V_{LR}$, we can write
it in matrix form as
\begin{align}
\ket{\psi} = \left(\begin{array}{rrr}
X & XU_{2} & XU_{3} \\
U_{2}X & U_{2}XU_{2} & U_{2}XU_{3} \\
U_{3}X & U_{3}XU_{2} & U_{3}XU_{3} \\
\end{array}\right). \label{eq:AXBform}
\end{align}
Since $\braket{G}{\psi}=0$ we additionally have $\tr[X]=0$.
From normalization we have $|X| = 1/3$. Now we calculate
\begin{eqnarray}
\mu := \bra{\psi}H_M \ket{\psi} =
\frac{|XU_2 - U_2X| +|XU_3 - U_3X|}{2}
\stackrel{\eq{qexp2}}{\geq} \frac{|X|}{2} 3c
= \frac{c}{2}
\label{eq:HM-exp-high}\end{eqnarray}
Setting $P_1=\tilde P_{LR}$ and $P_2 = H_M$ we can now apply
\lemref{two-proj} and find that the minimum eigenvalue of
$(\bbI - \tilde P_{LR}) + H_M$ is $\geq c/4$.
Finally, the second-smallest eigenvalue of $H_L+H_M+H_R$ is the
minimum of $\bra{\psi}H_L+H_M+H_R\ket{\psi}$ over all unit vectors
$\ket\psi$ satisfying $\braket{\psi}{G}=0$. For such a vector we have
\begin{eqnarray}
\bra{\psi}H_L+H_M+H_R\ket{\psi}
= \bra{\psi}(H_L+H_M+H_R+\proj G)\ket{\psi}
\geq \bra{\psi}\bbI - \tilde P_{LR} + H_M\ket\psi
\geq \frac c 4 \end{eqnarray}
Combined with \lemref{ground1} this shows that the gap is $\geq c/4$.
\end{proof}
\section{The abstract Hamiltonian can be implemented locally}
\label{sec:concreteH}
The Hamiltonian construction in Section~\ref{sec:HLMR} has very
interesting properties (a unique, very entangled ground state and a
constant gap), but the $H_L$ and $H_R$ terms act on particles of
arbitrary dimension. Alternatively, we can think of them as being
nonlocal Hamiltonians for a system of qubits. We now wish to
decompose them into local terms, acting on particles of dimension $O(1)$,
while retaining their desirable properties. This is done in two steps:
we first construct a Hamiltonian $H'_{LMR}$
with the desired
properties except the interactions are of polynomial strength, and then we correct this unphysical assumption and derive our final
Hamiltonian $H^{\text{gadget}}_{LMR}$.
We start by showing in Subsection \ref{sec:compute3} that $H_L$
can be made local. We do this using Kitaev's
circuit-to-Hamiltonian construction applied to the circuit computing
the application of the expander, padded with polynomially many
identity gates at the end of the computation. This derives a local
Hamiltonian $H^{\text{Kit}}_{L}$ with ground states very close to the
ground states of $H_L$ in Section~\ref{sec:HLMR}, tensored with some
state in an additional ancilla register. The price we pay in this construction
is an inverse polynomial gap instead of a constant one since Kitaev's
construction has an inverse polynomial gap. To avoid this, we multiply
the local interaction terms in $H_L^{\text{Kit}}$ by a polynomial
prefactor and arrive at a Hamiltonian $H'_{L}$ with a constant gap, as
stated in Claim~\ref{HclockClaim}. However, its terms have
polynomially large, unphysical norms.
Next, in Subsection~\ref{sec:double}, we show in
\thmref{HlocalProperties} that by using the above local construction
on
both sides of the $4$-particle chain of Section \ref{sec:HLMR},
without changing the middle interactions, we arrive at a Hamiltonian
whose strength of the middle interaction remains $O(1)$, while its unique
ground state retains all the desired properties of of the
four-particle Hamiltonian $H$ from Section~\ref{sec:HLMR}. Note that
the interaction terms which are not in the middle are still of
polynomial strength. This gives us a local Hamiltonian $H'_{LMR} =
H'_L + H'_R + H_M$ with a constant gap, and a unique, entangled ground
state, just as we had for $H$. We now wish to make the strength of the
interactions on both sides bounded as well.
In Theorem \ref{FinalClaim},
we decompose each high-norm local interaction term in $H'_L$ and $H'_R$ into many local, constant-norm terms, using the strengthening gadgets
of \cite{NagajCao}. Thus, we end up with a local Hamiltonian $H^{\text{gadget}}_{LMR}$ with all the desired properties of $H_{LMR}$ from Section~\ref{sec:HLMR}. We note that once again a price is to be paid:
in our final local, bounded-interactions Hamiltonian, each particle is involved in polynomially many 2-body interactions.
It remains open to make the degree of interaction bounded.
\subsection{Evaluating a quantum expander locally (3 computations in parallel)}
\label{sec:compute3}
Let us translate the Hamiltonian from Section~\ref{sec:HLMR} into a local one.
We start by mimicking $H_L$ by a sum of local terms.
The Hamiltonian $H_L$ acts on a space of dimension $3D$, and its ground states have form
\begin{align}
\ket{\Phi_x} = \frac{1}{\sqrt{3}} \sum_{i=1}^{3}
\ket{i} \otimes U_i \ket{x}. \label{phiX}
\end{align}
We will now enlarge our system and
find a local Hamiltonian $H'_L$, whose ground state
will be close to
\begin{align}
\ket{\Phi_x}\otimes \ket{w}, \label{extraTensor}
\end{align}
with $\ket{w}$ some state of an extra register.
\begin{claim}\label{cl:localexpander}
There exists a frustration-free local Hamiltonian with a constant spectral gap, set
on a chain of $2N+2$ constant-dimensional qudits, such that all ground
states are $\epsilon$-close to the form \eqref{extraTensor}, with
$\epsilon$ inverse polynomial in $N$. The local terms of the
Hamiltonian are of norm bounded by $O(\mathrm{poly}(N))$.
\label{HclockClaim}
\end{claim}
We prove Claim \ref{cl:localexpander} with $N'=\mathrm{poly}(n)$ qubits and
$5$-local interactions (in general geometry). This construction can
then be recast on a chain of $2N+2=\mathrm{poly}(n)$ qudits using \cite{AGIK}
or \cite{QMAcomplete1Dd8}.
There, the clock/data registers (with $N$ particles each) can be seen as sitting on top of each other, and pair clock/data particles into larger qudits. These will then sit on a chain $b_N$-$\cdots$-$b_1$-$\sigma_1$-$\sigma_2$-$b'_1$-$\cdots$-$b'_N$,
with qutrits $\sigma_1$, $\sigma_2$ in the middle.
We construct the Hamiltonian of Claim~\ref{HclockClaim} following Kitaev's Circuit-to-Hamiltonian construction \cite{KitaevBook}.
It allows us to write down a Hamiltonian whose ground states are the history states of a quantum computation $V$, i.e.
states of the form
\begin{align}
\ket{\psi^{\text{hist}}_{y}} = \frac{1}{\sqrt{T}} \sum_{t=0}^{T} \ket{t}_k \otimes V_{t} \dots V_1 \left( \ket{y}\otimes \ket{0}_q \right),
\end{align}
where $k$ is an extra ``clock'' register, $q$ is an ancilla register,
$\ket{y}$ is some initial state of a data register and $V_t$ are the gates of some circuit $V$, acting on the data register.
Our data register will contain $n$ data qubits (for simplicity, set $2^n = D$) and a ``control'' qutrit $a$.
We want to get the history state of the circuit $V$ with unitaries
\begin{align}
V_t = \sum_{i=1}^{3} \ket{i}\bra{i}_a \otimes U_{i,t},
\end{align}
for $t=1,\dots,\tau$. Here $U_{i,1},\ldots,U_{i,\tau}$ are the gates
that together implement $U_i$ from the quantum expander, including uncomputing
any changes to the ancilla register $q$ at the end. On top of this, we pad the circuit $V$ with many identity gates $V_t = \mathbb{I}$ for $\tau < t \leq T$, for some $T \gg \tau$, setting
$\epsilon = \frac{\tau}{T} = \frac{1}{\mathrm{poly}(n)}$.
We also require an extra clock register $k$ capable of locally implementing a clock with $T+1$ clock states, as well as an ancilla scratch register $q$.
The ground states (history states of $V$) for the new Hamiltonian $H'_L$
have form
\begin{align}
\ket{\Psi_x} &=
\frac{1}{\sqrt{T+1}} \sum_{t=0}^{T}
\ket{t}_{k} \otimes \label{groundL}
V_t \dots V_1 \left(\frac{1}{\sqrt{3}} \sum_{i=1}^{3} \ket{i}_a \ket{x}_{d} \ket{0\cdots 0}_{q} \right).
\end{align}
We will build $H^{\text{Kit}}_L = H_{\text{init}} + H_{\text{prop}}$ from two parts.
First, propagation-checking:
\begin{align}
H_{\text{prop}} &= \frac{1}{2} \sum_{t=1}^{T}
\left(
\ket{t-1}\bra{t-1}_k + \ket{t}\bra{t}_k
\right) \label{Hproj}
- \frac{1}{2} \sum_{t=1}^{T}
\left(
\ket{t-1}\bra{t}_k \otimes V_t^\dagger
+ \ket{t}\bra{t-1}_k \otimes V_t
\right).
\end{align}
Second, we need to ensure proper initialization by adding a projector that prefers a uniform superposition on the control qutrit when the clock register is $\ket{0}_k$ (we want all three computations to run on the same input).
Adding standard ancilla initialization-checking, we get
\begin{align*}
H_{\text{init}} &= \ket{0}\bra{0}_{k} \otimes \left[
\mathbb{I} - \ket{\alpha_3}\bra{\alpha_3}_{a}
+ \sum_{i=1}^{s} \ket{1}\bra{1}_{q_i} \right],
\end{align*}
with $\ket{\alpha_3} = \frac{1}{\sqrt{3}}\left(\ket{1}+\ket{2}+\ket{3}\right)$.
We can now implement the clock register and the corresponding projectors by a a 5-local,
unary clock with $T+1$ qubits \cite{KitaevBook}.
The Hamiltonian $H^{\text{Kit}}_L$ is positive-semidefinite, and frustration-free. It has a zero-energy state of the form \eqref{groundL} for any basis state $\ket{x}$ of the $n$ working qubits. Furthermore, the energy gap of $H^{\text{Kit}}_L$ to eigenstates with nonzero energy is $\Delta^{\text{Kit}}_L = \oo{T^{-2}}$ \cite{KitaevBook}.
Using the 1-D construction for a line of constant ($8$-dimensional) qudits from \cite{QMAcomplete1Dd8} based on \cite{AGIK}, which also has a gap that scales as an inverse polynomial in $T$, this results in a 1-D Hamiltonian with the properties we want.
Let us consider the ground states more closely.
For $t\geq \tau$,
the data register is in the desired state $\ket{\Phi_x}$ \eqref{phiX},
the ancilla register is uncomputed, and it is only the
clock register that changes. Recalling $T\gg \tau$, we realize that each $\ket{\Psi_x}$ can be rewritten as
\begin{align}
\ket{\Psi_x} &= \frac{1}{\sqrt{T}}
\sum_{t=1}^\tau \ket{\varphi_{x,t}}
+ \ket{\Phi_x}_{cd}\otimes \frac{1}{\sqrt{T}} \left(\sum_{t=\tau+1}^T \ket{t}_k \right)
\otimes \ket{0\cdots 0}_q
\nonumber\\
&= \sqrt{\epsilon} \, \ket{v_{x}} +
\sqrt{1-\epsilon} \, \ket{\Phi_{x}}_{cd} \otimes \ket{w}, \label{psiL}
\end{align}
with some normalized
vectors $\ket{v_x}$ and $\ket{w}$.
Each ground state is thus as close to $\ket{\Phi_x}\ket{w}$ \eqref{extraTensor} as we want, because we are free to choose $T$ as large a polynomial
as we want, making
$\epsilon=\tau/T$ an inverse polynomial as small as we want.
The gap of the Hamiltonian $H^{\text{Kit}}_L$ is however not constant.
We rescale the interaction strengths of all terms in $H^{\text{Kit}}_L$ by $T^2$ (or by a higher polynomial in $T$ for the 1-D construction),
and look at $H'_L = \mathrm{poly}(T) \cdot H^{\text{Kit}}_L$. This new $H'_L$ satisfies the requirements of
Claim~\ref{HclockClaim}.
\subsection{A local Hamiltonian with an entangled ground state}
\label{sec:double}
We now take two copies of the system from the previous Subsection, and
construct a Hamiltonian $H'_{LMR} = H'_L + H'_R + H_M$. We keep the
same two-qutrit middle term $H_M$ from Eq.~(\ref{HM}) in
\secref{HLMR}, but will replace the left and right terms $H_L,H_R$
with the construction from Subsection~\ref{sec:compute3}.
\begin{theorem}
\label{thm:HlocalProperties}
The 1-D qudit Hamiltonian $H'_{LMR} = H'_L + H'_R + H_M$ with terms of
norm $\mathrm{poly}(n)$ has a unique ground state, whose entanglement accross the middle cut
is at least $\Omega(\log(D))$, and a constant energy gap.
\end{theorem}
Unlike in \secref{HLMR}, this Hamiltonian is no longer frustration
free. However, a qualitatively similar version of the argument from
that section will work. One change is that we will work with an
approximate ground state. Define
\begin{align}
\ket{G'} &:= \frac{1}{\sqrt{D}}\sum_{x=1}^D \ket{\Phi_x}\ket{\Phi_x}\ket{w}\ket{w}
\label{groundLMR}\\
&=
\frac{1}{3\sqrt{D}} \left[\begin{array}{rrr}
\mathbb{I} & U_2 & U_3 \\
U_2 & U_2^2 & U_2 U_3 \\
U_3 & U_3 U_2 & U_3^2
\end{array}\right] \otimes \ket{w}\otimes\ket{w} =\ket{G}\ket{w}\ket{w},
\nonumber
\end{align}
with $\ket{\Phi_x}$ from \eqref{phiX}, $\ket{G}$ from
Lemma~\ref{lem:ground1}, and $\ket{w}$ a state of an ancilla register.
Thus the state $\ket{G'}$ is exactly the ground state we have in
Section~\ref{sec:entangledG}, with ancilla states added. It is not
the ground state of $H'_{LMR}$, nor can we even prove that it has low
energy. However, we will later construct a state $\ket{G_\varepsilon'}$ that
both has low enough energy to be close to the true ground state, and
is close enough to $\ket{G'}$ to have large entanglement.
The rest of our argument breaks up into the following subsidiary claims.
\begin{claim}\label{clm:GHG-small}
$\bra{G_\varepsilon'}H'_{LMR}\ket{G_\varepsilon'} \leq 1/\mathrm{poly}(n)$ (with
$\ket{G'_\varepsilon}$ defined later).
\end{claim}
\begin{claim}\label{clm:gap-large}
The second-smallest eigenvalue of $H'_{LMR}$ is $\geq
\Omega(1)$, implying that the gap is large.
\end{claim}
\begin{claim}\label{clm:entangled}
The ground state of $H'_{LMR}$ has large overlap with
$\ket{G'}$, and therefore high entanglement.
\end{claim}
We begin by showing a precise sense in which $H_L', H_R'$ give an
approximation of $H_L,H_R$.
Define $H'_{LR}:=H'_L+H'_R$
to be the Hamiltonian acting on the two
sides of the chain without interaction. The ground space
of $H'_{LR}$ is spanned by basis states of the form
\begin{align}
\ket{\Psi_x}\ket{\Psi_y} &=
\sqrt{\epsilon(2-\epsilon)}\, \ket{z^{\epsilon}_{xy}}
+ (1-\epsilon) \ket{\Phi_x}\ket{\Phi_y}\ket{w}^{\otimes 2},
\label{groundLReps}
\end{align}
where $\ket{\Psi_x}$ and likewise $\ket{\Psi_y}$ are given by
\eqref{psiL}, and $\epsilon = \frac{\tau}{T}$ is an inverse polynomial
which we can make as small as we want by increasing $T$ to a large
polynomial in $n$.
\begin{definition}\label{defs0} {\bf $S_0, S_\epsilon$:}
Define $S_0$ to be the space spanned by states of the form
$\ket{\Phi_x}\otimes\ket{\Phi_y}$, where $\ket{\Phi_x}$ (and likewise
$\ket{\Phi_y}$) are defined in \eqref{phiX}. Define $S_\epsilon$ to
be the space spanned by all states of the form with $\ket{\Psi_x}
\otimes \ket{\Psi_y}$, defined in \eqref{psiL}. Define the
corresponding projectors to be $P_0, P_\varepsilon$.
\end{definition}
\begin{claim} \label{clm:s0}
Let $\ket{s_\epsilon}$ be any state in
$S_\epsilon$; then there exists a state $\ket{s_0}\in S_0$ such that
\begin{equation}
\ket{s_\epsilon}=\sqrt{1-\epsilon'^2}\ket{s_0}\ket{w}\ket{w}+\epsilon'\ket{\epsilon},
\end{equation}
for $\epsilon'\le 2\epsilon$, and $\ket{\epsilon}$ orthogonal to
$\ket{s_0}\ket{w}\ket{w}$. As a result \begin{eqnarray} \frac{1}{2}\norm{
\proj{s_0} - \proj{s_\varepsilon}}_1 \leq 2\varepsilon .
\label{eq:s0-seps}\end{eqnarray}
\end{claim}
\begin{proof}
The proof follows by direct calculation, using Definition
\ref{defs0} of $S_0$ and $S_\epsilon$ and the observation that the
normalized error vectors $\ket{z_{xy}^\epsilon}$ from \eqref{psiL}
are all orthogonal for different pairs of $x,y$. This follows from
their definition as history states of different initial vectors, as
seen in \eqref{psiL}.
\end{proof}
As a direct consequence of \clmref{s0} we establish
\clmref{GHG-small}. Indeed, $\ket{G'}\in S_0 \otimes \ket w \otimes \ket w$
and by \clmref{s0} there exists an $\varepsilon$-close state $\ket{G_\varepsilon'}\in S_\varepsilon$ in
the ground space of $H_{LR}'$. Thus
\begin{eqnarray} \bra{G'_\varepsilon}H_L' + H_M + H_R'\ket{G'_\varepsilon} =
\bra{G'_\varepsilon}H_M\ket{G'_\varepsilon} \leq 2\varepsilon \leq 1/\mathrm{poly}(n),
\label{eq:Geps-low-E}\end{eqnarray}
where in the last step we have used \eq{s0-seps} and $0\leq H_M\leq I$.
Now we consider gap.
The Hamiltonians $H'_L$ and $H'_R$ act on independent subspaces, while
both Hamiltonians have a constant energy gap above the ground state
subspace as we proved in Claim \ref{cl:localexpander}.
Therefore, $H'_{LR}$ has constant gap. Let us denote this
gap by $\Delta'_{LR}$, so that (using also the fact that $H_{LR}'$ has
lowest eigenvalue 0) we have the operator inequality
\begin{eqnarray} H'_{LR} \geq \Delta'_{LR} (\bbI - P_\varepsilon).
\label{eq:prime-op-ineq}\end{eqnarray}
Continuing along the lines of the proof of \lemref{gap1}, define
$\tilde P_\varepsilon := P_\varepsilon - \proj{G_\varepsilon'}$, $\tilde P_0 = P_0 - \proj{G}$
and define $\tilde S_\varepsilon, \tilde S_0$ to be their supports.
Now calculate
\begsub{mu-prime}
\mu &:= \min \{\bra{\psi}H_M\ket{\psi} : \ket\psi \in \tilde S_\varepsilon,
\braket{\psi}{\psi}=1\} \\
&\stackrel{\eq{s0-seps}}{\geq}
\min \{\bra{\psi}H_M\ket{\psi} : \ket\psi \in \tilde S_0\otimes\ket{w}\otimes\ket{w},
\braket{\psi}{\psi}=1\} - 2\varepsilon\\
&\stackrel{\eq{HM-exp-high}}{\geq} \frac{c}{2} - 2\varepsilon
\endsub
Denote the second-smallest eigenvalue of a Hermitian matrix $X$ by
$\lambda_2(X)$. A variant of the Courant-Fischer min-max principle gives
the following variational characterization of $\lambda_2(X)$:
\begin{eqnarray} \lambda_2(X) = \sup_{\ket{v}} \lambda_{\min}(X + \proj v).\end{eqnarray}
We can apply this to our problem by observing that
\ba
\lambda_{2}(H_{LMR}')
& = \sup_{\ket{v}} \lambda_{\min}(H_{LMR}' + \proj{v})
\\ & \geq \lambda_{\min}(H_{LMR}' + \Delta'_{LR} \proj{G_\varepsilon'})
\\ & \stackrel{\eq{prime-op-ineq}}{\geq}
\Delta'_{LR} \lambda_{\min}(\bbI - \tilde P_\varepsilon + H_M)
\\ & \stackrel{\text{\lemref{two-proj}}}{\geq}\;
\frac{c}{4}-\varepsilon \geq \Omega(1)
\ea
This establishes \clmref{gap-large}.
To complete the proof, we need to show that the ground state is highly
entangled. Two challenges which complicate the usual continuity
arguments are that $H_{LMR}'$ has a large norm and a large ancilla
dimension. We sidestep these as follows.
Let $\ket{\gamma}$ denote the ground state of $H_{LMR}'$. Adjust its
overall phase so that $\braket{\gamma}{G_\varepsilon'}$ is real, implying
$\ket{G_\varepsilon'} = \sqrt{1-\delta}\ket\gamma +
\sqrt{\delta}\ket{\gamma^\perp}$ for some orthogonal state
$\ket{\gamma^\perp}$. Then
\begin{eqnarray} \varepsilon \stackrel{\eq{Geps-low-E}}{\geq}
\bra{G_\varepsilon'}H'_{LMR}\ket{G_\varepsilon'} =
(1-\delta)\bra{\gamma}H'_{LMR}\ket{\gamma} +
\delta\bra{\gamma^\perp}H'_{LMR}\ket{\gamma^\perp}
\geq \delta \L(\frac c 4 -\varepsilon\right),\end{eqnarray}
where this last inequality follows from the fact that $H_{LMR}'$ is
positive semi-definite and has gap $\geq c/4-\varepsilon$. Thus $\delta \leq
5\varepsilon/c$ (assuming $\varepsilon \leq c/20$).
Combining this with previous facts we have \begin{eqnarray} \ket\gamma
\approx_\delta \ket{G_\varepsilon'}\approx_{2\varepsilon} \ket{G'} =\ket{G}\otimes\ket w
\otimes \ket w\label{eq:approx-chain}\end{eqnarray} and the latter state is highly
entangled according to \lemref{entangledstate}. Assume WLOG that
$\braket{\gamma}{G'}$ is real and nonnegative. Thus $\braket{g}{G'}
\geq 1-\varepsilon'$ for some $\varepsilon' = \mathrm{poly}(1/n)$. However, the large
dimension of the ancilla states means we cannot directly use Fannes's
inequality. Let the Schmidt decomposition of $\ket{\gamma}$ be
$$\ket\gamma = \sum_i \sqrt{\lambda_i} \ket{L_i} \otimes\ket{R_i},$$
with $\lambda_1 \geq \lambda_2 \geq \ldots \geq 0$ and $\sum_i
\lambda_i=1$.
Then
$$1-\varepsilon' \leq \braket{\gamma}{G'}
\leq \frac{1}{\sqrt{D}} \sum_{i=1}^D \sqrt{\lambda_i}.$$
Let $r$ be the largest $r$ for which $\lambda_r \geq \kappa/D$ for
$\kappa>1$ to be chosen later. By normalization, $r \leq D/\kappa
\leq D$. Let $\beta = \sum_{i\leq r} \lambda_i$.
Then
$$\sqrt{D}(1-\varepsilon') \leq \sum_{i=1}^D \sqrt{\lambda_i}
\leq \sqrt{\frac{D}{\kappa}}\sum_{i\leq r}\lambda_i +
\sqrt{D}\sum_{i=r+1}^D \lambda_i
\leq \sqrt{D}(\beta\kappa^{-1/2} + (1-\beta)).$$
The second inequality follows from $\sqrt{\lambda_i}\leq
\sqrt{\frac{D}{\kappa}}\lambda_i$ for $i\leq r$ in the first term and
Cauchy-Schwarz in the second term.
Rearranging we find that $\beta \leq \varepsilon'(1-\kappa^{-1/2})$.
We conclude that the entropy of entanglement is
$$\sum_i \lambda_i \log\L(\frac{1}{\lambda_i}\right)
\geq \sum_{i>r}\lambda_i \log\L(\frac{D}{\kappa}\right)
\geq (1-\varepsilon'(1-\kappa^{-1/2})) \log\L(\frac{D}{\kappa}\right).$$
Optimizing over $\kappa$ we find that the entanglement is
$(1-o(1))\log(D)$.
This concludes the proof of \clmref{entangled} and therefore \thmref{HlocalProperties}.
\subsection{Decomposing the Hamiltonian $H'_L$ into $O(1)$-strength interaction terms}
\label{sec:gadgets}
We now handle the problem of large interaction norm. The interactions
in the Hamiltonian $H'_L$ have norm
$\mathrm{poly}(T)$. Each such term can be decomposed using the
strengthening quantum gadget construction by Nagaj and
Cao~\cite{NagajCao}, into $\mathrm{poly}(T)$
bounded-strength interactions acting on the original set of qudits
plus $\mathrm{poly}(T)$ extra ancilla qubits. The gap of this new $H'_L$ will
remain a constant, while any state in its ground state will now be
$1/\mathrm{poly}(T)$ close to some $\ket{\Psi_x} \ket{0\cdots 0}_{\textrm{new
ancillas}}$, with $\ket{\Psi_x}$ from \eqref{psiL}. This also
implies that each (less than a small constant energy) state of $H'_L$
is $1/\mathrm{poly}(T)$ close to the state $\ket{\Phi_x} \ket{w} \ket{0\cdots
0}_{\textrm{new ancillas}}$ for some $x$. However, this is just what
we had in \eqref{extraTensor}, with an expanded ancilla register state
$\ket{w'} = \ket{w}\ket{0 \cdots 0}$. Therefore, all of the arguments
of Section~\ref{sec:double} go through, and we have shown that
\begin{theorem}
There exists a 2-body Hamiltonian on $n$ qudits, whose terms are of
$O(1)$ norm. The interaction graph is as in Figure~\ref{fig-intro},
where the two particles on the two sides of the cut
are qutrits. All particles are involved in at most
$\mathrm{poly}(n)$ interactions. Moreover,
the Hamiltonian is gapped with a unique ground state,
such that the entanglement entropy across the middle
cut scales as $\Omega\left(n^{c}\right)$ for some $0<c<1$.
\label{FinalClaim}
\end{theorem}
\section{Entanglement testing and ground states of Hamiltonians}
\label{sec:comm}
\begin{figure}
\begin{center}
\includegraphics[width=13cm]{HtoEPRtest}%
\caption{EPR testing procedures. a) The test from Section~\ref{sec:EPRprotocol}.
b) A modified test with two ancillas.
c) The Hamiltonian from Section~\ref{sec:HLMR} can be also easily recast as an EPR testing protocol.}%
\label{fig:commgames}\end{center}
\end{figure}
We now connect our two results more directly, by providing an alternative derivation of the results in Section \ref{sec:HLMR}.
Starting from the
EPR testing protocol
of Section~\ref{sec:EPRprotocol}, we turn it into a non-local Hamiltonian violating
the generalized area law, using
Kitaev's circuit-to-Hamiltonian construction.
In fact, we use a slight variant
of the EPR testing protocol, which uses two ancillas (see
Figure~\ref{fig:commgames}b), as it translates to a Hamiltonian more easily.
We first describe the modified EPR testing protocol.
Alice has two registers, $L\otimes a_L$, and Bob has two registers denoted
$a_R\otimes R$, where $R,L$ are of large dimension $D$ and
$a_R$,$a_L$ are of constant dimension $d$.
They wish to check whether their joint state $\ket{\psi}_{LR}$ on registers $L \otimes R$ is maximally entangled.
First, Alice and Bob pre-share a maximally entangled state on $a_L\otimes a_R$:
\begin{align}
\ket{\phi_{d}} = \frac{1}{\sqrt{d}}\sum_{i=1}^{d}\ket{i}_{a_L}\ket{i}_{a_R} \label{entancillas}
\end{align}
Second, Alice applies the unitary $W=\sum_{i=1}^{d} \proj i \otimes U_i$
to $a_L \otimes L$, and Bob applies $W^*$ to $a_R\otimes R$.
Finally, they apply a projective measurement on $a_L\otimes a_R$
of the state $\ket{\phi_d}$ \eqref{entancillas}.
It is not difficult to see that this too is an EPR testing protocol; the
test passes with probability close to 1 if and only if the original state $\ket{\psi}_{LR}$
was very close to the maximally entangled state.
To encode this protocol into a Hamiltonian
via the circuit-to-Hamiltonian construction, we use {\it two} independent, two-step clocks.
(We will think of the circuit $W$ as well as $W^*$ as applied in a single
time step).
The Hamiltonian will thus act on four registers, $L,R$ and two enlarged
registers, $A_L=a_L\otimes C_L$ and $A_R= a_R\otimes C_R$ with $C_L,C_R$
being the two $2$-dimensional spaces of the two clocks, respectively.
We write the basis states of $A_L, A_R$ as $\ket{0,i}$ and $\ket{1,i}$
for $i\in \{1,..,d\}$.
The Hamiltonian consists of the following terms. An ``initialization'' and ``output'' term
on $A_L\otimes A_R$:
\begin{align}
H_{M} &= \sum_{s = 0}^{1}\sum_{i,j = 1}^{d} \ket{s,i}\bra{s,i}_{A_L}\otimes
\ket{s,j}\bra{s,j}_{A_R}
- \sum_{s = 0}^{1}\sum_{i = 1}^{d} \ket{s,i}\bra{s,i}_{A_L}\otimes \ket{s,i}\bra{s,i}_{A_R}
\end{align}
whose ground states have the form $\ket{1,i}\ket{0,j}$ and $\ket{0,i}\ket{1,j}$ for any $i,j$, but more importantly
$\frac{1}{\sqrt{d}}\sum_{i=1}^{d} \ket{0,i}\ket{0,i}$ and $\frac{1}{\sqrt{d}}\sum_{i=1}^{d} \ket{1,i}\ket{1,i}$.
These two states are maximally entangled states of the ancillas when
the ``clocks'' are both $0$ (initialization) or both $1$ (output).
Second, we have the ``left-computation-checking'' Hamiltonian, which acts on the registers $a_L$ and $L$:
\begin{align}
H_L &= \frac{1}{2} \sum_{i=1}^{d} \left(\ket{0i}\bra{0i} +
\ket{1i}\bra{1i}\right)_{a_L} \otimes \mathbb{I}_{L}
- \frac{1}{2} \sum_{i=1}^{d} \ket{1i}\bra{0i} \otimes W
- \frac{1}{2} \sum_{i=1}^{d} \otimes
\ket{0i}\bra{1i} \otimes W^\dagger.
\end{align}
Similarly, we define $H_R$, the ``right-computation-checking'' Hamiltonian which acts on $A_R$ and $R$, replacing $W$ by $W^*$ and $W^\dagger$ by $W^T$:
\begin{align}
H_R &= \frac{1}{2} \sum_{i=1}^{d} \left(\ket{0i}\bra{0i} +
\ket{1i}\bra{1i}\right)_{a_R} \otimes \mathbb{I}_{R}
- \frac{1}{2} \sum_{i=1}^{d} \ket{1i}\bra{0i} \otimes W^*
- \frac{1}{2} \sum_{i=1}^{d} \ket{0i}\bra{1i} \otimes W^T.
\end{align}
The final Hamiltonian, $H=H_{M}+H_{L}+H_{R}$ is our desired counterexample.
We claim that its unique, frustration-free ground state
is the ``history'' state
\begin{align}
\ket{\Psi} &=
\frac{1}{\sqrt{d}} \sum_{i=1}^{d}
\left(
\ket{0,i}_{a_L} + \left(W\otimes \mathbb{I}\right) \ket{1,i}_{a_L}
\right)
\left(
\ket{0,i}_{a_R} + \left(\mathbb{I} \otimes W^* \right) \ket{1,i}_{a_R}
\right)
\ket{\phi_{D}}_{LR}. \nonumber
\end{align}
It is not difficult to check that this is a maximally entangled state of
dimension $dD$, by observing that the Schmidt rank is $dD$
and the coefficients are uniform.
\section{Discussion, related work and Open Questions}
\label{sec:discussion}
We have shown that in both the context of EPR testing, as well as
ground states of Hamiltonians, constant resources suffice to enforce
what seems to be a global property. Our results are reminiscent in
spirit to the classical PCP theorem, or more generally to property
testing. The common theme is that a small amount of resources
(bits checked, Hamiltonian interactions, qubits transmitted, etc.)
serve to verify the properties of some large object. However, the fact
that such highly non-local properties as global entanglement can be
detected using local resources seems rather counter-intuitive, and
calls for further investigation in other contexts.
Our results leave many questions open. Below we discuss them as well as the broader context of these results.
{~}
\noindent{\bf The Area Law question}
Of course, the major open question of the 2-D area law, which was the
main motivation for this work, is left wide open. A more modest goal
would be to reduce the degree in our construction to a constant. Such
a step already seems to require significant progress in our
understanding of related notions, e.g., parallel
circuit-to-Hamiltonian constructions (see e.g.,\cite{RecentTerhal}),
and quantum expanders which are geometrically constrained, as well as
the notion of quantum degree reduction, as a possible route towards
quantum PCP~\cite{AAV13}. Alternatively, it might be true that the
generalized area law does in fact hold with bounded-degree
bounded-strength Hamiltonians.
Indeed, such a conjecture is not unplausible, and could potentially be
motivated by the following intuition. The area law had been long
believed, without proof, to be related to another very important
physical property of gapped Hamiltonians: the exponential decay of
correlations in the ground state. This means that a Hamiltonian has an
associated correlation length $\xi$ such that
$|\ip{AB}-\ip{A}\ip{B}| \leq \|A\| \, \|B\| \, e^{-\ell/\xi}$ where
$\ip{X}:=\bra{G} X \ket{G}$ and $A,B$ are observables separated by a distance $\ell$.
Such an exponential decay is
known to hold on a lattice of any constant dimension, and in fact in
any constant-degree graph~\cite{HastingsK06}
\footnote{Why don't our constructions contradict this, since they will
have large amounts of entanglement in the ground state? Our
large-dimension construction in \secref{HLMR} does fit the criteria
of \cite{HastingsK06} to have constant correlation length, but there
the entire graph has constant diameter. Our construction in
\secref{concreteH} has large degree and so the correlation-length
bound from \cite{HastingsK06} is also growing with the system size.}
It is perhaps natural to conjecture that if correlations in gapped
Hamiltonians are in this way ``local'', entanglement is also local;
One way to quantify this is with the area law conjecture. However, we
stress that only in 1-D this implication is known to hold~\cite{BrandaoH12}.
Our results (in particular \thmref{HlocalProperties}) provide a
counterexample to another possible version of the area law: one in
which the interaction degree is bounded, but the norms of the
interactions are required to be bounded only accross the cut, and
otherwise they can be polynomially large. The rationale of this
condition is that large norm terms on each side of the cut should only
increase the entanglement within the two regions on each side of the
cut and therefore by monogamy of entanglement only {\it decrease} the
entanglement across the cut. Our counterexample suggests that the
above monogamy-of-entanglement argument is too naive.
Another possibile version of the area law that might still hold is that a subsystem with
dimension $d_i$ at distance $\ell_i$ from the cut can contribute at most
$\log(d_i) e^{-\ell_i/\xi}$ entanglement, where $\xi$ is the correlation
length.
Attempting to strengthen our counter-example may either rule
these conjectures out or, in failing to do so, give a hint of how
they might be proved.
We note that the implications of an(y of the above forms of an) area
law for general systems are not yet fully understood. For ground
states of gapped one-dimensional systems, proving an area law was an
important step towards proving that they can be efficiently
described~\cite{Hastings09} and that these descriptions can be found
efficiently~\cite{AradKLV12area,LandauVV14}. For Hamiltonians on
general graphs, a partition into pieces with subvolume entanglement
scaling (i.e. region $S$ has $\leq \varepsilon |S|$ entropy) would imply a
classical description accurate enough to be incompatible with the
quantum PCP conjecture~\cite{BH-product}. Since entropy is a way to
count effective degrees of freedom, another interpretation of area
laws is that a quantum system can equivalently be represented by a
theory living on its boundary. This idea is known as the holographic
principle, and is currently a major conjecture in quantum field
theory~\cite{Swingle12}. It remains to be clarified whether an area law of any of
the suggested forms can lead to a more succinct description of the
ground states.
{~}
\noindent{\bf Related work}
There are several related works that we would like to mention. First,
Gottesman and Hastings \cite{GottesmanH10}, Irani \cite{Irani10}, and
Movassagh and Shor \cite{MS14} have examined qudit chains with highly
entangled ground states for Hamiltonians whose gaps are
inverse-polynomial. To the best of our knowledge, our results cannot
be derived in a straightforward manner from these works. Here, we
focus on spin chains with a constant gap. One can attempt to get a
constant-gap version of the above constructions by using the
strengthening gadgets of Nagaj and Cao~\cite{NagajCao} as we did in
this paper. However, this fails to provide the desired counterexample,
since these gadgets introduce a complicated geometry of interactions,
and we would need to apply them for every edge. Thus, the size of the
cut in the resulting graph would no longer be small. It is crucial
that in our present construction, the middle link is unchanged; only
the rest of the interactions need to be strengthened by gadgets.
We mention another relevant prior work \cite{Hastings-expander1, Hastings-notes}, which
described a state on a one-dimensional chain with a large amount of
entanglement across cuts (say $\log n$) but only short-range correlations. The
claim about decaying correlations here is rather subtle: two regions
that are separated by a distance $\ell$ from each other and $\ell'$
from the boundary have correlation no greater than $e^{-\ell/\xi}(1 + n
e^{-\ell'/\xi})$. In this way it avoids contradicting the relation
between decaying correlation and area law from \cite{BrandaoH12}. See
\cite{Hastings-notes} for further discussion. This result is
incomparable to ours because the states in question are not ground
states of a gapped $O(1)$-local Hamiltonian.
{~}
\noindent{\bf More general implications}
Finally, we believe that our results point at a potentially useful
link between two seemingly unrelated topics. Our paper shows that a
counterexample to the generalized area law can be derived from an
entanglement testing protocol of limited communication and converting
it into a Hamiltonian using Kitaev's circuit-to-Hamiltonian
construction. Our area-law violating Hamiltonian can be viewed as a
``tester'' of its highly entangled ground state, where the norm of the
Hamiltonian terms along the cut corresponds to the communication
complexity of the protocol. Can any area-law-violating Hamiltonian be
connected to an entanglement-testing protocol with communication
pattern corresponding to the interaction graph of the Hamiltonian?
More generally, in what ways can Hamiltonians be viewed as {\it
testers} for their ground states? Whether such a ``translation''
always exists between entanglement testing protocols of limited
communication, and entangled ground states of Hamiltonians with
limited interactions between different parts of the system, remains to
be explored. Making such an equivalence rigorous might open up a
whole new set of tools to studying the area law question, and more
generally, help develop better intuition for local Hamiltonians and
their ground states. A related question is whether EPR testing is in
fact {\it equivalent} in some sense to the property of being a quantum
expander.
\section{Acknowledgements}
\label{sec:thanks}
The authors thank the Simons Institute (the Quantum Hamiltonian
Complexity program) where part of this work was done. DA acknowledges
the support of ERC grant 030-8301 and BSF grant 037-8574. AWH was
funded by NSF grant CCF-1111382 and ARO contract W911NF-12-1-0486. ZL
was supported by NSF Grants CCF-0905626 and CCF-1410022 and Templeton
Grants 21674 and 52536. DN thanks the Slovak Research and Development
Agency grant APVV-0808-12 QIMABOS. MS is supported by the NSF Grant
No. CCF-0832787, ``Understanding, Coping with, and Benefiting from,
Intractability'' and by CISE/MPS 1246641. UV was supported by ARO
Grant W911NF-12-1-0541, NSF Grant CCF-0905626, and Templeton Grants
21674 and 52536.
|
\section{{ Introduction.}}
In this note I would like to introduce a new approach to
(or rather a new language for) representation theory of groups.
Namely, I propose to consider a (complex) representation of a group $G$
as a sheaf on some geometric object.
This point of view necessarily leads to a conclusion that
the standard approach to (continuous)
representations of algebraic groups should be modified.
Let us start with a local or finite field $F$
and fix an algebraic group ${\mathcal G}$ defined over $F$.
In the standard approach
we consider the set $G = {\mathcal G}(F)$ of $F$-points of ${\mathcal G}$
as a topological group
and study an appropriate category $Rep(G)$ of continuous
representations of $G$.
\smallskip
The main goal of this note is to explain that
this approach is philosophically inconsistent.
In fact I will describe how to extend the category $Rep(G)$ to
some larger category ${\mathcal M}({\mathcal G}, F)$ that better
corresponds to our intuitive understanding of representations of $G$.
We will see that this category can be naturally described as a
product of categories $Rep(G_i)$ over all pure inner forms of the group $G$.
On the level of simple objects this means that
$Irr({\mathcal M}({\mathcal G}, F)) \approx \coprod Irr(G_i)$.
This agrees with observation by several mathematicians
(e.g. by D. Vogan \cite{Vog})
that when we classify irreducible representations
it is better to work with the union of sets $Irr(G_i)$ for several forms
of the group $G$ than with one set $Irr(G)$.
\subsection{{ Representations and sheaves on Stacks. }}
In order to describe the category ${\mathcal M}(G)$ I propose to
consider representations as sheaves on algebraic stacks.
Stacks play a more and more important role in
contemporary Mathematics. Since they are not yet the common language in
representation theory I will recall some basic notions related to stacks.
\smallskip
Informally, stack is a ``space" $X$ such that every point
$x \in X$ is endowed with a group $G_x$ of automorphisms of
inner degrees of freedom at this point.
We see that in order to consider stacks we should first fix a
Geometric Environment,
i.e. a category ${\mathcal S}$ of spaces on which we
model our stacks.
In fact the category ${\mathcal S}$ should be considered together with some Grothendieck topology.
The standard term for such category ${\mathcal S}$ is ``site".
\smallskip
Usually one works with the following sites:
\smallskip
(i) Category of schemes over a field $F$ (with etale or smooth topology).
(ii) Category of smooth manifolds (with usual topology).
(iii) Category of locally compact Hausdorff topological spaces with usual
topology (or some natural subcategory of it, e.g. the category of totally disconnected spaces).
(iv) Category $Sets$ of sets with discrete topology.
\subsection{{ Groupoids. }}
Stacks modeled on the site ${\mathcal S} = Sets$ are {\textbf {groupoids}}. Let me remind that,
by definition, groupoid is a category in which all morphisms are isomorphisms.
Groupoids represent rather elementary examples of stacks. However they exhibit many
features of the general case. Also in this case the general ideas that I would like to explain
are much easier to understand.
For this reason I would like to discuss this case in some detail.
\subsubsection{{Examples of groupoids. }}
To every discrete group $G$ we assign
the {\textbf{ basic groupoid}} $BG = pt / G$ as follows:
\smallskip
An object of the category $BG$ is a $G$-torsor $T$ (i.e. a non-empty $G$-set on which
$G$ acts transitively and free).
The morphisms in this category are morphisms of $G$-sets.
\smallskip
More generally, given an action of the group $G$ on a set $Z$ we define
the {\textbf {quotient groupoid}} $BG(Z) = Z / G$ as follows:
Object of $BG(Z)$ is a $G$-torsor $T$ equipped with a $G$-morphism $\nu: T \to Z$.
Morphisms are morphisms of $G$-sets over $Z$.
\subsubsection{ Representations as sheaves on groupoids.}
Given a groupoid ${\mathcal X}$ it is natural to think about it
as a geometric object (some kind of a space). Then it is natural to consider sheaves on this space.
\smallskip
We define a {\textbf {sheaf}} $R$ (of complex vector spaces) on a groupoid ${\mathcal X}$
to be a functor $R: {\mathcal X} \to Vect$, where $Vect$ is the category of complex
vector spaces (we will see later why this notion is natural).
We denote by $Sh({\mathcal X})$ the category of sheaves on ${\mathcal X}$.
\smallskip
{\textbf {Claim.}} The category $Sh(BG)$ is naturally equivalent to the category $Rep(G)$.
More generally, for every $G$-set $Z$ the category $Sh(BG(Z))$ is naturally
equivalent to the category $Sh_G(Z)$ of $G$-equivariant sheaves on $Z$
(see \ref{equiv-sheaves}).
\smallskip
This gives us a ``geometric" description of the category $Rep(G)$. This construction, that
is very elementary in case of groupoids, is the basis of the approach that I describe in this note.
\subsection{{ Topological groupoids.}}
Let $G$ be a locally compact group. For technical reasons let us assume that it is
totally disconnected. In this case we also can define the basic
groupoid $BG$ and quotient groupoids $BG(Z)$ as stacks modeled on the site of
locally compact spaces. For any stack ${\mathcal X}$
of this type we will define a category $Sh({\mathcal X})$ of sheaves on ${\mathcal X}$.
Using these constructions we can interpret the category
$Rep(G)$ as the category of sheaves on the stack $BG$.
\smallskip
One of the ways to think about a stack modeled on the site of locally compact spaces
is to interpret it as a topological groupoid.
For example, using this interpretation
it is easy to show that for a $G$-space $Z$ the category $Sh(BG(Z))$ of sheaves
on the quotient stack
$BG(Z) = Z / G$ is naturally
equivalent to the category $Sh_G(Z)$ of $G$-equivariant sheaves
on $Z$ (see \ref{equiv-sheaves}).
Technically, working with topological groupoids is a little difficult.
Later we describe another way to define the category $Sh({\mathcal X})$ of
sheaves on ${\mathcal X}$ for stacks of this type that is technically simpler
(see section \ref{F-sheaves}).
\subsection{{ Algebraic groups and stacks.}}
Let us consider more interesting case of an algebraic group ${\mathcal G}$
over a local (or finite) field $F$.
By analogy with the discrete case we define the {\textbf {basic
stack}} $B{\mathcal G} = pt/{\mathcal G}$.
This is an algebraic stack over the field $F$ (i.e. it is modeled on the site ${\mathcal S}$ of
schemes over $F$ (see section \ref{stacks})).
\smallskip
For any algebraic stack ${\mathcal X}$ over $F$ we will construct the category
$Sh({\mathcal X})$ of sheaves of complex vector spaces on ${\mathcal X}$.
The informal idea is that $F$-points
${\mathcal X}(F)$ of the stack ${\mathcal X}$ form a groupoid and a sheaf $R$ on ${\mathcal X}$
is just a sheaf $R$ on this groupoid.
For finite fields this works fine. For local fields we should take into
account the topology of the groupoid ${\mathcal X}(F)$. \smallskip
\smallskip
Given an algebraic group ${\mathcal G}$ over $F$ we define
the category ${\mathcal M} = {\mathcal M}({\mathcal G}, F)$ as the category $Sh(B{\mathcal G})$ of
sheaves on the algebraic stack $B{\mathcal G}$.
I call the objects of this category {\textbf {stacky $G$-modules}.}
\subsubsection{Two competing definitions}
Now starting with algebraic group ${\mathcal G}$ over $F$ we can consider
two competing definitions of a representation.
\smallskip
{\textbf Definition}. {\textbf 1.} Category $Rep(G)$ \ obtained from ${\mathcal G}$ by a chain of constructions
\smallskip
\quad \quad ${\mathcal G} \quad \Longrightarrow$ \quad group $ G = {\mathcal G}(F) \quad \Longrightarrow$
\quad groupoid ${\mathcal Y} = BG$ $ \quad \Longrightarrow \quad Sh({\mathcal Y})$
\medskip
{\textbf Definition}. {\textbf 2.} Category ${\mathcal M}({\mathcal G} ,F)$ obtained from ${\mathcal G}$ by a chain of constructions
\smallskip
\quad \quad ${\mathcal G} \quad \Longrightarrow$ \quad stack ${\mathcal X} = B{\mathcal G}
\quad \Longrightarrow$ \quad groupoid
${\mathcal X}(F)$ \quad $ \Longrightarrow \quad Sh({\mathcal X}(F))$
\smallskip
The subtle point is that the groupoids ${\mathcal Y}$ and ${\mathcal X}(F)$ are not
always equivalent.
So the category ${\mathcal M}({\mathcal G}, F) = Sh({\mathcal X}(F))$ might be
not equivalent to the category $Rep(G) = Sh({\mathcal Y})$.
\smallskip
The standard notion of a continuous representation of $G$ is based on definition 1.
\textbf{The main goal} of this note is to convince the reader that from
many points of view definition 2 is much more appropriate than the standard definition 1.
\subsection{{ Vogan's picture.}} \label{Vogan}
One advantage of this definition is that it gives an explanation to
representations that appear in Vogan's interpretation of the Langlands correspondence.
Let me remind what is Vogan's suggestion (see \cite{Vog}, Conjecture 4.15 or \cite{ABV},
Theorem 1.18).
Vogan tried to describe the Langlands correspondence in the following explicit way.
Let $F$ be a local field, ${\mathcal G}$ a reductive algebraic group over $F$ and $G$
the group of its $F$-points. Consider on one side the set $Irr(G)$ of equivalence
classes of irreducible representations of $G$. On the other side consider the
set $Lan({\mathcal G} , F)$ of Langlands'
parameters $\Phi$ defined in terms of the dual group $L({\mathcal G})$ and the field $F$.
Vogan tried to construct a canonical bijection between these two sets.
He realized that in many
cases this can not work since the set $Irr(G)$ is just too small. But he also discovered that if
we replace this set by the disjoint union $\coprod Irr(G_i)$ of corresponding sets for all pure inner
forms $G_i$ of the group $G$ then this set has correct size (note that some of these forms
might be isomorphic -- then the ``same" representation will appear in this list several times).
In fact in many cases Vogan was
able to describe a bijection of this set with the set of Langlands' parameters.
\smallskip
In the language I propose Vogan's conjecture can be formulated as follows.
First of all, instead of the group ${\mathcal G}$
we consider the algebraic stack $B{\mathcal G}$.
Instead of representations of the group $G = {\mathcal G}(F)$ let us study the
category ${\mathcal M}:= Sh(B{\mathcal G})$ of sheaves on this stack
(i.e. the category of stacky $G$-modules).
Since for any pure inner form ${\mathcal G}'$
of the group ${\mathcal G}$ the stack $B{\mathcal G}'$ is equivalent to $B{\mathcal G}$, the category
${\mathcal M}$ depends only on the pure inner class of ${\mathcal G}$.
Our goal is to parameterize the set $Irr({\mathcal M})$ of isomorphism classes
of simple objects in ${\mathcal M}$.
We will see in \ref{sheaves-algebraic-stacks} that this set can be described
as a disjoint union over all pure inner forms ${\mathcal G}_i$ of the group
${\mathcal G}$ of the sets $Irr(G_i)$, where $G_i = {\mathcal G}_i(F)$.
Then the conjecture 4.15 in \cite{Vog} is essentially the statement
that the set $Irr({\mathcal M})$
is in natural bijection with the set $Lan({\mathcal G}, F)$ of Langlands' parameters.
\smallskip
{\textbf{Remarks}.} \textbf{1.} In fact Vogan formulated his conjectures only for
pure inner forms of quasi-split groups. Later they were generalized by Kaletha
to other forms (see \cite{Kal}).
\textbf{2.} It would be interesting to understand whether one has
a \textbf{canonical} bijection between these sets.
The constructions proposed in \cite{Vog}, \cite{ABV}
are not quite canonical -- they depend on choice of some ``Whittaker data".
\smallskip
I think that in fact they are not in canonical bijection. Namely,
I think that the set of Langlands' parameters $Lan({\mathcal G}, F)$
should be slightly modified -- and after this the resulting set $Lan'({\mathcal G},F)$
will be in a canonical bijection with the set $Irr({\mathcal M})$.
Possible suggestions for this modified set of Langlands' parameters
can be found in \cite{BG} and sources listed there.
\subsection{About this note.}
In section \ref{sheaves} \ I remind basic facts about sheaves that are relevant
in Representation Theory.
In particular, I discuss the notion of an equivariant sheaf on a topological $G$-space $Z$.
This notion plays central role in any geometric approach to Representation Theory.
This is also an important computational tool -- in fact later on when I talk about
explicit description
of some object I mean a description in the language of equivariant sheaves.
In section \ref{example} \ I describe a striking example that shows that
the standard definition of representations is not a good one.
In subsequent sections I make various comments on the notions we discussed above,
indicate how to formulate precise technical definitions that describe these notions
and describe some technical tools that help to make computations with them.
In section \ref{groupoids} \ I discuss the case of groupoids.
In section \ref{stacks} \ I shortly describe how to
give a technical definition of a stack.
In section \ref{F-sheaves} \ I
discuss a technical definition of the category $Sh({\mathcal X})$ of sheaves of
complex vector spaces on
a stack ${\mathcal X}$ and describe how to make computations with them. In particular,
I explain how one can describe the category $Sh({\mathcal X})$ in
terms of equivariant sheaves on topological spaces.
In section \ref{AG} \ I explain the algebro-geometric structure
that describes the relation between groupoids ${\mathcal X}(F)$ for different fields $F$.
\medskip
This note is an expanded version of the lecture that I gave at
the Fourth Conference of Tsinghua Sanya International Mathematics Forum (TSIMF)
in December 2013. I would like to thank TSIMF organizers for the invitation.
\smallskip
Main ideas about stacks in representations theory came out of
my research in the framework of the ERC grant 291612. My research was supported
by the grant 533/14 of Israel Science Foundation.
Much of my work on this subject was done during my visits to MPIM, Bonn.
I would like to thank MPIM for very creative atmosphere.
I would like to thank A. Vistoli for his remarks about stacks that helped me to
understand this notion. I also thank D. Zagier for critical remarks about an
earlier version of this note.
I would like to thank the referee for several useful suggestions.
\section{ Sheaves and equivariant sheaves.} \label{sheaves}
\subsection{Sheaves relevant in representation theory.}
Let $F$ be a finite or local field, ${\mathcal G}$ an algebraic group over $F$.
We consider the group $G = {\mathcal G}(F)$ of its $F$-points as a topological group.
In order to study representations of the group $G$ we usually place this group in some
geometric environment.
Namely we fix a site ${\mathcal S}$ appropriate for the group $G$ and for every space $Z \in {\mathcal S}$
we consider some category $Sh(Z)$ of sheaves of complex vector spaces on $Z$.
\smallskip
Let us describe this in more detail.
\smallskip
(i) Let $F$ be a finite field. We consider the site ${\mathcal S}$ of finite discrete sets.
For a space $Z \in {\mathcal S}$ we denote by $Sh(Z)$ the category of all sheaves of
complex vector spaces on $Z$.
\smallskip
(ii) Let $F$ be a local non-Archimedean field. We consider the site ${\mathcal S}$ that consists
of Hausdorff locally compact totally disconnected spaces $Z$ with countable base of open subsets
($l$-spaces in terminology of \cite{BZ}).
Given an $l$-space $Z$ we consider the category $Sh(Z)$ of all sheaves of complex vector
spaces on $Z$ (in \cite{BZ} they were called $l$-sheaves).
\smallskip
(iii) Let $F$ be a local Archimedean field, i.e. $F = {\mathbf R}} \def\Q{\mathbf {Q}} \def\P{{\mathbf P}$ or $F = {\mathbf C}} \def\V{{\mathbf V}$. In this case I see several
candidates
for the site ${\mathcal S}$.
We can consider ${\mathcal S}$ to be the category of smooth manifolds.
For a manifold $Z$ the category $Sh(Z)$ is the category of sheaves of ${\mathcal O}_Z$-modules,
where ${\mathcal O}_Z$ is the sheaf of smooth complex valued functions on $Z$.
\smallskip
Another possibility is to work with the site of Nash manifolds and some sheaves of Schwartz functions
(or distributions) on these manifolds (see \cite{AG}).
I do not know what is the correct approach to this case. So in what
follows I mostly deal with finite and local non-Archimedean fields.
\medskip
{\textbf {Remarks}.} {\textbf 1.} In this note for simplicity I discuss only representations over the field $k = {\mathbf C}} \def\V{{\mathbf V}$
of complex numbers. In case when $F$ is a finite or non-Archimedean field we can consider other
fields $k$ of coefficients.
If $F$ is a local non-Archimedean field and $k$ is an extension of $F$ we can try to include the
theory of locally analytic representations by considering the site ${\mathcal S}$ of analytic
manifolds over $F$
and categories $Sh(Z)$ of sheaves of modules over the sheaf ${\mathcal O}_Z$
of locally analytic functions on $Z$.
It seems that the stacky language that I will introduce might be applicable
and useful in this theory.
\smallskip
{\textbf 2.} The structures I described are important when we are trying to specify the category
$Rep(G)$ of representations of the group $G$ that we would like to study.
In case when the group $G$ is defined over a finite or local non-Archimedean field $F$
there is a consensus what is the ``correct" category of representations $Rep(G)$.
Namely, in case of a finite field we consider the category of all representations and
in case of a local non-Archimedean field we consider the category of smooth
representations (smooth means that every vector has open stabilizer in the group $G$).
The situation for real groups is different -- the correct choice of the appropriate
category $Rep(G)$ is a very
non-trivial question. In case of reductive groups this was done by Casselman-Wallach
(see \cite{Cass},\cite{W}, \cite{BK}).
For general real algebraic group I do not know a good candidate for this category.
In this note I am not going to discuss this tricky question. For this reason I will mostly deal with finite
and local non-Archimedean fields.
\subsection{Equivariant sheaves.} \label{equiv-sheaves}
From now on we assume that all topological spaces we consider are $l$-spaces
(locally compact, totally disconnected with countable base of open subsets).
Let $G$ be a topological group and $Z$ a $G$-space, i.e. $Z$ is equipped with
a continuous action $a: G \times Z \to Z$.
We denote by $Sh_G(Z)$ the category
of $G$-equivariant sheaves (of complex vector spaces) on $Z$.
This category will play central role in what follows.
\smallskip
Recall, that a $G$-equivariant sheaf is a sheaf $R$ on $Z$ equipped with
an isomorphism ${\alpha}} \def\bet{{\beta}} \def\gam{{\gamma}: a^*(R) \to pr_Z^*(R)$ of two liftings of $R$ to the
space $G \times Z$ satisfying some natural cocycle condition (this condition is that after the lifting
to the space $G \times G \times Z$ two morphisms $\nu, \mu: (m \cdot a)^*(R) \to pr^*(R)$ of
sheaves naturally constructed from ${\alpha}} \def\bet{{\beta}} \def\gam{{\gamma}$
should coincide (see details of the definition and discussion in \cite{BL}, Part1, section $0$).
\smallskip
{\textbf {Remark}.} In case of real groups we can work with the site ${\mathcal S}$ of
smooth manifolds and with sheaves of ${\mathcal O}$-modules. In this case
the definition of equivariant sheaves formally looks exactly the same, but has
quite different geometric meaning.
The reason is that the pullback functor $a^*$ in the category of ${\mathcal O}$-modules
is quite different from the pullback functor in the category of sheaves.
\medskip
Let me remind two standard facts about equivariant sheaves
(here we assume that $G$ is an $l$-group).
\smallskip
{\textbf {Fact 1}.} Let $Z$ be a point. Then the category $Sh_G(Z)$ is equivalent
to the category $Rep(G)$ of {\textbf {smooth}} representations of $G$.
\smallskip
{\textbf {Fact 2}.} Suppose that $Z$ is a quotient space of the group $G$.
Fix a point $z \in Z$ and denote by $H$ its stabilizer in $G$. Then we have natural
equivalences of categories
$Sh_G(Z) \approx Sh_H(z) \approx Rep(H)$.
\section{An example.} \label{example}
In this section I describe a striking example that illustrates what is wrong with
the standard approach.
\subsection{ Heuristics.} My example is based on the following heuristic geometric principle.
Let $a: {\mathcal G} \times {\mathcal Z} \to {\mathcal Z}$ be a transitive action of an algebraic
group ${\mathcal G}$ on an algebraic variety ${\mathcal Z}$.
Passing to $F$-points we get a continuous action $a: G \times Z \to Z$.
This action is usually not transitive. In many cases we can write $Z$ as a
union of open orbits $Z = \bigsqcup Z_i$, $i = 1,...,n$.
\smallskip
{\textbf {Heuristic Geometric Principle}.}
\smallskip
1. The space $Z$ is ``good", i.e. it is easy to describe.
\smallskip
2. Every individual orbit $Z_i$ might be a ``bad" space, that
means that it is difficult to describe.
\smallskip
{\textbf {Illustration}.} Consider the space $V$ of real symmetric $8 \times 8$ matrices and
try to give explicit descriptions of subsets $Z, Z_3 \subset V$ that describe non-degenerate
quadratic forms and quadratic forms of signature $3$ on ${\mathbf R}} \def\Q{\mathbf {Q}} \def\P{{\mathbf P}^8$ respectively.
\subsection{{ An example -- representations of orthogonal groups.}}
Fix an $n$-dimensional vector space $V$ over $F$.
The group $G = GL(V )$ acts on the space $Z$ of non-degenerate quadratic forms.
Fix a form $Q \in Z$ and denote by $H$ its stabilizer in $G$ (i.e. $H$ is the orthogonal group $O(Q)$).
Let us denote by $Z_Q$ the $G$-orbit of $Q$ in $Z$. Then we have an equivalence of categories
$Rep(H) \approx Sh_G(Z_Q)$.
\medskip
According to the heuristic geometric principle the space $Z_Q$ might be (and often is)
a ``bad" space. This means that the category $Rep(H) \approx Sh_G(Z_Q)$ might be a
``bad" category (i.e. very difficult to describe).
In other words, the ``natural" problem of classification of irreducible representations
of the orthogonal group $H = O(Q)$ turns out to be not that natural.
\smallskip
However we see that the bad category $Rep(H) \approx Sh_G(Z_Q)$ can be naturally
extended to a larger category
${\mathcal M} := Sh_G(Z)$ of all $G$-equivariant sheaves on the good algebraic space $Z$.
We can expect (and this is really the case) that this larger category ${\mathcal M}$ is a ``good" category.
\smallskip
Note that we have a natural decomposition ${\mathcal M} \approx \prod Sh_G(Z_i)$,
where $Z_i$ are $G$-orbits in $Z$. In particular the set $Irr({\mathcal M})$ of isomorphism classes of simple
objects in ${\mathcal M}$ is a disjoint union of sets $Irr(Sh_G(Z_i))$.
It seems reasonable to assume that the classification of
simple objects of the category
${\mathcal M}$ might be relatively simple problem, but then to sort out
which of them are related
to the orbit $Z_Q$ might turn out to be much more difficult problem
(and it is not clear whether this problem is a meaningful one).
\medskip
The example we are considering suggests a certain pattern that seems to work
also in the general case.
Namely, we see that if $G$ is a group of points of an algebraic group
then the category $Rep(G)$ of its representations might be a bad category,
but we can include it as a direct factor into some larger good
category ${\mathcal M}(G)$.
\smallskip
I had this example in mind for some time until I realized how one can define
this larger category using sheaves on stacks.
\section{{ Some remarks about groupoids}} \label{groupoids}
\subsection{{Equivalence of groupoids}}
We know that if two objects of some category are isomorphic
then it is better to consider them as two realizations of the same geometric structure.
Similarly,
if two groupoids ${\mathcal X}$ and $ {\mathcal Y}$ are equivalent (as categories) we
can assume that they represent
two realizations of the same geometric structure.
\smallskip
\smallskip
A subtle point here is that the equivalences between these groupoids form a groupoid.
This means that if we fix an equivalence $Q: {\mathcal X} \to {\mathcal Y}$ then this equivalence itself has
automorphisms, and it is not immediately clear how we should think about them.
\smallskip
Similarly, if we would like to show that two groupoids are {\textbf {canonically}}
equivalent we have to construct an equivalence between them and show that this equivalence
is defined up to a canonical isomorphism.
\smallskip
{\textbf {Example}.} Consider an action $a: G \times Z \to Z$.
Let us define a groupoid $BG_0(Z)$ as follows:
\smallskip
$ Ob(BG_0(Z)) = Z, Mor(BG_0(Z)) = G \times Z$, where morphism $(g,z)$ is a morphism from the object $z$ to the object $gz$.
\smallskip
The groupoid $BG_0(pt)$ we denote by $BG_0$. \smallskip
\smallskip
{\textbf {Claim}.} The groupoid $BG_0(Z)$ is canonically equivalent to the groupoid $BG(Z)$.
\smallskip
The groupoid $BG_0(Z)$ might be considered as a
``matrix" version of the groupoid $BG(Z)$. It is better suited for computations.
\smallskip
\subsection{{ Theory of groups and theory of groupoids.}}
I would like to explain that the theories describing groups and groupoids
are essentially equivalent.
\medskip
{\textbf {Proposition}.} {\textbf 1.} Every groupoid ${\mathcal X}$ is canonically decomposed
as a disjoint union of connected groupoids
(groupoid is connected if all its objects are isomorphic).
{\textbf 2.} A connected groupoid ${\mathcal Y}$ is equivalent to the basic
groupoid for some group $G$.
\smallskip
This result shows that any question about groupoids can be reduced
to a question in group theory.
In other words, the difference between theories of groups and groupoids
is in their emphasis.
In my opinion the relation between the theory of groupoids and the group theory
is very similar to the relation between linear algebra and matrix calculus.
While these two theories are basically equivalent, clearly linear algebra is much more intuitive.
So I expect that eventually the stacky approach will become a standard tool in representation theory.
\smallskip
\subsubsection{{ Equivalence between groups and connected groupoids.}}
The group $G$ corresponding to a connected groupoid ${\mathcal Y}$
is not defined canonically.
It depends on a choice of an object $Y \in {\mathcal Y}$.
Namely, given an object $Y$ we can define a group $G = G_Y$ by $G:= Aut(Y)$.
Then we get canonical equivalence of categories $Q = Q_Y: {\mathcal Y} \to BG$,
defined by $Q(X) = Mor(Y,X)$.
If we pick another object $Y'$ we get a different group $G'$ and a
different equivalence $Q': {\mathcal Y} \to BG'$.
Note that any choice of an isomorphism $\nu: Y \to Y'$
defines natural isomorphisms $G \backsimeq G'$
and $Q \backsimeq Q'$. However there is no preferred choice for such an isomorphism $\nu$.
\subsection{Examples of groupoids.}
The next three constructions show that in Mathematics we usually
encounter groupoids and not groups.
\subsubsection{{ Multiplicative groupoid of a category.}}
{\textbf {Construction I.}} \ Starting with any category $C$ we construct the
{\textbf {multiplicative groupoid}} $C^* = Iso(C)$
that has the same collection of
objects as category $C$ and isomorphisms of $C$ as morphisms.
\smallskip
\medskip
{\textbf {Example 1.}} $C = Finsets$ -- the category of finite sets.
\smallskip
In this case the groupoid $Iso(C)$ is essentially the collection of all
symmetric groups $S_n$.
\medskip
{\textbf {Example 2.}} $C = Vect_k$ -- the category of finite dimensional vector
spaces over a field $k$.
\smallskip
In this case the groupoid $Iso(C)$ describes the collection of
groups $GL(n,k)$ for all $n$.
\subsubsection{{ Poincar\'e groupoid.}}
{\textbf {Construction II.}} \ Poincar\'e groupoid $Poin(X)$ of a topological space $X$.
\smallskip
Objects of $Poin(X)$ are points of $X$. Morphisms from $x$ to $y$ are homotopy
classes of paths from $x$ to $y$.
\smallskip
If the space $X$ is path connected then the groupoid $Poin(X)$ is connected.
For any point $x \in X$ the group
$Aut_{Poin(X)}(x)$ is the fundamental group $\pi_1(X,x)$.
This shows that the Poincar\'e groupoid is more basic notion than the fundamental group.
\smallskip
\subsubsection{{ Galois groupoid.}}
{\textbf {Construction III. }} \ Galois groupoid $Gal(F)$ of a field $F$.
\smallskip
Objects of the groupoid $Gal(F)$ are field extensions $F \to \Om$ such that
$\Om$ is an algebraic closure of $F$.
Morphisms are isomorphisms of field extensions.
\smallskip
The groupoid $Gal(F)$ is connected.
If we fix an algebraic closure $\Om$ then by definition the group $Aut_{Gal(F)}(\Om)$
is the absolute Galois group $Gal(\Om / F)$.
Again we see that the notion of Galois groupoid is more basic than
the notion of Galois group.
\smallskip
Note that the constructions of Poincar\'e and Galois groupoids are very similar.
\smallskip
\section{ What is a stack ?} \label{stacks}
Let us fix some site ${\mathcal S}$. I would like to describe
the notion of a stack ${\mathcal X}$ modeled on ${\mathcal S}$.
I assume two features of this notion.
\smallskip
\smallskip
1. For every two stacks ${\mathcal X}, {\mathcal Y}$ the collection of morphisms
from ${\mathcal X}$ to ${\mathcal Y}$
forms a groupoid $Mor({\mathcal X},{\mathcal Y})$.
2. Every object $S \in {\mathcal S}$ is a stack.
\smallskip
The natural idea is to characterize a stack ${\mathcal X}$ by the collection of groupoids
${\mathcal X}(S):= Mor(S, {\mathcal X})$ for all objects $S \in {\mathcal S}$.
\smallskip
In fact usually it is enough to know the groupoids ${\mathcal X}(S)$
for objects $S$ in some subcategory ${\mathcal B} \subset {\mathcal S}$ provided it is large enough.
For example, if ${\mathcal S}$ is the category of schemes we can restrict everything to the
subcategory ${\mathcal B}$ of affine schemes.
\subsection{{ Informal technical definition of a stack.}}
Fix a large subcategory ${\mathcal B} \subset {\mathcal S}$. We define a stack ${\mathcal X}$ over the
site ${\mathcal S}$ to be the following collection of data:
\smallskip
(i) To every object $S \in {\mathcal B}$ we assign a groupoid ${\mathcal X}(S)$
(ii) To every morphism $\nu: S \to S'$ in ${\mathcal B}$ we assign a functor
${\mathcal X}(S') \to {\mathcal X}(S)$
(iii) To every composition of morphisms in ${\mathcal B}$ we assign an
isomorphism of the corresponding functors.
\smallskip
This data should satisfy a variety of compatibility conditions . These include
(i) Compatibility conditions for isomorphisms we have chosen.
(ii) Descent properties for
morphisms and for objects with respect to the Grothendieck topology on the site ${\mathcal S}$.
(iii) Some finiteness conditions.
(iv) We also usually assume that the stack ${\mathcal X}$ is
dominated by some object $Z \in {\mathcal S}$ (for example the quotient stack $Z / G$ that
we will describe bellow is dominated by an object $Z$).
\smallskip
A relatively elementary exposition of stacks one can find in \cite{Fan}.
More detailed and more sophisticated exposition see in \cite{Vis}.
\subsection{{Stacks modeled on the site $Sets$}}
Consider the site ${\mathcal S} = Sets$. To describe a stack ${\mathcal Y}$ modeled on ${\mathcal S}$ we have to
assign to every set $S \in {\mathcal S}$ a groupoid ${\mathcal Y}(S)$.
\smallskip
{\textbf{Example.}} Given a group $G \in {\mathcal S}$ we construct the basic stack $BG$ as follows:
For any set $S$ the objects of the groupoid $BG(S)$ are principle $G$-bundles $P$ over $S$
and morphisms are $G$-morphisms over $S$.
Let me remind that principle $G$-bundle over $S$
is a $G$-set $P$ equipped with a $G$-morphism $p: P \to S$, where $G$ acts trivially on $S$,
that locally on $S$ is isomorphic to a trivial $G$-bundle $pr: G \times U \to U$.
\smallskip
More generally, if $Z$ is a $G$-space we define the quotient stack $BG(Z)$ as follows:
For every set $S$ the objects of the groupoid $BG(Z)(S)$ are principle $G$-bundles over $S$ equipped with
a $G$-morphism to $Z$ and morphisms are morphisms of $G$-bundles over $Z$.
\medskip
Note that in case of sets we can restrict everything to a subcategory $B \subset {\mathcal S}$ that
contains just one object $pt$ -- this is a big enough subcategory. This shows that
every stack ${\mathcal Y}$ modeled
on the site ${\mathcal S}$ can be completely described by the groupoid ${\mathcal X} = {\mathcal Y}(pt)$.
This explains why the stacks in this case can be considered as groupoids.
\subsection{Stacks modeled on the site $S$ of $l$-spaces.}
Consider the site ${\mathcal S}$ of $l$-spaces. Let $G \in {\mathcal S}$ be an $l$-group.
We define the basic stack $BG$ as follows:
For every $S \in {\mathcal S}$ the objects of the groupoid $BG(S)$ are principal $G$-bundles over $S$ and
morphisms are morphisms of $G$-bundles.
In this case again a $G$-bundle is a morphism of $G$-spaces $p:P \to S$ that locally in $S$ is
isomorphic to the trivial $G$-bundle $pr: G \times U \to U$.
\smallskip
Similarly one defines a stack $BG(Z) = Z/G$
for a $G$-space $Z$.
\smallskip
In case of the site of $l$-spaces we can restrict everything to the
subcategory $B \subset {\mathcal S}$ consisting
of compact spaces. Using this fact it is easy to check that the basic stack $BG$
can be described in terms of a topological groupoid ${\mathcal X} = BG(pt)$.
Up to equivalence this groupoid can be explicitly described as follows: \
${\mathcal X}$ has one object $X$ and its automorphism group $Aut(X)$ is
the topological group $G$.
\subsection{Stacks modeled on the site of $F$-schemes.}
\subsubsection{Basic stack $B{\mathcal G}$.}
Let us fix a field $F$ and consider the site ${\mathcal S}$ of schemes over $F$.
Let ${\mathcal G} \in {\mathcal S}$ be an algebraic group defined over $F$.
We define the basic stack $B{\mathcal G}$
in the same way as before.
Namely, for an affine $F$-scheme $S$ an object of the groupoid $B{\mathcal G}(S)$
is a principal $G$-bundle $P$ over $S$. Morphisms are morphisms of $G$-bundles.
\smallskip
In this case the notion of a principal $G$-bundle is more subtle. Namely, we should
consider morphisms $p: P \to S$ of $G$-schemes that should be locally trivial in etale topology.
This is more sophisticated notion.
\subsubsection{Torsors.}
Let us consider in more detail the case of a point, i.e. assume that $S = Spec(F)$.
In this case we have to consider a morphism $p: T \to Spec(F)$, where $T$ is an $F$-scheme
with an action of the group $G$
that is locally trivial in etale topology (such $T$ is called a \textbf {$G$-torsor}).
\smallskip
The condition of local triviality is equivalent to the statement that after the base change
from $S = Spec(F)$ to $S' = Spec(F')$, where $F'$ is an algebraic closure of $F$, the torsor $T'$
will be isomorphic as a $G'$-space to the trivial $G'$ torsor $G'$.
\smallskip
Here is a typical example of torsors.
\smallskip
{\textbf{Example.}} Let $K$ be a finite separable field extension of $F$.
Multiplicative groups $K^*$ and $F^*$ we consider as $F$-points of
algebraic groups ${\mathcal K}^*, {\mathcal F}^*$ defined over $F$. Then we have the norm
morphism of algebraic groups
$N: {\mathcal K}^* \to {\mathcal F}^*$. We denote
by $G$ the kernel of this morphism.
This is an algebraic group defined over $F$.
For any point $a \in F^*$ we consider the variety ${\mathcal T}_a = N^{-1}(a) \subset {\mathcal K}^*$.
This variety is defined over $F$ and is a $G$-torsor.
Two torsors ${\mathcal T}_a, {\mathcal T}_b$ are isomorphic iff the quotient $a/b$ lies in the image of $K^*$ under the norm map
$N: K^* \to F^*$.
\subsubsection{Quotient stack $B{\mathcal G}({\mathcal Z})$.}
\smallskip
Let ${\mathcal G}$ act on a scheme ${\mathcal Z} \in {\mathcal S}$. \
We define the quotient stack
${\mathcal X} = B{\mathcal G}({\mathcal Z}) = {\mathcal Z} / {\mathcal G}$ as follows:
\smallskip
For an affine scheme $S \in {\mathcal S}$ the category $B{\mathcal G}(Z)(S)$ consists of principal $G$-bundles
$p: P \to S$ equipped with a $G$-morphism $ P \to Z$ and morphisms are morphisms of
principal $G$-bundles over $Z$.
\smallskip
\section{{ Sheaves on stacks. }} \label{F-sheaves}
\subsection{{ Technical definition of sheaves on stacks.}}
Let us fix a site ${\mathcal S}$. In situations we consider we have the notion of sheaves
(of complex vector spaces) on spaces $Z \in {\mathcal S}$. In other word, to every space
$Z \in {\mathcal S}$ we assign a category $Sh(Z)$, every morphism $\nu: Z \to W$ in
${\mathcal S}$ induces a functor $\nu^*: Sh(W) \to Sh(Z)$ with a compatibility isomorphisms
for the products of two morphisms.
\medskip
{\textbf{Basic examples.}} {\textbf 1.} Let $S$ be the site of $l$-spaces. To every space $Z \in {\mathcal S}$
we assign the category $Sh(Z)$ of $l$-sheaves (i.e sheaves of complex vector spaces) on $Z$;
the functor $\nu^*$ is the usual pullback functor.
\smallskip
{\textbf 2.} Fix a local non-Archimedean field $F$ and consider the site ${\mathcal S}$ of $F$-schemes of finite type.
For every scheme $Z \in {\mathcal S}$ we consider the set $Z(F)$ of its $F$-points as an $l$-space
and we set $Sh(Z): = Sh(Z(F))$ (i.e. a sheaf $R$ on $Z$ is just a sheaf of complex vector spaces
on the $l$-space $Z(F)$).
\smallskip
We would like to extend these categories of sheaves to stacks modeled on the site ${\mathcal S}$.
In other words, given a stack ${\mathcal X}$ we would like in some natural way
to assign to it a category $Sh({\mathcal X})$ so that for a space $Z \in {\mathcal S}$
considered as a stack it will be the category $Sh(Z)$. Let me indicate some general
strategy how to construct the category $Sh({\mathcal X})$.
\medskip
Suppose we already have some notion of sheaves on stacks.
Fix a sheaf $R$ on a stack ${\mathcal X}$.
Then for any space $S \in {\mathcal S}$ and any point $p \in {\mathcal X}(S) = Mor(S, {\mathcal X})$
we get a sheaf $R_p = p^*(F) \in Sh(S)$.
We also get a family of isomorphisms connecting these sheaves.
Now we can try to use these sheaves and isomorphisms to
characterize the sheaf $R$ as follows.
\smallskip
{\textbf {Informal definition}.} A sheaf $R$ on the stack ${\mathcal X}$ is a collection of
sheaves $R_p \in Sh(S)$
for all spaces $S \in {\mathcal S}$ and all
morphisms $p: S \to {\mathcal X}$ and a collection of isomorphisms satisfying correct
compatibility relations.
\smallskip
{\textbf{Claim.}}{\textbf 1.} Let ${\mathcal S}$ be the site of sets, ${\mathcal Y}$ a stack modeled on ${\mathcal S}$.
Then the category $Sh({\mathcal Y})$ is naturally equivalent to the category $Sh({\mathcal X}) := Funct({\mathcal X}, Vect)$,
where ${\mathcal X}$ is the groupoid ${\mathcal Y}(pt)$ that characterizes stack $ {\mathcal Y}$.
\smallskip
{\textbf 2.} Let ${\mathcal S}$ be the site of $l$-spaces. Consider an action of an $l$-group $G$ on an
$l$-space $Z$ and denote by $BG(Z)$ the quotient stack, Then the category $Sh(BG(Z))$ is
naturally equivalent to the category $Sh_G(Z)$ of $G$-equivariant sheaves on $Z$.
\medskip
{\textbf{Remark}.} In some cases the pullback functor is defined only partially (for some good morphisms).
This happens, for example, if we consider the site of manifolds and sheaves of distributions on them.
In this case pullback functor is well defined only for submersions.
In these cases we often can define appropriate notion of a sheaf for a stack ${\mathcal X}$ that is dominated by an
object $Z \in {\mathcal S}$, i.e. has a morphism $p: Z \to {\mathcal X}$, provided that this morphism $p$ is good.
The quotient stacks that we are interested in usually have this property.
\subsection{{ How to describe sheaves on an algebraic stack. }} \label{sheaves-algebraic-stacks}
Let ${\mathcal X}$ be an algebraic stack over $F$. I would like to give a
convenient description of
sheaves on the stack ${\mathcal X}$ in terms of equivariant sheaves.
I will do this for the case of a quotient stack ${\mathcal X} \approx {\mathcal Z}/{\mathcal G}$.
\smallskip
Let $T_1,...T_r$ be representatives of isomorphisms classes of ${\mathcal G}$-torsors.
They are described by elements of $H^1(Gal(F), {\mathcal G}(\bar{F}))$ (for simplicity we assume that this
set is finite; this is always the case when $char(F) = 0$).
For every index $i$ consider the group ${\mathcal G}_i = Aut_{\mathcal G}(T_i)$ -- this gives us
the collection of all pure inner forms of the group $G$. Also we consider a ${\mathcal G}_i$-scheme
${\mathcal Z}_i = Mor_{\mathcal G}(T_i, {\mathcal Z})$.
Passing to $F$ points we construct a topological group $G_i = {\mathcal G}_i(F)$
and a topological $G_i$-space $Z_i = {\mathcal Z}_i(F)$.
\smallskip
{\textbf {Claim}.} The category $Sh({\mathcal X})$ of sheaves on the stack ${\mathcal X}$
is naturally equivalent to $ \prod Sh_{G_i}(Z_i)$ (product of categories of equivariant sheaves).
\smallskip
In particular we see that the collection of simple
objects of the category $Sh({\mathcal X})$ is a disjoint union of
collections of simple $G_i$-equivariant sheaves on $Z_i$.
\smallskip
{\textbf{Remark}.} In case when ${\mathcal Z} = pt$ we see that $Irr({\mathcal M}({\mathcal G},F)) = \coprod Irr(G_i)$.
This means that the set $Irr({\mathcal M}({\mathcal G}, F))$ of equivalence classes of irreducible stacky $G$-modules
can be naturally described as the disjoint union $ \coprod Irr(G_i)$ taken over all pure inner forms
${\mathcal G}_i$ of the group ${\mathcal G}$.
\subsection{{ Reduction to the case of the group $GL(n)$.}} \label{GL(n)}
Let me present one more description of the category of
sheaves on a quotient stack ${\mathcal X} = {\mathcal Z} / {\mathcal G}$
that is often convenient in computations.
\smallskip
{\textbf {Construction}}. \ Suppose that ${\mathcal G}$ is a linear algebraic group.
Then we can imbed it into a group
${\mathcal P}$ isomorphic to $GL(n)$.
\smallskip
Using this we can realize our quotient stack ${\mathcal X} = {\mathcal Z} / {\mathcal G}$
as the quotient stack ${\mathcal W} / {\mathcal P}$, where ${\mathcal W} = {\mathcal P} \times_{{\mathcal G}} {\mathcal Z}$.
\medskip
The group ${\mathcal P}$ has only one pure inner form
(this is Hilbert 90 theorem). This implies that
the category $Sh({\mathcal X})$ can be realized as the category
$Sh_P(W)$ of $P$-equivariant sheaves on $W$, where $P = {\mathcal P}(F)$
and $W = {\mathcal W}(F)$.
\smallskip
{\textbf{Remark}.} The example in section \ref{example} is a special case
of this construction.
\smallskip
\section{{ Algebro-geometric structure of the groupoids ${\mathcal X}(F)$.}} \label{AG}
Let ${\mathcal X}$ be an algebraic stack.
Am important role in the study of sheaves on the stack
${\mathcal X}$ plays the fact that the collection of groupoids
${\mathcal X}(F)$ for different fields $F$ has an algebro-geometric structure.
\medskip
{\textbf {Proposition}.} Let $L \supset F$ be a finite Galois field extension and
$\Gamma = Gal(L / F)$ its Galois group. Then the group ${\Gamma}} \def\Om{{\Omega}$ acts on the
groupoid ${\mathcal X}(L)$ and the {\textbf {fixed point groupoid}} ${\mathcal X}(L)^\Gamma$ is
naturally equivalent to the groupoid ${\mathcal X}(F)$.
\medskip
Here the fixed point groupoid ${\mathcal X}(L)^{\Gamma}} \def\Om{{\Omega}$ is defined in a standard categorical
manner. Its object is an object $X$ of the groupoid ${\mathcal X}(L)$ equipped with
a collection of isomorphisms ${\alpha}} \def\bet{{\beta}} \def\gam{{\gamma}_\gam: X \to \gam(X)$
satisfying natural compatibility conditions.
\medskip
This proposition is just a reformulation of the descent property for the stack ${\mathcal X}$.
|
\section{What are braids?}
\subsection{A brief introduction to the braid group}
\label{sec:braidgroup}
\index{braid!group|(}%
\index{braid!geometric|(}%
Braids are collections of strings anchored between two planes,
\begin{figure}
\begin{center}
\subfigure[]{
\includegraphics[height=.2\textheight]{geombraid}
\label{fig:geombraid}
}\hspace{.25\textwidth}
\subfigure[]{
\includegraphics[height=.2\textheight]{sbdiag}
\label{fig:sbdiag}
}
\end{center}
\caption{(a) A geometric braid with five strings. (b) The corresponding
braid diagram.}
\end{figure}
as in Fig.~\ref{fig:geombraid}. More precisely this is called a
\emph{geometric braid}. In this case we can think of the vertical axis as
`time' and the horizontal axes as `space.' The strings cannot occupy the same
point in space at the same time, and they can't reverse direction (i.e., they
can't go back in time). We consider two braids to be equal if their strings
can be deformed into each other, with no string crossing another. The points
on the bottom plane where the strings emanate in Fig.~\ref{fig:geombraid} are
the same as the points on the top plane. In that sense braids naturally
represent \emph{periodic orbits} of two-dimensional dynamical systems. When
slicing the braid horizontally at any given time we get a collection of points
in the 2D plane, corresponding to the strings. We call these points
\emph{punctures} or \emph{particles}. %
\index{punctures}
In fact the set of all braids with a given number of strings, and the same
anchorpoints, form a \emph{group}. The group multiplication law is simply to
lay one braid after another; it is easy to see that this is associative. The
identity braid consists of straight strings that are unentangled with each
other. The inverse of a braid is obtained by reversing time. (See the book
by \citeauthor{Birman1975} for more details.) \index{braid!geometric|)}%
\index{braid!algebraic|(}%
A convenient way to represent a braid is to deform it %
as in Fig,~\ref{fig:sbdiag}. Here the geometric braid is combed such that
only one crossing occurs at a time, and each crossing fits in a time interval
of the same length. Such a picture is called a \emph{braid diagram}. The
operation of the~$i$th string (counted from left to right) being exchanged
with the~$(i+1)$th string, such that the left string passes over the right, is
called a %
\index{braid!generator|(}%
\emph{generator} of the braid group, as is denoted~$\sigma_i$. Any element of
the braid group on~$\nn$ strings can be written as a product of the
generators~$\{\sigma_1,\ldots,\sigma_{\nn-1}\}$ and their inverses. Thus, the
braid in Fig.~\ref{fig:sbdiag} can be written
\begin{equation}
\sigma_3\,\sigma_2^{-1}\sigma_1^{-1}\,\sigma_2\sigma_3^{-1}
\sigma_2\sigma_1^{-1}\sigma_3^{-1}\,,
\label{eq:sbdiag}
\end{equation}
where we read the generators from left to right, and from bottom to top in
Fig.~\ref{fig:sbdiag}. A braid written in terms of generators as
in~\eqref{eq:sbdiag} is called an \emph{algebraic braid}.
\index{braid!algebraic|)}%
\begin{figure}
\begin{center}
\subfigure[]{
\raisebox{.09\textwidth}{%
\includegraphics[width=.35\textwidth]{13is31}}
\label{fig:13is31}
}\hspace{.1\textwidth}
\subfigure[]{
\includegraphics[width=.35\textwidth]{121is212}
\label{fig:121is212}
}
\end{center}
\caption{(a) Generators that don't share a string commute. (b) The braid
relation.}
\label{fig:relations}
\end{figure}
The generators obey some special rules, called \emph{relations}, by virtue of
arising from geometrical braids. The relations are
\begin{equation}
\sigma_j\sigma_k=\sigma_k\sigma_j, \quad \lvert j-k\rvert > 1; \qquad
\sigma_j\sigma_k\sigma_j = \sigma_k\sigma_j\sigma_k, \quad \lvert j-k\rvert=1.
\label{eq:relations}
\end{equation}
The first type of relation, depicted in Fig.~\ref{fig:13is31}, says that
generators commute if they don't share a string. The second type, often
called the braid relation, reflects the equality of the two braids shown in
Fig.~\ref{fig:121is212}. While it's obvious from Fig.~\ref{fig:relations}
that the relations~\eqref{eq:relations} are satisfied, it is far less obvious
that those are the \emph{only} relations that hold, as proved by
\citet{Artin1947}. Hence, the relations~\eqref{eq:relations} fully
characterize the braid group for~$\nn$ strings, denoted~$B_\nn$.
\index{braid!generator|)}
\index{braid!group|)}%
\subsection{Constructing a braid from orbit data}
\label{sec:braidfromorbitdata}
\index{braid!from data|(}%
\begin{figure}
\begin{center}
\subfigure[]{
\includegraphics[width=.3\textwidth]{crossing}
\label{fig:crossing}
}\hspace{.2\textwidth}
\subfigure[]{
\includegraphics[width=.3\textwidth]{crossing2}
\label{fig:crossing2}
}
\end{center}
\caption{(a) The two types of generators that occur when particles exchange
positions. (b) A particle exchanging position twice in a row with another
leads to the generators~$\sigma_i$ followed by~$\sigma_i^{-1}$, which
cancel each other. [After~\citet{Thiffeault2005}.]}
\end{figure}
So far we have regarded braids as geometrical objects, and showed how to turn
them into algebraic objects by writing them in terms of generators. But in
practice how do we create braids from particle orbits? That is, if we have
two-dimensional continuous trajectory data arising from some dynamical system,
how do we turn this data into a braid?
\index{projection line|(}%
The technique to do this was described in \citet{Thiffeault2005}. We define
an arbitrary line in the 2D plane, called the \emph{projection line}, and look
at the order of the particles projected along that line. We label each
particle according to its order when projected on this line. In
Fig.~\ref{fig:crossing}, for example, we see three particles labeled~$i$,
$i+1$, $i+2$, where the projection line is the horizontal.
\index{crossing|(}%
A \emph{crossing} occurs whenever any two particles exchange position (only
adjacent particles can exchange position at a given time, since the
trajectories are continuous). For particle~$i$ exchanging position with
particle~$i+1$, we assign a generator~$\sigma_i$ to the crossing if the
particle on the left passes above the one on the right, and~$\sigma_i^{-1}$ if
it passes below (Fig.~\ref{fig:crossing}). As we watch the trajectories
unfold, we construct an algebraic braid as a sequence of generators, with each
generator corresponding to a crossing, and their order determined by when each
crossing occurs. Note that when two particles exchange position along the
crossing line twice as in Fig.~\ref{fig:crossing2}, without exchanging
positions vertically, then the two crossings yield the generators~$\sigma_i$
followed by~$\sigma_i^{-1}$, which cancel. %
\index{projection line|)}%
\index{crossing|)}%
The outcome of this procedure is not a true geometric braid, since the
particles do not necessarily end up in the same positions as they started.
(If the particles are part of a periodic orbit, then we do obtain a geometric
braid.) We do however obtain an algebraic braid, and as long as we are
careful to take long enough trajectories the fact that the geometric braid
does not close will not be too consequential.
\index{braid!from data|)}%
\section{A tour of \texttt{braidlab}}%{\lstinline{braidlab}}
\label{sec:tour}
You will need access to a recent version of Matlab to use \texttt{braidlab}}%{\lstinline{braidlab}. See
Appendix~\ref{sec:install} for instructions on how to install \texttt{braidlab}}%{\lstinline{braidlab}\ on
your machine.
\subsection{The \texttt{braid}}%{\lstinline{braid}\ class}
\label{sec:braidclass}
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class|(}
\subsubsection{Constructor and elementary operations}
\texttt{braidlab}}%{\lstinline{braidlab}\ defines a number of classes, most importantly \texttt{braid}}%{\lstinline{braid}\ and \texttt{loop}}%{\lstinline{loop}.
The braid~$\sigma_1\sigma_2^{-1}$ is constructed with %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!constructor|(}
\begin{lstlisting}[frame=single,framerule=0pt]
>> a = braid([1 -2])
a = < 1 -2 >
\end{lstlisting}
which defaults to the minimum required strings,~$3$. The same braid
on~$4$ strings is constructed with
\begin{lstlisting}[frame=single,framerule=0pt]
> a4 = braid([1 -2],4)
a4 = < 1 -2 >
\end{lstlisting}
Two braids can be multiplied: %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!multiplication (\lstinline{*})}
\begin{lstlisting}[frame=single,framerule=0pt]
>> a = braid([1 -2]); b = braid([1 2]);
>> a*b, b*a
ans = < 1 -2 1 2 >
ans = < 1 2 1 -2 >
\end{lstlisting}
Powers %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!power (\lstinline{^})}%
can also be taken, including the inverse: %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!inverse (\lstinline{inv})}%
\begin{lstlisting}[frame=single,framerule=0pt]
>> a^5, inv(a), a*a^-1
ans = < 1 -2 1 -2 1 -2 1 -2 1 -2 >
ans = < 2 -1 >
ans = < 1 -2 2 -1 >
\end{lstlisting}
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!identity braid}%
Note that this last expression is the identity braid, but is not simplified.
The method \lstinline{compact} attempts to simplify the braid: %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!compact@\lstinline{compact}}
\begin{lstlisting}[frame=single,framerule=0pt]
>> compact(a*a^-1)
ans = < e >
\end{lstlisting}
The method \lstinline{compact} is based on the heuristic algorithm
of~\citet{Bangert2002}, since finding the braid of minimum length in the
standard generators is in general difficult~\citep{Paterson1991}. Hence,
there is no guarantee that in general \lstinline{compact} will find the
identity braid, even though it does so here. To really test if a braid is the
identity (trivial braid), use the method \lstinline{istrivial}: %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!istrivial@\lstinline{istrivial}}
\begin{lstlisting}[frame=single,framerule=0pt]
>> istrivial(a*a^-1)
ans = 1
\end{lstlisting}
The number of strings is %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!number of strings (\lstinline{n})}
\begin{lstlisting}[frame=single,framerule=0pt]
>> a.n
ans = 3
\end{lstlisting}
Note that
\index{help@\lstinline{help}|(}
\begin{lstlisting}[frame=single,framerule=0pt]
>> help braid
\end{lstlisting}
describes the class \texttt{braid}}%{\lstinline{braid}. To get more information on the \texttt{braid}}%{\lstinline{braid}\
constructor, invoke%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!constructor}%
\begin{lstlisting}[frame=single,framerule=0pt]
>> help braid.braid
\end{lstlisting}
which refers to the method \texttt{braid}}%{\lstinline{braid}\ within the class \texttt{braid}}%{\lstinline{braid}. %
\index{help@\lstinline{help}|)}%
(Use \lstinline{methods(braid)} to list all the methods in the class.) There
are other ways to construct a \texttt{braid}}%{\lstinline{braid}, such as using random %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!constructor!random braid} generators, here a
braid with~$5$ strings and~$10$ random generators:
\begin{lstlisting}[frame=single,framerule=0pt]
>> braid('Random',5,10)
ans = < 1 4 -4 2 4 -1 -2 4 4 4 >
\end{lstlisting}
The constructor can also build some standard braids: %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!constructor!half-twist}
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!constructor!knots}
\index{knot!braid representative}
\begin{lstlisting}[frame=single,framerule=0pt]
>> braid('HalfTwist',5)
ans = < 4 3 2 1 4 3 2 4 3 4 >
>> braid('8_21')
ans = < 4 3 2 1 4 3 2 4 3 4 >
\end{lstlisting}
In Section~\ref{sec:braidfromdata} we will show how to construct a braid from
a trajectory data set. \index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!constructor|)}
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!equality (\lstinline{==})}%
The \texttt{braid}}%{\lstinline{braid}\ class handles equality of braids:
\begin{lstlisting}[frame=single,framerule=0pt]
>> a = braid([1 -2]); b = braid([1 -2 2 1 2 -1 -2 -1]);
>> a == b
ans = 1
\end{lstlisting}
These are the same braid, even though they appear different from their
generator sequence~\citep{Birman1975}. Equality is determined efficiently
by %
\index{loop!coordinates}%
\index{Dynnikov coordinates|see{loop coordinates}}%
acting on loop coordinates~\citep{Dynnikov2002}, as described by
\citet{Dehornoy2008}. See Sections~\ref{sec:loop}--\ref{sec:loopcoords} for
more details. If for some reason lexicographic (generator-per-generator)
equality of braids is needed, use the method \lstinline{lexeq(b1,b2)}. %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!lexeq class@\lstinline{lexeq}}
We can extract a subbraid %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!subbraid class@\lstinline{subbraid}|(} by
choosing specific strings: for example, if we take the~$4$-string
braid~$\sigma_1\sigma_2\sigma_3^{-1}$ and discard the third string, we
obtain~$\sigma_1\sigma_2^{-1}$:
\begin{lstlisting}[frame=single,framerule=0pt]
>> a = braid([1 2 -3]);
>> subbraid(a,[1 2 4])
ans = < 1 -2 >
\end{lstlisting}
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!subbraid class@\lstinline{subbraid}|)}
The opposite of subbraid is the \emph{tensor product}, the larger braid
obtained by laying two braids side-by-side \citep{KasselTuraev}: %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!tensor@\lstinline{tensor}|(}%
\begin{lstlisting}[frame=single,framerule=0pt]
>> a = braid([1 2 -3]); b = braid([1 -2]);
>> tensor(a,b)
ans = < 1 2 -3 5 -6 >
\end{lstlisting}
Here, the tensor product of a 4-braid and a 3-braid has 7 strings. The
generators $\sigma_1\sigma_2^{-1}$ of \lstinline{b} became
$\sigma_5\sigma_6^{-1}$ after re-indexing so they appear to the right of
\lstinline{a}.
\subsubsection{Topological entropy and complexity}
\label{sec:entropy}
There are a few methods that exploit the connection between braids and
homeomorphisms \index{homeomorphism} of the punctured disk. %
\index{disk, punctured}%
\index{punctures}
Braids label \emph{isotopy classes} %
\index{homeomorphism!isotopy classes} of homeomorphisms, so we can assign a
topological entropy %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!entropy@\lstinline{entropy}|(}%
\index{braid!entropy|(}%
\index{topological entropy|see{braid entropy}}%
\index{entropy|see{braid entropy}}%
to a braid:
\begin{lstlisting}[frame=single,framerule=0pt]
>> entropy(braid([1 2 -3]))
ans = 0.8314
\end{lstlisting}
\index{action!of braid on loop}%
The entropy is computed by iterated action on a loop~\citep{Moussafir2006}.
This can fail if the braid is finite-order %
\index{braid!finite-order|(}%
or has very low entropy:
\begin{lstlisting}[frame=single,framerule=0pt]
>> entropy(braid([1 2]))
Warning: Failed to converge to requested tolerance; braid is likely finite-order or has low entropy. Returning zero entropy.
ans = 0
\end{lstlisting}
To force the entropy to be computed using the Bestvina--Handel train track
algorithm~\citep{Bestvina1995}, %
\index{Bestvina--Handel algorithm|(}%
we add an optional \lstinline{'Method'} parameter:
\begin{lstlisting}[frame=single,framerule=0pt]
>> entropy(braid([1 2]),'Method','trains')
ans = 0
\end{lstlisting}
\index{braid!finite-order|)}%
Note that for large braids the Bestvina--Handel algorithm is impractical. But
when applicable it can also determine the Thurston--Nielsen type %
\index{braid!Thurston--Nielsen type}%
of the braid~\citep{Fathi1979,Thurston1988,Casson1988,Boyland1994}:
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!tntype@\lstinline{tntype}|(}%
\index{braid!pseudo-Anosov|(}%
\index{braid!reducible|(}%
\begin{lstlisting}[frame=single,framerule=0pt]
>> tntype(braid([1 2 -3]))
ans = pseudo-Anosov
>> tntype(braid([1 2]))
ans = finite-order
>> tntype(braid([1 2],4))
ans = reducible
\end{lstlisting}
\index{braid!pseudo-Anosov|)}%
\index{braid!reducible|)}%
\texttt{braidlab}}%{\lstinline{braidlab}\ uses Toby Hall's implementation of the Bestvina--Handel
algorithm~\citep{HallTrain}. %
\index{Bestvina--Handel algorithm|)}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!tntype@\lstinline{tntype}|)}%
The topological entropy is a measure of braid complexity that relies on %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!entropy@\lstinline{entropy}|)}%
\index{braid!entropy|)}%
iterating the braid. It gives the maximum growth rate of a `rubber band'
anchored on the braid, as the rubber band slides up many repeated copies of
the braid. For finite-order braids, %
\index{braid!finite-order}%
this will converge to zero. The \emph{geometric complexity} %
\index{braid!complexity|(}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!complexity@\lstinline{complexity}|(}%
\index{geometric complexity|see{braid complexity}}%
\index{complexity|see{braid complexity}}%
of a braid~\citep{Dynnikov2007}, is defined in terms of the $\log_2$ of the
number of intersections of a set of curves with the real axis, after one
application of the braid:
\begin{lstlisting}[frame=single,framerule=0pt]
>> complexity(braid([1 -2]))
ans = 2
>> complexity(braid([1 2]))
ans = 1.5850
\end{lstlisting}
See Section~\ref{sec:loop} or `\lstinline{help braid.complexity}' for details
on how the geometric complexity is computed. %
\index{braid!complexity|)}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!complexity@\lstinline{complexity}|)}%
\subsubsection{Representation and invariants}
There are a few remaining methods in the braid class, which we describe
briefly. The reduced Burau matrix
representation~\citep{Burau1936,Birman1975} %
\index{braid!Burau representation|(}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!burau@\lstinline{burau}|(}%
of a braid is obtained with the method \lstinline{burau}:
\begin{lstlisting}[frame=single,framerule=0pt]
>> burau(braid([1 -2]),-1)
ans = 1 -1
-1 2
\end{lstlisting}
where the last argument ($-1$) is the value of the parameter~$t$ in the
Laurent polynomials %
\index{Laurent polynomials|(}%
that appear in the entries of the Burau matrices. With access to Matlab's
wavelet toolbox, %
\index{Matlab!wavelet toolbox}%
\index{laurpoly@\lstinline{laurpoly}|(}%
we can use actual Laurent polynomials as the entries:
\begin{lstlisting}[frame=single,framerule=0pt]
>> B = burau(braid([1 -2]),laurpoly(1,1))
| - z^(+1) z^(+1) |
| |
B = | |
| |
| - 1 + 1 - z^(-1) |
\end{lstlisting}
but the matrix is now given as a cell array%
\index{Matlab!cell array}%
\index{cell array|see{Matlab cell array}}%
\footnote{%
\index{Matlab!cell array}%
A Matlab cell array is similar to a numeric array, except that its entries
can hold any data, not just numeric. The entries are indexed as
\lstinline{a\{1,2\}} rather than \lstinline{a(1,2)}, and matrix operations
like multiplication are not defined.} %
, %
each entry containing a \lstinline{laurpoly} object:
\begin{lstlisting}[frame=single,framerule=0pt]
>> B{2,2}
ans(z) = + 1 - z^(-1)
\end{lstlisting}
Instead of \lstinline{laurpoly} objects, we can use Matlab's symbolic
toolbox: \index{Matlab!symbolic toolbox}%
\begin{lstlisting}[frame=single,framerule=0pt]
>> B = burau(braid([1 -2]),sym('t'))
B = [ -t, t]
[ -1, 1 - 1/t]
\end{lstlisting}
where now \lstinline{B} is a matrix of \lstinline{sym} objects:
\begin{lstlisting}[frame=single,framerule=0pt]
>> B(2,2)
ans = 1 - 1/t
\end{lstlisting}
\index{braid!Lawrence--Krammer representation|(}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!lk@\lstinline{lk}|(}%
Another well-known homological representation of braid groups is the
Lawrence--Krammer representation \citep{Lawrence1990,Bigelow2001}. It is
given in terms of two parameters, usually denoted~$t$ and~$q$:
\begin{lstlisting}[frame=single,framerule=0pt]
>> K = lk(braid([1 -2]),sym('t'),sym('q'))
K = [ -(q-1)^2/q - q*t*(q-1), -(q-1)/q, (q^2-q+1)/q^2]
[ -q^2*t, 0, 0]
[ -(q-1)/(q*t), -1/(q*t), (q-1)/(q^2*t)]
\end{lstlisting}
In this case there we cannot use \lstinline{laurpoly} entries, since the
representation involves Laurent polynomials in two symbols. For this reason,
and because its size grows more rapidly with the number of strings (matrices
of dimension $\tfrac12\nn(\nn-1)$), the Lawrence--Krammer representation is
very slow to compute for large braids. %
\index{braid!Lawrence--Krammer representation|)}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!lk@\lstinline{lk}|)}%
The reduced Burau matrix of a braid can be used to compute the
\emph{Alexander--Conway polynomial} (or Alexander polynomial for short) %
\index{Alexander--Conway polynomial|(}%
of its closure. For instance, the trefoil knot is given by the closure of
the %
\index{knot!trefoil}%
\index{knot!Alexander polynomial}%
braid~$\sigma_1^3$ \citep{AlexanderPolynomial}, which gives a Laurent
polynomial
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!alexpoly@\lstinline{alexpoly}|(}%
\begin{lstlisting}[frame=single,framerule=0pt]
>> alexpoly(braid([1 1 1]))
ans(z) = + z^(+2) - z^(+1) + 1
\end{lstlisting}
The figure-eight knot is the closure of~$(\sigma_1\sigma_2^{-1})^2$:
\index{knot!figure-eight}
\begin{lstlisting}[frame=single,framerule=0pt]
>> alexpoly(braid([1 -2 1 -2]))
ans(z) = - 1 + 3*z^(-1) - z^(-2)
\end{lstlisting}
This can be `centered' so that it satisfies~$p(z)=\pm p(1/z)$:
\begin{lstlisting}[frame=single,framerule=0pt]
>> alexpoly(braid([1 -2 1 -2]),'Centered')
ans(z) = - z^(+1) + 3 - z^(-1)
\end{lstlisting}
The centered Alexander polynomial is a knot invariant, \index{knot!invariant}
so it can be used to determine when two knots are not the same. For knots,
the centered polynomial is guaranteed to have integral powers. For links,
such as the Hopf link consisting of two singly-linked loops, it might not:
\begin{lstlisting}[frame=single,framerule=0pt]
>> alexpoly(braid([1 1]),'Centered')
Error using braidlab.braid/alexpoly
Polynomial with fractional powers. Remove 'Centered' option or use the symbolic toolbox.
\end{lstlisting}
\index{Matlab!symbolic toolbox}%
Fractional powers cannot be represented with a \lstinline{laurpoly} object. %
\index{laurpoly@\lstinline{laurpoly}|)}%
\index{Laurent polynomials|)}%
In that case we can drop the \lstinline{Centered} option, which yields the uncentered
polynomial $1-z$. Alternatively, we can switch to using a variable from the
symbolic toolbox:
\begin{lstlisting}[frame=single,framerule=0pt]
>> alexpoly(braid([1 1]),sym('x'),'Centered')
ans = 1/x^(1/2) - x^(1/2)
\end{lstlisting}
\index{sym@\lstinline{sym}}%
which can represent fractional powers. This polynomial
satisfies~$p(x)=-p(1/x)$.
\index{braid!Burau representation|)}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!burau@\lstinline{burau}|)}%
\index{Alexander--Conway polynomial|)}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!alexpoly@\lstinline{alexpoly}|)}%
The method \lstinline{perm} %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!perm@\lstinline{perm}|(}%
gives the permutation of strings corresponding to a braid: %
\begin{lstlisting}[frame=single,framerule=0pt]
>> perm(braid([1 2 -3]))
ans = 2 3 4 1
\end{lstlisting}
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!perm@\lstinline{perm}|)}%
If the strings are unpermuted, then the braid is \emph{pure}, %
\index{braid!pure}%
which can also be tested with the method \lstinline{ispure}. %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!ispure@\lstinline{ispure}}%
Finally, the \emph{writhe} %
\index{braid!writhe}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!writhe@\lstinline{writhe}|(}%
of a braid is the sum of the powers of its generators. The writhe of
$\sigma_1^{+1}\sigma_2^{+1}\sigma_3^{-1}$ is $+1+1-1 = 1$:
\begin{lstlisting}[frame=single,framerule=0pt]
>> writhe(braid([1 2 -3]))
ans = 1
\end{lstlisting}
The writhe is a braid invariant.
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!writhe@\lstinline{writhe}|)}%
\subsubsection{The \texttt{annbraid}}%{\lstinline{anbraid}\ subclass}
\label{sec:annbraid}
\index{annbraid@\lstinline{annbraid}|(}
\index{punctures!in annulus|(}
\begin{figure}
\begin{center}
\subfigure[]{
\includegraphics[width=.4\textwidth]{annulardomain}
\label{fig:annulardomain}
}\hspace{1em}
\subfigure[]{
\raisebox{3.75em}{\includegraphics[width=.5\textwidth]{annulardomain2}}
\label{fig:annulardomain2}
}
\end{center}
\caption{(a) Punctures in an annular domain, with two generators. The
generator~$\Sigma_\nn$ is unique to the annulus. (b) The punctures
rearranged with the center of the annulus as an extra puncture on the
right, showing how the generator~$\Sigma_\nn$ can be deformed in terms of
standard generators as in~\eqref{eq:Sigman}.}
\end{figure}
Often it is useful to consider braids in an annular domain, as in
Fig.~\ref{fig:annulardomain}. Since it doesn't change the overall topology,
it is convenient to rearrange the punctures as in
Fig.~\ref{fig:annulardomain2}. The center of the annulus becomes an extra
puncture, but that extra puncture is fixed. For~$\nn$ moving punctures, the
braid group on the annulus has~$\nn$
generators~$\{\Sigma_1,\ldots,\Sigma_\nn\}$, one more than for the standard
braid group owing to the fixed puncture. These are related to the standard
generators by~$\Sigma_i=\sigma_i$ for~$1 \le i <\nn$, and
\begin{equation}
\Sigma_\nn = \sigma_\nn^2 \sigma_{\nn-1} \cdots \sigma_2\,\sigma_1\,
\sigma_2^{-1} \cdots \sigma_{\nn-1}^{-1} \sigma_\nn^{-2}\,.
\label{eq:Sigman}
\end{equation}
This last generator effectively exchanges puncture~$\nn$ with puncture~$1$,
exploiting the annular topology.
\texttt{braidlab}}%{\lstinline{braidlab}\ support annular braids with the subclass \texttt{annbraid}}%{\lstinline{anbraid}, derived from
\texttt{braid}}%{\lstinline{braid}. The syntax for creating an annular braid is
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = annbraid([1 2 -3])
b = < 1 2 -3 >*
\end{lstlisting}
The asterisk indicates that this is an annular braid, which has an extra fixed
puncture on the right (called the basepoint). Hence, the braid has~$4$
punctures, as indicated by
\begin{lstlisting}[frame=single,framerule=0pt]
>> b.n
ans = 4
\end{lstlisting}
but only~$3$ punctures can move, as returned by \lstinline{nann}, the number
of annular punctures:
\begin{lstlisting}[frame=single,framerule=0pt]
>> b.nann
ans = 3
\end{lstlisting}
Many of the methods work described for the \lstinline{braid} class can be
applied to \texttt{annbraid}}%{\lstinline{anbraid}{}s. For instance,
\begin{lstlisting}[frame=single,framerule=0pt]
>> entropy(braid([1 -2])))
ans = 0.9624
>> entropy(annbraid([1 -2])))
ans = 1.7627
\end{lstlisting}
The annular braid has more entropy, since curves grow faster by getting
entangled on the extra puncture.
\begin{figure}
\begin{center}
\includegraphics[width=.7\textwidth]{annbraid_s1s-2}
\end{center}
\caption{The output of \lstinline{plot(annbraid([1 -2]))}. The green strand
represents the center of the annulus, as in
Fig.~\ref{fig:annulardomain2}. (See Section~\ref{sec:prop} for how to plot
braids sideways.)}
\label{fig:annbraid_s1s-2}
\end{figure}
The annular braid is shown in Fig.~\ref{fig:annbraid_s1s-2}.
\index{punctures!in annulus|)}
\index{annbraid@\lstinline{annbraid}|)}
\subsection{Constructing a braid from data}
\label{sec:braidfromdata}
\subsubsection{An example}
\label{sec:braidfromdataex}
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!constructor!from data|(}%
\index{braid!from data|(}%
One of the main purposes of \texttt{braidlab}}%{\lstinline{braidlab}\ is to analyze two-dimensional
trajectory data using braids. We can assign a braid to trajectory data by
looking for \emph{crossings} %
\index{crossing}%
along a projection line (see \citet{Thiffeault2005,Thiffeault2010} and
Section~\ref{sec:braidfromdata}). %
\index{projection line|(}%
The \texttt{braid}}%{\lstinline{braid}\ constructor allows us to do this easily.
The folder \lstinline{testsuite/testcases} %
\index{testsuite}%
contains a dataset of trajectories, from laboratory data for granular
media~\citep{Puckett2012}. We load the
data:
\begin{lstlisting}[frame=single,framerule=0pt]
>> clear; load testdata
>> whos
Name Size Bytes Class Attributes
XY 9740x2x4 623360 double
ti 1x9740 77920 double
\end{lstlisting}
Here \lstinline{ti} is the vector of times, and \lstinline{XY} is a
three-dimensional array: its first component specifies the timestep,
its second specifies the $X$ or $Y$ coordinate, and its third
specifies one of the~$4$ particles. Figure~\ref{fig:testdata_trajs3}
shows
\begin{figure}
\begin{center}
\subfigure[]{
\includegraphics[height=.3\textheight]{testdata_trajs3}
\label{fig:testdata_trajs3}
}\hspace{1em}
\subfigure[]{
\includegraphics[height=.3\textheight]{testdata_trajs}
\label{fig:testdata_trajs}
}\hspace{1em}
\subfigure[]{
\includegraphics[height=.3\textheight]{testdata_braid}
\label{fig:testdata_braid}
}\hspace{1em}
\subfigure[]{
\includegraphics[height=.3\textheight]{testdata_braidY}
\label{fig:testdata_braidY}
}
\end{center}
\caption{(a) A dataset of four trajectories, (b) projected along the~$X$ axis.
(c) The compacted braid~$ \sigma_2^{-2} \sigma_1^{-1} \sigma_2^{-1}
\sigma_1^{-5} \sigma_3 \sigma_1^{-1}\sigma_3\sigma_2\sigma_1$ corresponding
to the~$X$ projection in (b). (d) The compacted
braid~$\sigma_3^{-4}\sigma_1\sigma_3^{-1}\sigma_1\sigma_3^{-3}$
corresponding to the~$Y$ projection, with closure enforced. %
\index{braid!closure}%
The braids in (c) and (d) are conjugate.\index{braid!conjugate}}
\end{figure}
the~$X$ and~$Y$ coordinates of these four trajectories, with time
plotted vertically. Figure~\ref{fig:testdata_trajs} shows the same
data, but projected along the~$X$ direction. To construct a braid
from this data, we simply execute %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!length@\lstinline{length}}
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = braid(XY);
>> b.length
ans = 894
\end{lstlisting}
This is a very long braid! But Figure~\ref{fig:testdata_trajs} suggests that
this is misleading: many of the crossings %
\index{crossing}%
are `wiggles' that cancel each other out. Indeed, if we attempt to shorten
the braid: %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!compact@\lstinline{compact}}%
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = compact(b)
b = < -2 -2 -1 -2 -1 -1 -1 -1 -1 3 -1 3 2 1 >
>> b.length
ans = 14
\end{lstlisting}
we find the number of generators (the length) has dropped to~$14$! We can
then plot this shortened braid as a braid diagram using \lstinline{plot(b)} %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!plot@\lstinline{plot}} to produce
Figure~\ref{fig:testdata_braid}. The braid diagram allows us to see some
topological information clearly, such as the fact that the second and third
particles undergo a large number of twists around each other; we can check
this by creating a subbraid %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!subbraid class@\lstinline{subbraid}}%
with only those two strings:
\begin{lstlisting}[frame=single,framerule=0pt]
>> subbraid(b,[2 3])
ans = < -1 -1 -1 -1 -1 -1 -1 -1 >
\end{lstlisting}
which shows that the winding number between these two strings is~$-4$.
\subsubsection{Changing the projection line and enforcing closure}
\label{sec:projection}
The braid in the previous section was constructed from the data by assuming a
projection along the~$X$ axis (the default). We can choose a different
projection by specifying an optional angle for the projection line; for
instance, to project along the~$Y$ axis we invoke
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = braid(XY,pi/2);
>> b.length
ans = 673
>> b.compact
ans = < -3 -3 -3 -3 1 -3 -3 -3 -3 >
\end{lstlisting}
In general, a change of projection line only changes the braid by
conjugation~\citep{Boyland1994,Thiffeault2010}. We can test for
conjugacy: %
\index{braid!conjugate|(}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!compact@\lstinline{compact}|(}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!conjtest@\lstinline{conjtest}|(}%
\begin{lstlisting}[frame=single,framerule=0pt]
>> bX = compact(braid(XY,0)); bY = compact(braid(XY,pi/2));
>> conjtest(bX,bY)
ans = 0
\end{lstlisting}
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!compact@\lstinline{compact}|)}%
The braids are not conjugate. This is because our trajectories do not
form a `true' braid: the final points do not correspond exactly with
the initial points, as a set. If we truly want a
rotationally-conjugate braid out of our data, we need to enforce a
closure method: %
\index{braid!closure|(}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!closure@\lstinline{closure}|(}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!compact@\lstinline{compact}}%
\index{crossing!in braid closure|(}
\begin{lstlisting}[frame=single,framerule=0pt]
>> XY = closure(XY);
>> bX = compact(braid(XY,0)), bY = compact(braid(XY,pi/2))
bX = < -2 -2 -1 -2 -1 -1 -1 -1 -1 3 -1 3 2 1 >
bY = < -3 -3 -3 -3 1 -3 1 -3 -3 -3 >
\end{lstlisting}
This default closure simply draws line segments from the final points to the
initial points in such a way that no new crossings are created in the~$X$
projection. %
\index{crossing!in braid closure|)}%
Hence, the $X$-projected braid \lstinline{bX} is unchanged by the closure, but
here the $Y$-projected braid \lstinline{bY} is longer by one generator
(\lstinline{bY} is plotted in Figure~\ref{fig:testdata_braidY}). This is
enough to make the braids conjugate:
\begin{lstlisting}[frame=single,framerule=0pt]
>> [~,c] = conjtest(bX,bY)
c = < 3 2 >
\end{lstlisting}
\index{tilde@tilde (\lstinline{~}), as return argument}
where the optional second argument \lstinline{c} is the conjugating
braid, as we can verify:
\begin{lstlisting}[frame=single,framerule=0pt]
>> bX == c*bY*c^-1
ans = 1
\end{lstlisting}
There are other ways to enforce closure of a braid (see
\lstinline{help closure}), in particular
\lstinline{closure(XY,'MinDist')}, which minimizes the total distance
between the initial and final points.
\index{projection line|)}%
\index{braid!closure|)}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!closure@\lstinline{closure}|)}%
Note that \lstinline{conjtest} uses the library \emph{CBraid} \citep{CBraid} %
\index{CBraid}%
to first convert the braids to Garside canonical form \citep{Birman2005}, %
\index{braid!Garside form}%
then to determine conjugacy. This is very inefficient, so is impractical for
large braids.
\index{braid!conjugate|)}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!conjtest@\lstinline{conjtest}|)}%
\subsubsection{The \lstinline{databraid} subclass}
\label{sec:databraid}
In some instances when dealing with data it is important to know the
\emph{crossing times}, %
\index{crossing!times|(}%
that is, the times at which two particles exchanged position along the
projection line. %
\index{projection line}%
A braid object does not keep this information, but there is an object that
does: a \lstinline{databraid}. %
\index{databraid class@\lstinline{databraid} class|(}%
Its constructor takes an optional vector of times as an argument, and it has a
data member \lstinline{tcross} %
\index{tcross@\lstinline{tcross}|(}%
that retains the crossing times. Using the same data \lstinline{XY} from
before, sampled at times \lstinline{ti}, we have
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = databraid(XY,ti);
>> b.tcross(1:3)
ans = 870.9010
872.1758
887.0089
\end{lstlisting}
\index{tcross@\lstinline{tcross}|)}%
Storing crossing times enables us to truncate generators in a
\lstinline{databraid} by retaining only those with crossing times within a
desired interval (see \lstinline{databraid.trunc}). There are always exactly
as many crossing times as generators in the braid. %
\index{databraid class@\lstinline{databraid} class!trunc@\lstinline{trunc}}%
\index{crossing!times|)}%
Many operations that can be done to a \lstinline{braid} also work on a
\lstinline{databraid}, with a few differences:
\begin{itemize}
\item\lstinline{compact} %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!compact@\lstinline{compact}}%
works a bit differently. It is less effective than \lstinline{braid.compact}
since it must preserve the order of generators in order to maintain the
ordering of the crossing times.
\item Equality testing checks if two \hbox{\lstinline{databraid}s} are
lexicographically equal (i.e., generator-by-generator) and that their
crossing times all agree. This is very restrictive. To check if the
underlying braids are equal, first convert the \hbox{\lstinline{databraid}s}
to \hbox{\lstinline{braid}s} by using the method
\lstinline{databraid.braid}.
\item Multiplication of two \hbox{\lstinline{databraid}s} is only defined if
the crossing times of the first braid are all earlier than the second.
\item Powers and inverses of \hbox{\lstinline{databraid}s} are not defined,
since this would break the time-ordering of crossings.
\item%
\index{Finite Time Braiding Exponent (FTBE)|(}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!entropy@\lstinline{entropy}}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!complexity@\lstinline{complexity}}%
The entropy of a \lstinline{databraid} is an ambiguously defined concept.
While entropy of certain braids can be computed non-iteratively, e.g., in
\citet{Hall2009}, in general it is only estimated by an iterative process.
Iterations rely on taking powers of the braid, which is not defined for
\hbox{\lstinline{databraid}s}. The functions \hbox{\lstinline{entropy}} and
\hbox{\lstinline{complexity}} can still be used by converting
\lstinline{databraid} objects to \lstinline{braid} objects; however, this
should be avoided in favor of the appropriate concept for
\hbox{\lstinline{databraid}s},
the Finite Time Braiding Exponent (FTBE) \citep{Thiffeault2005,Budisic2015}.
A \lstinline{databraid}~$b_T$ recorded over a time interval of length~$T$
has an FTBE defined by
\begin{equation}
\FTBE(b_{T})
= \frac{1}{T}\, \log \frac{ \lvert b_{T}\, \ell \rvert}{\lvert\ell\rvert},
\label{eq:ftbe}
\end{equation}
where $\ell$ is a loop given by the generating set for the fundamental group
of the disk with $n$ \index{punctures} punctures (see
Section~\ref{sec:loopcoords}), and~$b_{T} \ell$ is that loop transformed by
a single application of the braid. Here~$\lvert\cdot\rvert$ is a measure of
the length of the loop. Unless specified otherwise, \texttt{braidlab}}%{\lstinline{braidlab}\ calculates
\(T\) as the time elapsed between the first and last crossing in the
braid. This duration could be much smaller than the length of trajectories
analyzed, e.g., when no crossings occur near the beginning or the end of
trajectories. To set the custom value of \(T\) and other options, see
documentation of the method \lstinline{databraid.ftbe}.
\end{itemize}
\index{databraid class@\lstinline{databraid} class|)}%
\index{databraid class@\lstinline{databraid} class!ftbe@\lstinline{ftbe}}%
\index{Finite Time Braiding Exponent (FTBE)|)}
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!constructor!from data|)}
\index{braid!from data|)}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class|)}
\subsection{The \texttt{loop}}%{\lstinline{loop}\ class}
\label{sec:loop}
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class|(}
\subsubsection{Loop coordinates}
\label{sec:loopc}
A simple closed loop on a disk with~$5$ punctures %
\index{disk, punctured}%
\index{punctures}%
is shown in Figure~\ref{fig:dynn_loop}.
\begin{figure}
\begin{center}
\subfigure[]{
\includegraphics[height=.22\textheight]{dynn_loop}
\label{fig:dynn_loop}
}\hspace{1em}
\subfigure[]{
\includegraphics[height=.22\textheight]{dynn_def}
\label{fig:dynn_def}
}
\end{center}
\caption{(a) A simple close loop in a disk with~$\nn=5$ punctures.
(b) Definition of intersection numbers~$\mu_i$ and~$\nu_i$.
[From~\citet{Thiffeault2010}.] \index{loop!intersection numbers}}
\end{figure}
We consider equivalence classes of such loops under homotopies %
\index{loop!homotopy classes}%
relative to the punctures. %
\index{punctures}%
In particular, the loops are \emph{essential}, %
\index{loop!essential}%
meaning that they are not null-homotopic or homotopic to the boundary or a
puncture. The \emph{intersection numbers} %
\index{loop!intersection numbers|(}%
\index{intersection numbers|see{loop intersection numbers}}%
are also shown in Figure~\ref{fig:dynn_loop}: these count the minimum number
of intersections of an equivalence class of loops with the fixed vertical
lines shown. For~$\nn$ punctures, we define the intersection numbers~$\mu_i$
and~$\nu_i$ in Figure~\ref{fig:dynn_def}.
Any given loop will lead to a unique set of intersection numbers, but
a general collection of intersection numbers do not typically
correspond to a loop. %
\index{loop!coordinates|(}
It is therefore more convenient to define
\begin{equation}
\ac_\ip = \tfrac12\l(\mu_{2\ip} - \mu_{2\ip-1}\r), \qquad
\bc_\ip = \tfrac12\l(\nu_\ip - \nu_{\ip+1}\r), \qquad
\ip=1,\ldots,\nn-2.
\end{equation}
We then combine these in a vector of length~$(2\nn-4)$,
\begin{equation}
\abv = (\ac_1,\ldots,\ac_{\nn-2},\bc_1,\ldots,\bc_{\nn-2}),
\label{eq:abvdef}
\end{equation}
which gives the \emph{loop coordinates} (or \emph{Dynnikov coordinates}) for
the loop. (Some authors such as~\citet{Dehornoy2008} give the coordinates
as~$(\ac_1,\bc_1,\ldots,\ac_{\nn-2},\bc_{\nn-2})$.) There is now a bijection
between~$\mathbb{Z}^{2\nn-4}$ and essential simple closed
loops~\citep{Dynnikov2002,Moussafir2006,Hall2009,Thiffeault2010}. Actually, %
\index{multiloop|see{loop, multi-}}%
\index{loop!multi-} \emph{multiloops}: loop coordinates can describe unions of
disjoint loops (see Section~\ref{sec:loopcoords}).%
\footnote{%
Here we use multiloop \index{loop!multi-} as a convenient mnemonic. The
technical term is \emph{integral lamination}: %
\index{integral lamination|see{loop, multi-}}%
a set of disjoint non-homotopic simple closed curves~\citep{Moussafir2006}.}
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!constructor|(}
Let's create the loop in Figure~\ref{fig:dynn_loop} as a \texttt{loop}}%{\lstinline{loop}\ object:
\begin{lstlisting}[frame=single,framerule=0pt]
>> l = loop([-1 1 -2 0 -1 0])
l = (( -1 1 -2 0 -1 0 ))
\end{lstlisting}
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!constructor|)}
Figure~\ref{fig:dynn_loop2} shows the output of the \lstinline{plot(l)} %
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!plot@\lstinline{plot}}
command. We can convert from loop coordinates\index{loop!coordinates} to
intersection numbers with
\begin{lstlisting}[frame=single,framerule=0pt]
>> intersec(l)
ans = 2 0 1 3 4 0 2 2 4 4
\end{lstlisting}
which returns~$\mu_1\dots\mu_{2n-4}$ followed by~$\nu_1\dots\mu_{n-1}$, as
defined in Figure~\ref{fig:dynn_def}.
\index{loop!intersection numbers|)}%
We can also extract the loop coordinates from a \texttt{loop}}%{\lstinline{loop}\ object using the
methods \lstinline{a}, \lstinline{b}, and \lstinline{ab}: %
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!%
abab@\lstinline{a}, \lstinline{b}, \lstinline{ab}}%
\begin{lstlisting}[frame=single,framerule=0pt]
>> l = loop ([-1 1 -2 0 -1 0]);
>> l.a
ans = -1 1 -2
>> l.b
ans = 0 -1 0
>> [a,b] = l.ab
a = -1 1 -2
b = 0 -1 0
\end{lstlisting}
As for braids, \lstinline{l.n} returns the number of punctures (or strings). %
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!number of punctures (\lstinline{n})}
\subsubsection{Acting on loops with braids}
\label{sec:actingonloops}
Now we can act on this loop with braids. %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!action on \texttt{loop}}%{\lstinline{loop}\ (\lstinline{*})|(}%
For example, we define the braid
\lstinline{b} to be~$\sigma_1^{-1}$ with~$5$ strings, corresponding to the~$5$
punctures, %
\index{punctures}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!multiplication (\lstinline{*})|(}
and then act on the loop \lstinline{l} by using the multiplication operator:
\begin{figure}
\begin{center}
\subfigure[]{
\includegraphics[width=.6\textwidth]{dynn_loop2}
\label{fig:dynn_loop2}
}\hspace{1em}
\subfigure[]{
\includegraphics[width=.6\textwidth]{dynn_loop2_sigm1}
\label{fig:dynn_loop2_sigm1}
}
\end{center}
\caption{(a) The loop \lstinline{((-1 1 -2 0 -1 0))}. (b) The braid generator
$\sigma_1^{-1}$ applied to the loop in (a).}
\end{figure}
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = braid([-1],5);
>> b*l
ans = (( -1 1 -2 1 -1 0 ))
\end{lstlisting}
Figure~\ref{fig:dynn_loop2_sigm1} shows \lstinline{plot(b*l)}. The first and
second punctures %
\index{punctures|(}%
were interchanged counterclockwise (the action of~$\sigma_1^{-1}$), dragging
the loop along. %
\index{loop!coordinates|)}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!multiplication (\lstinline{*})|)}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!action on \texttt{loop}}%{\lstinline{loop}\ (\lstinline{*})|)}%
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!minlength@\lstinline{minlength}|(}
\index{loop!minimum length} The minimum length of an equivalence class of
loops is determined by assuming the punctures are one unit of length apart and
have zero size. After pulling tight the loop on the punctures, %
\index{punctures|)}%
it is then made up of unit-length segments. The minimum length is thus an
integer. For the loop in Figure~\ref{fig:dynn_loop2},
\begin{lstlisting}[frame=single,framerule=0pt]
>> minlength(l)
ans = 12
\end{lstlisting}
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!minlength@\lstinline{minlength}|)}
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!intaxis@\lstinline{intaxis}|(}
Another useful measure of a loop's complexity is its minimum intersection
number with the real axis~\citep{Moussafir2006,Hall2009,Thiffeault2010}, which
for this loop is the same as its minimum length:
\begin{lstlisting}[frame=single,framerule=0pt]
>> intaxis(l)
ans = 12
\end{lstlisting}
The \lstinline{intaxis} method is used to measure a braid's geometric
complexity, %
\index{braid!complexity|(}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!complexity@\lstinline{complexity}|(}%
as defined by~\citet{Dynnikov2007}.
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!intaxis@\lstinline{intaxis}|)}
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!constructor|(}
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!vectorized|(}
Sometimes we wish to study a large set of different loops. The loop
constructor vectorizes:
\begin{lstlisting}[frame=single,framerule=0pt]
>> ll = loop([-1 1 -2 0; 1 -2 3 4])
ll = (( -1 1 -2 0 ))
(( 1 -2 3 4 ))
\end{lstlisting}
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!constructor|)}
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!minlength@\lstinline{minlength}|(}
We can then, for instance, compute the length of every loop:
\begin{lstlisting}[frame=single,framerule=0pt]
>> minlength(ll)
ans = 14
34
\end{lstlisting}
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!minlength@\lstinline{minlength}|)}
or even act on all the loops with the same braid:
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!action on \texttt{loop}}%{\lstinline{loop}\ (\lstinline{*})|(}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!multiplication (\lstinline{*})|(}
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = braid([1 -2]);
>> b*ll
ans = (( 2 1 -2 1 ))
(( 5 -2 -3 11 ))
\end{lstlisting}
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!multiplication (\lstinline{*})|)}
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!action on \texttt{loop}}%{\lstinline{loop}\ (\lstinline{*})|)}%
Some commands, such as \lstinline{plot}, do not vectorize. Different loops
can then be accessed by indexing, such as~\lstinline{plot(ll(2))}.
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!vectorized|)}
The \lstinline{entropy} method %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!entropy@\lstinline{entropy}|(}%
\index{braid!entropy|(}%
of the \lstinline{braid} class
(Section~\ref{sec:braidclass}) computes the topological entropy of a braid by
repeatedly acting on a loop, and monitoring the growth rate of the loop. For
example, let us compare the entropy obtained by acting~$100$ times on an
initial loop, compared with the \lstinline{entropy} method:
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = braid([1 2 3 -4]);
>> log(minlength(b^100*l)/minlength(l)) / 100
ans = 0.7637
>> entropy(b)
ans = 0.7672
\end{lstlisting}
The entropy value returned by \lstinline{entropy(b)} is more precise,
since that method monitors convergence and adjusts the number of
iterations accordingly. %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!entropy@\lstinline{entropy}|)}%
\index{braid!entropy|)}%
\subsection{Loop coordinates for a braid}
\label{sec:loopcoords}
\index{braid!loop coordinates|(}
\index{loop!coordinates|(}
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!constructor|(}
The command \lstinline{loop(n,'BasePoint')} returns a \emph{canonical set of
loops} for~$n$ punctures:%
\index{punctures|(}%
\begin{lstlisting}[frame=single,framerule=0pt]
>> l = loop(5,'BP')
ans = (( 0 0 0 0 -1 -1 -1 -1 ))*
\end{lstlisting}
\index{loop!multi-|(} This multiloop is depicted in
Figure~\ref{fig:fundloops}, with basepoint puncture shown in green. The
\lstinline{*} indicates that this loop has a basepoint. Note that the
multiloop returned by \hbox{\lstinline{loop(5,'BP')}} actually has 6
punctures! The rightmost puncture is meant to represent the boundary of a
disk, %
\index{disk, punctured}%
or a base point for the fundamental group on a sphere with $n$ punctures. The
loops form a (nonoriented) generating set for the fundamental group of the
disk with $n$ punctures. The extra puncture thus plays no role dynamically,
and \lstinline{l.n} returns~5. If you want the true total number of
punctures, including the base point, use \lstinline{l.totaln}. %
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!totaln@\lstinline{totaln}}
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class!constructor|)}
\index{punctures|)}%
\index{loop!multi-|)}
The canonical set of loops allows us to define loop coordinates for a braid,
which is a unique normal form.
\begin{figure}
\begin{center}
\subfigure[]{
\includegraphics[width=.7\textwidth]{fundloops}
\label{fig:fundloops}
}\hspace{1em}
\subfigure[]{
\includegraphics[width=\textwidth]{fundloops_act}
\label{fig:fundloops_act}
}
\end{center}
\caption{(a) The multiloop created by \lstinline{loop(5,'BP')}, with basepoint
puncture in green. (b) The multiloop \lstinline{b*loop(5,'BP')}, where
\lstinline{b} is the braid
$\sigma_1\sigma_2\sigma_3\sigma_4^{-1}$. \index{loop!multi-}}
\end{figure}
The canonical loop coordinates for braids exploit the fact that two braids are
equal if and only if they act the same way on the fundamental group of the
disk \citep{Dehornoy2008}. Hence, if we take a braid and act on
\lstinline{loop(5,'BP')},
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = braid([1 2 3 -4]);
>> b*loop(5,'BP')
ans = (( 0 0 3 -1 -1 -1 -4 3 ))*
\end{lstlisting}
then the set of numbers \lstinline{(( 0 0 3 -1 -1 -1 -4 3 ))*} can be thought
of as \emph{uniquely} characterizing the braid. It is this property that is
used to rapidly determine equality of braids. (The loop
\lstinline{b*loop(5,'BP')} is plotted in Figure~\ref{fig:fundloops_act}.) The
same loop coordinates for the braid can be obtained without creating an
intermediate loop with %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!loopcoords@\lstinline{loopcoords}}
\begin{lstlisting}[frame=single,framerule=0pt]
>> loopcoords(b)
ans = (( 0 0 3 -1 -1 -1 -4 3 ))*
\end{lstlisting}
\index{loop class@\texttt{loop}}%{\lstinline{loop}\ class|)}
\index{loop!coordinates|)}
\index{braid!loop coordinates|)}
\jlt{Next section: Braid from random walks? Compute runs of same gen.}
\section{The effective linear action and its cycles}
\label{sec:elacycles}
\subsection{Effective linear action}
\label{sec:linact}
\index{effective linear action|(}%
\index{linear action|see{effective linear action}}%
\index{action!effective linear|(}%
\index{action!of braid on loop|(}%
In Section~\ref{sec:actingonloops} we introduced the action of a
braid~$\gamma$ on a loop~$\abv$.
Here~$\abv=(\ac_1,\ldots,\ac_{\nn-2},\bc_1,\ldots,\bc_{\nn-2})$ is a vector of
coordinates for the loop, %
\index{loop!coordinates}%
defined in Section~\ref{sec:loopc}. We write~$\abv' = \gamma\cdot\abv$ for
the new, updated coordinates after the action. These updated coordinates are
given by composing the action of individual generators.
\index{update rules|see{action}}
\index{action!update rules|(}
For~$1 < \ip < \nn-1$, we can express the update rules for the braid group
generator~$\sigma_\ip$ acting on~$\abv$ as
\begin{subequations}
\begin{align}
\acnew_{\ip-1} &= \ac_{\ip-1} - \pos{\bc_{\ip-1}}
- \pos{\l(\pos{\bc_\ip} + \cc_{\ip-1}\r)}\,,\\
\bcnew_{\ip-1} &= \bc_\ip + \neg{\cc_{\ip-1}}\,,\\
\acnew_\ip &= \ac_\ip - \neg{\bc_\ip}
- \neg{\l(\neg{\bc_{\ip-1}} - \cc_{\ip-1}\r)}\,,\\
\bcnew_\ip &= \bc_{\ip-1} - \neg{\cc_{\ip-1}}\,,
\end{align}
\label{eq:ur}%
\end{subequations}
where
\begin{equation}
\cc_{\ip-1} = \ac_{\ip-1} - \ac_\ip - \pos{\bc_\ip} + \neg{\bc_{\ip-1}}\,.
\label{eq:ccdef}
\end{equation}
Coordinates not listed (i.e., $\ac_k$ and~$\bc_k$ for~$k\ne\ip$ or~$\ip-1$)
are unchanged. The superscripts~${}^{+/-}$ are defined as
\begin{equation}
\pos\fc \ldef \max(\fc,0),\qquad
\neg\fc \ldef \min(\fc,0).
\end{equation}
(See~\citet{Thiffeault2010} for the update rules for the
generators~$\sigma_1$, $\sigma_{\nn-1}$, and the inverse generators. The
update rules are in several other papers but use different conventions.)
\index{action!update rules|)}
Notice that the action~\eqref{eq:ur} is \emph{piecewise-linear} in the loop
coordinates: once the~${}^{+/-}$ operators are resolved, what is left is a
linear operation on the vector~$\abv$. We can thus write
\begin{equation}
\abv' = M(\gamma,\abv)\cdot\abv,
\qquad
M(\gamma,\abv) \in \mathrm{SL}_{2\nn-4}(\mathbb{Z}),
\end{equation}
where the dot now denotes the standard matrix product. Here~$M(\gamma,\abv)$
is the \emph{effective linear action} of the braid~$\gamma$ on the
loop~$\abv$.
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!action on \texttt{loop}}%{\lstinline{loop}\ (\lstinline{*})|(}%
Let's show an example using \texttt{braidlab}}%{\lstinline{braidlab}. We take the
braid~$\sigma_1\sigma_2^{-1}$ and the loop with coordinates~$\ac_1=0$,
$\bc_1=-1$. The action is
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = braid([1 -2]); l = loop([0 -1]);
>> lp = b*l
lp = (( 1 -1 ))
\end{lstlisting}
The effective linear action can be obtained by requesting a second output
argument from the result of~\lstinline{*}:
\begin{lstlisting}[frame=single,framerule=0pt]
>> [lp,M] = b*l; full(M)
ans = 1 -1
0 1
\end{lstlisting}
Note that the effective linear action \lstinline{M} is by default returned as
a sparse matrix, %
\index{sparse matrix}%
which it often is when dealing with many strands. We use \lstinline{full} %
\index{full@\lstinline{full}}%
to convert it back into a regular full matrix. We can then verify that the
matrix product of \lstinline{M} and the column vector of coordinates
\lstinline{l.coords'} is the same as the action \lstinline{lp = b*l}:
\begin{lstlisting}[frame=single,framerule=0pt]
>> M*l.coords'
ans = 1
-1
>> lp.coords'
ans = 1
-1
\end{lstlisting}
The difference is that \lstinline{M} may only be applied \emph{to this
specific loop} (or a loop that happens to share the same effective linear
action).
A common thing to do is to find the effective linear action on the canonical
set \lstinline{loop(b.n,'BP')} (see Section~\ref{sec:loopcoords}):
\begin{lstlisting}[frame=single,framerule=0pt]
>> [~,M] = b*loop(b.n,'BP'); full(M)
ans = 0 0 -1 0
0 1 0 1
0 1 1 1
1 -1 -1 0
\end{lstlisting}
The canonical set assumes an extra puncture, so the matrix dimension is larger
by~$2$.
The effective linear action doesn't seem to offer much at this point. Its
real advantage will become apparent in Section~\ref{sec:cycles}, when we find
that it can achieve periodic limit cycles.
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!action on \texttt{loop}}%{\lstinline{loop}\ (\lstinline{*})|)}%
\index{action!of braid on loop|)}%
\subsection{Limit cycles of the effective linear action}
\label{sec:cycles}
\index{cycle!of effective linear action|(}
\index{effective linear action!cycle|(}
The effective linear action has a very interesting behavior when a braid is
iterated on some initial loop. Consider the following example:
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = braid([1 -2]); l = loop([1 1]);
>> [l,M] = b*l; l, full(M)
l = (( 3 -1 ))
M = 2 1
-1 0
\end{lstlisting}
Now repeat this last command:
\begin{lstlisting}[frame=single,framerule=0pt]
>> [l,M] = b*l; l, full(M)
l = (( 7 -4 ))
M = 2 -1
-1 1
\end{lstlisting}
And again:
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = braid([1 -2]); l = loop([1 1]);
>> [l,M] = b*l; l, full(M)
l = (( 18 -11 ))
M = 2 -1
-1 1
\end{lstlisting}
The effective linear action \lstinline{M} has not changed. In fact it has
achieved a fixed point: %
\index{fixed point|see{cycle}}%
running the same command again will change the loop, but the linear action
will remain the same forever. \texttt{braidlab}}%{\lstinline{braidlab}\ can automate the iteration with the
method \lstinline{cycle}. %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!cycle@\lstinline{cycle}|(}%
Figure~\ref{fig:efflinact1} shows the output of
\begin{figure}
\begin{center}
\subfigure[]{
\includegraphics[height=.35\textheight]{efflinact1}
\label{fig:efflinact1}
}\hspace{.5em}
\subfigure[]{
\includegraphics[height=.35\textheight]{efflinact2}
\label{fig:efflinact2}
}\hspace{.5em}
\subfigure[]{
\includegraphics[height=.35\textheight]{efflinact3}
\label{fig:efflinact3}
}
\end{center}
\caption{The plot produced by \lstinline{cycle(b,'Plot')} %
for (a) \lstinline{b = braid([1 -2])}; (b) \lstinline{b = braid([1 2 3])};
(c) \lstinline{b = braid('Psi',11)}.}
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!cycle@\lstinline{cycle}}%
\label{fig:efflinact}
\end{figure}
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = braid([1 -2]); M = cycle(b,'Plot');
\end{lstlisting}
The member function \lstinline{cycle} iterates the braid on an initial loop,
taken to be the canonical set \lstinline{loop(b.n,'BP')}. The vertical axis
in Fig.~\ref{fig:efflinact1} shows the elements of the effective linear action
as a function of iterates of the braid. The matrix of the action is flattened
into a vector of length~$4^2$, where~$4$ is the dimension the initial loop
\lstinline{loop(b.n,'BP')}. It is evident that the fixed point is reached
rapidly, since the `stripes' stop changing.
\index{braid!pseudo-Anosov|(}%
Such fixed points of the effective linear action are ubiquitous for braids
corresponding to a pseudo-Anosov isotopy class, such
as~$\sigma_1\sigma_2^{-1}$. %
In general, instead of a fixed point we may find a \emph{limit cycle} of some
period. \citet{Yurttas2014_preprint} discussed these limit cycles for
pseudo-Anosov braids: they occur when the unstable foliation falls on the
boundary of the linear regions of the update rules. We can reproduce her
example with the following:%
\footnote{To get exactly the same matrices, we use the
braid~$\sigma_1^{-1}\sigma_2^{-1}\sigma_3^{-1}\sigma_4$ rather than
her~$\sigma_1\sigma_2\sigma_3\sigma_4^{-1}$, since her generators rotate the
punctures counterclockwise.}
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = braid([-1 -2 -3 4]);
>> M = b.cycle(loop(b.n),'Iter')
M = [6x6 double] [6x6 double]
\end{lstlisting}
The option \lstinline{'Iter'} tells \lstinline{cycle} to compute an individual
matrix for each iterate of the cycle, rather than the net product of all the
matrices in the cycle. The output is a cell array %
\index{Matlab!cell array}%
of two~$6$ by~$6$ matrices, corresponding to the period-2 cycle:
\begin{lstlisting}[frame=single,framerule=0pt]
>> full(M{1}), full(M{2})
ans = -1 1 0 0 0 0
0 0 0 1 1 0
0 0 2 -1 -1 1
0 0 0 0 1 0
-1 0 1 -1 -1 1
0 0 1 0 0 1
ans = 0 0 0 1 0 0
0 0 0 1 1 0
0 0 2 -1 -1 1
-1 1 0 -1 1 0
0 -1 1 0 -1 1
0 0 1 0 0 1
\end{lstlisting}
as given by \citet{Yurttas2014_preprint}. Note that we use an initial loop
for~$\nn$ punctures (\lstinline{loop(b.n)}) without base point, rather than
the default, to reproduce her example exactly. For the pseudo-Anosov case,
any initial loop will give the same matrices.
What is more surprising is that these limit cycles occur for finite-order
braids as well. Figure~\ref{fig:efflinact2} is produced by
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = braid([1 2 3]); [~,period] = cycle(b,'Plot')
period = 4
\end{lstlisting}
Indeed, staring at the pattern in Fig.~\ref{fig:efflinact2} it is easy to see
that the effective action does achieve a limit cycle of period~$4$. This
braid is definitely not pseudo-Anosov: it is finite-order.
\index{braid!finite-order}%
However, we do not expect such limit cycles to be unique in the
non-pseudo-Anosov case.
Pseudo-Anosov braids %
\index{braid!pseudo-Anosov}%
can achieve longer cycles, which \texttt{braidlab}}%{\lstinline{braidlab}\ can find:
Figure~\ref{fig:efflinact3} is the plot produced by
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = braid('Psi',11); [M,period] = cycle(b,'Plot');
\end{lstlisting}
The period here is~$5$, and the matrix~$M$ is~$20$ by~$20$. The braid %
\index{psi braids@$\psi$ braids}%
\index{braid!entropy!minimum}%
\lstinline{braid('Psi',11)} is the braid~$\psi_{11}$ in the notation of
\citet{Venzke_thesis}. It is a pseudo-Anosov braid with low %
\index{braid!pseudo-Anosov}%
\index{dilatation}%
dilatation~\citep{Hironaka2006,Thiffeault2006}, conjectured to be the lowest
possible for~$11$ strings.%
\footnote{The dilatation of a braid is the exponential of its entropy.}
The braids~$\psi_\nn$ are known to have to lowest dilatation for~$\nn$ string
for $\nn\le8$~\citep{LanneauThiffeault2011_braids}.
The largest eigenvalue of the matrix \lstinline{M} gives us the
\index{dilatation}%
dilatation of the braid, which in itself is not a real improvement over our
earlier entropy iterative algorithm (Section~\ref{sec:entropy}). However,
with the matrix in hand we can find the characteristic polynomial:%
\footnote{Matlab's symbolic toolbox %
\index{Matlab!symbolic toolbox}%
is needed for \lstinline{poly2sym} and \lstinline{factor}.}
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = braid('Psi',7); [M,period] = cycle(b);
>> factor(poly2sym(charpoly(M)))
ans = (x^2 + 1)*(x^3 - x^2 - 1)*(x^3 + x - 1)*(x - 1)^2*(x + 1)^2
\end{lstlisting}
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!cycle@\lstinline{cycle}|)}%
Compare this to the known polynomial that gives the dilation:
\index{psiroots@\lstinline{psiroots}}
\begin{lstlisting}[frame=single,framerule=0pt]
>> factor(poly2sym(psiroots(7,'Poly')))
ans = (x + 1)*(x^3 - x^2 - 1)*(x^3 + x - 1)
\end{lstlisting}
(The function \lstinline{psiroots} returns the roots and characteristic
polynomial of a~$\psi$ braid; this is useful for testing purposes.) Note that
the factor whose largest root is the dilatation, %
\index{dilatation}%
\lstinline{x^3 - x^2 - 1}, appears in both polynomials. This is not always
the case, though the dilatation has to be a root of both polynomials.
\index{braid!pseudo-Anosov|)}%
To our knowledge, the existence of these limit cycles has not been fully
explained (except in the pseudo-Anosov case by \citet{Yurttas2014_preprint}).
They seem to occur for \emph{any} braid, regardless of its %
\index{homeomorphism!isotopy classes}%
isotopy class. In that sense they could provide an alternative to the the
Bestvina--Handel train track algorithm~\citep{Bestvina1995}, %
\index{Bestvina--Handel algorithm}%
which is used to compute the isotopy class of a braid.
\index{cycle!of effective linear action|)}%
\index{effective linear action|)}%
\index{effective linear action!cycle|)}
\index{action!effective linear|)}%
\index{limit cycle|see{cycle}}
\section{An example: Taffy pullers}
\label{sec:taffy}
\index{taffy pullers|(}
Taffy pullers are a class of devices designed to stretch and fold soft candy
repeatedly \citep{MattFinn2011_silver}. The goal is to aerate the taffy.
Since many folds are required, the process has been mechanized using fixed and
moving rods. The two most typical designs are shown in
Figure~\ref{fig:taffy}: the one in
\begin{figure}
\begin{center}
\subfigure[]{
\includegraphics[height=.2\textheight]{taffy_3rods}
\label{fig:taffy_3rods}
}\hspace{1em}
\subfigure[]{
\includegraphics[height=.2\textheight]{taffy_4rods}
\label{fig:taffy_4rods}
}
\end{center}
\caption{(a) Three-rod taffy puller. (b) Four-rod taffy puller.}
\label{fig:taffy}
\end{figure}
Figure~\ref{fig:taffy_3rods} has a single fixed rod (gray) and two moving
rods, each rotating on a different axis. The design in
Figure~\ref{fig:taffy_3rods} has four moving rods, sharing two axes of
rotation. (There are several videos of taffy pullers on
\href{http://www.youtube.com/watch?v=6QkGp2qBbn4}{YouTube}.)
Let's use \texttt{braidlab}}%{\lstinline{braidlab}\ to analyze the rod motion. From the folder
\lstinline{doc/examples}, run the command%
\footnote{When using the parallel code, the generator sequences in this
section may sometimes differ from run-to-run, due to the simultaneous
crossings. However, the braids themselves are still equal, after the
relations~\eqref{eq:relations} are taken into account.}
\index{taffy@\lstinline{taffy}|(}%
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = taffy('3rods')
b = < -2 1 1 -2 >
\end{lstlisting}
which also produces Figure~\ref{fig:taffy_3rods}. The Thurston--Nielsen %
\index{braid!Thurston--Nielsen type}%
type and topological entropy %
\index{braid!entropy}%
of this braid are %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!tntype@\lstinline{tntype}|(}%
\begin{lstlisting}[frame=single,framerule=0pt]
>> [t,entr] = tntype(b)
t = pseudo-Anosov
entr = 1.7627
\end{lstlisting}
One would expect a competent taffy puller to be pseudo-Anosov, %
\index{braid!pseudo-Anosov}%
as this one is. It implies that there is no `bad' initial condition where a
piece of taffy never gets stretched, or stretches slowly. A reducible or
finite-order braid would indicate poor design. The entropy is a measure of
the taffy puller's effectiveness: it gives the rate of growth of curves
anchored on the rods. Thus, the length of the taffy is multiplied
(asymptotically) by $\ee^{1.7627} \simeq 5.828$ for each full period of rod
motion. Needless to say, this leads to extremely rapid growth, since after
$10$ periods the taffy length has been multiplied by roughly $10^7$.
The design in Figure~\ref{fig:taffy_4rods} can be plotted and analyzed with
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = taffy('4rods')
b = < 1 3 2 2 1 3 >
\end{lstlisting}
When we apply \lstinline{tntype} to this braid we find the braid is
pseudo-Anosov with exactly the same entropy as the 3-rod taffy puller,
$1.7627$. There is thus no obvious advantage to using more rods in this case.
A simple modification of the 4-rod design in Figure~\ref{fig:taffy_4rods} is
shown in Figure~\ref{fig:taffy_6rods-bad}.
\begin{figure}
\begin{center}
\subfigure[]{
\includegraphics[height=.2\textheight]{taffy_6rods-bad}
\label{fig:taffy_6rods-bad}
}\hspace{1em}
\subfigure[]{
\includegraphics[height=.2\textheight]{taffy_6rods}
\label{fig:taffy_6rods}
}
\end{center}
\caption{(a) A six-rod taffy puller based on Figure~\ref{fig:taffy_4rods},
with two added fixed rods (gray). This is a poor design, since it leads to
a reducible braid. %
\index{braid!reducible}%
(b) Same as (a), but with the same radius of motion for all the moving rods.
The braid is in this case pseudo-Anosov, with larger entropy than the 4-rod
design.}
\label{fig:taffy_6rods-improved}
\end{figure}
The only change is to extend the rotation axles into two extra fixed rods
(shown in gray). The resulting braid is
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = taffy('6rods-bad')
b = < 2 1 2 4 5 4 3 3 2 1 2 4 5 4 >
\end{lstlisting}
with Thurston--Nielsen type
\index{braid!reducible|(}%
\begin{lstlisting}[frame=single,framerule=0pt]
>> tntype(b)
ans = reducible
\end{lstlisting}
There are reducing curves in this design: simply wrap a loop around the left
gray rod and the inner red rod, and it will rotate without stretching. %
\index{braid!reducible|)}%
To avoid this, we extend the radius of motion of the inner rods to equal that
of the outer ones, and obtain the design shown in
Figure~\ref{fig:taffy_4rods}. The corresponding braid is
\begin{lstlisting}[frame=single,framerule=0pt]
>> b = taffy('6rods')
b = < 3 2 1 2 4 5 4 3 3 2 1 2 5 4 5 3 >
\end{lstlisting}
with Thurston--Nielsen type and entropy
\begin{lstlisting}[frame=single,framerule=0pt]
>> [t,entr] = tntype(b)
t = pseudo-Anosov
entr = 2.6339
\end{lstlisting}
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!tntype@\lstinline{tntype}|)}%
The fixed rods have increased the entropy by~$50\%$! This sounds like a
fairly small change, but what it means is that this 6-rod design achieves
growth of~$10^7$ in about~$6$ iterations rather than~$10$. Alexander Flanagan
constructed this six-rod device while an undergraduate student at the
University of Wisconsin -- Madison, but as far as we know this new design has
not yet been used in commercial applications.
The symmetric design of the taffy pullers illustrates one pitfall when
constructing braids. If we give an optional projection angle %
\index{projection line!bad choice of angle|(}%
of~$\pi/2$ to \lstinline{taffy}:
\begin{lstlisting}[frame=single,framerule=0pt]
>> taffy('4rods',pi/2)
Error using braidlab.braid/colorbraiding
Paths of particles 2 and 1 have a coincident projection.
Try changing the projection angle.
\end{lstlisting}
This corresponds to using the $y$ (vertical) axis to compute the braid, but as
we can see from Figure~\ref{fig:taffy_4rods} this is a bad choice, since all
the rods are initially perfectly aligned along that axis. The braid obtained
would depend sensitively on numerical roundoff when comparing the rod
projections.%
\footnote{\texttt{braidlab}}%{\lstinline{braidlab}\ uses a property \lstinline{BraidAbsTol} to determine when
coordinates are close enough to be considered coincident, with a default
value of~$10^{-10}$. See Section~\ref{sec:prop} for how to set global
properties.}
Instead of attempting to construct the braid, \texttt{braidlab}}%{\lstinline{braidlab}\ returns an error and
asks the user to modify the projection axis. A tiny change in the projection
line is sufficient to break the symmetry:
\begin{lstlisting}[frame=single,framerule=0pt]
>> taffy('4rods',pi/2 + .01)
ans = < -2 2 1 3 2 -3 -1 3 1 2 1 3 >
>> compact(ans)
ans = < 3 1 2 2 3 1 >
\end{lstlisting}
\index{taffy@\lstinline{taffy}|)}%
which is actually equal to the braid formed from projecting on the~$x$ axis,
though it need only be conjugate %
\index{braid!conjugate}%
(see Section~\ref{sec:braidfromdata}). %
\index{projection line!bad choice of angle|)}%
\index{taffy pullers|)}
\section{Side note: On filling-in punctures}
\index{punctures!filling-in|(}
Recall the command~\lstinline{subbraid}
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!subbraid class@\lstinline{subbraid}}%
from Section~\ref{sec:braidclass}. We
took the~$4$-string braid~$\sigma_1\sigma_2\sigma_3^{-1}$ and discarded the
third string, to obtain~$\sigma_1\sigma_2^{-1}$:
\begin{lstlisting}[frame=single,framerule=0pt]
>> a = braid([1 2 -3]);
>> b = subbraid(a,[1 2 4])
b = < 1 -2 >
\end{lstlisting}
\begin{figure}
\begin{center}
\subfigure[]{
\includegraphics[width=.22\textwidth]{s1s2s-3_diagram}
\label{fig:s1s2s-3_diagram}
}\hspace{5em}
\subfigure[]{
\includegraphics[width=.22\textwidth]{s1s-2_diagram}
\label{fig:s1s-2_diagram}
}
\end{center}
\caption{Removing the third string from the braid
(a)~$\sigma_1\sigma_2\sigma_3^{-1}$ yields the braid
(b)~$\sigma_1\sigma_2^{-1}$.}
\label{fig:subbraid}
\end{figure}
The braids \lstinline{a} and \lstinline{b} are shown in
Fig.~\ref{fig:subbraid}; their entropies are %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!entropy@\lstinline{entropy}|(}%
\index{braid!entropy|(}%
\begin{lstlisting}[frame=single,framerule=0pt]
>> a.entropy, b.entropy
ans = 0.8314
ans = 0.9624
\end{lstlisting}
Note that the entropy of the subbraid~\lstinline{b} is \emph{higher} than the
original braid. This is counter-intuitive: shouldn't removing strings cause
loops to shorten, therefore lowering their growth?\footnote{In fact, the
entropy obtained by the removal of a string is constrained by the minimum
possible entropy %
\index{braid!entropy!minimum}%
for the remaining number of strings
\citep{Song2002,Hironaka2006,Thiffeault2006,
Ham2007,Venzke_thesis,LanneauThiffeault2011_braids}. So here the entropy
of the 3-braid could only be zero or $\ge 0.9624$.}
In some sense this must be true: consider the rod-stirring device shown in
Fig.~\ref{fig:s1s2s-3_no_text}, where the rods move according the to
braid~$\sigma_1\sigma_2\sigma_3^{-1}$.
\begin{figure}
\begin{center}
\subfigure[]{
\includegraphics[height=.3\textheight]{s1s2s-3_no_text}
\label{fig:s1s2s-3_no_text}
}\hspace{2em}
\subfigure[]{
\includegraphics[height=.3\textheight]{s1s2s-3_4_diagram}
\label{fig:s1s2s-3_4_diagram}
}\hspace{2em}
\subfigure[]{
\includegraphics[height=.3\textheight]{s1s-2s1s-2s1s2_diagram}
\label{fig:s1s-2s1s-2s1s2_diagram}
}
\end{center}
\caption{(a) The mixing protocol specified by the
braid~$\sigma_1\sigma_2\sigma_3^{-1}$ \citep{Thiffeault2008b}. The inset
shows how the rods are moved. (b)~The pure \index{braid!pure}
braid~$(\sigma_1\sigma_2\sigma_3^{-1})^4$. (c)~The
braid~$(\sigma_1\sigma_2^{-1})^2\sigma_1\sigma_2$, obtained by removing the
third string from~(b).}
\end{figure}
Removing the third string can be regarded as \emph{filling-in} the third
puncture (rod); clearly then the material line can be shortened, leading to a
decrease in entropy.
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!perm@\lstinline{perm}|(}%
The flaw in the argument is that even though we can remove any string, we
cannot fill in a puncture that is permuted, since the resulting braid does not
define a homeomorphism on the filled-in surface. To remedy this, let us take
enough powers of the braid~$\sigma_1\sigma_2\sigma_3^{-1}$ to ensure that the
third puncture returns to its original position, using the method
\lstinline{perm} to find the permutation induced by the braid:
\begin{lstlisting}[frame=single,framerule=0pt]
>> perm(a)
ans = 2 3 4 1
\end{lstlisting}
The permutation is cyclic (it can be constructed with exactly one cycle), so
the fourth power should do it:
\begin{lstlisting}[frame=single,framerule=0pt]
>> perm(a^4)
ans = 1 2 3 4
\end{lstlisting}
This is now a pure braid: all the strings return to their original position
(Fig.~\ref{fig:s1s2s-3_4_diagram}). Now here's the surprise: the subbraid
obtained by removing the third string from \lstinline{a^4} is
\begin{lstlisting}[frame=single,framerule=0pt]
>> b2 = subbraid(a^4,[1 2 4])
b2 = < 1 -2 1 -2 1 2 >
\end{lstlisting}
which is \emph{not} \lstinline{b^4} (Fig.~\ref{fig:s1s-2s1s-2s1s2_diagram})!
However, now there is no paradox in the entropies:\footnote{\citet{Song2005}
showed that the entropy of a pure braid \index{braid!pure} is greater
than~$\log(2+\sqrt5) \simeq 1.4436$, if it is nonzero.}
\index{braid!finite-order|(}%
\begin{lstlisting}[frame=single,framerule=0pt]
>> entropy(a^4), entropy(b2)
ans = 3.3258
Warning: Failed to converge to requested tolerance; braid is likely finite-order or has low entropy. Returning zero entropy.
ans = 0
\end{lstlisting}
\texttt{braidlab}}%{\lstinline{braidlab}\ has trouble computing the entropy because the braid \lstinline{b2}
appears to be finite-order. Indeed, the braid \lstinline{b2} is conjugate
to~$\sigma_1^2$: %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!compact@\lstinline{compact}}%
\begin{lstlisting}[frame=single,framerule=0pt]
>> c = braid([2 -1],3);
>> compact(c*b2*c^-1)
ans = < 1 1 >
\end{lstlisting}
showing that its entropy is indeed zero.
\index{braid!finite-order|)}%
The moral is: when filling-in punctures, make sure that the strings being
removed are permuted only among themselves. For very long, random braids, we
still expect that removing a string will decrease the entropy, since the
string being removed will have returned to its initial position many times. %
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!entropy@\lstinline{entropy}|)}%
\index{braid!entropy|)}%
\index{braid class@\texttt{braid}}%{\lstinline{braid}\ class!perm@\lstinline{perm}|)}%
\index{punctures!filling-in|)}
\section{Setting global properties}
\label{sec:prop}
Braids have been studied for a long time and by many communities, so several
different conventions were bound to emerge. \texttt{braidlab}}%{\lstinline{braidlab}\ has some reasonable
default conventions, but perhaps you'd be happier if it used your favorite
one. Luckily, \texttt{braidlab}}%{\lstinline{braidlab}\ has you covered.
\index{prop@\lstinline{prop}|(}%
To see the properties available and their current values, use the
\lstinline{prop} command:
\begin{lstlisting}[frame=single,framerule=0pt]
>> prop
ans = GenRotDir: 1
GenLoopActDir: 'lr'
GenPlotOverUnder: 1
BraidAbsTol: 1.0000e-10
BraidPlotDir: 'bt'
LoopCoordsBasePoint: 'right'
\end{lstlisting}
To set a property, use something like \lstinline{prop('BraidPlotDir','lr')}.
This will plot braids from left-to-right from now on, as in
Fig.~\ref{fig:fulltwist_lr}.
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{fulltwist_lr}
\end{center}
\caption{\lstinline{plot(braid('FullTwist',5))} with the property
\lstinline{'BraidPlotDir'} set to \hbox{\lstinline{'lr'}}.}
\label{fig:fulltwist_lr}
\end{figure}
The default is \lstinline{'bt'} for bottom-to-top. Other possible values are
\lstinline{'tb'} for top-to-bottom and \lstinline{'rl'} for right-to-left.
See \lstinline{help prop} for more information on properties and allowed
values.
\index{prop@\lstinline{prop}|)}
\section*{Acknowledgments}
\addcontentsline{toc}{section}{Acknowledgments}
The development of \texttt{braidlab}}%{\lstinline{braidlab}\ was supported by the US National Science
Foundation, under grants DMS-0806821 and CMMI-1233935. The authors thank
Michael Allshouse and Margaux Filippi for extensive testing and comments.
Michael Allshouse also contributed some of the code. Alexander Flanagan
helped with testing and the taffy puller research. James Puckett and Karen
Daniels provided the test data from their granular medium
experiments~\citep{Puckett2012}; \texttt{braidlab}}%{\lstinline{braidlab}\ uses Toby Hall's \emph{Train}
\citep{HallTrain}; Jae Choon Cha's \emph{CBraid} \citep{CBraid}; Juan
Gonz\'{a}lez-Meneses's \emph{Braiding} \citep{Braiding}; John D'Errico's
\emph{Variable Precision Integer Arithmetic} \citep{vpi}; Markus Buehren's
\emph{assignmentoptimal} \citep{assignmentoptimal}; Jakob Progsch's
\emph{ThreadPool} \citep{ThreadPool}; and John R.\ Gilbert's function for
computing the Smith Normal Form of a matrix \citep{snf}.
|
\section{Introduction}
Roughly two-thirds of disk galaxies have bars, of which half have strong
bars today \citep{eskridge00, barazza08}.
The bar fraction seems to be
a strong function of redshift; it decreases by a factor of three
from $z = 0$ to $z = 0.8$ \citep{sheth08}.
Bars can affect galaxy properties
by driving secular evolution
(e.g. \cite{kormendy79, combes81, pfenniger90, combes93,
debattista04, kk04, athanassoula05, athanassoula13, kormendy13}),
and thus a number of simulations have
performed to study their formation and evolution processes.
Most of the simulations employ idealized initial conditions for isolated
galaxies in order to answer specific questions (e.g.
\cite{combes90, debattista00, athanassoula03, martinez-valpuesta06, amr13}).
A general picture obtained by these studies is that a bar continuously grows
and slows down with time as its angular momentum gets transferred to the
dark halo and the outer disk
(e.g. \cite{weinberg85, combes93, athanassoula03}).
In reality however disks and dark halos keep growing while bars
form and evolve.
The dark halos resulting from cosmological simulations are
naturally triaxial \citep{frenk88, jing02} and abundant in substructure
\citep{moo99, kly99, aquarius},
both of which should exert torque on bars.
Galaxy interactions could trigger bar formation \citep{miwa98, berentzen04}.
Moreover, even disk orientation changes with time owing to the misalignment
of the angular momentum vector of the newly accreting gas
\citep{oka05, okamoto13}.
Early attempts to investigate the effects of cosmological growth of galaxies
on the bar evolution are made either by simplifying the initial density
perturbations \citep{heller07, romano-diaz08} or by embedding a live disk
in a growing dark halo \citep{curir06}.
\citet{kraljic12} use a cosmological $N$-body simulation as boundary
conditions for sticky particle simulations and follow galaxy and bar
evolution in evolving halos.
They confirm that the bar fraction in fact decreases with increasing
redshift.
\citet{scannapieco12} analyze the bars in fully self-consistent
simulations of galaxy formation at $z = 0$ and compare their properties
with those in the idealized simulations.
Only recently have cosmological simulations with sufficient resolution to
follow evolution of detailed structure in disk galaxies become possible
\citep{eris, okamoto13, magicc, marinacci14}.
An important difference of these simulations from the idealized simulations is
that they invoke much stronger stellar feedback than the idealized simulations
normally assume, otherwise simulations form
too many stars for a given halo mass (e.g. \cite{aquila, okamoto13, okamoto14}).
The feedback is so strong that it can lower the central stellar and dark matter
density during the galaxy evolution \citep{duffy10, governato12}, and hence
orbits of stars should be affected.
Two Milky Way-mass galaxies formed in the cosmological simulations by
\citet{okamoto13} offer a unique opportunity to investigate the evolution
of bars in the context of the $\Lambda$-cold dark matter ($\Lambda$CDM)
cosmology.
One galaxy is clearly barred from $z \simeq 1$ to $0$, while the other has a
much weaker bar.
In this paper we compare and contrast these two galaxies focusing on the
evolution of their bars.
We also compare their properties with those obtained by idealized simulations.
This paper is organized as follows. In section~2 we briefly describe the
simulations and galaxies we analyze. We present our results in section~3
We then discuss them and summarize our main conclusion in section~4.
\section{Sample galaxies}
We study the evolution of the bars formed in two smoothed particle hydrodynamic
(SPH) simulations of galaxy formation in a $\Lambda$CDM universe.
The cosmological parameters employed in these simulations are:
$\Omega_0 = 0.25$, $\Omega_\Lambda = 0.75$, $\Omega_\mathrm{b} = 0.045$,
$\sigma_8 = 0.9$, $n_\mathrm{s} = 0.9$, and a Hubble constant of
$H_0 = 100~h$~km~s$^{-1}$~Mpc$^{-1}$, where $h = 0.73$.
The galaxies are selected from a cosmological periodic box of a side
length of $100~h^{-1}$~Mpc and called `Aq-C' and `Aq-D' according to
the labeling system of the Aquarius project \citep{aquarius}.
The dark matter particle masses are $2.6 \times 10^5$ and
$2.2 \times 10^5$~M$_\odot$ for Aq-C and Aq-D, respectively, and
the original SPH particle masses are $5.8 \times 10^4$ and
$4.8 \times 10^4$~M$_\odot$, respectively.
At $z < 3$ the gravitational softening lengths are fixed in physical
coordinates as $\epsilon = 0.257$ and $0.240$~kpc respectively in
Aq-C and Aq-D. The simulations include radiative cooling, photo-heating
by the ultra-violet background \citep{ogt08}, star formation, timed release
of energy, mass, and metals by type II and Ia supernovae and AGB stars
\citep{ofjt10, okamoto13}.
It should be noted that the mass resolution of our simulations are as
good as that in the high resolution idealized simulations, but the
spatial resolution is still poor compared with that of the recent
idealized simulations ($50$~pc in \cite{amr13}).
The way of implementing feedback is a key to reproduce observed properties
of galaxies \citep{ofjt10, okamoto14}.
In our simulations, the energetic feedback from supernovae is modelled
as winds. A gas particle may receive an amount of energy, $\Delta E$, form
type II supernovae during a time-step, $\Delta t$.
We add this gas particle to winds with a probability,
$p_\mathrm{w} = \Delta E/[(1/2) m_\mathrm{gas} v_\mathrm{w}^2]$,
where $m_\mathrm{gas}$ is the mass
of the gas particle and $v_\mathrm{w}$ is the initial wind speed.
The initial wind speed, $v_\mathrm{w}$ is given as $v_\mathrm{w} = 5 \sigma$,
where $\sigma$ is the one-dimensional velocity dispersion of the dark matter
particles around the gas particle.
Doing this ensures that the wind mass generated by a type~II supernova
is proportional to $\sigma^{-2}$, and therefore less mass galaxies
blow more winds per unit star formation than there more massive
counterparts.
This feedback model explains luminosities and metallicities of galaxies
ranging form the Local Group satellites \citep{ofjt10} to more massive
galaxies \citep{okamoto14}.
The supernova feedback implemented this way is the strongest among the
simulations presented in \citet{aquila}, and the resulting stellar mass of
the galaxies is consistent with what expected by the abundance matching for
their halo masses. \citep{guo10, behroozi13, moster13}.
The simulations follow the evolution of the stellar populations throughout
their entire life. Therefore the star particles lose their masses not
only by the type II supernovae and mass loss from the massive stars but also
by type Ia supernovae and the stellar mass loss from the intermediate mass
stars. The feedback from old stellar populations by type Ia supernovae is
also considered.
The stellar masses of these galaxies at $z = 0$ are
$4.0 \times 10^{10}$~M${_\odot}$ (Aq-C) and
$3.1 \times 10^{10}$~M${_\odot}$ (Aq-D);
hence their masses are close to that of the Milky Way.
The global evolution and properties of these galaxies are fully
described in \citet{okamoto13}.
We here only show the evolution of the circular velocity profiles
which is the most relevant to the current study.
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig01.eps}
\end{center}
\caption{
Evolution of the circular velocity profiles. The upper and lower panels
show Aq-C and Aq-D, respectively.
From left to right, the circular velocities at $z = 1.5$, $1.0$, $0.5$,
and $0$ are plotted as functions of radius.
The blue solid lines are the circular velocities calculated from the
azimuthally averaged radial accelerations in the disk planes, and the
green dotted, red dashed, and cyan dot-dashed lines indicate
the contribution of the dark matter, baryons, and stellar components,
respectively.
The vertical dotted lines indicate the radii at which
$v_\mathrm{c}(r) = 1.1 v_\mathrm{c, b}(r)$,
where $v_\mathrm{c, b}(r)$ is the contribution of the baryonic matter to
the circular velocity at the radius, $r$.
}
\label{fig:vc}
\end{figure*}
In order to characterize the mass distribution and the angular frequency of
circular orbits, we calculate the circular velocity from the azimuthally
averaged radial acceleration in the disk plane, $a_r(r)$,
as $v_\mathrm{c}(r)^2 = - r a_r(r)$.
We measure the radial acceleration at equally spaced 128 azimuthal positions
at each radius in the disk plane to determine $a_r(r)$.
We show the evolution of the circular velocity measured this way in
figure~\ref{fig:vc}.
We find that the central circular velocity continuously
decreases with time in both galaxies. This is presumably due to the stellar
mass loss and feedback since we do not find this behavior in simulations
without stellar mass loss nor feedback.
This drop changes the resonance structure as we will show later.
We also find that Aq-C has more centrally concentrated mass distribution
than Aq-D.
The vertical line indicates the radius inside which a baryonic component
dominates as $v_\mathrm{c}(r) < 1.1 v_\mathrm{c, b}(r)$,
where $v_\mathrm{c, b}(<r)$ is the contribution of
the baryons (gas and stars) to the circular velocity;
the disks are bar unstable
when this condition is satisfied \citep{eln93}, although this criterion
is only appropriate for a pure $N$-body cold disk in a rigid halo
(e.g. \cite{athanassoula08}).
This radius is almost constant with redshift ($\simeq 1$~kpc) in both
galaxies.
Similarly, the radius at which the contribution of the baryons is equal
to that of the dark matter is also constant with redshift ($\simeq 4$~kpc),
indicating that the stellar mass fractions as functions of radius do not
evolve strongly with redshift.
Aq-C's circular velocity curve is more resembling to the universal
rotation curve proposed by \citet{salucci07} than Aq-D's.
\section{Results}
In this section, we first describe how we identify stellar bars in the
simulated galaxies and then undertake detailed studies of the evolution
of the bars.
In the following analyses, the $z$-direction is chosen to be parallel
to the angular momentum vector of stars within 5 per cent of the virial
radius\footnote{The virial radius is calculated based on the spherical
collapse model \citep{ecf96}.} at given redshift unless otherwise stated.
Note that the disk orientation significantly changes with redshift
\citep{oka05, okamoto13}.
We use stars with $|z| < 1~h^{-1}$~kpc to compute the surface stellar
density to avoid the contamination of the stars in the satellite galaxies.
\subsection{Bar identification}
The strength, length, and angle of a bar can be parameterized by the
amplitude and phase of its Fourier component, defined by expressing the
surface stellar density as a Fourier series,
\begin{equation}
\frac{\Sigma(r, \phi)}{\bar{\Sigma}(r)}
= 1 + \sum_{m = 1}^\infty A_m(r) \cos[m\{\phi - \phi_m(r)\}],
\end{equation}
where $\Sigma(r, \phi)$ is the surface stellar density, $\phi$ is the azimuthal
angle, $\bar{\Sigma}(r)$ is the azimuthally averaged surface stellar density at
radius $r$, and $A_m$ and $\phi_m$ are the amplitude and phase of the
$m$-th Fourier component, respectively.
Before we perform this analysis, we remove stellar clumps in order for the
$m = 2$ component not to include contributions of them, which are
particularly abundant in Aq-D \citep{okamoto13}.
The details of this procedure and its effect are described in
appendix~\ref{app:clump}. In short, the removal of the clumps does not affect
our results presented in this paper.
We calculate the Fourier series for the face-on projections.
Typically, the amplitude of the $m = 2$ mode, $A_2(r)$, has a peak if a
bar exists. The phase, $\phi_2(r)$, should be constant in the bar
region.
We first identify the radius at which $A_2(r)$ takes the maximum value,
$A_2^\mathrm{max}$, and we call this radius $r_2^\mathrm{max}$.
The bar angle is defined as
$\phi_\mathrm{bar} = \phi_2(r_2^\mathrm{max})$.
We utilize the radial profile of the phase of the $m = 2$ Fourier
component to determine the bar length.
Outside the bar region, the phase should show large variation
owing to a spiral structure or by the absence of clear structure.
We thus define the bar length, $r_\mathrm{bar}$, as the maximum radius
where the bar angle
$\phi_\mathrm{bar}$ and the phase of the $m = 2$ component differ by
less than $\Delta \phi$.
We employ $\Delta \phi = 0.1 \pi$ in this paper.
\citet{am02} present a number of ways to measure bar length;
each method has its own pros and cons.
\citet{scannapieco12} compare three methods among them, which are readily
applicable to disks obtained by cosmological simulations.
We also compare these three methods in appendix~\ref{app:bar_length};
the first method is what we have described above,
the second method uses the $m = 2$ amplitude profiles, and the
third method compares the surface stellar density profile along the bar
major axis with that along the bar minor axis.
We find that the bar lengths obtained by these three methods agree
reasonably well.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig02.eps}
\end{center}
\caption{The amplitude and phase profiles of the $m = 2$ mode.
The upper and lower six panels show Aq-C and Aq-D, respectively.
From left to right, the results at $z = 1$, 0.5, and 0 are presented.
The bar length, $r_\mathrm{bar}$, is indicated by the vertical red dashed
line in each panel.
The horizontal green dot-dashed lines indicate the values of
$A_2^\mathrm{max}$ and $\phi_\mathrm{bar}$ in the panels for amplitude and
phase, respectively.
}
\label{fig:m2}
\end{figure}
We demonstrate our method in figure~\ref{fig:m2} where we plot the
amplitude and phase profiles of the $m = 2$ Fourier component at $z = 1$, 0.5,
and 0.
We find that the amplitude, $A_2(r)$, has a clear peak
at each redshift and the phase, $\phi_2(r)$, is nicely constant around the peak.
We hence conclude that both $A_2^\mathrm{max}$ and $\phi_\mathrm{bar}$ are
robustly defined by this method.
On the other hand, the bar length, $r_\mathrm{bar}$, has weak dependence on
the value of $\Delta \phi$, and thus it is less robustly defined\footnote{Any methods described in appendix~\ref{app:bar_length} introduce a tolerance
parameter to define the bar length.}.
In Aq-D, the amplitude, $A_2(r)$, has two peaks within the bar length
at $z = 0$.
Such a feature is hardly seen either in real galaxies (e.g.
\cite{buta06, elmegreen07, buta09}) or in the idealized simulations
(e.g. \cite{am02}).
We therefore examine whether the outer peak is a different component from
the inner component, which is coincidentally aligned with the inner component
in appendix~\ref{app:outer}.
We find that the amplitude of the outer peak shows a rapid time variation
and the outer component always has almost the same phase as the inner
component.
By inspecting the surface density distribution, we speculate that
the outer $m = 2$ peak is induced by interactions with the debris of
the stellar clumps in the outer region of the bar,
which is orbiting with the different angular velocity from the bar.
We thus hold our definition of the bar length throughout this paper.
We will later show how the bar length changes if we exclude the outer component.
It should be noted that we always use the inner component to define the
bar angle and amplitude of Aq-D, and therefore the treatment of
the outer component does not affect most of the results and
discussions presented in this paper.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig03.eps}
\end{center}
\caption{
Contour maps of the face-on stellar surface density of the galaxies.
We have removed the clumps from the density distributions as described
in appendix~\ref{app:clump}.
The upper and lower panels correspond to Aq-C and Aq-D, respectively,
and the galaxies at $z = 1$, 0.5, and 0 are shown from left to right.
The contour levels are logarithmic.
The black solid lines indicate the bars whose length and angle are defined
as described in the text.
}
\label{fig:contour}
\end{figure}
The bars identified this way are overplotted on the face-on surface
stellar density maps in figure~\ref{fig:contour}.
The surface stellar density maps confirm that Aq-C always has a longer bar
than Aq-D, while Aq-D has a better defined bar at $z = 1$, which is evident
from figure~\ref{fig:m2} where the value of $A_2^\mathrm{max}$ of Aq-D
at $z = 1$ is larger than that of Aq-C.
We find that the central surface stellar density decreases with time
in both galaxies.
We also find that Aq-D has more structure than Aq-C such as spiral arms.
Such structure in Aq-D is mainly induced by the interactions with the clumps
(see appendix~\ref{app:clump}),
which we have erased in figure~\ref{fig:contour}.
The clumps sink to the central region \citep{okamoto13},
and it can change the bar properties by adding mass and changing
the angular momentum of the central region.
By comparing the bar lengths and angles indicated by the straight
lines in figure~\ref{fig:contour} with
the shapes of the isodensity contours, we conclude that our method
identifies the bars in the simulated galaxies well.
\subsection{Redshift evolution}
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig04.eps}
\end{center}
\caption{
Time evolution of the bar length, $r_\mathrm{bar}$, and the amplitude,
$A_2^\mathrm{max}$.
The upper panel shows the bar lengths as functions of time (or redshift).
The red solid and blue dotted lines indicate Aq-C and Aq-D, respectively.
The black dashed line represents the bar length of Aq-D when we exclude
the outer component.
The amplitude, $A_2^\mathrm{max}$, is shown in the lower panel.
}
\label{fig:zevo}
\end{figure}
We now investigate the redshift evolution of the bars.
We plot the bar lengths, $r_\mathrm{bar}$, and the bar amplitudes,
$A_2^\mathrm{max}$, as functions of cosmic time and redshift in
figure~\ref{fig:zevo}.
We find that Aq-C's bar becomes longer and stronger with time.
This behavior is usually seen for bars in early-type
disk galaxies \citep{combes93} or bars in galaxies with centrally
concentrated dark halos in the idealized
simulations \citep{athanassoula03}.
This is consistent with the fact that Aq-C has a massive bulge
\citep{okamoto13} and its halo shows the high central concentration
(figure~\ref{fig:vc}).
Contrarily, Aq-D's bar does not show such strong evolution in
the length and its amplitude is almost constant until $t \sim 9.8$~Gyr,
around which it sharply drops to the lower value.
We also compute the bar length in Aq-D by excluding the
component corresponding to the outer peak.
To do so, we find the radius between the inner and outer peaks
at which the amplitude, $A_2(r)$, takes the minimum value.
If this radius is shorter than the bar length defined by the
phase profile, we employ this radius as the bar length.
The bar length defined this way exhibits very weak time evolution.
The comparison between two bar lengths with the different definitions
implies that
the outer region of Aq-D's bar is violently disturbed
by the interactions with the stellar clumps.
We will show both bar lengths when the bar length is matter.
\subsection{Pattern speed of the bars} \label{sec:pattern}
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig05.eps}
\end{center}
\caption{
Time evolution of the bar pattern speed, $\Omega_\mathrm{bar}$,
and the amplitude, $A_2^\mathrm{max}$, around $z = 1$, 0.5, and 0
(from left to right).
We show Aq-C and Aq-D in the upper and lower six panels, respectively.
The bar pattern speed and the amplitude are respectively shown
in the upper and lower three panels of each group of panels
as functions of cosmic time.
}
\label{fig:pattern}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig06.eps}
\end{center}
\caption{
Same as figure~\ref{fig:pattern} but now as functions of the
bar angle, $\phi_\mathrm{bar}$.
}
\label{fig:patternphi}
\end{figure}
The time-steps used to output the simulation snapshots are
too large to measure the bar pattern speed.
Moreover, the change of the disk orientation with time \citep{okamoto13}
makes it difficult to measure the pattern speed.
We hence restart the simulations around $z = 1$, 0.5, and 0, and then
follow their evolution for a short period of time during which
the orientation of the disks hardly changes;
when the rotation axis of the disk is most misaligned with the initial one, the cosine of the angle between them is $0.996$,
which occurs in the simulation for Aq-C at $z \sim 1$.
In other simulations this value is always greater than $0.999$.
The disk does not tumble during these periods, and thus the misalignment
monotonically increases with time.
We employ the output time-step of $10$~Myr for these additional simulations.
In figure~\ref{fig:pattern} we show the bar pattern speed,
$\Omega_\mathrm{bar} \equiv \dot{\phi}_\mathrm{bar}$, and the amplitude,
$A_2^\mathrm{max}$, as functions of time around $z = 1$, 0.5 and 0.
Firstly, we find that the pattern speed of the bars slows down
from $z = 1$ to 0 in both galaxies.
The long-term behavior of Aq-C's bar is consistent with what we expect
from the results of the idealized simulations, that is, the bar becomes
stronger and longer as it rotates more slowly.
Aq-D's bar however does not change its amplitude between
$z \simeq 1$ and 0.5 while it's pattern speed decreases.
the bar amplitude becomes smaller from $z \simeq 0.5$ to $0$.
The pattern speed and the amplitude of Aq-C's bar display large
oscillations at $z \simeq 1$ and 0.5. Interestingly, the
short-term behavior is similar to the long-term one, i.e. the bar gets
stronger when the pattern speed decreases. This oscillation
becomes smaller with time and almost vanishes at $z \simeq 0$.
Such oscillations have also been observed in idealized simulations
\citep{dubinski09, amr13}.
\citet{amr13} suggest that the oscillation is caused by the
interaction between a bar and a host triaxial halo.
We will investigate halo properties in section~\ref{sec:halo}.
The small oscillations seen in the pattern speed of Aq-D's bar do not
correlate with the oscillations of the bar amplitude; the amplitude is
more or less constant during each period of time.
The time evolution of the direction of the rotation axis is not the
reason for the oscillations because the angle with the initial
rotation axis ($z$-direction) is sufficiently small.
Moreover this angle monotonically increases with time and does not
oscillate.
We suspect that the oscillation of the pattern speed of Aq-D's bar is
caused by the interactions between the bar and the clumps.
To see the short-term behaviors more closely, we now plot
the pattern speed and the amplitude as functions of the bar angle,
$\phi_\mathrm{bar}(t)$, in figure~\ref{fig:patternphi}.
We find that the periods of the oscillations in the amplitude
and the pattern speed are half a revolution period of the bar
when they show large oscillations (at $z \simeq 1$ and $0.5$ in Aq-C).
On the other hand, we do not see such a periodicity for Aq-D's bar.
\subsection{Which component obtains the angular momentum?} \label{sec:torque}
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig07.eps}
\end{center}
\caption{
The $z$-component of the specific gravitational torque acting on
the stars in the bar regions at $z \simeq 1$, 0.5, and 0.
The upper and lower panels show Aq-C and Aq-D, respectively.
The red circles indicate torque from all the particles
within virial radius, $r_\mathrm{vir}$.
The cyan squares, green stars, and blue diamonds respectively
show the contributions of the dark matter, stars, and gas.
The vertical dotted lines in the upper left and the upper middle
panels indicate the bar angle at which the bar pattern speed is
the smallest.
}
\label{fig:torque}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{fig08.eps}
\end{center}
\caption{
Radial profiles of the halo triaxiality.
The left panels show the halo properties of Aq-C and the right
ones show those of Aq-D.
From top to bottom, we display the radial profiles of the
minor-to-major axial ratio, $c/a$, the intermediate-to-major
axial ratio, $b/a$, the directional between the disc
ration axis and the halo minor axis, and the azimuthal direction
of the halo major axis.
From left to right, we show the halos at $z = 1$, $0.5$, and $0$
for each galaxy.
}
\label{fig:shape_diff}
\end{figure*}
Since the bar pattern speed slows down from $z = 1$ to 0,
we expect that the angular momentum of the bars is transferred to
other components, such as outer disks or dark halos.
In order to identify which component plays the most important
role in spinning down the bars, we try to measure the torque
from each component acting on the bars.
Doing this is however not straightforward because defining
particles that constitute the bar is not a simple task.
It is also difficult to relate the measured torque to the
change in the bar pattern speed even if we somehow define
the particles that belong to the bar, since
the angular momentum of the bar,
$L_\mathrm{bar} = I_\mathrm{bar} \Omega_\mathrm{bar}$,
where $I_\mathrm{bar}$ is the moment of the inertia of the bar,
is different from the total angular momentum of the particles
that constitute the bar.
We therefore take a simpler approach.
We define the bar region as a disk with the radius, $r_\mathrm{bar}$,
and the height 1~$h^{-1}$~kpc, and then compute the torques acting on
the star particles within this region.
Since the $z$-component of the torque acting on an axisymmetric
component is zero, the torque mainly operates on the bar as long as the bar
is the most significant non-axisymmetric structure in the region.
We have checked that we obtain qualitatively equivalent results when we
calculate torques on the particles along the bar.
We have also confirmed the results does not change qualitatively if we employ
the bar length by excluding the outer $m = 2$ component for Aq-D.
In figure~\ref{fig:torque}, we show the $z$-component of the specific
torque from the particles within the virial radius acting on the stars
in the bar region as a function of the bar angle, $\phi_\mathrm{bar}(t)$.
We find that the torque from the dark halo dominates the total
torque in Aq-C and it is negative on average; the torque from the dark
matter spins down the stars in the bar region more strongly at higher
redshift.
In Aq-C, the torque from the dark matter shows the periodic change with
the period of half a bar revolution period.
This behavior suggests that the torque is exerted by the anisotropic
distribution of the dark matter.
Although the period with which the torque changes is the same as the
period with which the pattern speed changes in Aq-C at $z \simeq 1$ and
$0.5$, the change in the angular momentum of the stars within
the bar region is not directly reflected to the bar pattern speed.
In the panels of $z \simeq 1$ and $0.5$ for Aq-C, we indicate the
bar angle where the bar pattern speed becomes the smallest.
We find that, against the intuition, the torque takes the largest
negative value where the bar pattern speed becomes the smallest.
The pattern speed should be the smallest where the torque changes its
sign from minus to plus
if the angular momentum of the stars within the bar region were
directly reflected to the bar pattern speed.
Moreover, the pattern speed of Aq-C's bar at $z \simeq 0$ is almost
constant while the torque show the same level of variation as at
$z \simeq 1$ and 0.
This is a direct consequence of the fact that the bar angular momentum,
$I_\mathrm{bar} \Omega_\mathrm{bar}$, is different physical quantity
from the total angular momentum of the particles belonging to the
bar as we mentioned earlier.
In Aq-D, the amplitude of the specific torque is much smaller than
that in Aq-C. The long-term evolution of the bar pattern speed in
Aq-D is thus much more moderate than Aq-C.
The torque is not always dominated by the contribution of the
dark matter, for example the torque is dominated by the contribution
of the stars at $z \simeq 0.5$.
This is consistent with the fact that the central density of the dark
halo of Aq-D is much lower than that of Aq-C and thus there is less
dark matter to absorb the angular momentum of the bar.
\citet{athanassoula03} show that the angular momentum is emitted
from near-resonant material at the inner Lindblad resonance (ILR),
i.e. a bar, and absorbed by mainly by near-resonant material at the
corotation and the outer Lindblad resonance in the halo;
the near-resonant material in the outer disk also absorbs the angular
momentum, but the role the outer disk plays is much less significant than
the halo.
This picture is in good agreement with the analytic calculations
\citep{lynden-bell72, tremaine84}.
Our results show that the halo is the main absorber of the angular momentum
when a significant amount of the angular momentum of the bar is transferred
from the bar. The results thus qualitatively agree with those by the idealized
simulations and the prediction of the analytic calculations.
\subsection{Halo properties} \label{sec:halo}
\citet{amr13} claim that the oscillations seen in the bar amplitude can
be understood by the interaction between a bar and its host triaxial halo.
They find that $A_2^\mathrm{max}$ has a minimum and the halo
intermediate-to-major axial ratio, $b/a$, has a maximum when
the bar and halo major axis are aligned. On the other hand,
$A_2^\mathrm{max}$ has a maximum and the halo $b/a$ has a minimum
when the bar and halo major axis are perpendicular.
These results imply that the triaxial halo plays a role of an outer bar
and the basic building blocks of the bar are loops.
We thus investigate the halo triaxiality in this section.
We have removed the subhalos identified by {\scriptsize SUBFIND}
\citep{spr01} in order to consider only the smooth component of
the dark halos.
The results are however almost identical to the case in which
we use the dark matter distribution as it is, except at the
outer parts of the dark halos.
In figure~\ref{fig:shape_diff}, we show the radial profiles of
the halo triaxiality, where each radial bin contains 20000 dark
matter particles.
The inner halo has a higher sphericity than the outer halo and
the inner halo's minor axis is well aligned with the disk rotation axis.
The direction of the minor axis of the outer halo is almost independent
from that of the inner halo.
The baryonic contraction due to gas cooling makes the inner halo
more spherical than dark matter only simulations \citep{kazantzidis04, bai05}.
The tidal torque between the disk and the inner halo aligns the rotation
axis and the halo minor axis \citep{bai05}.
We also find that Aq-D's halo is more spherical than Aq-C's.
This explains the small amplitude of the specific torque from
dark matter acting on the bar in Aq-D.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig09.eps}
\end{center}
\caption{
Time evolutions of the bar amplitude and the halo triaxiality of Aq-C
around $z = 1$, $0.5$ and $0$ (from left to right).
From top to bottom, we present the bar amplitude, $A_2^\mathrm{max}$,
the minor-to-major halo axial ratio, $c/a$, the intermediate-to-major
halo axial ratio, $b/a$, and the azimuthal direction of the halo major
axis, $\phi_\mathrm{major}$.
We also show the axis ratio of the dark matter ring in the disk plane and
the azimuthal direction of its major axis by the red dashed lines.
The vertical dashed lines indicate the time when the directions of the halo
major axis and the bar are aligned and the vertical dotted lines indicate
the time when the directions of the major axis of the dark matter ring and bar
are aligned.
}
\label{fig:halo_oscillation}
\end{figure}
We now explore the time oscillations of the halo triaxiality of Aq-C that
shows oscillations in the bar amplitude and pattern speed at high redshift.
\citet{am02} claim that the halo triaxiality should be measured
in a density bin instead of a radial bin since the former is noisier
than the latter.
It is however not the case in our simulations probably because
the dark matter density distribution in cosmological halos are not as
smooth as that in the idealized simulations even if we remove
the subhalos.
We hence measure the triaxiality of a spherical shell of radius
between $r_\mathrm{bar}$ and $2 r_\mathrm{bar}$, where we employ
the mean bar length during each simulation as $r_\mathrm{bar}$.
We first sort the dark matter particles by the distance from the center
and identify the ranking of the particles that lie between
$r_\mathrm{bar}$ and $2 r_\mathrm{bar}$ in the first snapshot.
In the rest of the snapshots, we use the dark matter particles
that have the same ranking in radius as the particles selected in the
first snapshot to measure the orientation and the shape of the halo.
Our results are not sensitive to the choice of the radius of the shell
as long as that lies between $r_\mathrm{bar}$ and $3 r_\mathrm{bar}$.
Within $r_\mathrm{bar}$, the direction of the halo major axis largely
oscillates, but not rotates. In the outer part,
($r \gtrsim 3 r_\mathrm{bar}$), the direction of the major axis of the
halo is independent of that of the inner part
(figure~\ref{fig:shape_diff}).
Since the minor axis of the halo is not parallel to the rotation axis
of the disk as shown in figure~\ref{fig:shape_diff}, the azimuthal
direction of the major axis of the halo can be different from that of the
major axis of the dark matter distribution in the disk plane as we will
show later.
We thus measure the axis ratio of the dark matter ring of $r_\mathrm{bar} <
r < 2 r_\mathrm{bar}$ and $|z| < 0.1 h^{-1}$~kpc and the direction of its
major axis, too.
In figure~\ref{fig:halo_oscillation}, we show the time evolution of the bar
amplitude, the minor-to-major halo axial ratio,
and the intermediate-to-major halo axial ratio,
together with the axial ratio of the dark matter ring.
We also show the azimuthal directions of the major axes of the dark matter
shell and ring. The directions of these axes are better aligned at lower
redshift simply because the directions of the halo minor axis and the disk
rotation axis are better aligned at lower redshift as shown in
figure~\ref{fig:shape_diff}.
We find that the axial ratios, $b/a$, of the shell and ring oscillate with
the same frequency as the bar oscillation.
The minor-to-major axial ratio, $c/a$, also shows similar oscillations to
$b/a$.
These results agree with the finding by \citet{amr13}.
The bar amplitude is however largest when the bar and the major axis
of the dark matter ring are parallel. This behavior contradicts the
result by \citet{amr13} and the loop concept that predicts a loop
corresponding to an inner bar is less elongated when the inner
and outer bars are aligned \citep{maciejewski00, maciejewski07}.
The simulation results by \citet{heller01} and \citet{heller07b}
are more analogous to ours. They find that the inner bar/ring component
is more elongated when it is parallel to the outer bar.
On the other hand, the halo is more axisymmetric, i.e. $c/a$ and $b/a$ are
large, when the bar and the major axis of the dark matter ring are aligned.
This result is consistent with the behaviors of the halo triaxiality
found by \citet{amr13}.
Further studies are needed to understand the origin of the
disagreement among the simulations.
\subsection{Interaction with the $m = 4$ Fourier mode} \label{sec:m4}
In this subsection, we investigate the second most significant Fourier
mode, the $m = 4$ mode, and explore possible interaction between
the $m = 2$ and $m = 4$ Fourier modes.
In figure~\ref{fig:m4}, we plot the amplitude and the phase profiles of
the $m = 4$ Fourier mode. Since the phase of the $m = 4$ mode,
$\phi_4(r)$, is defined between 0 and $\pi/2$,
we also plot $\phi_4(r) + \pi/2$ to investigate the alignment between $m = 4$ and $2$ modes. Note that as Aq-C's snapshot at $z \simeq 1$, we chose
the one in which the misalignment between the two components is evident.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig10.eps}
\end{center}
\caption{The amplitude and phase profiles of the $m = 2$ and $m = 4$ modes.
The upper and lower six panels show Aq-C and Aq-D, respectively.
The blue solid and red dotted lines respectively represent the
$m = 2$ and $m = 4$ modes.
The amplitude and phase are shown in the upper and lower panels
of each group of panels.
From left to right, the results at $z = 1$, 0.5, and 0 are presented.
For the phase, $\phi_4(r)$, we also plot $\phi_4(r) + \pi/2$ in
order to compare it with the $m = 2$ mode.
}
\label{fig:m4}
\end{figure}
We find that the relative importance of the $m = 4$ component to the
$m = 2$ component is larger in Aq-C than in Aq-D.
This result is consistent with those obtained by many simulations.
In \citet{am02}, the relative importance of the higher order even moments
is higher in a galaxy with a stronger bar.
The same trend is seen in cosmological bars \citep{scannapieco12}.
In Aq-C, the maxima of $A_4(r)$ occur at almost the same radii as
the maxima of $A_2(r)$, i.e. at $r_2^\mathrm{max}$.
Usually in idealized simulations, the maxima of the amplitudes of the higher
order even moments occur considerably larger radii than $r_2^\mathrm{max}$
in galaxies with massive halos \citep{am02, athanassoula03}.
\citet{scannapieco12} obtain the same result as ours for the same Aquarius
halo, Aq-C, in spite of their lower resolution and weaker feedback,
i.e. a heavier disk, than ours \citep{sca09}.
This agreement implies that the halo formation history and its shape play
an important role in shaping substructure of a galaxy.
In Aq-C, the $m = 2$ component is largely misaligned with the $m = 4$
component at $z = 1$ and is slightly misaligned at $z = 0.5$.
At $z = 0$ the two modes are perfectly aligned with each other.
On the other hand, the two modes are always aligned in Aq-D.
It is interesting that the bar and the $m = 4$ component are
misaligned when the bar shows large oscillations in its pattern speed
and amplitude.
Next, we explore the relation between the pattern speed and amplitude
of the $m = 2$ and $m = 4$ components.
To calculate the pattern speed of the $m = 4$ component,
we define $A_4^\mathrm{max}$ and $r_4^\mathrm{max}$ by exactly the same way
as we defined $A_2^\mathrm{max}$ and $r_2^\mathrm{max}$ and then we
define the phase of the $m = 4$ component as $\phi_4(r_4^\mathrm{max})$.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig11.eps}
\end{center}
\caption{
Same as figure~\ref{fig:pattern} but we now show the pattern speed
and amplitude of the $m = 4$ components as well.
The $m = 2$ and $m = 4$ components are respectively represented by
the blue solid and red dotted lines.
}
\label{fig:m4pattern}
\end{figure}
In figure~\ref{fig:m4pattern}, we show the pattern speed of the $m = 2$
component, $\Omega_\mathrm{bar} \equiv \dot{\phi}_\mathrm{bar}$, and
the $m = 4$ component, $\dot{\phi}_4(r_4^\mathrm{max})$.
We find that the pattern speed of the $m = 4$ component also shows large and
periodic oscillation at $z \simeq 1$ in Aq-C.
This oscillation becomes much smaller at $z \simeq 0.5$ at which
the misalignment between the two modes is small (see figure~\ref{fig:m4}).
Once the phases of the two components are perfectly aligned with each other,
the pattern speed of the both components becomes almost constant and of
course the two components have the same pattern speed.
The amplitude of the $m = 4$ component also varies with the same frequency
as the pattern speed. As for the $m = 2$ component, the amplitude becomes
large when the pattern speed is small and vise versa.
The frequency of the oscillation of the pattern speed of the $m = 4$
component is clearly higher than that of the $m = 2$ component.
We find that the period of the oscillations of
the pattern speed and the amplitude of the $m = 4$ component is
a quarter of its revolution period.
As we have shown, the $m = 2$ and $m = 4$ components in Aq-D are always
aligned with each other. In this case, the pattern speed and the amplitude of
both $m = 2$ and $m = 4$ components do not show the periodic oscillations,
which are seen in Aq-C at $z \simeq 1$ and 0.5.
The high-frequency oscillations are probably caused by the interactions
with clumps as we have already discussed for the $m = 2$ mode.
Our results suggest that the pattern speed of the bar oscillates if its
phase is misaligned with the phase of the $m = 4$ component in the same
region.
\subsection{Resonances and the bar length}
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig12.eps}
\end{center}
\caption{
Behavior of $\Omega - \kappa/2$, $\Omega - \kappa /4$ and $\Omega$
at $z = 1$, 0.5, and 0.
The blue dashed, green dotted, and red dot-dashed lines respectively
represent $\Omega - \kappa/2$, $\Omega - \kappa/4$, and $\Omega$.
The upper and lower panels indicate Aq-C and Aq-D, respectively, and
redshifts are $1$, $0.5$, and $0$ from left to right.
We show the bar pattern speed around these redshifts by the horizontal
black solid lines, whose lengths correspond to the bar lengths.
For Aq-D, we also show the case in which we define the bar lengths
by excluding the outer component (horizontal yellow solid lines).
}
\label{fig:resonance}
\end{figure}
Finally we investigate the resonance structure of the simulated galaxies.
In figure~\ref{fig:resonance} we show the behaviors of $\Omega - \kappa/2$,
$\Omega - \kappa/4$, and $\Omega$ as functions of radius,
where $\Omega$ is the angular frequency of a circular orbit and $\kappa$
is the radial angular frequency.
The sharp rise of the $\Omega - \kappa/2$ curves towards the center is
due to the gravitational softening.
Since the spatial extent of this region is much smaller than the scale
we are interested in, we will ignore the innermost 2:1 resonance due to
this sharp rise in the following discussion.
Each $\Omega - \kappa/2$ curve has a peak and thus these galaxies
have double ILRs if the bar pattern speed is smaller than the
peak value (and if we ignore the inner most ILR).
The peak values decrease with time since the central density of
the galaxies decrease with time due to the stellar evolution and the
feedback (see figure~\ref{fig:vc}).
The density structure does not change on the short time-scale during
which we measure the pattern speed of the bars around $z = 1$, 0.5, and 0.
We show the bar pattern speed and the bar length measured around these
redshifts by the horizontal lines.
From the behavior of the bar pattern speed and the amplitude in Aq-C
around $z = 1$ and 0.5, we speculate that
the smaller the pattern speed is, the larger the amplitude is
if the $\Omega - \kappa/2$ curve is fixed \citep{kormendy13}.
On the long time-scale, $\Omega - \kappa/2$ curve is lowered as shown in
figure~\ref{fig:resonance}.
In Aq-C, the bar pattern speed strongly decreases from $z = 1$ to 0.
As a result, the bar amplitude becomes larger with time
(figure~\ref{fig:zevo}).
In Aq-D, the slow down rate of the bar pattern speed is much lower than that
in Aq-C. Consequently, the peak value of $\Omega - \kappa/2$ curve decreases
as fast as or faster than the bar pattern speed.
The bar amplitude in this case does not change much or becomes smaller
as shown in figure~\ref{fig:zevo}.
At $z \simeq 0$, there is no 2:1 resonances in many cases if we ignore the
inner most resonance.
As a result, the bar amplitude at $z = 0$ is smaller than that at $z = 0.5$.
We will discuss the sharp decline in the amplitude of Aq-D's bar at
$t \sim 9.8$~Gyr shown in figure~\ref{fig:zevo} in the next section.
The bar lengths are around the 4:1 resonances in both galaxies at
all redshifts.
Even if we define the length of Aq-D's bar by excluding the outer component,
the bar looks terminated at this resonance except at $z \simeq 0$.
\citet{patsis97} study the orbital structure in the potential model of
NGC 4314 and predict that the longest stable periodic orbits are
found at the 4:1 resonance.
Our cosmological simulations confirm their prediction.
\section{Discussion and conclusions}
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{fig13.eps}
\end{center}
\caption{
Evolution of the bar pattern speed and amplitude of Aq-D
around $t \simeq 9.8 Gyr$.
In the top panel, we show the evolution of the bar pattern speed
and the peak value of the $\Omega - \kappa/2$ curve by the blue solid and
black dotted lines respectively.
In the middle panel we also plot the specific angular momentum of the
stars in the bar region ($r < 1.68$~kpc and $|z| < 1~h^{-1}$~kpc).
In the bottom panel, we show the bar amplitude $A_2^\mathrm{max}(t)$.
}
\label{fig:resoevo}
\end{figure}
We have investigated the cosmological evolution of bars by utilizing
the two Milky Way-mass galaxies formed in the fully self-consistent
simulations of the galaxy formation by \citet{okamoto13}.
The evolution qualitatively agrees with what is expected from idealized
simulations, that is, the bar in the galaxy having more centrally
concentrated mass distribution exhibit stronger evolution than the
other \citep{combes93, athanassoula03}.
This picture is in good agreement with observations (e.g. \cite{cheung13}).
The bar in Aq-C receives the large negative torque from the dark matter
and is significantly slowed down $z = 1$ to 0.
As the bar pattern speed decreases, the bar becomes stronger and longer.
On the other hand, the torque acting on Aq-D's bar is much weaker by
lacking materials that absorb the angular momentum of the bar.
The pattern speed of Aq-C's bar violently oscillates at $z \simeq 1$.
The oscillation becomes smaller at $z \simeq 0.5$ and disappears at
$z \simeq 0$. The period of the oscillation as a function
of the bar angle is $\pi$, indicating the interaction with the halo's
quadrapole moment.
The amplitude of the bar also shows the oscillations with the same frequency
as those of the pattern speed.
When the bar rotates slower, it becomes stronger and vise versa.
This short-term behavior is consistent with the long-term behavior in Aq-C,
i.e. the slower the bar rotation is, the lager its amplitude is.
These oscillations correlate with the oscillations in the halo triaxiality
as pointed out by \citet{amr13}. When the bar and the halo major axis
are parallel, the halo is more spherical and the bar is strongest.
The former result is agree with the simulations by \citet{amr13}, but
the latter is opposite to them. The origin of the disagreement is
unclear.
The evolution of the bar in Aq-D seems to contradict this scenario.
While the pattern speed of the bar decreases with time, its amplitude
does not increase from $z = 1$ to 0.5 and even decreases from
$z = 0.5$ to 0.
We speculate that this is because the $\Omega - \kappa/2$ curve is
lowered with time (figure~\ref{fig:resonance}) due to the decreasing
central density with time (figure~\ref{fig:vc}),
which is most likely caused by the mass loss and feedback from the
stellar populations.
The inclusion of the mass loss through entire life of stellar
populations and strong feedback
is one of the biggest differences between our cosmological simulations
and idealized simulations of isolated galaxies.
The instantaneous recycling approximation is often used in
idealized simulations and old stars do not lose their mass even
when gas cooling, star formation,
and feedback are included in idealized simulations (e.g. \cite{amr13}).
Thanks to the efficient angular momentum transfer from the bar to the
dark matter in Aq-C, the bar pattern speed drops faster than
the $\Omega - \kappa/2$ curve, and thus the bar becomes stronger.
On the other hand, the slowdown rate of Aq-D's bar is as low as
the decreasing rate of the peak value of the $\Omega - \kappa/2$ curve.
In Aq-D, the amplitude of the bar sharply drops at $t \sim 9.8$~Gyr.
To see why the bar is weakened, we measure the bar pattern speed
and the peak value of the $\Omega - \kappa/2$ curve,
$(\Omega - \kappa/2)_\mathrm{peak}$, around this epoch.
In figure~\ref{fig:resoevo}, we show the evolution of the bar pattern speed
by restarting the simulations from several snapshots and compare it with
the evolution of $(\Omega - \kappa/2)_\mathrm{peak}$.
We find that the bar pattern speed is in fact an increasing function of time
on average until $t \simeq 9.9$~Gyr, while
$(\Omega - \kappa/2)_\mathrm{peak}$ slowly decreases with time.
As the bar pattern speed approaches $(\Omega - \kappa/2)_\mathrm{peak}$,
the bar amplitude decreases.
We also show the evolution of the specific angular momentum of the stars
in the bar region ($r < 1.68$~kpc).
The specific angular momentum in the bar region sharply increases with
time until $t \simeq 9.9$~Gyr.
The angular momentum is most presumably brought by the clumps.
\citet{okamoto13} has shown that a non-negligible amount of mass is added to
Aq-D's bulge by the clumps during this epoch.
The high frequency oscillations in the bar pattern speed is likely to be
cause by the interactions with the clumps.
At $t \gtrsim 9.9$~Gyr, the bar pattern speed often exceeds
$(\Omega - \kappa/2)_\mathrm{peak}$.
These results support the idea that the bar amplitude is determined by
the relation between the pattern speed and the $\Omega - \kappa/2$ curve.
Interestingly, the oscillations in the bar pattern speed and amplitude
in Aq-D's bar are observed only when the $m = 2$ component is misaligned
with the $m = 4$ component.
The oscillations become smaller as the two components get aligned with
each other.
The $m = 4$ components have the comparable spatial size to the
$m = 2$ components in both galaxies.
The bar lengths seem to be determined by the 4:1 resonances.
Our cosmological simulations thus confirm the prediction by
\citet{patsis97}.
Our results are also consistent with the observation of
NGC~253, which has a bar whose length coincides
with 4:1 resonance \citep{sorai00}.
In summary, using high-resolution cosmological simulations of disk galaxy formation, we show that the bar evolution in the cosmological simulations is qualitatively consistent with that obtained by idealized simulations of isolated disk galaxies and observations.
We find that the strong feedback and continuous mass loss from stellar populations significantly lowers the central density with time and hence changes the resonance structure. We also find that the clumps formed in the disk can spin up the bar and can weaken its strength.
Our sample is too small to understand why the oscillations in bar amplitude
correlate with those in halo triaxiality differently from
the idealized simulations by \citet{amr13} and thus we leave it for future
studies.
\bigskip
We are grateful to the anonymous referee for the careful reading of
the manuscript and the thoughtful comments.
We would like to thank Masafumi Noguchi and Junich Baba for helpful
discussion.
We also thank John Kormendy, Francoise Combes, and Shunsuke Hozumi
for useful comments on the manuscript.
Numerical simulations were carried out with Cray XC30 in CfCA at NAOJ
and T2K-Tsukuba in Center for Computational Sciences at University
of Tsukuba.
TO acknowledges the financial support of Japan Society for the Promotion of
Science (JSPS) Grant-in-Aid for Young Scientists (B: 24740112).
|
\section{Introduction}
\subsection{Finite vs. infinite index}
In \cite{MR696688}, Jones pioneered the modern theory of subfactors.
Starting with a finite index {\rm II}$_1$-subfactor $A_0\subseteq A_1$, he used his basic construction to construct the Jones tower $(A_n)_{n\geq 0}$ iteratively by adding the Jones projections $(e_n)_{n\geq 1}$, which satisfy the Temperley-Lieb relations.
Jones used these Temperley-Lieb algebras to show that the index lies in the range $\set{4\cos^2(\pi/n)}{n\geq 3}\cup [4,\infty)$, and he found a hyperfinite subfactor for each allowed index value.
A finite index subfactor is studied by analyzing its standard invariant, the two towers of finite dimensional centralizer algebras $(A_i'\cap A_j)_{i=0,1;j\geq 0}$.
The standard invariant has been axiomatized in three different ways: Ocneanu's paragroups \cite{MR996454}, Popa's $\lambda$-lattices \cite{MR1334479}, and Jones' planar algebras \cite{math/9909027}.
Some finite index results generalize to infinite index subfactors.
Discrete, irreducible, ``depth $2$" subfactors correspond to outer (cocycle) actions of Kac algebras \cite{MR1055223,MR1387518}.
The classical Galois correspondence also holds for outer actions of infinite discrete groups and minimal actions of compact groups \cite{MR1622812}.
Burns, in his Ph.D. thesis \cite{1111.1362}, studied extremality and rotations for infinite index subfactors, as the key ingredient in proving isotopy invariance for Jones' planar algebras in \cite{math/9909027} is the rotation operator (also known to Ocneanu \cite{MR1317353}).
Essentially, Burns observed that for infinite index subfactors, the centralizer algebras $A_0'\cap A_{n}$ and the central $L^2$-vectors $A_0'\cap L^2(A_n)$ do not coincide.
Using this observation, the second author generalized the work of Burns in \cite{MR3040370}, where he gave a planar calculus for an arbitrary index {\rm II}$_1$-factor bimodule $\sb{A}H_A$.
Setting $H^n=\bigotimes_A^n H$, he found two planar operads acting on the centralizer algebras $\cQ_n=A'\cap (A\op)'\cap B(H^n)$ and the central $L^2$-vectors $\cP_n=A'\cap H^n$ respectively whose actions are compatible.
We recover the subfactor case when $A=A_0$ and $H=L^2(A_1)$.
Interestingly, this planar structure was discovered without the use of Jones' basic construction and without the resulting Jones projections.
Hence we have one possible definition for the standard invariant of an infinite index subfactor, or a {\rm II}$_1$-factor bimodule: the centralizer algebras $\cQ_\bullet=(\cQ_n)_{n\geq 0}$ and the central $L^2$-vectors $\cP_\bullet=(\cP_n)_{n\geq 0}$, together with their compatible planar calculi.
\subsection{The simplest possible standard invariant}
The Jones subfactors with index at most 4 discovered in \cite{MR696688} have the simplest possible standard invariants; they consist entirely of the Temperley-Lieb algebras generated by the Jones projections.
Since these projections are always contained in the centralizer algebras, the Temperley-Lieb standard invariant is always contained within the standard invariant of a finite index subfactor.
Hence each subfactor planar algebra has a canonical Temperley-Lieb planar subalgebra.
In \cite{MR1198815}, for every index greater than $4$, Popa found a (non-hyperfinite) subfactor whose standard invariant is only Temperley-Lieb, and his methods led to his famous subfactor reconstruction theorem \cite{MR1334479}.
An important open question is to determine for which indices greater than 4 there is a hyperfinite subfactor whose standard invariant is Temperley-Lieb.
The main motivation for this article is the following question.
\begin{quest}
For infinite index subfactors, what is the simplest possible standard invariant?
\end{quest}
When the index is infinite, one still has a Jones tower $(A_n)_{n\geq 0}$ of type {\rm II} factors, but $A_n$ is type {\rm II}$_\infty$ for $n\geq 2$ (see Section \ref{sec:SubfactorBackground}).
In this case, Burns showed in \cite{1111.1362} that the odd canonical trace-preserving operator-valued weight $T_{2n+1} : A_{2n+1}\to A_{2n}$ is a conditional expectation, which results in an odd Jones projection $e_{2n+1}\in \cQ_{n+1}$.
We immediately see that $\dim(\cQ_n)\geq n$, since the abelian algebra generated by the odd Jones projections is contained in $\cQ_n$.
However, the odd Jones projections actually give us non-abelian structure as well.
\begin{thm}
The odd Jones projections are equivalent in $\cQ_n$.
Hence $\cQ_n$ is not abelian for $n\geq 2$.
\end{thm}
We prove this result in more generality for the case of a {\rm II}$_1$-factor bimodule $\sb{A}H_A$ containing a distinguished central vector $\zeta$, so $\cP_n\neq (0)$.
This is the natural analog of the bimodule $H=L^2(A_1)$ with distinguished $A_0$-central vector $\widehat{1}$.
We give the odd Jones projections for such bimodules in Section \ref{sec:BimoduleRepresentations}.
\subsection{GICAR and planar rook algebras and categories}
The Temperley-Lieb algebras appear implicitly in Lieb's ice-type model in statistical mechanics \cite{LiebIceModel,MR0498284}, \cite[Section 2.5]{JonesPANotes}.
The canonical algebra generated by the odd Jones projections together with the partial isometries witnessing the equivalences is actually another well-studied canonical operator algebra which arises in the study of fermions.
\begin{thm}\label{thm:Main2}
The gauge-invariant canonical anticommutation relations algebra $\GICAR(\cH_n)$ (also known as the fermion algebra) where $\dim(\cH_n)=n$ is represented faithfully in $\cQ_n$ as the odd Jones projections and the partial isometries between them.
\end{thm}
For finite index subfactors, this map was constructed by Connes and Evans in their work on representations of the Virasoro algebra \cite{MR990778}.
Our map is the bimodule analog, which is independent of von Neumann dimension.
In fact, there is a simple proof of the existence of such an injection, although further analysis is needed to show the image is correct.
Our distinguished $A$-central vector $\zeta\in H$ yields an $A-A$ bimodule isomorphism $H\cong L^2(A)\oplus K$ where $L^2(A)\cong \overline{A\zeta}^{\|\cdot\|_2}$ and $K\cong\{\zeta\}^\perp$.
By the binomial theorem,
$$
\bigotimes_{A}^n H
\cong
(L^2(A)\oplus K)^{\otimes_A n}
\cong
\bigoplus_{j=0}^n {n\choose j} K^j
$$
where $K^0=L^2(A)$, and $K^j=\bigotimes_A^j K$.
The obvious intertwiners amongst the $n\choose j$ copies of $K^j$ give a canonical inclusion
$$
\bigoplus_{j=0}^n M_{n\choose j}(\mathbb{C})
\hookrightarrow
\End_{A-A}\left( \bigotimes^n_A H\right).
$$
The left hand side above is isomorphic to $\GICAR(\cH_n)$.
We compute our map explicitly in Section \ref{sec:RectangularGICARReps}, and we show it is compatible with the towers $\GICAR(\cH_\bullet)=(\GICAR(\cH_n))_{n\geq 0}$ and $\cQ_\bullet$, along with their standard representations.
The GICAR tower arises from choosing an orthonormal basis $(\xi_n)$ of an infinite dimensional separable $\cH$, and setting $\cH_n=\spann\{\xi_1,\dots, \xi_n\}$.
Again, we do so in more generality:
\begin{thm}
The tower $\GICAR(\cH_\bullet)$ fits naturally into a ``rectangular" $*,\otimes$-category $\sR\sG$ which acts faithfully as $A-A$ bimodule maps amongst the $H^n$'s. This action extends the faithful representation from Theorem \ref{thm:Main2}.
\end{thm}
For a finite index subfactor, the image of Connes and Evans' map, which is also the map from Theorem \ref{thm:Main2}, consists of the Kauffman diagrams in the Temperley-Lieb algebra with only shaded caps and cups \cite[Lemma 4.2, Theorem 4.3 ]{MR990778}.
Contracting shaded regions, we obtain a diagrammatic algebra which also appears in the literature as the planar rook algebra (see Section \ref{sec:PlanarRookAlgebras}).
$$
\begin{tikzpicture}[rectangular]
\clip (2.5,1.1) --(-1.1,1.1) -- (-1.1,-1.1) -- (2.5,-1.1);
\filldraw[shaded] (1.8,-1)--(.8,1) -- (1.2,1) -- (2.2,-1);
\filldraw[shaded] (0,1)--(0,-1) -- (.4,-1) -- (.4,1);
\filldraw[shaded] (.8,-1) arc (180:0:.3cm);
\filldraw[shaded] (-.8,-1) arc (180:0:.3cm);
\filldraw[shaded] (-.8,1) arc (-180:0:.3cm);
\filldraw[shaded] (1.6,1) arc (-180:0:.3cm);
\draw [very thick] (2.4,1) --(-1,1) -- (-1,-1) -- (2.4,-1)--(2.4,1);
\end{tikzpicture}
\longleftrightarrow
\begin{tikzpicture}[baseline=-.1cm, scale=1.5]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\BrokenString{(-.4,0)}
\String{(-.2,0)}
\PartialIsoStar{(0,0)}
\end{tikzpicture}
\longleftrightarrow
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (3.2,-.2) -- (3.2,1.2) -- (-.2,1.2);
\draw (1,0) -- (1,1);
\draw (2,1) -- (3,0);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (2,0) circle (.1cm);
\filldraw (3,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\filldraw (2,1) circle (.1cm);
\filldraw (3,1) circle (.1cm);
\end{tikzpicture}
\right)$$
The representation theory of these diagrammatic algebras was studied in \cite{MR2541502}, where they showed the Bratteli diagram for the tower of algebras resulting from the right inclusion is Pascal's Triangle.
Of course this also follows from the isomorphism with the tower of GICAR algebras (see Sections \ref{sec:GICAR}, \ref{sec:PlanarRookCategories}, and \ref{sec:DigrammaticFermions}), and we get a diagrammatic representation of the infinite dimensional GICAR algebra in Section \ref{sec:DigrammaticFermions}.
We remark that Bigelow-Ramos-Yi showed that the Jones and Alexander polynomials can be recovered via traces on the planar rook algebras \cite{MR2978881}.
Just as there is an annular version of the Temperley-Lieb category, there is an annular GICAR category $\sA\sG$, which contains the rectangular GICAR category $\sR\sG$.
$$
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (0:1.15cm) -- (0:1.5cm);
\filldraw (0:1.15cm) circle (.05cm);
\draw (60:1.15cm) -- (60:1.5cm);
\filldraw (60:1.15cm) circle (.05cm);
\draw (60:0.5cm) .. controls ++(60:.4cm) and ++(-60:.4cm) .. (120:1.5cm);
\draw (180:0.5cm) -- (180:.85cm);
\filldraw (180:.85cm) circle (.05cm);
\draw (180:1.15cm) -- (180:1.5cm);
\filldraw (180:1.15cm) circle (.05cm);
\draw (0:0.5cm) .. controls ++(0:.4cm) and ++(120:.4cm) .. (300:1.5cm);
\draw (240:0.5cm) -- (240:.85cm);
\filldraw (240:.85cm) circle (.05cm);
\draw (240:1.15cm) -- (240:1.5cm);
\filldraw (240:1.15cm) circle (.05cm);
\node at (30:.75cm) {$\star$};
\node at (30:1.25cm) {$\star$};
\draw[very thick, red] (30:.5cm) -- (30:1.5cm);
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
\longleftrightarrow
\begin{tikzpicture}[rectangular]
\clip (1.9,1.1) --(-1.1,1.1) -- (-1.1,-1.1) -- (1.9,-1.1);
\draw (-.6,1)--(-.6,.35);
\filldraw (-.6,.35) circle (.05cm);
\draw (.2,1)--(.2,.35);
\filldraw (.2,.35) circle (.05cm);
\draw (.6,1)--(.6,.35);
\filldraw (.6,.35) circle (.05cm);
\draw (1.4,1)--(1.4,.35);
\filldraw (1.4,.35) circle (.05cm);
\draw (-.2,1)--(-.6,-1);
\draw (1,1)--(1.4,-1);
\draw (.1,-1)--(.1,-.35);
\filldraw (.1,-.35) circle (.05cm);
\draw (.7,-1)--(.7,-.35);
\filldraw (.7,-.35) circle (.05cm);
\draw [very thick] (1.8,1) --(-1,1) -- (-1,-1) -- (1.8,-1)--(1.8,1);
\end{tikzpicture}
$$
Diagrams in $\sA\sG$ are obtained from diagrams in $\sR\sG$ by tensoring the morphisms with themselves around the outside, i.e., gluing the rectangles into annuli, and then allowing for rotation.
This has two consequences:
\begin{enumerate}[(1)]
\item
$\sA\sG$ is no longer a tensor category, and
\item
$\sA\sG$ must act on the spaces obtained from the $H^n$'s by tensoring themselves on the outside, i.e., the invariant vectors of the bimodules.
\end{enumerate}
Using the Burns rotations studied in \cite{MR3040370}, we get the following theorem:
\begin{thm}
There is an action of the annular GICAR category $\sA\sG$ as maps amongst the sequence of central $L^2$-vectors $\cP_\bullet$.
\end{thm}
However, this action is not necessarily faithful, and there are subfactor examples where it is completely degenerate.
This is in stark contrast to the finite index case, where the action of the annular Temperley-Lieb category is never degenerate.
In Section \ref{sec:Representations}, we compute the representation theory of $\sR\sG$ and $\sA\sG$ in the spirit of Graham and Lehrer's cellular algebras \cite{MR1376244}, as was done for the affine and annular Temperley-Lieb categories in \cite{MR1659204} and \cite{MR1929335} respectively.
\subsection{Examples}
By Theorem \ref{thm:Main2}, we see that $\cQ_n$ must contain $\GICAR(\cH_n)$.
In Examples \ref{ex:OuterZAction} and \ref{ex:OuterZAction2}, we give an example of a {\rm II}$_1$-factor bimodule with a distinguished central vector such that $\cQ_n$ is exactly the image of $\GICAR(\cH_n)$ and $\dim(\cP_n)=1$ for all $n\geq 0$.
However, this example does not come from a subfactor, and at this point, we do not have such an example.
We note that in the subfactor case, $\cP_{2n} \cong L^2(\cQ_n,\Tr_n)$ \cite[Remark 4.27]{MR3040370}, and the only Hilbert-Schmidt element in $\cQ_n$ in the image of $\GICAR(\cH_n)$ is the product of all the odd Jones projections, which can be identified with the element $\widehat{1}\otimes \cdots \otimes \widehat{1}\in H^n$ (see Example \ref{ex:SInfinity}).
In \cite{MR3040370} it was shown that when $H=L^2(A_1)$ for the subgroup-subfactor $A_0=R\rtimes \Stab(1)\subset R\rtimes S_\infty=A_1$, we have $\dim(\cQ_n)<\infty$ and $\dim(\cP_n)=1$ for all $n\geq 0$.
\subsection{Outline}
In Section \ref{sec:Fermions}, we give a background on fermionic Fock space and the CAR and GICAR algebras along with planar rook algebras.
In Section \ref{sec:GICARCategories}, we define the diagrammatic annular and rectangular planar rook categories and the abstract annular and rectangular GICAR categories, and we show they are respectively equivalent.
We then give the classification of the finite dimensional Hilbert space representations of the annular and rectangular categories in Section \ref{sec:Representations}.
We give the background necessary for our {\rm II}$_1$-factor bimodule and subfactor representations of these categories in Section \ref{sec:InfiniteBackground}, and we construct these representations in Section \ref{sec:BimoduleRepresentations}.
\subsection{Future research}
We will continue to search for an example of an infinite index subfactor with the simplest possible standard invariant, or to attempt to show no such example exists.
In the recent article \cite{1110.5671}, the authors clarify the connection between bifinite Hilbert bimodules and two-sided dualizability.
Given an infinite index {\rm II}$_1$-subfactor $A\subset B$, the standard bimodule $\sb{A}L^2(B)_B$ is finite on only one side, so we have only one-sided duals.
In future work, we will clarify the connection between one-sided finite Hilbert bimodules and one-sided dualizability.
We will work with an operator-valued index for bimodules over finite von Neumann algebras which may be infinite in several distinct ways.
It would be interesting if there were different types of one-sided duals associated to the different flavors of one-sided finite index bimodules.
\subsection{Acknowledgements}
This work was completed in three installments:
while David Penneys visited Vanderbilt University in Spring 2011 (thanks to Dietmar Bisch and Jesse Peterson);
while both authors visited Institut Henri Poincar\'{e} during the 2011 trimester on von Neumann algebras and ergodic theory of groups actions (thanks to the organizers Damien Gaboriau, Sorin Popa, and Stefaan Vaes); and
during the Summer of 2013.
The second author would like to thank Michael Hartglass and James Tener for helpful conversations.
David Penneys was supported in part by the Natural Sciences and Engineering Research Council of Canada.
Both authors were also supported by
NSF DMS grant 0856316
and
DOD-DARPA grants HR0011-11-1-0001 and HR0011-12-1-0009.
\section{Fermions and planar rook algebras}\label{sec:Fermions}
In this section, we give the background material on fermionic Fock space, the CAR and GICAR algebras, and planar rook algebras.
\subsection{Fermions, CAR, and GICAR}\label{sec:GICAR}
We take the following definitions from \cite[Chapter 18]{JonesVNA}.
Suppose $\cH$ is a Hilbert space.
\begin{defn}
The $n$-th exterior power of $\cH$ is $\Lambda^n\cH=p_n\bigotimes^n \cH$, where $p_n$ is the projection given by
$$
p_n(\xi_1\otimes\cdots \otimes \xi_n ) = \frac{1}{n!}\sum_{\sigma\in S_n} (-1)^{\sign(\sigma)} \xi_{\sigma(1)}\otimes\cdots\otimes \xi_{\sigma(n)}.
$$
The fermionic Fock space $\cF(\cH)$ is given by $\cF(H)=\bigoplus_{n\geq 0} \Lambda^n \cH$.
Given $\xi_1,\dots,\xi_n\in\cH$, we set
$$
\xi_1\wedge\cdots\wedge \xi_n = \sqrt{n!}\,\, p_n(\xi_1\otimes\cdots\otimes \xi_n).
$$
The inner product on $\cF(\cH)$ is given by
$$
\langle
\eta_1\wedge\cdots \wedge \eta_n,
\xi_1\wedge\cdots\wedge \xi_n
\rangle
=
\det\big((\langle \eta_i,\xi_j\rangle)_{i,j}\big).
$$
For $f\in \cH$, the left creation operator $a(f)$ is given by the unique linear extension of
$$
a(f)(\xi_1\wedge\cdots \wedge \xi_n) = f\wedge \xi_1\wedge\cdots\wedge \xi_n),
$$
and its adjoint is given by
$$
a(f)^*(\xi_1\wedge\cdots \wedge \xi_n) = \sum_{i=1}^n (-1)^{i+1} \langle \xi_i,f\rangle(\xi_1\wedge\cdots \wedge\widehat{\xi_i}\wedge\cdots\wedge \xi_n).
$$
\end{defn}
\begin{rem}
The wave function of several fermions is antisymmetric, so the exterior power $\Lambda^n(\cH)$ describes $n$ identical fermions.
The fermionic Fock space $\cF(\cH)$ is used to treat a countably infinite family of fermions.
\end{rem}
\begin{defn}
If $\cH$ is a complex vector space, the canonical anticommutation relations algebra $\CAR(\cH)$ is the unital $*$-algebra with generators $a(f)$ for $f\in \cH$ subject to the following relations:
\begin{align}
&\text{The map $f\mapsto a(f)$ is linear.}
\tag{CAR1}
\label{rel:CAR1}
\\
&\text{$a(f)a(g)+a(g)a(f)=0$ for all $f,g\in\cH$.}
\tag{CAR2}
\label{rel:CAR2}
\\
&\text{$a(f)a(g)^*+a(g)^*a(f)=\langle f,g\rangle 1_{B(\cH)}$ for all $f,g\in\cH$.}
\tag{CAR3}
\label{rel:CAR3}
\end{align}
\end{defn}
\begin{fact}
There is a unique C* norm and normalized trace on $\CAR(\cH)$.
\end{fact}
\begin{defn}
Given a $u\in U(\cH)$, the Bogoliubov automorphism $\alpha_u$ of $\CAR(\cH)$ is given by $\alpha_u(a(f))=a(uf)$ for all $f\in\cH$.
\end{defn}
\begin{defn}
The gauge-invariant canonical anticommutation relations algebra $\GICAR(\cH)$ is $\CAR(\cH)^{U(1)}$, where $U(1)$ is the scalars acting by Bogoliubov automorphisms on $\CAR(\cH)$.
\end{defn}
\begin{fact}\label{fact:ChooseBasis}
Suppose $\cH$ is separable and infinite dimensional with a fixed choice of orthonormal basis $(\xi_i)_{i\geq 1}$. Let $\cH_n=\spann\{\xi_1,\dots, \xi_n\}$, and define $A_n = \CAR(\cH_n)$ and $G_n=A_n^{U(1)}=\GICAR(\cH_n)$.
We use the abbreviation $a_i=a(\xi_i)$ and $a_i^*=a(\xi_i)^*$ for all $i\geq 1$.
The inclusion $\cH_n\hookrightarrow \cH_{n+1}$ induces inclusions of algebras $A_n\hookrightarrow A_{n+1}$ and $G_n\hookrightarrow G_{n+1}$.
A straightforward calculation (e.g., see \cite[Examples III.5.4-5]{MR1402012}) shows
\begin{align*}
A_n &=C^*\{a_1,\dots, a_n,a_1^*,\dots,a_n^*\}
\cong \bigotimes^n_{k=1} M_2(\mathbb{C})\cong M_{2^n}(\mathbb{C}) \text{ and}\\
G_n &=\spann\set{a_{i_k}\cdots a_{i_1}a_{j_1}^*\cdots a_{j_k}^*}{i_1<\cdots <i_k,\,\, j_1<\cdots <j_k,\,\,k\in\mathbb{N}}
\cong \bigoplus_{k=0}^n M_{n\choose k}(\mathbb{C}),
\end{align*}
where the inclusion $A_n\hookrightarrow A_{n+1}$ is given by $x\mapsto x\otimes 1$, and the Bratteli diagram for the tower $(G_n)_{n\geq 1}$ is Pascal's triangle.
\[
\xymatrix@R=5pt@C=5pt{
&&&&1\ar@{-}[dr]\ar@{-}[dl] \\
&&&1\ar@{-}[dr]\ar@{-}[dl] & & 1\ar@{-}[dr]\ar@{-}[dl] \\
&&1\ar@{-}[dr]\ar@{-}[dl] && 2\ar@{-}[dr]\ar@{-}[dl] && 1\ar@{-}[dr]\ar@{-}[dl] \\
&1\ar@{..}[dr]\ar@{..}[dl]\ &&3\ar@{..}[dr]\ar@{..}[dl]\ && 3\ar@{..}[dr]\ar@{..}[dl]\ && 1\ar@{..}[dr]\ar@{..}[dl]\\
&&&&&&&&&&
}
\]
\end{fact}
The following facts are well known about the GICAR algebras.
We provide a short proof for the convenience of the reader.
Let $\cH$, $(\xi_i)_{i\geq 1}$, $\cH_n$, and $G_n$ be as in Fact \ref{fact:ChooseBasis}.
\begin{thm}\label{thm:GICARRepresentations}
\mbox{}
\begin{enumerate}[(1)]
\item
The representation of $G_n$ on $\Lambda^k \cH_n$ is irreducible.
\item
The left regular representation of $G_n$ breaks up as
$$
G_n \cong \bigoplus_{k=0}^n {n\choose k} \Lambda^k\cH_n.
$$
Thus the complete list of irreducible representations of $G_n$ is $\set{\Lambda^k\cH_n}{k=0,\dots, n}$.
\item
When restricted to the image of $G_{n-1}$ in $G_{n}$,
$$
\Lambda^{k}\cH_{n}\cong \Lambda^{k-1}\cH_{n-1}\oplus \Lambda^{k}\cH_{n-1},
$$
where $\Lambda^{k-1}\cH_{n-1}=(0)$ if $k=0$ and $\Lambda^{k+1}\cH_n=(0)$ if $k=n$.
\end{enumerate}
\end{thm}
\begin{proof}
\mbox{}
\begin{enumerate}[(1)]
\item
This is straightforward.
One can use that any vector of the form $\xi_{i_1}\wedge \cdots \wedge \xi_{i_k}$ with $i_1<\cdots <i_k$ and $k\leq n$ generates $\Lambda^k \cH_n$ as a $G_n$-module.
\item
By Relations \eqref{rel:CAR2}-\eqref{rel:CAR3}, for each $i=1,\dots, n$, the operators $a_ia_i^*$ are commuting projections.
The words in $a_ia_i^*$ and $a_j^*a_j=1-a_ja_j^*$ for which all subscripts $1,\dots, n$ appear give the $2^n$ minimal projections in $G_n$.
Thus there are exactly $n\choose k$ minimal projections in $G_n$ with exactly $k$ projections $a_ja_j^*$ appearing in the word.
For each of these minimal projections $p$, the left $G_n$-module $G_np$ is isomorphic to $\Lambda^k\cH_n$ via the map
$$
p=\prod_{\ell=1}^k a_{i_\ell}a_{i_\ell}^* \prod_{\ell=1}^{n-k} a_{j_\ell}^*a_{j_\ell}\longmapsto \xi_{i_1}\wedge \cdots \wedge \xi_{i_k}.
$$
Since minimal projections in a multi-matrix algebra generate equivalent representations if and only if the projections are equivalent, the result follows from Fact \ref{fact:ChooseBasis}. The last statement follows from the Artin-Wedderburn Theorem.
\item
We have two invariant subspaces of $\Lambda^k\cH_n$ under the action of $G_{n-1}$, namely
$$
\hspace{-.5cm}
\spann\set{\xi_{i_1}\wedge \cdots\wedge \xi_{i_k}}{i_1<\cdots <i_k<n}
\text{ and }
\spann\set{\xi_{i_1}\wedge \cdots\wedge \xi_{i_k}}{i_1<\cdots <i_k=n}.
$$
Both subspaces are irreducible under the action of $G_{n-1}$ as in (1), the latter because $G_{n-1}$ never moves $\xi_n$.
The first is isomorphic to $\Lambda^{k}\cH_n$ since $\xi_n$ never appears, and the second is isomorphic to $\Lambda^{k-1}\cH_n$ since we may ignore the $\xi_n$ which never moves.
If $k=n$, the first subspace is $(0)$, and if $k=1$, $G_{n-1}$ acts as the zero algebra on the second subspace.
\qedhere
\end{enumerate}
\end{proof}
\begin{rems}
\mbox{}
\begin{enumerate}[(1)]
\item
Theorem \ref{thm:GICARRepresentations} part (3) gives another proof that the Bratteli diagram for the tower $(G_n)_{n\geq 0}$ where $G_0=\mathbb{C}$ is Pascal's Triangle.
\item
Remark \ref{rem:DiagrammaticGICARRep} gives a nice diagrammatic description of the representations in Theorem \ref{thm:GICARRepresentations} part (2).
\end{enumerate}
\end{rems}
\subsection{Rook monoids and planar rook algebras}\label{sec:PlanarRookAlgebras}
\begin{defn}
Let $R_n$ be the set of all $n\times n$ zero-one matrices with at most one entry equal to one in each row and column.
Then $R_n$ is a monoid under matrix multiplication.
In \cite{MR1939108}, the author named $R_n$ the \underline{rook monoid}, since the matrices are in one-to-one correspondence with placements of non-attacking rooks on an $n\times n$ chessboard.
\end{defn}
\begin{ex}
The rook monoid $R_2$ consists of the following matrices
$$
R_2=
\left\{
\begin{pmatrix}
0 & 0\\
0 & 0
\end{pmatrix}
\,,\,
\begin{pmatrix}
1 & 0\\
0 & 0
\end{pmatrix}
\,,\,
\begin{pmatrix}
0 & 1\\
0 & 0
\end{pmatrix}
\,,\,
\begin{pmatrix}
0 & 0\\
1 & 0
\end{pmatrix}
\,,\,
\begin{pmatrix}
0 & 0\\
0 & 1
\end{pmatrix}
\,,\,
\begin{pmatrix}
0 & 1\\
1 & 0
\end{pmatrix}
\,,\,
\begin{pmatrix}
1 & 0\\
0 & 1
\end{pmatrix}
\right\}.
$$
\end{ex}
In \cite{MR2541502}, a diagrammatic description of the rook monoid was given as follows. Since each matrix in $R_n$ has at most one 1 in each row and column, we can identify it with a bipartite graph on two rows of $n$ vertices such that each node has degree 0 or 1.
If the $(i,j)$-th entry of $x\in R_n$ is 1, then we connect the $i$-th node on the top row to the $j$-th node on the bottom row.
For example, the matrices in $R_2$ are identified with the following diagrams:
$$
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (1.2,-.2) -- (1.2,1.2) -- (-.2,1.2);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\end{tikzpicture}
\right)
\,,\,
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (1.2,-.2) -- (1.2,1.2) -- (-.2,1.2);
\draw (0,0) -- (0,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\end{tikzpicture}
\right)
\,,\,
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (1.2,-.2) -- (1.2,1.2) -- (-.2,1.2);
\draw (1,0) -- (0,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\end{tikzpicture}
\right)
\,,\,
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (1.2,-.2) -- (1.2,1.2) -- (-.2,1.2);
\draw (0,0) -- (1,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\end{tikzpicture}
\right)
\,,\,
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (1.2,-.2) -- (1.2,1.2) -- (-.2,1.2);
\draw (1,0) -- (1,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\end{tikzpicture}
\right)
\,,\,
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (1.2,-.2) -- (1.2,1.2) -- (-.2,1.2);
\draw (0,0) -- (1,1);
\draw (1,0) -- (0,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\end{tikzpicture}
\right)
\,,\,
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (1.2,-.2) -- (1.2,1.2) -- (-.2,1.2);
\draw (0,0) -- (0,1);
\draw (1,0) -- (1,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\end{tikzpicture}
\right)
$$
Multiplicaiton then corresponds to vertical concatenation of diagrams up to isotopy, where we contract any edge which does not reach the other side, and we delete the middle nodes, e.g.,
\begin{align*}
\begin{pmatrix}
0 & 1 & 0\\
0 & 0 & 0\\
0 & 0 & 1
\end{pmatrix}
\begin{pmatrix}
0 & 0 & 0\\
1 & 0 & 0\\
0 & 0 & 0
\end{pmatrix}
&=
\begin{pmatrix}
1 & 0 & 0\\
0 & 0 & 0\\
0 & 0 & 0
\end{pmatrix}
\\&\updownarrow\\
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (2.2,-.2) -- (2.2,1.2) -- (-.2,1.2);
\draw (0,1) -- (1,0);
\draw (2,0) -- (2,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (2,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\filldraw (2,1) circle (.1cm);
\end{tikzpicture}
\right)
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (2.2,-.2) -- (2.2,1.2) -- (-.2,1.2);
\draw (0,0) -- (1,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (2,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\filldraw (2,1) circle (.1cm);
\end{tikzpicture}
\right)
&=
\left(
\begin{tikzpicture}[baseline=.5cm, scale=.6]
\clip (-.2,-.2) -- (2.2,-.2) -- (2.2,2.2) -- (-.2,2.2);
\draw (0,0) -- (1,1);
\draw (0,2) -- (1,1);
\draw (2,1) -- (2,2);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (2,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\filldraw (2,1) circle (.1cm);
\filldraw (0,2) circle (.1cm);
\filldraw (1,2) circle (.1cm);
\filldraw (2,2) circle (.1cm);
\end{tikzpicture}
\right)
=
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (2.2,-.2) -- (2.2,1.2) -- (-.2,1.2);
\draw (0,0) -- (0,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (2,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\filldraw (2,1) circle (.1cm);
\end{tikzpicture}
\right),
\end{align*}
and the adjoint corresponds to vertical reflection
$$
\begin{pmatrix}
0 & 1\\
0 & 0
\end{pmatrix}
^*
=
\begin{pmatrix}
0 & 0\\
1 & 0
\end{pmatrix}
\longleftrightarrow
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (1.2,-.2) -- (1.2,1.2) -- (-.2,1.2);
\draw (0,1) -- (1,0);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\end{tikzpicture}
\right)
^*
=
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (1.2,-.2) -- (1.2,1.2) -- (-.2,1.2);
\draw (0,0) -- (1,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\end{tikzpicture}
\right).
$$
\begin{defn}
The \underline{planar rook monoid} \cite{MR2541502} $P_n$ consists of the subset of $R_n$ for which the corresponding graphs are planar.
For example,
$$
P_2 =
\left\{
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (1.2,-.2) -- (1.2,1.2) -- (-.2,1.2);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\end{tikzpicture}
\right)
\,,\,
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (1.2,-.2) -- (1.2,1.2) -- (-.2,1.2);
\draw (0,0) -- (0,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\end{tikzpicture}
\right)
\,,\,
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (1.2,-.2) -- (1.2,1.2) -- (-.2,1.2);
\draw (1,0) -- (0,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\end{tikzpicture}
\right)
\,,\,
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (1.2,-.2) -- (1.2,1.2) -- (-.2,1.2);
\draw (0,0) -- (1,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\end{tikzpicture}
\right)
\,,\,
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (1.2,-.2) -- (1.2,1.2) -- (-.2,1.2);
\draw (1,0) -- (1,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\end{tikzpicture}
\right)
\,,\,
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (1.2,-.2) -- (1.2,1.2) -- (-.2,1.2);
\draw (0,0) -- (0,1);
\draw (1,0) -- (1,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\end{tikzpicture}
\right)
\right\}
$$
The \underline{planar rook algebra} $\mathbb{C} P_n$ is the complex $*$-algebra spanned by $P_n$.
\end{defn}
\begin{fact}
The representation theory of $\mathbb{C} P_n$ was classified in \cite{MR2541502}.
Moreover, it was shown that $\mathbb{C} P_n \cong \bigoplus_{k=0}^n M_{n\choose k}(\mathbb{C})$, and the Bratteli diagram for the tower of algebras $(\mathbb{C} P_n)_{n\geq 0}$ is Pascals' Triangle, where the unital inclusion $\mathbb{C} P_n \hookrightarrow \mathbb{C} P_{n+1}$ is given by adding a through string on the right:
$$
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (1.2,-.2) -- (1.2,1.2) -- (-.2,1.2);
\draw (0,0) -- (1,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\end{tikzpicture}
\right)
\longmapsto
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (2.2,-.2) -- (2.2,1.2) -- (-.2,1.2);
\draw (0,0) -- (1,1);
\draw (2,0) -- (2,1);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (2,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\filldraw (2,1) circle (.1cm);
\end{tikzpicture}
\right)
$$
Hence the tower $(G_n)_{n\geq 0}$ is isomorphic to the tower $(\mathbb{C} P_n)_{n\geq 0}$.
\end{fact}
We give an independently found short proof of the isomorphism of towers in Proposition \ref{prop:Pascal} using a notational trick due to Bigelow.
After we establish that the towers are isomorphic, we immediately get the representation theory of the $\mathbb{C} P_n$ from the well-known representation theory of the GICAR algebras given in Theorem \ref{thm:GICARRepresentations}.
In Remark \ref{rem:DiagrammaticGICARRep}, we give a diagrammatic description of these representations in the spirit of Graham and Lehrer's cellular algebras \cite{MR1376244}.
\begin{rem}
We discovered these diagrams in a completely different way.
The Temperley-Lieb diagrams in $TL_{2n}(\delta)$ with only shaded caps and cups are in one-to-one correspondence with the diagrams in $P_n$. One sees this by contracting the cups and caps to nodes and contracting shaded regions to lines as in Figure \ref{fig:ShadedTLDiagrams}. To make the multiplication agree on the nose, one must include a factor of $\delta$ for each maxima in the Temperley-Lieb diagram. (Note that the number of maxima must equal the number of minima).
\begin{figure}[!ht]
$$
\begin{tikzpicture}[rectangular]
\clip (2.5,1.1) --(-1.1,1.1) -- (-1.1,-1.1) -- (2.5,-1.1);
\filldraw[shaded] (1.8,-1)--(.8,1) -- (1.2,1) -- (2.2,-1);
\filldraw[shaded] (0,1)--(0,-1) -- (.4,-1) -- (.4,1);
\filldraw[shaded] (.8,-1) arc (180:0:.3cm);
\filldraw[shaded] (-.8,-1) arc (180:0:.3cm);
\filldraw[shaded] (-.8,1) arc (-180:0:.3cm);
\filldraw[shaded] (1.6,1) arc (-180:0:.3cm);
\draw [very thick] (2.4,1) --(-1,1) -- (-1,-1) -- (2.4,-1)--(2.4,1);
\end{tikzpicture}
\longleftrightarrow
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (3.2,-.2) -- (3.2,1.2) -- (-.2,1.2);
\draw (1,0) -- (1,1);
\draw (2,1) -- (3,0);
\filldraw (0,0) circle (.1cm);
\filldraw (1,0) circle (.1cm);
\filldraw (2,0) circle (.1cm);
\filldraw (3,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\filldraw (2,1) circle (.1cm);
\filldraw (3,1) circle (.1cm);
\end{tikzpicture}
\right)$$
\caption{TL diagrams with only shaded cups/caps and planar rook diagrams}
\label{fig:ShadedTLDiagrams}
\end{figure}
Note that this is the same map $G_n \to TL_{2n}(\delta)$ found by Evans and Connes \cite[Theorem 4.3]{MR990778} without the use of Kauffman diagrams \cite{MR899057}.
They showed this map is injective regardless of $\delta$ by verifying the minimal projections in $G_n$ (see Proposition \ref{prop:Pascal}) map to nonzero orthogonal projections in $TL_{2n}(\delta)$.
We will show a modification of this map works for infinite index subfactors (see Theorem \ref{thm:GICARinQ}).
\end{rem}
\section{Annular and rectangular GICAR categories}\label{sec:GICARCategories}
Just as the Temperley-Lieb algebras can be thought of as a category, so can the planar rook algebras.
We discuss two realizations of this category, which we show are equivalent:
a diagrammatic category, which we call the rectangular planar rook category,
and an abstract category via generators and relations, which we call the rectangular GICAR category.
We also have the notion of the annular planar rook and GICAR categories, which we show are equivalent.
Along the way, we will take a brief detour to discuss a diagrammatic representation of the GICAR algebra.
\begin{nota}
We denote categories using the sans-serif font $\sA\sB\sC\dots$
Given a category $\sC$, we write $X,Y\in\sC$ to denote $X,Y$ are objects in $\sC$, and we write $\sC(X,Y)$ for the space of morphisms from $X$ to $Y$.
If the objects in $\sC$ are symbols of the form $[n]$ for $n\geq 0$, we simply write $\sC(m,n)$ for $\sC([m],[n])$.
We further simplify notation by writing $\sC_n$ for $\sC(n,n)$.
\end{nota}
\subsection{Annular and rectangular planar rook categories}\label{sec:PlanarRookCategories}
\begin{defn}
The \underline{annular planar rook category} $\sA\sP$ is the following small involutive category:
\itt{Objects} $[n]$ for $n\geq 0$, and
\itt{Morphisms}
$\sA\sP(m,n)$ is all $\mathbb{C}$-linear combinations of isotopy classes of tangles on annuli with decoration as follows.
\begin{itemize}
\item There are $m$ marked points on the inner boundary, called the inner points, and $n$ marked points on the external boundary, called the outer points.
\item Each marked point is connected to exactly one string.
Each string is connected to at least one and at most two marked boundary points.
Strings do not intersect.
No string may connect two inner points or two outer points.
Hence there are three possibilities for strings:
\begin{enumerate}[(1)]
\item
A \underline{through string} connects an inner and an outer boundary point.
\item
A \underline{cap} is a string that only connects to an inner boundary point.
\item
A \underline{cup} is a string that only connects to an outer boundary point.
\end{enumerate}
We draw a dark circle on the end of a non-through string to denote that that end does not attach to another boundary point.
\item There is a distinguished interval on each boundary disk, marked by a $\star$.
\end{itemize}
\itt{Composition} Composition is the $\mathbb{C}$-linear extension of insertion of annuli, making sure the boundary points line up, as do the distinguished intervals.
When we get a floating string (a string connected to no boundary points), we just remove it.
$$
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (0:0.5cm) -- (0:1.5cm);
\draw (60:0.5cm) -- (60:1.5cm);
\draw (180:0.5cm) .. controls ++(180:.4cm) and ++(-60:.4cm) .. (120:1.5cm);
\draw (120:0.5cm) -- (120:.85cm);
\filldraw (120:.85cm) circle (.05cm);
\draw (180:1.15cm) -- (180:1.5cm);
\filldraw (180:1.15cm) circle (.05cm);
\draw (240:0.5cm) .. controls ++(240:.4cm) and ++(120:.4cm) .. (300:1.5cm);
\draw (300:0.5cm) -- (300:.85cm);
\filldraw (300:.85cm) circle (.05cm);
\draw (240:1.15cm) -- (240:1.5cm);
\filldraw (240:1.15cm) circle (.05cm);
\node at (30:.75cm) {$\star$};
\node at (210:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
\,\circ\,
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (0:0.5cm) -- (0:1.5cm);
\draw (60:0.5cm) -- (60:1.5cm);
\draw (180:0.5cm) .. controls ++(180:.4cm) and ++(-60:.4cm) .. (120:1.5cm);
\draw (180:1.15cm) -- (180:1.5cm);
\filldraw (180:1.15cm) circle (.05cm);
\draw (240:0.5cm) .. controls ++(240:.4cm) and ++(120:.4cm) .. (300:1.5cm);
\draw (240:1.15cm) -- (240:1.5cm);
\filldraw (240:1.15cm) circle (.05cm);
\node at (30:.75cm) {$\star$};
\node at (210:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
=
\begin{tikzpicture}[annular]
\clip (0,0) circle (2.6cm);
\draw (0:0.5cm) -- (0:1.5cm) .. controls ++(0:.8cm) and ++(120:.8cm) .. (300:2.5cm);
\draw (0:2.15cm) -- (0:2.5cm);
\filldraw (0:2.15cm) circle (.05cm);
\draw (60:0.5cm) -- (60:1.5cm) .. controls ++(60:.8cm) and ++(-60:.8cm) .. (120:2.5cm);
\draw (60:2.15cm) -- (60:2.5cm);
\filldraw (60:2.15cm) circle (.05cm);
\draw (180:0.5cm) .. controls ++(180:.4cm) and ++(-60:.4cm) .. (120:1.5cm) -- (120:1.85cm);
\filldraw (120:1.85cm) circle (.05cm);
\draw (180:1.15cm) -- (180:2.5cm);
\filldraw (180:1.15cm) circle (.05cm);
\draw (240:0.5cm) .. controls ++(240:.4cm) and ++(120:.4cm) .. (300:1.5cm) -- (300:1.85cm);
\filldraw (300:1.85cm) circle (.05cm);
\draw (240:1.15cm) -- (240:2.5cm);
\filldraw (240:1.15cm) circle (.05cm);
\node at (30:.75cm) {$\star$};
\node at (210:1.25cm) {$\star$};
\node at (210:1.75cm) {$\star$};
\node at (30:2.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\draw [very thick] (0,0) circle (2.5cm);
\end{tikzpicture}
=
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (0:1.15cm) -- (0:1.5cm);
\filldraw (0:1.15cm) circle (.05cm);
\draw (60:1.15cm) -- (60:1.5cm);
\filldraw (60:1.15cm) circle (.05cm);
\draw (60:0.5cm) .. controls ++(60:.4cm) and ++(-60:.4cm) .. (120:1.5cm);
\draw (180:0.5cm) -- (180:.85cm);
\filldraw (180:.85cm) circle (.05cm);
\draw (180:1.15cm) -- (180:1.5cm);
\filldraw (180:1.15cm) circle (.05cm);
\draw (0:0.5cm) .. controls ++(0:.4cm) and ++(120:.4cm) .. (300:1.5cm);
\draw (240:0.5cm) -- (240:.85cm);
\filldraw (240:.85cm) circle (.05cm);
\draw (240:1.15cm) -- (240:1.5cm);
\filldraw (240:1.15cm) circle (.05cm);
\node at (30:.75cm) {$\star$};
\node at (30:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
$$
\itt{Adjoint} The adjoint is the conjugate-linear extension of flipping the tangle inside out.
$$
\left(
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (0:0.5cm) -- (0:1.5cm);
\draw (60:0.5cm) -- (60:1.5cm);
\draw (180:0.5cm) .. controls ++(180:.4cm) and ++(-60:.4cm) .. (120:1.5cm);
\draw (120:0.5cm) -- (120:.85cm);
\filldraw (120:.85cm) circle (.05cm);
\draw (180:1.15cm) -- (180:1.5cm);
\filldraw (180:1.15cm) circle (.05cm);
\draw (240:0.5cm) .. controls ++(240:.4cm) and ++(120:.4cm) .. (300:1.5cm);
\draw (300:0.5cm) -- (300:.85cm);
\filldraw (300:.85cm) circle (.05cm);
\draw (240:1.15cm) -- (240:1.5cm);
\filldraw (240:1.15cm) circle (.05cm);
\node at (30:.75cm) {$\star$};
\node at (210:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
\right)^*
=
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (0:0.5cm) -- (0:1.5cm);
\draw (60:0.5cm) -- (60:1.5cm);
\draw (120:0.5cm) .. controls ++(120:.4cm) and ++(0:.4cm) .. (180:1.5cm);
\draw (120:1.15cm) -- (120:1.5cm);
\filldraw (120:1.15cm) circle (.05cm);
\draw (180:.5cm) -- (180:.85cm);
\filldraw (180:.85cm) circle (.05cm);
\draw (300:0.5cm) .. controls ++(300:.4cm) and ++(60:.4cm) .. (240:1.5cm);
\draw (300:1.1cm) -- (300:1.5cm);
\filldraw (300:1.15cm) circle (.05cm);
\draw (240:.5cm) -- (240:.85cm);
\filldraw (240:.85cm) circle (.05cm);
\node at (30:1.25cm) {$\star$};
\node at (210:0.75cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
$$
\end{defn}
\begin{rem}
Unlike the annular Temperley-Lieb category, $\sA\sP_n$ is finite dimensional for all $n\geq 0$ due to the absence of non-contractible closed loops.
\end{rem}
We now count the number of annular tangles in $\sA\sP(m,n)$.
\begin{defn}
Let $N(m,n;k)$ be the number of annular tangles in $\sA\sP_n$ with $m$ inner points, $n$ outer points, and $k$ through strings.
Let $N(m,n)=\sum_{k=0}^{\min\{m,n\}} N(m,n;k)$.
\end{defn}
\begin{rem}\label{rem:CountTangles}
Note that
\begin{itemize}
\item
$N(m,n;k)=N(n,m;k)$ for all $m,n,k$, so we only need to count when $k\leq m\leq n$,
\item
$N(0,n)=1$ for al $n\geq 0$, and
\item
$N(m,n;0)=1$ for all $m,n\geq 0$.
\end{itemize}
\end{rem}
\begin{lem}\label{lem:CountTangles}
If $1\leq k\leq m\leq n$, then $\displaystyle N(m,n;k)=m{n\choose k}{m-1\choose k-1}$.
\end{lem}
\begin{proof}
Draw an annulus with $m$ inner points and $n$ outer points.
Fix the outer $\star$.
There are exactly $n\choose k$ ways to connect $k$ through strings to the $n$ outer points.
Equivalently, there are exactly $n\choose k$ choices for the cup positions.
Let us examine one of these choices more closely.
Look at the first through string connected to an outer point counting clockwise from the outer $\star$.
Follow the through string inward, and put the inner star on the interval to the left of this inner point, so that the region meeting the outer $\star$ meets the inner $\star$.
We now see there are exactly $m-1 \choose k-1$ ways to connect the remaining through strings to the remaining inner points.
Equivalently, there are exactly $m-1\choose k-1$ choices of the cap positions.
Fix such a choice of cap position, which we will call the tangle's initial cap position.
Note that given an annular tangle in $\sA\sP(m,n)$, the cup positions and the initial cap positions only depend on the outer $\star$.
Hence we get $m$ distinct tangles as we shift the inner $\star$ clockwise.
In summary, for each of the $n\choose k$ cup positions and for each of the resulting $m-1 \choose k-1$ initial cap positions, there are $m$ distinct tangles.
The formula follows.
\end{proof}
Remark \ref{rem:CountTangles} and Lemma \ref{lem:CountTangles} now prove the following.
\begin{prop}\label{prop:CountTangles}
If $1\leq m\leq n$, then
$\displaystyle
N(m,n)=1+\sum_{k=1}^m m{n\choose k}{m-1\choose k-1}.
$
\end{prop}
We will determine the algebra structure of $\sA\sP_n$ at the end of this subsection in Proposition \ref{prop:AnnularAlgebraStructure}.
We first treat the rectangular planar rook category as a warmup.
\begin{defn}
The \underline{rectangular planar rook category} $\sR\sP$ is the subcategory of $\sA\sP$ such that $\sR\sP(m,n)$ is the $\mathbb{C}$-linear combinations of diagrams in $\sA\sP(m,n)$ such that the region meeting the internal $\star$ also meets the external $\star$.
For example,
$$
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (0:1.15cm) -- (0:1.5cm);
\filldraw (0:1.15cm) circle (.05cm);
\draw (60:1.15cm) -- (60:1.5cm);
\filldraw (60:1.15cm) circle (.05cm);
\draw (60:0.5cm) .. controls ++(60:.4cm) and ++(-60:.4cm) .. (120:1.5cm);
\draw (180:0.5cm) -- (180:.85cm);
\filldraw (180:.85cm) circle (.05cm);
\draw (180:1.15cm) -- (180:1.5cm);
\filldraw (180:1.15cm) circle (.05cm);
\draw (0:0.5cm) .. controls ++(0:.4cm) and ++(120:.4cm) .. (300:1.5cm);
\draw (240:0.5cm) -- (240:.85cm);
\filldraw (240:.85cm) circle (.05cm);
\draw (240:1.15cm) -- (240:1.5cm);
\filldraw (240:1.15cm) circle (.05cm);
\node at (30:.75cm) {$\star$};
\node at (30:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
\in\sR\sP(4,6),
\text{ but }
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (0:0.5cm) -- (0:1.5cm);
\draw (60:0.5cm) -- (60:1.5cm);
\draw (180:0.5cm) .. controls ++(180:.4cm) and ++(-60:.4cm) .. (120:1.5cm);
\draw (120:0.5cm) -- (120:.85cm);
\filldraw (120:.85cm) circle (.05cm);
\draw (180:1.15cm) -- (180:1.5cm);
\filldraw (180:1.15cm) circle (.05cm);
\draw (240:0.5cm) .. controls ++(240:.4cm) and ++(120:.4cm) .. (300:1.5cm);
\draw (300:0.5cm) -- (300:.85cm);
\filldraw (300:.85cm) circle (.05cm);
\draw (240:1.15cm) -- (240:1.5cm);
\filldraw (240:1.15cm) circle (.05cm);
\node at (30:.75cm) {$\star$};
\node at (210:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
\notin\sR\sP(6,6).
$$
Each such morphism can be represented by a rectangular tangle rather than an annular tangle as follows.
First, cut along a path from the internal $\star$ to the external $\star$ which does not meet any strings.
Second, isotope the resulting diagram into a rectangle with lower and upper boundary points so that the inner boundary points of the annulus are now the lower boundary points of the rectangle, and the outer boundary points of the annulus are now the upper boundary points of the rectangle.
\begin{figure}[!ht]
$$
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (0:1.15cm) -- (0:1.5cm);
\filldraw (0:1.15cm) circle (.05cm);
\draw (60:1.15cm) -- (60:1.5cm);
\filldraw (60:1.15cm) circle (.05cm);
\draw (60:0.5cm) .. controls ++(60:.4cm) and ++(-60:.4cm) .. (120:1.5cm);
\draw (180:0.5cm) -- (180:.85cm);
\filldraw (180:.85cm) circle (.05cm);
\draw (180:1.15cm) -- (180:1.5cm);
\filldraw (180:1.15cm) circle (.05cm);
\draw (0:0.5cm) .. controls ++(0:.4cm) and ++(120:.4cm) .. (300:1.5cm);
\draw (240:0.5cm) -- (240:.85cm);
\filldraw (240:.85cm) circle (.05cm);
\draw (240:1.15cm) -- (240:1.5cm);
\filldraw (240:1.15cm) circle (.05cm);
\node at (30:.75cm) {$\star$};
\node at (30:1.25cm) {$\star$};
\draw[very thick, red] (30:.5cm) -- (30:1.5cm);
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
\longmapsto
\begin{tikzpicture}[rectangular]
\clip (1.9,1.1) --(-1.1,1.1) -- (-1.1,-1.1) -- (1.9,-1.1);
\draw (-.6,1)--(-.6,.35);
\filldraw (-.6,.35) circle (.05cm);
\draw (.2,1)--(.2,.35);
\filldraw (.2,.35) circle (.05cm);
\draw (.6,1)--(.6,.35);
\filldraw (.6,.35) circle (.05cm);
\draw (1.4,1)--(1.4,.35);
\filldraw (1.4,.35) circle (.05cm);
\draw (-.2,1)--(-.6,-1);
\draw (1,1)--(1.4,-1);
\draw (.1,-1)--(.1,-.35);
\filldraw (.1,-.35) circle (.05cm);
\draw (.7,-1)--(.7,-.35);
\filldraw (.7,-.35) circle (.05cm);
\draw [very thick] (1.8,1) --(-1,1) -- (-1,-1) -- (1.8,-1)--(1.8,1);
\end{tikzpicture}
$$
\caption{Cutting an annulus to get a rectangle}
\label{fig:CuttingAndGluing}
\end{figure}
Composition of annuli then corresponds to stacking rectangles,
$$
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\BrokenBottomString{(-.4,0)}
\String{(-.2,0)}
\BrokenString{(0,0)}
\String{(.2,0)}
\end{tikzpicture}
\,\circ\,
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\BrokenTopString{(-.4,0)}
\String{(-.2,0)}
\PartialIsoStar{(0,0)}
\end{tikzpicture}
\,=\,
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.8}{-.2}{-.4}{}
\draw[very thick] (-.6,0) -- (.4,0);
\BrokenBottomString{(-.4,.4)}
\String{(-.2,.4)}
\BrokenString{(0,.4)}
\String{(.2,.4)}
\BrokenTopString{(-.4,-.4)}
\String{(-.2,-.4)}
\PartialIsoStar{(0,-.4)}
\end{tikzpicture}
\,=\,
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\String{(-.4,0)}
\BrokenString{(-.2,0)}
\BrokenString{(0,0)}
\end{tikzpicture}
$$
and the adjoint operation corresponds to vertical flipping of rectangles.
$$
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\BrokenTopString{(-.4,0)}
\String{(-.2,0)}
\PartialIsoStar{(0,0)}
\end{tikzpicture}^{\,*}
\,=\,
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\BrokenBottomString{(-.4,0)}
\String{(-.2,0)}
\PartialIso{(0,0)}
\end{tikzpicture}
$$
Viewing morphisms in $\sR\sP$ as rectangular tangles, we can endow $\sR\sP$ with a tensor structure.
The tensor product of objects is $[m]\otimes [n]=[m+n]$, and the tensor product of morphisms is the $\mathbb{C}$-linear extension of horizontal join.
$$
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\BrokenBottomString{(-.4,0)}
\String{(-.2,0)}
\BrokenString{(0,0)}
\String{(.2,0)}
\end{tikzpicture}
\,\otimes\,
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\BrokenTopString{(-.4,0)}
\String{(-.2,0)}
\PartialIsoStar{(0,0)}
\end{tikzpicture}
\,=\,
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{.8}{}
\BrokenBottomString{(-.4,0)}
\String{(-.2,0)}
\BrokenString{(0,0)}
\String{(.2,0)}
\BrokenTopString{(.4,0)}
\String{(.6,0)}
\PartialIsoStar{(.8,0)}
\end{tikzpicture}
$$
\end{defn}
\begin{rem}
Obviously $\sR\sP_n\cong \mathbb{C} P_n$ by contracting cups and caps and trading the external boundary for nodes at the marked boundary points.
$$
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\BrokenTopString{(-.4,0)}
\String{(-.2,0)}
\PartialIsoStar{(0,0)}
\end{tikzpicture}
\longleftrightarrow
\left(
\begin{tikzpicture}[baseline=.2cm, scale=.6]
\clip (-.2,-.2) -- (3.2,-.2) -- (3.2,1.2) -- (-.2,1.2);
\draw (1,0) -- (1,1);
\draw (2,1) -- (3,0);
\filldraw (1,0) circle (.1cm);
\filldraw (2,0) circle (.1cm);
\filldraw (3,0) circle (.1cm);
\filldraw (0,1) circle (.1cm);
\filldraw (1,1) circle (.1cm);
\filldraw (2,1) circle (.1cm);
\filldraw (3,1) circle (.1cm);
\end{tikzpicture}
\right)
$$
We use different diagrams for morphisms in $\sR\sP_n$ than the usual diagrams for $P_n$ to utilize a notational trick of Bigelow (see Definition \ref{defn:DottedStrand}).
\end{rem}
\begin{prop}\label{prop:Pascal}
As a complex $*$-algebra, $\sR\sP_n\cong \bigoplus_{k=0}^n M_{n\choose k}(\mathbb{C})$. Moreover, the Bratteli diagram for the tower of finite dimensional algebras $(\sR\sP_n)_{n\geq 0}$ under the right inclusion (adding a through string to the right) is given by Pascal's Triangle.
\end{prop}
\begin{proof}
Let $0\leq k\leq n$. There are exactly $n\choose k$ diagrams with exactly $k$ through strings in $\sR\sP_n$.
Hence $\dim_\mathbb{C}(\sR\sP_n)=\sum_{k=0}^n {n \choose k}=2^n$.
However, it is important to note that diagrams with exactly $k$ through strings are not orthogonal to diagrams with exactly $j$ through strings for $j\neq k$.
To fix this problem, we make the following definition.
\begin{defn}\label{defn:DottedStrand}
As in \cite[Section 3]{MR2925434}, we let the dotted strand denote the following morphism in $\sR\sP_1$:
$$
\eOnePerp=\identity-\eOne\,.
$$
We then have the following relations in $\sR\sP$:
\begin{align}
\confetti &= 1
\tag{$\sR\sP$1}\label{rel:RP1}\\
\doubleDotStrand &= \eOnePerp
\tag{$\sR\sP$2}\label{rel:RP2}\\
\doubleDotEnd &= 0.
\tag{$\sR\sP$3}\label{rel:RP3}
\end{align}
\end{defn}
\begin{rem}
Under the identification of these diagrams with those in the Temperley-Lieb category with only shaded cups and caps in Figure \ref{fig:ShadedTLDiagrams}, the broken strand corresponds to the Jones projection $e_1$, and the dotted strand corresponds to the Jones-Wenzl projection $\jw{2}=1-e_1$.
\end{rem}
With the use of the dotted strand, we find $2^n$ minimal orthogonal projections in $\sR\sP_n$ given by the simple tensors composed entirely of
$$
\eOne\, \text{ and }\,\eOnePerp\,.
$$
The diagrams with exactly $k$ through strings, all of which are dotted, span a full matrix algebra $M_{n\choose k}(\mathbb{C})$.
Hence $\sR\sP_n$ is isomorphic to the orthogonal direct sum $ \bigoplus_{k=0}^n M_{n\choose k} (\mathbb{C})$.
We now look at the right inclusion $\sR\sP_n\hookrightarrow \sR\sP_{n+1}$ given by adding a string to the right. Since
$$
\identity = \eOne + \eOnePerp\,,
$$
we see that the right inclusion maps each minimal projection in $\sR\sP_n$ to the sum of exactly two minimal projections in $\sR\sP_{n+1}$.
More precisely, each minimal projection in the simple summand corresponding to $M_{n\choose k}(\mathbb{C})$ maps to the sum of two minimal projections, one in $M_{{n+1}\choose k}(\mathbb{C})$, and one in $M_{{n+1} \choose {k+1}}(\mathbb{C})$. Hence the Bratteli diagram is as claimed.
\end{proof}
\begin{rem}
We give an explicit formula for the resulting isomorphism of towers $(G_n)_{n\geq 0}\cong (\sR\sP_n)_{n\geq 0}$ in Theorem \ref{thm:IsoOfTowers}.
\end{rem}
\begin{cor}
$\displaystyle\dim(\sR\sP_n)=\sum_{k=0}^n {n\choose k}^2$.
\end{cor}
We now determine the algebra structure of $\sA\sP_n$.
The dotted strand will be of great use to us.
\begin{prop}\label{prop:AnnularAlgebraStructure}
As a complex $*$-algebra,
$\displaystyle\sA\sP_n \cong \mathbb{C}\oplus \bigoplus_{k=1}^n k M_{n\choose k}(\mathbb{C})$.
\end{prop}
\begin{rem}\label{rem:BinomialIdentity}
Note that we have the identity
$$
n {n\choose k}{n-1\choose k-1}=k{n\choose k}^2,
$$
so the formula in Proposition \ref{prop:AnnularAlgebraStructure} is consistent with Proposition \ref{prop:CountTangles}.
\end{rem}
\begin{proof}[Proof of Proposition \ref{prop:AnnularAlgebraStructure}]
Using Bigelow's dotted strand, consider the annular tangles which give minimal projections in $\sR\sP_n$ under the cutting operation in Figure \ref{fig:CuttingAndGluing}.
These annular tangles are orthogonal projections in $\sA\sP_n$, but the only one that remains minimal is the one with only broken strings and no dotted through strings.
Now given a projection $p_k$ with $k$ dotted through strings, the $k$ powers of the 1-click rotation tangle $\rho$ (see Figure \ref{fig:rotation}) can be compressed by $p_k$ to give $k$ distinct tangles $p_k \rho^i p_k$ for $i=1\dots, k$.
\begin{figure}[!ht]
$$
\rho =
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (216:0.5cm) .. controls ++(216:.6cm) and ++(324:.4cm) .. (144:1.5cm);
\draw (144:0.5cm) .. controls ++(144:.6cm) and ++(252:.4cm) .. (72:1.5cm);
\draw (72:0.5cm) .. controls ++(72:.6cm) and ++(180:.4cm) .. (0:1.5cm);
\draw (0:0.5cm) .. controls ++(0:.6cm) and ++(108:.4cm) .. (-72:1.5cm);
\draw (-72:0.5cm) .. controls ++(-72:.6cm) and ++(36:.4cm) .. (-144:1.5cm);
\node at (158:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
$$
\caption{The one click rotation $\rho\in \sA\sP_5$}
\label{fig:rotation}
\end{figure}
Now if $\omega_k$ is a $k$-th root of unity, we get a projection
$$
p_k^\omega=\frac{1}{k}\sum_{i=0}^{k-1} \omega_k^{-i} p_k \rho^i p_k
$$
which lives under $p_k$. Distinct $\omega$ give distinct projections, since $\rho(p_k^\omega)=\omega p_k^\omega$, so $p_k$ splits into $k$ non-zero orthogonal projections.
Now using the usual partial isometries from $\sR\sP_n$ in annular form, we see that splitting each projection with $k\geq 2$ dotted through strings into $k$ orthogonal summands also splits the corresponding copy of $M_{n\choose k}(\mathbb{C})$ in $\sR\sP_n$ into $k$ copies of $M_{n\choose k}(\mathbb{C})$ in $\sA\sP_n$, which results in the claimed decomposition.
By Remark \ref{rem:BinomialIdentity}, we must have all the minimal projections, since the dimension count agrees with Proposition \ref{prop:CountTangles}.
\end{proof}
\subsection{Annular and rectangular GICAR categories}
\begin{defn}
The \underline{annular GICAR category} $\sA\sG$ is the following small involutive category.
\itt{Objects} symbols $[n]$ for $n\geq 0$.
\itt{Morphisms}
The morphisms of $\sA\sG$ are $\mathbb{C}$-linear combinations of the words $*$-generated by the maps
\begin{align*}
\alpha_i : [n] &\longrightarrow [n+1] \text{ for }i=1,\dots, n+1\text{ and }n\geq 0\\
\alpha_i^* : [n] &\longrightarrow [n-1] \text{ for }i=1,\dots, n\text{ and }n\geq 1\\
\tau :[n] & \longrightarrow [n] \text{ for }n\geq 0
\end{align*}
subject to the relations
\begin{align}
\alpha_i\alpha_{j-1}&=\alpha_j\alpha_i
\text{ and }
\alpha_i^*\alpha_j^* = \alpha_{j-1}^*\alpha_i^*
\text{ for all }i<j
\label{rel:AG1}
\tag{$\sA\sG1$}
\\
\alpha_{i}^*\alpha_{j}
&=
\begin{cases}
\alpha_{j+1}\alpha_{i}^* &\text{if }i<j\\
\id_{[n]} & \text{if } i=j\\
\alpha_j \alpha_{i+1}^* & \text{if } i>j
\end{cases}
\label{rel:AG2}
\tag{$\sA\sG2$}
\\
\tau^*&=\tau^{-1}\text{ and }\tau^n=\id_{[n]}
\label{rel:AG3}
\tag{$\sA\sG3$}
\\
\alpha_i \tau &= \tau \alpha_{i-1}\text{ and }\alpha_i^*\tau = \tau\alpha_{i-1}^* \text{ for all }i=2,\dots, n.
\label{rel:AG4}
\tag{$\sA\sG4$}
\end{align}
\itt{Composition} The composition in $\sA\sG$ is the concatenation of words.
\itt{Adjoint} The adjoint of a word $w=\ell_1\dots \ell_n$ where the letters $\ell_k\in\{\alpha_i,\alpha_j^*,\tau \}$ is given by $w^*=\ell_n^*\cdots \ell_1^*$.
\end{defn}
\begin{rem}
$\sA\sG$ is the full subcategory of ${\sf a\Delta}$ in \cite{MR2903179} generated by $\tau,\alpha_i,\alpha_j^*$, after replacing the $\alpha_i$'s by the $\beta_{2i}$'s appearing there.
\end{rem}
\begin{prop}\label{prop:ExtraRelation}
The additional relations
\begin{equation}
\alpha_1=\tau\alpha_n
\text{ and }
\alpha_1^*\tau = \alpha_n^*
\label{rel:AG5}
\tag{$\sA\sG5$}
\end{equation}
hold in $\sA\sG$.
\end{prop}
\begin{proof}
By Relations \eqref{rel:AG3} and \eqref{rel:AG4}, we have
$$
\alpha_1 = \tau^{n+1}\alpha_1 = \tau^n \alpha_2\tau=\cdots=\tau\alpha_n\tau^n=\tau\alpha_n.
$$
Now take adjoints. (Note there is a typo in the proof of this relation in \cite[Proposition 3.6.(1)]{MR2903179}).
\end{proof}
\begin{rem}
By the results of \cite{MR2903179}, there is a $*$-equivalence of categories $\sA\sP\cong \sA\sG$.
We provide a short proof of this fact for the convenience of the reader along the same line of reasoning as \cite{MR2903179}.
\end{rem}
\begin{prop}\label{prop:Standard}
Suppose $w\in \sA\sG(m,n)$ is a word in the $\alpha_i,\tau,\alpha_j^*$.
Then $w$ can be written uniquely in the standard form
$$
w=\alpha_{i_k}\cdots \alpha_{i_1} \tau^r \alpha_{j_1}^*\cdots \alpha_{j_\ell}^*
$$
where $i_1< \cdots <i_k$, $0\leq r < m-k$, and $j_1<\cdots < j_\ell$.
In particular, the words in standard form give bases for $\sG(m,n)$, and thus $\dim_\mathbb{C}(\sA\sG(m,n))<\infty$ for all $m,n$.
\end{prop}
\begin{proof}
Using Relations \eqref{rel:AG1}-\eqref{rel:AG5}, first push all the $\alpha_i$'s all the way to the left and all the $\alpha_j^*$'s all the way the right, leaving the $\tau$'s in the middle. Then use Relation \eqref{rel:AG1} to put the $\alpha_i$'s in decreasing order and the $\alpha_j^*$'s in increasing order. Finally, use Relation \eqref{rel:AG3} to reduce the number of $\tau$'s in the middle.
\end{proof}
\begin{thm}\label{thm:Equivalence}
There is a $*$-equivalence of categories $\sA\sP\cong \sA\sG$.
\end{thm}
\begin{proof}
We construct a $*$-functor $\Psi:\sA\sG\to \sA\sP$.
First, define $\Psi([n])=[n]$.
Next, we define $\Psi$ on the morphisms $\alpha_i,\tau,\alpha_j^*$.
\begin{itemize}
\item
$\alpha_i\in \sA\sG(n,n+1)$ maps to the tangle in $\sR\sP$ with $n$ inner points, $n+1$ outer points, a cap attached to outer boundary point $i$, and all other boundary points are connected by through strings so that the region meeting the internal $\star$ also meets the external $\star$.
$$
\underset{\Psi(\alpha_1\in \sA\sG(4,5))}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (120:1.15cm) -- (120:1.5cm);
\filldraw (120:1.15cm) circle (.05cm);
\draw (60:0.5cm) -- (60:1.5cm);
\draw (0:0.5cm) -- (0:1.5cm);
\draw (-60:0.5cm) -- (-60:1.5cm);
\draw (-120:0.5cm) -- (-120:1.5cm);
\node at (180:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
\,,\,
\underset{\Psi(\alpha_2\in \sA\sG(4,5))}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (120:0.5cm) -- (120:1.5cm);
\draw (60:1.15cm) -- (60:1.5cm);
\filldraw (60:1.15cm) circle (.05cm);
\draw (0:0.5cm) -- (0:1.5cm);
\draw (-60:0.5cm) -- (-60:1.5cm);
\draw (-120:0.5cm) -- (-120:1.5cm);
\node at (180:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
\,,\,
\underset{\Psi(\alpha_3\in \sA\sG(4,5))}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (120:0.5cm) -- (120:1.5cm);
\draw (60:0.5cm) -- (60:1.5cm);
\draw (0:1.15cm) -- (0:1.5cm);
\filldraw (0:1.15cm) circle (.05cm);
\draw (-60:0.5cm) -- (-60:1.5cm);
\draw (-120:0.5cm) -- (-120:1.5cm);
\node at (180:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
\,,\,
\underset{\Psi(\alpha_4\in \sA\sG(4,5))}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (120:0.5cm) -- (120:1.5cm);
\draw (60:0.5cm) -- (60:1.5cm);
\draw (0:0.5cm) -- (0:1.5cm);
\draw (-60:1.15cm) -- (-60:1.5cm);
\filldraw (-60:1.15cm) circle (.05cm);
\draw (-120:0.5cm) -- (-120:1.5cm);
\node at (180:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
\,,\,
\underset{\Psi(\alpha_5\in \sA\sG(4,5))}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (120:0.5cm) -- (120:1.5cm);
\draw (60:0.5cm) -- (60:1.5cm);
\draw (0:0.5cm) -- (0:1.5cm);
\draw (-60:0.5cm) -- (-60:1.5cm);
\draw (-120:1.15cm) -- (-120:1.5cm);
\filldraw (-120:1.15cm) circle (.05cm);
\node at (180:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
$$
\item
$\alpha_j^*\in\sA\sG(n,n-1)$ maps to $\Psi(\alpha_j) ^*\in\sA\sP(n-1,n)$.
$$
\underset{\Psi(\alpha_1^*\in \sA\sG(5,4))}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (120:0.5cm) -- (120:.85cm);
\filldraw (120:.85cm) circle (.05cm);
\draw (60:0.5cm) -- (60:1.5cm);
\draw (0:0.5cm) -- (0:1.5cm);
\draw (-60:0.5cm) -- (-60:1.5cm);
\draw (-120:0.5cm) -- (-120:1.5cm);
\node at (180:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
\,,\,
\underset{\Psi(\alpha_2^*\in \sA\sG(5,4))}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (120:0.5cm) -- (120:1.5cm);
\draw (60:0.5cm) -- (60:.85cm);
\filldraw (60:.85cm) circle (.05cm);
\draw (0:0.5cm) -- (0:1.5cm);
\draw (-60:0.5cm) -- (-60:1.5cm);
\draw (-120:0.5cm) -- (-120:1.5cm);
\node at (180:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
\,,\,
\underset{\Psi(\alpha_3^*\in \sA\sG(5,4))}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (120:0.5cm) -- (120:1.5cm);
\draw (60:0.5cm) -- (60:1.5cm);
\draw (0:0.5cm) -- (0:.85cm);
\filldraw (0:.85cm) circle (.05cm);
\draw (-60:0.5cm) -- (-60:1.5cm);
\draw (-120:0.5cm) -- (-120:1.5cm);
\node at (180:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
\,,\,
\underset{\Psi(\alpha_4^*\in \sA\sG(5,4))}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (120:0.5cm) -- (120:1.5cm);
\draw (60:0.5cm) -- (60:1.5cm);
\draw (0:0.5cm) -- (0:1.5cm);
\draw (-60:0.5cm) -- (-60:.85cm);
\filldraw (-60:.85cm) circle (.05cm);
\draw (-120:0.5cm) -- (-120:1.5cm);
\node at (180:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
\,,\,
\underset{\Psi(\alpha_5^*\in \sA\sG(5,4))}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (120:0.5cm) -- (120:1.5cm);
\draw (60:0.5cm) -- (60:1.5cm);
\draw (0:0.5cm) -- (0:1.5cm);
\draw (-60:0.5cm) -- (-60:1.5cm);
\draw (-120:0.5cm) -- (-120:.85cm);
\filldraw (-120:.85cm) circle (.05cm);
\node at (180:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
$$
\item
$\tau\in \sA\sG_n$ maps to the counter-clockwise one click rotation in $\sA\sP_n$,
and $\tau^*=\tau^{-1}$ maps to the clockwise one click rotation.
$$
\underset{\Psi(\tau\in \sA\sG(5,5))}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (216:0.5cm) .. controls ++(216:.6cm) and ++(324:.4cm) .. (144:1.5cm);
\draw (144:0.5cm) .. controls ++(144:.6cm) and ++(252:.4cm) .. (72:1.5cm);
\draw (72:0.5cm) .. controls ++(72:.6cm) and ++(180:.4cm) .. (0:1.5cm);
\draw (0:0.5cm) .. controls ++(0:.6cm) and ++(108:.4cm) .. (-72:1.5cm);
\draw (-72:0.5cm) .. controls ++(-72:.6cm) and ++(36:.4cm) .. (-144:1.5cm);
\node at (158:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
\text{ and }
\underset{\Psi(\tau^*\in \sA\sG(5,5))}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (72:0.5cm) .. controls ++(72:.6cm) and ++(324:.4cm) .. (144:1.5cm);
\draw (0:0.5cm) .. controls ++(0:.6cm) and ++(252:.4cm) .. (72:1.5cm);
\draw (-72:0.5cm) .. controls ++(-72:.6cm) and ++(180:.4cm) .. (0:1.5cm);
\draw (-144:0.5cm) .. controls ++(-144:.6cm) and ++(108:.4cm) .. (-72:1.5cm);
\draw (144:0.5cm) .. controls ++(144:.6cm) and ++(36:.4cm) .. (-144:1.5cm);
\node at (-158:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
$$
\end{itemize}
One sees that these tangles satisfy the relations of $\sA\sG$ by drawing the appropriate diagrams.
We define $\Psi^{-1}$ by its $\mathbb{C}$-linear extension on tangles from $\sA\sP$.
Given an annular tangle $T\in \sA\sP(m,n)$, there is a unique $r$ satisfying
$$
0\leq r< \#(\text{through strings of }T)
$$
which is the number of through strings to the left of the inner $\star$ that one must cross to get to the region which meets the outer $\star$.
We call this $r$ the relative star position of $T$.
Now, $\Psi^{-1}(T)\in \sA\sG(m,n)$ is the word in standard form
$$
\Psi^{-1}(T)=\alpha_{i_k}\cdots \alpha_{i_1}\tau^r\alpha_{j_1}^*\cdots \alpha_{j_\ell}^*
$$
where $j_1<\cdots < j_\ell$ are the positions of the caps of $T$, $r$ is the relative star position, and $i_1<\cdots < i_k$ are the positions of the cups of $T$.
That $\Psi^{-1}\circ \Psi= \id_{\sA\sG}$ and $\Psi\circ\Psi^{-1}=\id_{\sA\sP}$ follows immediately.
\end{proof}
\begin{defn}
The \underline{rectangular GICAR category} $\sR\sG$ is the subcategory of $\sA\sG$ such that $\sR\sG(m,n)$ consists of all $\mathbb{C}$-linear combinations of words $w$ on $\alpha_i,\tau,\alpha_j^*$ such that in the standard form of $w$ afforded by Proposition \ref{prop:Standard}, no $\tau$ appears, i.e., $r=0$.
\end{defn}
\begin{thm}\label{thm:RectangularEquivalence}
There is a $*$-equivalence of categories $\sR\sP\cong \sR\sG$.
\end{thm}
\begin{proof}
First, it is clear the functor $\Psi$ constructed in Theorem \ref{thm:Equivalence} restricts to a $*$-equivalence $\sR\sP\cong \sR\sG$.
In fact,
\begin{itemize}
\item
$\alpha_i\in\sR\sG(n,n+1)$ maps to the diagram with $n$ lower boundary points, $n+1$ upper boundary points, a cup attached to lower boundary point $i$, and all other boundary points connected by undotted through strings.
$$
\hspace{-.5cm}
\underset{\Psi(\alpha_1\in\sR\sG(3,4))}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\BrokenTopString{(-.4,0)}
\String{(-.2,0)}
\String{(0,0)}
\String{(.2,0)}
\end{tikzpicture}
}
\,,\,
\underset{\Psi(\alpha_2\in\sR\sG(3,4))}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\String{(-.4,0)}
\BrokenTopString{(-.2,0)}
\String{(0,0)}
\String{(.2,0)}
\end{tikzpicture}
}
\,,\,
\underset{\Psi(\alpha_3\in\sR\sG(3,4))}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\String{(-.4,0)}
\String{(-.2,0)}
\BrokenTopString{(0,0)}
\String{(.2,0)}
\end{tikzpicture}
}
\,,\,
\underset{\Psi(\alpha_4\in\sR\sG(3,4))}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\String{(-.4,0)}
\String{(-.2,0)}
\String{(0,0)}
\BrokenTopString{(.2,0)}
\end{tikzpicture}
}
$$
\item
$\alpha_j^*\in\sR\sG(n,n-1)$ maps to $\Psi(\alpha_j)^*\in \sR\sP(n-1,n)$.
\begin{align*}
&
\underset{\Psi(\alpha_1^*\in\sR\sG(4,3))}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\BrokenBottomString{(-.4,0)}
\String{(-.2,0)}
\String{(0,0)}
\String{(.2,0)}
\end{tikzpicture}
}
\,,\,
\underset{\Psi(\alpha_2^*\in\sR\sG(4,3))}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\String{(-.4,0)}
\BrokenBottomString{(-.2,0)}
\String{(0,0)}
\String{(.2,0)}
\end{tikzpicture}
}
\,,\,
\underset{\Psi(\alpha_3^*\in\sR\sG(4,3))}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\String{(-.4,0)}
\String{(-.2,0)}
\BrokenBottomString{(0,0)}
\String{(.2,0)}
\end{tikzpicture}
}
\,,\,
\underset{\Psi(\alpha_4^*\in\sR\sG(4,3))}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{0}{}
\String{(-.4,0)}
\String{(-.2,0)}
\String{(0,0)}
\BrokenBottomString{(.2,0)}
\end{tikzpicture}
}
\qedhere
\end{align*}
\end{itemize}
\end{proof}
\begin{rem}
We can now pull back the tensor structure on $\sR\sP$ to get a tensor structure on $\sR\sG$.
The tensor product of objects is $[m]\otimes [n]=[m+n]$, and the tensor product of morphisms in standard form
\begin{align*}
\varphi&=
\alpha_{i_k}\cdots \alpha_{i_1}\alpha_{j_1}^*\cdots \alpha_{j_\ell}^*
\in \sR\sG(m_1,n_1)
\text{ and }
\\
\psi&=
\alpha_{i_{k'}'}\cdots a_{i_{1}'}\alpha_{j_{1}'}^*\cdots \alpha_{j_{\ell'}'}^*
\in \sR\sG(m_2,n_2)
\end{align*}
is given by
$$
\varphi\otimes \psi=
\alpha_{i_{k'}'+n_1}\cdots \alpha_{i_{1}'+n_1}\alpha_{i_k}\cdots \alpha_{i_1}
\alpha_{j_1}^*\cdots \alpha_{j_\ell}^*\alpha_{j_{1}'+m_1}^*\cdots \alpha_{j_{\ell'}'+m_1}^*
\in \sR\sG(m_1+m_2,n_1+n_2).
$$
With this tensor structure, the functor $\Psi$ in Theorem \ref{thm:RectangularEquivalence} is a $*,\otimes$-functor.
\end{rem}
\section{Representation theory of the GICAR categories}\label{sec:Representations}
We now compute the representation theory of the GICAR categories $\sR\sG\cong \sR\sP$ and $\sA\sG\cong \sA\sP$ in the spirit of Graham and Lehrer's theory of cellular algebras \cite{MR1376244,MR1659204}.
\subsection{A diagrammatic representation of the GICAR algebra}\label{sec:DigrammaticFermions}
We first give a diagrammatic description of the GICAR algebra acting on fermionic Fock space using the diagrams from $\sR\sP$ so that we may use Bigelow's dotted strand (Definition \ref{defn:DottedStrand}).
These diagrams implicitly appear in \cite{MR990778}, while our diagrams for fermionic Fock space arise from the cellular structure in the spirit of \cite{MR1376244,MR1659204}.
Our diagrammatic representation of $\GICAR(\cH)$ relies on choosing an orthonormal basis of $\cH$.
This should neither surprise nor worry the reader for the following reason.
Recall from Fact \ref{fact:ChooseBasis} that we must choose an orthonormal basis to show that $\CAR(\cH)\cong \bigotimes^\infty M_2(\mathbb{C})$ and to show that the Bratteli diagram for the tower of algebras $(G_n=A_n^{U(1)})_{n\geq 0}$ is given by Pascal's Triangle.
Since our diagrams in $\sR\sP$ are equivalent to those for the $\mathbb{C} P_n$'s, we are relying on the AF structure of $\GICAR(\cH)$, which relies on the choice of basis.
Suppose $\cH$ is a separable, infinite dimensional Hilbert space with a fixed choice of orthonormal basis $(\xi_i)_{i\geq 1}$.
Define $\cH_n = \spann\{\xi_1,\dots, \xi_n\}$.
We use the abbreviations $a_i=a(\xi_i)$ and $a_j^*=a(\xi_j)^*$.
Recall the following facts about fermionic Fock space and the GICAR algebra.
\begin{facts}
\mbox{}
\begin{enumerate}[(1)]
\item
An orthonormal basis of $\cF(\cH)$ is given by symbols of the form $\xi_{i_1}\wedge \cdots \wedge \xi_{i_n}$ for $i_1<i_2<\cdots < i_n$ together with the vacuum vector $\Omega$.
\item
By \cite[Lemma 2.2]{MR990778}, $\GICAR(\cH_n)$ has a presentation as a $*$-algebra with generators $f_i$ for $i=1,\dots, n$ and $u_i$ for $i=1,\dots, n-1$ and relations:
\begin{align}
&\text{$f_i=f_i^*=f_i^2$}
\tag{G1}
\label{rel:G1}
\\
&\text{$[f_i,f_j]=0$ if $j\neq i$}
\tag{G2}
\label{rel:G2}
\\
&\text{$[u_i,f_j]=0$ if $j\neq i,i+1$}
\tag{G3}
\label{rel:G3}
\\
&\text{$[u_i,u_j]=[u_i,u_j^*]=0$ if $|i-j|\geq 2$}
\tag{G4}
\label{rel:G4}
\\
&\text{$u_i^*u_i=f_{i+1}(1-f_i)$ and $u_iu_i^*=f_i(1-f_{i+1})$}
\tag{G5}
\label{rel:G5}
\end{align}
The isomorphism is given by $f_i\mapsto a_i^*a_i$ and $u_i\mapsto a_i^*a_{i+1}$.
\end{enumerate}
\end{facts}
We now construct a diagrammatic Hilbert space $\cD_n$ on which we represent $\sR\sP_n$.
We then give a spatial isomorphism $\Theta_n: \cF(\cH_n)\to \cD_n$ and a $*$-isomorphism of algebras $\theta_n: G_n\to \sR\sP_n$ which intertwines the actions, i.e., for all $\eta,\zeta\in \cF(\cH_n)$ and all $x,y\in G_n$, we have
\begin{align}
\langle \eta,\zeta\rangle_{\cF(\cH_n)}
&=
\langle \Theta_n(\eta),\Theta_n(\zeta)\rangle_{\cD_n},
\label{rel:D1}
\tag{D1}
\\
\theta_n(xy^*)
&=
\theta_n(x)\theta_n(y)^*,
\text{ and}
\label{rel:D2}
\tag{D2}
\\
\Theta_n(x\eta)
&=
\theta_n(x)\Theta_n(\eta)
\label{rel:D3}
\tag{D3}
\end{align}
\begin{defn}
For $n\geq 0$, let $\cD_n$ be the complex span of diagrams in $\sR\sP$ with $n$ top boundary points, at most $n$ bottom boundary points, and no caps, such that all through strings are dotted.
Define an inner product on $\cD_n$ by declaring the diagrammatic basis of $\sR\sP$ to be orthonormal.
Let $\sR\sP_n$ act on $\cD_n$ by the usual composition of maps in $\sR\sP$.
\end{defn}
\begin{defn}
We define $\Theta_n: \cF(\cH_n)\to \cD_n$ as follows.
Let $\Theta_n(\xi_{i_1}\wedge\cdots \wedge \xi_{i_k})$ where $i_1<\cdots <i_k$ and $k\leq n$ be the diagram with $n$ upper boundary points, $n-k$ lower boundary points, cups in the $i_\ell$-th positions for all $\ell=1,\dots, k$, and all other strings are dotted through strings.
For example, when $k\leq 2$, we have
\begin{align*}
\Omega&\longmapsto
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{.2}{.2}{}
\DottedString{(-.4,0)}
\node at (.05,0) {$\cdots$};
\DottedString{(.4,0)}
\node at (.4,.6) {\scriptsize{$n$}};
\end{tikzpicture}\,,
\\
\xi_i&\longmapsto
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{.2}{1.4}{}
\DottedString{(-.4,0)}
\node at (.05,0) {$\cdots$};
\DottedString{(.4,0)}
\coordinate (a) at (.6,0);
\BrokenTopString{(a)};
\node at ($(a)+(0,.6)$) {\scriptsize{$i$}};
\DottedString{($(a)+(.2,0)$)}
\node at ($(a)+(.65,0)$) {$\cdots$};
\DottedString{($(a)+(1,0)$)}
\node at ($(a)+(1,.6)$) {\scriptsize{$n$}};
\end{tikzpicture}\,,\text{ and}
\\
\xi_i\wedge\xi_j &\longmapsto
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{.2}{2.6}{}
\DottedString{(-.4,0)}
\node at (.05,0) {$\cdots$};
\DottedString{(.4,0)}
\coordinate (a) at (.6,0);
\BrokenTopString{(a)};
\node at ($(a)+(0,.6)$) {\scriptsize{$i$}};
\DottedString{($(a)+(.2,0)$)}
\node at ($(a)+(.65,0)$) {$\cdots$};
\DottedString{($(a)+(1,0)$)}
\coordinate (b) at ($(a)+(1.2,0)$);
\BrokenTopString{(b)};
\node at ($(b)+(0,.6)$) {\scriptsize{$j$}};
\DottedString{($(b)+(.2,0)$)}
\node at ($(b)+(.65,0)$) {$\cdots$};
\DottedString{($(b)+(1,0)$)}
\node at ($(b)+(1,.6)$) {\scriptsize{$n$}};
\end{tikzpicture}
\text{ for }i<j.
\end{align*}
\end{defn}
\begin{thm}\label{thm:IsoOfTowers}
Define the map $\theta_n : G_n \to \sR\sP_n$ by
$$
a_i^*a_i \mapsto
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{.3}{1.1}{}
\String{(-.5,0)};
\node at (-.15,0) {$\cdots$};
\coordinate (a) at (.4,0);
\String{($(a)+(-.2,0)$)};
\DottedString{(a)};
\node at ($(a)+(0,.6)$) {\scriptsize{$i$}};
\String{($(a)+(.2,0)$)};
\node at ($(a)+(.55,0)$) {$\cdots$};
\String{($(a)+(.9,0)$)};
\node at ($(a)+(.9,.6)$) {\scriptsize{$n$}};
\end{tikzpicture}\,,
\text{ and }
a_i^*a_{i+1} \mapsto
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{.3}{1.3}{}
\String{(-.5,0)};
\node at (-.15,0) {$\cdots$};
\coordinate (a) at (.4,0);
\String{($(a)+(-.2,0)$)};
\PartialDottedIsoStar{(a)}
\node at ($(a)+(0,.6)$) {\scriptsize{$i$}};
\String{($(a)+(.4,0)$)};
\node at ($(a)+(.75,0)$) {$\cdots$};
\String{($(a)+(1.1,0)$)};
\node at ($(a)+(1.1,.6)$) {\scriptsize{$n$}};
\end{tikzpicture}\,.
$$
Then $\Theta_n$ and $\theta_n$ satisfy Equations \eqref{rel:D1}-\eqref{rel:D3}.
\end{thm}
\begin{proof}
First, Equation \eqref{rel:D1} holds since $\Theta_n$ is a spatial isomorphism which maps an orthonormal basis to an orthonormal basis.
Second, by verifying that Relations \eqref{rel:G1}-\eqref{rel:G5} hold for $f_i=\theta_n(a_i^*a_i)$ and $u_i=\theta_n(a_i^*a_{i+1})$, we see that $\theta_n$ is an injective $*$-algebra homomorphism, since $\dim(G_n)=\dim(\sR\sP_n)$ by Proposition \ref{prop:Pascal}.
It remains to show Equation \eqref{rel:D3}.
Since Relations \eqref{rel:D1}-\eqref{rel:D2} hold, it suffices to verify Equation \eqref{rel:D3} when $x$ is one of $a_i^*a_i,a_i^*a_{i+1}$, and $\eta$ is of the form $\xi_{i_1}\wedge\cdots \wedge \xi_{i_k}$ where $i_1<\cdots <i_k$.
Clearly
\begin{align*}
a_i^*a_i (\xi_{i_1}\wedge\cdots \wedge \xi_{i_k})
&=
\begin{cases}
0 & \text{if }i\in \{i_1,\dots, i_k\}\\
\xi_{i_1}\wedge\cdots \wedge \xi_{i_k} & \text{otherwise, and}
\end{cases}
\\
a_i^*a_{i+1} (\xi_{i_1}\wedge\cdots \wedge \xi_{i_k})
&=
\begin{cases}
0 & \text{if }i\notin \{i_1,\dots, i_k\}\text{ and }\\ &i+1\in \{i_1,\dots, i_k\}\\
\xi_{i_1}\wedge\cdots\wedge \widehat{\xi_i}\wedge \xi_{i+1}\wedge\cdots \wedge \xi_{i_k} & \text{otherwise.}
\end{cases}
\end{align*}
The rest is straightforward using Relations \eqref{rel:RP1}-\eqref{rel:RP3}.
\end{proof}
\begin{rem}
There is an easy graphical description of the inner product.
If $\eta,\zeta$ are single diagrams in $\cD_n$, we look at the composite $\zeta^*\eta$ in $\sR\sP$, which is well-defined since $\eta,\zeta$ both have $n$ top boundary points.
We then use Relations \eqref{rel:RP1}-\eqref{rel:RP3}.
If we get a non-zero composite, then $\zeta^*\eta$ consists of only dotted through strings.
Thus it must be the case that $\eta=\zeta$, since the cap positions must agree, and $\langle \eta,\zeta\rangle = 1$.
We leave it to the reader to extend this discussion to a formal definition of the graphical inner product.
\end{rem}
\begin{rem}
The following diagram commutes where the maps $G_n\to G_{n+1}$ and $\cH_n\to \cH_{n+1}$ are the usual inclusions, the map $\sR\sP_n\to \sR\sP_{n+1}$ is the right inclusion, and the map $\cD_n\to \cD_{n+1}$ is adding a dotted through string on the right.
\[
\xymatrix{
G_n\ar[rr]\ar[dd]^{\theta_n} \ar[dr]&& G_{n+1}\ar[dd]^(.6){\theta_{n+1}}\ar[dr]
\\
& B(\cF(\cH_n))\ar[dd]^(.6){\Ad(\Theta_{n})}\ar[rr] && B(\cF(\cH_{n+1}))\ar[dd]^{\Ad(\Theta_{n+1})}
\\
\sR\sP_n\ar[rr]\ar[dr] && \sR\sP_{n+1}\ar[dr]
\\
& B(\cD_n)\ar[rr] && B(\cD_{n+1})
}
\]
\end{rem}
\begin{ex}
We have the following diagrammatic representation of $G_3$ on $\cF(\cH_3)$.
\begin{align*}
\sR\sP(3,3)&=C^*\left\{
\underset{a_1a_1^*}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{0}{0}{}
\BrokenString{(-.2,0)}
\String{(0,0)}
\String{(.2,0)}
\end{tikzpicture}
}
\,,\,
\underset{a_1a_2^*}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{0}{0}{}
\PartialDottedIso{(-.2,0)}
\String{(.2,0)}
\end{tikzpicture}
}
\,,\,
\underset{a_2a_1^*}{
\begin{tikzpicture}[baseline=-.1cm,]
\nbox{unshaded}{(0,0)}{.4}{0}{0}{}
\PartialDottedIsoStar{(-.2,0)}
\String{(.2,0)}
\end{tikzpicture}
}
\,,\,
\underset{a_2a_2^*}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{0}{0}{}
\String{(-.2,0)}
\BrokenString{(0,0)}
\String{(.2,0)}
\end{tikzpicture}
}
\,,\,
\underset{a_2a_3^*}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{0}{0}{}
\PartialDottedIso{(0,0)}
\String{(-.2,0)}
\end{tikzpicture}
}
\,,\,
\underset{a_3a_2^*}{
\begin{tikzpicture}[baseline=-.1cm,]
\nbox{unshaded}{(0,0)}{.4}{0}{0}{}
\PartialDottedIsoStar{(0,0)}
\String{(-.2,0)}
\end{tikzpicture}
}
\,,\,
\underset{a_3a_3^*}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{0}{0}{}
\String{(-.2,0)}
\String{(0,0)}
\BrokenString{(.2,0)}
\end{tikzpicture}
}
\right\}
\text{ and }
\\
\cD_3&=
\spann
\left\{
\underset{\Omega}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{0}{0}{}
\DottedString{(-.2,0)};
\DottedString{(0,0)};
\DottedString{(.2,0)};
\end{tikzpicture}
}
\,,\,
\underset{\xi_1}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{0}{0}{}
\BrokenTopString{(-.2,0)};
\DottedString{(0,0)};
\DottedString{(.2,0)};
\end{tikzpicture}
}
\,,\,
\underset{\xi_2}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{0}{0}{}
\DottedString{(-.2,0)};
\BrokenTopString{(0,0)};
\DottedString{(.2,0)};
\end{tikzpicture}
}
\,,\,
\underset{\xi_3}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{0}{0}{}
\DottedString{(-.2,0)};
\DottedString{(0,0)};
\BrokenTopString{(.2,0)};
\end{tikzpicture}
}
\,,\,
\underset{\xi_1\wedge \xi_2}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{0}{0}{}
\BrokenTopString{(-.2,0)};
\BrokenTopString{(0,0)};
\DottedString{(.2,0)};
\end{tikzpicture}
}
\,,\,
\underset{\xi_1\wedge \xi_3}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{0}{0}{}
\BrokenTopString{(-.2,0)};
\DottedString{(0,0)};
\BrokenTopString{(.2,0)};
\end{tikzpicture}
}
\,,\,
\underset{\xi_2\wedge \xi_3}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{0}{0}{}
\DottedString{(-.2,0)};
\BrokenTopString{(0,0)};
\BrokenTopString{(.2,0)};
\end{tikzpicture}
}
\,,\,
\underset{\xi_1\wedge \xi_2\wedge \xi_3}{
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{0}{0}{}
\BrokenTopString{(-.2,0)};
\BrokenTopString{(0,0)};
\BrokenTopString{(.2,0)};
\end{tikzpicture}
}
\right\}.
\end{align*}
However, note that we need a linear combination of diagrams to represent $a_1a_3^*$ and $a_3a_1^*$.
Using Relations \eqref{rel:CAR2}-\eqref{rel:CAR3}, we have
\begin{align*}
a_1a_3^*
&=a_1a_2^*a_2a_3^*+a_1a_2a_2^*a_3^*
=a_1a_2^*a_2a_3^*+a_2a_1a_3^*a_2^*
=a_1a_2^*a_2a_3^*-a_2a_3^*a_1a_2^*
\\
&\longmapsto
\begin{tikzpicture}[baseline=-.1cm]
\draw[very thick] (-.4,-.8) rectangle (.4,.8);
\draw[very thick] (-.4,0) -- (.4,0);
\PartialDottedIso{(-.2,.4)}
\String{(.2,.4)}
\PartialDottedIso{(0,-.4)}
\String{(-.2,-.4)}
\end{tikzpicture}
-
\begin{tikzpicture}[baseline=-.1cm]
\draw[very thick] (-.4,-.8) rectangle (.4,.8);
\draw[very thick] (-.4,0) -- (.4,0);
\PartialDottedIso{(-.2,-.4)}
\String{(.2,-.4)}
\PartialDottedIso{(0,.4)}
\String{(-.2,.4)}
\end{tikzpicture}
=
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{.2}{.2}{}
\draw (-.4,-.4) -- (0,.4);
\draw (0,-.4) -- (.4,.4);
\draw (-.4,.4) -- (-.4,.15);
\draw (.4,-.4) -- (.4,-.15);
\filldraw (-.4,.15) circle (.05cm);
\filldraw (.4,-.15) circle (.05cm);
\filldraw (-.2,0) circle (.05cm);
\filldraw (.2,0) circle (.05cm);
\end{tikzpicture}
-
\begin{tikzpicture}[baseline=-.1cm]
\nbox{unshaded}{(0,0)}{.4}{.2}{.2}{}
\draw (-.4,-.4) -- (.4,.4);
\draw (-.4,.4) -- (-.4,.15);
\draw (.4,-.4) -- (.4,-.15);
\draw (.1,-.4) -- (.1,-.15);
\draw (-.1,.4) -- (-.1,.15);
\filldraw (-.4,.15) circle (.05cm);
\filldraw (.4,-.15) circle (.05cm);
\filldraw (.1,-.15) circle (.05cm);
\filldraw (-.1,.15) circle (.05cm);
\filldraw (0,0) circle (.05cm);
\end{tikzpicture}
\,.
\end{align*}
\end{ex}
\begin{rem}
At this point, we do not know if it is possible to use these diagrams or similar ones to represent $\CAR(\cH)$ on $\cF(\cH)$.
One might be tempted to define $a_1$ by a cup in the first position on the top.
In order for $a_1$ to kill $\xi_1$, we must connect a dotted string to the first lower point.
However, if this dotted string connected to upper point $i$, the image of $a_1$ would never contain an antisymmetric tensor containing a $\xi_i$, which is absurd.
\end{rem}
\begin{rem}\label{rem:DiagrammaticGICARRep}
We now get a nice diagrammatic description of the representations given in Theorem \ref{thm:GICARRepresentations} part (2).
There are exactly $n\choose k$ minimal projections in $\sR\sP_n$ with exactly $k$ through strings, all of which are dotted. Each of these minimal projections $p$ generates a copy of $\Lambda^{n-k}\cH_n$ as $\sR\sP_np$, where we ignore the bottom of the broken strands. For example, if we have the minimal projection with $k$ broken strings on the left and $n-k$ doted through strings on the right,
$$
\sR\sP_n
\left(
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.6}{.8}{};
\BrokenString{(-.8,0)};
\node at (-.35,0) {$\dots$};
\BrokenString{(0,0)};
\DottedString{(.2,0)};
\node at (.65,0) {$\dots$};
\DottedString{(1,0)};
\end{tikzpicture}
\right)
\cong
\sR\sP_n
\left(
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.6}{.8}{};
\BrokenTopString{(-.8,0)};
\node at (-.35,0) {$\dots$};
\BrokenTopString{(0,0)};
\DottedString{(.2,0)};
\node at (.65,0) {$\dots$};
\DottedString{(1,0)};
\end{tikzpicture}
\right)
\cong
G_n(
\xi_1\wedge \cdots \wedge \xi_k)
=
\Lambda^k \cH_n.
$$
It would also be interesting to describe these representations using a unitary version of Graham and Lehrer's theory of cellular algebras \cite{MR1376244,MR1659204}.
\end{rem}
\subsection{Representations of small involutive categories}
We now discuss the representation theory of small involutive categories, where for simplicity, we work with finite dimensional complex Hilbert spaces.
Our treatment is along the lines of \cite[Sections 2-3]{MR1929335}. We provide proofs for completeness.
\begin{defn}
Suppose $\sC$ is a small involutive category whose hom spaces are finite dimensional complex vector spaces.
A \underline{Hilbert $\sC$-module} is a $*$-functor $V: \sC\to {\sf Hilb}$, the category of finite dimensional Hilbert spaces and linear maps.
We denote $V(n)$ by $V_n$ for $n\in\sC$, and we just use the notation $c$ for $V(c)\in B(V_m,V_n)$ when $c\in\sC(m,n)$.
This means for all $\xi\in V_m$ and $\eta\in V_n$, we have
$$
\langle c \xi, \eta\rangle_{V_n}=\langle \xi, c^* \eta\rangle_{V_m}.
$$
\end{defn}
\begin{rem}
Sometimes one defines a $\sC$-module as a functor originating in $\sC^{\sf op}$, e.g., simplicial sets are functors ${\sf \Delta^{op}}\to {\sf Set}$.
Since $\sC$ is involutive, $\sC\cong \sC^{\sf op}$ via the involution, so we just use covariant functors.
\end{rem}
\begin{defn}
We call a Hilbert $\sC$-module
\begin{itemize}
\item
\underline{indecomposable} if it cannot be written as the direct sum of two non-trivial orthogonal submodules, or
\item
\underline{irreducible} if it has no proper submodules, i.e., any non-zero element in $V_m$ for any $m$ generates all of $V$.
\end{itemize}
\end{defn}
\begin{lem}\label{lem:IrreducibleAndIndecomposable}
Suppose $V$ is a Hilbert $\sC$-module.
Then $V$ is irreducible if and only if $V$ is indecomposable.
\end{lem}
\begin{proof}
Suppose $V$ is indecomposable, and let $\xi\in V_n$ for some $n\in\sC$.
Define Hilbert $\sC$-modules $W,W^\perp$ by $W_m=\sC(m,n)\xi$ and $W_m^\perp$ is the usual orthogonal complement for all $m\geq 0$.
Then $V=W\oplus W^\perp$, so $W^\perp$ must be the zero module, i.e., $W_m=V_m$ for all $m$.
Thus $V$ is irreducible.
Suppose now that $V$ is irreducible, and suppose $V=W\oplus X$ for orthogonal Hilbert $\sC$-modules $W$ and $X$.
Suppose $\xi\in W_m$ is non-zero for some $m\in\sC$.
Let $\eta\in X_n$.
Then for all $c\in\sC(m,n)$, we have $\langle c\xi,\eta \rangle_{V_n} = 0$, but $\sC(m,n)\xi=V_n$, so $\eta=0$.
Hence $X$ is the zero module, and $V$ is indecomposable.
\end{proof}
\begin{lem}\label{lem:Orthogonal}
Suppose $V$ is a Hilbert $\sC$-module.
Suppose $W_m$ and $X_m$ are orthogonal $\sC(m,m)$-invariant subspaces of $V_m$. Then $\sC(W)$ is orthogonal to $\sC(X)$.
\end{lem}
\begin{proof}
If $\xi=c_1\xi_0$ for some $c_1\in\sC(m,n)$ and $\xi_0\in W_m$, and $\eta=c_2 \eta_0$ for some $c_2\in \sC(m,n)$ and $\eta_0\in X_m$, then
$$
\langle \xi,\eta \rangle_{V_n} = \langle c_1\xi_0, c_2\eta_0\rangle_{V_n} = \langle c_2^*c_1 \xi_0, \eta_0\rangle_{V_m}=0
$$
since $c_2^*c_1\in \sC(m,m)$ and $W_m$ is $\sC(m,m)$-invariant.
\end{proof}
\begin{lem}\label{lem:IrreducibleSubmodule}
Suppose $W\subset V_m$ is an irreducible $\sC(m,m)$-module for some $m$. Then $W_n=\sC(W)_n$ is irreducible for all $n$.
\end{lem}
\begin{proof}
Suppose $\xi\in W_n$ is nonzero, and let $\eta\in W_n$ be another vector.
Write $\xi=c_1 \xi_0$ and $\eta= c_2 \eta_0$ for $\eta_0,\xi_0\in W$ and $c_1,c_2\in\sC(m,n)$.
Then $c_1^*\xi = c_1^*c_1 \xi_0\in W$ is non-zero, so there is a $c_3\in \sC(m,m)$ with $c_3c_1^*\xi=\eta_0$.
Then $c_2c_3c_1^*\xi=\eta$, and $W_n$ is irreducible.
\end{proof}
\begin{assume}\label{assume:Generating}
We now assume that our Hilbert $\sC$-module $V$ satisfies the following generating property:
\begin{itemize}
\item
For any two objects $m,n\in\sC$ such that $V_m,V_n\neq (0)$, the image of $\sC(m,n)$ in $B(V_m,V_n)$ contains a non-zero map.
\end{itemize}
\end{assume}
\begin{lem}\label{lem:AllIrreducible}
The following are equivalent.
\begin{enumerate}[(1)]
\item
$V_m$ is an irreducible $\sC(m,m)$-module for all $m$,
and
\item
$V$ is irreducible.
\end{enumerate}
\end{lem}
\begin{proof}
\itm{(1)\Rightarrow (2)}
Suppose $\xi\in V_m$ and $\eta\in V_n$ are nonzero.
There is a nonzero map $c_2\in\sC(m,n)$ by Assumption \ref{assume:Generating}, so there is an $\zeta\in V_m$ such that $c_2 \zeta\in V_n\setminus\{0\}$.
Since $V_m$ is irreducible, there is a $c_1\in \sC(m,m)$ such that $c_1\xi=\zeta$.
Since $V_n$ is irreducible, there is a $c_3\in \sC(n,n)$ such that $\eta=c_3c_2\zeta=c_3c_2c_1\xi$.
\itm{(2)\Rightarrow (1)}
We prove the contrapositive.
Suppose $V_m$ has a non-zero proper $\sC(m,m)$-module $W_m$ for some $m$.
Since $\sC(m,m)$ acts as a $*$-algebra of bounded operators on $V_m$, $W_m^\perp$ is also a non-zero proper submodule of $V_m$.
Applying $\sC$ to $W_m$ and $W_m^\perp$ yields two proper $\sC$-submodules of $V$ which are orthogonal by Lemma \ref{lem:Orthogonal}.
\end{proof}
\begin{assume}\label{assume:ObjectsAreN}
We now assume that the objects of $\sC$ are the symbols $[n]$ for $n\geq 0$, which come with the usual total order on $\mathbb{N}\cup\{0\}$.
Moreover, we assume that for $m\leq n$, there is a monomorphism in $\sC(m,n)$, so that Assumption \ref{assume:Generating} is satisfied.
\end{assume}
\begin{ex}
The (annular) Temperley-Lieb and the (annular) GICAR categories satisfy Assumption \ref{assume:ObjectsAreN}.
\end{ex}
\begin{defn}
The \underline{weight} $\wt(V)$ of a $\sC$-module $V$ is the least number $n$ such that $V_n$ is non-zero.
Elements of $V_{\wt(V)}$ are called \underline{lowest weight vectors}.
\end{defn}
\begin{prop}\label{prop:IrreducibleDecomposition}
Every Hilbert $\sC$-module has a canonical decomposition as an orthogonal direct sum of irreducible Hilbert $\sC$-modules.
\end{prop}
\begin{proof}
First decompose $V_{\wt(V)}$ into an orthogonal direct sum of irreducible $\sC_{\wt(V)}$-modules.
The direct summands generate irreducible $\sC$-modules by Lemmas \ref{lem:IrreducibleSubmodule} and \ref{lem:AllIrreducible}, all of which are mutually orthogonal by \ref{lem:Orthogonal}.
The orthogonal complement of these $\sC$-modules have higher weight, so we are finished by an induction argument.
\end{proof}
\begin{lem}\label{lem:ExtendMorphism}
If $V,W$ are two Hilbert $\sC$-modules with $V$ irreducible, and $\theta : V_m \to W_m$ is a non-zero homomorphism of $\sC_m$-modules, then $\theta$ extends uniquely to an injective homomorphism $\Theta$ of Hilbert $\sC$-modules.
\end{lem}
\begin{proof}
Suppose $\xi\in V_n$.
Then $\xi=a \eta$ for some $a\in \sC(m,n)$ and $\eta\in V_m$ since $V$ is irreducible.
We claim $\Theta(\xi)=c\theta(\eta)$ gives a well-defined map of $\sC$-modules.
If $\xi=b\eta$ for another $b\in \sC(m,n)$, then for any $c\in \sC(m,n)$ and $\zeta\in V_m$, we have
$$
\langle a\theta(\eta), c\zeta\rangle_{V_n}
= \langle c^*a \theta(\eta),\zeta\rangle
= \langle \theta(c^*b \eta),\zeta\rangle
= \langle c^*b \theta(\eta),\zeta\rangle
= \langle b \theta(\eta),c \zeta\rangle,
$$
so $a\theta(\eta)=b\theta(\eta)$, so $\Theta$ is well-defined.
By construction $\Theta$ is a $\sC$-module map, and it is injective since $V$ is irreducible.
The uniqueness of the extension $\Theta$ of $\theta$ is obvious.
\end{proof}
\subsection{The representation theory of $\sA\sG\cong \sA\sP$}
Since the annular GICAR category $\sA\sG\cong \sA\sP$ satisfies Assumption \ref{assume:Generating}, the lemmas from the last subsection apply, and we easily obtain the complete classification of the representations of $\sA\sG\cong \sA\sP$.
We work with $\sA\sP$ so we can work modulo the ideal of diagrams without the maximal number of through strings.
We give two equivalent constructions of the irreducible modules; the first follows the technique of \cite{MR1929335}, and the second uses the algebra decomposition of $\sA\sP_k$ given in Proposition \ref{prop:AnnularAlgebraStructure}.
\begin{nota}
We identify the maps $\alpha_i,\alpha_j^*,\tau\in \sA\sG$ with their images in $\sA\sP$ under the equivalence $\Psi$ given in Theorem \ref{thm:Equivalence}.
\end{nota}
\begin{defn}
Let $\sA\sI(k,m)$ be the space spanned by tangles in $\sA\sP(k,m)$ with fewer than $k$ through strings, so $\sA\sI(k,m)=(0)$ if $m<k$.
We will use the usual abbreviation $\sA\sI_k = \sA\sI(k,k)$.
Note that $\sA\sI_k$ has codimension $k$ in $\sA\sP_k$, and $\sA\sP_k/\sA\sI_k \cong \mathbb{C}[\tau]\cong \mathbb{C}[\mathbb{Z}/k]$.
\end{defn}
\begin{prop}\label{prop:AnnularDescend}
Let $V$ be a Hilbert $\sA\sP$-module, and let $W_k$ be the $\sA\sP_k$-submodule of $V_k$ generated by all the $\sA\sP$-submodules of $V$ with weight less than $k$. Then
$$
W_k^\perp = \bigcap_{a\in \sA\sI_k}\ker(a).
$$
\end{prop}
\begin{proof}
Suppose $\xi\in W_k^\perp$, and let $a\in \sA\sI_k$.
Then $a$ is a linear combination of elements of the form $b^*c$ where $b,c\in \sA\sP(k,m)$ with $m<k$.
For any $\eta\in V_k$, we have $b\eta\in V_m$ with $m<k$, $c^*(b\eta)\in W_k$.
Hence $\langle a \xi, \eta\rangle = \langle \xi ,c^* b\eta\rangle=0$, and thus $a\xi=0$.
Suppose $\xi\in \ker(a)$ for all $a\in \sA\sI_k$, and let $\eta\in W_k$. Then $\eta$ is a linear combination of elements of the form $b \zeta$ where $\zeta\in V_m$ and $b$ a single diagram in $\sA\sP(m,k)$ with $m<k$.
\begin{claim}
There is a diagram $c\in\sA\sI_k$ such that $b=cb$.
\end{claim}
\begin{proof}
Let $c$ be the projection in $\sA\sP_k$ corresponding to a projection in $\sR\sP_k$ under the cutting operation in Figure \ref{fig:CuttingAndGluing} such that $c$ has only non-dotted through strings in the same positions as the (dotted) through strings of $b$.
Clearly $b=cb$, and since $b\in\sA\sP(m,k)$, $c$ has at most $m<k$ through strings, and thus $c\in \sA\sI_k$.
\end{proof}
By the claim, $\langle \xi, \eta\rangle = \langle \xi,cb\zeta\rangle = \langle c^*\xi,b\zeta\rangle=0$, and $\xi\in W_k^\perp$.
\end{proof}
\begin{cor}\label{cor:AnnularDecend}
If $V$ is irreducible of weight $k$, then $V_k$ descends to an irreducible $\sA\sP_k/\sA\sI_k\cong \mathbb{C}[\tau]\cong \mathbb{C}[\mathbb{Z}/k]$-module, all of which are one-dimensional.
\end{cor}
\begin{defn}
Since the action of $\sA\sP(m,n)$ on $\sA\sP(k,m)$ can only decrease the number of through strings, the action maps $\sA\sI(k,m)$ to $\sA\sI(k,n)$, and thus the action descends to an action of $\sA\sP(m,n)$ on $\sA\sP(k,m)/\sA\sI(k,m)$.
Note that the diagrams of $\sA\sP(k,m)$ with exactly $k$ through strings descend to a basis of $\sA\sP(k,m)/\sA\sI(k,m)$ under the canonical surjection.
Hence we may think of the action of $\sA\sP(m,n)$ on $\sA\sP(k,m)/\sA\sI(k,m)$ as the usual action in $\sA\sP$, except with the rule that if a composite has fewer than $k$ through strings, then we get zero.
Now $\langle \tau\rangle \cong \mathbb{Z}/k$ acts on $\sA\sP(k,m)/\sA\sI(k,m)$ by internal rotation, freely permuting the diagrammatic basis, and the internal rotation action commutes with the external $\sA\sP(m,n)$ action.
This means $\sA\sP(k,m)/\sA\sI(k,m)$ splits into the eigenspaces of the rotation $\tau$.
Let $V^{k,\omega}_m$ be the eigenspace of $\sA\sP(k,m)/\sA\sI(k,m)$ associated to the rotational eigenvalue $\omega$.
\end{defn}
\begin{prop}\label{prop:DimV}
$\displaystyle\dim(V^{k,\omega}_m)=
\begin{cases}
0 &\text{if }m<k\\
\displaystyle{n\choose k} &\text{if }m\geq k.
\end{cases}
$
\end{prop}
\begin{proof}
For $k\leq m$, the action of $\mathbb{Z}/k$ on $\sA\sP(k,m)/\sA\sI(k,m)$ is free, and
$$
\dim(\sA\sP(k,m)/\sA\sI(k,m))=N(k,m;k)=
\begin{cases}
\displaystyle k{m\choose k}{k-1 \choose k-1}=k{m\choose k} & \text{if }k>0\\
1 & \text{if }k=0
\end{cases}
$$
by Remark \ref{rem:CountTangles} and Lemma \ref{lem:CountTangles}.
\end{proof}
\begin{defn}\label{defn:VBasis}
Let
$$
\xi_k^\omega=
\frac{1}{Z}
\sum_{j=1}^k \omega^{-j}\tau^j
\in\sA\sP_k.
$$
where $Z$ is a normalization constant to be determined later.
Note that $\tau \xi_k^\omega=\omega \xi^\omega_k$, and $\alpha_i^*\xi^\omega_k=0$ for all $i=1,\dots, k$, since it is in $V^{k,\omega}_{k-1}=(0)$.
By Proposition \ref{prop:DimV}, a basis for $V^{k,\omega}_m$ is given by $\set{\alpha_{i_{m-k}}\cdots\alpha_{i_{1}}\xi^\omega_k}{i_1<\cdots < i_{m-k}}$.
For example, $V^{2,\omega}_4$ has the following diagrammatic basis:
$$
\left\{
\underset{\alpha_2\alpha_1\xi^\omega_2}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\AnnularBrokenTopString{135}
\AnnularBrokenTopString{45}
\AnnularString{-45}
\AnnularString{-135}
\node at (0:0) {$\xi^\omega_2$};
\node at (180:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
\,,\,
\underset{\alpha_3\alpha_1\xi^\omega_2}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\AnnularBrokenTopString{135}
\AnnularString{45}
\AnnularBrokenTopString{-45}
\AnnularString{-135}
\node at (0:0) {$\xi^\omega_2$};
\node at (180:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
\,,\,
\underset{\alpha_4\alpha_1\xi^\omega_2}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\AnnularBrokenTopString{135}
\AnnularString{45}
\AnnularString{-45}
\AnnularBrokenTopString{-135}
\node at (0:0) {$\xi^\omega_2$};
\node at (180:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
\,,\,
\underset{\alpha_3\alpha_2\xi^\omega_2}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\AnnularString{135}
\AnnularBrokenTopString{45}
\AnnularBrokenTopString{-45}
\AnnularString{-135}
\node at (0:0) {$\xi^\omega_2$};
\node at (180:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
\,,\,
\underset{\alpha_4\alpha_2\xi^\omega_2}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\AnnularString{135}
\AnnularBrokenTopString{45}
\AnnularString{-45}
\AnnularBrokenTopString{-135}
\node at (0:0) {$\xi^\omega_2$};
\node at (180:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
\,,\,
\underset{\alpha_4\alpha_3\xi^\omega_2}{
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\AnnularString{135}
\AnnularString{45}
\AnnularBrokenTopString{-45}
\AnnularBrokenTopString{-135}
\node at (0:0) {$\xi^\omega_2$};
\node at (180:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
}
\right\}.
$$
\end{defn}
\begin{defn}\label{defn:AnnularInnerProduct}
We define an inner product on $V^{k,\omega}$ as follows. Let $\tr$ be a faithful trace on $\sA\sP_k/\sA\sI_k\cong \mathbb{C}[\tau]\cong \mathbb{Z}/n$, and extend $\tr$ to $\sA\sP_k$ via the quotient map.
Given $a,b\in\sA\sP(k,m)$, $b^*a\in\sA\sP_k$, so we define an inner product on $\sA\sP_m$ by $\langle a,b\rangle_m = \tr(b^*a)$.
It is clear that $\langle \cdot,\cdot\rangle$ satisfies $\langle ab,c\rangle = \langle b,a^*c\rangle$, and the rotation $\tau$ is clearly unitary, so the decomposition into the $V^{k,\omega}$ is orthogonal.
We can now choose the $Z$ in Definition \ref{defn:VBasis} to be any scalar for which $\|\xi_k^\omega\|^2=\langle \xi^\omega_k ,\xi^\omega_k\rangle = 1$.
We immediately get that the that basis of $V^{k,\omega}_m$ given in Definition \ref{defn:VBasis} is orthonormal.
\end{defn}
\begin{prop}\label{prop:DetermineInnerProduct}
All inner products in $V^{k,\omega}$ are determined by:
\begin{enumerate}[(1)]
\item $\langle \xi^\omega_k ,\xi^\omega_k\rangle = 1$,
\item $\alpha_i^*\xi_k^\omega=0$ for all $i=1,\dots, k$, i.e., $\xi^\omega_k$ is \underline{uncappable}, and
\item $\tau \xi_k^\omega = \omega \xi^\omega_k$.
\end{enumerate}
\end{prop}
\begin{proof}
$V^{k,\omega}_m$ is spanned by elements of the form $a \xi^\omega_k$ with $a$ a single diagram in $\sA\sP(k,m)$.
Since $\xi^\omega_k$ is uncappable, we see that $\langle a \xi^\omega_k , b\xi^\omega_k\rangle$ is zero unless all through strings connected to one $\xi^\omega_k$ attach to the other.
When $\xi^\omega_k$ is not capped off, we use the rotational eigenvector property to end up with some power of $\omega$ times $\langle \xi^\omega_k, \xi^\omega_k\rangle$.
\end{proof}
\begin{prop}\label{prop:UniqueAnnularModule}
Under this inner product, $V^{k,\omega}$ is an irreducible Hilbert $\sA\sP$-module of weight $k$.
Moreover, any irreducible Hilbert $\sA\sP$-module of weight $k$ is isomorphic to some $V^{k,\omega}$.
\end{prop}
\begin{proof}
We know $V^{k,\omega}$ is a Hilbert $\sA\sP$-module, since we have exhibited an orthonormal basis. Irreducibility now follows by Lemmas \ref{lem:IrreducibleSubmodule} and \ref{lem:AllIrreducible} since $V^{k,\omega}=\sA\sP(\xi^\omega_k)$.
If $W$ is another irreducible $\sA\sP$-module of weight $k$, then it is generated by a lowest weight rotational eigenvector vector $\eta\in W_k$ by Lemma \ref{lem:AllIrreducible} and Corollary \ref{cor:AnnularDecend}.
Let $\omega$ be the rotational eigenvalue, and without loss of generality, assume $\|\eta\|_{W_k}=1$.
Define $\theta : V^{k,\omega}_k \to W_k$ by $\theta(\xi^\omega_k) = \eta$, which is a non-zero homomorphism of $\sA\sP_k$-modules.
By Lemma \ref{lem:ExtendMorphism}, $\theta$ extends uniquely to an injective homomorphism $\Theta: V^{k,\omega}\to W$
which preserves the inner product by Proposition \ref{prop:DetermineInnerProduct}.
It is clear $\Theta$ is an isomorphism, as we can construct its inverse similarly.
\end{proof}
\begin{rem}\label{rem:EquivalentDefinition}
We get the following equivalent characterization of $V^{k,\omega}$.
Let $\dot{\tau}$ be the dotted rotation operator, e.g.,
$$
\begin{tikzpicture}[annular]
\clip (0,0) circle (1.6cm);
\draw (216:0.5cm) .. controls ++(216:.6cm) and ++(324:.4cm) .. (144:1.5cm);
\draw (144:0.5cm) .. controls ++(144:.6cm) and ++(252:.4cm) .. (72:1.5cm);
\draw (72:0.5cm) .. controls ++(72:.6cm) and ++(180:.4cm) .. (0:1.5cm);
\draw (0:0.5cm) .. controls ++(0:.6cm) and ++(108:.4cm) .. (-72:1.5cm);
\draw (-72:0.5cm) .. controls ++(-72:.6cm) and ++(36:.4cm) .. (-144:1.5cm);
\filldraw (18:1cm) circle (.05cm);
\filldraw (90:1cm) circle (.05cm);
\filldraw (162:1cm) circle (.05cm);
\filldraw (-126:1cm) circle (.05cm);
\filldraw (-54:1cm) circle (.05cm);
\node at (158:.75cm) {$\star$};
\node at (180:1.25cm) {$\star$};
\draw [very thick] (0,0) circle (.5cm);
\draw [very thick] (0,0) circle (1.5cm);
\end{tikzpicture}
=\dot{\tau}\in \sA\sP_5.
$$
Recall from Proposition \ref{prop:AnnularAlgebraStructure} that
$$
\sA\sP_k \cong \mathbb{C} \oplus \bigoplus_{j=1}^k j M_{k\choose j}(\mathbb{C}),
$$
where the $j$ matrix algebras of size $k\choose j$ correspond to the annuli with $j$ dotted through strings.
This means that the $k$ powers of $\dot{\tau}$ correspond to $k$ copies of $\mathbb{C}$.
Define a non-faithful trace $\tr$ on $\sA\sP_k$ by mapping the minimal projections
$$
p_k^\omega=
\frac{1}{k}
\sum_{j=1}^k \omega^{-j}\dot{\tau}^j
\in\sA\sP_k
$$
to 1 and mapping all other minimal projections to zero.
Then under the usual GNS sesquilinear form $\langle a,b\rangle = \tr(b^*a)$, the minimal projections $p_k^\omega$ are orthonormal, and all other minimal projections have length zero.
Note that $\tau p_k^\omega=\omega \xi^\omega_k$, and $\alpha_i^*p_k^\omega=0$ for all $i=1,\dots, k$.
Now we can extend this sesquilinear form to $\sA\sP(k,m)$ as in Definition \ref{defn:AnnularInnerProduct}, since for any $a,b\in\sA\sP(k,m)$, $b^*a\in\sA\sP_k$.
Let $V^{k}_m$ be the quotient of $\sA\sP(k,m)$ by the radical of the sesquilinear form, so that $V^k_m$ is a Hilbert space, which naturally carries an action of $\sA\sP(m,n)$.
The decomposition of $V^k_k=\spann\set{p_k^\omega}{\omega^k=1}$ into orthogonal eigenspaces for the rotation is easy.
Set $V^{k,\omega}_k=\spann\{p_k^\omega\}$, and let $V^{k,\omega}$ be the irreducible Hilbert $\sA\sP$-submodule generated by $p_k^\omega$.
Proposition \ref{prop:UniqueAnnularModule} implies this definition is equivalent to the previous definition via the isomorphism which maps $p_k^\omega$ to $\xi^\omega_k$.
\end{rem}
\subsection{The representation theory of $\sR\sG\cong \sR\sP$}
We now do the same for $\sR\sG\cong \sR\sP$.
The proofs of the propositions in the subsection are similar to the proofs from the last subsection, and they will be omitted.
\begin{nota}
We now identify $\alpha_i,\alpha_j^*\in\sR\sG$ with their image in $\sR\sP$ under the restriction of the equivalence $\Psi$ as in Theorem \ref{thm:RectangularEquivalence}.
\end{nota}
\begin{defn}
Let $\sR \sI_n\subset \sR\sP_n$ denote the ideal generated by the diagrams with fewer than $n$ through strings.
Note that $\sR\sI_n$ has codimension one in $\sR\sP_n$.
\end{defn}
\begin{prop}
Let $V$ be a Hilbert $\sR\sP$-module, and let $W_k$ be the $\sR\sP_k$-submodule of $V_k$ generated by all the $\sR\sP$-submodules of $V$ with weight less than $k$. Then
$$
W_k^\perp = \bigcap_{a\in \sR\sI_k}\ker(a).
$$
\end{prop}
\begin{cor}\label{cor:RectangularDecend}
If $V$ is irreducible of weight $k$, then $V_k$ descends to an irreducible $\sR\sP_k/\sR\sI_k\cong \mathbb{C}$-module, which is one dimensional.
\end{cor}
\begin{defn}
We define the irreducible Hilbert $\sR\sP$-modules $V^k$ using the technique of Remark \ref{rem:EquivalentDefinition}.
For $k\geq 0$, let
$$
p_k=
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{.2}{};
\DottedString{(-.4,0)};
\node at (.05,0) {$\dots$};
\node at (.4,.6) {\scriptsize{$k$}};
\DottedString{(.4,0)};
\end{tikzpicture}
\in\sR\sP_k,
$$
and note that $\alpha_i^*p_k=0$ for all $i=1,\dots, k$.
Define a sesquilinear form on $\sR\sP_k$ by declaring the minimal projection $p_k$ to be a unit vector, and all other minimal projections have length zero.
Extend the sesquilinear form to $\sR\sP(k,m)$ as before, and let $V^k_m$ be the quotient of $\sR\sP(k,m)$ by the radical of the form.
Then $V^k$ is an irreducible Hilbert $\sR\sP$-module of weight $k$.
\end{defn}
\begin{prop}
$
\displaystyle
\dim(V^k_m) =
\begin{cases}
0 &\text{if }m<k\\
\displaystyle{n\choose k} &\text{if }m\geq k.
\end{cases}
$
\end{prop}
\begin{prop}\label{prop:DetermineRectangularInnerProduct}
All inner products in $V^{k}$ are determined by:
\begin{enumerate}[(1)]
\item $\langle p_k ,p_k\rangle = 1$, and
\item $\alpha_i^*p_k=0$ for all $i=1,\dots, k$, i.e., $p_k$ is uncappable.
\end{enumerate}
\end{prop}
\begin{prop}
$V^k$ is the the unique irreducible Hilbert $\sR\sP$-module of weight $k$ up to isomorphism.
\end{prop}
\begin{rem}
It would be interesting to fully compute the decomposition of the Temperley-Lieb algebras $TL_n(\delta)$ as irreducible $\sA\sG$ and $\sR\sG$-modules.
For example, while there is only one Temperley-Lieb module for Temperley-Lieb, there are many GICAR modules.
It is well known that the $n$-th Catalan number counts the number of non-crossing partitions of $\{1,\dots,n\}$.
Using this fact, we expect to find for each $n\in \bbN$ a new low-weight generator corresponding to the partition which includes all $\{1,\dots, n\}$ and to use these elements to decompose the $TL_n(\delta)$ into irreducible modules by applying rotational symmetries.
We leave this for another time.
\end{rem}
\section{Hilbert bimodules and {\rm II}$_1$-subfactors}\label{sec:InfiniteBackground}
We now have a brief interlude to introduce the background necessary to construct our {\rm II}$_1$-factor bimodule and subfactor representations of the annular and rectangular GICAR categories.
\subsection{Hilbert bimodules}\label{sec:BimoduleBackground}
We refer to \cite[Section 2]{MR3040370} for the background on Hilbert bimodules. We rapidly introduce our notation and conventions.
\begin{nota}\mbox{}
\begin{itemize}
\item
$A$ is a {\rm II}$_1$-factor.
\item
$H$ is an $A-A$ Hilbert bimodule.
\item
$D({\sb{A}H})$ is the set of left $A$-bounded vectors.
\begin{itemize}
\item
For each $\eta\in D({\sb{A}H})$ the right multiplication operator $R(\eta): L^2(A)\to H$ is the unique extension of $\widehat{a}\mapsto a\eta$.
\item
On $D({\sb{A}H})$, the $A$-valued inner product ${\sb{A}\langle} \eta_1,\eta_2\rangle = JR(\eta_1)^*R(\eta_2)J\in A$ is $A$-linear on the \underline{left}.
\item
An $\sb{A}H$ basis (which exists by \cite{MR561983}) is a set of vectors $\{\alpha\}\subset D({\sb{A}H})$ such that
$$
\sum\limits_\alpha R(\alpha)R(\alpha)^* = 1_H \Longleftrightarrow \sum_\alpha {\sb{A}\langle}\eta ,\alpha\rangle\alpha = \eta\text{ for all }\eta\in D({\sb{A}H}).
$$
\item
The canonical normal, faithful, semifinite (n.f.s.) trace on $A'\cap B(H)$ is given by $\Tr_{A'\cap B(H)}(x)= \sum_\alpha \langle x \alpha,\alpha\rangle$ where $\{\alpha\}$ is any $\sb{A}H$ basis.
\end{itemize}
\item
$D(H_A)$ is the set of right $A$-bounded vectors.
\begin{itemize}
\item
For each $\xi\in D(H_A)$ the left multiplication operator $L(\xi): L^2(A)\to H$ is the unique extension of $\widehat{a}\mapsto \xi a$.
\item
On $D(H_A)$, the $A$-valued inner product $\langle \xi_1|\xi_2\rangle_A = L(\xi_1)^*L(\xi_2)\in A$ is $A$-linear on the \underline{right}.
\item
An $H_A$ basis (which exists by \cite{MR561983}) is a set of vectors $\{\beta\}\subset D(H_A)$ such that
$$
\sum\limits_\beta L(\beta)L(\beta)^* = 1_H\Longleftrightarrow \sum_\beta \beta \langle \beta|\xi\rangle_A = \xi \text{ for all } \xi\in D(H_A).
$$
\item
The canonical n.f.s. trace on on $(A\op)'\cap B(H)$ is given by $\Tr_{(A\op)'\cap B(H)}(x)= \sum_\beta \langle x \beta,\beta\rangle$ where $\{\beta\}$ is any $H_A$ basis.
\end{itemize}
\item
$H^n = \bigotimes^n_A H$, and we use the convention $H^0=L^2(A)$.
\begin{itemize}
\item
For $\eta\in D(H^k_A)$ and $\xi\in D({\sb{A}H^n})$, we denote their relative tensor product in $H^{k+n}$ by $\eta\otimes \xi$.
\item
For each $\eta\in D({\sb{A}H})$, the right creation operator $R_{\eta} : H^n \to H^{n+1}$ is the unique extension of $\zeta\mapsto \zeta\otimes \eta$ for $\zeta\in D(H^n_A)$.
Its adjoint is given by $R_{\eta}^* (\zeta\otimes \xi)=\zeta {\sb{A}\langle } \xi,\eta\rangle$ for $\zeta,\xi$ appropriate bounded vectors.
\item
For each $\xi_1\in D(H_A)$, the right creation operator $L_{\xi_1} : H^n \to H^{n+1}$ is the unique extension of $\xi_2\mapsto \xi_1\otimes \xi_2$ for $\xi_w\in D({\sb{A}H^n})$.
Its adjoint is given by $L_{\xi}^* (\eta\otimes \zeta)=\langle \xi| \eta\rangle_A \zeta$ for $\eta,\zeta$ appropriate bounded vectors.
\item
If $x\in (A\op)'\cap B(H^k)$ and $y\in A'\cap B(H^n)$, the operator $x\otimes_A y \in B(H^{k+n})$ given by the unique extension of $\xi\otimes \eta\mapsto (x\eta)\otimes(y\xi)$ where $\eta\in D(H^k_A)$ and $\xi\in D({\sb{A}H^n})$ is well-defined and bounded, and $\|x\otimes_A y\|_\infty\leq \|x\|_\infty\|y\|_\infty$.
\end{itemize}
\item
$B^n=D(\sb{A}H^n)\cap D(H^n_A)$, the simultaneously left and right-bounded vectors, which are dense in $H^n$ \cite[Lemma 1.2.2]{correspondences}.
Also, we use the convention that $B=B^1$, and we note $B^0=A$.
{\textbf{Note:}} In the case that $H$ is obtained from a {\rm II}$_1$-superfactor $A_1$ of $A=A_0$, $B$ will have a different meaning.
However, all statements we make about $B$ will still hold for either definition of $B$.
See Remark \ref{rem:NewDefinitionForB} in Subsection \ref{sec:SubfactorBackground}.
\item
Fix $\{\alpha\}\subset B$ an $\sb{A}H$ basis (see Lemma \ref{lem:BasesInB} below).
We have
$$\{\alpha^n\}=\set{\alpha_{1}\otimes\cdots\otimes\alpha_{n} }{\alpha_{i}\in\{\alpha\}\text{ for all } i=1,\dots,n}\subset B^n$$
is the corresponding $\sb{A}H^n$ basis, since $R_{\alpha_{1}\otimes\cdots\otimes \alpha_{n}} =R_{\alpha_{1}}\cdots R_{\alpha_{n}}$.
Similarly, we let $\{\beta\}\subset B$ be an $H_A$ basis, and we have the corresponding $H_A^n$ basis $\{{\beta}^n\}\subset B^n$.
\item
$C_n = (A^{\OP})'\cap B(H^n)$, the commutant of the right $A$-action on $H^n$, which has a canonical trace $\Tr_n=\sum_{{\beta}^n} \langle \, \cdot \, {\beta}^n,{\beta}^n\rangle$.
\begin{itemize}
\item
The inclusion $C_n\hookrightarrow C_{n+1}$ is given by $x\mapsto x\otimes_A \id_H$.
\item
The unique trace preserving n.f.s. operator valued weight\\
$T_{n+1} : (C_{n+1}^+,\Tr_{n+1})\to (\widehat{C_{n}^+},\Tr_{n})$ is given by $x\mapsto \sum_\beta R_\beta^* x R_\beta$.
\end{itemize}
\item
$C_n^{\OP} = A'\cap B(H^n)$, which has a canonical trace $\Tr_n^{\OP}=\sum_{{\alpha}^n} \langle \, \cdot \, {\alpha}^n,{\alpha}^n\rangle$.
\begin{itemize}
\item
The inclusion $C_n\op\hookrightarrow C_{n+1}\op$ is given by $y\mapsto \id_H\otimes_A y$.
\item
The unique trace preserving n.f.s. operator valued weight\\
$T_{n+1}\op: ((C_{n+1}\op)^+,\Tr_{n+1}\op)\to (\widehat{(C_{n}\op)^+},\Tr_{n}\op)$ is given by $y\mapsto \sum_\alpha L_\alpha^* y L_\alpha$.
\end{itemize}
\item
The \underline{standard invariant} of $H$ is the sequences of centralizer algebras $\cQ_n=C_n\cap C_n\op$ and central $L^2$-vectors $\cP_n = A'\cap H^n=\set{\zeta\in H^n}{a\zeta=\zeta a \text{ for all }a\in A}$.
\begin{itemize}
\item
The planar calculus of \cite{MR3040370} acts on the $\cQ_n$ and $\cP_n$.
Note that the $\cQ_n$ naturally act on the $\cP_n$.
\item
Note that $\cP_0=A'\cap L^2(A) = \mathbb{C}\widehat{1}$.
\end{itemize}
\end{itemize}
\end{nota}
The next lemma was used without proof in \cite{MR3040370}; due to its importance, we provide a proof for the convenience of the reader.
The proof uses ideas similar to \cite[Lemma 1.2.2]{correspondences} and \cite[Proposition 3.2.19]{1111.1362}.
The latter proves this result in the subfactor case (but $B$ has a different meaning - see Remark \ref{rem:NewDefinitionForB}).
\begin{lem}\label{lem:BasesInB}
There exist $\sb{A}H$ and $H_A$-bases which are subsets of $B$.
\end{lem}
\begin{proof}
We show there is a $H_A$-basis $\{\gamma\}\subset B$, and the other case is similar.
First, let $\{\beta_i\}$ be an orthogonal $H_A$-basis, so the $L(\beta_i)L(\beta_i)^*$ are projections which sum to $1_H$, and $L(\beta_i)^*L(\beta_j)=\langle \beta_i|\beta_j\rangle_A=0$ if $i\neq j$.
This exists by \cite[Proposition 2.2]{MR1387518}.
Moreover, $C_1=\set{L(\eta)L(\xi)^*}{\eta,\xi\in D(H_A)}''$ by \cite{MR561983}.
For each $i$, $T_1(L(\beta_i)L(\beta_i)^*)$ has finite trace, and thus has a spectral resolution
$$
T_1(L(\beta_i)L(\beta_i)^*)=\int_0^\infty \lambda \, de_\lambda^i.
$$
Mimicking \cite[Proposition 3.2.19]{1111.1362}, set $\gamma_{i,1}=e_1^i \beta$ and $\gamma_{i,n} = (e_{n}^i-e_{n-1}^i)\beta$ for $n\geq 2$.
Thus $T_1(L(\gamma_{i,n})L(\gamma_{i,n})^*)$ has norm at most $n$, and $L(\beta_i)L^2(A) = \bigoplus_{n\in \bbN} L(\gamma_{i,n})L^2(A)$.
Moreover, the $\gamma_{i,n}$ are orthogonal, since $L(\gamma_{i,m})^*L(\gamma_{j,n})=L(\beta_i)^*e_me_nL(\beta_j)$, which is zero unless $i=j$ and $m=n$.
Thus $\{\gamma_{i,n}\}$ is an $H_A$-basis.
Finally, to show each $\gamma_{i,n}\in D({\sb{A}H})$, for each $a\in A$, we have
\begin{align*}
\|a\gamma_{i,n}\|_H^2
&=
\| aL(\gamma_{i,n}) \widehat{1}\|_H^2
=
\tr_A(L(\gamma_{i,n})^*a^*aL(\gamma_{i,n}))
=
\Tr_{1}(aL(\gamma_{i,n})L(\gamma_{i,n})^*a^*)\\
&=
\tr_A(a T_1(L(\gamma_{i,n})L(\gamma_{i,n})^*)a^*)
\leq
n\tr_A(aa^*)
=
n\|\widehat{a}\|_{L^2(A)}^2.
\qedhere
\end{align*}
\end{proof}
\begin{defn}\label{defn:symmetric}
$H$ is called \underline{symmetric} if there is a conjugate-linear isomorphism $J: H\to H$ such that $J(a\xi b)= b^* (J\xi) a^*$ for all $a,b\in A$ and $\xi\in H$ and $J^2=\id_H$.
\end{defn}
\begin{rem}\label{rem:symmetric}
If $H$ is symmetric, then for $n\geq 1$, $H^n$ is symmetric with conjugate-linear isomorphism $J_n: H^n\to H^n$ given by the extension of
$$
J_n( \xi_1\otimes \cdots \otimes \xi_n)=(J\xi_1)\otimes\cdots \otimes (J\xi_n).
$$
for $\xi_i\in B$ for all $i$. Note that $J_nAJ_n=A^{\OP}$, $J_n C_nJ_n=C_n^{\OP}$, and $J_n B^n=B^n$. On $B(H^n)$, we define $j_n$ by $j_n(x)=J_nx^*J_n$. Note that $j_n^2=\id$ and $\Tr_n=\Tr_n^{\OP}\circ j_n$.
If $H$ is not symmetric, then in general, $C_n^{\OP}$ is not the opposite algebra of $C_n$, e.g. $\sb{R\otimes 1} L^2(R\otimes R)_{R\otimes R}$ where $R$ is the hyperfinite {\rm II}$_1$-factor.
\end{rem}
\subsection{Arbitrary index {\rm II}$_1$-subfactors}\label{sec:SubfactorBackground}
Suppose $A=A_0$ is contained in a {\rm II}$_1$-factor $A_1$.
The $A-A$ bimodule $H=L^2(A_1)$ is the motivating example for this paper.
\begin{rem}\label{rem:NewDefinitionForB}
In the case that $H=L^2(A_1)$, we no longer use the notation $B$ for $D({\sb{A}L^2(A_1)})\cap D(L^2(A_1)_A)$.
Rather, we set $B=A_1$, which is obviously still dense in $L^2(A_1)$.
Note that the bi-bounded vectors $D({\sb{A}L^2(A_1)})\cap D(L^2(A_1)_A)$ do not agree with the image of $B$ in $L^2(B)$ when $[B:A]=\infty$.
However, all statements we made concerning the $B$ in the previous subsection still apply for $B=A_1$.
\end{rem}
We can perform the Jones basic construction to get another type {\rm II} factor $\langle B, e_A\rangle$, which is type {\rm II}$_1$ if and only if $[B: A]<\infty$ \cite{MR696688}.
When $[B:A]<\infty$, we can form the Jones tower, and the higher relative commutants form a planar algebra \cite{math/9909027}, which always includes a Temperley-Lieb planar subalgebra.
When $[B:A]=\infty$, $\langle B, e_A\rangle$ is a {\rm II}$_\infty$-factor, and we must be more careful.
Detailed analysis of this situation was started in \cite{MR1387518,1111.1362},
and planar structure was given for the centralizer algebras and central $L^2$-vectors in \cite{MR3040370}.
We rapidly recall the necessary background from \cite{MR1387518,1111.1362}.
\begin{defn}
Suppose $M$ is a semifinite von Neumann algebra with n.f.s. trace $\Tr_M$. Define
\begin{align*}
\mathfrak{n}_{\Tr_M}&=\set{x\in M}{\Tr_M(x^*x)<\infty}\text{ and}\\
\mathfrak{m}_{\Tr_M}&= \mathfrak{n}_{\Tr_M}^*\mathfrak{n}_{\Tr_M} = \spann\set{x^* y}{x,y\in \mathfrak{n}_{\Tr_M}}.
\end{align*}
Suppose $N$ is a von Neumann subalgebra of $M$ with n.f.s. trace $\Tr_N$. Then by \cite{MR534673}, there is a unique trace-preserving n.f.s. operator valued weight $T: M^+\to \widehat{N^+}$, the extended positive cone of $N$. We define
\begin{align*}
\mathfrak{n}_{T}&=\set{x\in M}{T(x^*x)\in N^+}\text{ and}\\
\mathfrak{m}_{T}&= \mathfrak{n}_{T}^*\mathfrak{n}_{T} = \spann\set{x^* y}{x,y\in \mathfrak{n}_{T}}.
\end{align*}
\end{defn}
Suppose $A_0=A\subset B=A_1$ is an infinite index {\rm II}$_1$-subfactor.
First, recall that $A_0,A_1$ have normal, faithful, finite normalized traces $\tr_0,\tr_1$ respectively, and $\tr_1|_{A_0}=\tr_0$.
We have $T_1=E_1 :A_1 \to A_0$ is the normal, faithful conditional expectation, which is implemented by the Jones projection $e_1\in A_0'\cap B(L^2(A_1,\tr_1))$.
The basic construction of $A_0\subset A_1$ is $A_2$, a type {\rm II}$_\infty$-factor, and the canonical n.f.s. trace $\Tr_2$ on $A_2$ satisfies
$$
\Tr_2(xe_1 y^*)=\Tr_2(L(\widehat{x})L(\widehat{y})^*)=\tr_1(xy^*)
$$
for all $x,y\in A_1$. (Note that in the notation of the previous subsection, $A_2=C_1$, and $\Tr_2=\Tr_1$.)
We form the $L^2$-space in the usual way as the closure of $\mathfrak{n}_{\Tr_2}$.
Now the unique trace-preserving operator valued weight $T_2: A_2^+\to \widehat{A_1^+}$ is not a conditional expectation. For $x\in \mathfrak{n}_{T_2}\subset A_2$, we define left creation operators $\Lambda_{T_2}(x) : L^2(A_1)\to L^2(A_2)$ which commute with the right $A_1$-action by $\widehat{y}\mapsto \widehat{xy}$ for $y\in \mathfrak{n}_{\tr_1}=A_1$ which is well-defined and bounded since $xy\in \mathfrak{n}_{T_2}\cap \mathfrak{n}_{\Tr_2}$:
$$
\|xy\|_2^2=\Tr_2((yy^*)^{1/2} x^*x(yy^*)^{1/2})=\tr_1((yy^*)^{1/2} T(x^*x)(yy^*)^{1/2})\leq \|T(x^*x)\|_\infty \|y\|_2^2.
$$
Moreover, the maps $\Lambda_{T_2}(x)$ satisfy
\begin{itemize}
\item $\Lambda_{T_2}(x)^*\widehat{y}=\widehat{T_2(x^*y)}$ for all $x\in\mathfrak{n}_{T_2}$ and $y\in \mathfrak{n}_{\Tr_2}$, and
\item $\Lambda_{T_2}(x)^*\Lambda_{T_2}(y) = E_1(x^*y)$ for all $x,y\in\mathfrak{n}_{T_2}$.
\end{itemize}
To iterate the basic construction, note that the modular conjugation $J_2$ on $\mathfrak{n}_{\Tr_2}$ extends to an anti-linear unitary. Hence we may define the basic construction by $A_3=J_2 (A_1'\cap B(L^2(A_2,\Tr_2)))J_2$.
It follows from \cite{MR1387518} that
$$
A_3= \set{\Lambda_{T_2}(x)\Lambda_{T_2}(y)^*}{x,y\in\mathfrak{n}_{T_{2}}}''.
$$
The canonical n.f.s. trace $\Tr_3$ on $A_3$ is given by
$$
\Tr_3(\Lambda_{T_2}(x)\Lambda_{T_2}(y)^*)=\tr_1(\Lambda_{T_2}(y)^*\Lambda_{T_2}(x))=\tr_1(y^*x),
$$
and the unique trace-preserving n.f.s. operator valued weight satisfies
$$
T_3(\Lambda_{T_2}(x)\Lambda_{T_2}(x)^*)=xx^*.
$$
By \cite[Section 3.2.1]{1111.1362}, $\Tr_3|_{A_2^+}=\Tr_2$, so $T_3=E_3$ is a conditional expectation, which is implemented by a Jones projection $e_3\in A_2'\cap B(L^2(A_3,\Tr_3))$.
One continues this process as in \cite{MR1387518,1111.1362} to get a tower of type {\rm II} factors $(A_n,\Tr_n)_{n\geq 0}$ together with
\begin{itemize}
\item
the conjugate-linear unitary $J_n$ on $L^2(A_n,\Tr_n)$ extending the adjoint on $\mathfrak{n}_{\Tr_n}$,
\item
the basic construction
$$
A_{n+1}
=J_n (A_{n-1}'\cap B(L^2(A_n,\Tr_n))J_n
= \set{\Lambda_{T_2}(x)\Lambda_{T_2}(y)^*}{x,y\in\mathfrak{n}_{T_{n}}}'',
$$
\item
the operator valued weights $T_{n+1}: A_{n+1}^+\to \widehat{A_{n}^+}$,
\item
the left creation operators $\Lambda_{T_n}(x): L^2(A_{n-1},\Tr_{n-1})\to L^2(A_n,\Tr_n)$ for $x\in \mathfrak{n}_{T_n}$ which commute with the right $A_{n-1}$ action,
\item
the n.f.s. traces $\Tr_{n+1}$ satisfying
\begin{align*}
\Tr_{n+1}(\Lambda_{T_n}(x)\Lambda_{T_n}(y)^*)&=\Tr_{n-1}(\Lambda_{T_n}(y)^*\Lambda_{T_n}(x))=\Tr_{n-1}(y^*x)\text{ for all }x,y\in\mathfrak{n}_{T_n},
\end{align*}
\end{itemize}
We have the following facts due to \cite{MR1387518,1111.1362}.
\begin{facts}
\mbox{}
\begin{enumerate}[(1)]
\item
$L^2(A_n)\cong \bigotimes_A^n L^2(B)$ via isomorphisms $\theta_n$ (see Subsection \ref{sec:RealizeOddJones}).
\item
Writing $B^n=\bigotimes^n_A\widehat{B}$ (again, this notation differs from Subsection \ref{sec:BimoduleBackground}), $B^n$ is dense in $\bigotimes_A^n L^2(B)$.
\item (Multistep basic construction \cite{MR1387518}) For $0\leq k\leq n$, the inclusions $A_{n-k}\subseteq A_n\subseteq A_{n+k}$ are standard, i.e.
$$
(\id_k\otimes_A J_{n-k} A_{n-k}J_{n-k})'=J_n(A_{n-k}\otimes_A \id_k)'J_n \cong A_{n+k}.
$$
\item (Shifts \cite{MR1387518})For $0\leq k\leq n$, we let $j_n(x)=J_n x^* J_n$ for $x\in B(L^2(A_n,\Tr_n))$. Then $j_n$ is an anti-isomorphism of $A_{n-k}$ onto $A_{n+k}'$. Hence if we compose $j_n$ and $j_{n+1}$, we get an isomorphism:
$$
j_{n+1}j_n (A_{n-k}'\cap A_n)\underset{anti}{\cong} j_{n+1} (A_{n}'\cap A_{n+k} )\underset{anti}{\cong} A_{n-k+2}'\cap A_{n+2}.
$$
\item
(Odd Jones projections \cite[Section 3.2.1]{1111.1362}) For all $n\in\mathbb{N}$, $\Tr_{2n}|_{A_{2n-1}^+}=\infty$ and $\Tr_{2n+1}|_{A_{2n}^+}=\Tr_{2n}$. Therefore, $T_{2n+1}: A_{2n+1}\to A_{2n}$ is a conditional expectation, which gives rise to the odd Jones projection $e_{2n+1}$.
When we realize $A_{2n}$ acting on $L^2(A_{n},\Tr_{n})$ from the multistep basic construction,
$e_{2n-1}=J_{n} e_1 J_{n}$.
\end{enumerate}
\end{facts}
\begin{rem}
For $x_1,\dots x_n\in B$, we write $\widehat{x_1}\otimes\cdots \otimes \widehat{x_n} \in B^n$ omitting the subscript $A$ to distinguish between operators and vectors, such as
$x\otimes_A \id_1$ and $\widehat{x}\otimes \widehat{1}$ for $x\in B$. One is left multiplication by $x\in B$ on $L^2(B)\otimes_A L^2(B)$, and the other is $\theta_2^{-1}(\widehat{xe_1})$.
\end{rem}
\subsection{Identifying the Jones projections}\label{sec:RealizeOddJones}
We now identify the Jones projections acting on $\bigotimes_A^n L^2(B)$ via $\theta_n$.
We recall Burns' definition of the isomorphisms $\theta_n : \bigotimes^n_A L^2(B) \to L^2(A_n)$ \cite[Section 3.2.2]{1111.1362}.
Note that our numbering differs from Burns' numbering in that we start with $A_0\subset A_1$.
Also, Burns' definition is more general in that he works with arbitrary type {\rm II} factors, and he defines a more general set of isomorphisms.
We provide a simplified definition for the reader's convenience.
\begin{defn}
The isomorphisms $\theta_n : \bigotimes^n_A L^2(B) \to L^2(A_n)$ are composites of other known isomorphisms. We define:
\begin{itemize}
\item
$v_{k+1} : L^2(A_{k})\underset{A_{k-1}}{\otimes} L^2(A_{k})\to L^2(A_{k+1})$ by $\eta\otimes J_{k}\xi \mapsto L(\eta)L(\xi)^*$ for $\eta,\xi\in D(L^2(A_k)_{A_{k-1}})$,
\item
$\iota_{k} : L^2(A_{k})\underset{A_k}{\otimes}L^2(A_{k})\to L^2(A_{k})$ by $\widehat{x}\otimes \widehat{y} \mapsto \widehat{xy}$ for $x,y\in\mathfrak{n}_{\Tr_n}$,
\item
$\displaystyle
\psi_{k,n}=\left(\bigotimes_{A_k}^{n-1} v_{k+1} \right) \circ \left(\id_k \underset{A_{k-1}}{\otimes} \left(\bigotimes_{A_{k-1}}^{n-2} \iota_k^* \right) \underset{A_{k-1}}{\otimes} \id_k\right)
$
which doubles, regroups, and contracts, i.e., $\psi_{k,n}$ is the composite map
$$
\hspace{-1.2cm}
\bigotimes_{A_{k-1}}^{n} L^2(A_k)
\to
\bigotimes_{A_{k-1}}^{n} \left(L^2(A_k)\underset{A_k}{\otimes} L^2(A_k)\right)
\cong
\bigotimes_{A_{k}}^{n-1} \left(L^2(A_k)\underset{A_{k-1}}{\otimes} L^2(A_k)\right)
\to
\bigotimes_{A_k}^{n-1} L^2(A_{k+1}),
$$
and
\item
for $n\geq 2$,
$\theta_n : \bigotimes^n_{A} L^2(B) \to L^2(A_{n})$ by $\psi_{n-1,2}\circ \psi_{n-3,3}\circ\cdots \circ \psi_{2,n-1}\circ\psi_{1,n}$.
\end{itemize}
Note that $\theta_n$ is compatible with $J : \bigotimes^n_A L^2(B)\to \bigotimes^n_A L^2(B)$ by $\xi_1\otimes \cdots\otimes \xi_n \mapsto J_B\xi_n \otimes\cdots \otimes J_B\xi_1$ since the $v_{k+1}$, $\iota_k$, and $\psi_{k,j}$ are also.
\end{defn}
\begin{lem}\label{lem:IdentifyOddJonesProjections}
When we use $\theta_n$ to transport the action of $A_{2n}$ to $\bigotimes^n_{A} L^2(B)$, the Jones projection $e_1$ maps to $e_1\otimes_A \id_{n-1}$ and the Jones projection $e_{2n-1}$ maps to $\id_{n-1} \otimes_A e_1$.
\end{lem}
\begin{proof}
The result follows from \cite[Lemma 3.3.20]{1111.1362} and the compatibility of $\theta_n$ and $J$.
\end{proof}
\begin{prop}\label{prop:IdentifyOddJonesProjections}
When we use $\theta_n$ to transport the action of $A_{2n}$ to $\bigotimes^n_{A} L^2(B)$, then for $1\leq i\leq n$, we may identify the Jones projection $e_{2i-1}$ with $\id_{i-1} \otimes_A e_1\otimes_A \id_{n-i}$ .
\end{prop}
\begin{proof}
We use strong induction on $n$. The case $n=1$ is trivial. Suppose the result holds for all $0\leq i \leq n$. Now use $\theta_{n+1}$ to realize the action of $A_{2n+2}$ on $\bigotimes^{n+1}_A L^2(B)$.
In the notation of Subsection \ref{sec:BimoduleBackground}, the inclusion $A_{2n}\hookrightarrow A_{2n+2}$ transports to the inclusion $C_n\hookrightarrow C_{n+1}$ which is given by $x\mapsto x\otimes_A \id_{1}$, so the result is true for all $1\leq i < n+1$ by the associativity of the relative tensor product of $A-A$ bilinear operators. The result for $i=n+1$ now follows by Lemma \ref{lem:IdentifyOddJonesProjections}.
\end{proof}
\section{Representations via subfactors and bimodules}\label{sec:BimoduleRepresentations}
The rectangular GICAR category $\sR\sG$ acts on tensor powers of a {\rm II}$_1$-factor bimodule which contains a copy of the trivial bimodule.
The action of the tensor category $\sR\sG$ is compatible with the tensor structure of the tensor category of bimodules.
One can imagine that the annular GICAR category $\sA\sG$ is obtained from $\sR\sG$ by tensoring the morphisms with themselves around the outside, i.e., gluing the rectangles into annuli (the opposite of Figure \ref{fig:CuttingAndGluing}).
This no longer leaves us with a tensor category, and thus $\sA\sG$ must act on the spaces obtained from the bimodules by tensoring themselves on the outside, i.e., the invariant vectors of the bimodules.
\subsection{Rectangular GICAR representations}\label{sec:RectangularGICARReps}
Let $A$ be a {\rm II}$_1$-factor and let $H$ be a Hilbert $A-A$ bimodule.
We assume the following.
\begin{assume}\label{assume:ContainsTrivial}
Suppose $H$ is not the trivial bimodule, but $H$ contains a distinguished copy of the trivial bimodule, i.e., $H\cong L^2(A)\oplus K$ for a non-zero $A-A$ Hilbert bimodule $K$.
\end{assume}
\begin{rem}\label{rem:ACentral}
Containing a distinguished copy of the trivial bimodule is equivalent to the existence of a distinguished central $L^2$-vector $\zeta\in \cP_1$ with $\langle \zeta|\zeta\rangle_A = 1_A$.
Note that all central $L^2$-vectors are automatically $A-A$ bounded.
See \cite[Sections 3.3-3.4]{MR3040370} for more details.
For all $A-A$ bounded $\kappa\in K \neq (0)$, we have $\langle \kappa |\zeta\rangle_A = 0$.
\end{rem}
Below is the main theorem of this subsection, which is implied by Theorem \ref{thm:GICARinRH}.
\begin{thm}\label{thm:GICARinQ}
There is a faithful $*,\otimes$-representation of the rectangular GICAR category $\sR\sG$ as $A-A$ bimodule maps between the $H^n$, which is independent of the left and right von Neumann dimension of $H$.
\end{thm}
Of particular importance is the following corollary, which tells us some basic structure of the centralizer algebras $\cQ_n$.
\begin{cor}
The $G_n$ embed faithfully in the centralizer algebras $\cQ_n$, so $\cQ_n$ is nonabelian for $n\geq 2$.
\end{cor}
Using the binomial theorem, it is easy to see how the algebras $G_n$ should arise as intertwiners among the $H^n$. For $n\geq 0$, let $K^n=\bigotimes_A^n K$, where as usual, $K^0=L^2(A)$. Then we have
$$
H^n\cong \left(L^2(A)\oplus K\right)^{\otimes n}\cong \bigoplus_{j=0}^n {n \choose j} K^j,
$$
and we get a canonical inclusion $G_n\hookrightarrow \End_{A-A}(H^n)$.
If $K^j$ is irreducible and distinct for all $j\in\mathbb{N}$, then $G_n\cong \End_{A-A}(H^n)$ for all $n\geq 0$.
\begin{ex}\label{ex:OuterZAction}
Let $\sigma : \mathbb{Z}\to \Out(R)$ be an outer action, where $R$ is the hyperfinite {\rm II}$_1$-factor.
We denote $\sigma(n)$ by $\sigma^n$.
Let $K=L^2(A)_\sigma$, where the action is given by $a\cdot \widehat{b}\cdot c = (ab\sigma(c))^{\widehat{\,\,}}$.
Recall that $K^j\cong L^2(A)_{\sigma^j}$ for all $j\geq 0$.
Hence each $K^j$ is irreducible and distinct, and
$\End_{R-R}\left(\bigotimes^n_R(L^2(R)\oplus K)\right)\cong G_n$ for all $n\geq 0$.
\end{ex}
\begin{quests}
Is there such a $K$...
\begin{itemize}
\item
which is symmetric?
\item
which is of the form $L^2(B)\ominus L^2(A)$ for a (necessarily infinite index) {\rm II}$_1$-subfactor $A\subset B$?
\end{itemize}
\end{quests}
With more care, we obtain a faithful representation of the entire rectangular GICAR category $\sR\sG$ as $A-A$ bimodule maps among the $H^n$'s.
\begin{rem}
Recall that $H\cong L^2(A)\oplus K$, where the copy of $L^2(A)$ corresponds to the distinguished central $L^2$-vector $\zeta\in\cP_1$.
Note that $L(\zeta): L^2(A)\to H$ can be viewed as the inclusion, and $L(\zeta)^* : H\to L^2(A)$ behaves like the Jones projection for a {\rm II}$_1$-subfactor.
More precisely, if $\xi\in B=D({\sb{A}H})\cap D(H_A)$, then $L(\zeta)^*\xi=\langle \zeta|\xi\rangle_A$ defines an element of $A$.
\end{rem}
\begin{nota}
We write $e_A=L(\zeta)^*$ and $e_A^*=L(\zeta)$.
Note that $e_A,e_A^*$ are $A-A$ bilinear since $\zeta\in\cP_1$.
\end{nota}
\begin{defn}
Given an $A-A$ bimodule $H$, we define the rectangular bimodule category $\sR(H)$ as the following small involutive tensor category:
\itt{Objects} $H^n$ for $n\geq 0$.
\itt{Tensoring objects}
Connes relative tensor product. Note that $H^m\otimes_A H^n \cong H^{m+n}$.
The associators are the unique extensions of the obvious associators on the subspaces of bounded vectors $B^n$.
\itt{Morphisms}
For $1\leq i\leq n$, define the maps $\a_i^*: H^n \to H^{n-1}$ by the following commutative diagrams:
\[
\xymatrix{
H^n \ar@{<->}[rr]^(.3)\cong\ar[d]_{\a_i^*} &&H^{i-1} \otimes_A H \otimes_A H^{n-i} \ar[d]^{\id_{i-1}\otimes_A e_A \otimes_A \id_{n-i}}\\
H^{n-1}\ar@{<->}[rr]^(.3)\cong &&H^{i-1}\otimes_A L^2(A) \otimes_A H^{n-i}.
}
\]
The horizontal arrows are the associator isomorphisms.
For $1\leq i\leq n$, we define the maps $\a_j: H^n \to H^{n-1}$ similarly, but replacing $e_A$ with $e_A^*$.
The morphisms of $\sR(H)$ are $\mathbb{C}$-linear combinations of all composites of the $\a_i,\a_j^*$.
Note that these morphisms are all $A-A$ bimodule maps.
\itt{Composition} composition of operators.
\itt{Adjoint} adjoint of operators.
\end{defn}
We have the following explicit characterization of the maps $\a_i,\a_j^*$.
\begin{prop}\label{prop:RelativeTensorMaps}
The maps $a_i$, $a_i^*$ are given by the unique extensions of
\begin{align*}
\a_i(\xi_1\otimes\cdots \otimes \xi_n) & = \xi_1\otimes\cdots \otimes \xi_{i-1} \otimes \zeta \otimes \xi_i\otimes\cdots\otimes \xi_n\tag{creation}\\
\a_i^*(\xi_1\otimes\cdots \otimes \xi_n) & =\xi_1\otimes\cdots\otimes \xi_{i-1} \otimes e_A(\xi_i)\xi_{i+1}\otimes\cdots\otimes \xi_n \tag{annihilation}
\end{align*}
where $\xi_j\in B$ for all $j$.
\end{prop}
\begin{proof}
Since $\zeta\in \cP_1$, the right hand side of the first formula is well-defined.
Since $\xi_j$ is $A$-bounded, $e_A(\xi_j)=\langle \zeta|\xi_j\rangle_A$ defines an element of $A$.
Since $\zeta\in\cP_1$ is $A$-central, $e_A$ is $A-A$ bilinear, and the right hand side of the second formula is well-defined.
The rest is a straightforward calculation.
\end{proof}
Compare Relations \eqref{rel:AG1}-\eqref{rel:AG2} and Proposition \ref{prop:Standard} with Proposition \ref{prop:Relations}.
\begin{prop}\label{prop:Relations}
\mbox{}
\begin{enumerate}[(1)]
\item The words on $\a_i,\a_j^*$ satisfy the following relations:
\begin{enumerate}[(i)]
\item $\a_i\a_{j-1} = \a_{j}\a_i$ and $\a_i^* \a_{j}^* = \a_{j-1}^* \a_{i}^*$ for all $i<j$,
\item $\displaystyle \a_{i}^*\a_{j}=
\begin{cases}
\a_{j+1}\a_{i}^* &\text{if }i<j\\
\id_{n} & \text{if } i=j\\
\a_j \a_{i+1}^* & \text{if } i>j
\end{cases}$ \hspace{.2in} and
\hspace{.2in}
\item $\a_i\a_i^* =\id_{i-1}\underset{A}{\otimes}\, e_A^*e_A \underset{A}{\otimes} \id_{n-i}$ for all $i\leq n$.
\end{enumerate}
\item Each word in the $a_i,a_j^*$ has a unique standard form
$$
\a_{i_k}\cdots \a_{i_1}\a_{j_1}^*\cdots \a_{j_\ell}^*
$$
where $i_1<\cdots <i_{k}$ and $j_1<\cdots j_\ell$.
\end{enumerate}
\end{prop}
\begin{proof}
Straightforward from Proposition \ref{prop:RelativeTensorMaps}.
\end{proof}
Comparing $(1.iii)$ in Proposition \ref{prop:Relations} with Proposition \ref{prop:IdentifyOddJonesProjections}, we make the following definition.
\begin{defn}[Odd Jones projections]\label{def:OddJones}
For $1\leq i\leq n$, define $e_{2i-1}=\a_i\a_i^*$.
\end{defn}
\begin{cor}
For $1\leq i,j \leq n$, the $\a_i\a_j^*\in \cQ_n$ witness the von Neumann equivalence of the projections $e_{2i-1},e_{2j-1}\in \cQ_n$.
Thus once we know $e_{2i-1}\neq e_{2j-1}$ (which follows from Corollary \ref{cor:LinearlyIndependent}), $\cQ_n$ is not abelian for $n\geq 2$.
\end{cor}
\begin{proof}
By Proposition \ref{prop:Relations},
\begin{align*}
(\a_i\a_j^*)(\a_j\a_i^*)
&= \a_i\a_i^* = e_{i-1}\otimes_A e_1\otimes_A \id_{n-i} =e_{2i-1}\text{ and}\\
(\a_j\a_i^*)(\a_i\a_j^*)
&= \a_j\a_j^* = e_{j-1}\otimes_A e_1\otimes_A \id_{n-j} =e_{2j-1}.
\end{align*}
\end{proof}
\begin{rem}
Suppose $H=L^2(B)$ where $A\subset B$ is a {\rm II}$_1$-subfactor.
In this case, since $H^n\cong L^2(A_n)$, we have $\cQ_n\cong A_0'\cap A_{2n}$, and the odd Jones projections in Definition \ref{def:OddJones} agree with Burns' odd Jones projections via Proposition \ref{prop:IdentifyOddJonesProjections}.
Thus $A_0'\cap A_{2n}$ is not abelian for $n\geq 2$.
\end{rem}
\begin{lem}\label{lem:Kappa}
There is a $\kappa\in K$ with $\|\kappa\|_K=1$ such that
\begin{enumerate}[(1)]
\item
$\kappa\in D(H_A)$ and $\langle \kappa |\kappa\rangle_A = 1_A$, or
\item
$\kappa\in D({\sb{A}K})$ and ${\sb{A}\langle} \kappa,\kappa\rangle = 1_A$.
\end{enumerate}
Hence the simple relative tensors consisting of only $\kappa$'s and $\zeta$'s are well-defined vectors in $H^n$.
\end{lem}
\begin{proof}
Since $K$ is a non-zero $A-A$ bimodule, $\dim_{A-}({\sb{A}K})\dim_{-A}(K_A)\geq 1$.
Suppose $\dim_{-A}(K_A)\geq 1$, and choose a submodule $M_A\subset K_A$ such that $\dim_{-A}(M_A)=1$.
Then there is a spatial isomorphism $\phi : L^2(A)\to M$ which intertwines the right $A$-actions. Let $\kappa=\phi(\widehat{1})$.
Then $\kappa\in D(M_A)\subset D(K_A)$, $\phi=L(\kappa)$, and $L(\kappa)^*L(\kappa)=\langle \kappa|\kappa\rangle_A=1_A$.
If $\dim_{A-}({\sb{A}K})\geq 1$, then a similar argument finds a $\kappa$ such that ${\sb{A}\langle} \kappa,\kappa\rangle = 1_A$.
The last assertion follows from the fact that for $n\geq 2$, $H^n$ is the completion of the algebraic tensor product $D(H_A)\odot_A H^{n-1}$ with the inner product $\langle \eta_1 \odot \xi_1 , \eta_2\odot \xi_2\rangle = \langle \langle \eta_2|\eta_1\rangle_A \xi_1, \xi_2\rangle_{H^{n-1}}$, and similarly for left modules.
\end{proof}
\begin{prop}\label{prop:injective}
Suppose
$$
x=\a_{i_k}\cdots \a_{i_{1}}\a_{j_{1}}^*\cdots \a_{j_\ell}^*\in \sR(H)(n,n-\ell+k)
$$
is in the standard form of Proposition \ref{prop:Relations}. Then there are $\xi\in B^n$ and $\eta\in B^{n-\ell+k}$ such that $\langle x\xi,\eta\rangle = 1$ and $\langle y\xi,\eta\rangle=0$ for all words $y\in\sR(H)(n,n-\ell+k)$ on the $\a_i,\a_j^*$ whose standard form has length at least $\ell+k$.
\end{prop}
\begin{proof}
Choose $\kappa$ as in Lemma \ref{lem:Kappa}.
Let
\begin{itemize}
\item $\xi\in B^n$ be the simple relative tensor with $\zeta$'s in positions $j_1<\dots<j_\ell$ and $\kappa$'s in the other positions, and
\item $\eta\in B^{n-\ell+k}$ be the simple realtive tensor with $\zeta$'s in positions $i_1<\cdots <i_k$ and $\kappa$'s in the other positions.
\end{itemize}
Then by Lemma \ref{lem:Kappa},
$$
\langle x\xi,\eta\rangle=\|\underbrace{\kappa \otimes\cdots \otimes\kappa}_{n-\ell\text{ vectors}} \|_{H^{n-\ell}}^2=1.
$$
Suppose $y\in \sR(H)(n,n-\ell+k)$ is a word on the $\a_i,\a_j^*$ with $\langle y\xi,\eta\rangle\neq 0$, and write $y$ in standard form
$$
y= \a_{i'_{k'}}\cdots \a_{i'_{1}}\a_{j'_{1}}^*\cdots \a_{j'_{\ell'}}^*.
$$
Since $e_A(\kappa)=0$, we must have $i'_1,\cdots i'_{k'}\in\{i_1,\dots,i_k\}$ and $j'_1,\cdots j'_{\ell'}\in \{j_1,\dots, j_\ell\}$, so $k'\leq k$ and $\ell'\leq \ell$. Moreover, if $k'=k$ and $\ell'=\ell$, then $y=x$.
\end{proof}
\begin{cor}\label{cor:LinearlyIndependent}
The words on $\a_i,\a_j^*$ in standard form in $\sR(H)(m,n)$ are a basis.
\end{cor}
\begin{proof}
We already know such words span by Proposition \ref{prop:Relations}.
Suppose
$$
0=\sum_{i=1}^k \lambda_i w_i\in \sR(H)(m,n)
$$
where $w_i\in\sR(H)(m,n)$ are distinct words on the $\a_i,\a_j^*$ in standard form, ordered by increasing word length.
We show by induction on $k$ that all the $\lambda_i$'s are zero.
If $k=1$, this is trivial, since $w\neq 0$ for all words $w$ by Proposition \ref{prop:injective} (there is a linear functional which separates $w$ from $0$).
Suppose now that $k>1$. Since the standard form word length of $w_1$ is minimal, by Proposition \ref{prop:injective}, there are $\xi\in B^m$ and $\eta\in B^n$ such that $\langle w_i \xi,\eta\rangle = \delta_{1,i}$.
This means
$$
\lambda_1 =\sum_{i=1}^k \lambda_i \langle w_i \xi,\eta\rangle = \left \langle \sum_{i=1}^k \lambda_i w_i \xi,\eta\right\rangle =0.
$$
We are finished by the induction hypothesis.
\end{proof}
\begin{thm}\label{thm:GICARinRH}
The $*,\otimes$-functor $\Phi: \sR\sG \to \sR(H)$ given by $[n]\mapsto H^n$ for $n\geq 0$ and
$$
\sR\sG(n,n+1)\ni\alpha_i\longmapsto \a_i\in \sR(H)(n,n+1) \text{ for }1\leq i\leq n+1
$$
defines an equivalence of involutive tensor categories.
\end{thm}
\begin{proof}
By Proposition \ref{prop:Relations}, the relations of $\sR\sG$ are satisfied in $\sR(H)$, so $\Phi$ is well-defined.
By definition $\Phi$ preserves the adjoint, and it is easy to check that $\Phi$ preserves $\otimes$.
Since the words on $\alpha_i,\alpha_j^*$ in $\sR\sG(m,n)$ in standard form are a basis for $\sR\sG(m,n)$ by Proposition \ref{prop:Standard}, Corollary \ref{cor:LinearlyIndependent} shows that $\Phi$ is fully faithful and essentially surjective.
\end{proof}
\begin{rem}
The involutive tensor category $\sR\sG\cong\sR\sP$ is positive, i.e., if $x\in\sR\sG(m,n)$ and $x^*x=0\in\sR\sG(m,m)$, then $x=0$.
This can be shown using the standard form in Proposition \ref{prop:Standard}, or using positivity of $\sR(H)$ which comes for free.
If $x\in \sR\sG(m,n)$ with $x^*x=0$, then $\Phi(x^*x)=0$, so $\Phi(x)=0$ as $\sR(H)$ is positive. Hence $x=0$ as $\Phi$ is injective on hom spaces.
\end{rem}
\begin{rem}\label{rem:PlanarCalculus}
The planar calculus of \cite{MR3040370} is compatible with diagrams in $\sR\sP$. The value of a free-floating strand is
$$
\confetti =
\begin{cases}
\cwBrokenLoop &= \Tr_1(e_1) = \dim_{-A}(L^2(A))=1\\
\ccwBrokenLoop &= \Tr_1\op(e_1) = \dim_{A-}(L^2(A))=1,
\end{cases}
$$
and the value of the dotted closed oriented loops are
\begin{align*}
\cwDottedLoop &= \Tr_1(1-e_1)=\dim_{-A}(K) \text{ and }
\\
\ccwDottedLoop &=\Tr_1\op(1-e_1)=\dim_{A-}(K).
\end{align*}
Thus if $0\leq k\leq n$ and $q\in\sR\sP_n$ is a minimal projection with exactly $j$ dotted through strings, then
\begin{align*}
\Tr_n(\Phi\circ\Psi^{-1}(q))
&=
\dim_{-A}(K)^{j}
\\
\Tr_n\op(\Phi\circ\Psi^{-1}(q))
&=
\dim_{A-}(K)^{j} .
\end{align*}
\end{rem}
\subsection{Annular GICAR representations}
Let $H$ be as in Assumption \ref{assume:ContainsTrivial}.
Then $\cP_n\neq (0)$ for all $n\geq 0$, since it contains the vector $\zeta\otimes \cdots \otimes \zeta$.
\begin{defn}[{\cite[Section 4]{MR3040370}}]
A Hilbert $A-A$ bimodule $H$ is called \underline{extremal} if $\Tr_1=\Tr_1\op$ on $\cQ_1$.
A \underline{Burns rotation} is an operator $\rho:\cP_n\to\cP_n$ such that for all $\zeta\in\cP_n$ and $b_1,\dots, b_n\in B$, we have
$$
\langle \rho(\zeta) , b_1\otimes\cdots \otimes b_n\rangle = \langle \zeta, b_2\otimes\cdots \otimes b_n\otimes b_1\rangle.
$$
An \underline{opposite Burns rotation} is defined similarly:
$$
\langle \rho\op(\zeta) , b_1\otimes\cdots \otimes b_n\rangle = \langle \zeta, b_n\otimes b_1\otimes\cdots \otimes b_{n-1}\rangle.
$$
Note that if such a $\rho$ exists, then it is unique, and $\rho^n=\id_{\cP_n}$. In this case, $\rho^{-1}=\rho\op$.
\end{defn}
Recall the following theorems.
\begin{thm}[{\cite[Theorem 4.7]{MR3040370}}]
The following are equivalent:
\begin{enumerate}[(1)]
\item
$H$ is extremal.
\item
$H^n$ is extremal for all $n\geq 1$.
\item
$H^n$ is extremal for some $n\geq 1$.
\end{enumerate}
\end{thm}
\begin{thm}[{\cite[Theorems 4.11, 4.20, and 4.28]{MR3040370}}]
If $H$ is extremal, then the Burns rotation $\rho=\sum_\beta L_\beta R_\beta^*$ converges strongly on $\cP_n$ for all $n\geq 2$.
Moreover, $\rho^{-1}=\rho^*$ is given by the strongly convergent sum $\sum_\alpha R_\alpha L_\alpha^*$.
Conversely, if a unitary Burns rotation $\rho$ exists on $\cP_{2n}$ and $H$ is symmetric, then $H^n$ is extremal.
\end{thm}
We now impose the following assumption.
\begin{assume}\label{assume:Extremal}
Suppose $H$ is extremal, so that the Burns rotation $\rho=\sum_\beta L_\beta R_\beta^*$ converges strongly on $\cP_n$ for all $n\geq 2$.
\end{assume}
\begin{defn}
Given an $A-A$ bimodule $H$, we define the annular bimodule category $\sA(H)$ as the following small involutive category:
\itt{Objects} $\cP_n$ for $n\geq 0$.
\itt{Morphisms}
For $1\leq i\leq n+1$, the maps $\a_i: H^n\to H^{n+1}$ descend to maps $\cP_n \to \cP_{n+1}$ since they are $A-A$ bilinear, i.e.,
for all $x\in A$ and $\xi\in \cP_n$,
$$
x(\a_i(\xi))=\a_i(x\xi)= \a_i(\xi x)=(\a_i(\xi))x.
$$
A similar statement holds for the $\a_j^*$.
For $n=0$, let $\rho=\id_{L^2(A)}$, and for $n\geq 1$, let $\rho$ be the Burns rotation, which preserves $\cP_n$.
The morphisms of $\sA(H)$ are $\mathbb{C}$-linear combinations of all composites of the $\a_i,\a_j^*,\rho$.
\itt{Composition} composition of operators.
\itt{Adjoint} adjoint of operators.
\end{defn}
The main theorem of this subsection is as follows.
\begin{thm}\label{thm:GICARinP}
The $*$-functor $\Phi_{\sA}: \sA\sG \to \sA(H)$ given by $\alpha_i\mapsto \a_i$ and $\tau\mapsto \rho$ defines a $*$-representation.
\end{thm}
\begin{proof}
Note that Relations \eqref{rel:AG1} and \eqref{rel:AG2} automatically hold in $\sA(H)$ by Proposition \ref{prop:Relations}.
It remains to show Relations \eqref{rel:AG3} and \eqref{rel:AG4} hold. Since $\rho$ is periodic and unitary, Relation \eqref{rel:AG3} follows for $\Phi_{\sA}(\tau)=\rho$ immediately. Suppose $2\leq i\leq n$. Then for all $\xi\in \cP_n$ and $b_1,\dots, b_{n-1}\in B$, by the definition of the Burns rotation, we have
\begin{align*}
\langle \Phi_{\sA}(\alpha_i^*\tau)(\xi),b_1\otimes \cdots \otimes b_{n-1}\rangle
&=
\langle \Phi_{\sA}(\alpha_i)^*\Phi_{\sA}(\tau)(\xi),b_1\otimes \cdots \otimes b_{n-1}\rangle
\\
&=
\langle \rho(\xi),\a_i(b_1\otimes \cdots \otimes b_{n-1})\rangle
\\
&=
\langle \rho(\xi),b_1\otimes \cdots \otimes b_{i-1}\otimes \zeta \otimes b_i \otimes \cdots \otimes b_{n-1} \rangle
\\
&=
\langle \xi,b_2 \cdots \otimes b_{i-1}\otimes \zeta \otimes b_i \otimes \cdots \otimes b_{n-1}\otimes b_1 \rangle
\\
&=
\langle \xi,\a_{i-1}(b_2 \otimes\cdots \otimes b_{n-1}\otimes b_1) \rangle
\\
&=
\langle\Phi_{\sA}(\alpha_{i-1})^*( \xi),b_2 \otimes\cdots \otimes b_{n-1}\otimes b_1 \rangle
\\
&=
\langle\rho(\Phi_{\sA}(\alpha_{i-1}^*)(\xi)),b_1 \otimes\cdots \otimes b_{n-1}\rangle
\\
&=
\langle\Phi_{\sA}(\tau\alpha_{i-1}^*)( \xi),b_1 \otimes\cdots \otimes b_{n-1}\rangle.
\end{align*}
The relation $\alpha_i\tau=\tau\alpha_{i-1}$ is similar, and Relation \eqref{rel:AG4} holds.
\end{proof}
\begin{rem}
The representation of Theorem \ref{thm:GICARinP} is not necessarily faithful as we will see in Examples \ref{ex:SInfinity} and \ref{ex:OuterZAction2}.
\end{rem}
Just as every subfactor planar algebra decomposes as an orthogonal direct sum of irreducible annular Temperley-Lieb modules, so does the sequence of central $L^2$-vectors $(\cP_n)_{n\geq 0}$ under our $\sA\sG$-action afforded by Theorem \ref{thm:GICARinP}.
Just as the empty diagram generates an annular Temperley-Lieb module for a subfactor planar algebra \cite{MR1929335},
the vector $\widehat{1}\in \cP_0=A'\cap L^2(A)$ always generates an annular GICAR module.
However, this $\sA\sG$-module is trivial, since the only $\sA\sG$-consequence of $\widehat{1}\in\cP_0$ in $\cP_n$ is the $n$-fold tensor product of $\zeta$.
In stark comparison with finite index subfactors, it may be the case that $(\cP_n)_{n\geq 0}$ only consists of the trivial $\sA\sG$-module when $A_0\subset A_1$ is a non-trivial subfactor!
\begin{ex}[{\cite[Section 5]{MR3040370}}]\label{ex:SInfinity}
Consider $A_0=R\rtimes \Stab(1)\subset R\rtimes S_\infty=A_1$ where $\Stab(1)$ is the stabilizer of 1 under the action of $S_\infty$ on $\mathbb{N}$.
Then $\dim_\mathbb{C}(\cP_n)=1$ for all $n\geq 0$.
More precisely, for $n\geq 1$,
$$
\cP_n=A_0'\cap \bigotimes^n_{A_0} L^2(A_1)
=\spann\left\{
\underbrace{\widehat{1}\otimes \cdots \otimes \widehat{1}}_{n\text{ vectors}}
\right\}
\cong \mathbb{C} \,
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{.2}{};
\BrokenTopString{(-.4,0)};
\node at (.05,0) {$\dots$};
\BrokenTopString{(.4,0)};
\node at (.4,.6) {\scriptsize{$n$}};
\end{tikzpicture}\,.
$$
However, note that although $\dim_\mathbb{C}(\cQ_n)<\infty$ for all $n$, the dimension grows superfactorially, which is much faster than $\dim(G_n)=\sum_{k=0}^n {n\choose k}^2$. Thus $A_0\subset A_1$ does not have trivial standard invariant, i.e., it is not the infinite index analog of the Temperley-Lieb subfactors.
\end{ex}
\begin{ex}\label{ex:OuterZAction2}
Recall Example \ref{ex:OuterZAction}, i.e. $H=L^2(R)\oplus L^2(R)_{\sigma}$ for an outer action $\sigma:\mathbb{Z}\to \Out(R)$, where we denote $\sigma^n=\sigma(n)$.
In this case, when $n\geq 1$, $K^n=L^2(R)_{\sigma^n}$ has no central vectors, so
$$
\cP_n=A'\cap H^n=\spann\left\{
\underbrace{\zeta\otimes \cdots \otimes \zeta}_{n\text{ vectors}}
\right\}
\cong \mathbb{C} \,
\begin{tikzpicture}[baseline=-.1cm]
\nbox{}{(0,0)}{.4}{.2}{.2}{};
\BrokenTopString{(-.4,0)};
\node at (.05,0) {$\dots$};
\BrokenTopString{(.4,0)};
\node at (.4,.6) {\scriptsize{$n$}};
\end{tikzpicture}\,,
$$
where $\zeta$ is the image of $\widehat{1}\in L^2(R)$ inside $H$.
This bimodule has trivial standard invariant by Example \ref{ex:OuterZAction}, but it does not come from a {\rm II}$_1$-subfactor.
\end{ex}
\bibliographystyle{amsalpha}
{\footnotesize
|
\section{Introduction}
Studies of nuclear effects in minimum bias (MB)
\pPb\ collisions for charged particles~\cite{alice_RpA_new}, heavy
flavor and jets show that the observed strong suppression in
\PbPb\ collisions is due to final state effects. Centrality dependent
measurements of the nuclear modification factor require the
determination of the average number of binary collisions, \Ncoll, for
each centrality class. Moreover, it has been recognized that the study
of \pPb\ collisions is interesting on its own, with several
measurements~\cite{Abelev:2012ola,Abelev:2013bla,Abelev:2013haa,ABELEV:2013wsa}
of particle production in the low and intermediate \pt\ region that
can not be explained by an incoherent superposition of
\pp\ collisions, but rather call for coherent and collective effects.
To determine centrality in ALICE we use as many detectors as
possible~\cite{AlicePerf}, in various rapidity
regions~\cite{Alice:Centrality}. Particle production measured by
detectors at mid-rapidity can be modeled with a negative binomial
distribution (NBD), while the zero-degree energy measures the slow
nucleons emitted in the nucleus fragmentation process, which we model
with a Slow Nucleon Model (SNM) \cite{AToia:2014}. These models (NBD
and SNM) are coupled to a \pPb\ Glauber MC and \Ncoll\ are obtained
for each centrality class determined by slicing the experimental
distribution in percentiles of the hadronic cross-section. The
\Ncoll\ values are similar for different estimators within the
systematic error from the Glauber parameters and a MC closure test
with HIJING.
However, in order to use these \Ncoll\ values in a $\rpa (\pt , {\rm
cent}) = \frac{{\rm d}N^{\rm pPb}_{\rm cent}/{\rm d}\pt} { \langle
N_{\rm coll}^{\rm cent} \rangle {\rm d}N^{\rm pp}/ {\rm d}\pt}$
calculation, one needs to take into account the bias arising when
sampling the \pA\ events in centrality classes. In \pPb\ collisions,
the range of multiplicities used to select a centrality class is of
similar magnitude as the fluctuations, with the consequence that a
centrality selection based on multiplicity may select a biased sample
of nucleon-nucleon collisions. In essence, by selecting high (low)
multiplicity one chooses not only large (small) average \Npart\, but
also positive (negative) multiplicity fluctuations. These fluctuations
are partly related to qualitatively different types of collisions,
described in all recent Monte Carlo generators by fluctuations of the
number of particle sources via multi-parton interaction. Concerning
the nuclear modification factor other types of bias have been
discussed: the jet-veto effect, due to the trivial correlation between
the centrality estimator and the presence of a high-\pt\ particles in
the event; the geometric bias, resulting from the mean impact
parameter between nucleons rising for most peripheral events. Studies
of particle production and centrality determination have already been
presented by ALICE \cite{AToia:2014}; here we focus on the results
obtained with a new approach, described in the following sections.
\section{The hybrid method}\label{sec:hybrid}
The hybrid method aims at providing a data driven and
unbiased centrality determination. We give priority to a centrality
selection with minimal bias and, therefore, use the signal in the Zero
Degree Calorimeter (ZNA). In this case we cannot establish a direct
connection to the collision geometry but we can study the correlation
of two or more observables that are causally disconnected after the
collision, e.g because they are well separated in rapidity.
\begin{figure}
\begin{center}
\includegraphics*[width=0.495\textwidth]{scalingvsdNdeta_pre.eps}
\includegraphics*[width=0.35\textwidth]{fit_linear_pre.eps}
\caption{Left: Normalized signal from various observables versus the
normalized charged-particle density, fit with a lixsnear function of
\Npart. Right: Results from the fits as a function of the
pseudorapidity covered by the various observables. The red
horizontal lines indicate the ideal \Npart\ and \Ncoll\ geometrical
scalings. The most central point is excluded from the fit, to avoid
pile-up effects.}
\label{fig:hybrid}
\end{center}
\end{figure}
Charged particle multiplicity is dominated by soft particles while
hard processes are expected to scale with \Ncoll. In centrality
classes selected by ZNA, we study the dependence of various
observables in different $\eta$ and \pt\ regions on the charged
particle density at mid-rapidity. In order to compare different
observables on the same scale and also to neglect efficiency and
acceptance, we normalize the values in different classes by the
corresponding MB value. The correlation of the signals to the
mid-rapidity particle density (Fig.\ref{fig:hybrid} left) exhibits a
clear dependence on the rapidity. The slope of the signals with
\dNdeta\ decreases towards the proton direction (C-side). In the
Wounded Nucleon Model, \Npart\ is expressed in terms of target and
projectile participants. The particle density at mid-rapidity is
proportional to \Npart, whereas at higher rapidities the model
predicts a dependence on a linear combination of the number of target
and projectile participants with coefficients which depend on the
rapidity. Close to Pb-rapidity a linear wounded target nucleon scaling
(\Nparttar\ = \Npart\ - 1) is expected.
In order to further quantify the trends of the observables and to
relate them with geometrical quantities, such as \Npart, one can adopt
the WNM model and make the assumption that the charged particles
density at mid-rapidity is proportional to \Npart\ and relate the
other observables to \Npart\ assuming linear dependence, parameterized
with \Npart\ - $\alpha$. The results for $\alpha$
(Fig.\ref{fig:hybrid} right) indicate a monotonic change of the
scaling with rapidity. The red horizontal lines show the ideal
geometrical scalings. In Pb-going direction (negative $\eta_{\rm CMS}$
in the figure) the values of $\alpha$ reach the ones obtained for
charged-particle production at high-\pt. In contrast, in the
proton-going direction, $\alpha$ is much lower, indicating strong
suppression of the charged-particle production with centrality with
respect to \Npart-scaling. The correlation between the ZDC energy and
any variable in the central part shows unambiguously the connection of
these observables to geometry. Our data are overlaid with the
corresponding fit parameters derived from PHOBOS in d--Au collisions
at $\sqrt{s_{NN}}$ = 200 GeV. The comparison shows a good agreement
over a wide $\eta$ range, with some deviations at large negative
pseudorapidity. In particular, the $\eta$ region covered by the
innermost ring of the VZERO-A detector corresponds to the target
fragmentation region where extended longitudinal scaling was observed
at RHIC~\cite{Back:2004mr}.
Exploiting the findings from the correlation analysis described, we
make use of observables that are expected to scale as a linear
function of \Ncoll\ or \Npart, to calculate \Ncoll:
\begin{itemize}
\item $\Ncoll^{\rm mult}$: the multiplicity at mid-rapidity
proportional to the \Npart;
\item $\Ncoll^{\rm Pb-side}$: the target-going multiplicity
proportional to \Nparttar;
\item $\Ncoll^{\rm high-\pt}$: the yield of high-\pt particles at
mid-rapidity is proportional to \Ncoll.
\end{itemize}
These scalings can be used as an ansatz to calculate \Ncoll, rescaling
the MB value $\Ncoll^{\rm MB}$ by the ratio of the normalized signals
to the MB one. We therefore obtain 3 sets of values of \Ncoll, whose relative
difference does not exceed 10\%. This confirms the consistency of the
used assumptions, although it does not constitute a proof that any or
all of the assumptions are valid.
\section{Physics Results}
\subsection{Nuclear Modification Factors}
As already discussed in \cite{AToia:2014}, the \qpa\ calculated with
multiplicity based estimator (shown in Fig.\ref{fig:QpAhybrid} for
CL1, where centrality is based on the tracklets measured in
$|\eta|<1.4$) widely spread between centrality classes. They also exhibit a
negative slope in \pt, mostly in periphearl events, due to the jet
veto bias, as jet contribution increases with \pt. The \qpa\ compared
to G-PYTHIA, a toy MC which couples Pythia to a p-Pb Glauber MC, show
a good agreement, everywhere in 80-100\%, and in general at high-\pt,
demonstrating that the proper scaling for high-\pt\ particle
production is an incoherent superposition of \pp\ collisions, provided
that the biased introduced by the centrality selection is properly
taken into account,as eg in G-PYTHIA. For ZNA centrality classes,
while no bias is expected on the multiplicity or high-Q$^2$ processes
is expected and indeed the classes present spectra which are much more
similar to each other, the absolute values of the spectra at high-\pt
indicate the presence of a bias, not due to the event selection but
because of inaccurate \Ncoll\ values calculated with the SNM.
With the hybrid method, using either the assumption on mid-rapidity
multiplicity proportional to \Npart, or forward multiplicity
proportional to \Nparttar, shown in Fig.~\ref{fig:QpAhybrid}, result in
consistent \qpa, also consistent with one for all centrality classes,
also observed for MB collisions,
indicating the absence of initial state effects. The observed Cronin
enhancement is stronger in central collisions and nearly absent in
peripheral collisions. The enhancement is also weaker at LHC energies
compared to RHIC energies.
\begin{figure}[t!f]
\centering
\includegraphics[width=0.4\textwidth]{QpPbCL1GlauberPythiarebinned.eps}
\includegraphics[width=0.4\textwidth]{HybNpartTargQpPb.eps}
\caption{\qpa\ calculated with CL1 estimator (left), the lines are
the G-PYTHIA calculations; with the hybrid method (right), spectra
are measured in ZNA-classes and \Ncoll\ are obtained with the
assumption that forward multiplicity is proportional to \Nparttar.
\label{fig:QpAhybrid}}
\end{figure}
\subsection{Charged particle density}
Charged particle density is also studied as a function of $\eta$, for
different centrality classes, with different estimators. In peripheral
collisions the shape of the distribution is almost fully symmetric and
resembles what is seen in proton-proton collisions, while in central
it becomes progressively more asymmetric, with an increasing excess of
particles produced in the direction of the Pb beam. We have quantified
the evolution plotting the asymmetry between the proton and lead peak
regions, as a function of the yield around the center of mass (see
Fig.~\ref{fig:dndetaNpart2}, left): the increase of the asymmetry is
different for the different estimators. Fig.~\ref{fig:dndetaNpart2}
right shows \Nch\ at mid-rapidity divided by \Npart\ as a function of
\Npart\ for various centrality estimators. For Multiplicity-based
estimators (CL1, V0M, V0A) the charged particle density at mid
rapidity increases more than linearly, as a consequence of the strong
multiplicity bias. This trend is absent when \Npart\ are calculated
with the Glauber-Gribov model, which shows a relatively constant
behavior, with the exception of the most peripheral point. For ZNA,
there is a clear sign of saturation above \Npart\ = 10, due to the
saturation of forward neutron emission. None of these curves points
towards the \pp\ data point. In contrast, the results obtained with
the hybrid method, using either \Nparttar-scaling at forward rapidity
or \Ncoll-scaling for high-\pt\ particles give very similar trends,
and show a nearly perfect scaling with \Npart, which naturally reaches
the \pp\ point. This indicates the sensitivity of the \Npart-scaling
behavior to the Glauber modeling, and the importance of the
fluctuations of the nucleon-nucleon collisions themselves.
\begin{figure}[t!f]
\centering
\includegraphics[width=0.55\textwidth]{Nch_Asym_pre.eps}
\includegraphics[width=0.4\textwidth]{NchNpart1_pre.eps}
\caption{Left: Asymmetry of particle yield, as a function of the
pseudorapidity density at mid-rapidity for various centrality
classes and estimators. Right: Pseudorapidity density of charged
particles at mid-rapidity per participant as a function of
\Npart\ for various centrality estimators.
\label{fig:dndetaNpart2}}
\end{figure}
\section{Conclusions}
Multiplicity Estimators induce a bias on the hardness of the pN
collisions. When using them to calculate centrality-dependent \qpa,
one must include the full dynamical bias. However, using the
centrality from the ZNA estimator and our assumptions on particle
scaling, an approximate independence of the multiplicity measured at
mid-rapitity on the number of participating nucleons is observed.
Furthermore, at high-\pt\ the \pPb\ spectra are found to be consistent
with the pp spectra scaled by the number of binary nucleon--nucleon
collisions for all centrality classes. Our findings put strong
constraints on the description of particle production in high-energy
nuclear collisions.
|
\section{Introduction}
\begin{figure}
\centering
\includegraphics[bb=0 45 595 667,width=0.73\columnwidth ,clip]{j1104ua.eps} \\
\includegraphics[bb=0 45 595 693,width=0.73\columnwidth ,clip]{j1104kb.eps} \\
\includegraphics[bb=0 45 595 693,width=0.73\columnwidth ,clip]{j1104qd.eps} \\
\caption{Images of Mrk\,421 for the first observing epoch at 15\,GHz (top image), the second observing epoch at 24\,GHz (central image), and the third observing epoch at 43\,GHz (bottom image). Levels are drawn at $(-1, 1, 2, 4...) \times$ the lowest contour (that is, at 1.0 mJy/beam for 15 and 24\,GHz images and at 0.65 mJy/beam for the 43\,GHz image) increasing by factors of 2. The restoring beam, shown in the bottom left corner, has a value of 1.05~mas $\times$ 0.66~mas, 0.67~mas $\times$ 0.40~mas, and 0.33~mas $\times$ 0.19~mas for 15, 24, and 43\,GHz, respectively. The overlaid color maps show the linearly polarized intensity, and bars represent the absolute orientation of the EVPAs.}
\label{maps}
\end{figure}
In the family of active galactic nuclei (AGN), blazars are the most powerful objects. Their relativistic jets are closely aligned with the line of sight. For this particular geometry, we observe that the radiation originates in the jet that points in our direction because of Doppler boosting effects. The emission in these objects is dominated by nonthermal radiation from relativistic electrons interacting with the magnetic field.
Mrk\,421 is one of the nearest \citep[$z=0.03$,][]{deVaucouleurs1991} and brightest blazars and is therefore suitable for probing and investigating the physics of the innermost regions of relativistic jets. For this reason, Mrk\,421 has been intensively studied throughout the electromagnetic spectrum, especially since it was detected at TeV energies in 1992 \citep{Punch1992}.
In general, the spectral energy distribution (SED) of blazars is dominated by the emission of the jet and consists of two separate components: a low-frequency hump, due to synchrotron emission by relativistic electrons within the jet, and a high-energy hump, that is commonly assumed to be inverse Compton (IC) scattering. In BL Lacs the high-energy IC component is commonly interpreted as synchrotron-self-Compton emission \citep[SSC, see][]{Abdo2011,Tavecchio2001}, resulting from the upscatter of the synchrotron photons off the jet electrons. For Mrk\,421, as for most TeV blazars \citep{Piner2013, Tiet2012}, the synchrotron hump peaks at soft X-rays, and for this reason, it is classified as a high-synchrotron-peaked (HSP) blazar \citep{Abdo2011}.
Since the emission from these objects is dominated by nonthermal radiation, studying their polarization properties can provide important information on the magnetic field structure and the emission mechanisms. Furthermore, thanks to the multi-wavelength observations, we can investigate the location where the radiation is produced.
We observed Mrk\,421 with the Very Long Baseline Array (VLBA) at 15, 24, and 43\,GHz, both in total and polarized intensity. These datasets belong to a multi-frequency campaign, that was carried out during 2011, which also involved observations at submm (SMA), cm (e.g., F-GAMMA, Medicina), optical/IR (GASP), UV/X-ray ({\it Swift}, RXTE, MAXI), and $\gamma$-ray ({\it Fermi}, MAGIC, VERITAS) wavelengths.
In two previous works we presented the complete analysis of observations in total intensity at 15 and 24\,GHz \citep{Lico2012} and at 43\,GHz \citep{Blasi2013}. We constrained some physical parameters such as the Doppler factor (in the radio emission region we found $\delta_r \sim 3$, while in the high-energy emission region we found $\delta_{\rm h.e.} \sim 14$), the viewing angle ($2^\circ < \theta < 5^\circ$), apparent speeds (no significant motion was detected in the jet), and the brightness temperature ($T_{\rm B,var} \sim 2.1\times 10^{10}$~K).
In this work we present an analysis of the observations in linearly polarized intensity, which allows us to determine some physical parameters such as the degree of polarization and the absolute orientation of the electric vector position angle (EVPA), and to obtain some useful information on the magnetic field topology. We also used the data collected by the Large Area Telescope (LAT) onboard the {\em Fermi} satellite to produce $\gamma$-ray light curves for all of 2011.
This paper is structured as follows: in Section 2 we describe the VLBA and {\em Fermi} observations and analysis, and we introduce the methods used for determining the EVPA orientation. In Section 3 we report the results of this work, and in Section 4 we discuss these results and interpret them in the astrophysical context. Throughout the paper we use the following conventions for cosmological parameters: $H_0=70$ km sec$^{-1}$ Mpc$^{-1}$, $\Omega_M=0.25$ and $\Omega_\Lambda=0.75$, in a flat Universe.
All angles are measured from north through east. At the redshift of the target, 1 mas corresponds to 0.59 pc.
\begin{table*}
\caption{Final EVPA rotations for 15, 24, and 43\,GHz (Cols. 3, 6, and 9, respectively); numbers in boldface refer to the comparison with JVLA values. We also report the relative rotations, obtained by comparing antenna tables for consecutive epochs, and the reference antennas used for the phase calibration.}
\label{tabrotations}
\centering
\tiny
\begin{tabular}{ccccccccccc}
\hline
\hline
Epoch & MJD & Final $\Delta$\tablefootmark{a} & $\Delta$D-terms\tablefootmark{b} & Reference & Final $\Delta$\tablefootmark{a} & $\Delta$D-terms\tablefootmark{b} & Reference & Final $\Delta$\tablefootmark{a} & $\Delta$D-terms\tablefootmark{b} & Reference\tablefootmark{c} \\
year/month/day & & (deg) & (deg) & antenna & (deg) & (deg) & antenna & (deg) & (deg) & antenna \\
\hline
& & $15\,GHz$ & & & $24\,GHz$ & & & $43\,GHz$ & & \\
\hline
2011/01/14 & 55575 & $-$21.7 & & PT & $-$8.5 & & PT & $-$3.2 & & PT \\
2011/02/25 & 55617 & $-$21.7 & 0 & PT & $-$8.5 & 0 & PT & $-$3.2 & 0 & PT \\
2011/03/29 & 55649 & \textbf{$-$21.7}& 0 & PT & \textbf{$-$8.5} & 0 & PT & \textbf{$-$3.2} & 0 & PT \\
2011/04/25 & 55675 & $-$21.7 & 0 & PT & $-$8.5 & 0 & PT & $-$3.2 & 0 & PT \\
2011/05/31 & 55712 & $-$21.7 & 0 & PT & $-$8.5 & 0 & PT & $-$3.2 & 0 & PT \\
2011/06/29 & 55741 & \textbf{22.2} & 45 & OV & $-$91 & $-$82.5 & OV & $-$25.2 & $-$22 & KP\\
2011/07/28 & 55770 & 22.2 & 0 & OV & $-$153.5 & $-$62.5 & KP & $-$25.2 & 0 & KP\\
2011/08/29 & 55802 & 85.2 & 63 & KP & $-$153.5 & 0 & KP & $-$25.2 & 0 & KP\\
2011/09/28 & 55832 & 157.2 & 72 & PT & $-$6.5 & $-$33 & PT & 6.8 & 32 & PT\\
2011/10/29 & 55863 & 157.2 & 0 & PT & $-$6.5 & 0 & PT & 6.8 & 0 & PT\\
2011/11/28 & 55893 & \textbf{25.5} & 45 & OV & \textbf{$-$79.7}& $-$67.7 & OV & \textbf{$-$49.2} & $-$51 & OV\\
2011/12/23 & 55918 & $-$19.5 &-45 & PT & $-$9.7 & $-$110 & PT & 0.8 & 50 & KP\\
\hline
\hline\\
\end{tabular}
\tablefoot{
\begin{tiny}
\newline
\tablefoottext{a}{Final rotation to apply to obtain the absolute EVPA orientation.}\\
\tablefoottext{b}{Relative rotation obtained by comparing antenna table of two consecutive epochs.}\\
\tablefoottext{c}{PT = Pie Town, KP = Kitt Peak, OV = Owens Valley.}\\
\end{tiny}
}
\end{table*}
\begin{figure
\includegraphics[bb=77 360 539 720, width=1.0\columnwidth ,clip]{Core_15GHz.ps} \\
\includegraphics[bb=77 360 539 720, width=1.0\columnwidth ,clip]{Core_24GHz.ps} \\
\includegraphics[bb=77 360 539 720, width=1.0\columnwidth ,clip]{Core_43GHz.ps} \\
\caption{Evolution with time of some physical parameters of Mrk\,421 in the core region at 15\,GHz (upper frame), 24\,GHz (middle frame), and 43\,GHz (lower frame). For each frame we report from the first to the fourth panel the light curves for the total intensity, the polarized flux density, the fractional polarization, and the EVPA values. Triangles in the lower frame (first panel) represent the VLBA 43\,GHz observations of Mrk\,421 provided by the VLBA-BU-BLAZAR program.}
\label{plots_core}
\end{figure}
\section{Observations and data reduction}
\subsection{VLBA data: details and analysis}
We observed Mrk\,421 with the VLBA once per month throughout 2011 at three frequencies: 15, 24, and 43\,GHz. Observations were carried out in total and polarized intensity (in right and left circular polarization). For the details of the observations at 15\,GHz and 24\,GHz, see \citet{Lico2012}, and for the 43\,GHz observations, see \citet{Blasi2013}. Our 43\,GHz dataset has been expanded by adding data from 11 epochs provided by the VLBA-BU-BLAZAR program\footnote{\url{http://www.bu.edu/blazars/VLBAproject.html}} (see Table~\ref{bostondata}). We updated the total intensity flux densities here with respect to our previous works: for the main campaign datasets, we made a new and independent calibration (for the sake of a homogeneous polarization calibration, see below), which resulted in new values entirely consistent with the previous ones within the uncertainties; for the supplementary Boston University datasets, the flux density scale was adjusted by a scaling factor obtained from comparing of the flux densities between the VLBA and the single-dish Mets\"ahovi near-simultaneous datasets, resulting in a significantly improved accuracy of the flux density scale.
For the calibration, the fringe-fitting procedure, and the detection of cross-polarized fringes we used the software package Astronomical Image Processing System (AIPS) \citep{Greisen2003}. We produced the cleaned and final images with the software package DIFMAP \citep{Shepherd1997}. The polarization D-terms were determined with the task LPCAL in AIPS. We compared antenna tables (which provide the phase and amplitude values for the R and L circular polarization for each antenna) at two consecutive epochs to determine the relative EVPA rotations using the IDL routines developed by J.~L.~G\'omez.
We calibrated the instrumental polarization using the strong (flux density $> 1$ Jy) and structureless source J1310+3220. This source also provides good coverage of the parallactic angle ($>100^\circ$) and has negligible polarization on large scales, which is important because of the very different angular resolution of the VLBA with respect to the Karl G. Jansky Very Large Array (JVLA).
\subsection{Polarization calibration}
\label{methods}
In general, to obtain the absolute orientation of the EVPAs (defined as $\chi =0.5 \times \arctan(U/Q)$, where $U$ and $Q$ are Stokes flux densities) in very long baseline interferometric (VLBI) observations, a comparison with quasi-simultaneous single-dish or JVLA observation is required. This is because of the lack of polarization calibrators with stable EVPAs on (sub)milliarcsecond scales.
To determine the EVPA absolute orientation, we used the method developed by \citet{Leppanen1995}, which makes use of the instrumental polarization parameters (the so-called D-terms). This method provides us with an independent way of calibrating the absolute right-left (R-L) circular polarization phase offset; it is based on the assumption that the D-terms change slowly with time \citep[see][]{Gomez2002}. For most of the antennas we found that D-terms remain stable during the whole 12-month observing period.
The method consists of comparing the D-terms for each antenna in consecutive epochs, which yields the relative rotation in right (R) and left (L) circular polarization. In other words, the phase offset in R and L between two epochs is provided by the phase difference of the D-terms.
We note that since the VLBA consists of ten antennas, twenty different values are involved in the comparison to determine the relative rotation between two epochs, (two R-L values for each antenna). This guarantees an accurate and reliable determination of the phase offset in case of an antenna failure.
The relative rotations obtained for 15, 24, and 43\,GHz are reported in Cols. 4, 7, and 10 in Table~\ref{tabrotations}. For a specific epoch the relative rotation may differ at different frequencies because it depends on the reference antenna used for the calibration; the reference antennas used at the different frequencies are listed in Cols. 5, 8, and 11 in the same table.
After determining the relative rotations, we set the absolute EVPA calibration for one epoch by comparison with a JVLA observation provided by the POLCAL program\footnote{\url{http://www.aoc.nrao.edu/~smyers/evlapolcal/polcal_master.html}} (values in boldface in Table~\ref{tabrotations}). Then we determined the absolute orientation for all the EVPAs by applying the relative rotations obtained from the D-terms. For example, after fixing a value in the third column for the 15\,GHz data by the comparison with JVLA, we obtained the other values by summing the previous value in the same column and the relative rotation for the same epoch in Col. 4. For example, at 15\,GHz in Col. 3 in Table~\ref{tabrotations} for the 11th observing epoch (November 2011) we obtain the value of $25.5^\circ$ by the comparison with JVLA. Then, to obtain the absolute rotation for the consecutive epoch (December 2011), we just add the relative rotation to the value of $25.5^\circ$ (fourth column), which in this case is $-45^\circ$, and we obtain a final rotation of $-19.5^\circ$. The relative rotation between two epochs is $0^\circ$ for the same reference antenna at both epochs.
At 15\,GHz we have three JVLA measurements taken during 2011, which enabled two cross checks. The values obtained with the D-terms method and those from the comparison with the JVLA agree within $5^\circ$. For the 24 and 43\,GHz observations we have two JVLA measurements, which we also cross-checked, finding very good agreement within $5^\circ$. This confirms the validity and accuracy of the D-terms calibration method reported previously by \citet{Gomez2002}.
\subsection{Determination of uncertainties}
Error bars for the total intensity flux density ($S$) and the linearly polarized emission (defined as $P=\sqrt{Q^2+U^2}$) were calculated by considering a calibration uncertainty $\sigma_c$ of about $10\%$ of the flux density and a statistical error provided by the map rms noise.
To determine the statistical error for the jet emission we also took the number of beams into account:
\begin{equation}
\label{err_flux}
\Delta_S= \sqrt{\sigma_c^2 + \left(\sqrt{\frac{\mbox{box size}}{\mbox{beam size}}} \times rms\right)^2}.
\end{equation}
The box size term is defined in Sect.~\ref{morphology}.
The uncertainties in fractional polarization (defined as $m=P/S$) were calculated from error propagation theory:
\begin{equation}
\Delta m = \frac{1}{S} \sqrt{\sigma_P^2 + \left(\frac{P}{S} \times \sigma_S\right)^2},
\end{equation}
where $\sigma_S$ and $\sigma_P$ represent the uncertainties in $S$ and $P$.
The uncertainties in EVPA values were calculated by taking into account all of these contributions:
\begin{equation}
\Delta \chi = \sqrt{\sigma_\mathrm{cal}^2+\sigma_\mathrm{D-terms}^2+\sigma_\mathrm{JVLA}^2+\sigma_{\chi}^2},
\end{equation}
where $\sigma_\mathrm{cal}$ is the scatter of the value measured on the polarization map, $\sigma_\mathrm{D-terms}$ is the calibration uncertainty introduced when we compare D-terms at different epochs to obtain the relative rotations by using the antenna tables, and $\sigma_\mathrm{JVLA}$ is the $5^\circ$ mean difference between the JVLA and the D-terms methods. The last term $\sigma_{\chi}$ is the uncertainty in $\chi$ calculated from error propagation theory:
\begin{equation}
\sigma_{\chi} = \frac{0.5}{Q^2+U^2} \sqrt{U^2 \Delta_{Q}^2 + Q^2 \Delta_{U}^2},
\end{equation}
where $\Delta_Q$ and $\Delta_U$ are the uncertainties in the $Q$ and $U$ Stokes flux densities, which are calculated using Eq.~(\ref{err_flux}).
Since $\Delta_Q \sim \Delta_U$ \citep{fanti2001}, the formula becomes
\begin{equation}
\sigma_{\chi} = \frac{0.5 \times \Delta_Q }{\sqrt{Q^2+U^2}} = 0.5 \times \frac{\Delta_Q }{P}.
\end{equation}
To determine the polarization parameters we did not include the random noise correction \citep{wardle1974}; this contribution is always within the uncertainties.
\subsection{{\em Fermi}-LAT data: selection and analysis}
The {\em Fermi}-LAT is a pair-conversion telescope operating from 20 MeV to $>$ 300 GeV. Further details about the {\em Fermi}-LAT are given in \citet{Atwood2009}. The LAT data reported here were collected from 2011 January 1 (MJD 55562) to December 31 (MJD 55926). During this time, the {\em Fermi} observatory operated almost entirely in survey mode. The analysis was performed with the software package \texttt{ScienceTools} version v9r32p5. The LAT data were extracted within a region of $20^{\circ}$ radius centered on the location of Mrk\,421. Only events belonging to the `Source' class were used. The time intervals collected when the rocking angle of the LAT was greater than 52$^{\circ}$ were rejected. In addition, a cut on the zenith angle ($<100^{\circ}$) was applied to reduce contamination from the Earth limb $\gamma$ rays, which are produced by cosmic rays interacting with the upper atmosphere.
The spectral analysis was performed with the instrument response functions \texttt{P7REP\_SOURCE\_V15} using an unbinned maximum-likelihood method implemented in the Science tool \texttt{gtlike}. A Galactic diffuse emission model and isotropic component, which is the sum of an extragalactic and residual cosmic-ray background, were used to model the background\footnote{\url{http://fermi.gsfc.nasa.gov/ssc/data/access/lat/BackgroundModels.html}}.
The normalizations of the two components in the background model were allowed to vary freely during the spectral fitting.
\begin{table}
\caption{VLBA 43\,GHz data provided by the Boston University blazar monitoring program.}
\label{bostondata}
\centering
\tiny
\begin{tabular}{c c c c}
\hline\hline
Epoch & MJD & S\tablefootmark{a} & $\sigma_S$\tablefootmark{b} \\
year/month/day & & (mas) & (mas) \\
\hline
2011/01/02 & 55563 & 328.0 & 21.9 \\
2011/02/04 & 55597 & 355.3 & 23.7 \\
2011/03/01 & 55621 & 415.9 & 27.7 \\
2011/04/21 & 55673 & 324.6 & 21.6 \\
2011/05/22 & 55703 & 275.0 & 18.3 \\
2011/06/12 & 55724 & 223.6 & 14.9 \\
2011/07/21 & 55763 & 177.0 & 11.8 \\
2011/08/23 & 55796 & 220.3 & 14.7 \\
2011/09/16 & 55820 & 162.7 & 10.8 \\
2011/10/16 & 55850 & 189.0 & 12.6 \\
2011/12/02 & 55897 & 199.6 & 13.3 \\
\hline
\hline
\end{tabular}
\tablefoot{
\begin{tiny}
\newline
\tablefoottext{a}{Flux density in mJy.}\\
\tablefoottext{b}{Estimated errors for the flux density.}\\
\end{tiny}
}
\end{table}
We analyzed a region of interest of $10^{\circ}$ radius centered on the location of Mrk\,421.
We evaluated the significance of the $\gamma$-ray signal from the sources by means of the maximum-likelihood test statistic TS = 2$\Delta$log(likelihood) between models with and without a point source at the position of Mrk\,421 \citep{Mattox1996}. The source model used in \texttt{gtlike} includes all of the point sources from the second {\em Fermi}-LAT catalog \citep[2FGL;][]{Nolan2012} as well as in a preliminary third {\em Fermi}-LAT catalog (Ackermann et al., in prep.) that fall within $15^{\circ}$ of the source.
The spectra of these sources were parametrized by power-law functions, except for 2FGL\,J1015.1+4925, for which we used a log-parabola as in the 2FGL catalog. A first maximum-likelihood analysis was performed to remove from the model the sources with TS $<$ 25 and/or the predicted number of counts based on the fitted model $N_{pred} < 3$. A second maximum-likelihood analysis was performed on the updated source model. In the fitting procedure, the normalization factors and the photon indices of the sources lying within 10$^{\circ}$ of Mrk\,421 were left as free parameters. For the sources located between 10$^{\circ}$ and 15$^{\circ}$ from our target, we kept the normalization and the photon index fixed to the values from the 2FGL catalog.
The systematic uncertainty on the effective area\footnote{\url{http://fermi.gsfc.nasa.gov/ssc/data/analysis/LAT_caveats.html}} is $10\%$ below 100 MeV \citep{Ackermann2012}, decreasing linearly with the logarithm of energy to $5\%$ between 316 MeV and 10 GeV, and increasing linearly with the logarithm of energy up to $15\%$ at 1 TeV.
\begin{figure}
\includegraphics[bb=77 360 539 720, width=1.0\columnwidth ,clip]{JET_15GHz.ps} \\
\includegraphics[bb=77 360 539 720, width=1.0\columnwidth ,clip]{JET_24GHz.ps} \\
\caption{Evolution with time of some physical parameters in the jet region at 15\,GHz (upper frame) and at 24\,GHz (lower frame). For each frame we report from the first to the fourth panel the light curves for the total intensity, the polarized flux density, the fractional polarization, and the EVPA values.}
\label{plots_jet}
\end{figure}
\begin{figure*}
\includegraphics[width=0.67\columnwidth]{RM_core_Jan_deg.ps}
\includegraphics[width=0.67\columnwidth]{RM_core_Feb_deg.ps}
\includegraphics[width=0.67\columnwidth]{RM_core_Mar_deg.ps} \\
\includegraphics[width=0.67\columnwidth]{RM_core_Apr_deg.ps}
\includegraphics[width=0.67\columnwidth]{RM_core_May_deg.ps}
\includegraphics[width=0.67\columnwidth]{RM_core_Jun_deg.ps} \\
\includegraphics[width=0.67\columnwidth]{RM_core_Jul.ps}
\includegraphics[width=0.67\columnwidth]{RM_core_Aug_deg.ps}
\includegraphics[width=0.67\columnwidth]{RM_core_Sep.ps} \\
\includegraphics[width=0.67\columnwidth]{RM_core_Oct_deg.ps}
\includegraphics[width=0.67\columnwidth]{RM_core_Nov_deg.ps}
\includegraphics[width=0.67\columnwidth]{RM_core_Dec_deg.ps} \\
\caption{EVPAs versus $\lambda^2$ linear fits for all the observing epochs. We also show the 15\,GHz EVPAs with empty symbols as measured in the maps, i.e.\ before the rotation of $90^\circ$ applied for July and September to account for opacity effects.}
\label{rm_core}
\end{figure*}
\section{Results}
\subsection{Images and morphology}
\label{morphology}
In Fig.~\ref{maps} we show a sample of three polarization images of Mrk\,421 at 15\,GHz (upper panel), 24\,GHz (middle panel), and 43\,GHz (lower panel) produced with DIFMAP and IDL\footnote{\url{http://www.exelisvis.com/ProductsServices/IDL.aspx}}. To improve the sensitivity to the extended jet emission, we restored the images with natural weighting. The contours show the total intensity, the overlaid color maps show the linearly polarized intensity. Bars represent the absolute orientation of the EVPAs.
We note that the position of the peak in the linearly polarized and total intensity emission images do not always coincide. For example, in the first observing epoch at 15\,GHz (upper panel in Fig.~\ref{maps}), the peak is 4.93 mJy/beam, and it is not coincident with the total intensity peak, instead it lies in the jet about $1$ mas from the core region.
From the total intensity images at all three frequencies, we clearly detect a well-defined and collimated one-sided jet structure, emerging from a compact nuclear region that extends about 4.5 mas (2.7 pc), with a position angle (PA) of about $-35^\circ$. This agrees well with the result of \citet{Giroletti2004a}.
At 15\,GHz the linearly polarized emission extends for about 1 mas from the core region, allowing us to distinguish core and jet emission. The outer polarized emission is too faint to be detected. The polarized emission in the jet becomes fainter at higher frequencies: at 24\,GHz it is still clearly detected, but at 43\,GHz we only detect polarized emission from the core region; this is probably due to sensitivity limitations.
In our previous works we represented the total intensity jet structure with Gaussian components. However, this approach is not reliable in the case of polarized emission. We instead estimated the jet polarized flux density as the difference between the total amount and the contribution of the core region. In practice, we first measured the total polarized flux density $P_\mathrm{tot}$ at each epoch by setting a box containing the entire polarized region; we adjusted the size of the box depending on the extension of the polarized emission. We then determined the core contribution $P_\mathrm{core}$ from the value of the polarized flux density at the position of the total intensity peak. Finally, we estimated the jet polarized flux density as $P_\mathrm{jet}=P_\mathrm{tot}-P_\mathrm{core}\times\cos(\chi_\mathrm{core}-\chi_\mathrm{jet})$. We also determined the jet EVPA directly on the image at the location where the jet polarized flux density is highest.
The 43\,GHz images reveal a transverse structure in the inner part of the jet in the form of limb brightening in the polarized emission. This transverse structure is clearly revealed in the April 2011 polarization image (see lower panel in Fig.~\ref{maps}). During this epoch the polarization emission peak is in the core region and is $\sim8$ mJy/beam, while in the limbs the polarization peak is $\sim2.3$ mJy/beam. The limb-brightening structure is also detected in March and May 2011, but it is less pronounced, while it is absent from all the other epochs.
\begin{figure}
\includegraphics[width=1.0\columnwidth]{RM+chi.ps} \\
\caption{Upper panel: time evolution for the RM values derived for the core region using 15, 24, and 43\,GHz data. Lower panel: time evolution for the intrinsic values of the polarization angle obtained from the $\lambda^2$ fits.}
\label{rm}
\end{figure}
\subsection{Radio light curves and evolution of polarization angle}
In Fig.~\ref{plots_core}, we show the light curves for the core region of Mrk\,421 during 2011 at 15\,GHz (upper frame), 24\,GHz (middle frame), and 43\,GHz (lower frame). For each frame, we show in four panels (from top to bottom) the total intensity and the polarized flux density, the fractional polarization, and the EVPAs.
The plots for the jet region are shown in Fig.~\ref{plots_jet} at 15\,GHz (upper frame) and 24\,GHz (lower frame). No polarized emission from the jet is apparent in the 43\,GHz images. All of these values are reported in Table~\ref{table_data}.
\begin{figure*}
\includegraphics[width=1.0 \columnwidth]{Mrk421_LAT.ps} %
\includegraphics[width=1.0\columnwidth]{flux_index.ps}
\caption{0.1-100 GeV flux in units of 10$^{-8}$ ph cm$^{-2}$ s$^{-1}$ (top panel) and the spectral photon index from a power-law fit (bottom panel) for Mrk\,421 for time intervals on one week during 2011 (from MJD 55562 to MJD 55926). The dashed line in both panels represents the mean value. In the right frame we show the scatter plot of the photon index vs the $\gamma$-ray flux.}
\label{fermi_plot}
\end{figure*}
\subsubsection{Core region}
As we noted in our previous works \citep{Lico2012, Blasi2013}, for all the three frequencies the total intensity light curves peak around the end of February (MJD $\sim$55617), with a decrease until July 28 (MJD 55770) followed by a slight increase in the last part of the year.
In the core region the linearly polarized emission at 15\,GHz and 24\,GHz is significantly variable ($>3\sigma$ difference between the highest and the lowest value), but we observe the most significant variation at 43\,GHz, where we clearly detect a peak of $12.5$ mJy/beam during the third observing epoch (MJD $\sim$55649). This behavior may be connected with the enhanced activity in the $\gamma$-ray light curve between MJD 55562 and MJD 55660, as described in Sect.~\ref{fermi_sec}.
The core polarization fraction has a mean value of $1\%$, in agreement with other studies of this source \citep[e.g.,][]{Marscher2002, Pollack2003}. At 43\,GHz the polarization percentage is higher, around $2\%$, reaching $\sim4\%$ during the third observing epoch, which is close in time with the enhanced activity at high energy.
In the bottom panel of each frame of Fig.~\ref{plots_core}, we also show the trend of the polarization angle for the core region of Mrk\,421 during 2011 at 15\,GHz (upper frame), 24\,GHz (middle frame), and 43\,GHz (lower frame). The EVPAs vary. For most of the year, they show values varying between $110^\circ$ and $150^\circ$. In between some pairs of epochs, in particular at 15\,GHz, they change by about $90^\circ$, which has no clear connection with the EVPA variation at 24 and 43\,GHz or with the polarized flux density trend.
\begin{table}
\caption{Rotation measure and intrinsic polarization angle values for the core region.}
\label{rm_table}
\centering
\tiny
\begin{tabular}{c c c c c c}
\hline\hline
Epoch & MJD & RM\tablefootmark{a} & $\sigma_{RM}$\tablefootmark{b} & $\chi$\tablefootmark{c} & $\sigma_{\chi}$\tablefootmark{d} \\
year/month/day & & (rad\,m$^{-2}$) & (rad\,m$^{-2}$) & ($^\circ$) & ($^\circ$) \\
\hline
2011/01/14 & 55576 & $-710$ & 650 & 127 & 10\\
2011/02/25 & 55617 & $-450$ & 590 & 132 & 9\\
2011/03/29 & 55649 & 1450 & 880 & 148 & 13\\
2011/04/25 & 55676 & 1940 & 750 & 132 & 11\\
2011/05/31 & 55713 & $-1060$ & 850 & 163 & 10\\
2011/06/29 & 55742 & $-1570$ & 790 & 141 & 12\\
2011/07/28 & 55771 & $-2930$ & 1380 & 178 & 20\\
2011/08/29 & 55803 & $-3640$ & 930 & 168 & 13\\
2011/09/28 & 55833 & $-3500$ & 1250 & 180 & 18\\
2011/10/29 & 55863 & $-1310$ & 750 & 135 & 12\\
2011/11/28 & 55893 & $-1970$ & 1180 & 147 & 17\\
2011/12/23 & 55918 & $-2370$ & 930 & 168 & 13\\
\hline
\hline
\end{tabular}
\tablefoot{
\begin{tiny}
\newline
\tablefoottext{a}{Rotation measure in rad\,m$^{-2}$.}\\
\tablefoottext{b}{Estimated error for the rotation measure.}\\
\tablefoottext{c}{Intrinsic polarization angle in deg.}\\
\tablefoottext{d}{Estimated error for the intrinsic polarization angle.}\\
\end{tiny}
}
\end{table}
\subsubsection{Jet region}
The light curves for the jet region, extending to about 1 mas from the core at 15 and 24\,GHz, are shown in Fig.~\ref{plots_jet}. The situation is very different in the jet from the core region. The total intensity flux density for the extended region does not show any significant variation.
The polarized flux density shows some variability, reaching a peak during the fourth epoch at 15 and 24\,GHz.
The polarization percentage in the jet region is about $16\%$; this higher degree of polarization with distance from the core seems to be a common feature in blazars \citep{Lister2001}. Finally, the EVPAs are also quite stable, fluctuating weakly around a value of $60^\circ$, i.e., roughly perpendicular to the jet position angle ($\sim-35^{\circ}$).
\subsection{Faraday rotation analysis}
\label{faraday}
A polarized wave propagating through a magnetized plasma is affected by Faraday rotation. As a consequence, the observed polarization angle ($\chi_\mathrm{obs}$) appears rotated with respect to its intrinsic value ($\chi_\mathrm{int}$). This effect is described by a linear relationship between $\chi_\mathrm{obs}$ and the observing wavelength squared ($\lambda^2$):
\begin{equation}
\chi_\mathrm{obs}=\chi_\mathrm{int} + RM\times \lambda^2,
\end{equation}
where RM represents the rotation measure, a quantity related to the electron density $n_e$ (cm$^{-3}$), the parallel to the line of sight, aberrated by relativistic motion, component of the magnetic field $\textbf{B}_{\parallel}$ (milligauss), and the path length $dl$ (parsecs):
\begin{equation}
RM = 812 \int n_e \textbf{B}_{\parallel} \cdot dl \ \ \ [\mbox{rad \ m}^{-2}].
\end{equation}
For the core region, where EVPA values at 15, 24, and 43\,GHz for each observing epoch are available (see Table~\ref{table_data}), we performed linear fits of EVPAs versus $\lambda^2$, obtaining the RM and $\chi_\mathrm{int}$ values and uncertainties reported in Table~\ref{rm_table}. Since the two flips observed at 15 GHz in July and September strongly suggest optically thin-thick transitions, we carried out RM fits using EVPA values rotated by $90^\circ$ at 15 GHz in these epochs. All of these linear fits are reported in Fig.~\ref{rm_core}; in some cases they show significant scatter about a linear trend (e.g., in February 2011), and in other cases they agree well with a linear behavior (e.g., in December 2011).
The time evolution of RM values and of the intrinsic EVPA values are reported in the upper and lower panels of Fig.~\ref{rm}, respectively.
The RM values are distributed across a wide range of values, spanning $(-3640 \pm 930)$ to $(+1940\pm750)$ rad\,m$^{-2}$. However, the uncertainties are often very large, with many values being consistent with 0 within $1\sigma$ or $2\sigma$. It is thus difficult to provide accurate values and even more difficult to claim significant variability during the year.
For the intrinsic values of the polarization angle, the results {\bf tend to} reflect the roughly stable behavior observed for the 43\,GHz EVPAs, with a value of about 150$^\circ$, that is, roughly parallel to the jet axis (see Fig.~\ref{rm}). However, there is some residual, mildly significant variability ($F_{var}$ is $0.10 \pm 0.04$).
\subsection{$\gamma$-ray light curves from {\em Fermi} data}
\label{fermi_sec}
After integrating over the period 2011 January 1 - December 31, we obtain a fit with a power-law model in the 0.1-100 GeV energy range that results in TS = 8728 ($\sim 93 \, \sigma$), with an average flux of ($17.4 \pm 0.5) \times10^{-8}$ ph cm$^{-2}$ s$^{-1}$, a photon index of $\Gamma = 1.77 \pm 0.02$, and an apparent isotropic $\gamma$-ray luminosity of $7.5 \times10^{44}$ erg s$^{-1}$, which is fully compatible with the values obtained over the first two years of {\em Fermi} operation \citep{Nolan2012}. In Fig.~\ref{fermi_plot} we show the $\gamma$-ray flux using time bins of one week (top panel) and the photon index variation (bottom panel) over the entire observing period. For each time bin, the spectral parameters for Mrk\,421 and for all the sources within a radius of 10$^{\circ}$ were left free to vary.
In the $\gamma$-ray light curve we identify three peaks: a main peak (P1) in the first observing period (2011 March 5-11, MJD 55625-55631) with a subsequent decrease to the lowest flux level in 2011 June, followed by two other peaks in the final observing period (P2 in 2011 September 3-16, MJD 55807-55820; P3 in 2011 November 12-18, MJD 55877-55883).
The daily peak flux during P1 is $(38\pm11)\times10^{-8}$ ph cm$^{-2}$ s$^{-1}$, observed on 2011 March 7, corresponding to an apparent isotropic $\gamma$-ray luminosity of $1.6\times10^{45}$ erg s$^{-1}$.
The daily peak flux during P2 and P3 is $(37\pm12)\times10^{-8}$ ph cm$^{-2}$ s$^{-1}$ and $(30\pm9)\times10^{-8}$ ph cm$^{-2}$ s$^{-1}$, observed on September 8 and November 13, respectively.
Spectral hardening during flares has been seen in some blazars, FSRQs in particular \citep[e.g., PKS 1510$-$089;][]{D'Ammando2011}. This sort of behavior is not without precedent in BL Lacs, but it is rare \citep[e.g.,][]{Gasparrini2011,Raiteri2013}. Consistent with the trend for BL Lacs in general, and for Mrk 421 in particular (e.g., Abdo et al 2011), we do not detect any spectral hardening in Mrk 421 during the periods of enhanced $\gamma$-ray activity, and no obvious relation is observed between the photon index versus the $\gamma$-ray flux (see Fig.~\ref{fermi_plot}). Furthermore, the spectral index is generally compatible with the average, ranging from $1.4$ to $2.2$.
\subsection{Correlation between radio and $\gamma$-ray data}
\label{corr_sec}
The comparison of the 15, 24, and 43 GHz radio core and $\gamma$-ray light curves suggests similar trends. We calculated the correlation coefficient between the radio core flux density at each epoch and frequency and the $\gamma$-ray flux during the weekly period containing the radio observations. We report these results in Table~\ref{t.coeff}. All of the radio data sets show a strong mutual correlation, with coefficients in the range $0.88-0.96$. The correlation coefficients between radio and $\gamma$-ray data are lower, in the range $0.42-0.46$; the strongest correlation is found with the data at 43\,GHz, where the largest number of data points is available.
To assess the significance of this correlation and to determine a possible time lag, we calculated the discrete cross-correlation function (DCF) between the radio core flux density at 43\,GHz and the $\gamma$-ray flux.
The results of the correlation analysis are shown in Fig.~\ref{DCF}. \\
To compute the DCF, we used the algorithm developed by \citet{Edelson1988}, and to determine the significance of the correlation, we followed the approach recommended by \citet{Chatterjee2008} and \citet{Max-Moerbeck2010}. As recommended by \citet{Timmer1995}, we simulated 3000 light curves with the same mean and standard deviation as the observed light curves.
The power spectral density (PSD), corresponding to the power in the variability of emission as a function of timescale, is represented by a power-law $\mbox{PSD} \propto f^{-\beta}$, where $f$ is the inverse of the timescale.
For each set of simulations, $\beta$ varies from 1 to 2.5 in steps of 0.5. The curves obtained from different combinations of different PSD slopes, with a confidence level > 99.7\%, are shown in gray in Fig.~\ref{DCF}.
We investigated the delay over a range of $\pm100$ days, with a bin of 15 days. The highest value for the correlation (0.54) is found for zero delay. As is clear from Fig.~\ref{DCF}, this peak does not have a high significance level
for all combinations of the different PSD slopes used here. It has a significance level $> 99.7\%$ for the combination of $\beta$ lower than 1.5 for the $\gamma$-ray PSD and $\beta$ ranging from 1 to 2.5 for the radio PSD. This agrees with the fact that we observe the strongest variability on shorter timescales for the $\gamma$-ray emission (flatter $\beta$) and on longer timescales for the radio emission (steeper $\beta$).
\begin{figure}
\includegraphics[bb=50 0 504 360,width=1.0\columnwidth ,clip]{DCF.eps}
\caption{Discrete cross-correlation function between the $\gamma$-ray and 43 GHz radio light curves (black curve). The gray curves represent the 99.7\% confidence limits relative to stochastic variability, obtained from the combination of different power spectral density slopes. See Sect. \ref{corr_sec} for more details.}
\label{DCF}
\end{figure}
\section{Discussion}
For many decades, Mrk\,421 has been a target of regular monitoring at radio frequencies, showing only moderate variability \citep{aller1999,Venturi2001,Fan2007,Nieppola2009}. Multi-wavelength campaigns have failed to reveal significant correlations between variability at high (or very high) energy and radio wavelengths \citep{Abdo2011,Acciari2011}. The variability detected within the present campaign and its possible connection to the $\gamma$-ray light curve are therefore of great interest, especially considering that they foreshadowed the dramatic and unprecedented radio and $\gamma$-ray flares in 2012 \citep{D'Ammando2012,Richards2013}.
\subsection{Possible radio and $\gamma$-ray connection}
We find significant variability in the first months of 2011 at all three observing frequencies (15, 24, and 43\,GHz) in the total intensity emission ($>3\sigma$ difference between the highest and lowest value). During the same observing period, enhanced activity also occurred at high energies. In Fig.~\ref{norm_light} we show the radio and $\gamma$-ray light curves normalized to the peak value.
In the radio band, the increase of the flux density is stronger at 43\,GHz, where the fractional variability amplitude $F_{var}$ \citep{Edelson2002, Vaughan2003} is $0.24 \pm 0.06$, while at longer wavelengths it corresponds to $0.13 \pm 0.04$ and $0.18 \pm 0.05$ at 15\,GHz and 24\,GHz. In the $\gamma$-ray light curve $F_{var}$ is $0.17 \pm 0.04$, on a weekly timescale. In both radio and $\gamma$-ray energy bands, the peak value is reached close in time, although the sampling interval of our observations does not allow us to determine the date of the radio burst with a better accuracy than a few weeks.
\begin{table}
\caption{Pearson correlation coefficients between radio flux density and $\gamma$-ray photon flux.}
\label{t.coeff}
\centering
\begin{tabular}{c c}
\hline\hline
Data pairs & $r$ \\
\hline
$r_{15-24}$ & 0.93\\
$r_{15-43}$ & 0.88 \\
$r_{24-43}$ & 0.96 \\
$r_{15-\gamma}$ & 0.44\\
$r_{24-\gamma}$ & 0.42\\
$r_{43-\gamma}$ & 0.46\\
\hline
\hline
\end{tabular}
\end{table}
The main peak in the radio light curve (around MJD 55621) occurs close in time with the first $\gamma$-ray peak (MJD 55627); no clear radio counterpart was observed for the second and third $\gamma$-ray enhanced activity episodes. Overall, we find a good correlation between the low-frequency and high-energy emission, as shown by the high and statistically significant value of the correlation coefficient. This could indicate a co-location of the radio and $\gamma$-ray emission regions, and a size as compact as $c\, \Delta t\, \delta \sim 5.3\times10^{16}\times \delta$ cm, assuming $\Delta t \sim 3$ weeks.
On the other hand, the study of the possible delay between radio and $\gamma$-ray light curves, based on the DCF, did not provide significant constraints on the lag between the radio and $\gamma$-ray light curves. In fact, much longer data trains are necessary to reach highly significant lag values \citep[for a detailed discussion of the significance of radio-gamma light curve correlation and lags, see][]{Max-Moerbeck2013}. We furthermore note that in a recent study of the radio and $\gamma$-ray correlated variability in {\it Fermi} bright blazars based on 3.5 years of dense monitoring, \citet{Fuhrmann2014} did not find significant correlation between the 7mm and the $\gamma$-ray data for Mrk\,421.
The proximity between the radio and $\gamma$-ray enhanced activity at the beginning of 2011 differs from what was observed in 2012, where the $\gamma$-ray flare \citep{D'Ammando2012} led the radio burst by about 40 days at 15\,GHz \citep{Richards2013}. However, the enhanced activities observed in 2011 and 2012 are different. The latter is characterized by a $\gamma$-ray apparent luminosity of $6.7\times 10^{45}$ erg s$^{-1}$, that is, about four times more powerful than the former, with simultaneous tentative detection in the TeV energy band \citep{Bartoli2012}. The different behavior of the two flares may suggest different regions for the $\gamma$-ray activity: upstream of the radio emission in the 2012 flare, while downstream along the jet where the emission is not opaque at the radio frequencies in the 2011 episode.
\begin{figure}
\includegraphics[width=1.0\columnwidth]{norm_light.ps}
\caption{Radio light curves normalized to the peak value. Triangles, squares, and circles represent 15, 24, and 43\,GHz data, respectively. All error bars (not shown) are about 0.1 mJy. The stars (labeled with P1, P2 and P3) represent the three main peaks in the $\gamma$-ray light curve normalized to the highest value.}
\label{norm_light}
\end{figure}
Interestingly, after the main peak in February 2011, which was observed both at radio and high energies, the polarized flux density at 43\,GHz increases rapidly (Fig.~\ref{plots_core}) and peaks on 2011 March 29 (MJD 55649). Simultaneously the polarization percentage increases at 43\,GHz and reaches a peak value of $3.6\%$, followed by a gradual decrease to the mean value ($\sim 1\%$). A similar behavior was observed by \citet{Piner2005} after an enhanced activity at TeV energies, when the polarization percentage of the core reached $\sim5\%$ at 22\,GHz. This behavior is less pronounced in our data at 15 and 24\,GHz.
In the jet region the polarized flux density and the fractional polarization reach a peak on 2011 April 25 (MJD 55675) at 15 and 24\,GHz. These peaks occur just after the enhanced activity observed in the core region in total intensity emission. This behavior could be associated with a propagating disturbance, but no new components were detected in the total intensity images.
\citet{Blasi2013} investigated the behavior in the optical regime during 2011, by considering optical data provided by the Steward Observatory of the University of Arizona. We found a peak in the V magnitude occurring during the same period. The optical polarization increased from $\sim1\%$ to $\sim7\%$ and the optical polarization angle reduced its scatter. This may indicate that the magnetic field underwent a transition from a tangled state to a more regular one, in correspondence to the observed enhanced activity.
\subsection{EVPA variation and magnetic field topology}
To fully understand the possible relation between the enhanced activity and the physical changes occurring in the source, it is important to investigate the EVPA trend in the different regions of the source. So far, no obvious relation has been found between EVPA variation and high-energy activity.
In the jet region, extending about 1 mas from the VLBI core, the EVPAs maintain a stable value of about $60^\circ$ during the entire observing period, roughly perpendicular to the inner jet position angle, which is $\sim-35^{\circ}$. This implies that the magnetic field is parallel to the jet axis, which is a rare behavior in BL Lac objects. However, we note that previous VLBI observations of Mrk\,421 had indicated this peculiar magnetic field configuration in its parsec scale jet \citep{Piner2005}. This configuration could be caused either by velocity shear across the jet or by a helical magnetic field with a pitch angle smaller than $45^\circ$ \citep{Wardle2013}. At 24\,GHz, we observe an EVPA rotation of about $60^\circ$ at the end of 2011 (MJD 55893). This behavior may be associated with the propagation of a shock, but no new component and no significant motion were detected in total intensity images. We note that the nondetection of the jet polarized emission at the following epoch (MJD 55918) results in a lack of the EVPA, which does not allow us to confirm the possible EVPA rotation.
In the core region, the EVPA trend is more complex. At 43\,GHz, where the Faraday rotation effects are weaker, the EVPAs are roughly stable in a range of values between $\sim100^\circ$ and $\sim150^\circ$, that is, roughly parallel to the jet axis. This trend is then reflected in the intrinsic values of the polarization angle, obtained from the $\lambda^2$ fits (see the lower panel in Fig.~\ref{rm}). Assuming that this emission region is optically thin at 43\,GHz, the EVPA configuration reflects a magnetic field nearly orthogonal to the jet axis, which is typical of the emission from a transverse shock and different from what is observed in the outer jet. A similar behavior was reported in \citet{Piner2005}: they found parallel EVPAs both in the core and in the innermost jet component (at $\sim$0.3 mas from the core). The extent of the validity of the optically thin emission assumption can be debated. However, the average core spectral index between 15 and 43\,GHz in our data is $\alpha = 0.20\pm 0.13 $ \citep[see also][]{Blasi2013}, suggesting that at 43\,GHz the emission region is predominantly optically thin.
Considering the lower frequencies, in particular at 15\,GHz, we find at least two clear flips of the core EVPAs by about $90^\circ$ (from parallel to perpendicular to the jet, in July and September, with a smaller change also in March). A similar behavior for the EVPAs in the VLBI core region of Mrk\,421 was observed also by \citet{Charlot2006}. They measured a rotation of $\sim60^\circ$ during a three-month period at 15 and 22\,GHz. This behavior is the typical signature of a transition between an optically thin and thick regime, in which the EVPA swings from perpendicular to parallel to the magnetic field direction. Indeed, both in July and September the spectral index flattens ($\Delta \alpha_\mathrm{Jul-Jun}^{15-43}=-0.17$ and $\Delta \alpha_\mathrm{Sep-Aug}^{15-43}=-0.13$). To account for these variations solely by an opacity effect, we adopted in this discussion the core 15 GHz EVPAs rotated by $90^\circ$ for the July and September data, as we already did in determining the RM fits in Sect.~\ref{faraday}.
To explain the observed variability at longer wavelengths, a possible physical scenario might be an association with opacity effects and variable Faraday rotation. The RM variability might be related to changes in the accretion rate, as suggested also by the similar trends found for the RM and the core flux density evolution; for viable explanations of the RM sign change in relativistic jets, see also \citet{OSullivan2009} and \citet{Gomez2011}.
Finally, residual variability ($F_{var}$ is $0.10 \pm 0.04$) is present in the intrinsic (Faraday de-rotated) EVPAs. This intrinsic variability can be the consequence of a blend of variable cross-polarized subcomponents within the beam, whose relative contributions to the total polarization properties vary as a function of time. This structure explains both the variations of the intrinsic EVPA values, which are integrated over the VLBI core region, and the low polarization of the core, resulting from significant cancellation from subcomponents with different EVPAs.
The cross-polarized subcomponents can be explained by two different physical scenarios: (1) the jet base might be turbulent, so that the subcomponents would be turbulent cells with random field direction; (2) the core magnetic field might be perpendicular to the jet, but with much of the emission optically thick at 15\,GHz, giving rise to EVPAs parallel to the jet in the optically thin part and EVPAs perpendicular to the jet in the optically thick part.
In either scenario the intrinsic EVPA variations are the direct consequence of changes in the relative weight of one component with respect to the other(s). The second scenario seems to be favored by the observed behavior at 43\,GHz (EVPAs stable in the range 100$^\circ$-150$^\circ$).
The relevance of subcomponent blending can be directly seen in the April data (MJD 55675): the 43\,GHz polarization image clearly shows a transverse EVPA distribution and a limb-brightened structure in the polarized emission (see the lower panel in Fig.~\ref{maps}) in the inner part of the jet. We argue that velocity gradients in this region might cause Kelvin-Helmholtz instabilities that would heat the boundary layer, causing the limb brightening. A similar transverse structure for this source was observed by \citet{Piner2010} at 22\,GHz. These authors noted that the EVPAs were almost parallel along the jet axis and became orthogonal toward the jet edges. This is similar to what we observe in the EVPAs in the western limb. This may reflect a spine plus layer polarization structure and seems to be a common feature in TeV blazars, such as Mrk\,501 \citep[e.g.,][]{Giroletti2008, Piner2010}.
The observed transverse EVPA rotation implies a complex topology for the magnetic field. For example, \citet{Lyutikov2005} argued that such a transverse EVPA distribution may be consistent with a large scale helical magnetic field in a resolved cylindrical jet.
This transverse polarization structure would reflect the intrinsic magnetic field geometry, and not the propagation of a shock arising from the core and interacting with the surrounding medium \citep{Lyutikov2005, Gabuzda2004}.
Finally, we observe that, contrary to what was found by \citet{Piner2010}, no increase of the fractional polarization towards the jet edges is observed in our data. This may be due to the high activity state of Mrk\,421, causing an enhancement of the polarized emission from the core region.
\section{Summary and conclusions}
We presented a detailed analysis of new VLBA polarization observations of Mrk\,421 at 15, 24, and 43\,GHz during 2011.
During this observing period the source showed significant radio flux density variability. In particular, in the first part of 2011 we detected a prominent peak in the radio light curves and an enhanced $\gamma$-ray activity occurring close in time. A good correlation was found between the radio core and the $\gamma$-ray light curves. Just after this enhanced activity we detected a peak at 43\,GHz in the polarized flux density and an increase in the fractional polarization.
We found the core region to be polarized by about $1\%$ while the jet region showed an average fractional polarization of $\sim 16\%$. In the jet region EVPAs were roughly orthogonal to the jet position angle, implying a magnetic field parallel to the jet axis. In the core region, the EVPAs varied between $\sim100^\circ$ and $\sim150^\circ$ at 43\,GHz, but they were more variable at lower frequencies, in particular at 15\,GHz, because of opacity effects and variable Faraday rotation.
Near the VLBI radio core we confirmed the presence of a limb-brightened structure in polarized emission and a transverse EVPA distribution, as previously reported by \citet{Piner2010} and \citet{Giroletti2008}.
Thanks to these multi-frequency and multi-epoch datasets we constrained polarization parameters, and we proposed the following scenario to explain the observed properties: The parallel magnetic field in the jet region could be caused by velocity shear across the jet. These velocity gradients might cause Kelvin-Helmholtz instabilities that would heat the boundary layer, causing the observed limb brightening. This physical mechanism may have been favored by the enhanced activity, that foreshadowed the appearance of the limb brightened structure.
To explain the intrinsic EVPA variations in the core region, we proposed subcomponents within the VLBI core region, with different polarization properties, which also result in the low degree of polarization.
The polarimetric information presented in this paper complements the variability analysis discussed in \citet{Lico2012} and in \citet{Blasi2013}. The measurements are from a multi-frequency campaign carried out during 2011, which involves observations throughout the electromagnetic spectrum. For this reason, our results provide a starting point for future broadband analyses that could improve our knowledge about the emission of TeV blazars.
\begin{table*}
\begin{center}
\begin{tiny}
\caption{Summary of the total and polarized intensity parameters plotted in Figs.~\ref{plots_core} and \ref{plots_jet}.}
\label{table_data}
\begin{tabular}{ccccccccccc}
\hline\hline
Epoch & Frequency & Region & S\tablefootmark{a} & $\sigma_S$\tablefootmark{b} & P\tablefootmark{c} & $\sigma_P$\tablefootmark{d} & m\tablefootmark{e} & $\sigma_m$\tablefootmark{f} & $\chi$\tablefootmark{g} & $\sigma_{\chi}$\tablefootmark{h} \\
year/month/day & (GHz) & & (mJy) & (mJy) & (mJy) & (mJy) & (\%) & (\%) & (deg) & (deg) \\
\hline\hline
&&&&&&&&&&\\
2011/01/14 & 15 & Core & 332.2 & 33.2 & 4.9 & 0.5 & 1.5 & 0.2 & 108.3 & 8.7 \\
& & Jet & 60.5 & 6.1 & 8.9 & 1.0 & 14.8 & 2.3 & 58.3 & 7.5 \\
& 24 & Core & 318.9 & 31.9 & 2.7 & 0.3 & 0.8 & 0.1 & 126.5 & 8.8 \\
& & Jet & 47.4 & 4.8 & 6.4 & 0.9 & 13.6 & 2.3 & 58.5 & 9.0 \\
& 43 & Core & 291.2 & 29.1 & 2.2 & 0.3 & 0.8 & 0.1 & 117.8 & 10.9 \\
2011/02/25 & 15 & Core & 392.9 & 39.3 & 4.2 & 0.4 & 1.1 & 0.2 & 117.3 & 7.8 \\
& & Jet & 43.4 & 4.4 & 6.5 & 0.8 & 15.0 & 2.5 & 55.3 & 8.1 \\
& 24 & Core & 419.9 & 42.0 & 4.2 & 0.5 & 1.0 & 0.1 & 143.5 & 8.2 \\
& & Jet & 45.1 & 4.6 & 5.7 & 0.9 & 12.6 & 2.3 & 55.5 & 8.7 \\
& 43 & Core & 408.6 & 40.9 & 2.6 & 0.3 & 0.6 & 0.1 & 113.8 & 10.2 \\
2011/03/29 & 15 & Core & 399.6 & 40.0 & 2.6 & 0.3 & 0.7 & 0.1 & 177.3 & 12.4 \\
& & Jet & 55.7 & 5.6 & 7.6 & 0.9 & 13.7 & 2.2 & 56.3 & 12.6 \\
& 24 & Core & 364.7 & 36.5 & 5.5 & 0.6 & 1.5 & 0.2 & 172.5 & 12.3 \\
& & Jet & 39.6 & 4.0 & 11.5 & 1.3 & 29.1 & 4.4 & 61.5 & 12.5 \\
& 43 & Core & 346.8 & 34.7 & 12.5 & 1.4 & 3.6 & 0.5 & 142.8 & 13.1 \\
2011/04/25 & 15 & Core & 367.4 & 36.7 & 5.1 & 0.5 & 1.4 & 0.2 & 174.3 & 10.5 \\
& & Jet & 47.3 & 4.8 & 16.5 & 1.8 & 34.8 & 5.1 & 58.3 & 8.9 \\
& 24 & Core & 328.8 & 32.9 & 6.5 & 0.7 & 2.0 & 0.3 & 157.5 & 10.5 \\
& & Jet & 44.0 & 4.5 & 16.4 & 1.7 & 37.3 & 5.5 & 72.5 & 10.5 \\
& 43 & Core & 306.7 & 30.7 & 8.1 & 0.8 & 2.6 & 0.4 & 130.8 & 11.8 \\
2011/05/31 & 15 & Core & 344.4 & 34.4 & 1.1 & 0.2 & 0.3 & 0.1 & 130.3 & 13.8 \\
& & Jet & 43.7 & 4.4 & 5.4 & 0.8 & 12.3 & 2.2 & 51.3 & 9.6 \\
& 24 & Core & 301.4 & 30.1 & 3.0 & 0.3 & 1.0 & 0.1 & 168.5 & 10.1 \\
& & Jet & 35.3 & 3.6 & 3.7 & 0.7 & 10.4 & 2.2 & 66.5 & 11.4 \\
& 43 & Core & 246.8 & 24.7 & 3.4 & 0.4 & 1.4 & 0.2 & 149.8 & 10.1 \\
2011/06/29 & 15 & Core & 264.6 & 26.5 & 1.8 & 0.2 & 0.7 & 0.1 & 102.2 & 11.4 \\
& & Jet & 55.6 & 5.6 & 8.9 & 1.1 & 16.0 & 2.5 & 61.2 & 11.6 \\
& 24 & Core & 219.5 & 22.0 & 2.2 & 0.3 & 1.0 & 0.2 & 144.0 & 14.7 \\
& & Jet & 37.3 & 3.9 & 6.0 & 1.0 & 16.2 & 3.2 & 46.0 & 13.0 \\
& 43 & Core & 176.0 & 17.6 & 2.4 & 0.4 & 1.3 & 0.2 & 129.8 & 11.5 \\
2011/07/28 & 15 & Core & 246.8 & 24.7 & 2.6 & 0.4 & 1.1 & 0.2 & 196.2 & 20.2 \\
& & Jet & 45.9 & 4.7 & 6.4 & 1.1 & 13.9 & 2.7 & 56.2 & 20.2 \\
& 24 & Core & 215.3 & 21.5 & 4.0 & 0.5 & 1.9 & 0.3 & 165.5 & 20.4 \\
& & Jet & 43.3 & 4.4 & 1.8 & 1.1 & 4.2 & 2.5 & 52.5 & 26.3 \\
& 43 & Core & 195.9 & 19.6 & 3.4 & 0.5 & 1.7 & 0.3 & 159.8 & 20.0 \\
2011/08/29 & 15 & Core & 299.2 & 29.9 & 4.4 & 0.5 & 1.5 & 0.2 & 91.2 & 13.4 \\
& & Jet & 47.3 & 4.8 & 12.6 & 1.5 & 26.6 & 4.2 & 63.2 & 14.4 \\
& 24 & Core & 259.5 & 25.9 & 2.6 & 0.4 & 1.0 & 0.2 & 115.5 & 13.4 \\
& & Jet & 38.9 & 4.0 & 7.9 & 1.2 & 20.3 & 3.7 & 49.5 & 14.3 \\
& 43 & Core & 216.1 & 21.6 & 4.7 & 0.5 & 2.2 & 0.3 & 172.8 & 13.7 \\
2011/09/28 & 15 & Core & 275.4 & 27.5 & 2.5 & 0.3 & 0.9 & 0.1 & 181.2 & 18.1 \\
& & Jet & 38.7 & 3.9 & 3.8 & 0.8 & 9.7 & 2.2 & 57.2 & 18.8 \\
& 24 & Core & 258.7 & 25.9 & 4.0 & 0.5 & 1.6 & 0.2 & 175.5 & 17.9 \\
& & Jet & 42.2 & 4.3 & 5.3 & 1.0 & 12.7 & 2.8 & 42.5 & 19.2 \\
& 43 & Core & 227.0 & 22.7 & 3.0 & 0.4 & 1.3 & 0.2 & 149.8 & 18.5 \\
2011/10/29 & 15 & Core & 267.8 & 26.8 & 4.7 & 0.5 & 1.8 & 0.3 & 108.2 & 10.5 \\
& & Jet & 53.1 & 5.4 & 5.5 & 0.9 & 10.3 & 2.0 & 65.2 & 11.1 \\
& 24 & Core & 259.7 & 26.0 & 1.8 & 0.3 & 0.7 & 0.1 & 103.5 & 13.3 \\
& & Jet & 34.1 & 3.5 & 6.8 & 0.9 & 19.9 & 3.4 & 43.5 & 12.1 \\
& 43 & Core & 233.9 & 23.4 & 2.5 & 0.4 & 1.1 & 0.2 & 140.8 & 11.5 \\
2011/11/28 & 15 & Core & 303.8 & 30.4 & 2.6 & 0.4 & 0.9 & 0.1 & 106.5 & 17.1 \\
& & Jet & 43.1 & 4.4 & 7.0 & 1.1 & 16.2 & 3.1 & 83.5 & 17.6 \\
& 24 & Core & 304.1 & 30.4 & 2.1 & 0.3 & 0.7 & 0.1 & 109.3 & 18.5 \\
& & Jet & 40.7 & 4.3 & 2.9 & 1.0 & 7.0 & 2.6 & 115.3 & 19.6 \\
& 43 & Core & 278.5 & 27.9 & 3.1 & 0.4 & 1.1 & 0.2 & 152.8 & 17.2 \\
2011/12/23 & 15 & Core & 303.8 & 30.4 & 2.9 & 0.4 & 1.0 & 0.2 & 114.5 & 14.1 \\
& & Jet & 46.3 & 4.7 & 3.6 & 0.9 & 7.8 & 2.0 & 64.5 & 13.6 \\
& 24 & Core & 284.9 & 28.5 & 1.9 & 0.3 & 0.7 & 0.1 & 144.3 & 13.8 \\
& & Jet & 25.8 & 2.7 & 0.4 & 0.6 & 1.6 & 2.3 & - & - \\
& 43 & Core & 202.9 & 20.3 & 3.3 & 0.4 & 1.6 & 0.3 & 162.8 & 12.5 \\
\hline
\end{tabular}
\end{tiny}
\tablefoot{
\begin{tiny}
\newline
\tablefoottext{a}{Flux density in mJy.}\\
\tablefoottext{b}{Estimated errors for the flux density.}\\
\tablefoottext{c}{Polarized flux density in mJy.}\\
\tablefoottext{d}{Estimated errors for the polarized flux density.}\\
\tablefoottext{e}{Fractional polarization.}\\
\tablefoottext{f}{Estimated errors for the fractional polarization.}\\
\tablefoottext{g}{EVPAs.}\\
\tablefoottext{h}{Estimated errors for the EVPA.}
\end{tiny}
}
\end{center}
\end{table*}
\begin{acknowledgement}
\begin{small}
The \textit{Fermi} LAT Collaboration acknowledges generous ongoing support
from a number of agencies and institutes that have supported both the
development and the operation of the LAT as well as scientific data analysis.
These include the National Aeronautics and Space Administration and the
Department of Energy in the United States, the Commissariat \`a l'Energie Atomique
and the Centre National de la Recherche Scientifique / Institut National de Physique
Nucl\'eaire et de Physique des Particules in France, the Agenzia Spaziale Italiana
and the Istituto Nazionale di Fisica Nucleare in Italy, the Ministry of Education,
Culture, Sports, Science and Technology (MEXT), High Energy Accelerator Research
Organization (KEK) and Japan Aerospace Exploration Agency (JAXA) in Japan, and
the K.~A.~Wallenberg Foundation, the Swedish Research Council and the
Swedish National Space Board in Sweden.
Additional support for science analysis during the operations phase is gratefully
acknowledged from the Istituto Nazionale di Astrofisica in Italy and the Centre National d'\'Etudes Spatiales in France. \\
This work is based on observations obtained through the BG207 VLBA project, which makes use of the Swinburne University of Technology software correlator, developed as part of the Australian Major National Research Facilities Programme and operated under licence \citep{Deller2011}. The
National Radio Astronomy Observatory is a facility of the National Science Foundation operated
under cooperative agreement by Associated Universities, Inc. For this paper we made use of the NASA/IPAC Extragalactic Database NED which is operated by the JPL, Californian Institute of Technology, under contract with the National Aeronautics and Space Administration. We acknowledge financial contribution from grant PRIN-INAF-2011. KVS and YYK are partly supported by the Russian Foundation for Basic Research (project 13-02-12103). KVS is also supported by the RFBR grant 14-02-31789. YYK is also supported by the Dynasty Foundation. The research at Boston University was supported in part by NASA through Fermi grants NNX08AV65G, NNX08AV61G, NNX09AT99G, NNX09AU10G, and NNX11AQ03G, and by US National Science Foundation grant AST-0907893. Part of this work was supported by the Marco Polo program of the University of Bologna and the COST Action MP0905 "Black Holes in a Violent Universe". The research at the Instituto de Astrofisica de Andalucia was supported in part by the Spanish Ministry of Economy and Competitiveness grant AYA2010-14844 and by the Regional Government of Andalucía (Spain) grant P09-FQM-4784.
We thank Dr. Astrid Peter for the language editing work which improved the text of the present manuscript and the anonymous referee for the valuable comments and suggestions.
\end{small}
\end{acknowledgement}
|
\section{Introduction}
Let $E = \big(\mathbb R^n, \|\cdot\|\big)$ be an $n$-dimensional space equipped with a quasi-norm $\|\cdot\|$, and let $X$ be a random vector in $E$. The present note is concerned with small ball estimates of $X$, i.e., estimates of the form
\begin{align} \label{small ball def}
\mathbb P\big(\|X\| \le t\big) \le \varphi(t),
\end{align}
where $\varphi(t) \to 0$ as $t\to 0$.
\smallskip
Estimates of the form \eqref{small ball def} have been studied under different assumptions on $E$ and $X$. One direction is the case when $E = \ell_2^n$, i.e., when $\|\cdot\| = |\cdot|_2$ is the Euclidean norm, and $X$ is assumed to be log-concave or, more generally, $\kappa$-concave. Recall that a log-concave vector is a vector that satisfies that for every $A,B\subseteq\mathbb R^n$ and every $\lambda \in [0,1]$,
\begin{align*}
\mathbb P\big(X\in \lambda A + (1-\lambda) B\big) \ge \mathbb P\big(X\in A\big)^{\lambda}\cdot\mathbb P\big(X\in B\big)^{1-\lambda}.
\end{align*}
For such vectors it was shown in \cite{Pao12} that
\begin{align*}
\mathbb P\big(|X|_2 \le \sqrt n t \big) \le \big(Ct\big)^{C'\sqrt n},
\end{align*}
and this result was later generalized in \cite{AGLLOPT12} to $\kappa$-concave vectors.
\smallskip
Another direction which has been studied is the case when $X$ is a Gaussian vector, and $\|\cdot\|$ is a \emph{general} norm. For example, in \cite{LO05} it was shown that if $X$ is a centered Gaussian vector and $\|\cdot\|$ is a norm on $\mathbb R^n$ with unit ball $K$ such that its $n$-dimensional Gaussian measure, denoted $\gamma_n(K)$, is less than $1/2$, then
\begin{align*}
\mathbb P\big(\|X\| \le t \big) \le \big(2t\big)^{\frac{\omega^2}{4}}\gamma_n(K),
\end{align*}
where $\omega$ is the inradius of $K$. See also \cite{LS01} for an earlier survey of the subject.
\smallskip
Finally, let us mention that small ball estimates play a r\^ole in other problems, such as invertibility of random matrices and convex geometry. See e.g. \cite{GM04, RV09, PP13}.
\smallskip
While the above results have a more geometric flavor, in the present note we will try to present a more analytic approach. For a random vector $X$, let $\phi_X$ be its characteristic function ,i.e.,
\begin{align*}
\phi_{X}(\xi) = \mathbb E\exp\big(i\langle \xi,X\rangle\big).
\end{align*}
Recall the following result:
\begin{thm} \cite[Theorem 3.1]{FG11} \label{thm fourier}
Assume that $\|\cdot\|$ is a quasi-norm on $\mathbb R^n$ with unit ball $K$. Then for every $t>0$,
\begin{align} \label{not smooth}
\mathbb P\big(\|X\| \le t\big) \le |K|\left(\frac{t}{2\pi}\right)^{n}\int_{\mathbb R^n}\big|\phi_X(\xi)\big|d\xi.
\end{align}
\end{thm}
Theorem \ref{thm fourier} says that one can obtain small ball estimates by estimating the $L_1$ norm of the characteristic function of the random vector. Moreover, one can consider a ``smoothed" version of Theorem \ref{thm fourier}: consider instead of $X$ the random vector $X + tG$, where $G$ is a standard Gaussian vector in $\mathbb R^n$ which is independent of $X$. Since $\|\cdot \|$ is a quasi-norm on $\mathbb R^n$, there exists a constant $C_K>0$ such that
\begin{align} \label{const quasi}
\|x + y\| \le C_K(\|x\| + \|y\|), ~~ x,y \in \mathbb R^n.
\end{align}
Therefore,
\begin{align*}
\mathbb P\big(\|X + tG\| \le 2C_Kt\big) & \ge \mathbb P\big(\|X\| \le t \wedge \|G\| \le 1\big)
\\ & = \mathbb P\big(\|X\| \le t\big)\cdot \mathbb P\big(\|G\| \le 1\big)
\\ & = \mathbb P\big(\|X\| \le t\big)\cdot \gamma_n(K),
\end{align*}
where $\gamma_n(\cdot)$ is the $n$-dimensional Gaussian measure. Thus, Theorem \ref{thm fourier} implies
\begin{align*}
\mathbb P\big(\|X\|\le t\big) \le \frac{\mathbb P\big(\|X + tG\| \le 2C_Kt\big)}{\gamma_n(K)} \le \frac{|K|}{\gamma_n(K)}\left(\frac{C_Kt}{\pi}\right)^n\int_{\mathbb R^n}\big|\phi_{X + tG}(\xi)\big|d\xi.
\end{align*}
Using the independence of $X$ and $G$,
\begin{align} \label{small ball smoothed}
\nonumber \mathbb P\big(\|X\|\le t\big) & \le \frac{|K|}{\gamma_n(K)}\left(\frac{C_Kt}{\pi}\right)^n\int_{\mathbb R^n}\big|\phi_{X}(\xi)\big|\phi_{tG}(\xi)d\xi
\\ & = \frac{|K|}{\gamma_n(K)}\left(C'_Kt\right)^n\int_{\mathbb R^n}\big|\phi_{X}(\xi)\big|e^{-\frac{t^2|\xi |_2^2}{2}}d\xi.
\end{align}
Inequality \eqref{small ball smoothed} enables one to obtain small ball estimates in cases where \eqref{not smooth} cannot be applied. We use it for two different sets of examples. In the first set of examples we consider continuous random vector under certain assumptions on their characteristic functions (which are nothing but the Fourier transform of their density functions). This is discussed in Section \ref{sec cont}, Theorem \ref{thm decay}. In the second set of examples, we consider random vectors $X$ of the form
\begin{align*}
X = \sum_{i = 1}^N\alpha_i a_i,
\end{align*}
where the $a_i$ are fixed vectors in $\mathbb R^n$ and the $\alpha_i$'s are i.i.d. random variables that satisfy a certain anti-concentration condition. This problem and its applications have been studied by many authors, first in the one dimensional case (i.e., when $n=1$) and later in the multidimensional case. See \cite{FS07, RV08, RV09, TV092, TV091, TV12, Ngu121, NV13} and the reference therein for more information on this subject. In the case $E = \ell_2^n$, the problem of finding a small ball estimate have been previously considered in \cite{FS07, RV09} and is called a Littlewood-Offord type estimate. Here such an estimate is obtained for a general quasi-norm. This is discussed in Section \ref{sec lo}, Theorem \ref{thm lo}. We recall that in \cite{RV09}, Littlewood-Offord estimates were used to estimate the smallest singular value of a rectangular matrix, where the smallest singular value of a matrix $A$ is defined as $\inf_{|x|_2 =1}|Ax|_2$. It would be interesting to try and use Theorem \ref{thm lo} to estimate $\inf_{\|x\|=1}\|Ax\|$ where now $\|\cdot\|$ is a general norm, or even quasi-norm. See e.g. \cite{LL15} for some recent results in this direction.
\smallskip
\noindent{\bf Notation.} In this note $C$, $C'$, etc. always denote absolute constants. $\|\cdot\|$ denotes a quasi-norm with unit ball $K$. $|\cdot|_2$ denotes the Euclidean norm on $\mathbb R^n$. $B(x,r)$ denotes the closed ball around $x$ with radius $r$ with respect to the Euclidean norm. $\gamma_n(\cdot)$ denotes the $n$-dimensional Gaussian measure.
\section{Small ball estimates for continuous random vectors} \label{sec cont}
In this section we consider continuous random vectors, i.e., vectors with density function $f_X$. For such vectors we have
\begin{align*}
\phi_X(\xi) = \mathbb E\exp\big(i\langle \xi, X\rangle \big) = \int_{\mathbb R^n}e^{i\langle \xi, x\rangle}f_X(x)dx = \hat f(\xi).
\end{align*}
We can rewrite \eqref{small ball smoothed} in the following way:
\begin{align*}
\mathbb P\big(\|X\| \le t \big) \le \frac{|K|}{\gamma_n(K)}\left(C_K't\right)^n\int_{\mathbb R^n}\big|\hat f_X(\xi)\big|e^{-\frac{t^2|\xi|_2^2}{2}}d\xi.
\end{align*}
This suggests that small ball estimates are related to weighted norms of $\hat f_X$ which are in turn known to be related to smoothness properties of $f_X$. We will study small ball estimates in terms of Sobolev norms.
\subsection{Small ball estimates and Sobolev norm}
Recall the definition of Sobolev norm: if $\mathcal F$ is the Fourier transform on $\mathbb R^n$, then
\begin{align} \label{def sobolev}
\|f\|_{\beta,p} = \|f\|_{H_{\beta,p}(\mathbb R^n)} = \left\|\mathcal F^{-1} \left( \left(1 + |\xi|^2 \right)^{\beta/2}\hat f \right)\right\|_{L_p \left(\mathbb R^n \right)},
\end{align}
where $p\in (1,\infty)$ and $\beta>0$.
\begin{thm} \label{thm decay}
Assume that $X$ is a random vector in $\mathbb R^n$. Assume that $1 < p\le 2$. Then for every quasi-norm $\|\cdot\|$ on $\mathbb R^n$ with unit ball $K$,
\begin{align} \label{small ball sobolev}
\mathbb P \left(\|X\| \le t \right) \le C_K'^n\frac{|K|}{\gamma_n(K)} \|f_X\|_{\beta, p}\cdot M(\beta,p,n,t) .
\end{align}
If $pt^2\le 2$, then
\begin{align*}
M(\beta,p,n,t) \le
\begin{cases}
2^{\frac n{2p}-\frac{\beta}{2}}|\mathbb S^{n-1}|^{1/p}\Gamma\left(\frac{n-\beta p}{2}\right)^{1/p}p^{\frac{\beta}{2}-\frac{n}{2p}}\cdot t^{\beta + \frac{n}{p'}} & 2 < n-\beta p,
\\
2^{\frac n{2p}-\frac{\beta}{2}}|\mathbb S^{n-1}|^{1/p}\left(\log\left(\frac {2e}{pt^2}\right)\right)^{1/p}p^{\frac{\beta}{2}-\frac{n}{2p}}\cdot t^{\beta + \frac{n}{p'}} & 0 < n-\beta p \le 2,
\\
|\mathbb S^{n-1}|^{1/p}\left(\log\left(\frac {2e}{pt^2}\right)\right)^{1/p}t^n & n- \beta p \le 0,
\end{cases}
\end{align*}
where $p' = p/(p-1)$. Otherwise, if $pt^2\ge 2$, then
\begin{align*}
M(\beta,p,n,t) \le
\begin{cases}
|\mathbb S^{n-1}|^{1/p}\left(2e^{-\frac{pt^2}{18}} + \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\Gamma\left(\frac{n-\beta p}{2}\right)\right)^{1/p}t^n & 2 \le pt^2 \le n-\beta p,
\\
3^{1/p}|\mathbb S^{n-1}|^{1/p}e^{-\frac{t^2}{18}}t^n & n-\beta p \le pt^2 \le n,
\\
|\mathbb S^{n-1}|^{1/p}\left(\frac {2n} {pt^2}\log\left(\frac{ept^2}{n}\right)\right)^{\frac{n}{2p}}t^n & n \le pt^2.
\end{cases}
\end{align*}
\end{thm}
The main tool in the proof of Theorem \ref{thm decay} is the following lemma.
\begin{lem} \label{lem lp norm}
Let $p \in [1,2]$. If $pt^2 \le 2$ then
\begin{align*}
\frac{\big\| \left(1 + |\xi|^2 \right)^{-\frac{\beta}{2}}e^{-\frac{t^2|\xi|^2}{2}}\big\|_{L_p \left(\mathbb R^n \right)}^p }{|\mathbb S^{n-1}|}
\le
\begin{cases}
\Gamma \left(\frac{n-\beta p}{2} \right)\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}} & \beta p <n-2 \\
\log \left(\frac{2e}{pt^2}\right)\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}& n-2\le \beta p <n \\
\log \left(\frac{2e}{pt^2}\right)& \beta p \ge n.
\end{cases}
\end{align*}
Otherwise, if $pt^2 \ge 2$ then
\begin{align*}
\frac{\big\| \left(1 + |\xi|^2 \right)^{-\frac{\beta}{2}}e^{-\frac{t^2|\xi|^2}{2}}\big\|_{L_p \left(\mathbb R^n \right)}^p }{|\mathbb S^{n-1}|}
\le
\begin{cases}
2e^{-\frac{pt^2}{18}} + \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\Gamma\left(\frac{n-\beta p}{2}\right) & 2 \le pt^2 \le n-\beta p,
\\
3e^{-\frac{pt^2}{18}}& n-\beta p \le pt^2 \le n,
\\
\left(\frac {2n} {pt^2}\log\left(\frac{ept^2}{n}\right)\right)^{n/2} & n \le pt^2.
\end{cases}
\end{align*}
\end{lem}
As part of the proof of Lemma \ref{lem lp norm}, we need the following.
\begin{prp} \label{prop x alpha}
Assume that $x\ge \alpha \ge 1$. Then
\begin{align*}
\int_x^{\infty}r^{\alpha-1}e^{-r}dr \le \frac{2^{\alpha + 1}x^{\alpha}e^{-x}}{\alpha}.
\end{align*}
\end{prp}
\begin{proof}
We have
\begin{align} \label{iden split}
\nonumber\int_x^{\infty}r^{\alpha-1}e^{-r}dr & = e^{-x}\int_0^{\infty}(u + x)^{\alpha-1}e^{-u}du
\\ & = e^{-x}\left[\int_0^{x}(u + x)^{\alpha-1}e^{-u}du + \int_x^{\infty}(u + x)^{\alpha-1}e^{-u}du\right].
\end{align}
Now,
\begin{align*}
\int_0^x(u + x)^{\alpha-1}e^{-u}du \le \int_0^x(u + x)^{\alpha-1}du = \frac{x^{\alpha}\left(2^{\alpha}-1\right)}{\alpha} \le \frac{2^{\alpha}x^{\alpha}}{\alpha}.
\end{align*}
For the second integral, since $x + u\le 2u$ we have
\begin{align*}
\int_x^{\infty}(u + x)^{\alpha-1}e^{-u}du \le 2^{\alpha-1}\int_x^{\infty}u^{\alpha-1}e^{-u}du.
\end{align*}
Altogether, we get in \eqref{iden split},
\begin{align*}
\int_x^{\infty}r^{\alpha-1}e^{-r}dr \le \frac{2^{\alpha}x^{\alpha}e^{-x}}{\alpha} + 2^{\alpha-1}e^{-x}\int_x^{\infty}r^{\alpha-1}e^{-r}dr.
\end{align*}
Since $x \ge \alpha$, we have, $2^{\alpha-1}e^{-x} \le 1/2$, which completes the proof.
\end{proof}
We can now proceed to the proof of Lemma \ref{lem lp norm}
\begin{proof}[Proof of Lemma \ref{lem lp norm}]
To estimate the norm, notice that $ \left(1 + |\xi|^2 \right)^{-\beta/2} \le \min\left(1,|\xi|^{-\beta}\right)$, and so using polar coordinates
$$
\left\| \left(1 + |\xi|^2 \right)^{-\beta/2}e^{-\frac{t^2|\xi|_2^2}{2}}\right\|_{L_p(\mathbb R^n)}^p \le |\mathbb S^{n-1}|\int_0^{\infty}r^{n-1}\min\left(1,r^{-\beta p}\right)e^{-\frac{pt^2r^2}{2}}dr .
$$
Now,
\begin{align}
& \nonumber \int_0^{\infty}r^{n-1}\min\left(1,r^{-\beta p}\right)e^{-\frac{pt^2r^2}{2}}dr
\\
& \nonumber \qquad = \int_0^{1}r^{n-1}e^{-\frac{pt^2r^2}{2}}dr + \int_{1}^{\infty}r^{n-1-\beta p}e^{-\frac{pt^2r^2}{2}}dr
\\
& \qquad \label{splitting at 1} = \int_0^{1}r^{n-1}e^{-\frac{pt^2r^2}{2}}dr + \frac 1 2 \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\int_{\frac{pt^2}{2}}^{\infty}r^{\frac{n-\beta p}{2}-1}e^{-r}dr.
\end{align}
\noindent{\bf Case 1: Assume $pt^2\le 2$. } To bound the first term in \eqref{splitting at 1}, use the trivial bound
\begin{align} \label{bound gamma}
\int_0^1r^{n-1}e^{-\frac{pt^2r^2}{2}}dr \le \int_0^1r^{n-1}~dr = \frac 1 n.
\end{align}
To bound the second term, note first that
\begin{align} \label{splitting gamma}
&\int_{\frac{pt^2}2}^{\infty}r^{\frac{n-\beta p}{2}-1}e^{-r}dr = \int_{\frac{pt^2}2}^{1}r^{\frac{n-\beta p}{2}-1}e^{-r}dr + \int_{1}^{\infty}r^{\frac{n-\beta p}{2}-1}e^{-r}dr.
\end{align}
Since $pt^2 \le 2$,
\begin{align} \label{bound int r small}
\nonumber \int_{\frac{pt^2}2}^1r^{\frac{n-\beta p}2-1}e^{-r}dr & \le \int_{\frac{pt^2}2}^1r^{\frac{n-\beta p}{2}-1}dr
\\ & =
\begin{cases}
\frac{2}{n-\beta p} \left(1-\left(\frac{pt^2}{2}\right)^{\frac{n-\beta p}{2}} \right) \quad & \beta p \neq n, \\
\log \left(\frac{2}{pt^2}\right) \quad & \beta p = n, \\
\end{cases}
\end{align}
and also
\begin{align} \label{bound int r large}
\int_1^{\infty}r^{\frac{n-\beta p}2-1}e^{-r}dr \le
\begin{cases}
1 & \frac{n-\beta p}2-1 \le 0, \\
\Gamma \left(\frac{n-\beta p}{2} \right) & \frac{n-\beta p}2-1>0.
\end{cases}
\end{align}
Plugging \eqref{bound int r small} and \eqref{bound int r large} into \eqref{splitting gamma}, we get
\begin{multline*}
\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\int_{\frac{pt^2}{2}}^{\infty}r^{\frac{n-\beta p}{2}-1}e^{-r}dr
\le
\\
\begin{cases}
\frac{2}{n-\beta p} \left(\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}-1 \right) + \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\Gamma \left(\frac{n-\beta p}{2} \right) & \beta p <n-2,
\\
\frac{2}{n-\beta p} \left(\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}-1 \right) + \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}} & n-2\le \beta p <n,
\\
\log \left(\frac{2e}{pt^2} \right) & \beta p = n,
\\
\frac{2}{\beta p -n} \left(1-\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}} \right) + \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}} & \beta p >n.
\end{cases}
\end{multline*}
Now, if $a\ge 1$ then we have
\begin{align*}
\left|\frac{a^x-1}{x}\right| \le
\begin{cases}
a^x \log a & 0<x\le 1, \\
a^x & x\ge 1.
\end{cases}
\end{align*}
Thus, when $\beta p <n-2$ we have
\begin{multline*}
\frac{2}{n-\beta p} \left(\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}-1 \right) + \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\Gamma \left(\frac{n-\beta p}{2} \right)
\\ \le \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\left(1 + \Gamma \left(\frac{n-\beta p}{2} \right) \right)
\le 2\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\Gamma \left(\frac{n-\beta p}{2} \right).
\end{multline*}
When $n-2 \le \beta p <n$ we have
\begin{multline*}
\frac{2}{n-\beta p} \left(\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}-1 \right) + \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}
\\ \le \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\log \left(\frac{2}{pt^2} \right) + \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}
= \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\log \left(\frac{2e}{pt^2} \right).
\end{multline*}
Also, when $\beta p >n$ we use the fact that when $0< a \le 1$ and $x>0$,
\begin{align*}
\frac{1-a^x}{x} \le \log \left(\frac 1 a \right),
\end{align*}
and get
\begin{align*}
\frac{2}{\beta p-n} \left(1-\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}} \right) + \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}} &\le \log \left(\frac{2}{pt^2} \right) + \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}
\\ & \le \log \left(\frac{2}{pt^2} \right) + 1
\\ & = \log \left(\frac{2e}{pt^2} \right).
\end{align*}
Altogether,
\begin{align*}
\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\int_{\frac{pt^2}{2}}^{\infty}r^{\frac{n-\beta p}{2}-1}e^{-r}dr \le
\begin{cases}
2\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\Gamma \left(\frac{n-\beta p}{2} \right) & \beta p <n-2, \\
\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\log \left(\frac{2e}{pt^2} \right) & n-2\le \beta p <n, \\
\log \left(\frac{2e}{pt^2} \right)& \beta p \ge n.
\end{cases}
\end{align*}
Plugging this into \eqref{splitting at 1} and using \eqref{bound gamma}, we get
\begin{align*}
\int_0^{\infty}r^{n-1}\min\left(1,r^{-p\beta}\right)e^{-\frac{pt^2r^2}{2}}dr \le
\begin{cases}
\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\Gamma \left(\frac{n-\beta p}{2} \right) & \beta p <n-2 \\
\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\log \left(\frac{2e}{pt^2} \right)\ & n-2\le \beta p <n \\
\log \left(\frac{2e}{pt^2}\right)& \beta p \ge n,
\end{cases}
\end{align*}
which completes the proof in the case $pt^2 \le 2$.
\noindent{\bf Case 2: Assume $pt^2\ge 2$. }
To estimate the first term in \eqref{splitting at 1}, we consider two different cases. If $pt^2 \ge n$, choose
\begin{align*}
r_0 = \sqrt{\frac{n}{pt^2}\log\left(\frac{pt^2}{n}\right)},
\end{align*}
and then
\begin{align*}
\int_0^1r^{n-1}e^{-\frac{pt^2r^2}{2}}dr & = \int_0^{r_0}r^{n-1}e^{-\frac{pt^2r^2}{2}}dr + \int_{r_0}^1r^{n-1}e^{-\frac{pt^2r^2}{2}}dr
\\ & \le \int_0^{r_0}r^{n-1}~dr + \int_{r_0}^1re^{-\frac{pt^2r^2}{2}}dr
\\ & \le \frac{r_0^n}{n} + \frac{1}{pt^2}e^{-\frac{pt^2r_0^2}{2}}
\\ & = \frac 1 n\left(\frac n {pt^2}\log\left(\frac{pt^2}{n}\right)\right)^{n/2} + \frac{1}{pt^2}\left(\frac n {pt^2}\right)^{n/2}
\\ & \le \left(\frac n {pt^2}\log\left(\frac{ept^2}{n}\right)\right)^{n/2}.
\end{align*}
Otherwise, if $pt^2\le n$, choose
\begin{align*}
r_0 = e^{-\frac{pt^2}{n}}.
\end{align*}
Note that we have, say, $r_0 \ge 1/3$. Then, since $1-e^{-x}\le x$,
\begin{align*}
\int_0^1r^{n-1}e^{-\frac{pt^2r^2}{2}}dr & \le \frac{r_0^n}{n} + \frac{1}{pt^2}\left(e^{-\frac{pt^2r_0^2}{2}}-e^{-\frac{pt^2}{2}}\right)
\\ & \le \frac 1 ne^{-pt^2} + \frac {e^{-\frac{pt^2r_0^2}{2}}} {pt^2}\cdot \frac{pt^2(1-r_0^2)}{2}
\\ & \le \frac 1 n e^{-pt^2} + \frac{pt^2}{n} e^{-\frac{pt^2}{18}}
\\ & \le 2e^{-\frac{pt^2}{18}}.
\end{align*}
For the first term in \eqref{splitting at 1} we thus have (assuming that $n \ge 2$),
\begin{align} \label{1st term t large}
\int_0^1 r^{n-1}e^{-\frac{pt^2r^2}{2}}dr \le
\begin{cases}
\left(\frac n {pt^2}\log\left(\frac{ept^2}{n}\right)\right)^{n/2} & pt^2\ge n,
\\
2e^{-\frac{pt^2}{18}} & pt^2 \le n.
\end{cases}
\end{align}
If $n-\beta p \le 2$ then
\begin{align} \label{n small}
\int_1^{\infty}r^{n-\beta p-1}e^{-\frac{pt^2r^2}{2}}dr \le \int_1^{\infty}re^{-\frac{pt^2r^2}{2}}dr = \frac{e^{-\frac{pt^2}{2}}}{pt^2}\le e^{-\frac{pt^2}{2}}.
\end{align}
Otherwise, if $n-\beta p \ge 2$, then again we consider two different cases. If $pt^2\le n-\beta p$ , we have
\begin{align} \label{n large t small}
\nonumber \frac 1 2 \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\int_{\frac{pt^2}{2}}^{\infty}r^{\frac{n-\beta p}{2}-1}e^{-r}dr & \le \frac 1 2\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\int_{0}^{\infty}r^{\frac{n-\beta p}{2}-1}e^{-r}dr
\\ & = \frac 1 2\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\Gamma\left(\frac{n-\beta p}{2}\right).
\end{align}
Otherwise, suppose that we still have $n-\beta p \ge 2$, but now $pt^2 \ge n-\beta p$. Then by Proposition \ref{prop x alpha}, we have
\begin{align} \label{n large t large}
\frac 1 2\left(\frac 2 {pt^2}\right)^{\frac{n-\beta p}{2}}\int_{\frac{pt^2}{2}}^{\infty}r^{\frac{n-\beta p}{2}-1}e^{-r}dr \le \frac {2^{\frac{n-\beta p}{2}}e^{-\frac{pt^2}{2}}}{n-\beta p} \le 2^{\frac{n-\beta p}{2}-1}e^{-\frac{pt^2}{2}} .
\end{align}
Altogether, combining \eqref{n small}, \eqref{n large t small} and \eqref{n large t large}, we obtain
\begin{align} \label{2nd term t large}
\frac 1 2\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\int_{\frac{pt^2}{2}}^{\infty}r^{\frac{n-\beta p}{2}-1}e^{-r}dr
\le
\begin{cases}
2^{\frac{n-\beta p}{2}}e^{-\frac{pt^2}{2}}& 2\le n-\beta p \le pt^2,
\\
\left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\Gamma\left(\frac{n-\beta p}{2}\right) & 2 \le pt^2 \le n-\beta p ,
\\
e^{-\frac{pt^2}{2}} & n- \beta p \le2 \le pt^2.
\end{cases}
\end{align}
Plugging \eqref{1st term t large} and \eqref{2nd term t large} into \eqref{splitting at 1} gives
\begin{multline*}
\int_0^{\infty}r^{n-1}\min\left(1, r^{-\beta p}\right)e^{-\frac{pt^2r^2}{2}}dr
\le
\\
\begin{cases}
2e^{-\frac{pt^2}{18}} + \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\Gamma\left(\frac{n-\beta p}{2}\right) & 2 \le pt^2 \le n-\beta p,
\\
2e^{-\frac{pt^2}{18}} + 2^{\frac{n-\beta p}{2}}e^{-\frac{pt^2}{2}}& 2 \le n-\beta p \le pt^2 \le n,
\\
\left(\frac n {pt^2}\log\left(\frac{ept^2}{n}\right)\right)^{n/2} + 2^{\frac{n-\beta p}{2}}e^{-\frac{pt^2}{2}}& 2\le n-\beta p \le n \le pt^2,
\\
2e^{-\frac{pt^2}{18}} + e^{-\frac{pt^2}{2}} & n- \beta p \le 2 \le pt^2 \le n,
\\ \left(\frac n {pt^2}\log\left(\frac{ept^2}{n}\right)\right)^{n/2} + e^{-\frac{pt^2}{2}} & n- \beta p \le 2 \le n \le pt^2 .
\end{cases}
\end{multline*}
In order to simplify the last expression, first notice that when $n \le pt^2$, we have
\begin{align*}
e^{-\frac{pt^2}{2}} \le \left(\frac n {pt^2}\log\left(\frac{ept^2}{n}\right)\right)^{n/2}.
\end{align*}
Also, we have that whenever $pt^2\ge n-\beta p \ge 2$, since we have that $1- \log 2 >1/4$ we get the following estimate,
\begin{align*}
2^{\frac{n-\beta p}{2}}e^{-\frac{pt^2}{2}} \le e^{-\frac{pt^2}{2}\left(1 - \log 2\right)} \le e^{-\frac{pt^2}{8}} \le e^{-\frac{pt^2}{18}}.
\end{align*}
Hence, we conclude that,
\begin{align*}
\int_0^{\infty}r^{n-1}\min\left(1, r^{-\beta p}\right)e^{-\frac{pt^2r^2}{2}}dr
\le
\begin{cases}
2e^{-\frac{pt^2}{18}} + \left(\frac{2}{pt^2}\right)^{\frac{n-\beta p}{2}}\Gamma\left(\frac{n-\beta p}{2}\right) & 2 \le pt^2 \le n-\beta p,
\\
3e^{-\frac{pt^2}{18}}& n-\beta p \le pt^2 \le n,
\\
\left(\frac {2n} {pt^2}\log\left(\frac{ept^2}{n}\right)\right)^{n/2} & n \le pt^2.
\end{cases}
\end{align*}
This completes the proof of Lemma \ref{lem lp norm}.
\end{proof}
We are now in a position to prove Theorem \ref{thm decay}.
\begin{proof}[Proof of Theorem \ref{thm decay}]
we have,
\begin{align*}
\int_{\mathbb R^n}\big|\hat f_X (\xi )\big|e^{-\frac{t^2|\xi|_2^2}{2}}d\xi \le
\left\| \left(1 + |\xi|^2 \right)^{\beta/2}\hat f_X\right\|_{L_{p'}(\mathbb R^n)} ~ \left\| \left(1 + |\xi|^2 \right)^{-\beta/2}e^{-\frac{t^2|\xi|_2^2}{2}}\right\|_{L_p(\mathbb R^n)}.
\end{align*}
Since $1< p\le 2$, $\mathcal F:L_{p} \to L_{p'}$ is bounded with norm 1. Hence,
\begin{eqnarray*}
\left\| \left(1 + |\xi|^2 \right)^{\beta/2}\hat f_X\right\|_{L_{p'}(\mathbb R^n)} & = & \left\|\mathcal F \left(\mathcal F^{-1} \left( \left(1 + |\xi|^2 \right)^{\beta/2}\hat f_X \right) \right)\right\|_{L_{p'}(\mathbb R^n)}
\\ & \le & \left\|\mathcal F^{-1} \left( \left(1 + |\xi|^2 \right)^{\beta/2}\hat f_X \right)\right\|_{L_p(\mathbb R^n)}
\\ & \stackrel{\eqref{def sobolev}}{ = } & \|f_X\|_{\beta,p}.
\end{eqnarray*}
Altogether,
$$
\int_{\mathbb R^n}\big | \hat f_X (\xi)\big |e^{-\frac{t^2|\xi|_2^2}{2}}d\xi \le\left\|\ \left(1 + |\xi|^2 \right)^{-\beta/2}e^{-\frac{t^2|\xi|_2^2}{2}}\right\|_{L_p(\mathbb R^n)}\|f_X\|_{\beta,p}.
$$
Now use Lemma \ref{lem lp norm} to complete the proof.
\end{proof}
\smallskip
\subsection{Sobolev Embeddings and Theorem \ref{thm decay}}
We did not study the sharpness of Theorem \ref{thm decay}. In some cases, Sobolev Embedding Theorems can imply simpler proofs and better dependence in $t$. We are grateful for the referee who pointed this to us.
\noindent{\bf The case $n-\beta p <0$.}
In this case we can write
\begin{align}\label{beta p large}
\mathbb P\big(\|X\| \le t\big) = \int_{tK}f_X(x)dx \le |K|t^n\|f_X\|_{\infty}.
\end{align}
Since we assumed in Theorem \ref{thm decay} that $p\le 2$, we have just as in the proof of Theorem \ref{thm decay}
\begin{align}\label{bound sup}
\nonumber \|f_X\|_{\infty} \le \int_{\mathbb R^n}|\hat f_X(\xi)|d\xi & \le \left\|\left(1+|\xi|^2\right)^{-\frac{\beta}{2}}\right\|_{L_p(\mathbb R^n)}\left\|\left(1+|\xi|^2\right)^{\frac{\beta}{2}}\hat f_X(\xi)\right\|_{L_{p'}(\mathbb R^n)}
\\ & \le \left\|\left(1+|\xi|^2\right)^{-\frac{\beta}{2}}\right\|_{L_p(\mathbb R^n)}\|f_X\|_{\beta,p}.
\end{align}
Now,
\begin{align}\label{bound int potential}
\nonumber \left\|\left(1+|\xi|^2\right)^{-\frac{\beta}{2}}\right\|_{L_p(\mathbb R^n)} & = \left(\int_{\mathbb R^n}\frac{d\xi}{\left(1+|\xi|^2\right)^{\frac{\beta p}{2}}}\right)^{\frac 1 p}
\\ \nonumber & = |\mathbb S^{n-1}|^{\frac 1 p}\left(\int_0^{\infty}\frac{r^{n-1}dr}{\left(1+r^2\right)^{\frac{\beta p}{2}}}\right)^{\frac 1 p}
\\ & \le |\mathbb S^{n-1}|\left(\frac 1 {\beta p-n}+\frac 1 n\right)^{\frac 1 p}.
\end{align}
Plugging \eqref{bound sup} and \eqref{bound int potential} into \eqref{beta p large}, we get
\begin{align*}
\mathbb P\big(\|X\| \le t\big) \le |\mathbb S^{n-1}|\left(\frac 1 {\beta p-n}+\frac 1 n\right)^{\frac 1 p}|K|t^{n}\|f_X\|_{\beta,p}.
\end{align*}
While the bound gives a better dependence on $t$ when $t$ is small (as it removes the $\log$ term), its dependence on the other parameters can be worse as the implied constant tends to infinity as $\beta p \to n$.
\noindent{\bf The case $n-\beta p > 0$.}
The Sobolev Embedding Theorem (see e.g. \cite[Ch. 9]{Bre11} for the case where $\beta$ is an integer)
\begin{align*}
\|f\|_{L_q(\mathbb R^n)} \le C(\beta,n)\|f\|_{\beta,p},
\end{align*}
where $q = \frac{n p }{n- \beta p}$. Thus, we have
\begin{align}\label{beta p small}
\nonumber \mathbb P\big(\|X\| \le t\big) & = \int_{tK}f_X(x)dx \le |tK|^{\frac 1 {q'}}\|f_X\|_{L_q(\mathbb R^n)}
\\ & \le C(\beta,n)|K|^{\frac 1 {q'}}t^{\beta+\frac n {p'}}\|f_X\|_{\beta,p},
\end{align}
Note that once again, \eqref{beta p small} gives a better dependence on $t$ for small values of $t$. However, the term $|K|^{\frac 1 {q'}}$ might be worse than the term that appears in Theorem \ref{thm decay}.
\noindent{\bf The case $n-\beta p =0$.}
In this case we have
\begin{align*}
\|f\|_{L_q(\mathbb R^n)} \le C(\beta,n,q)\|f\|_{\beta,p}.
\end{align*}
where now $q\ge p$ and $C(\beta,n,q) \to \infty$ as $q\to \infty$. As before, using the Sobolev Embedding Theorem, we get
\begin{align*}
\mathbb P\big(\|X\| \le t\big) \le C(\beta,n,q)|K|^{\frac 1 {q'}}t^{\beta+\frac n {p'}}\|f_X\|_{\beta,p},
\end{align*}
which gives a better dependence in $t$ when $t$ is small, but possibly a worse dependence on the other parameters.
\section{Littlewood-Offord type estimates} \label{sec lo}
Let $a_1, \ldots, a_N$ be (deterministic) vectors in $\mathbb R^n$, and denote by $A$ the $N\times n$ matrix whose rows are $a_1, \ldots, a_N$. Let $\delta_1,\dots, \delta_N$ be i.i.d. random variables for which there exists $b \in (0,1)$ such that
\begin{align} \label{eq:small ball xi}
\sup_{x \in \mathbb R} \mathbb P \left( | \delta_i - x| \le 1 \right) \le 1-b .
\end{align}
Now, consider the random vector
\begin{align} \label{def x}
X = \sum_{k = 1}^N\delta_k a_k.
\end{align}
As in \cite{FS07, RV09}, the small ball estimate of $X$ involves the least common denominator of the matrix $A$. Thus, for $\alpha > 0$ and $\gamma\in \left(0,1 \right)$, define
\begin{align} \label{def lcd}
{\mathrm{LCD}}_{\alpha, \gamma} \left(A \right) \stackrel{\mathrm{def}}{ = } \inf \left\{ |\theta|_2 : \theta \in \mathbb R^n, d_2\left(A \theta, \mathbb Z^n \right) < \min\left(\gamma |A \theta| _2, \alpha \right) \right\}.
\end{align}
Recall that $|\cdot|_2$ denotes the Euclidean norm.
\begin{thm} \label{thm lo}
Let $X$ be defined as in \eqref{def x}, and assume that the $N\times n$ matrix $A$ satisfies $|A\theta|_2 \ge |\theta|_2$ for all $\theta$ in $\mathbb R^n$. Assume also that $t \ge \frac{\sqrt n}{\mathrm{LCD}_{\alpha,\gamma}(A)}$. Then
$$
\mathbb P \left(\|X\| \le t \right) \le \frac{|K|}{\gamma_n \left(K \right)} \left(\frac{C_K}{\pi}\right)^n \left( \left(\frac{t}{\gamma \sqrt b} \right)^n + \exp \left(-b \alpha^2 \right) \right),
$$
where $C_K$ is again the quasi-norm constant from \eqref{const quasi}.
In particular, for any $p > 0$, if we let $|x|_p = \left(\sum_{j=1}^n|x_j|^p\right)^{1/p}$ for $x = (x_1,\dots,x_n)\in \mathbb R^n$, then
$$
\mathbb P \left( |X|_p \le t n^{1/p} \right) \le \left(C\cdot C_p\right)^n \left( \left(\frac{t}{\gamma \sqrt b} \right)^n + \exp \left(-b \alpha^2 \right) \right) ,
$$
where $C_p = \min\left\{2^{1/p-1},1\right\}$.
\end{thm}
The first step of the proof is to estimate the small ball probability using the integer structure of the vectors $a_i$. To do that, for a given $\theta \in \mathbb R^n$, define
\begin{align} \label{def f}
f(\theta) \stackrel{\mathrm{def}}{ = } \inf_{m\in \mathbb Z^N}\left| \frac z t A\theta-m\right|_2.
\end{align}
\begin{lem}[Small ball estimate in terms of integer structure] \label{integ struc}
Let $X$ be a random vector as in \eqref{def x} and let $t > 0$. Then
$$
\mathbb P \left(\|X\| \le t \right) \le \frac{|K|}{\gamma_n(K)}\left(C'_Kt\right)^n\cdot \sup_{z \ge \frac{1}{2 \pi}}\int_{\mathbb R^n}e^{-4bf(\theta)^2-{|\theta|_2^2}/{2}}d\theta.
$$
\end{lem}
\begin{proof}
By \eqref{small ball smoothed} we have
$$
\mathbb P \left( \|X\| \le t \right) \le \frac{|K|}{\gamma_n \left(K \right)} \left(C'_Kt\right)^n\int_{\mathbb R^n} \big|\phi_{X}(\xi)\big|e^{-\frac{t^2|\xi|_2^2}{2}} d\xi.
$$
Setting $\theta = t\xi$,
\begin{align} \label{change var}
t^n \int_{\mathbb R^n} \big| \phi_{X} (\xi) \big|e^{-\frac{t^2|\xi|_2^2}{2}} d\xi = \int_{\mathbb R^n} \big| \phi_{X} ({\theta}/{t}) \big| e^{-\frac{ |\theta|_2^2}{2}} d\theta.
\end{align}
Using the definition of $X$, and the independence of $\delta_1,\ldots,\delta_N$, we have
\begin{multline} \label{char prod}
\left| \phi_{X} \left({\theta}/{t}\right) \right| = \mathbb E \exp\left(i\left\langle \sum_{i = 1}^N\delta_i a_i, \theta/t\right\rangle\right)
\\ = \prod_{k = 1}^N \mathbb E\exp\left(i\delta_k \frac{\langle a_k,\theta\rangle}{t} \right) = \prod_{k = 1}^N \left| \phi_\delta\left( \frac{\langle \theta, a_k \rangle}{t} \right) \right|,
\end{multline}
where $\delta$ is an independent copy of $\delta_1,\dots,\delta_N$. In order to estimate the right side of \eqref{char prod}, follow the conditioning argument that was used in \cite{FS07, RV09}. Let $\delta'$ be an independent copy of $\delta$, and denote by $\bar\delta$ the symmetric random variable $\delta-\delta'$. We have, $|\phi_{\delta} \left(\xi \right)|^2 = \mathbb E\cos \left(\xi\bar\delta \right)$. Using the inequality $|x|\le \exp \left(- \left(1-x^2 \right)/2 \right)$, which is valid for all $x\in\mathbb R$, we obtain
\begin{align} \label{bound char delta}
|\phi_{\delta} \left(\xi \right)| \le \exp \left( - \frac{\left(1-\mathbb E\cos \left(\xi\bar\delta \right) \right)}{2} \right).
\end{align}
By assumption \eqref{eq:small ball xi} it follows that $\mathbb P \left(|\bar\delta| \ge 1 \right)\ge b$. Therefore, by conditioning on $\bar\delta$, we get
\begin{align*}
1-\mathbb E\cos \left(\xi\bar\delta \right) & \ge \mathbb P \left(|\bar\delta|\ge 1 \right) \cdot \mathbb E \left( 1 - \cos \left(\xi \bar \delta \right) \Big| | \bar \delta | \ge 1 \right)
\\ & \ge b\mathbb E \left( 1 - \cos \left(\xi \bar \delta \right) \Big| | \bar \delta | \ge 1 \right).
\end{align*}
By the fact that $1 - \cos \theta \geq \frac{2}{\pi^2} \theta^2$, for any $|\theta|\le \pi$, we have for any $\theta\in\mathbb R$,
$$
1 - \cos \theta \ge\frac{2}{\pi^2} \min_{m \in \mathbb Z} |\theta - 2\pi m|^2.
$$
Hence,
\begin{align*}
1-\mathbb E\cos \left(\xi\bar\delta \right) & \ge \frac {2b} {\pi^2} \cdot \mathbb E\left( \min_{m \in \mathbb Z} \big| {\xi\bar\delta} - 2\pi m \big|^2 \Big| \big|{\bar\delta} \big| \ge 1 \right)
\\ & = 8b \cdot \mathbb E\left( \min_{m \in \mathbb Z} \big| {\xi\bar\delta} - m \big|^2 \Big| \big| {\bar\delta} \big| \ge 1/2\pi \right).
\end{align*}
Plugging this into \eqref{bound char delta} gives
\begin{align} \label{bound with exp}
\big|\phi_{\delta}(\xi)\big| \le \exp\left(-4b \mathbb E\left( \min_{m \in \mathbb Z} \big| {\xi\bar\delta} - m \big|^2 \Big| \big| {\bar\delta} \big| \ge 1/2\pi \right)\right).
\end{align}
Replacing the conditional expectation with supremum over all the possible values $z \ge 1/2\pi$ and using Jensen's inequality, we get
\begin{eqnarray}
\nonumber && \int_{\mathbb R^n} \left|\phi_X\left({\theta}/{t}\right)\right|e^{-{|\theta|_2^2}/{2}}d\theta
\\ \nonumber && \qquad \stackrel{\eqref{char prod}}{ = } \int_{\mathbb R^n}\prod_{k = 1}^N\left|\phi_{\delta}\left(\frac{\langle \theta,a_k\rangle}{t}\right)\right|e^{-{|\theta|_2^2}/{2}}d\theta \\
\nonumber && \qquad \stackrel{\eqref{bound with exp}}{\le} \int_{\mathbb R^n} \exp \left( -4b\cdot \mathbb E \left( \sum_{k = 1}^N \min_{m \in \mathbb Z} \left| {\frac{\langle \theta, a_k \rangle}{t} \bar\delta} - m \right|^2 \Bigg| \big| {\bar\delta} \big| \ge 1/2\pi \right) -{|\theta|_2^2}/{2}\right) d\theta \\
\nonumber && \qquad \le \mathbb E \left[ \int_{\mathbb R^n} \exp \left( -4b \min_{m \in \mathbb Z^N} \left| {\frac{\bar\delta}{t} A\theta} - m \right|_2^2 -{|\theta|_2^2}/{2}\right) d\theta \Bigg| \big| {\bar\delta} \big| \ge 1/2\pi \right] \\
\nonumber & & \qquad \le \sup_{z \ge 1/2\pi} \int_{\mathbb R^n} \exp \left( -4b f \left(\theta \right)^2-{|\theta|_2^2}/{2} \right)d\theta .
\end{eqnarray}
Using \eqref{change var} the result follows.
\end{proof}
Define the set
$$
T_s \stackrel{\mathrm{def}}{ = } \left\{\theta\in \mathbb R^n : f(\theta) \le s\right\}.
$$
The next step in the proof is to rewrite the integral that appears in Lemma \ref{integ struc} in the following way:
\begin{multline} \label{int level set}
\int_{\mathbb R^n} \exp \left( -4b f \left(\theta \right)^2 \right)\exp \left( -|\theta|_2^2 /2 \right) d\theta
\\ = \int_{\mathbb R^n} \int_{s \ge f \left(\theta \right)} 8bs \exp \left( -4b s^2 \right) ds \exp \left(-|\theta|_2^2 /2 \right) d\theta
\\ = \left(2\pi \right)^{n/2} \int_0^{\infty} 8bs \exp \left(-4bs^2 \right) \gamma_n \left(T_s\right) ds,
\end{multline}
which means that we have to bound $\gamma_n(T_s)$. To do that, we start with the following covering lemma.
\begin{lem}[Covering of $T_s$] \label{lem covering}
Let $\alpha > 0$ and $\gamma \in (0,1)$. Assume that $t \ge \frac{\sqrt n}{\mathrm{LCD}_{\alpha, \gamma}(A)}$. Assume also that $|A\theta|_2 \ge |\theta|_2$ for all $\theta \in \mathbb R^n$. If $0 \le s \le \alpha/2$, then there exist vectors $\{x_i\}_{i \in I}\subseteq \mathbb R^n$ such that
\begin{align} \label{good covering}
T_s \subseteq \bigcup_{i\in I}B(x_i, r) \text{ and } |x_i-x_{i'}|_2 \ge R, \forall i \neq i',
\end{align}
where $r = \frac{2st}{\gamma z}$ and $R = \frac{\sqrt n}{z}$. Moreover, for any $j \ge 1$,
\begin{align} \label{small shells}
{\mathrm{card}} \left( \left\{ i \in I : j R \le |x_i|_2 < \left( j + 1 \right)R \right\} \right) \le n2^n \left(j + 1 \right)^{n-1}.
\end{align}
\end{lem}
\begin{proof}
Let $\theta_1, \theta_2 \in T_s$. By \eqref{def f}, there exists $p_1, p_2 \in \mathbb Z^N$ such that
$$
\left| \frac{z}{t} A \theta_1 - p_1\right| \le s \quad \hbox{and} \quad \left| \frac{z}{t} A \theta_2 - p_2\right| \le s.
$$
By the triangle inequality,
$$
\left| \frac{z}{t} A \left(\theta_1-\theta_2 \right) - \left(p_1-p_2 \right) \right| \le 2s ,
$$
which means that $d_2 \left(A\tau, \mathbb Z^N \right) \le 2s \le \alpha$, where $\tau = z\left(\theta_1 - \theta_2 \right)/t$. By \eqref{def lcd} this implies that either
$$
|\tau|_2 \ge {\mathrm{LCD}}_{\alpha, \gamma} \left(A \right) ,
$$
or
$$
\alpha \ge 2 s \ge d_2\left(A\tau, \mathbb Z^N \right) \ge \min\left( \gamma |A \tau|_2, \alpha \right) = \gamma |A\tau|_2.
$$
By the assumptions that $|A \tau|_2 \ge |\tau|_2$ and ${\mathrm{LCD}}_{\alpha, \gamma} \left(A \right) \ge \sqrt n / t$, we conclude that
$$
\hbox{either } \quad |\theta_1 - \theta_2|_2 \ge \frac{\sqrt n}{z} = : R \quad \hbox{ or } \quad |\theta_1 - \theta_2|_2 \le \frac{2s t}{\gamma z} = : r.
$$
Hence, $T_s$ can be covered by a union of Euclidean balls of radius $r$ whose centers are $R$-separated, which proves \eqref{good covering}. Next, for $j\ge 1$, let
$$
M_j \stackrel{\mathrm{def}}{ = } {\mathrm{card}} \left( \left\{i \in I : jR \le |x_i|_2 \le (j + 1)R \right\} \right).
$$
To estimate $M_j$, use a well-known volumetric argument. Indeed, since $\{x_i \}_{i \in I}$ are $R$-separated, we know that the Euclidean balls $B \left(x_i, R/2 \right)$ are disjoint and contained in the shell
$$
\left\{y \in \mathbb R^n : \left(j-1/2 \right)R \le |y|_2 \le \left(j + 3/2 \right) R\right\}.
$$
Hence, taking the volume,
\begin{align*}
M_j \left( \frac{R}{2} \right)^n & \le R^n \left( \left(j + 3/2 \right)^n - \left(j-1/2 \right)^n \right)
\\ & = R^n \left(j + 1/2 \right)^n \left( \left(1 + \frac{2}{2j + 1} \right)^n - \left(1- \frac{2}{2j + 1} \right)^n \right).
\end{align*}
Since for every $x \in \left(0,1 \right)$, we have $ \left(1 + x \right)^n - \left(1-x \right)^n \le 2nx \left(1 + x \right)^{n-1}$, we conclude that
$$
M_j \le n2^n \left(j + 1 \right)^{n-1}.
$$
This completes the proof.
\end{proof}
Using the covering lemma, we can now prove the required volume estimate.
\begin{cor} \label{cor measure}
Let $r$ and $R$ be as in Lemma \ref{lem covering}. If $R \ge 2 r$, then
$$
\gamma_n \left(T_s \right) \le \left(\frac{Cr}{R} \right)^n = \left( \frac{2 C t s}{\gamma \sqrt n} \right)^n.
$$
\end{cor}
\begin{proof}
Let $y \in \mathbb R^n$, we have
$$
\gamma_n \left( B \left(y,r \right) \right) = \frac{1}{ \left(2\pi \right)^{n/2}} \int_{|y-x|_2 \le r} e^{-\frac{|x|_2^2}{2}} dx.
$$
Since $|x|_2^2 + |x-y|_2^2 = \frac{1}{2} \left(|y|_2^2 + |2x-y|_2^2 \right) \ge \frac{1}{2} |y|_2^2$,
$$
\gamma_n \left( B \left(y,r \right) \right) \le \frac{1}{ \left(2\pi \right)^{n/2}} e^{- \frac{|y|_2^2}{4}} \int_{|y-x|_2 \le r} e^{\frac{|y-x|_2^2}{2}} dx.
$$
Therefore, if $|y|_2 \ge R\ge 2r$,
\begin{align} \label{eq:gaussian}
\nonumber \gamma_n \left( B \left(y,r \right) \right) & \le \frac{1}{ \left(2\pi \right)^{n/2}} \exp \left(- \frac{|y|_2^2}{4} \right) e^{r^2/2} |B \left(0,r \right)|
\\ & \le \frac{1}{ \left(2\pi \right)^{n/2}} \exp \left( - \frac{|y|_2^2}{8} \right) |B \left(0,r \right)|.
\end{align}
Assume that $s$ is such that $r \le R/2$, i.e. $4 t s\le \gamma \sqrt n$, then by \eqref{good covering}
$$
\gamma_n \left(T_s \right) \le \sum_{i \in I} \gamma_n \left( B \left(x_i, r \right) \right) \le \sum_{j = 0}^{ \infty} \sum_{i \in I : jR \le |x_i|_2 < \left(j + 1 \right)R} \gamma_n \left( B \left(x_i, r \right) \right).
$$
Also, for $j \ge 1$, we have by \eqref{small shells}
$$
{\mathrm{card}} \left( \left\{i \in I :j R \le |x_i|_2 < \left(j + 1 \right)R \right\} \right) \le C^n j^{n-1}.
$$
By $ \left(\ref{eq:gaussian} \right)$,
$$
\gamma_n \left( B \left(x_i, r \right) \right) \le \frac{1}{ \left(2\pi \right)^{n/2}} \exp \left( - \frac{j^2 R^2}{8} \right) |B (0,r )|.
$$
Hence,
\begin{align} \label{bound gamma t}
\nonumber\gamma_n \left(T_s \right) & \le \gamma_n \left( B \left(0,r \right) \right) + \sum_{j = 1}^{ \infty} \left(\frac{C^2}{2 \pi} \right)^{n/2} j^{n-1}\exp \left( - \frac{j^2 R^2}{8} \right) |B \left(0,r \right)| \\
& \le \frac{ |B \left(0,r \right)| }{ \left(2\pi \right)^{n/2}} \left(1 + C^n \sum_{j = 1}^{ \infty} j^{n-1} \exp \left( - \frac{j^2 R^2}{8} \right) \right).
\end{align}
The function $v \mapsto v^{n-1} e^{-v^2 R^2 /8}$ is decreasing for $v \ge 2 \sqrt n / R$. By comparing series with integrals,
\begin{align*}
\sum_{j = 1}^{\infty} j^{n-1} \exp \left( - \frac{j^2 R^2}{8} \right) & \le \left( \frac{2 \sqrt n}{R} \right)^n + \int_0^{\infty} v^{n-1} e^{-v^2 R^2 /8} dv \\
& = \left( \frac{2 \sqrt n}{R} \right)^n + \frac{8^{n/2}}{R^n} \int_0^{\infty} u^{\frac{n-1}{2} - 1} e^{-u} du \le \left( \frac{C n^{1/2}}{R} \right)^{n}.
\end{align*}
Since $z \ge 1/ 2\pi$, we have $R \le 2\pi \sqrt n$, so that
$$
\left(1 + C^n \sum_{j = 1}^{\infty} j^{n-1} \exp \left( - \frac{j^2 R^2}{8} \right) \right) \le \left( \frac{C_1 n^{1/2}}{R}\right)^{n}.
$$
Moreover, it is well-known that $|B \left(0,r \right)| \le C_2^n n^{-n/2} r^n$ which implies by \eqref{bound gamma t} that
$$
\gamma_n \left(T_s \right) \le \left(\frac{Cr}{R} \right)^n = \left( \frac{2 C t s}{\gamma \sqrt n} \right)^n,
$$
which completes the proof.
\end{proof}
We are now in a position to prove Theorem \ref{thm lo}.
\begin{proof}[Proof of Theorem \ref{thm lo}]
By Lemma \ref{integ struc} and \eqref{int level set}, to have a small ball estimate it is enough to evaluate the integral
$$
\int_0^{\infty}8bs\exp\left(-4bs^2\right)\gamma_n\left(T_s\right)ds.
$$
We have,
\begin{align*}
&\int_0^{\infty}8bs\exp\left(-4bs^2\right)\gamma_n\left(T_s\right)ds
\\ & \quad \quad = \int_0^{\alpha/2}8bs\exp\left(-4bs^2\right)\gamma_n\left(T_s\right)ds + \int_{\alpha/2}^{\infty}8bs\exp\left(-4bs^2\right)\gamma_n\left(T_s\right)ds
\\ & \quad \quad \le \int_0^{\alpha/2}8bs\exp\left(-4bs^2\right)\gamma_n\left(T_s\right)ds + \exp\left(-b\alpha^2\right).
\end{align*}
Assume first that $\alpha \le \frac{\gamma \sqrt n}{2t}$ so that for any $s \le \alpha/2$ we have $R \ge 2r$. By Corollary \ref{cor measure},
$$
\gamma_n \left(T_s\right) \le \left( \frac{2Cts}{\gamma \sqrt n} \right)^n ,
$$
and so
\begin{align*}
\int_0^{\alpha/2} 8bs \exp \left(-4bs^2 \right) \gamma_n \left(T_s\right) ds & \le\int_0^{\alpha/2} 8bs \exp \left(-4bs^2 \right) \left( \frac{2 C t s}{\gamma \sqrt n} \right)^n ds \\
& \le 8b \left(\frac{2Ct}{\gamma \sqrt n} \right)^n \int_0^{ \infty} s^{n + 1} e^{-4bs^2} ds \\
& = \left(\frac{Ct}{\gamma \sqrt b \sqrt n} \right)^n \int_0^{ \infty} u^{n/2} e^{-u} du \\
& \le \left(\frac{C' t}{\gamma \sqrt b} \right)^n.
\end{align*}
Assume otherwise that $\alpha \ge \frac{\gamma \sqrt n}{2t} \stackrel{\mathrm{def}}{ = } \alpha_0$. Then, as before,
$$
\int_0^{\infty} 8bs \exp \left(-4bs^2 \right) \gamma_n \left(T_s\right) ds \le \int_0^{\alpha_0/2} 8bs \exp \left(-4bs^2 \right) \gamma_n \left(T_s\right) ds + \exp \left(-b \alpha_0^2 \right).
$$
For $s \le \alpha_0/2$ we do exactly the same computation as in the first case and obtain
$$
\int_0^{\infty} 8bs \exp \left(-4bs^2 \right) \gamma_n \left(T_s\right) ds \le \left(\frac{C' t}{\gamma \sqrt b} \right)^n + \exp \left(-b \alpha_0^2 \right).
$$
In this case, we also have
$$
\exp \left(-b \alpha_0^2 \right) = \exp \left(-\frac{b \gamma^2 n}{4 t^2} \right) \le \left(\frac{C t}{\gamma \sqrt b} \right)^n.
$$
This concludes the fact that
$$
\int_{\mathbb R^n} \exp \left( -4b f \left(\theta \right)^2 \right)\exp \left( -|\theta|_2^2 /2 \right) d\theta \le \left(\frac{C t}{\gamma \sqrt b} \right)^n + \exp \left(-b \alpha^2 \right).
$$
Using Lemma \ref{integ struc}, Theorem \ref{thm lo} follows.
\end{proof}
\bibliographystyle{amsalpha}
|
\section{Introduction}
\label{sec:intro}
In many machine learning applications, in the process of training a high-accuracy classifier,
the primary bottleneck in time and effort is often the annotation of the large quantities of data required
for supervised learning.
Active learning is a protocol designed to reduce this cost by allowing the
learning algorithm to sequentially identify highly-informative data points to be annotated.
In the specific model we study below, called \emph{pool-based} active learning,
the learning algorithm is initially given access to a large pool of unlabeled data points,
which are considered inexpensive and abundant.
It is then able to select any unlabeled data point from the pool and request to observe its label.
Given the label of this data point, the algorithm can then select another unlabeled data point
to be labeled, and so on. This interactive process continues until at most some prespecified
number of rounds is reached, at which time the algorithm must halt and produce a classifier.
This contrasts with \emph{passive learning}, where the learning algorithm would be given
access to a number of \emph{random} labeled data points. The hope is that, by sequentially
selecting the data points to be labeled, the active learning algorithm can direct the annotation
effort toward only the highly-informative data points, given the information already gathered
by the previously-labeled data points, and thereby reduce the total number of data points
required to produce a classifier capable of predicting the labels of new instances with at most
a desired error rate.
This model of active learning has been successfully applied to a variety of learning problems,
often with significant reductions in the number of label observations required to obtain a
given error rate for the resulting classifier \citep*[see][for a survey of several such applications]{settles:12}.
This article studies the theoretical capabilities of active learning, regarding the number
of label requests sufficient to learn a classifier to a desired error rate, known as the \emph{label complexity}.
There is now a substantial literature on this subject \citep*[see][for a survey of known results]{hanneke:survey},
but on the important question of \emph{optimal} performance in the general setting, the gaps present
in the literature are quite substantial in some cases. In this work, we address this question by carefully
studying the \emph{minimax} performance. Specifically, we are interested in the \emph{minimax label complexity},
defined as the smallest (over the choice of active learning algorithm) worst-case
number of label requests sufficient for the active learning algorithm to produce a classifier of a specified error rate,
in the context of various noise models (e.g., Tsybakov noise, bounded noise, agnostic noise, etc.).
We derive upper and lower bounds on the minimax label complexity for several noise models,
which reveal a variety of interesting and (in some cases) surprising observations.
Furthermore, in establishing the upper bounds, we propose a novel active learning strategy,
which often achieves significantly smaller label complexities than the active learning methods
studied in the prior literature.
\subsection{The Prior Literature on the Theory of Active Learning}
Before getting into the technical details, we first review some background information
about the prior literature on the theory of active learning. This will also allow us to
introduce the key contributions of the present work.
The literature on the theory of active learning began with studies of the \emph{realizable case},
a setting in which the labels are assumed to be consistent with some classifier in a known hypothesis class,
and have no noise \citep*{cohn:94,freund:97,dasgupta:04,dasgupta:05}. In this simple setting,
\citet*{dasgupta:05} supplied the first general analysis of the label complexity of active learning,
applicable to arbitrary hypothesis classes. However, \citet*{dasgupta:05} found that there are a range
of minimax label complexities, depending on the structure of the hypothesis class, so that even
among hypothesis classes of roughly the same minimax sample complexities for passive learning,
there can be widely varying minimax label complexities for active learning. In particular, he found
that some hypothesis classes (e.g., interval classifiers) have minimax label complexity essentially no better than that of passive learning,
while others have a minimax label complexity exponentially smaller than that of passive learning (e.g., threshold classifiers).
Furthermore, most nontrivial hypothesis classes of interest in learning theory seem to fall into the former category,
with minimax label complexities essentially no better than passive learning. Fortunately, \citet*{dasgupta:05}
also found that in some of these hard cases, it is still possible to show improvements over passive learning
under restrictions on the data distribution.
Stemming from these observations, much of the literature on active learning in the
realizable case has focused on describing various special conditions under which the label complexity of
active learning is significantly better than that of passive learning: for instance, by placing restrictions
on the marginal distribution of the unlabeled data \citep*[e.g.,][]{dasgupta:05b,balcan:07,el-yaniv:12,balcan:13,hanneke:survey},
or abandoning the minimax approach by expressing the label complexity with an explicit dependence on the optimal classifier \citep*[e.g.,][]{dasgupta:05,hanneke:10a,hanneke:thesis,hanneke:12a}.
In the general case, such results have been abstracted into various distribution-dependent (or sometimes data-dependent)
\emph{complexity measures}, such as the \emph{splitting index} \citep*{dasgupta:05},
the \emph{disagreement coefficient} \citep*{hanneke:07b,hanneke:thesis},
the \emph{extended teaching dimension growth function} \citep*{hanneke:07a},
and the related \emph{version space compression set size} \citep*{el-yaniv:10,el-yaniv:12,hanneke:14a}.
For each of these, there are general upper bounds (and in some cases, minimax lower bounds) on the label complexities
achievable by active learning methods in the realizable case, expressed in terms of the complexity measure.
By expressing bounds on the label complexity in terms of these quantities, the analysis of the label complexities
achievable by active learning methods in the realizable case has been effectively reduced to the problem of
bounding one of these complexity measures.
In particular, these complexity measures are capable of exhibiting a range of behaviors, corresponding to the
range of label complexities achievable by active learning. For certain values of the complexity measures,
the resulting bounds reveal significant improvements over the minimax sample complexity of passive learning,
while for other values, the resulting bounds are essentially no better than the
minimax sample complexity of passive learning.
Moving beyond these initial studies of the realizable case, the more-recent literature has developed active
learning algorithms that are provably robust to label noise. This advance was initiated by the seminal
work of \citet*{balcan:06,balcan:09} on the $A^{2}$ (Agnostic Active) algorithm, and continued by a
number of subsequent works \citep*[e.g.,][]{dasgupta:07,balcan:07,castro:06,castro:08,hanneke:07a,hanneke:09a,hanneke:thesis,hanneke:11a,hanneke:12a,minsker:12,koltchinskii:10,beygelzimer:09,beygelzimer:10,hsu:thesis,ailon:12,hanneke:12b}.
When moving into the analysis of label complexity in noisy settings, the literature continues to follow the same intuition
from the realizable case: that is, that there should be some active learning problems that are inherently hard, sometimes no better than passive learning,
while others are significantly easier, with significant savings compared to passive learning. As such, the general label complexity bounds
proven in noisy settings have tended to follow similar patterns to those found in the realizable case. In some scenarios, the
bounds reflect interesting savings compared to passive learning, while in other scenarios the bounds do not reflect any
improvements at all. However, unlike the realizable case, these upper bounds on the label complexities of the various proposed methods
for noisy settings lacked complementary minimax lower bounds showing that they were accurately describing the fundamental
capabilities of active learning in these settings.
For instance, in the setting of Tsybakov noise, there are essentially only two types of general lower bounds on the minimax label complexity in the prior literature:
(1) lower bounds that hold for all nontrivial hypothesis classes of a given VC dimension, which therefore reflect a kind of best-case scenario \citep*{hanneke:11a,hanneke:survey},
and (2) lower bounds inherited from the realizable case (which is a special case of Tsybakov noise).
In particular, both of these lower bounds are always smaller than the minimax sample complexity of passive learning under Tsybakov noise.
Thus, although the upper bounds on the label complexity of active learning in the literature are sometimes no better than the minimax sample complexity of passive learning,
the existing lower bounds are unable to confirm that active learning truly cannot outperform passive learning in these scenarios.
This gap in our understanding of active learning with noise has persisted for a number of years now, without really receiving a good
explanation for why the gap exists and how it might be closed.
In the present work, we show that there is a very good reason for why better lower bounds have not been
discovered in general for the noisy case. For certain ranges of the noise parameters (corresponding to the high-noise regime),
these simple lower bounds are actually \emph{tight} (up to certain constant and logarithmic factors):
that is, the upper bounds can actually be reduced to nearly match these basic lower bounds.
Proving this surprising fact requires the introduction of a new type of active learning strategy, which selects its
queries based on both the structure of the hypothesis class and the estimated variances of the labels.
In particular, in these high-noise regimes, we find that \emph{all} hypothesis classes of the same VC dimension
have essentially the same minimax label complexities (up to logarithmic factors), in stark contrast to the
well-known differentiation of hypothesis classes observed in the realizable case by \citet*{dasgupta:05}.
For the remaining range of the noise parameters (the low-noise regime), we argue that the label complexity
takes a value sometimes larger than this basic lower bound, yet still typically smaller than the known upper bounds.
In this case, we further argue that the minimax label complexity is well-characterized by a simple combinatorial
complexity measure, which we call the \emph{star number}.
In particular, these results reveal that for nonextremal parameter values, the minimax label complexity of active
learning under Tsybakov noise with \emph{any} VC class is \emph{always} smaller than that of passive learning,
a fact not implied by any results in the prior literature.
We further find that the star number can be
used to characterize the minimax label complexities for a variety of other noise models. Interestingly,
we also show that almost all of the distribution-dependent or data-dependent complexity measures from the prior literature on the label complexity
of active learning are exactly \emph{equal} to the star number when maximized over the choice of distribution or data set (including all of those mentioned above).
Thus, the star number represents a unifying core concept within these disparate styles of analysis.
\subsection{Our Contributions}
Below, we summarize a few of the main contributions and interesting implications of this work.
\begin{itemize}
\item[$\bullet$] We develop a general noise-robust active learning strategy, which unlike previously-proposed general methods,
selects its queries based on both the structure of the hypothesis class \emph{and} the estimated variances of the labels.
\item[$\bullet$] We obtain the first near-matching general distribution-free upper and lower bounds on the minimax label complexity of active learning, under a variety of noise models.
\item[$\bullet$] In many cases, the upper bounds significantly improve over the best upper bounds implied by the prior literature.
\item[$\bullet$] The upper bounds for Tsybakov noise \emph{always} reflect improvements over the minimax sample complexity of passive learning (for non-extremal noise parameter values), a feat not previously known to be possible.
\item[$\bullet$] In high-noise regimes of Tsybakov noise, our results imply that
all hypothesis classes of a given VC dimension have roughly the same minimax label complexity (up to logarithmic factors),
in contrast to well-known results for bounded noise. This fact is not implied by any results in the prior literature.
\item[$\bullet$] We express our upper and lower bounds on the label complexity in terms of a simple combinatorial complexity measure, which we refer to as the \emph{star number}.
\item[$\bullet$] We show that for any hypothesis class, almost every complexity measure proposed to date in the active learning literature has worst-case value exactly \emph{equal} to the star number,
thus unifying the disparate styles of analysis in the active learning literature.
We also prove that the doubling dimension is bounded if and only if the star number is finite.
\item[$\bullet$] For most of the noise models studied here, we exhibit examples of hypothesis classes spanning the gaps between
the upper and lower bounds, thus demonstrating that the gaps cannot generally be reduced (aside from logarithmic factors)
without introducing additional complexity measures.
\item[$\bullet$] We prove a separation result for Tsybakov noise vs the Bernstein class condition, establishing
that the respective minimax label complexities can be significantly different. This contrasts with passive learning, where
they are known to be equivalent up to a logarithmic factor.
\end{itemize}
The algorithmic techniques underlying the proofs of the most-interesting of our upper bounds involve a combination
of the disagreement-based strategy of \citet*{cohn:94} (and the analysis thereof by \citealp*{hanneke:11a}, and \citealp*{hanneke:14a}),
along with a repeated-querying technique of \citet*{kaariainen:06}, modified to account for variations in label variances
so that the algorithm does not waste too many queries determining the optimal classification of highly-noisy points;
this modification represents the main algorithmic innovation in this work. In a supporting role, we also rely on
auxiliary lemmas on the construction of $\varepsilon$-nets and $\varepsilon$-covers based on random samples, and the
use of these to effectively discretize the instance space.
The mathematical techniques underlying the proofs of the lower bounds are largely taken directly from the
literature. Most of the lower bounds are established by a combination of a technique originating with \citet*{kaariainen:06}
and refined by \citet*{beygelzimer:09} and \citet*{hanneke:11a,hanneke:survey}, and a technique of \citet*{raginsky:11} for
incorporating a complexity measure into the lower bounds.
We note that, while the present work focuses on the distribution-free setting, in which the marginal distribution over the
instance space is unrestricted, our results reveal that low-noise settings can still benefit from distribution-dependent
analysis, as expected given the aforementioned observations by \citet*{dasgupta:05} for the realizable case. For instance,
under Tsybakov noise, it is often possible to obtain stronger upper bounds in low-noise regimes under
assumptions restricting the distribution of the unlabeled data \citep*[see e.g.,][]{balcan:07}. We leave for future work the important problem of
characterizing the minimax label complexity of active learning in the general case for an arbitrary fixed marginal distribution
over the instance space.
\subsection{Outline}
The rest of this article is organized as follows.
Section~\ref{sec:definitions} introduces the formal setting and basic notation used throughout,
followed in Section~\ref{sec:noise-models} with the introduction of the noise models studied in this work.
Section~\ref{sec:star} defines a combinatorial complexity measure -- the star number -- in terms of which we will express the label complexity bounds below.
Section~\ref{sec:main} provides statements of the main results of this work: upper and lower bounds on the minimax label complexities of active learning under each of the noise models defined in Section~\ref{sec:noise-models}.
That section also includes a discussion of the results, and a brief sketch of the arguments underlying the most-interesting among them.
Section~\ref{sec:passive} compares the results from Section~\ref{sec:main} to the known results on the minimax sample complexity of passive learning, revealing which scenarios yield improvements of active over passive.
Next, in Section~\ref{sec:complexities}, we go through the various results on the label complexity of active learning from the literature, along with their
corresponding complexity measures (most of which are distribution-dependent or data-dependent).
We argue that all of these complexity measures are exactly equal to the star number when maximized over the
choice of distribution or data set. This section also relates the star number to the well-known concept of \emph{doubling dimension},
in particular showing that the doubling dimension is bounded if and only if the star number is finite.
We note that the article is written with the intention that it be read in-order; for instance,
while Appendix~\ref{app:main-proofs} contains proofs of the results in Section~\ref{sec:main},
those proofs refer to quantities and results introduced in Sections \ref{sec:passive} and \ref{sec:complexities}
(which follow Section~\ref{sec:main}, but precede Appendix~\ref{app:main-proofs}).
\section{Definitions}
\label{sec:definitions}
The rest of this paper makes use of the following formal definitions.
There is a space $\mathcal X$, called the \emph{instance space}.
We suppose $\mathcal X$ is equipped with a $\sigma$-algebra ${\cal B}_{\mathcal X}$, and for simplicity we will assume $\{\{x\} : x \in \mathcal X\} \subseteq {\cal B}_{\mathcal X}$.
There is also a set $\mathcal Y = \{-1,+1\}$, known as the \emph{label space}.
Any measurable function $h : \mathcal X \to \mathcal Y$ is called a \emph{classifier}.
There is an arbitrary set $\mathbb{C}$ of classifiers, known as the \emph{hypothesis class}.
To focus on nontrivial cases, we suppose $|\mathbb{C}| \geq 3$ throughout.
For any probability measure $P$ over $\mathcal X \times \mathcal Y$ and any $x \in \mathcal X$, define $\eta(x; P) = \P( Y = +1 | X=x )$ for $(X,Y) \sim P$,
and let $f^{\star}_{P}(x) = {\rm sign}( 2\eta(x;P) - 1 )$ denote the \emph{Bayes optimal classifier},\footnote{Since
conditional probabilities are only defined up to probability zero differences, there can be multiple valid functions $\eta(\cdot;P)$ and $f^{\star}_{P}$,
with any two such functions being equal with probability one. As such, we will interpret statements such as ``$f^{\star}_{P} \in \mathbb{C}$'' to mean that
\emph{there exists} a version of $f^{\star}_{P}$ contained in $\mathbb{C}$, and similarly for other claims and conditions for $f^{\star}_{P}$ and $\eta(\cdot;P)$.}
where ${\rm sign}(t) = +1$ if $t \geq 0$, and ${\rm sign}(t) = -1$ if $t < 0$.
Define the \emph{error rate} of a classifier $h$ with respect to $P$ as
${\rm er}_{P}(h) = P((x,y) : h(x) \neq y)$.
In the learning problem, there is a \emph{target distribution} $\mathcal{P}_{XY}$ over $\mathcal X \times \mathcal Y$,
and a \emph{data sequence} $(X_{1},Y_{1}),(X_{2},Y_{2}),\ldots$, which are independent $\mathcal{P}_{XY}$-distributed random variables.
However, in the active learning protocol, the $Y_{i}$ values are initially ``hidden'' until individually requested by the algorithm (see below).
We refer to the sequence $X_{1},X_{2},\ldots$ as the \emph{unlabeled data sequence}.\footnote{Although, in practice, we would expect to have
access to only a finite number of unlabeled samples, we expect this number would often be quite large (as unlabeled samples are considered
inexpensive and abundant in many applications). For simplicity, and to focus the analysis purely on the number of \emph{labels} required
for learning, we approximate this scenario by supposing an \emph{inexhaustible} source of unlabeled samples. We leave open the question
of the number of unlabeled samples sufficient to obtain the minimax label complexity; in particular, we expect the number of such samples
used by the methods obtaining our upper bounds to be quite large indeed.}
We will sometimes denote by $\mathcal{P}$ the marginal distribution of $\mathcal{P}_{XY}$ over $\mathcal X$: that is, $\mathcal{P}(\cdot) = \mathcal{P}_{XY}(\cdot \times \mathcal Y)$.
In the \emph{pool-based active learning} protocol,\footnote{Although technically we study the pool-based active learning protocol,
all of our results apply equally well to the stream-based (selective sampling) model of active learning
(in which the algorithm must decide whether or not to request the label $Y_{i}$ before
observing any $X_{j}$ with $j > i$ or requesting any $Y_{j}$ with $j > i$).}
we define an \emph{active learning algorithm} $\mathcal A$ as an algorithm taking as input a budget $n \in \mathbb{N} \cup \{0\}$,
and proceeding as follows.
The algorithm initially has access to the unlabeled data sequence $X_{1},X_{2},\ldots$.
If $n > 0$, the algorithm may then select an index $i_{1} \in \mathbb{N}$ and request to observe the label $Y_{i_{1}}$.
The algorithm may then observe the value of $Y_{i_{1}}$, and if $n \geq 2$, then based on both the unlabeled sequence and this new observation $Y_{i_{1}}$,
it may select another index $i_{2} \in \mathbb{N}$ and request to observe $Y_{i_{2}}$.
This continues for a number of rounds at most $n$ (i.e., it may request at most $n$ labels),
after which the algorithm must halt and produce a classifier $\hat{h}_{n}$.
More formally, an active learning algorithm is defined by a random sequence $\{i_{t}\}_{t=1}^{\infty}$ in $\mathbb{N}$, a random variable $N$ in $\mathbb{N}$,
and a random classifier $\hat{h}_{n}$, satisfying the following properties.
Each $i_{t}$ is conditionally independent from $\{(X_{i},Y_{i})\}_{i=1}^{\infty}$ given $\{i_{j}\}_{j=1}^{t-1}$, $\{Y_{i_{j}}\}_{j=1}^{t-1}$, and $\{X_{i}\}_{i=1}^{\infty}$.
The random variable $N$ always has $N \leq n$, and for any $k \in \{0,\ldots,n\}$,
$\mathbbm{1}[ N = k ]$ is independent from $\{(X_{i},Y_{i})\}_{i=1}^{\infty}$ given $\{i_{j}\}_{j=1}^{k}$, $\{Y_{i_{j}}\}_{j=1}^{k}$, and $\{X_{i}\}_{i=1}^{\infty}$.
Finally, $\hat{h}_{n}$ is independent from $\{(X_{i},Y_{i})\}_{i=1}^{\infty}$ given $N$, $\{i_{j}\}_{j=1}^{N}$, $\{Y_{i_{j}}\}_{j=1}^{N}$, and $\{X_{i}\}_{i=1}^{\infty}$.
We are now ready for the definition of our primary quantity of study: the minimax label complexity.
In the next section, we define several well-known noise models as specifications of the set $\mathbb{D}$ referenced in this definition.
\begin{definition}
\label{def:LC}
For a given set $\mathbb{D}$ of probability measures on $\mathcal X \times \mathcal Y$,
$\forall \varepsilon \geq 0$, $\forall \delta \in [0,1]$,
the \emph{minimax label complexity} (of active learning) under $\mathbb{D}$ with respect to $\mathbb{C}$,
denoted $\Lambda_{\mathbb{D}}(\varepsilon,\delta)$,
is the smallest $n \in \mathbb{N} \cup \{0\}$ such that there exists an active learning algorithm $\mathcal A$
with the property that, for every $\mathcal{P}_{XY} \in \mathbb{D}$,
the classifier $\hat{h}_{n}$ produced by $\mathcal A(n)$ based on the (independent $\mathcal{P}_{XY}$-distributed)
data sequence $(X_{1},Y_{1}),(X_{2},Y_{2}),\ldots$ satisfies
\begin{equation*}
\P\left( {\rm er}_{\mathcal{P}_{XY}}\left( \hat{h}_{n} \right) - \inf_{h \in \mathbb{C}} {\rm er}_{\mathcal{P}_{XY}}(h) > \varepsilon \right) \leq \delta.
\end{equation*}
If no such $n$ exists, we define $\Lambda_{\mathbb{D}}(\varepsilon,\delta) = \infty$.
\end{definition}
Following \citet*{vapnik:71,anthony:99}, we say a collection of sets $\mathcal T \subseteq 2^{\mathcal X}$
\emph{shatters} a sequence $S \in \mathcal X^{k}$ (for $k \in \mathbb{N}$) if $\{ A \cap S : A \in \mathcal T \} = 2^{S}$.
The \emph{VC dimension} of $\mathcal T$ is then defined as the largest $k \in \mathbb{N} \cup \{0\}$ such that
there exists $S \in \mathcal X^{k}$ shattered by $\mathcal T$; if no such largest $k$ exists, the VC dimension is
defined to be $\infty$.
Overloading this terminology, the VC dimension of a set $\H$ of classifiers is defined as
the VC dimension of the collection of sets $\{ \{x : h(x) = +1\} : h \in \H \}$.
Throughout this article,
we denote by $d$ the VC dimension of $\mathbb{C}$.
We are particularly interested in the case $d < \infty$, in which case $\mathbb{C}$ is called a \emph{VC class}.
For any set $\H$ of classifiers,
define ${\rm DIS}(\H) = \{x \in \mathcal X : \exists h,g \in \H \text{ s.t. } h(x) \neq g(x)\}$, the \emph{region of disagreement} of $\H$.
Also, for any classifier $h$, any $r \geq 0$, and any probability measure $P$ on $\mathcal X$,
define ${\rm B}_{P}(h,r) = \{g \in \mathbb{C} : P(x : g(x) \neq h(x)) \leq r\}$, the \emph{$r$-ball centered at $h$}.
Before proceeding, we introduce a few additional notational conventions that help to simplify the theorem statements and proofs.
For any $\mathbb{R}$-valued functions $f$ and $g$, we write
$f(x) \lesssim g(x)$ (or equivalently $g(x) \gtrsim f(x)$) to express the fact
that there is a \emph{universal} finite numerical constant $c > 0$ such that $f(x) \leq c g(x)$.
For any $x \in [0,\infty]$, we define ${\rm Log}(x) = \max\{\ln(x),1\}$, where $\ln(0) = -\infty$ and $\ln(\infty) = \infty$.
For simplicity, we define $\frac{\infty}{{\rm Log}(\infty)} = \infty$,
but in any other context, we always define $0 \cdot \infty = 0$, and also define $\frac{a}{0} = \infty$ for any $a > 0$.
For any function $\phi : \mathbb{R} \to \mathbb{R}$, we use the notation
``$\lim_{\gamma \to 0} \phi(\gamma)$'' to indicating taking the limit as $\gamma$ approaches $0$
\emph{from above}: i.e., $\gamma \downarrow 0$.
For $a,b \in \mathbb{R}$, we denote $a \land b = \min\{a,b\}$ and $a \lor b = \max\{a,b\}$.
Finally, we remark that some of the claims below technically require additional qualifications
to guarantee measurability of certain quantities (as is typically the case in empirical process theory);
see \citet*{blumer:89,van-der-Vaart:96,van-der-Vaart:11} for some discussion of this issue.
For simplicity, we do not mention these issues in the analysis below;
rather, we implicitly qualify all of these results with the
condition that $\mathbb{C}$ is such that all of the random variables and events
arising in the proofs are measurable.
\section{Noise Models}
\label{sec:noise-models}
We now introduce the noise models under which we will study the minimax label complexity of active learning.
These are defined as sets of probability measures on $\mathcal X \times \mathcal Y$, corresponding to specifications of the
set $\mathbb{D}$ in Definition~\ref{def:LC}.
\begin{itemize}
\item[$\bullet$] (Realizable Case) Define ${\rm RE}$ as the collection of $\mathcal{P}_{XY}$ for which $f^{\star}_{\mathcal{P}_{XY}} \in \mathbb{C}$ and $2\eta(\cdot; \mathcal{P}_{XY})-1 = f^{\star}_{\mathcal{P}_{XY}}(\cdot)$ (almost everywhere w.r.t. $\mathcal{P}$).
\item[$\bullet$] (Bounded Noise) For $\beta \in [0,1/2)$, define ${\rm BN}(\beta)$ as the collection of joint distributions $\mathcal{P}_{XY}$ over $\mathcal X \times \mathcal Y$ such that
$f^{\star}_{\mathcal{P}_{XY}} \in \mathbb{C}$ and
\begin{equation*}
\mathcal{P}\left( x : \left| \eta(x;\mathcal{P}_{XY}) - 1/2 \right| \geq 1/2 - \beta \right) = 1.
\end{equation*}
\item[$\bullet$] (Tsybakov Noise)
For $a \in [1,\infty)$ and $\alpha \in (0,1)$, define
${\rm TN}(a, \alpha)$ as the collection of joint distributions $\mathcal{P}_{XY}$ over $\mathcal X \times \mathcal Y$ such that
$f^{\star}_{\mathcal{P}_{XY}} \in \mathbb{C}$ and
$\forall \gamma > 0$,
\begin{equation*}
\mathcal{P}\left( x : \left| \eta(x;\mathcal{P}_{XY}) - 1/2 \right| \leq \gamma \right) \leq a^{\prime} \gamma^{\alpha/(1-\alpha)},
\end{equation*}
where $a^{\prime} = (1-\alpha)(2\alpha)^{\alpha/(1-\alpha)}a^{1/(1-\alpha)}$.
\item[$\bullet$]
(Bernstein Class Condition)
For $a \in [1,\infty)$ and $\alpha \in [0,1]$, define
${\rm BC}(a, \alpha)$ as the collection of joint distributions $\mathcal{P}_{XY}$ over $\mathcal X \times \mathcal Y$ such that,
$\exists h_{\mathcal{P}_{XY}} \in \mathbb{C}$ for which $\forall h \in \mathbb{C}$,
\begin{equation*}
\mathcal{P}( x : h(x) \neq h_{\mathcal{P}_{XY}}(x) ) \leq a ( {\rm er}_{\mathcal{P}_{XY}}(h) - {\rm er}_{\mathcal{P}_{XY}}(h_{\mathcal{P}_{XY}}) )^{\alpha}.
\end{equation*}
\item[$\bullet$] (Benign Noise) For $\nu \in [0,1/2]$, define ${\rm BE}(\nu)$ as the collection of all joint distributions $\mathcal{P}_{XY}$ over $\mathcal X \times \mathcal Y$ such that $f^{\star}_{\mathcal{P}_{XY}} \in \mathbb{C}$ and ${\rm er}_{\mathcal{P}_{XY}}(f^{\star}_{\mathcal{P}_{XY}}) \leq \nu$.
\item[$\bullet$] (Agnostic Noise) For $\nu \in [0,1]$, define ${\rm AG}(\nu)$ as the collection of all joint distributions $\mathcal{P}_{XY}$ over $\mathcal X \times \mathcal Y$ such that $\inf_{h \in \mathbb{C}} {\rm er}_{\mathcal{P}_{XY}}(h) \leq \nu$.
\end{itemize}
It is known that ${\rm RE} \subseteq {\rm BN}(\beta) \subseteq {\rm BC}(1/(1-2\beta), 1)$,
and also
that ${\rm RE} \subseteq {\rm TN}(a, \alpha) \subseteq {\rm BC}(a, \alpha)$.
Furthermore, ${\rm TN}(a,\alpha)$ is equivalent to the conditions in ${\rm BC}(a,\alpha)$ being
satisfied for \emph{all} classifiers $h$, rather than merely those in $\mathbb{C}$ \citep*{mammen:99,tsybakov:04,boucheron:05}.
All of ${\rm RE}$, ${\rm BN}(\beta)$, and ${\rm TN}(a,\alpha)$ are contained in $\bigcup_{\nu < 1/2} {\rm BE}(\nu)$,
and in particular, ${\rm BN}(\beta) \subseteq {\rm BE}(\beta)$.
The realizable case is the simplest setting studied here, corresponding to the ``optimistic case'' of \citet*{vapnik:98}
or the PAC model of \citet*{valiant:84}.
The bounded noise model has been studied under various names \citep*[e.g.,][]{massart:06,gine:06,kaariainen:06,koltchinskii:10,raginsky:11};
it is sometimes referred to as \emph{Massart's noise condition}.
The Tsybakov noise condition was introduced by \citet*{mammen:99} in a slightly stronger form (in the related context of discrimination analysis)
and was distilled into the form stated above by \citet*{tsybakov:04}. There is now a substantial
literature on the label complexity under this condition, both for passive learning and active learning
\citep*[e.g.,][]{mammen:99,tsybakov:04,bartlett:06,koltchinskii:06,balcan:07,hanneke:11a,hanneke:12a,hanneke:survey,hanneke:12b}.
However, in much of this literature, the results are in fact established under the weaker assumption given
by the Bernstein class condition \citep*{bartlett:04}, which is known to be implied by the Tsybakov noise condition \citep*{mammen:99,tsybakov:04}.
For passive learning, it is known that the minimax sample complexities under Tsybakov noise and under the Bernstein class condition
are equivalent up to a logarithmic factor. Interestingly, our results below imply that this is not the case for active learning.
The benign noise condition \citep*[studied by][]{hanneke:thesis} requires only that the Bayes optimal classifier be contained
within the hypothesis class, and that the Bayes error rate be at most the value of the parameter $\nu$.
The agnostic noise condition (sometimes called \emph{adversarial noise} in related contexts)
is the weakest of the noise assumptions studied here, and admits any distribution for which
the best error rate among classifiers in the hypothesis class is at most the value of the parameter $\nu$.
This model has been widely studied in the literature, for both passive and active learning
\citep*[e.g.,][]{vapnik:71,vapnik:82,vapnik:98,kearns:94a,kalai:05,balcan:06,hanneke:07b,hanneke:07a,awasthi:14}.
\section{A Combinatorial Complexity Measure}
\label{sec:star}
There is presently a substantial literature on distribution-dependent bounds on the
label complexities of various active learning algorithms. These bounds are expressed in terms of a variety
of interesting complexity measures, designed to capture the behavior of each of these particular algorithms.
These measures of complexity include the disagreement coefficient \citep*{hanneke:07b}, the reciprocal of
the splitting index \citep*{dasgupta:05}, the extended teaching dimension growth function \citep*{hanneke:07a},
and the version space compression set size \citep*{el-yaniv:10,el-yaniv:12}. These quantities have been studied
and bounded for a variety of learning problems (see \citealp*{hanneke:survey}, for a summary).
They each have many interesting properties, and in general can exhibit a wide variety of behaviors,
as functions of the distribution over $\mathcal X$ (and in some cases, the distribution over $\mathcal X \times \mathcal Y$) and $\varepsilon$,
or in some cases, the data itself.
However, something remarkable happens when we maximize each of these complexity measures
over the choice of distribution (or data set): they all become equal to a simple and easy-to-calculate
combinatorial quantity (see Section~\ref{sec:complexities} for proofs of these equivalences).
Specifically, consider the following definition.\footnote{A similar notion previously appeared in a lower-bound
argument of \citet*{dasgupta:05}, including a kind of distribution-dependent version of the ``star set'' idea.
Indeed, we explore these connections formally in Section~\ref{sec:complexities},
where we additionally prove this definition is exactly equivalent to a quantity studied by \citet*{hanneke:07a}
(namely, the distribution-free version of the extended teaching dimension growth function), and has connections
to several other complexity measures in the literature.}
\begin{definition}
\label{def:star}
Define the \emph{star number} $\mathfrak{s}$ as the largest integer $s$
such that there exist $x_{1},\ldots,x_{s} \in \mathcal X$ and $h_{0},h_{1},\ldots,h_{s} \in \mathbb{C}$
with the property that $\forall i \in \{1,\ldots,s\}$, ${\rm DIS}(\{h_{0},h_{i}\}) \cap \{x_{1},\ldots,x_{s}\} = \{x_{i}\}$;
if no such largest integer exists, define $\mathfrak{s} = \infty$.
\end{definition}
For any set $\H$ of functions $\mathcal X \to \mathcal Y$,
any $t \in \mathbb{N}$, $x_{1},\ldots,x_{t} \in \mathcal X$, and $h_{0},h_{1},\ldots,h_{t} \in \H$,
we will say $\{x_{1},\ldots,x_{t}\}$ is a \emph{star set} for $\H$, \emph{witnessed by} $\{h_{0},h_{1},\ldots,h_{t}\}$,
if $\forall i \in \{1,\ldots,t\}$, ${\rm DIS}(\{h_{0},h_{i}\}) \cap \{x_{1},\ldots,x_{t}\} = \{x_{i}\}$.
For brevity, in some instances below, we may simply say that $\{x_{1},\ldots,x_{t}\}$ \emph{is a star set for} $\H$,
indicating that $\exists h_{0},h_{1},\ldots,h_{t} \in \H$ such that
$\{x_{1},\ldots,x_{t}\}$ is a star set for $\H$, witnessed by $\{h_{0},h_{1},\ldots,h_{t}\}$.
We may also say that $\{x_{1},\ldots,x_{t}\}$ \emph{is a star set for} $\H$ \emph{centered at} $h_{0} \in \H$
if $\exists h_{1},\ldots,h_{t} \in \H$ such that $\{x_{1},\ldots,x_{t}\}$ is a star set for $\H$, witnessed by $\{h_{0},h_{1},\ldots,h_{t}\}$.
For completeness, we also say that $\{\}$ (the empty sequence) is a star set for $\H$ (witnessed by $\{h_{0}\}$ for any $h_{0} \in \H$), for any nonempty $\H$.
In these terms, the star number of $\mathbb{C}$ is the maximum possible cardinality of a star set for $\mathbb{C}$,
or $\infty$ if no such maximum exists.
\paragraph{Remark:} The star number can equivalently be described as the maximum possible degree in the data-induced one-inclusion graph for $\mathbb{C}$ \citep*[see][]{haussler:94},
where the maximum is over all possible data sets and nodes in the graph.%
\footnote{The maximum degree in the one-inclusion graph was recently studied in the context of teaching complexity by \citet*{fan:12}.
However, using the data-induced one-inclusion graph of \citet*{haussler:94} (rather than the graph based on the full space $\mathcal X$) can substantially increase the maximum degree by omitting certain highly-informative points.}
To relate this to the VC dimension,
one can show
that the VC dimension is the maximum possible degree of a \emph{hypercube} in the data-induced one-inclusion graph
for $\mathbb{C}$ (maximized over all possible data sets).
From this, it is clear that $\mathfrak{s} \geq d$. Indeed,
any set $\{x_{1},\ldots,x_{k}\}$ shatterable by $\mathbb{C}$ is also a star set for $\mathbb{C}$,
since some $h_{0} \in \mathbb{C}$ classifies all $k$ points $-1$, and for each $x_{i}$,
some $h_{i} \in \mathbb{C}$ has $h_{i}(x_{i}) = +1$ while $h_{i}(x_{j}) = -1$ for every $j \neq i$
(where $h_{i}$ is guaranteed to exist by shatterability of the set).
On the other hand, there is no general upper bound on $\mathfrak{s}$ in terms of $d$,
and the gap between $\mathfrak{s}$ and $d$ can generally be infinite.
\paragraph{Examples:}
Before continuing, we briefly go through a few simple example calculations of the star number.
For the class of \emph{threshold} classifiers on $\mathbb{R}$ (i.e., $\mathbb{C} = \{ x \mapsto 2 \mathbbm{1}_{[t,\infty)}(x) - 1 : t \in \mathbb{R} \}$),
we have $\mathfrak{s} = 2$, as $\{x_{1},x_{2}\}$ is a star set for $\mathbb{C}$ centered at $2 \mathbbm{1}_{[t,\infty)}-1$ if and only if $x_{1} < t \leq x_{2}$,
and any set $\{x_{1},x_{2},x_{3}\}$ cannot be a star set for $\mathbb{C}$ centered at any given $2 \mathbbm{1}_{[t,\infty)}-1$ since, of the (at least) two of these points
on the same side of $t$, any threshold classifier disagreeing with $2 \mathbbm{1}_{[t,\infty)}-1$ on the one further from $t$ must also
disagree with $2 \mathbbm{1}_{[t,\infty)}-1$ on the one closer to $t$.
In contrast, for the class of \emph{interval} classifiers on $\mathbb{R}$ (i.e., $\mathbb{C} = \{ x \mapsto 2 \mathbbm{1}_{[a,b]}(x) - 1 : -\infty < a \leq b < \infty \}$),
we have $\mathfrak{s} = \infty$, since for \emph{any} distinct points $x_{0},x_{1},\ldots,x_{s} \in \mathbb{R}$,
$\{x_{1},\ldots,x_{s}\}$ is a star set for $\mathbb{C}$ witnessed by $\{ 2 \mathbbm{1}_{[x_{0},x_{0}]}-1, 2\mathbbm{1}_{[x_{1},x_{1}]}-1,\ldots,2\mathbbm{1}_{[x_{s},x_{s}]}-1 \}$.
It is an easy exercise to verify that we also have $\mathfrak{s}=\infty$
for the classes of \emph{linear separators} on $\mathbb{R}^{k}$ ($k \geq 2$) and axis-aligned rectangles on $\mathbb{R}^{k}$ ($k \geq 1$),
since the above construction for interval classifiers can be embedded into
these spaces, with the star set lying within a lower-dimensional manifold in $\mathbb{R}^{k}$ \citep*[see][]{dasgupta:04,dasgupta:05,hanneke:survey}.
As an intermediate case, where $\mathfrak{s}$ has a range of values,
consider the class of \emph{intervals of width at least $w \in (0,1)$} (i.e., $\mathbb{C} = \{ x \mapsto 2 \mathbbm{1}_{[a,b]}(x) - 1 : -\infty < a \leq b < \infty, b-a \geq w \}$),
for the space $\mathcal X = [0,1]$. In this case, we can show that $\lfloor 2/w \rfloor \leq \mathfrak{s} \leq \lfloor 2/w \rfloor + 2$, as follows.
We may note that letting $k = \lfloor 2/(w+\varepsilon) \rfloor + 1$ (for $\varepsilon > 0$),
and taking $x_{i} = (w+\varepsilon) (i-1) / 2$ for $1 \leq i \leq k$,
we have that $\{x_{1},\ldots,x_{k}\}$ is a star set for $\mathbb{C}$, witnessed by $\{ 2 \mathbbm{1}_{[-2w,-w]}-1, 2\mathbbm{1}_{[x_{1}-w/2,x_{1}+w/2]}-1,\ldots,2\mathbbm{1}_{[x_{k}-w/2,x_{k}+w/2]}-1 \}$.
Thus, taking $\varepsilon \to 0$ reveals that $\mathfrak{s} \geq \lfloor 2/w \rfloor$.
On the other hand, for any $k^{\prime} \in \mathbb{N}$ with $k^{\prime} > 2$, and points $x_{1},\ldots,x_{k^{\prime}} \in [0,1]$,
suppose $\{x_{1},\ldots,x_{k^{\prime}}\}$ is a star set for $\mathbb{C}$ witnessed by $\{h_{0},h_{1},\ldots,h_{k^{\prime}}\}$.
Without loss of generality, suppose $x_{1} \leq x_{2} \leq \cdots \leq x_{k^{\prime}}$.
First suppose $h_{0}$ classifies all of these points $-1$.
Note that, for any $i \in \{3,\ldots,k^{\prime}\}$, since the interval corresponding to $h_{i-1}$ has width at least $w$ and contains $x_{i-1}$ but not $x_{i-2}$ or $x_{i}$,
we have $x_{i}-x_{i-1} > \max\{ 0, w - (x_{i-1}-x_{i-2}) \}$.
Thus, $1 \geq \sum_{i=2}^{k^{\prime}} x_{i} - x_{i-1} > x_{2} - x_{1} + \sum_{i=3}^{k^{\prime}} \max\{ 0, w - (x_{i-1}-x_{i-2}) \} \geq (k^{\prime}-2) w - \sum_{i=3}^{k^{\prime}-1} x_{i} - x_{i-1} = (k^{\prime}-2) w - ( x_{k^{\prime}-1} - x_{2} )$,
so that $x_{k^{\prime}-1} - x_{2} > (k^{\prime}-2)w - 1$.
But $x_{k^{\prime}-1} - x_{2} \leq 1$, so that
$k^{\prime} < 2/w + 2$.
Since $k^{\prime}$ is an integer, this implies $k^{\prime} \leq \lfloor 2/w \rfloor + 2$.
For the remaining case, if $h_{0}$ classifies some $x_{i}$ as $+1$,
then let $x_{i_{0}} = \min\{ x_{i} : h_{0}(x_{i}) = +1 \}$ and $x_{i_{1}} = \max\{ x_{i} : h_{0}(x_{i}) = +1 \}$.
Note that, if $i_{0} > 1$, then for any $x < x_{i_{0}-1}$, any $h \in \mathbb{C}$ with $h(x_{i_{0}}) = h(x) = +1 \neq h_{0}(x)$ must have $h(x_{i_{0}-1}) = +1 \neq h_{0}(x_{i_{0}-1})$,
so that $\{x,x_{i_{0}-1}\} \subseteq {\rm DIS}(\{h,h_{0}\})$. Therefore, $\nexists x_{i} < x_{i_{0}-1}$
(since otherwise ${\rm DIS}(\{h_{i},h_{0}\}) \cap \{x_{1},\ldots,x_{k^{\prime}}\} = \{x_{i}\}$ would be violated),
so that $i_{0} \leq 2$. Symmetric reasoning implies $i_{1} \geq k^{\prime}-1$.
Similarly, if $\exists x \in (x_{i_{0}},x_{i_{1}})$, then any $h \in \mathbb{C}$ with $h(x) = -1 \neq h_{0}(x)$ must have either $h(x_{i_{0}}) = -1 \neq h_{0}(x_{i_{0}})$ or $h(x_{i_{1}}) = -1 \neq h_{0}(x_{i_{1}})$,
so that either $\{x,x_{i_{0}}\} \subseteq {\rm DIS}(\{h,h_{0}\})$ or $\{x,x_{i_{1}}\} \subseteq {\rm DIS}(\{h,h_{0}\})$. Therefore, $\nexists x_{i} \in (x_{i_{0}},x_{i_{1}})$
(since again, ${\rm DIS}(\{h_{i},h_{0}\}) \cap \{x_{1},\ldots,x_{k^{\prime}}\} = \{x_{i}\}$ would be violated),
so that $i_{1} \in \{i_{0},i_{0}+1\}$.
Combined, these facts imply $k^{\prime} \leq i_{1} + 1 \leq i_{0} + 2\leq 4 \leq \lfloor 2/w \rfloor + 2$.
Altogether, we have $\mathfrak{s} \leq \lfloor 2/w \rfloor + 2$.
\section{Main Results}
\label{sec:main}
We are now ready to state the main results of this article: upper and lower bounds on the minimax label complexities under the above noise models.
For the sake of making the theorem statements more concise, we abstract the dependence on logarithmic
factors in several of the upper bounds into a simple ``${\rm polylog}(x)$'' factor, meaning a value $\lesssim {\rm Log}^{k}(x)$, for some $k \in [1,\infty)$ (in fact, all of these results hold with values of $k \leq 4$);
the reader is referred to the proofs for a description of the actual logarithmic factors this ${\rm polylog}$ function represents,
along with tighter expressions of the upper bounds.
The formal proofs of all of these results are included in Appendix~\ref{app:main-proofs}.
\begin{theorem}
\label{thm:realizable}
For any $\varepsilon \in (0,1/9)$, $\delta \in (0,1/3)$,
\begin{equation*}
\max\left\{ \min\left\{\mathfrak{s}, \frac{1}{\varepsilon}\right\}, d, {\rm Log}\left(\min\left\{\frac{1}{\varepsilon}, |\mathbb{C}|\right\}\right)\right\}
\lesssim \Lambda_{{\rm RE}}(\varepsilon,\delta) \lesssim
\min\left\{ \mathfrak{s}, \frac{d}{\varepsilon}, \frac{\mathfrak{s} d}{{\rm Log}(\mathfrak{s})} \right\} {\rm Log}\left(\frac{1}{\varepsilon}\right).
\end{equation*}
\end{theorem}
\begin{theorem}
\label{thm:bounded}
For any $\beta \in [0,1/2)$, $\varepsilon \in (0,(1-2\beta)/24)$, $\delta \in (0,1/24]$,
\begin{multline*}
\frac{1}{(1-2\beta)^{2}} \max\left\{ \min\left\{\mathfrak{s}, \frac{1-2\beta}{\varepsilon}\right\} \beta {\rm Log}\left(\frac{1}{\delta}\right), d\right\}
\\ \lesssim \Lambda_{{\rm BN}(\beta)}(\varepsilon,\delta) \lesssim
\frac{1}{(1-2\beta)^{2}} \min\left\{ \mathfrak{s}, \frac{(1-2\beta) d}{\varepsilon} \right\} {\rm polylog}\left(\frac{d}{\varepsilon\delta}\right).
\end{multline*}
\end{theorem}
\begin{theorem}
\label{thm:tsybakov}
For any $a \in [4,\infty)$, $\alpha \in (0,1)$, $\varepsilon \in (0,1/(24 a^{1/\alpha}))$, and $\delta \in (0,1/24]$,
\\if $0 < \alpha \leq 1/2$,
\begin{equation*}
a^{2} \left(\frac{1}{\varepsilon}\right)^{2-2\alpha} \left(d + {\rm Log}\left(\frac{1}{\delta}\right)\right)
\lesssim \Lambda_{{\rm TN}(a,\alpha)}(\varepsilon,\delta)
\lesssim a^{2} \left(\frac{1}{\varepsilon}\right)^{2-2\alpha} d \cdot {\rm polylog}\left(\frac{d}{\varepsilon\delta}\right)
\end{equation*}
and if $1/2 < \alpha < 1$,
\begin{multline*}
a^{2} \left(\frac{1}{\varepsilon}\right)^{2-2\alpha} \max\left\{ \min\left\{ \mathfrak{s}, \frac{1}{a^{1/\alpha} \varepsilon}\right\}^{2\alpha-1} {\rm Log}\left(\frac{1}{\delta}\right), d\right\}
\\ \lesssim \Lambda_{{\rm TN}(a,\alpha)}(\varepsilon,\delta)
\lesssim a^{2} \left(\frac{1}{\varepsilon}\right)^{2-2\alpha} \min\left\{ \frac{\mathfrak{s}}{d}, \frac{1}{a^{1/\alpha}\varepsilon} \right\}^{2\alpha-1} d \cdot {\rm polylog}\left(\frac{d}{\varepsilon\delta}\right).
\end{multline*}
\end{theorem}
\begin{theorem}
\label{thm:diameter-localization}
For any $a \in [4,\infty)$, $\alpha \in (0,1)$, $\varepsilon \in (0,1/(24 a^{1/\alpha}))$, and $\delta \in (0,1/24]$,
\\if $0 \leq \alpha \leq 1/2$,
\begin{equation*}
a^{2} \left(\frac{1}{\varepsilon}\right)^{2-2\alpha} \left(d + {\rm Log}\left(\frac{1}{\delta}\right)\right)
\lesssim \Lambda_{{\rm BC}(a,\alpha)}(\varepsilon,\delta)
\lesssim a^{2} \left(\frac{1}{\varepsilon}\right)^{2-2\alpha} \!\!\min\left\{ \mathfrak{s}, \frac{1}{a \varepsilon^{\alpha}} \right\} d \cdot {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right),
\end{equation*}
and if $1/2 < \alpha \leq 1$,
\begin{multline*}
a^{2} \left(\frac{1}{\varepsilon}\right)^{2-2\alpha} \max\left\{\min\left\{ \mathfrak{s}, \frac{1}{a^{1/\alpha} \varepsilon}\right\}^{2\alpha-1} {\rm Log}\left(\frac{1}{\delta}\right), d\right\}
\\ \lesssim \Lambda_{{\rm BC}(a,\alpha)}(\varepsilon,\delta)
\lesssim a^{2} \left(\frac{1}{\varepsilon}\right)^{2-2\alpha} \min\left\{ \mathfrak{s}, \frac{1}{a \varepsilon^{\alpha}} \right\} d \cdot {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right).
\end{multline*}
\end{theorem}
\begin{theorem}
\label{thm:benign}
For any $\nu \in [0,1/2)$, $\varepsilon \in (0,(1-2\nu)/24)$, and $\delta \in (0,1/24]$,
\begin{equation*}
\frac{\nu^{2}}{\varepsilon^{2}} \left(d + {\rm Log}\left(\frac{1}{\delta}\right)\right) + \min\left\{\mathfrak{s},\frac{1}{\varepsilon}\right\}
\lesssim \Lambda_{{\rm BE}(\nu)}(\varepsilon,\delta)
\lesssim \left( \frac{\nu^{2}}{\varepsilon^{2}} d + \min\left\{\mathfrak{s},\frac{d}{\varepsilon}\right\}\right) {\rm polylog}\left(\frac{d}{\varepsilon\delta}\right).
\end{equation*}
\end{theorem}
\begin{theorem}
\label{thm:agnostic}
For any $\nu \in [0,1/2)$, $\varepsilon \in (0,(1-2\nu)/24)$, and $\delta \in (0,1/24]$,
\begin{multline*}
\frac{\nu^{2}}{\varepsilon^{2}} \left(d + {\rm Log}\left(\frac{1}{\delta}\right)\right) + \min\left\{\mathfrak{s},\frac{1}{\varepsilon}\right\}
\\ \lesssim \Lambda_{{\rm AG}(\nu)}(\varepsilon,\delta)
\lesssim \min\left\{ \mathfrak{s}, \frac{1}{\nu+\varepsilon} \right\} \left( \frac{\nu^{2}}{\varepsilon^{2}} + 1 \right) d \cdot {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right).
\end{multline*}
\end{theorem}
Here, we mention a few noteworthy
observations and comments regarding the above theorems.
We sketch the main innovations underlying the active learning algorithm achieving these upper bounds in Section~\ref{sec:algorithm}.
Sections \ref{sec:passive} and \ref{sec:complexities} include further detailed and thorough comparisons
of each of these results to those in the prior literature on passive and active learning.
\paragraph{Comparison to the previous best known results:}
Aside from Theorems \ref{thm:diameter-localization} and \ref{thm:agnostic}, each of the above
results offers some kind of refinement over the previous best known results on the label complexity of active learning.
Some of these refinements are relatively mild, such as those for the realizable case and bounded noise.
However, our refinements under Tsybakov noise and benign noise are far more significant.
In particular,
perhaps the most surprising and interesting of the above results are the upper bounds in Theorem~\ref{thm:tsybakov},
which can be considered the primary contribution of this work.
As discussed above, the prior literature on noise-robust active learning is largely rooted in the intuitions and techniques developed for the realizable case.
As indicated by Theorem~\ref{thm:realizable}, there is a wide spread of label complexities
for active learning problems in the realizable case, depending on the structure of the hypothesis class. In particular,
when $\mathfrak{s} < \infty$, we have $O({\rm Log}(1/\varepsilon))$ label complexity in the realizable case, representing a nearly-exponential
improvement over passive learning, which has $\tilde{\Theta}(1/\varepsilon)$ dependence on $\varepsilon$. On the other hand,
when $\mathfrak{s} = \infty$, we have $\Omega(1/\varepsilon)$ minimax label complexity for active learning, which is the same
dependence on $\varepsilon$ as known for passive learning (see Section~\ref{sec:passive}). Thus, for active learning in the realizable case,
some hypothesis classes are ``easy'' (such as threshold classifiers), offering strong improvements over passive learning,
while others are ``hard'' (such as interval classifiers), offering almost no improvements over passive.
With the realizable case as inspiration, the results in the prior literature on general noise-robust active learning have all continued to reflect
these distinctions, and the label complexity bounds in those works continue to exhibit this wide spread. In the case of Tsybakov
noise, the best general results in the prior literature \citep*[from][]{hanneke:12b,hanneke:survey} correspond to an upper bound of roughly
$a^{2} \left(\frac{1}{\varepsilon}\right)^{2-2\alpha} \min\left\{ \mathfrak{s}, \frac{1}{a \varepsilon^{\alpha}} \right\} d \cdot {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right)$
(after converting those complexity measures into the star number via the results in Section~\ref{sec:complexities} below).
When $\mathfrak{s} < \infty$, this has dependence $\tilde{\Theta}(\varepsilon^{2\alpha-2})$ on $\varepsilon$, which reflects a strong improvement
over the $\tilde{\Theta}(\varepsilon^{\alpha-2})$ minimax sample complexity of passive learning for this problem (see Section~\ref{sec:passive}).
On the other hand, when $\mathfrak{s} = \infty$, this bound is $\tilde{\Theta}(\varepsilon^{\alpha-2})$, so that as in the realizable case,
the bound is no better than that of passive learning for these hypothesis classes. Thus, the prior results in the literature
continue the trend observed in the realizable case, in which the ``easy'' hypothesis classes admit strong improvements
over passive learning, while the ``hard'' hypothesis classes have a bound that is no better than the sample complexity of passive learning.
With this as background, it comes as quite a surprise that the upper bounds in Theorem~\ref{thm:tsybakov} are
\emph{always} smaller than the corresponding minimax sample complexities of passive learning,
in terms of their asymptotic dependence on $\varepsilon$ for $0 < \alpha < 1$. Specifically, these upper bounds
reveal a label complexity $\tilde{O}( \varepsilon^{2\alpha-2} )$ when $\mathfrak{s} < \infty$, and $\tilde{O}( \varepsilon^{2\alpha-2} \lor (1/\varepsilon) )$ when $\mathfrak{s} = \infty$.
Comparing to the $\tilde{\Theta}(\varepsilon^{\alpha-2})$ minimax sample complexity of passive learning, the improvement for active learning
is by a factor of $\tilde{\Theta}( \varepsilon^{-\alpha} )$ when $\mathfrak{s} < \infty$, and by a factor of $\tilde{\Theta}( \varepsilon^{-\min\{\alpha,1-\alpha\}} )$ when $\mathfrak{s} = \infty$.
As a further surprise, when $0 < \alpha \leq 1/2$ (the high-noise regime),
we see that the distinctions between active learning problems of a given VC dimension essentially \emph{vanish} (up to logarithmic factors), so that the familiar
spread of label complexities from the realizable case is no longer present. Indeed, in this latter case, \emph{all} hypothesis classes
with finite VC dimension exhibit the strong improvements over passive learning, previously only known to hold for the ``easy''
hypothesis classes (such as threshold classifiers): that is, $\tilde{O}(\varepsilon^{2\alpha-2})$ label complexity.
Further examining these upper bounds, we see that the spread of label complexities between ``easy'' and ``hard'' hypothesis classes
increasingly re-emerges as $\alpha$ approaches $1$, beginning with $\alpha = 1/2$. This transition point is quite sensible, since
this is precisely the point at which the label complexity has dependence on $\varepsilon$ of $\tilde{\Theta}(1/\varepsilon)$,
which is roughly the same as the minimax label complexity of the ``hard'' hypothesis classes in the realizable case,
which is, after all, included in ${\rm TN}(a,\alpha)$. Thus, as $\alpha$
increases above $1/2$, the ``easy'' hypothesis classes (with $\mathfrak{s} < \infty$) exhibit stronger improvements over passive learning,
while the ``hard'' hypothesis classes (with $\mathfrak{s} = \infty$) continue to exhibit precisely this $\tilde{\Theta}\left(\frac{1}{\varepsilon}\right)$
behavior. In either case, the label complexity exhibits an improvement in dependence on $\varepsilon$ compared to passive learning for the same $\alpha$ value.
But since the label complexity of passive learning decreases to $\tilde{\Theta}\left(\frac{1}{\varepsilon}\right)$ as $\alpha \to 1$, we naturally have that
for the ``hard'' hypothesis classes, the gap between the passive and active label complexities shrinks as $\alpha$ approaches $1$.
In contrast, the ``easy'' hypothesis classes exhibit a gap between passive and active label complexities that
becomes more pronounced as $\alpha$ approaches $1$ (with a near-exponential improvement over passive learning exhibited in the limiting case, corresponding to bounded noise).
This same pattern is present, though to a lesser extent, in the benign noise case.
In this case, the best general results in the prior literature \citep*[from][]{dasgupta:07,hanneke:07a,hanneke:survey}
correspond to an upper bound of roughly $\min\left\{\mathfrak{s}, \frac{1}{\nu+\varepsilon}\right\} \left(\frac{\nu^{2}}{\varepsilon^{2}}+1\right) d \cdot {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right)$
(again, after converting those complexity measures into the star number via the results in Section~\ref{sec:complexities} below).
When $\mathfrak{s} < \infty$, the dependence on $\nu$ and $\varepsilon$ is roughly $\tilde{\Theta}\left(\frac{\nu^{2}}{\varepsilon^{2}}\right)$ (aside from logarithmic factors and constants, and for $\nu > \varepsilon$).
However, when $\mathfrak{s} = \infty$, this dependence becomes roughly $\tilde{\Theta}\left(\frac{\nu}{\varepsilon^{2}}\right)$,
which is the same as in the minimax sample complexity of passive learning (see Section~\ref{sec:passive}).
Thus, for these results in the prior literature, we again see that the ``easy'' hypothesis classes have a bound reflecting improvements over passive learning,
while the bound for the ``hard'' hypothesis classes fail to reflect any improvements over passive learning at all.
In contrast, consider the upper bound in Theorem~\ref{thm:benign}.
In this case, when $\nu \geq \sqrt{\varepsilon}$ (again, the high-noise regime), for \emph{all} hypothesis classes with finite VC dimension,
the dependence on $\nu$ and $\varepsilon$ is roughly $\tilde{\Theta}\left(\frac{\nu^{2}}{\varepsilon^{2}}\right)$ (aside from logarithmic factors and constants).
Again, this makes almost no distinction between ``easy'' hypothesis classes (with $\mathfrak{s} < \infty$) and ``hard'' hypothesis classes (with $\mathfrak{s}=\infty$),
and instead always exhibits the strongest possible improvements (up to logarithmic factors),
previously only known to hold for the ``easy'' classes (such as threshold classifiers):
namely, reduction in label complexity by roughly a factor of $1/\nu$ compared to passive learning.
The improvements in this case are typically milder than we found in Theorem~\ref{thm:tsybakov},
but noteworthy nonetheless. Again, as $\nu$ decreases below $\sqrt{\varepsilon}$, the distinction
between ``easy'' and ``hard'' hypothesis classes begins to re-emerge, with the harder classes maintaining a $\tilde{\Theta}\left(\frac{1}{\varepsilon}\right)$
dependence
(which is equivalent to the realizable-case label complexity for these classes, up to logarithmic factors),
while the easier classes continue to exhibit the $\tilde{\Theta}\left(\frac{\nu^{2}}{\varepsilon^{2}}\right)$ behavior,
approaching $O\left( {\rm polylog}\left(\frac{1}{\varepsilon}\right) \right)$ as $\nu$ shrinks.
\paragraph{The dependence on $\boldsymbol{\delta}$:}
One remarkable fact about $\Lambda_{{\rm RE}}(\varepsilon,\delta)$ is that there is \emph{no} significant dependence on
$\delta$ in the optimal label complexity for the given range of $\delta$.\footnote{We should expect
a more significant dependence on $\delta$ near $1$, since one case easily prove that $\Lambda_{{\rm RE}}(\varepsilon,\delta) \to 0$ as $\delta \to 1$.}
Note that this is not the case in noisy settings, where the lower bounds have an explicit dependence on $\delta$.
In the proofs, this dependence on $\delta$ is introduced via randomness of the labels.
However, as argued by \citet*{kaariainen:06},
a dependence on $\delta$ is sometimes still required in $\Lambda_{\mathbb{D}}(\varepsilon,\delta)$,
even if we restrict $\mathbb{D}$ to those $\mathcal{P}_{XY} \in {\rm AG}(\nu)$ inducing \emph{deterministic} labels:
that is, $\eta(x;\mathcal{P}_{XY}) \in \{0,1\}$ for all $x$.
\paragraph{Spanning the gaps:}
All of these results have gaps between the lower and upper bounds.
It is interesting to note that one can construct examples of hypothesis classes spanning these gaps,
for Theorems \ref{thm:realizable}, \ref{thm:bounded}, \ref{thm:tsybakov}, and \ref{thm:benign} (up to logarithmic factors).
For instance, for sufficiently large $d$ and $\mathfrak{s}$ and sufficiently small $\varepsilon$ and $\delta$, these upper bounds are tight (up to logarithmic factors) in the case where
$\mathbb{C} = \{ x \mapsto 2\mathbbm{1}_{S}(x)-1 : S \subseteq \{1,\ldots,\mathfrak{s}\}, |S|\leqd\}$, for $\mathcal X = \mathbb{N}$
(taking inspiration from a suggested modification by \citealp*{hanneke:survey}, of the proof of a related result of \citealp*{raginsky:11}).
Likewise, these lower bounds are tight (up to logarithmic factors) in the case that $\mathcal X = \mathbb{N}$ and
$\mathbb{C} = \{ x \mapsto 2\mathbbm{1}_{S}(x)-1 : S \in 2^{\{1,\ldots,d\}} \cup \{ \{i\} : d+1 \leq i \leq \mathfrak{s} \} \}$.\footnote{Technically,
for Theorems~\ref{thm:bounded} and \ref{thm:benign}, we require slightly stronger versions of the lower bound to establish
tightness for $\beta$ or $\nu$ near $0$: namely, adding the lower bound from Theorem~\ref{thm:realizable}
to these lower bounds. The validity of this stronger lower bound follows immediately from the facts that ${\rm RE} \subseteq {\rm BN}(\beta)$ and ${\rm RE} \subseteq {\rm BE}(\nu)$.}
Thus, these upper and lower bounds cannot be significantly refined (without loss of generality)
without introducing additional complexity measures to distinguish these cases.
For completeness, we include proofs of these claims in Appendix~\ref{app:gaps}.
It immediately follows from this (and monotonicity of the respective noise models in $\mathbb{C}$)
that the upper and lower bounds in Theorems \ref{thm:realizable}, \ref{thm:bounded}, \ref{thm:tsybakov}, and \ref{thm:benign}
are each sometimes tight in the case $\mathfrak{s} = \infty$, as limiting cases of the above constructions: that is, the upper bounds are tight (up to logarithmic factors) for
$\mathbb{C} = \{ x \mapsto 2\mathbbm{1}_{S}(x)-1 : S \subseteq \mathbb{N}, |S| \leq d \}$,
and the lower bounds are tight (up to logarithmic factors) for
$\mathbb{C} = \{ x \mapsto 2\mathbbm{1}_{S}(x)-1 : S \in 2^{\{1,\ldots,d\}} \cup \{ \{i\} : d+1 \leq i < \infty \} \}$.
It is interesting to note that the above space $\mathbb{C}$ for which the upper bounds are tight can be embedded in a variety of
hypothesis classes in common use in machine learning (while maintaining VC dimension $\lesssim d$ and star number $\lesssim \mathfrak{s}$):
for instance, in the case of $\mathfrak{s}=\infty$, this is true of linear separators in $\mathbb{R}^{3d}$ and axis-aligned rectangles in $\mathbb{R}^{2d}$.
It follows that the upper bounds in these theorems are tight (up to logarithmic factors) for each of these hypothesis classes.
\paragraph{Separation of $\boldsymbol{{\rm TN}(a,\alpha)}$ and $\boldsymbol{{\rm BC}(a,\alpha)}$:}
Another interesting implication of these results is a separation between the noise models ${\rm TN}(a,\alpha)$ and ${\rm BC}(a,\alpha)$
not previously noted in the literature. Specifically, if we consider any class $\mathbb{C}$ comprised of only the $\mathfrak{s}+1$ classifiers in Definition~\ref{def:star},
then one can show\footnote{Specifically, this follows by taking $\zeta = \frac{a}{2} (4\varepsilon)^{\alpha}$, $\beta = \frac{1}{2} - \frac{2}{a 4^{\alpha}} \varepsilon^{1-\alpha}$,
and $k = \min\left\{ \mathfrak{s}-1, \lfloor 1/\zeta \rfloor \right\}$ in Lemma~\ref{lem:rr11-star} of Appendix~\ref{sec:rr-lemma},
and noting that the resulting set of distributions ${\rm RR}(k,\zeta,\beta)$ is contained in ${\rm BC}(a,\alpha)$ for this $\mathbb{C}$.}
that (for $\mathfrak{s} \geq 3$), for any $\alpha \in (0,1]$, $a \in [4,\infty)$, $\varepsilon \in (0,1/(4a^{1/\alpha}))$, and $\delta \in (0,1/16]$,
\begin{equation*}
\Lambda_{{\rm BC}(a,\alpha)}(\varepsilon,\delta) \gtrsim a^{2} \left(\frac{1}{\varepsilon}\right)^{2-2\alpha}\min\left\{\mathfrak{s},\frac{1}{a\varepsilon^{\alpha}}\right\} {\rm Log}\left(\frac{1}{\delta}\right).
\end{equation*}
In particular, when $\mathfrak{s} > \frac{1}{a \varepsilon^{\alpha}}$, we have $\Lambda_{{\rm BC}(a,\alpha)}(\varepsilon,\delta) \gtrsim a \varepsilon^{\alpha-2} {\rm Log}(1/\delta)$,
which is larger than the upper bound on $\Lambda_{{\rm TN}(a,\alpha)}(\varepsilon,\delta)$.
Furthermore, when $\mathfrak{s} = \infty$, this lower bound has asymptotic dependence on $\varepsilon$ that is $\Omega( \varepsilon^{\alpha-2} )$,
which is the same dependence
found in the sample complexity of passive learning, up to a logarithmic factor (see Section~\ref{sec:passive} below).
Comparing this to the upper bounds in Theorem~\ref{thm:tsybakov},
which exhibit asymptotic dependence on $\varepsilon$ as $\Lambda_{{\rm TN}(a,\alpha)}(\varepsilon,\delta) = \tilde{O}( \varepsilon^{\min\{2\alpha-1,0\}-1} )$
when $\mathfrak{s} = \infty$, we see that for this class, any $\alpha \in (0,1)$ has
$\Lambda_{{\rm TN}(a,\alpha)}(\varepsilon,\delta) \ll \Lambda_{{\rm BC}(a,\alpha)}(\varepsilon,\delta)$.
One reason this separation is interesting is that most of the existing
literature on active learning under ${\rm TN}(a,\alpha)$ makes use
of the noise condition via the fact that it implies $\mathcal{P}(x : h(x) \neq f^{\star}_{\mathcal{P}_{XY}}(x)) \leq a ({\rm er}_{\mathcal{P}_{XY}}(h) - {\rm er}_{\mathcal{P}_{XY}}(f^{\star}_{\mathcal{P}_{XY}}))^{\alpha}$ for all $h \in \mathbb{C}$:
that is, ${\rm TN}(a,\alpha) \subseteq {\rm BC}(a,\alpha)$. This
separation indicates that, to achieve the optimal performance under ${\rm TN}(a,\alpha)$,
one needs to consider more-specific properties of this noise model, beyond
those satisfied by ${\rm BC}(a,\alpha)$.
Another reason this separation is quite interesting is that it
contrasts with the known results for \emph{passive} learning, where
(as we discuss in Section~\ref{sec:passive} below)
the sample complexities under these two noise models are \emph{equivalent}
(up to an unresolved logarithmic factor).
\paragraph{Gaps in Theorems~\ref{thm:diameter-localization} and \ref{thm:agnostic}, and related open problems:}
We conjecture that the dependence on $d$ and $\mathfrak{s}$ in the upper bounds of
Theorem~\ref{thm:diameter-localization} can be refined in general
(where presently it is linear in $\mathfrak{s} d$).
More specifically, we conjecture that the upper bound can be improved to
\begin{equation*}
\Lambda_{{\rm BC}(a,\alpha)}(\varepsilon,\delta) \lesssim a^{2} \left(\frac{1}{\varepsilon}\right)^{2-2\alpha} \min\left\{ \mathfrak{s}, \frac{d}{a \varepsilon^{\alpha}} \right\} {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right),
\end{equation*}
though it is unclear at this time as to how this might be achieved.
The above example (separating ${\rm BC}(a,\alpha)$ from ${\rm TN}(a,\alpha)$)
indicates that we generally cannot hope to reduce the upper bound on the label complexity for ${\rm BC}(a,\alpha)$
much beyond this.
As for whether the form of the upper bound on $\Lambda_{{\rm AG}(\nu)}(\varepsilon,\delta)$ in Theorem~\ref{thm:agnostic} can
generally be improved to match the form of the upper bound for $\Lambda_{{\rm BE}(\nu)}(\varepsilon,\delta)$, this remains a fascinating open question.
We conjecture that at least the dependence on $d$ and $\mathfrak{s}$ can be improved to some extent (where presently it is linear in $d \mathfrak{s}$).
\paragraph{Minutiae:}
We note that the restrictions to the ranges of $\varepsilon$ and $\delta$ in the above results are required only for the lower bounds (aside from $\delta \in (0,1]$, $\varepsilon > 0$),
as are the restrictions to the ranges of the parameters $a$, $\alpha$, and $\nu$, aside from the constraints in the definitions in Section~\ref{sec:noise-models};
the upper bounds are proven without any such restrictions in Appendix~\ref{app:main-proofs}.
Also, several of the upper bounds above (e.g., Theorems~\ref{thm:tsybakov} and \ref{thm:benign}) are slightly looser (by logarithmic factors)
than those actually proven in Appendix~\ref{app:main-proofs}, which are typically stated in a different form (e.g., with factors of
$d{\rm Log}\left(\frac{1}{\varepsilon}\right)+{\rm Log}\left(\frac{1}{\delta}\right)$, rather than simply $d \cdot {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right)$).
We state the weaker results here purely to simplify the theorem statements, referring the interested reader to the proofs for the refined versions.
However, aside from Theorem~\ref{thm:realizable}, we believe it is possible to further optimize the logarithmic factors in all of these upper bounds.
We additionally note that we can also obtain results by the subset relations between the noise models.
For instance, since ${\rm RE} \subseteq {\rm BN}(\beta) \subseteq {\rm BE}(\beta) \subseteq {\rm AG}(\beta)$,
in the case $\beta$ is close to $0$ we can increase the lower bounds in Theorems \ref{thm:bounded}, \ref{thm:benign}, and \ref{thm:agnostic}
based on the lower bound in Theorem~\ref{thm:realizable}: that is, for $\nu \geq \beta \geq 0$,
\begin{equation*}
\Lambda_{{\rm AG}(\nu)}(\varepsilon,\delta) \geq \Lambda_{{\rm BE}(\nu)}(\varepsilon,\delta) \geq \Lambda_{{\rm BN}(\beta)}(\varepsilon,\delta) \geq \Lambda_{{\rm RE}}(\varepsilon,\delta) \gtrsim \max\left\{ \min\left\{ \mathfrak{s}, \frac{1}{\varepsilon}\right\},d\right\}.
\end{equation*}
Similarly, since ${\rm RE}$ is contained in all of the noise models studied here,
${\rm Log}\left(\min\left\{\frac{1}{\varepsilon}, |\mathbb{C}|\right\}\right)$ can also be included as a lower bound in each of these results.
Likewise, in the cases that $a$ is very large or $\alpha$ is very close to $0$, we can get a more informative upper bound in Theorem~\ref{thm:tsybakov}
via Theorem~\ref{thm:benign}, since ${\rm TN}(a,\alpha) \subseteq {\rm BE}(1/2)$.
For simplicity, in many cases we have not explicitly included the various compositions of the above results that can be obtained in this way
(with only a few exceptions).
\subsection{The Strategy behind Theorems~\ref{thm:tsybakov} and \ref{thm:benign}}
\label{sec:algorithm}
The upper bounds in Theorems \ref{thm:tsybakov} and \ref{thm:benign} represent the main results of this work,
and along with the upper bound in Theorem~\ref{thm:bounded}, are based on a general argument with essentially
three main components. The first component is a more-sophisticated variant of a basic approach introduced
to the active learning literature by \citet*{kaariainen:06}: namely, reduction to the realizable case via repeatedly
querying for the label at a point in $\mathcal X$ until its Bayes optimal classification can be determined (based
on a sequential probability ratio test, as studied by \citealp*{wald:45,wald:47}).
Of course, in the present model of active learning,
repeatedly requesting a label $Y_{i}$ yields no new information beyond requesting $Y_{i}$ once, since we are
not able to resample from the distribution of $Y_{i}$ given $X_{i}$ (as \citealp*{kaariainen:06}, does). To resolve this,
we argue that it is possible to partition the space $\mathcal X$ into cells, in a way such that $f^{\star}_{\mathcal{P}_{XY}}$ is nearly constant
in the vast majority of cells (without direct knowledge of $f^{\star}_{\mathcal{P}_{XY}}$ or $\mathcal{P}$); this is essentially a data-dependent
approximation to the recently-discovered finite approximability property of VC classes \citep*{adams:12}.
Given this partition, for a given point $X_{i}$, we can find many other points $X_{j}$ in the same cell of the partition with $X_{i}$,
and request labels for these points until we can determine what the majority label for the cell is. We show that,
with high probability, this value will equal $f^{\star}_{\mathcal{P}_{XY}}(X_{i})$, so that we can effectively use these majority
labels in an active learning algorithm for the realizable case.
However, we note that in the case of ${\rm TN}(a,\alpha)$, if we simply apply
this repeated querying strategy to random $\mathcal{P}$-distributed samples, the resulting label complexity would be too large,
and we would sometimes expect to exhaust most of the queries determining the optimal labels in very \emph{noisy} regions
(i.e., in cells of the partition where $\eta(\cdot;\mathcal{P}_{XY})$ is close to $1/2$ on average). This is because Tsybakov's condition
allows that such regions can have non-negligible probability, and the number of samples required to determine the
majority value of a $\pm 1$ random variable becomes unbounded as its mean approaches zero. However, we can note
that it is also less important for the final classifier $\hat{h}$ to agree with $f^{\star}_{\mathcal{P}_{XY}}$ on these high-noise points than it is for low-noise
points, since classifying them opposite from $f^{\star}_{\mathcal{P}_{XY}}$ has less impact on the excess error rate ${\rm er}_{\mathcal{P}_{XY}}(\hat{h}) - {\rm er}_{\mathcal{P}_{XY}}(f^{\star}_{\mathcal{P}_{XY}})$.
Therefore, as the second main component of our active learning strategy, we take
a tiered approach to learning, effectively shifting the distribution $\mathcal{P}$ to favor points in cells with average $\eta(\cdot;\mathcal{P}_{XY})$
value further from $1/2$.
We achieve this by discarding a point $X_{i}$ if the number of queries exhausted toward determining the majority label in
its cell of the partition becomes excessively large, and we gradually decrease this threshold as the data set grows,
so that the points making it through this filter have progressively less and less noisy labels.
By choosing $\hat{h}$ to agree with the inferred $f^{\star}_{\mathcal{P}_{XY}}$ classification of every point passing this filter,
and combining this with the standard analysis of learning in the realizable case \citep*{vapnik:82,vapnik:98,blumer:89},
this allows us to provide a bound
on the fraction of points in $\mathcal X$ at a given level of noisiness (i.e., $|\eta(\cdot;\mathcal{P}_{XY})-1/2|$) on which the produced classifier $\hat{h}$ disagrees with $f^{\star}_{\mathcal{P}_{XY}}$,
such that this bound decreases as the noisiness decreases (i.e., as $|\eta(\cdot;\mathcal{P}_{XY})-1/2|$ increases). Furthermore, by discarding many of the
points in high-noise regions without exhausting too many label requests trying to determine their $f^{\star}_{\mathcal{P}_{XY}}$ classifications, we are able to reduce the total
number of label requests needed to obtain $\varepsilon$ excess error rate.
Already these two components comprise the essential strategy that achieves these upper bounds in the case
of $\mathfrak{s}=\infty$.
However, to obtain the stated dependence on $\mathfrak{s}$ in these bounds when $\mathfrak{s} < \infty$, we need to
introduce a third component: namely, using the inferred values of $f^{\star}_{\mathcal{P}_{XY}}(X_{i})$ in the context
of an active learning algorithm for the realizable case. For this, we specifically use the disagreement-based
strategy of \citet*{cohn:94} (known as CAL), which processes the unlabeled data in sequence, and requests to observe
the classification $f^{\star}_{\mathcal{P}_{XY}}(X_{i})$ if and only if $X_{i}$ is in the region of disagreement of the set of
classifiers in $\mathbb{C}$ consistent with all previously-observed $f^{\star}_{\mathcal{P}_{XY}}(X_{j})$ values.
Using a modification of a recent analysis of this algorithm by \citet*{hanneke:14a}
(applied to each tier of label-noise separately),
combined with the results below (in Section~\ref{sec:xtd}) relating the complexity measure
used in that analysis to the star number, we obtain the dependence on $\mathfrak{s}$ stated in the above results.
\section{Comparison to Passive Learning}
\label{sec:passive}
The natural baseline for comparison in active learning is the \emph{passive learning} protocol,
in which the labeled data are i.i.d. samples with common distribution $\mathcal{P}_{XY}$:
that is, the input to the passive learning algorithm is $(X_1,Y_1),\ldots,(X_n,Y_n)$.
In this context, the minimax sample complexity of passive learning, denoted $\mathcal{M}_{\mathbb{D}}(\varepsilon,\delta)$,
is defined as the smallest $n \in \mathbb{N} \cup \{0\}$ for which there exists a passive learning rule
mapping $(X_1,Y_1),\ldots,(X_n,Y_n)$ to a classifier $\hat{h} : \mathcal X \to \mathcal Y$ such that,
for any $\mathcal{P}_{XY} \in \mathbb{D}$, with probability at least $1-\delta$, ${\rm er}_{\mathcal{P}_{XY}}(\hat{h}) - \inf_{h \in \mathbb{C}} {\rm er}_{\mathcal{P}_{XY}}(h) \leq \varepsilon$.
Clearly $\Lambda_{\mathbb{D}}(\varepsilon,\delta) \leq \mathcal{M}_{\mathbb{D}}(\varepsilon,\delta)$ for any $\mathbb{D}$, since for every passive learning algorithm $\mathcal A$,
there is an active learning algorithm that requests $Y_{1},\ldots,Y_{n}$ and then runs $\mathcal A$ with $(X_{1},Y_{1}),\ldots,(X_{n},Y_{n})$
to determine the returned classifier.
One of the main interests in the theory of
active learning is determining the size of the gap between these two complexities, for various sets $\mathbb{D}$.
For the purpose of this comparison, we now review several results known to hold for $\mathcal{M}_{\mathbb{D}}(\varepsilon,\delta)$, for
various sets $\mathbb{D}$.
Specifically, the following bounds are known to hold for any choice of hypothesis class $\mathbb{C}$,
and for $\beta$, $a$, $\alpha$, $\nu$, $\varepsilon$, and $\delta$ as in the respective theorems from Section~\ref{sec:main}
\citep*{vapnik:71,vapnik:82,vapnik:98,blumer:89,ehrenfeucht:89,haussler:94,massart:06,hanneke:survey}.
\begin{itemize}
\item $\frac{1}{\varepsilon}\left( d + {\rm Log}\left(\frac{1}{\delta}\right) \right) \lesssim \mathcal{M}_{{\rm RE}}(\varepsilon,\delta) \lesssim \frac{1}{\varepsilon}\left( d {\rm Log}\left( \frac{1}{\max\{\varepsilon,\delta\}} \right) + {\rm Log}\left(\frac{1}{\delta}\right) \right)$.
\item $\frac{1}{(1-2\beta)\varepsilon}\left( d + {\rm Log}\left(\frac{1}{\delta}\right) \right) \lesssim \mathcal{M}_{{\rm BN}(\beta)}(\varepsilon,\delta) \lesssim \frac{1}{(1-2\beta)\varepsilon}\left( d {\rm Log}\left( \frac{1-2\beta}{\varepsilon} \right) + {\rm Log}\left(\frac{1}{\delta}\right) \right)$.
\item $\frac{a}{\varepsilon^{2-\alpha}} \left( d + {\rm Log}\left(\frac{1}{\delta}\right) \right) \lesssim \mathcal{M}_{{\rm TN}(a,\alpha)}(\varepsilon,\delta) \leq \mathcal{M}_{{\rm BC}(a,\alpha)} \lesssim \frac{a}{\varepsilon^{2-\alpha}} \left( d {\rm Log}\left( \frac{1}{a \varepsilon^{\alpha}} \right) + {\rm Log}\left(\frac{1}{\delta}\right) \right)$.
\item $\frac{\nu+\varepsilon}{\varepsilon^{2}}\left( d + {\rm Log}\left(\frac{1}{\delta}\right) \right) \lesssim \mathcal{M}_{{\rm BE}(\nu)}(\varepsilon,\delta) \leq \mathcal{M}_{{\rm AG}(\nu)}(\varepsilon,\delta) \lesssim \frac{\nu+\varepsilon}{\varepsilon^{2}} \left( d {\rm Log}\left(\frac{1}{\nu+\varepsilon}\right) + {\rm Log}\left(\frac{1}{\delta}\right) \right)$.
\end{itemize}
Let us compare these to the results for active learning in Section~\ref{sec:main} on a case-by-case basis.
In the realizable case, we observe clear improvements of active learning over passive learning in the case $\mathfrak{s} \ll \frac{d}{\varepsilon}$ (aside from logarithmic factors).
In particular, based on the upper and lower bounds for both passive and active learning, we may conclude that
$\mathfrak{s} < \infty$ is necessary and sufficient for the asymptotic dependence on $\varepsilon$ to satisfy
$\Lambda_{{\rm RE}}(\varepsilon,\cdot) = o(\mathcal{M}_{{\rm RE}}(\varepsilon,\cdot))$; specifically, when $\mathfrak{s} < \infty$, $\Lambda_{{\rm RE}}(\varepsilon,\cdot) = O({\rm Log}(\mathcal{M}_{{\rm RE}}(\varepsilon,\cdot)))$,
and when $\mathfrak{s} = \infty$, $\Lambda_{{\rm RE}}(\varepsilon,\cdot) = \Theta(\mathcal{M}_{{\rm RE}}(\varepsilon,\cdot))$.
For bounded noise, we have a similar asymptotic behavior. When $\mathfrak{s} < \infty$, again $\Lambda_{{\rm BN}(\beta)}(\varepsilon,\cdot) = O({\rm polylog}(\mathcal{M}_{{\rm BN}(\beta)}(\varepsilon,\cdot)))$,
and when $\mathfrak{s} = \infty$, $\Lambda_{{\rm BN}(\beta)}(\varepsilon,\cdot) = \tilde{\Theta}(\mathcal{M}_{{\rm BN}(\beta)}(\varepsilon,\cdot))$.
In terms of the constants, to obtain improvements over passive learning (aside from the effects of logarithmic factors),
it suffices to have $\mathfrak{s} \ll \frac{(1-2\beta) d}{\varepsilon}$, which is somewhat smaller (depending on $\beta$)
than was sufficient in the realizable case.
Under Tsybakov's noise condition, every $\alpha \in (0,1/2]$ shows an improvement in the upper bounds for active learning
over the lower bound for passive learning by a factor of roughly $\frac{1}{a \varepsilon^{\alpha}}$ (aside from logarithmic factors).
On the other hand, when $\alpha \in (1/2,1)$,
if $\mathfrak{s} < \frac{d}{a^{1/\alpha} \varepsilon}$, the improvement of active upper bounds over the passive lower bound is by a factor of roughly
$\frac{1}{a \varepsilon^{\alpha}} \left(\frac{d}{\mathfrak{s}}\right)^{2\alpha-1}$, while for $\mathfrak{s} \geq \frac{d}{a^{1/\alpha} \varepsilon}$,
the improvement is by a factor of roughly $\frac{1}{a^{\frac{1-\alpha}{\alpha}} \varepsilon^{1-\alpha}}$ (again, ignoring logarithmic factors in both cases).
In particular, for \emph{any} $\alpha \in (0,1)$, when $\mathfrak{s} < \infty$, the asymptotic dependence
on $\varepsilon$ satisfies $\Lambda_{{\rm TN}(a,\alpha)}(\varepsilon,\cdot) = \tilde{\Theta}\left( \varepsilon^{\alpha} \mathcal{M}_{{\rm TN}(a,\alpha)}(\varepsilon,\cdot) \right)$,
and when $\mathfrak{s} = \infty$, the asymptotic dependence on $\varepsilon$ satisfies
$\Lambda_{{\rm TN}(a,\alpha)}(\varepsilon,\cdot) = \tilde{\Theta}\left( \varepsilon^{\min\{\alpha,1-\alpha\}} \mathcal{M}_{{\rm TN}(a,\alpha)}(\varepsilon,\cdot) \right)$.
In either case, we have that for any $\alpha \in (0,1)$,
$\Lambda_{{\rm TN}(a,\alpha)}(\varepsilon,\cdot) = o( \mathcal{M}_{{\rm TN}(a,\alpha)}(\varepsilon,\cdot) )$.
For the Bernstein class condition, the gaps in the upper and lower bounds of Theorem~\ref{thm:diameter-localization}
render unclear the necessary and sufficient conditions for $\Lambda_{{\rm BC}(a,\alpha)}(\varepsilon,\cdot) = o(\mathcal{M}_{{\rm BC}(a,\alpha)}(\varepsilon,\cdot))$.
Certainly $\mathfrak{s} < \infty$ is a sufficient condition for this, in which case the improvements are by a factor of roughly $\frac{1}{a \varepsilon^{\alpha}}$.
However, in the case of $\mathfrak{s} = \infty$, the upper bounds do not reveal any improvements over those given above for $\mathcal{M}_{{\rm BC}(a,\alpha)}(\varepsilon,\delta)$.
Indeed, the example given above in Section~\ref{sec:main} reveals that, in some nontrivial cases, $\Lambda_{{\rm BC}(a,\alpha)}(\varepsilon,\delta) \gtrsim \mathcal{M}_{{\rm BC}(a,\alpha)}(\varepsilon,\delta) / {\rm Log}(1/\varepsilon)$,
in which case any improvements would be, at best, in the constant and logarithmic factors. Note that this example also presents an interesting contrast between active and passive learning,
since it indicates that in some cases $\Lambda_{{\rm BC}(a,\alpha)}(\varepsilon,\delta)$ and $\Lambda_{{\rm TN}(a,\alpha)}(\varepsilon,\delta)$ are quite different,
while the above bounds for passive learning reveal that $\mathcal{M}_{{\rm BC}(a,\alpha)}(\varepsilon,\delta)$ is equivalent to $\mathcal{M}_{{\rm TN}(a,\alpha)}(\varepsilon,\delta)$ up to constant and logarithmic factors.
In the case of benign noise, comparing the above bounds for passive learning to Theorem~\ref{thm:benign}, we see
that (aside from logarithmic factors) the upper bound for active learning improves over the lower bound for passive learning
by a factor of roughly $\frac{1}{\nu}$ when $\nu \geq \sqrt{\varepsilon}$. When $\nu < \sqrt{\varepsilon}$, if $\mathfrak{s} > \frac{d}{\varepsilon}$,
the improvements are by a factor of roughly $\frac{\nu+\varepsilon}{\varepsilon}$, and if $\mathfrak{s} \leq \frac{d}{\varepsilon}$, the improvements
are by roughly a factor of $\min\left\{\frac{1}{\nu}, \frac{(\nu+\varepsilon) d}{\varepsilon^{2} \mathfrak{s}}\right\}$ (again, ignoring logarithmic factors).
However, as has been known for this noise model for some time \citep*{kaariainen:06}, there are no gains in terms of
the asymptotic dependence on $\varepsilon$ for fixed $\nu$. However, if we consider $\nu_{\varepsilon}$ such that $\varepsilon \leq \nu_{\varepsilon} = o(1)$,
then for $\mathfrak{s} < \infty$ we have $\Lambda_{{\rm BE}(\nu_{\varepsilon})}(\varepsilon,\cdot) = \tilde{\Theta}( \nu_{\varepsilon} \mathcal{M}_{{\rm BE}(\nu_{\varepsilon})}(\varepsilon,\cdot) )$,
and for $\mathfrak{s} = \infty$ we have $\Lambda_{{\rm BE}(\nu_{\varepsilon})}(\varepsilon,\cdot) = \tilde{O}\left( \max\left\{\nu_{\varepsilon}, \frac{\varepsilon}{\nu_{\varepsilon}} \right\} \mathcal{M}_{{\rm BE}(\nu_{\varepsilon})}(\varepsilon,\cdot) \right)$.
Finally, in the case of agnostic noise, similarly to the Bernstein class condition, the gaps between the upper and lower
bounds in Theorem~\ref{thm:agnostic} render unclear precisely what types of improvements we can expect when $\mathfrak{s} > \frac{1}{\nu+\varepsilon}$,
ranging from the lower bound, which has the behavior described above for $\Lambda_{{\rm BE}(\nu)}$, to the upper bound, which reflects
no improvements over passive learning in this case. When $\mathfrak{s} < \frac{1}{\nu+\varepsilon}$, the upper bound for active learning
reflects an improvement over the lower bound for passive learning by roughly a factor of $\frac{1}{(\nu+\varepsilon)\mathfrak{s}}$ (aside from logarithmic factors).
It remains an interesting open problem to determine whether the stronger improvements observed for benign noise
generally also hold for agnostic noise.
\paragraph{A remark on logarithmic factors:}
It is known that the terms of the form ``$d {\rm Log}( x )$'' in each of the above upper bounds for passive learning can be refined
to replace $x$ with the maximum of the disagreement coefficient (see Section~\ref{sec:dc} below) over the distributions in $\mathbb{D}$
\citep*{gine:06,hanneke:12b,hanneke:survey}.
Therefore, based on the results in Section~\ref{sec:dc} relating the disagreement coefficient to the star number, we can
replace these ``$d {\rm Log}(x)$'' terms with ``$d {\rm Log}(\mathfrak{s} \land x)$''.
In the case of ${\rm BN}(\beta)$, \citet*{massart:06} and \citet*{raginsky:11} have argued that, at least in some cases,
this logarithmic factor can also be included in the lower bounds. It is presently not known whether this is the case
for the other noise models studied here.
\section{Connections to the Prior Literature on Active Learning}
\label{sec:complexities}
\begin{table}[t]
\centering
\begin{tabular}{|l|c|c|}
\hline
technique & source & relation to $\mathfrak{s}$ \\ \hline\hline
disagreement coefficient & \citep*{hanneke:07b} & $\sup\limits_{P} \theta_{P}(\varepsilon) = \mathfrak{s} \land \frac{1}{\varepsilon}$\\ \hline
splitting index & \citep*{dasgupta:05} & $\sup\limits_{h,P} \lim\limits_{\tau\to0} \left\lfloor \frac{1}{\rho_{h,P}(\varepsilon;\tau)} \right\rfloor = \mathfrak{s} \land \left\lfloor \frac{1}{\varepsilon} \right\rfloor$ \\ \hline
teaching dimension & \citep*{hanneke:07a} & ${\rm XTD}(\mathbb{C},m) = \mathfrak{s} \land m$\\ \hline
version space compression & \citep*{el-yaniv:10} & $\max\limits_{h \in \mathbb{C}} \max\limits_{\mathcal U \in \mathcal X^{m}} \hat{n}_{h}(\mathcal U) = \mathfrak{s} \land m$\\ \hline
doubling dimension & \citep*{long:07} & $\sup\limits_{h,P} D_{h,P}(\varepsilon) \!\in\! [1,O(d)] \log\!\left(\mathfrak{s} \land \frac{1}{\varepsilon}\right)$\\ \hline
\end{tabular}
\caption{Many of the complexity measures from the literature are related to the star number.}
\label{tab:complexities}
\end{table}
As mentioned, there is already a substantial literature bounding the label complexities of various active
learning algorithms under various noise models.
It is natural to ask how the results in the prior literature compare to those stated above.
However, as most of the prior results are $\mathcal{P}_{XY}$-dependent, the appropriate comparison
is to the worst-case values of those results: that is, maximizing the bounds over $\mathcal{P}_{XY}$ in
the respective noise model.
This section makes this comparison. In particular, we will see that the label complexity upper bounds above
for ${\rm RE}$, ${\rm BN}(\beta)$, ${\rm TN}(a,\alpha)$, and ${\rm BE}(\nu)$ all show some improvements over the
known results, with the last two of these showing the strongest improvements.
The general results in the prior literature each express
their label complexity bounds in terms of some kind of complexity measure. There are now
several such complexity measures in use, each appropriate for studying some family of active learning
algorithms under certain noise models. Most of these quantities are dependent on the distribution $\mathcal{P}_{XY}$
or the data, and their definitions are quite diverse. For some pairs of them, there are known inequalities
loosely relating them, while other pairs have defied attempts to formally relate the quantities.
The dependence on $\mathcal{P}_{XY}$ in the general results in the prior literature is
typically isolated to the various complexity measures they are expressed in terms of. Thus, the natural first
step is to characterize the worst-case values of these complexity measures, for any given
hypothesis class $\mathbb{C}$. Plugging these worst-case values into the original bounds then allows
us to compare to the results stated above.
In the process of studying the worst-case behaviors of these complexity measures,
we also identify a \emph{very} interesting fact that has heretofore gone unnoticed: namely, that
almost all of the complexity measures in the relevant prior literature on the label complexity of active learning
are in fact \emph{equal} to the star number when maximized over the choice of distribution or data set.
In some sense, this fact is quite surprising, as this seemingly-eclectic collection of complexity measures includes disparate
definitions and interpretations, corresponding to entirely distinct approaches to the analysis of the respective
algorithms these quantities are used to bound the label complexities of.
Thus, this equivalence is interesting in its own right; additionally, it plays an important role in our proofs
of the main results above, since it allows us to build on these diverse techniques from the prior literature
when establishing these results.
Each subsection below is devoted to a particular complexity measure from the prior literature on active learning,
each representing an established technique for obtaining label complexity bounds. Together, they represent a
summary of the best-known general results from the prior literature relevant to our present discussion.
In each case, we show the equivalence of the worst-case value of the complexity
measure to the star number, and then combine this fact with the known results to obtain the corresponding
bounds on the minimax label complexities implicit in the prior literature. In each case, we then compare this result to those
obtained above.
We additionally study the \emph{doubling dimension}, a quantity which has been used to bound the sample
complexity of passive learning, and can be used to provide a loose bound on the label complexity of certain
active learning algorithms. Below we argue that, when maximized over the choice of distribution, the doubling
dimension can be upper and lower bounded in terms of the star number. One immediate implication of these
bounds is that the doubling dimension is bounded if and only if the star number is finite.
Our findings on the relations of these various complexity measures to the star number are summarized in Table~\ref{tab:complexities}.
\subsection{The Disagreement Coefficient}
\label{sec:dc}
We begin with, what is perhaps the most well-studied complexity measure in the active learning literature: the \emph{disagreement coefficient} \citep*{hanneke:07b,hanneke:thesis}.
\begin{definition}
\label{def:dc}
For any $r_{0} \geq 0$, any classifier $h$, and any probability measure $\mathcal{P}$ over $\mathcal X$,
the disagreement coefficient of $h$ with respect to $\mathbb{C}$ under $\mathcal{P}$ is defined as
\begin{equation*}
\theta_{h,\mathcal{P}}(r_{0}) = \sup_{r > r_{0}} \frac{\mathcal{P}\left( {\rm DIS}\left( {\rm B}_{\mathcal{P}}\left(h, r \right) \right) \right)}{r} \lor 1.
\end{equation*}
Also, for any probability measure $\mathcal{P}_{XY}$ over $\mathcal X \times \mathcal Y$, letting $\mathcal{P}$ denote the marginal distribution of $\mathcal{P}_{XY}$ over $\mathcal X$,
and letting $h^{*}_{\mathcal{P}_{XY}}$ denote a classifier with ${\rm er}_{\mathcal{P}_{XY}}(h^{*}_{\mathcal{P}_{XY}}) = \inf_{h \in \mathbb{C}} {\rm er}_{\mathcal{P}_{XY}}(h)$ and $\inf_{h \in \mathbb{C}} \mathcal{P}(x : h(x) \neq h^{*}_{\mathcal{P}_{XY}}(x)) = 0$,\footnote{See \citet*{hanneke:12a} for a proof that such a classifier always exists (though not necessarily in $\mathbb{C}$).}
define the disagreement coefficient of the class $\mathbb{C}$ with respect to $\mathcal{P}_{XY}$ as $\theta_{\mathcal{P}_{XY}}(r_{0}) = \theta_{h^{*}_{\mathcal{P}_{XY}},\mathcal{P}}(r_{0})$.
\end{definition}
The disagreement coefficient is used to bound the label complexities of a family of active learning algorithms, described as \emph{disagreement-based}.
This line of work was initiated by \citet*{cohn:94}, who propose an algorithm effective in the realizable case.
That method was extended to be robust to label noise by \citet*{balcan:06,balcan:09},
which then inspired a slew of papers studying variants of this idea; the interested reader is referred to
\citet*{hanneke:survey} for a thorough survey of this literature.
The general-case label complexity analysis of disagreement-based active learning (in terms of the disagreement coefficient) was initiated in the work of \citet*{hanneke:07b,hanneke:thesis},
and followed up by many papers since then \citep*[e.g.,][]{dasgupta:07,hanneke:09a,hanneke:11a,hanneke:12a,koltchinskii:10,hanneke:12b}, as well as many works
characterizing the value of the disagreement coefficient under various conditions \citep*[e.g.,][]{hanneke:07b,friedman:09,hanneke:10a,wang:11,balcan:13,hanneke:survey};
again, see \citet*{hanneke:survey} for a thorough survey of the known results on the disagreement coefficient.
To study the worst-case values of the label complexity bounds expressed in terms of the disagreement coefficient, let us define
\begin{equation*}
\hhat{\dc}(\varepsilon) = \sup_{\mathcal{P}_{XY}} \theta_{\mathcal{P}_{XY}}(\varepsilon).
\end{equation*}
In fact, a result of \citet*[][Theorem 7.4]{hanneke:survey} implies that $\hhat{\dc}(\varepsilon) = \sup_{\mathcal{P}} \sup_{h \in \mathbb{C}} \theta_{h,\mathcal{P}}(\varepsilon)$,
so that this would be an equivalent way to define $\hhat{\dc}(\varepsilon)$, which can sometimes be simpler to work with.
We can now express the bounds on the minimax label complexity implied by the best general results to date in the prior literature on
disagreement-based active learning \citep*[namely, the results of][]{hanneke:11a,dasgupta:07,koltchinskii:10,hanneke:12b,hanneke:survey},
summarized as follows (see the survey of \citealp*{hanneke:survey}, for detailed descriptions of the best-known logarithmic factors in these results).
\begin{itemize}
\item $\Lambda_{{\rm RE}}(\varepsilon,\delta) \lesssim \hhat{\dc}(\varepsilon) d \cdot {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right)$.
\item $\Lambda_{{\rm BN}(\beta)}(\varepsilon,\delta) \lesssim \frac{1}{(1-2\beta)^{2}} \hhat{\dc}(\varepsilon/(1-2\beta)) d \cdot {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right)$.
\item $\Lambda_{{\rm TN}(a,\alpha)}(\varepsilon,\delta) \lesssim a^{2} \left(\frac{1}{\varepsilon}\right)^{2-2\alpha} \hhat{\dc}( a \varepsilon^{\alpha} ) d \cdot {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right)$.
\item $\Lambda_{{\rm BC}(a,\alpha)}(\varepsilon,\delta) \lesssim a^{2} \left(\frac{1}{\varepsilon}\right)^{2-2\alpha} \hhat{\dc}( a \varepsilon^{\alpha} ) d \cdot {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right)$.
\item $\Lambda_{{\rm BE}(\nu)}(\varepsilon,\delta) \lesssim \left(\frac{\nu^{2}}{\varepsilon^{2}}+1\right) \hhat{\dc}(\nu+\varepsilon) d \cdot {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right)$.
\item $\Lambda_{{\rm AG}(\nu)}(\varepsilon,\delta) \lesssim \left(\frac{\nu^{2}}{\varepsilon^{2}}+1\right) \hhat{\dc}(\nu+\varepsilon) d \cdot {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right)$.
\end{itemize}
In particular, these bounds on $\Lambda_{{\rm TN}(a,\alpha)}(\varepsilon,\delta)$, $\Lambda_{{\rm BC}(a,\alpha)}(\varepsilon,\delta)$, $\Lambda_{{\rm BE}(\nu)}(\varepsilon,\delta)$, and $\Lambda_{{\rm AG}(\nu)}(\varepsilon,\delta)$
are the best general-case bounds on the label complexity of active learning in the prior literature (up to logarithmic factors), so that any improvements over these should be considered
an interesting advance in our understanding of the capabilities of active learning methods.
To compare these results to those stated in Section~\ref{sec:main}, we need to relate $\hhat{\dc}(\varepsilon)$ to the star number.
Interestingly, we find that these quantities are \emph{equal} (for $\varepsilon = 0$).
Specifically, the following result describes the relation between these two quantities;
its proof is included in Appendix~\ref{app:dc}.
This connection also plays a role in the proofs of some of our results from Section~\ref{sec:main}.
\begin{theorem}
\label{thm:dc-star}
$\forall \varepsilon \in (0,1]$,
$\hhat{\dc}(\varepsilon) = \mathfrak{s} \land \frac{1}{\varepsilon}$ and $\hhat{\dc}(0) = \mathfrak{s}$.
\end{theorem}
With this result in hand, we immediately observe that several of the upper bounds from Section~\ref{sec:main} offer
refinements over those stated in terms of $\hhat{\dc}(\cdot)$ above. For simplicity, we do not discuss differences in the logarithmic factors here.
Specifically, the upper bound on $\Lambda_{{\rm RE}}(\varepsilon,\delta)$ in Theorem~\ref{thm:realizable} refines that stated here by replacing the factor
$\hhat{\dc}(\varepsilon) d = \min\left\{ \mathfrak{s} d, \frac{d}{\varepsilon} \right\}$ with the sometimes-smaller factor $\min\left\{ \mathfrak{s}, \frac{d}{\varepsilon} \right\}$.
Likewise, the upper bound on $\Lambda_{{\rm BN}(\beta)}(\varepsilon,\delta)$ in Theorem~\ref{thm:bounded} refines the result stated here,
again by replacing the factor $\hhat{\dc}(\varepsilon/(1-2\beta)) d = \min\left\{ \mathfrak{s} d, \frac{(1-2\beta) d}{\varepsilon} \right\}$
with the sometimes-smaller factor $\min\left\{ \mathfrak{s}, \frac{(1-2\beta)d}{\varepsilon} \right\}$.
On the other hand, Theorem~\ref{thm:tsybakov} offers a much stronger refinement over the result stated above.
Specifically, in the case $\alpha \leq 1/2$, the upper bound in Theorem~\ref{thm:tsybakov} completely \emph{eliminates}
the factor of $\hhat{\dc}(a \varepsilon^{\alpha})$ from the upper bound on $\Lambda_{{\rm TN}(a,\alpha)}(\varepsilon,\delta)$ stated here
(i.e., replacing it with a universal constant). For the case $\alpha > 1/2$, the upper bound on $\Lambda_{{\rm TN}(a,\alpha)}(\varepsilon,\delta)$
in Theorem~\ref{thm:tsybakov} replaces this factor of $\hhat{\dc}(a \varepsilon^{\alpha}) = \min\left\{ \mathfrak{s}, \frac{1}{a \varepsilon^{\alpha}} \right\}$
with the factor $\min\left\{ \frac{\mathfrak{s}}{d}, \frac{1}{a^{1/\alpha} \varepsilon} \right\}^{2\alpha-1}$,
which is always smaller (for small $\varepsilon$ and large $d$).
The upper bounds on $\Lambda_{{\rm BC}(a,\alpha)}(\varepsilon,\delta)$ and $\Lambda_{{\rm AG}(\nu)}(\varepsilon,\delta)$ in Theorems \ref{thm:diameter-localization} and \ref{thm:agnostic}
are equivalent to those stated here; indeed, this is precisely how these results are obtained in Appendix~\ref{app:main-proofs}.
We have conjectured above that at least the dependence on $d$ and $\mathfrak{s}$ can be refined, analogous to the refinements for the
realizable case and bounded noise noted above.
However, we \emph{do} obtain refinements for the bound on $\Lambda_{{\rm BE}(\nu)}(\varepsilon,\delta)$ in Theorem~\ref{thm:benign},
replacing the factor of $\left(\frac{\nu^{2}}{\varepsilon^{2}} + 1\right) \hhat{\dc}(\nu+\varepsilon) d = \left(\frac{\nu^{2}}{\varepsilon^{2}}+1\right)\min\left\{\mathfrak{s} d, \frac{d}{\nu+\varepsilon} \right\}$ in the upper bound here
with a factor $\frac{\nu^{2}}{\varepsilon^{2}} d + \min\left\{\mathfrak{s},\frac{d}{\varepsilon}\right\}$, which is sometimes significantly smaller (for $\varepsilon \ll \nu \ll 1$ and large $d$).
\subsection{The Splitting Index}
\label{sec:splitting}
Another, very different, approach to the design and analysis of active learning algorithms
was proposed by \citet*{dasgupta:05}: namely, the \emph{splitting} approach.
In particular, this technique has the desirable property that it yields distribution-dependent
label complexity bounds for the realizable case which, even when the marginal distribution $\mathcal{P}$
is held fixed, (almost) imply near-minimax performance.
The intuition behind this technique is that the objective in the realizable case (achieving error rate
at most $\varepsilon$) is typically well-approximated by the related objective of reducing the \emph{diameter}
of the version space (set of classifiers consistent with the observed labels) to size at most $\varepsilon$.
From this perspective, at any given time, the impediments to achieving this objective are clearly
identifiable: pairs of classifiers $\{h,g\}$ in $\mathbb{C}$ consistent with all labels observed thus far,
yet with $\mathcal{P}(x : h(x) \neq g(x)) > \varepsilon$. Supposing we have only a finite number of such classifiers
(which can be obtained if we first replace $\mathbb{C}$ by a fine-grained finite \emph{cover} of $\mathbb{C}$),
we can then estimate the \emph{usefulness} of a given point $X_{i}$ by the number of these
pairs it would be guaranteed to eliminate if we were to request its label (supposing the worse
of the two possible labels); by ``eliminate,'' we mean that at least one of the two classifiers
will be inconsistent with the observed label. If we always request labels of points guaranteed
to eliminate a large fraction of the surviving $\varepsilon$-separated pairs, we will quickly arrive
at a version space of diameter $\varepsilon$, and can then return any surviving classifier.
\citet*{dasgupta:05} further applies this strategy in tiers, first eliminating at least one classifier
from every $\frac{1}{2}$-separated pair, then repeating this for the remaining $\frac{1}{4}$-separated
pairs, and so on. This allows the label complexity to be \emph{localized}, in the sense that
the surviving $\Delta$-separated pairs we need to eliminate will be composed of classifiers
within distance $2\Delta$ of $f^{\star}_{\mathcal{P}_{XY}}$ (or the representative thereof in the initial
finite cover of $\mathbb{C}$).
The analysis of this method naturally leads to the following definition from \citet*{dasgupta:05}.
For any finite set $Q \subseteq \{\{h,g\} : h,g \in \mathbb{C}\}$ of unordered pairs of classifiers in $\mathbb{C}$,
for any $x \in \mathcal X$ and $y \in \mathcal Y$, let $Q_{x}^{y} = \{\{h,g\} \in Q : h(x) = g(x) = y\}$, and define
\begin{equation*}
{\rm Split}(Q,x) = |Q| - \max_{y \in \mathcal Y} |Q_{x}^{y}|.
\end{equation*}
This represents the number of pairs guaranteed to be eliminated (as described above) by
requesting the label at a point $x$.
The splitting index is then defined as follows.
\begin{definition}
\label{def:splitting}
For any $\rho,\Delta,\tau \in [0,1]$, a set $\H \subseteq \mathbb{C}$ is said to be $(\rho,\Delta,\tau)$-splittable under a probability measure $\mathcal{P}$ over $\mathcal X$ if,
for all finite $Q \subseteq \{\{h,g\} \subseteq \H : \mathcal{P}(x : h(x) \neq g(x)) \geq \Delta \}$,
\begin{equation*}
\mathcal{P}(x : {\rm Split}(Q,x) \geq \rho |Q|) \geq \tau.
\end{equation*}
For any classifier $h : \mathcal X \to \mathcal Y$, any probability measure $\mathcal{P}$ over $\mathcal X$, and any
$\varepsilon,\tau \in [0,1]$, the \emph{splitting index} is defined as
\begin{equation*}
\rho_{h,\mathcal{P}}(\varepsilon;\tau) = \sup\left\{ \rho \in [0,1] : \forall \Delta \geq \varepsilon, {\rm B}_{\mathcal{P}}(h,4\Delta) \text{ is } (\rho,\Delta,\tau)\text{-splittable under } \mathcal{P} \right\}.
\end{equation*}
\end{definition}
\citet*{dasgupta:05} proves a bound on the label complexity of a general active learning algorithm based on the above strategy, in the realizable case,
expressed in terms of the splitting index. Specifically, for any $\tau > 0$, letting $\rho = \rho_{f^{\star}_{\mathcal{P}_{XY}},\mathcal{P}}(\varepsilon/4;\tau)$,
\citet*{dasgupta:05} finds that for that algorithm to achieve error rate at most $\varepsilon$ with probability at least $1-\delta$,
it suffices to use a number of label requests
\begin{equation}
\label{eqn:dasgupta-splitting-bound}
\frac{d}{\rho} {\rm polylog}\left(\frac{d}{\varepsilon\delta\tau\rho}\right).
\end{equation}
The $\tau$ argument to $\rho_{h,\mathcal{P}}(\varepsilon;\tau)$ captures the trade-off between the number of label requests and the number of unlabeled samples available,
with smaller $\tau$ corresponding to the scenario where more unlabeled data are available, and a larger value of $\rho_{h,\mathcal{P}}(\varepsilon;\tau)$.
Specifically, \citet*{dasgupta:05} argues that $\tilde{O}\left(\frac{d}{\tau \rho}\right)$ unlabeled samples suffice to achieve the above result.
In our present model, we suppose an abundance of unlabeled data, and as such, we are interested in the behavior for very small $\tau$.
However, note that the logarithmic factors in the above bound have an inverse dependence on $\tau$, so that taking $\tau$ too small can
potentially increase the value of the bound. It is not presently known whether or not this is necessary (though intuitively it seems not to be).
However, for the purpose of comparison to our results in Section~\ref{sec:main}, we will ignore this logarithmic dependence on $1/\tau$,
and focus on the leading factor. In this case, we are interested in the value $\lim\limits_{\tau \to 0} \rho_{h,\mathcal{P}}(\varepsilon;\tau)$.
Additionally, to convert \eqref{eqn:dasgupta-splitting-bound} into a distribution-free bound for the purpose of comparison to the results in Section~\ref{sec:main},
we should minimize this value over the choice of $\mathcal{P}$ and $h \in \mathbb{C}$.
Formally, we are interested in the following quantity, defined for any $\varepsilon \in [0,1]$.
\begin{equation*}
\hhat{\rho}(\varepsilon) = \inf_{P} \inf_{h \in \mathbb{C}} \lim_{\tau \to 0} \rho_{h,P}(\varepsilon;\tau).
\end{equation*}
In particular, in terms of this quantity, the maximum possible value of the bound \eqref{eqn:dasgupta-splitting-bound} for a given hypothesis class $\mathbb{C}$
is at least
\begin{equation*}
\frac{d}{\hhat{\rho}(\varepsilon/4)} {\rm polylog}\left(\frac{d}{\varepsilon\delta}\right).
\end{equation*}
To compare this to the upper bound in Theorem~\ref{thm:realizable}, we need to
relate $\frac{1}{\hhat{\rho}(\varepsilon)}$ to the star number. Again, we find that these quantities are
essentially \emph{equal} (as $\varepsilon \to 0$), as stated in the following theorem.
\begin{theorem}
\label{thm:splitting-star}
$\forall \varepsilon \in (0,1]$,
$\left\lfloor \frac{1}{\hhat{\rho}(\varepsilon)} \right\rfloor = \mathfrak{s} \land \left\lfloor \frac{1}{\varepsilon} \right\rfloor$.
\end{theorem}
The proof of this result is included in Appendix~\ref{app:splitting}.
We note that the inequalities $\mathfrak{s} \land \left\lfloor\frac{1}{\varepsilon}\right\rfloor \leq \left\lfloor \frac{1}{\hhat{\rho}(\varepsilon)} \right\rfloor \leq \left\lfloor \frac{1}{\varepsilon} \right\rfloor$
were already implicit in the original work of \citet*[][Corollary 3 and Lemma 1]{dasgupta:05}.
For completeness (and to make the connection explicit), we include these arguments in the proof given in Appendix~\ref{app:splitting},
along with our proof that $\left\lfloor \frac{1}{\hhat{\rho}(\varepsilon)} \right\rfloor \leq \mathfrak{s}$ (which was heretofore unknown).
Plugging this into the above bound, we see that the maximum possible value of the bound \eqref{eqn:dasgupta-splitting-bound} for a given hypothesis class $\mathbb{C}$
is at least
\begin{equation*}
\min\left\{ \mathfrak{s} d, \frac{d}{\varepsilon} \right\} {\rm polylog}\left(\frac{d}{\varepsilon\delta}\right).
\end{equation*}
Note that the upper bound in Theorem~\ref{thm:realizable} refines this by reducing the first term in the ``$\min$'' from $\mathfrak{s} d$ to simply $\mathfrak{s}$.
\citet*{dasgupta:05} also argues for a kind of lower bound in terms of the splitting index,
which was reformulated as a lower bound on the minimax label complexity (for a fixed $\mathcal{P}$)
in the realizable case by \citet*{hanneke:12c,hanneke:survey}.
In our present distribution-free style of analysis, the implication of that result is the following lower bound.
\begin{equation*}
\Lambda_{{\rm RE}}(\varepsilon,\delta) \gtrsim \frac{1}{\hhat{\rho}(4\varepsilon)}.
\end{equation*}
Based on Theorem~\ref{thm:splitting-star}, we see that the $\min\left\{ \mathfrak{s}, \frac{1}{\varepsilon} \right\}$ term in the lower bound
of Theorem~\ref{thm:realizable} follows immediately from this lower bound. For completeness, in Appendix~\ref{app:main-proofs},
we directly prove this term in the lower bound, based on a more-direct argument than that used to establish the above lower bound.
We note, however, that \citet*[][Corollary 3]{dasgupta:05} also describes a technique for obtaining lower bounds, which is
essentially equivalent to that used in Appendix~\ref{app:main-proofs} to obtain this term (and furthermore, makes use of
a distribution-dependent version of the ``star'' idea).
The upper bounds of \citet*{dasgupta:05} have also been extended to the bounded noise setting.
In particular, \citet*{hanneke:12c} and \citet*{hanneke:survey} have proposed variants of the
splitting approach, which are robust to bounded noise. They have additionally bounded the label
complexities of these methods in terms of the splitting index. Similarly to the above discussion
of the realizable case, the worst-case values of these bounds for any given hypothesis class $\mathbb{C}$ are larger
than those stated in Theorem~\ref{thm:bounded} by factors related to the VC dimension (logarithmic
factors aside). We refer the interested readers to these sources for the details of those bounds.
\subsection{The Teaching Dimension}
\label{sec:xtd}
Another quantity that has been used to bound the label complexity of certain active learning
methods is the \emph{extended teaching dimension growth function}. This quantity was introduced
by \citet*{hanneke:07a}, inspired by analogous notions used to tightly-characterize the query complexity
of \emph{Exact} learning with membership queries \citep*{hegedus:95,hellerstein:96}. The term
\emph{teaching dimension} takes its name from the literature on Exact teaching \citep*{goldman:95},
where the teaching dimension characterizes the minimum number of well-chosen labeled data points
sufficient to guarantee that the only classifier in $\mathbb{C}$ consistent with these labels is the target function.
\citet*{hegedus:95} extends this to target functions not contained in $\mathbb{C}$, in which case the objective
is simply to leave at most one consistent classifier in $\mathbb{C}$; he refers to the minimum number of
points sufficient to achieve this as the \emph{extended teaching dimension}, and argues that this
quantity can be used to characterize the minimum number of \emph{membership queries} by a
learning algorithm sufficient to guarantee that the only classifier in $\mathbb{C}$ consistent with the
returned labels is the target function (which is the objective in the \emph{Exact} learning model).
\citet*{hanneke:07a} transfers this strategy to the statistical setting studied here (where the objective
is only to obtain excess error rate $\varepsilon$ with probability $1-\delta$, rather than exactly identifying
a target function). That work introduces empirical versions of the teaching dimension and
extended teaching dimension, and defines distribution-dependent bounds on these quantities.
It then proves upper and lower bounds on the label complexity in terms of these quantities.
For our present purposes, we will be most-interested in a particular distribution-free
upper bound on these quantities, called the \emph{extended teaching dimension growth function},
also introduced by \citet*{hanneke:06,hanneke:07a}. Since both this quantity and the star number are distribution-free,
they can be directly compared.
We introduce these quantities formally as follows.
For any $m \in \mathbb{N} \cup \{0\}$ and $S \in \mathcal X^{m}$, and for any $h : \mathcal X \to \mathcal Y$,
define the \emph{version space} $V_{S,h} = \{ g \in \mathbb{C} : \forall x \in S, g(x) = h(x) \}$ \citep*{mitchell:77}.
For any $m \in \mathbb{N}$ and $\mathcal U \in \mathcal X^{m}$,
let $\mathbb{C}[\mathcal U]$ denote an arbitrary subset of classifiers in $\mathbb{C}$ such that, $\forall h \in \mathbb{C}$, $|\mathbb{C}[\mathcal U] \cap V_{\mathcal U,h}| = 1$:
that is, $\mathbb{C}[\mathcal U]$ contains exactly one classifier from each equivalence class in $\mathbb{C}$ induced by the classifications of $\mathcal U$.
For any classifier $h : \mathcal X \to \mathcal Y$, define
\begin{equation*}
{\rm TD}(h,\mathbb{C}[\mathcal U],\mathcal U) = \min\{ t \in \mathbb{N} \cup \{0\} : \exists S \in \mathcal U^{t} \text{ s.t. } |V_{S,h} \cap \mathbb{C}[\mathcal U]| \leq 1 \},
\end{equation*}
the \emph{empirical teaching dimension} of $h$ on $\mathcal U$ with respect to $\mathbb{C}[\mathcal U]$.
Any $S \in \bigcup_{t} \mathcal U^{t}$ with
$|V_{S,h} \cap \mathbb{C}[\mathcal U]| \leq 1$ is called a \emph{specifying set} for $h$ on $\mathcal U$ with respect to $\mathbb{C}[\mathcal U]$;
thus, ${\rm TD}(h,\mathbb{C}[\mathcal U],\mathcal U)$ is the size of a \emph{minimal specifying set} for $h$ on $\mathcal U$ with respect to $\mathbb{C}[\mathcal U]$.
Equivalently, $S \in \bigcup_{t} \mathcal U^{t}$ is a specifying set for $h$ on $\mathcal U$ with respect to $\mathbb{C}[\mathcal U]$ if and only if
${\rm DIS}(V_{S,h}) \cap \mathcal U = \emptyset$.
Also define ${\rm TD}(h,\mathbb{C},m) = \max\limits_{\mathcal U \in \mathcal X^{m}} {\rm TD}(h,\mathbb{C}[\mathcal U],\mathcal U)$,
${\rm TD}(\mathbb{C},m) = \max\limits_{h \in \mathbb{C}} {\rm TD}(h,\mathbb{C},m)$ (the \emph{teaching dimension growth function}),
and ${\rm XTD}(\mathbb{C},m) = \max\limits_{h : \mathcal X \to \mathcal Y} {\rm TD}(h,\mathbb{C},m)$ (the \emph{extended teaching dimension growth function}).
\citet*{hanneke:07a} proves two upper bounds on the label complexity of active learning relevant to our present discussion.
They are summarized as follows (see the original source for the precise logarithmic factors).\footnote{Here we have simplified
the arguments $m$ to the ${\rm XTD}(\mathbb{C},m)$ instances compared to those of \citet*{hanneke:07a},
using monotonicity of $m \mapsto {\rm XTD}(\mathbb{C},m)$, combined with the basic observation that ${\rm XTD}(\mathbb{C},m k) \leq {\rm XTD}(\mathbb{C},m) k$ for any integer $k \geq 1$.}
\begin{itemize}
\item[$\bullet$] $\Lambda_{{\rm RE}}(\varepsilon,\delta) \lesssim {\rm XTD}\left( \mathbb{C}, \left\lceil \frac{1}{\varepsilon} \right\rceil \right) d \cdot {\rm polylog}\left(\frac{d}{\varepsilon\delta}\right)$.
\item[$\bullet$] $\Lambda_{{\rm AG}(\nu)}(\varepsilon,\delta) \lesssim \left(\frac{\nu^{2}}{\varepsilon^{2}}+1\right) {\rm XTD}\left( \mathbb{C}, \left\lceil \frac{1}{\nu+\varepsilon} \right\rceil \right) d \cdot {\rm polylog}\left(\frac{d}{\varepsilon\delta}\right)$.
\end{itemize}
Since ${\rm BE}(\nu) \subseteq {\rm AG}(\nu)$, we have the further implication that
\begin{equation*}
\Lambda_{{\rm BE}(\nu)}(\varepsilon,\delta) \lesssim \left(\frac{\nu^{2}}{\varepsilon^{2}}+1\right) {\rm XTD}\left( \mathbb{C}, \left\lceil \frac{1}{\nu+\varepsilon} \right\rceil \right) d \cdot {\rm polylog}\left(\frac{d}{\varepsilon\delta}\right).
\end{equation*}
Additionally, by a refined argument of \citet*{hegedus:95}, the ideas of \citet*{hanneke:07a} can be applied (see \citealp*{hanneke:06,hanneke:thesis}) to show that
\begin{equation*}
\Lambda_{{\rm RE}}(\varepsilon,\delta) \lesssim \frac{{\rm XTD}( \mathbb{C}, \lceil d/\varepsilon \rceil )}{\log_{2}({\rm XTD}(\mathbb{C}, \lceil d/\varepsilon \rceil))} d \cdot {\rm polylog}\left(\frac{d}{\varepsilon\delta}\right).
\end{equation*}
To compare these bounds to the results stated in Section~\ref{sec:main}, we will need to relate the quantity ${\rm XTD}(\mathbb{C},m)$ to the star number.
Although it may not be obvious from a superficial reading of the definitions,
we find that these quantities are \emph{exactly equal} (as $m\to\infty$). Thus, the
extended teaching dimension growth function is simply an alternative way of referring to the star number (and vice versa),
as they define the same quantity.\footnote{In this sense, the star number is not really a \emph{new} quantity
to the active learning literature, but rather a simplified definition for the already-familiar
extended teaching dimension growth function.}
This equivalence is stated formally in the following theorem, the proof of which is included in Appendix~\ref{app:xtd}.
\begin{theorem}
\label{thm:xtd}
$\forall m \in \mathbb{N}$, ${\rm XTD}(\mathbb{C}, m) = {\rm TD}(\mathbb{C},m) = \min\{\mathfrak{s}, m\}$.
\end{theorem}
We note that the inequalities $\min\{\mathfrak{s},m\} \leq {\rm TD}(\mathbb{C},m) \leq {\rm XTD}(\mathbb{C},m) \leq m$ follow readily from previously-established facts about the teaching dimension.
For instance, \citet*{fan:12} notes that the teaching dimension of any class is at least the maximum degree of its one-inclusion graph;
applying this fact to $\mathbb{C}[\mathcal U]$ and maximizing over the choice of $\mathcal U \in \mathcal X^{m}$, this maximum degree becomes $\min\{\mathfrak{s},m\}$ (by definition of $\mathfrak{s}$).
However, the inequality ${\rm XTD}(\mathbb{C},m) \leq \mathfrak{s}$ and the resulting fact that ${\rm XTD}(\mathbb{C},m) = {\rm TD}(\mathbb{C},m)$ are apparently new.
In fact, in the process of proving this theorem, we establish another remarkable fact:
that \emph{every} minimal specifying set is a star set. This is stated formally
in the following lemma, the proof of which is also included in Appendix~\ref{app:xtd}.
\begin{lemma}
\label{lem:spec-star-set}
For any $h : \mathcal X \to \mathcal Y$, $m \in \mathbb{N}$, and $\mathcal U \in \mathcal X^{m}$,
every minimal specifying set for $h$ on $\mathcal U$ with respect to $\mathbb{C}[\mathcal U]$
is a star set for $\mathbb{C} \cup \{h\}$ centered at $h$.
\end{lemma}
Using Theorem~\ref{thm:xtd}, we can now compare the results above to those in Section~\ref{sec:main}.
For simplicity, we will not discuss the differences in logarithmic factors here.
Specifically, Theorem~\ref{thm:realizable} refines the results here on $\Lambda_{{\rm RE}}(\varepsilon,\delta)$,
replacing a factor of $\min\left\{ {\rm XTD}(\mathbb{C},\lceil 1/\varepsilon \rceil) d, \frac{{\rm XTD}(\mathbb{C},\lceil d/\varepsilon \rceil) d}{\log({\rm XTD}(\mathbb{C},\lceil d/\varepsilon \rceil))} \right\} \approx \min\left\{ \mathfrak{s} d, \frac{d}{\varepsilon}, \frac{\mathfrak{s} d}{\log(\mathfrak{s})}, \frac{d^{2}}{\varepsilon \log(d/\varepsilon)} \right\}$ implied by the above results
with a factor of $\min\left\{ \mathfrak{s}, \frac{d}{\varepsilon}, \frac{\mathfrak{s} d}{\log(\mathfrak{s})} \right\}$, thus reducing the first term in the ``$\min$'' by a factor of $d$
(though see below, as \citealp*{hanneke:14a}, have already shown this to be possible, directly in terms of ${\rm XTD}(\mathbb{C},m)$).
Theorem~\ref{thm:xtd} further reveals that the above bound on $\Lambda_{{\rm AG}(\nu)}(\varepsilon,\delta)$ is equivalent (up to logarithmic factors)
to that stated in Theorem~\ref{thm:agnostic}. However, the bound on $\Lambda_{{\rm BE}(\nu)}(\varepsilon,\delta)$ in Theorem~\ref{thm:benign}
refines that implied above, replacing a factor $\left(\frac{\nu^{2}}{\varepsilon^{2}}+1\right) {\rm XTD}\left(\mathbb{C}, \left\lceil \frac{1}{\nu+\varepsilon} \right\rceil \right) d \approx \left(\frac{\nu^{2}}{\varepsilon^{2}}+1\right) \min\left\{ \mathfrak{s} d, \frac{d}{\nu+\varepsilon}\right\}$
with a factor $\frac{\nu^{2}}{\varepsilon^{2}}d + \min\left\{ \mathfrak{s}, \frac{d}{\varepsilon} \right\}$, which can be significantly smaller for $\varepsilon \ll \nu \ll 1$ and large $d$.
\citet*{hanneke:06,hanneke:07a} also proves a \emph{lower bound} on the label complexity
of active learning in the realizable case, based on the following modification of
the extended teaching dimension.
For any set $\H \subseteq \mathbb{C}$, classifier $h : \mathcal X \to \mathcal Y$, $m \in \mathbb{N}$, $\mathcal U \in \mathcal X^{m}$, and $\delta \in [0,1]$,
define the \emph{partial teaching dimension} as
\begin{equation*}
{\rm XPTD}(h,\H[\mathcal U],\mathcal U,\delta) = \min\{ t \in \mathbb{N} \cup \{0\} : \exists S \in \mathcal U^{t} \text{ s.t. } |V_{S,h} \cap \H[\mathcal U]| \leq \delta |\H[\mathcal U]|+1 \},
\end{equation*}
and let ${\rm XPTD}(\H,m,\delta) = \max\limits_{h : \mathcal X \to \mathcal Y} \max\limits_{\mathcal U \in \mathcal X^{m}} {\rm XPTD}(h,\H[\mathcal U],\mathcal U,\delta)$.
\citet*{hanneke:06,hanneke:07a} proves that
\begin{equation*}
\Lambda_{{\rm RE}}(\varepsilon,\delta) \geq \max_{\H \subseteq \mathbb{C}} {\rm XPTD}\left(\H, \left\lceil \frac{1-\varepsilon}{\varepsilon} \right\rceil, \delta\right).
\end{equation*}
The following result relates this quantity to the star number.
\begin{theorem}
\label{thm:xptd}
$\forall m \in \mathbb{N}$, $\forall \delta \in [0,1/2]$,
\begin{equation*}
\left\lceil (1-2\delta) \min\{ \mathfrak{s}, m \} \right\rceil
\leq \max_{\H \subseteq \mathbb{C}} {\rm XPTD}(\H,m,\delta)
\leq \left\lceil \left(1-\frac{\delta}{1+\delta}\right) \min\{\mathfrak{s},m\} \right\rceil.
\end{equation*}
\end{theorem}
The proof is in Appendix~\ref{app:xtd}.
Note that, combined with the lower bound of \citet*{hanneke:06,hanneke:07a},
this immediately implies the part of the lower bound in Theorem~\ref{thm:realizable}
involving $\mathfrak{s}$. In Appendix~\ref{app:main-proofs}, we provide a direct proof for
this term in the lower bound, based on an argument similar to that of \citet*{hanneke:07a}.
\subsubsection{The Version Space Compression Set Size}
\label{sec:hatn}
More-recently, \citet*{el-yaniv:10,el-yaniv:12,hanneke:14a} have studied a quantity $\hat{n}_{h}(\mathcal U)$
(for a sequence $\mathcal U \in \bigcup_{m} \mathcal X^{m}$ and classifier $h$),
termed the minimal \emph{version space compression set size}, defined as the size of the
smallest subsequence $S \subseteq \mathcal U$ for which $V_{S,h} = V_{\mathcal U,h}$.\footnote{The quantity
studied there is defined slightly differently, but is easily seen to be equivalent to this definition.}
It is easy to see that, when $h \in \mathbb{C}$, the version space compression set size is equivalent to the empirical teaching dimension:
that is, $\forall h \in \mathbb{C}$,
\begin{equation*}
\hat{n}_{h}(\mathcal U) = {\rm TD}(h,\mathbb{C}[\mathcal U],\mathcal U).
\end{equation*}
To see this, note that since $|V_{\mathcal U,h} \cap \mathbb{C}[\mathcal U]| = 1$,
any $S \subseteq \mathcal U$ with $V_{S,h} = V_{\mathcal U,h}$ has $|V_{S,h} \cap \mathbb{C}[\mathcal U]| = 1$,
and hence is a specifying set for $h$ on $\mathcal U$ with respect to $\mathbb{C}[\mathcal U]$.
On the other hand, for any $S \subseteq \mathcal U$, we (always) have $V_{S,h} \supseteq V_{\mathcal U,h}$,
so that if $|V_{S,h} \cap \mathbb{C}[\mathcal U]| \leq 1$, then $V_{S,h} \cap \mathbb{C}[\mathcal U] \supseteq V_{\mathcal U,h} \cap \mathbb{C}[\mathcal U]$
and $|V_{S,h} \cap \mathbb{C}[\mathcal U]| \geq |V_{\mathcal U,h} \cap \mathbb{C}[\mathcal U]| = 1 \geq |V_{S,h} \cap \mathbb{C}[\mathcal U]|$, which together imply
$V_{S,h} \cap \mathbb{C}[\mathcal U] = V_{\mathcal U,h} \cap \mathbb{C}[\mathcal U]$; thus, $V_{S,h} \subseteq \{ g \in \mathbb{C} : \forall x \in \mathcal U, g(x) = h(x) \} = V_{\mathcal U,h} \subseteq V_{S,h}$,
so that $V_{S,h} = V_{\mathcal U,h}$: that is, $S$ is a version space compression set.
Thus, in the case $h \in \mathbb{C}$, any version space compression set $S$ is a specifying set
for $h$ on $\mathcal U$ with respect to $\mathbb{C}[\mathcal U]$ and vice versa.
That $\hat{n}_{h}(\mathcal U) = {\rm TD}(h,\mathbb{C}[\mathcal U],\mathcal U)$ $\forall h \in \mathbb{C}$ follows immediately from this equivalence.
In particular, combined with Theorem~\ref{thm:xtd}, this implies that $\forall m \in \mathbb{N}$,
\begin{equation}
\label{eqn:hatn-star}
\max_{\mathcal U \in \mathcal X^{m}} \max_{h \in \mathbb{C}} \hat{n}_{h}(\mathcal U) = {\rm TD}(\mathbb{C},m) = \min\{\mathfrak{s},m\}.
\end{equation}
Letting $\hat{n}_{m} = \hat{n}_{f^{\star}_{\mathcal{P}_{XY}}}(\{X_{1},\ldots,X_{m}\})$,
\citet*{hanneke:14a} have shown that, in the realizable case, for the CAL active learning algorithm \citep*[proposed by][]{cohn:94}
to achieve error rate at most $\varepsilon$ with probability at least $1-\delta$, it suffices to use a budget $n$ of any size at least
\begin{equation*}
\max_{1 \leq m \leq M_{\varepsilon,\delta}}
\hat{n}_{m} \cdot {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right),
\end{equation*}
where $M_{\varepsilon,\delta} \lesssim \frac{1}{\varepsilon}\left(d{\rm Log}\left(\frac{1}{\varepsilon}\right)+{\rm Log}\left(\frac{1}{\delta}\right)\right)$ is a bound
on the sample complexity of passive learning by returning an arbitrary classifier in the version space
\citep*{vapnik:82,vapnik:98,blumer:89}.
They further provide a distribution-dependent bound (to remove the dependence on the data here)
based on confidence bounds on $\hat{n}_{m}$ (analogous to the aforementioned distribution-dependent
bounds on the empirical teaching dimension studied by \citealp*{hanneke:07a}).
For our purposes (distribution-free, data-independent bounds), we can simply take the maximum over possible data sets and possible $f^{\star}_{\mathcal{P}_{XY}}$ functions,
so that the above bound becomes
\begin{multline*}
\max_{x_{1},x_{2},\ldots \in \mathcal X} \max_{h \in \mathbb{C}} \max_{1 \leq m \leq M_{\varepsilon,\delta}} \hat{n}_{h}(\{x_{1},\ldots,x_{m}\}) {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right)
\\ = {\rm TD}\left(\mathbb{C}, M_{\varepsilon,\delta}\right) {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right)
\lesssim {\rm TD}\left(\mathbb{C}, \left\lfloor \frac{d}{\varepsilon} \right\rfloor \right) {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right).
\end{multline*}
Combining this with \eqref{eqn:hatn-star}, we find that the label complexity of CAL in the realizable case is at most
\begin{equation*}
\min\left\{\mathfrak{s}, \frac{d}{\varepsilon}\right\} {\rm polylog}\left(\frac{1}{\varepsilon\delta}\right),
\end{equation*}
which matches the upper bound on the minimax label complexity from Theorem~\ref{thm:realizable} up to logarithmic factors.
\subsection{The Doubling Dimension}
\label{sec:doubling}
Another quantity of interest in the learning theory literature is the \emph{doubling dimension},
also known as the \emph{local metric entropy} \citep*{lecam:73,yang:99b,gupta:03,long:09}.
Specifically, for any set $\H$ of classifiers, a set of classifiers ${\cal{G}}$ is an $\varepsilon$-cover of $\H$
(with respect to the $\mathcal{P}({\rm DIS}(\{\cdot,\cdot\}))$ pseudometric) if
\begin{equation*}
\sup_{h \in \H} \inf_{g \in {\cal{G}}} \mathcal{P}(x : g(x) \neq h(x)) \leq \varepsilon.
\end{equation*}
Let ${\cal{N}}(\varepsilon,\H,\mathcal{P})$ denote the minimum cardinality $|{\cal{G}}|$ over all $\varepsilon$-covers ${\cal{G}}$ of $\H$,
or else ${\cal{N}}(\varepsilon,\H,\mathcal{P}) = \infty$ if no finite $\varepsilon$-cover of $\H$ exists.
The doubling dimension (at $h$) is defined as follows.
\begin{definition}
\label{def:doubling}
For any $\varepsilon \in (0,1]$, any probability measure $P$ over $\mathcal X$, and any classifier $h$,
define
\begin{equation*}
D_{h,P}(\varepsilon) = \max_{r \geq \varepsilon} \log_{2}\left( {\cal{N}}\left(r/2, {\rm B}_{P}(h,r), P \right) \right).
\end{equation*}
\end{definition}
The quantity $D_{\varepsilon} = D_{f^{\star}_{\mathcal{P}_{XY}},\mathcal{P}}(\varepsilon)$ is known to be useful in bounding the sample complexity of passive learning.
Specifically, \citet*{long:07,long:09} have shown that there is a passive learning algorithm achieving
sample complexity $\lesssim \frac{D_{\varepsilon/4}}{\varepsilon} + \frac{1}{\varepsilon} \log\left(\frac{1}{\delta}\right)$
for $\mathcal{P}_{XY} \in {\rm RE}$.
Furthermore, though we do not go into the details here, by a combination of the ideas from \citet*{dasgupta:05}, \citet*{balcan:09}, and \citet*{hanneke:07b},
it is possible to show that a certain active learning algorithm achieves a label complexity $\lesssim 4^{D_{\varepsilon}} D_{\varepsilon} \cdot {\rm polylog}(\frac{1}{\varepsilon\delta})$
for $\mathcal{P}_{XY} \in {\rm RE}$, though this is typically a very loose upper bound.
To our knowledge, the question of the worst-case value of the doubling dimension for a given hypothesis class $\mathbb{C}$
has not previously been explored in the literature (though there is an obvious $O(d \log(1/\varepsilon))$ upper bound).
Here we obtain upper and lower bounds on this worst-case value, expressed in terms of the star number.
While this relation generally has a wide range (roughly a factor of $d$), it does have the interesting implication
that the doubling dimension is \emph{bounded} if and only if $\mathfrak{s} < \infty$.
Specifically, we have the following theorem, the proof of which is included in Appendix~\ref{app:doubling}.
\begin{theorem}
\label{thm:dd-star}
$\forall \varepsilon \in (0,1/4]$,
$\max\left\{ d, {\rm Log}\left( \mathfrak{s} \land \frac{1}{\varepsilon} \right) \right\} \lesssim \sup\limits_{P} \sup\limits_{h \in \mathbb{C}} D_{h,P}(\varepsilon)
\lesssim d {\rm Log}\left( \mathfrak{s} \land \frac{1}{\varepsilon} \right)$.
\end{theorem}
One can show that the gap between the upper and lower bounds on $\sup_{P} \sup_{h \in \mathbb{C}} D_{h,P}(\varepsilon)$ in this result
cannot generally be improved by much without sacrificing generality or introducing additional quantities.
Specifically, for the class $\mathbb{C}$ discussed in Appendix~\ref{app:lb-tight}, we have
$\sup_{P} \sup_{h \in \mathbb{C}} D_{h,P}(\varepsilon)$ $\leq \sup_{P} \log_{2}( {\cal{N}}(\varepsilon/2,\mathbb{C},P) ) \lesssim \max\left\{ d, {\rm Log}\left(\mathfrak{s} \land \frac{1}{\varepsilon}\right) \right\}$,
so that the lower bound above is sometimes tight to within a universal constant factor.
For the class $\mathbb{C}$ discussed in Appendix~\ref{app:ub-tight}, based on a result of \citet*[][Lemma 4]{raginsky:11},
one can show $\sup_{P} \sup_{h \in \mathbb{C}} D_{h,P}(\varepsilon) \gtrsim d {\rm Log}\left( \frac{\mathfrak{s}}{d} \land \frac{1}{\varepsilon} \right)$,
so that the above upper bound is sometimes tight, aside from a small difference in the logarithmic factor (dividing $\mathfrak{s}$ by $d$).
Interestingly, in the process of proving the upper bound in Theorem~\ref{thm:dd-star}, we also establish the following
inequality relating the doubling dimension and the disagreement coefficient, holding for any classifier $h$, any probability measure $\mathcal{P}$ over $\mathcal X$, and any $\varepsilon \in (0,1]$.
\begin{equation*}
D_{h,\mathcal{P}}(\varepsilon) \leq 2 d \log_{2}\left( 22 e^{2} \theta_{h,\mathcal{P}}(\varepsilon) \right).
\end{equation*}
This inequality may be of independent interest, as it enables comparisons between
results in the literature expressed in terms of these quantities. For instance,
it implies that in the realizable case, the passive learning sample complexity
bound of \citet*{long:09} is no larger than that of \citet*{gine:06} (aside
from constant factors).
\section{Conclusions}
\label{sec:conclusions}
In this work, we derived upper and lower bounds on the minimax label complexity of active
learning under several noise models. In most cases, these new bounds offer refinements
over the best results in the prior literature. Furthermore, in the case of Tsybakov
noise, we discovered the heretofore-unknown fact that the minimax label complexity of
active learning with VC classes is \emph{always} smaller than that of passive learning.
We expressed each of these bounds in terms of a simple combinatorial complexity measure,
termed the \emph{star number}. We further found that almost all of the distribution-dependent
and sample-dependent complexity measures in the prior active learning literature are exactly
equal to the star number when maximized over the choice of distribution or data set.
The bounds derived here are all distribution-free, in the sense that they are expressed without
dependence or restrictions on the marginal distribution $\mathcal{P}$ over $\mathcal X$. They are also worst-case
bounds, in the sense that they express the maximum of the label complexity over the distributions
in the noise model $\mathbb{D}$, rather than expressing a bound on the label complexity achieved by
a given algorithm as a function of $\mathcal{P}_{XY}$. As observed by \citet*{dasgupta:05}, there are some
cases in which smaller label complexities can be achieved under restrictions on the marginal
distribution $\mathcal{P}$, and some cases in which there are achievable label complexities which
exhibit a range of values depending on $\mathcal{P}_{XY}$ \citep*[see also][for further exploration of this]{hanneke:10a,hanneke:12a}.
Our results reveal that in some cases, such as Tsybakov noise with $\alpha \leq 1/2$,
these issues might typically not be of much significance (aside from logarithmic factors).
However, in other cases, particularly when $\mathfrak{s} = \infty$, the issue of expressing
distribution-dependent bounds on the label complexity is clearly an important one.
In particular, the question of the minimax label complexity of active learning under the
restrictions of the above noise models that explicitly fix the marginal distribution
$\mathcal{P}$ remains an important and challenging open problem. In deriving such bounds, the
present work should be considered a kind of guide, in that we should restrict our focus to
deriving distribution-dependent label complexity bounds with worst-case values that are
never worse than the distribution-free bounds proven here.
|
\section{Efficient computation in $D=1$}
\label{App:onedim}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=8.5cm]{MPSspectral.pdf}
\caption{ Graphical representations of physical quantities that can be efficiently computed in $D=1$. (a) Expectation value of MPO $\Gamma$ in the energy eigenstate $\ket{E_{\vec{\mu}}}$, (b) Second Renyi entropy of the eigenstate $ R_{2}(\rho^{A}_{\vec{\mu}}) \equiv \mbox{tr}\left((\rho_{\vec{\mu}}^A)^2 \right)$, (c) Spectral average of the observable $\sum_{\vec{\mu}} |\Gamma_{\vec{\mu}}|^2$. The cost of manipulating these objects is $O(N)$.
}
\label{Fig:observables1d}
\end{center}
\end{figure}
The tensor network of the projector $ \mathcal{P}_{\vec{\mu}}$ in Fig.~\ref{fig:SpectralTN3} can be used to efficiently compute many physical observables in $D=1$.
Three examples are shown in Fig.~\ref{Fig:observables1d}.
Fig.~\ref{Fig:observables1d}(a) is the graphical representation of the expectation value of a matrix product operator $\Gamma$ in the energy subspace labelled by $\vec{\mu}$, $\textrm{Tr} (\mathcal{P}_{\vec{\mu}} \Gamma)$.
Physically relevant $\Gamma$ include local operators like the magnetization, spatially separated operators, non-local string operators, a sum of local operators like the Hamiltonian, a projector onto a matrix product state etc.
We can also efficiently compute observables that are non-linear in $ \mathcal{P}_{\vec{\mu}}$, such as the second Renyi entropy (Fig.~\ref{Fig:observables1d}(b)).
In addition, we can use the representation in Fig.~\ref{fig:SpectralTN3} to compute an average over the entire energy spectrum by tracing over the $\mu$ indices.
For example, Fig.~\ref{Fig:observables1d}(c) is the graphical representation of the spectral average of the observable $|\bra{\vec{\mu}} \Gamma \ket{\vec{\mu}} |^2$ evaluated in each eigenstate (for simplicity, $N_{\mathcal{Z}}=N$).
Notice that the computation of the average would be inefficient if we had an independent tensor network for each energy eigenstate as there are exponentially many eigenstates; the efficiency here really stems from the spectral tensor network.
In Appendix~\ref{App:PerfectSampling}, we also discuss a different statistical scheme known as perfect sampling to evaluate averages over the entire energy spectrum.
\section{Efficient tensor networks for time evolution and Gibbs ensembles}
\label{App:UZ}
In systems with sIOM\xspace, the time evolution operator and the Gibbs density matrix also have efficient tensor network representations.
Let the Hamiltonian $H$ be a sum of terms of finite range $H=\sum_i h^{[i]}$.
We define the operator $\tilde{h}^{[i]}$ as the diagonal part of $h^{[i]}$ in the energy eigenbasis:
\begin{equation}\label{eq:hi}
\tilde{h}^{[i]} \equiv \sum_{\vec{\mu}} \mathcal{P}_{\vec{\mu}} h^{[i]} \mathcal{P}_{\vec{\mu}}.
\end{equation}
For simplicity, we assume that the number of sIOM\xspace, $N_{\mathcal{Z}}$ is equal to $N$ so that $\mathcal{P}_{\vec{\mu}}$ is a projector onto a unique energy eigenstate.
Now, $h^{[i]}$ and $\mathcal{P}^{(n)}_{\mu_n}$ both have finite support.
When their supports have no overlap:
\begin{align}
\sum_{\mu_n} \mathcal{P}^{(n)}_{\mu_n} h^{[i]} \mathcal{P}^{(n)}_{\mu_n} &= h^{[i]} \sum_{\mu_n} \mathcal{P}^{(n)}_{\mu_n} \mathcal{P}^{(n)}_{\mu_n} \\
&= h^{[i]}
\end{align}
Thus, Eq.~\eqref{eq:hi} can be simplified to:
\begin{align}
\tilde{h}^{[i]} = \sum_{\vec{a}} \mathcal{P}_{\vec{a}} h^{[i]} \mathcal{P}_{\vec{a}}.
\end{align}
where the vector $\vec{a}$ contains the $\mu_n$ of the sIOM\xspace whose support overlaps with the support of $h^{[i]}$.
The $\tilde{h}^{[i]}$ operators have a number of desirable properties such as: 1) they have finite support on the lattice, 2) they commute with one another on different sites, and 3) the Hamiltonian can be expressed as $H= \sum_{j} \tilde{h}^{[j]}$.
The unitary evolution operator for time $t$, $U(t) \equiv e^{-iHt}$, can thus be decomposed into a product of commuting local evolution operators:
\begin{align}
\label{Eq:UDecomp}
U(t) = \prod_{i} U^{[i]},\quad U^{[i]} \equiv e^{-i \tilde{h}^{[i]}t}, \quad[U^{[i]}, U^{[i']}]=0.
\end{align}
Each $U^{[i]}$ can be efficiently expressed as an MPO as the support of $U^{[i]}$ is finite.
Using Eq.~\eqref{Eq:UDecomp}, we can therefore build an efficient tensor network representation for the time evolution operator up to any time by multiplying the MPO's corresponding to all the $U^{[i]}$'s.
Observe that the argument above applies \emph{mutatis mutandis} to the Gibbs ensemble, $e^{-\beta H}$.
Thus, the Gibbs ensemble also has an efficient tensor network representation in systems with sIOM\xspace.
\section{Perfect sampling }
\label{App:PerfectSampling}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=6.5cm]{Perfectsampling.pdf}
\caption{ (a) $Q(\mu_1)$, the marginal distribution defined in Eq.~\eqref{Eq:Qmu1}. (b) The conditional probability $Q(\mu_2|\mu_1)$. The green triangles represent unit vectors of length $q$.}
\label{Fig:perfect}
\end{center}
\end{figure}
To evaluate quantities that are non-polynomial functions of the spectrum like connected correlators in $D=1$, perfect sampling is a useful tool.
Let the system be in a mixed state with weight $P({\vec{\mu}})$ for eigenstate $|E_{\vec{\mu}}\rangle$:
\begin{align}
\rho = \sum_{\vec{\mu}} P(\vec{\mu}) \mathcal{P}_{\vec{\mu}}
\end{align}
Let $P(\vec{\mu})$ have an efficient MPS representation (for example, the Gibbs distribution or the diagonal ensemble of a MPS $\ket{\psi}$).
Perfect sampling is a method to efficiently draw statistically independent configurations $\vec{\mu}$ according to the probability $P(\vec{\mu})$ at a cost $O(N)$.
Thus, we can generate $M$ independent samples at a cost $O(NM)$ and efficiently approximate quantities like the weighted connected correlator of an observable $O$ in $\rho$:
\begin{align*}
W &= \sum_{\vec{\mu}} P(\vec{\mu}) C_{OO}(\vec{\mu}) \approx \frac{1}{M} \sum_{\vec{\mu} \in \Omega} C_{OO}(\vec{\mu}),
\end{align*}
where $\Omega$ denotes a set of $M$ uncorrelated configurations $\vec{\mu}$ drawn according to the probability $P(\vec{\mu})$, and $C_{OO}(\vec{\mu})$ is defined as:
\begin{align*}
C_{OO}(\vec{\mu}) &\equiv \bra{E_{\vec{\mu}}} O(x) O(y) \ket {E_{\vec{\mu}}}_c \\
&= \textrm{Tr} (O(x) O(y) \mathcal{P}_{\vec{\mu}}) - \textrm{Tr} (O(x) \mathcal{P}_{\vec{\mu}}) \textrm{Tr}( O(y) \mathcal{P}_{\vec{\mu}} )
\end{align*}
For each $\vec{\mu}$, we can efficiently compute $C_{OO}(\vec{\mu})$, as $O$ and $\mathcal{P}_{\vec{\mu}}$ have MPO representations.
Perfect sampling generates $\Omega$ (see below), and thus we have an efficient statistical sign problem free scheme to approximate $W$.
The statistical error in the estimate decreases as $1/\sqrt{M}$.
Let us now generate $\Omega$ by adopting the scheme proposed by Ferris \cite{Ferris:2014qa}.
First, define the marginal distribution:
\begin{align}
\label{Eq:Qmu1}
Q(\mu_1,\ldots \mu_i) \equiv \sum_{\mu_{i+1}, \cdots, \mu_N} P(\mu_1,\mu_2,\mu_3,\cdots,\mu_N).
\end{align}
We begin by computing $Q(\mu_1)$ by tracing over $\mu_2 \ldots \mu_N$, as shown in Fig.~\ref{Fig:perfect} (a). We draw a value $\mu_1$ according to this distribution. Fixing the first index to be this value, we then compute the conditional probability of $\mu_2$ given $\mu_1$, $Q(\mu_2|\mu_1)$, by summing over $\mu_3, \ldots \mu_N$ (see Fig.~\ref{Fig:perfect} (b)). We then draw a value $\mu_2$ according to $Q(\mu_2|\mu_1)$. Note that as $Q(\mu_1,\mu_2) = Q(\mu_1)Q(\mu_2|\mu_1)$, we have randomly drawn a pair of value $(\mu_1, \mu_2)$ according to the probability $Q(\mu_1, \mu_2)$. We now repeat: we fix the first two indices to be the values we picked, compute $Q(\mu_3|\mu_1,\mu_2)$ etc. After $N$ such repetitions, we have randomly drawn a vector $\vec{\mu}$ with probability $P(\vec{\mu})$, as intended. By properly recycling partial sums on the MPS, this can be accomplished at a cost $O(N^2q^3\chi)$, where $\chi$ is the bond dimension of $P(\vec{\mu})$ assumed to be greater than $q$.
\section{Bond dimension vs $L$ for a quasi-local operator}
\label{App:mbvsL}
In the spin-$1/2$ system described in the numerical section in the main text, we constructed a quasi-local operator by conjugating $\sigma_4^z$ by a quasi-local unitary, $e^{iH_{XYZ}}$.
In this appendix, we explore the system size dependence of the bond dimension of the MPO representation of $\mathcal{Z}= e^{iH_{XYZ}} \sigma_{4}^z e^{-iH_{XYZ}}$.
Shown in Fig.~\ref{Fig:Ldep} is the bond dimension of the MPO representation of $\mathcal{Z}$ versus bond index for different $L$ at $f=10^{-3}$.
The pink and blue lines are the maximum possible bond dimensions at $L=12,13$.
We see that the bond index is small and completely independent of $L$.
This is numerical evidence for the efficient representation of a quasi-local operator by a MPO in the thermodynamic limit up to any required precision.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=6.5cm]{Applot.pdf}
\caption{Bond dimension of the MPO representation of $e^{iH_{XYZ}} \sigma_{4}^z e^{-iH_{XYZ}}$, $m(b)$ versus the bond index $b$ for truncation parameter $f=10^{-3}$ at different $L$. The maximum value of the bond dimension at $L=12,13$ is also shown for comparison.}
\label{Fig:Ldep}
\end{center}
\end{figure}
|
\section{Introduction}
Historically, fractional calculus emerged nearly in the same time as classical calculus as a natural extension of classic calculus. However, its application for problems like fractional partial differential equations (FPDEs) are less mature than that associated with classic calculus. Only during the last few decade has fractional calculus seen a broader application as a tool to describe a wide range of non-classical phenomena in the applied science and engineering, for example, the fractional Fokker-Planck equations for anomalous diffusion problems, continuous time random walk models with power law waiting time and/or jump length distribution (\cite{barkai, metzler}), and the subdiffusion and superdiffusion process.
With the increasing utilization of fractional calculus it is necessary to develop appropriate and robust numerical methods to solve FPDEs for practical application. A fundamental difference between problems in classic calculus and fractional calculus lies in the non-local nature of the factional operators and the dependance of history, causing the essential difficulties and challenges for numerical approximation.
In recent years, however, successful work has emerged to deal with discretizing fractional models by adapting traditional numerical schemes, including finite difference methods, finite element methods, and spectral methods. Meerschaert and Tadjeran in [12] firstly proposed a stable difference method -- the shifted Gr\"{u}nwald-Letnikov formula -- to approximate fractional advection-dispersion flow equations. Recently, Tian et al. \cite{tian12} put forward a higher order accurate numerical solution
method for the space fractional diffusion equation, referred to as the weighted and shifted Gr\"unwald difference operators. In \cite{xu1}, Li and Xu considered a space-time spectral method for solving the time fractional diffusion equation and Deng \cite{deng1} developed finite element methods for discretizing the space and time fractional Fokker-Planck equation. More recently, Xu and Hesthaven \cite{xu14} discussed stable multi-domain spectral penalty methods for FPDEs and Zayernouri and Karniadakis \cite{Zayernouri} analyzed fractional Sturm-Liouville eigen-problems.
As an alternative, discontinuous Galerkin methods have also emerged. In 2010, Deng and Hesthaven \cite{deng2} proposed a local discontinuous Galerkin method for the fractional diffusion equation, and offered stability analysis and error estimates, confirming that the schemes should exhibit optional order of convergence for the superdiffusion case. Almost in the same time, Ji and Tang \cite{xia} presented a purely qualitative study of the solution of spatial Caputo fractional problems in one and two dimensions using a high-order Runge-kutta discontinuous Galerkin methods, but did not offer theoretical results.
The main advantages of DG methods include geometric flexibility and the support of locally adapted resolution as well as excellent parallel efficiency. In \cite{xia}, the authors adopted the rectangular meshes to deal with the two-dimensional cases. This paper, as a successor to previous work \cite{deng2}, discusses how to approximate fractional diffusion equations with genuinely unstructured grids beyond one dimension and offer a theoretical analysis for the high dimensional case. To
focus on how to overcome the difficulties of developing LDG for fractional problems with unstructured meshes, we consider left Riemann-Liouville fractional equations and just the numerical experiments are provided for the right Riemann-Liouville fractional equations; in fact, the discussions for the right (or both left and right) Riemann-Liouville fractional equations are similar to the left ones.
The paper is organized as follows. In Section 2, we review the needed definitions of fractional operators and the fractional functional setting. In the following section, we propose our numerical schemes and show the detailed algorithm in computation. Section 4 gives the corresponding stability analysis and error estimates. The numerical results are provided in Section 5 and a few concluding remarks are given in the last section.
\section{Preliminaries}
In this section, we introduce some preliminary definitions of fractional derivatives and associated functional setting for
the subsequent numerical schemes and theoretical analysis.
First we recall some definitions of the fractional derivatives and integrals listed as follows:
\begin{itemize}
\item left Riemann-Liouville fractional derivative:
$$ _aD_x^{\alpha}u(x) = \frac{1}{\Gamma(n-\alpha)} \frac{d^n}{dx^n} \int_{a}^{x} (x- \xi)^{n - \alpha -1}u(\xi) d\xi$$
\item right Riemann-Liouville fractional derivative:
$$ _xD_b^{\alpha}u(x) = \frac{(-1)^{n}}{\Gamma(n-\alpha)} \frac{d^n}{dx^n} \int_{x}^{b} (\xi - n)^{n - \alpha -1}u(\xi) d\xi$$
\item left fractional integral:
$$ _aD_x^{-\alpha}u(x) = \frac{1}{\Gamma(\alpha)} \int_{a}^{x} {(x-\xi)^{\alpha-1}}{u(\xi)} d\xi$$
\item right fractional integral:
$$ _xD_b^{-\alpha}u(x) = \frac{1}{\Gamma(\alpha)} \int_{x}^{b} {(\xi-x)^{\alpha-1}}{u(\xi)} d\xi,$$
\end{itemize}
where $\Gamma(\cdot)$ denotes the Gamma function, $\alpha \in [n-1,n)$, $x\in(a,b)$, a and b can be $-\infty$ and $+\infty$ respectively.
By simple linear transformations, the fractional integrals have the equivalent forms, i.e.,
\begin{equation}
\begin{split}
_aD_x^{-\alpha} u(x) = \frac{1}{\Gamma(\alpha)} \left( \frac{x-a}{2} \right)^{\alpha} \int_{-1}^{1} (1-\eta)^{\alpha-1}u\left(\frac{x+a}{2} + \frac{x-a}{2}\eta\right) d\eta,\\
_xD_b^{-\alpha} u(x) = \frac{1}{\Gamma(\alpha)} \left( \frac{b-x}{2} \right)^{\alpha} \int_{-1}^{1} (1+\eta)^{\alpha-1}u\left(\frac{b+x}{2} + \frac{b-x}{2}\eta\right) d\eta.
\end{split}
\label{eq2:2}
\end{equation}
On the basis of (\ref{eq2:2}), we use the Gauss-Jacobi quadrature with weight functions $(1-\eta)^{\alpha-1}$ and $(1+\eta)^{\alpha-1}$ to solve the weakly singular fractional integrals in numerical computation. For the following Lemmas \ref{lemma2.1}, \ref{lem2.3}, Definitions \ref{defn2.2}, \ref{defn2.4}, and Theorem \ref{thm2.5}, we refer to \cite{Ervin:2006, deng1, deng2}.
\begin{lem}
\label{lemma2.1}
The left and right fractional integral operators are adjoint in the sense of the $L^2(a,b)$ inner product, i.e.,
\begin{equation}
(_aD_x^{-\alpha}u,v)_{L^2(a,b)} = (u,\, _{x}D_b^{-\alpha}v)_{L^2(a,b)} \quad \forall \alpha>0, a<b.
\end{equation}
\end{lem}
In order to carry out the analysis, we need to introduce the fractional integral space here.
\begin{defn}\label{defn2.2}
Let $\alpha >0$. Define the norm
\begin{equation}
\norm{u}_{H^{-\alpha}(\mathcal{R})} := \norm{ |\omega|^{-\alpha} \widehat{u}}_{L^2(\mathcal{R})},
\end{equation}
where $\widehat{u}(\omega)$ is the Fourier transform of $u(x)$. Let $H^{-\alpha}(\mathcal{R})$ denote the closure of $\cinf{\mathcal{R}}$ with respect to $\norm{\cdot}_{H^{-\alpha}(\mathcal{R})}$.
\end{defn}
\begin{lem} \label{lem2.3}
\begin{equation}
(_{-\infty}D_x^{-\alpha}u,\smallskip _xD_{\infty}^{-\alpha} u) = cos(\alpha \pi)\norm{_{-\infty}D_x^{-\alpha}u}^2_{L^2(\mathcal{R})} = cos(\alpha \pi) \norm{u}^2_{H^{-\alpha}(\mathcal{R})}.
\end{equation}
\end{lem}
\noindent Let us now restrict attention to the case in which $ supp(v) \subset \Omega = (a,b)$. Then $_{-\infty}D_x^{-\alpha }u = {_a}D_x^{-\alpha}$,
and $_{x}D_{\infty}^{-\alpha} = {_x}D_b^{-\alpha}$. Straightforward extension of the above yields.
\begin{defn} \label{defn2.4}
Define the space $H_0^{-\alpha}(\Omega)$ as the closure of $\cinf{\Omega}$ with respect to $\norm{\cdot}_{H^{-\alpha}(\mathcal{R})}$.
\end{defn}
The following theorem gives the inclusion relation between the fractional integral spaces with different $\alpha$.
\begin{thm}\label{thm2.5
If $-\alpha_2 < -\alpha_1 < 0$, then $H_0^{-\alpha_1}(\Omega)$ is embedded into
$H_0^{-\alpha_2}(\Omega)$ and $L^2(\Omega)$ is embedded into both of them.
\end{thm}
\subsection{Notations for DG methods}
We consider problems posed on the physical domain $\Omega$ with boundary $\partial\Omega$ and assume that this domain is well approximated by
the computational domain $\Omega_h$. Generally, we denote $\Omega_h$ as $\Omega$ when no misunderstanding is possible. This is a space filling triangulation composed of a collection of K geometry-conforming nonoverlapping elements, $D^k$, i.e.,
$$\Omega \simeq \Omega_h = \bigcup_{k=1}^{K}D^k,$$
and $\Gamma$ denotes the union of the boundaries of the elements $D^k$ of $\Omega_h$.
$\Gamma$ consists of two parts: the set of unique purely internal edges $\Gamma_i$ and the set of external edges $\Gamma_b=\partial \Omega$ of domain boundaries, and
$\Gamma = \Gamma_i\bigcup\Gamma_b$.
The shape of these elements can be arbitrary although we will mostly consider cases where they are d-dimensional curvilinear simplices. In the two-dimensional case, the planar triangles are adopted.
Now we introduce the broken Sobolev space for any real number $s$,
$$
H^s(\Omega_h) = \{v\in L^2(\Omega): \forall k=1,2,\ldots, K, v|_{D^k} \in H^s(D^k)\},
$$
equipped with the broken Sobolev norm:
$$
\norm{v}_{H(\Omega_h)} = \big(\sum_{k=1}^{K}\norm{v}^2_{H^s(D^k)}\big)^{1/2}.
$$
Here, we retain the same style with the traditional Sobolev space when no misunderstanding is possible. When $s=0$, $H^0(\Omega_h) = L^2(\Omega_h).$
We introduce the approximation spaces $V_h$. Before that, we first define the space of N-th order polynomials in two variables on the 2-simplex $D$, $P_N^2(D)$ such that the dimension of the approximation polynomial space is
$$ \dim P_N^2(D) = N_p = \binom{N+2}{N},
$$
being the minimum space in which $P_N^2(D)$ may be complete. Let us introduce the nodal set $\{{\textbf{\emph{x}}}_i\}_{i=1}^{N_p}$, which are the nodal points or collocation points on $D$. These points must be chosen carefully to ensure well conditioned operators. The polynomial space $P_N^2(D)$ can be illustrated as
\begin{eqnarray}
P_N^2(D) &=& \text{span} \{\ell_i(\textbf{\emph{x}}), i=1,2,\ldots, N_p \} \nonumber\\
&=& \text{span} \{x^iy^j \ |\ (i,j)\geqslant0, i+j \leqslant N, (x,y) \in D^k \}, \nonumber
\end{eqnarray}
where $\ell_i(\textbf{\emph{x}})$ denotes the two-dimensional multivariate Lagrange interpolation basis function. These two forms are equivalent since $P_N^2(D)$ is a finite dimensional polynomial space. For the explicit way to evaluate the genuinely two-dimensional Lagrange polynomials, we refer to \cite{jan1} and \cite{jan2}.
We also need to define a number of different inner products on the simplex, $D$. Consider the two continuous functions $f,g$, then the inner product, the associated $L^2$ norm and the inner product over the surface of D are defined as
$$
(f,g)_D = \int_D f(\textbf{\emph{x}})g(\textbf{\emph{x}}) d\textbf{\emph{x}}, \quad (f,f) = ||f||_D^2, \quad (f,g)_{\partial D} = \int_{\partial D} f(\textbf{\emph{x}}) g(\textbf{\emph{x}}) d s .
$$
The corresponding global broken measures, inner products and norms are
$$
(f,g)_{\Omega} = \sum_{k=1}^{K} (f,g)_{D^k}, \quad (f,f) = \sum_{k=1}^{K} ||f||_{D^k} = ||f||_{\Omega}, \quad (f,g)_{\Gamma} = \sum_{k=1}^{K} (f,g)_{\partial D^k}.
$$
Next, we introduce some notations to manipulate numerical fluxes. For $e\in \Gamma$, we refer to the exterior information by a superscript `+' and to the interior information by a superscript `-'.
Using these notations, it is useful to define the average
$$\{u\} = \frac{u^++u^-}{2} \quad \text{on} \ e\in \Gamma_i,$$
$$ \{u\} = u \quad \text{on} \ e \in \Gamma_b, $$
where $u$ can be both a scalar and a vector.
In a similar fashion, we also define the jumps along a unit normal $\textbf{n}$, as
$$[u]=\textbf{n}^+u^+ + \textbf{n}^-u^-, \ [\textbf{u}]=\textbf{n}^+\cdot\textbf{u}^+ + \textbf{n}^-\cdot\textbf{u}^- \quad \text{on} \ e\in \Gamma_i,$$
$$[u]=\textbf{n} u, \ [\textbf{u}]=\textbf{n} \cdot\textbf{u} \quad \text{on} \ e\in \Gamma_b.$$
Assume that the global solution can be approximated as
\begin{equation}
u(\textbf{\emph{x}},t) \simeq u_h(\textbf{\emph{x}},t) = \bigoplus_{k=1}^{K} u_h^{k}(\textbf{\emph{x}},t) \in V_h = \bigoplus_{k=1}^{K} P_N^2(D^k),
\end{equation}
where $ P_N^2(D^k)$ is the space of N-th order polynomials defined on $D^k$.
The local solution, $u(\textbf{\emph{x}},t)$ can be expressed by
\begin{equation}
u_h^k(\textbf{\emph{x}},t)= \sum_{i=1}^{N_p}u_h^k(\textbf{\emph{x}}_i,t) \ell_i^k(\textbf{\emph{x}}), \quad \textbf{\emph{x}} \in D^k,
\end{equation}
utilizing a nodal representation.
\section{The local nodal discontinuous Galerkin methods for fractional diffusion equations}
We consider two dimensional fractional problems
\begin{equation}
\frac{\partial u(\textbf{\emph{x}},t)}{\partial t} = d_1 \frac{\partial^{\alpha}u(\textbf{\emph{x}},t)}{\partial x^{\alpha}} + d_2 \frac{\partial^{\beta} u(\textbf{\emph{x}},t)}{\partial y^{\beta}} + f(\textbf{\emph{x}},t), \ \textbf{\emph{x}} = (x,y) \in \mathcal{R}^2,
\end{equation}
subject to appropriate boundary and initial conditions. Here, $\alpha, \beta \in (1,2]$, $d_1, d_2 > 0$, $f(\textbf{\emph{x}},t)$ is a source term, and $\frac{\partial^{\alpha}}{\partial x^{\alpha}}, \frac{\partial^{\beta}}{\partial y^{\beta}}$ denote the left Riemann-Liouville fractional derivatives.
For convenience and to enable the theoretical analysis, we restrict our problem to a homogeneous Dirichlet boundary condition on the form
\begin{equation}
\left\{ \begin{array}{ll}
\frac{\partial u(\textbf{\emph{x}},t)}{\partial t}
= d_1\frac{\partial}{\partial x}{_a}D_x^{\alpha-2}\frac{\partial}{\partial x}u(\textbf{\emph{x}},t) \\
\qquad \qquad + d_2\frac{\partial}{\partial y}{_c}D_y^{\beta-2}\frac{\partial}{\partial y}u(\textbf{\emph{x}},t) + f(\textbf{\emph{x}},t) & \textrm{$(\textbf{\emph{x}},t) \in \Omega \times[0,T] $},\\
u(\textbf{\emph{x}},0) = u_0(\textbf{\emph{x}}) & \textrm{$\textbf{\emph{x}} \in \Omega$},\\
u(\textbf{\emph{x}},t) = 0 & \textrm{$(\textbf{\emph{x}},t) \in \partial \Omega\times [0,T]$,}
\end{array} \right.
\label{eq3:2}
\end{equation}
where $\Omega = (a,b)\times(c,d)$. For the simplicity of theoretical analysis, we set $d_1 = d_2 = 1$ without the loss of generality.
\subsection{The primal formulation}
Following the standard approach for the development of local discontinuous Galerkin methods for problems with higher order derivatives, we introduce the auxiliary variables $\textbf{\emph{p}} = (p^x,p^y)$ and $\textbf{\emph{q}}= (q^x,q^y)$, and express the problem as
\begin{equation} \label{weakform}
\left\{ \begin{array}{ll}
\frac{\partial u(\textbf{\emph{x}},t)}{\partial t} = \nabla \cdot \textbf{\emph{q}} + f(\textbf{\emph{x}},t) &(\textbf{\emph{x}},t)\in \Omega \times [0,T],\\
\textbf{\emph{q}} = (\frac{\partial^{\alpha-2}p^x}{\partial x^{\alpha-2}}, \frac{\partial^{\beta-2}p^y}{\partial y^{\beta-2}}) &(\textbf{\emph{x}},t) \in \Omega\times[0,T],\\
\textbf{\emph{p}} = \nabla u &(\textbf{\emph{x}},t)\in \Omega \times [0,T],\\
u(\textbf{\emph{x}},0) = u_0(\textbf{\emph{x}}) &\textbf{\emph{x}}\in \Omega ,\\
u(\textbf{\emph{x}},t) = 0 &(\textbf{\emph{x}},t)\in \partial \Omega \times [0,T].\\
\end{array}\right.
\end{equation}
We set $h_k := \text{diam}(D^k)$ and $h := \text{max}_{k=1}^{K}h_k$. Since the fractional integral spaces are embedded in $L^2(\Omega)$, we could assume that the exact solution $(u,\textbf{\emph{p}},\textbf{\emph{q}})$ of (\ref{weakform}) belongs to
\begin{equation}
H^1(0,T;H^1(\Omega_h)) \times (L^2(0,T;L^2(\Omega_h)))^2 \times (L^2(0,T;H^1(\Omega_h)))^2.
\end{equation}
Next, we require that $(u,\textbf{\emph{p}},\textbf{\emph{q}})$ satisfies the local formulation
\begin{eqnarray}
\left(\frac{\partial u(\textbf{\emph{x}},t)}{\partial t},v\right)_{D^k} &=& \left(\textbf{n}\cdot \textbf{\emph{q}}, v\right)_{\partial D^k} - (\textbf{\emph{q}},\nabla v)_{D^k} + (f,v)_{D^k},\\
(\textbf{\emph{q}}, \pmb{\phi})_{D^k} &=& \left(\left(\frac{\partial^{\alpha-2}p^x}{\partial x^{\alpha-2}},\frac{\partial^{\beta-2}p^y}{\partial y^{\beta-2}}\right), \pmb{\phi}\right)_{D^k},\\
(\textbf{\emph{p}}, \pmb{\pi})_{D^k} &=& (u,\textbf{n}\cdot\pmb{\pi})_{\partial D^k} - (u, \nabla \cdot \pmb{\pi})_{D^k},\\
(u(\cdot,0), v)_{D^k} &=& (u_0(\cdot),v)_{D^k},
\end{eqnarray}
for all test functions $v \in H^1(\Omega_h)$, $\pmb{\phi} = (\phi^x, \phi^y) \in (L^2(\Omega_h))^2 = L^2(\Omega_h) \times L^2(\Omega_h)$, and $\pmb{\pi} =(\pi^x,\pi^y) \in (H^1(\Omega_h))^2 = H^1(\Omega_h) \times H^1(\Omega_h)$.
To complete the primal formulation of our numerical schemes, we need to introduce the finite dimensional subspace of $H^1(\Omega_h)$, i.e.,
$$V_h = \{v:\Omega_h \rightarrow \mathbb{R}\big|\ v|_{D^k} \in P_N^2(D^k), \ k=1,2,\cdots, K \},$$
and then restrict the trial and test functions $v$ to $V_h$,
$\pmb{\phi}, \pmb{\pi}$ to $ (V_h)^2 = V_h \times V_h$ respectively. Furthermore we define $u_h, \textbf{\emph{p}}_h, \textbf{\emph{q}}_h$ as the approximation of $u,\textbf{\emph{p}},\textbf{\emph{q}}$.
We then seek $(u_h,\textbf{\emph{p}}_h,\textbf{\emph{q}}_h) \in H^1(0,T;V_h)\times (L^2(0,T;V_h))^2 \times (L^2(0,T;V_h))^2$ such that for all $v \in V_h, \ \pmb{\phi}, \pmb{\pi} \in (V_h)^2 $ the following holds:
\begin{eqnarray}
\left(\frac{\partial u_h(\textbf{\emph{x}},t)}{\partial t},v\right)_{D^k} &=& (\textbf{n}\cdot \flux{\textbf{\emph{q}}}_h, v)_{\partial D^k} - (\textbf{\emph{q}}_h,\nabla v)_{D^k} + (f,v)_{D^k}, \label{eq3:9}\\
(\textbf{\emph{q}}_h, \pmb{\phi})_{D^k} &=& \left(\left(\frac{\partial^{\alpha-2}p_h^x}{\partial x^{\alpha-2}},\frac{\partial^{\beta-2}p_h^y}{\partial y^{\beta-2}}\right), \pmb{\phi}\right)_{D^k},\label{eq3:10}\\
(\textbf{\emph{p}}_h, \pmb{\pi})_{D^k} &=& (\flux{u}_h,\textbf{n}\cdot\pmb{\pi})_{\partial D^k} - (u_h, \nabla \cdot \pmb{\pi})_{D^k}. \label{eq3:11}
\end{eqnarray}
The form (\ref{eq3:9})-(\ref{eq3:11}), obtained after integration by parts once, is known as the weak form, which is used for theoretical analysis in the following. For the computational part,
we introduce the strong form, recovered by doing integration by parts once again partially, as
\begin{eqnarray}
\left(\frac{\partial u_h(\textbf{\emph{x}},t)}{\partial t},v\right)_{D^k} &=& (\nabla \cdot \textbf{\emph{q}}_h, v)_{D^k} - (\textbf{n}\cdot (\textbf{\emph{q}}_h - \flux{\textbf{\emph{q}}}_h), v)_{\partial D^k} ,\label{eq3:12} \\
(\textbf{\emph{q}}_h, \pmb{\phi})_{D^k} &=& \left(\left(\frac{\partial^{\alpha-2}p_h^x}{\partial x^{\alpha-2}},\frac{\partial^{\beta-2}p_h^y}{\partial y^{\beta-2}}\right), \pmb{\phi}\right)_{D^k},\label{eq3:13}\\
(\textbf{\emph{p}}_h, \pmb{\pi})_{D^k} &=& (\nabla u_h, \pmb{\pi})_{D^k} - (u_h - \flux{u}_h,\textbf{n}\cdot\pmb{\pi})_{\partial D^k}. \label{eq3:14}
\end{eqnarray}
The two formulations are mathematically equivalent but computationally different \cite{jan2}. The boundary condition $u|_{\Gamma_b}=0$ will be imposed on the 3rd equation above. To guarantee consistency, stability and optimal order of convergence of the formulation above, we must define the numerical flux $\flux{u}_h, \flux{\textbf{\emph{q}}}_h$ carefully.
In the two-dimensional case, we adopt the central flux, defined as
\begin{equation}
\flux{u}_h = \frac{u^+_h + u^-_h}{2}, \quad \flux{\textbf{\emph{q}}}_h =\frac{\textbf{\emph{q}}^+_h + \textbf{\emph{q}}^-_h}{2},
\end{equation}
at all internal edges, and at the external edges we use
\begin{equation}
\flux{u}_h = 0, \quad \flux{\textbf{\emph{q}}}_h = \textbf{\emph{q}}_h^+ = \textbf{\emph{q}}_h^-.
\end{equation}
\subsection{The semidiscrete scheme}
We will borrow the notations for various operators from \cite{jan1} to express the semidiscrete scheme.
Let us introduce some local and global vector and matrix notations to form the local statement
\begin{displaymath}
\begin{array}{ll}
\mathbf{u}_h^k = [u_1^k, u_2^k, \cdots, u^k_{N_p}]^T; & \bm{u}_h = [\bm{u}_h^1; \bm{u}_h^2; \cdots, \bm{u}_h^K]^T, \\
(\mathbf{p}^k)^x_h = [(p^k)^x_1, (p^k)^x_2, \cdots, (p^k)^x_{N_p}]^T; & \textbf{\emph{p}}_h^x = [(\textbf{\emph{p}}^x)_h^1; (\textbf{\emph{p}}^x)_h^2; \cdots; (\textbf{\emph{p}}^x)_h^K]^T, \\
(\mathbf{p}^k)^y_h = [(p^k)^y_1, (p^k)^y_2, \cdots, (p^k)^y_{N_p}]^T; & \textbf{\emph{p}}_h^y = [(\textbf{\emph{p}}^y)_h^1; (\textbf{\emph{p}}^y)_h^2; \cdots; (\textbf{\emph{p}}^y)_h^K]^T, \\
(\mathbf{q}^k)^x_h = [(q^k)^x_1, (q^k)^x_2, \cdots, (q^k)^x_{N_p}]^T; & \textbf{\emph{q}}_h^x = [(\textbf{\emph{q}}^x)_h^1; (\textbf{\emph{q}}^x)_h^2; \cdots; (\textbf{\emph{q}}^x)_h^K]^T, \\
(\mathbf{q}^k)^y_h = [(q^k)^y_1, (q^k)^y_2, \cdots, (p^k)^y_{N_p}]^T; & \textbf{\emph{q}}_h^y = [(\textbf{\emph{q}}^y)_h^1; (\textbf{\emph{q}}^y)_h^2; \cdots; (\textbf{\emph{q}}^y)_h^K]^T.
\end{array}
\end{displaymath}
We also have the local mass matrix $M^k$ with
$$ M_{ij}^k = (\ell_i^k(\textbf{\emph{x}}),\ell_j^k(\textbf{\emph{x}}) )_{D^k}$$
and the local spatial stiff matrix $S_x^k, S_y^k$ with the entries
$$
(S_x^k)_{ij} = \left( \frac{\partial \ell_j (\textbf{\emph{x}}) }{\partial x}, \ell_i(\textbf{\emph{x}}) \right)_{D^k}, \quad
(S_y^k)_{ij} = \left( \frac{\partial \ell_j (\textbf{\emph{x}}) }{\partial y}, \ell_i(\textbf{\emph{x}}) \right)_{D^k}.
$$
It is a little complex to compute the fractional spatial stiff matrices of (\ref{eq3:13}) and they need to
be stored globally. However, their elements depend on the affected regions only and they are sparse.
Next, we will discuss how to form the global fractional spatial stiffness matrix in detail.
Denote $\widetilde{S}_x$ and $\widetilde{S}_y$ as the global fractional spatial stiffness matrices in the $x$ and $y$ direction, respectively.
$\widetilde{S}_x$ and $\widetilde{S}_y$ are $K \times K$ block matrices and every element is a $N_p \times N_p$ matrix.
Here, we use numerical quadrature on a triangle to compute these factional stiffness matrices.
We only describe the specific procedure of forming the global fractional spatial stiff matrix $\widetilde{S}_x$ for simplicity.
$\widetilde{S}_y$ is obtained in the same fashion.
The integral of a function $u$ defined on the physical element $D$ can be approximated by using the Gauss quadrature rule on triangle, i.e.,
$$ \int_{D} u(\textbf{\emph{x}}) d\textbf{\emph{x}} = J\int_{I} u(\textbf{\emph{x}}(\bm{r})) d\bm{r} \simeq J\sum_{j=1}^{Q_D}u(\textbf{\emph{x}}(\bm{r}_j))w_j = J\sum_{j=1}^{Q_D}u(\textbf{\emph{x}} _j )w_j,$$
where $Q_D$ is the total number of Gauss quadrature weights or points and $J$ is the constant transformation Jacobian between the physical element and the reference element $I$.
The sets $\{ \bm{r}_j \}_{j=1}^{Q_D} \in I$ and $\{ \textbf{\emph{x}}_j \}_{j=1}^{Q_D} \in D$ are one-to-one correspondence by an affine map.
Thus, we have
\begin{equation}
\begin{split}
& (_aD_x^{\alpha-2}p_h^x,\ell^k(\textbf{\emph{x}}))_{D^k} \\ &\simeq J^k \sum_{j=1}^{Q_D} {_a}D_x^{\alpha-2}p_h^x \ell^k(\textbf{\emph{x}}_j)w_j \\
& = J^k \sum_{j=1}^{Q_D}\frac{1}{\Gamma(2-\alpha)}\bigg(\sum_{m\in A}\int_{x_{m-1}}^{x_m}(x_j-\xi)^{1-\alpha}
(p_h^x)^m(\xi,y_j)dx \\
&~~ ~ + \int_{x_{k-1}}^{x_j}(x_j-\xi)^{1-\alpha}(p_h^x)^k(\xi,y_j)dx\bigg) \ell^k(\textbf{\emph{x}}_j)w_j,
\end{split}
\end{equation}
where $A$ is the set of the indices of elements affected by each quadrature points on every triangle as sketched in Figure \ref{fig1}.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{Maxwell05.eps}
\caption{All triangles (the ones intersected with the line $AB$) in $x$ direction affected by one Gauss quadrature point P on triangle $D^k$: the ones intersected with $AP$ is for the left fractional derivative/integral and the ones intersected with $PB$ is for the right fractional derivative/integral.}
\label{fig1}
\end{figure}
By using (\ref{eq2:2}) and the definition of the extension of the basis functions, we rewrite the above equation
\begin{equation}
\begin{split}
& (_aD_x^{\alpha-2}p_h^x,\ell^k(\textbf{\emph{x}}))_{D^k} \\
& \simeq \frac{J^k}{\Gamma(2-\alpha)} \sum_{j=1}^{Q_D} \bigg[\sum_{m\in A}\bigg(\left(\frac{x_j-x_{m-1}}{2}\right)^{2-\alpha}\int_{-1}^{1}(1-\eta)^{1-\alpha}\\
&~~~ \cdot(p_h^x)^m\left(\frac{x_j+x_{m-1}}{2} + \frac{x_j-x_{m-1}}{2}\eta,y_j\right)d\eta \\
&~~~ - \left(\frac{x_j-x_{m}}{2}\right)^{2-\alpha}\int_{-1}^{1}(1-\eta)^{1-\alpha} \\
&~~~ \cdot(p_h^x)^m\left(\frac{x_j+x_{m}}{2} + \frac{x_j-x_{m}}{2}\eta,y_j\right)d\eta \bigg)\\
&~~~ + \left(\frac{x_j-x_{k}}{2}\right)^{2-\alpha}\int_{-1}^{1}(1-\eta)^{1-\alpha} \\
&~~~ \cdot(p_h^x)^k\left(\frac{x_j+x_{k}}{2} + \frac{x_j-x_{k}}{2}\eta,y_j\right)d\eta \bigg]\ \ell^k(\textbf{\emph{x}}_j)w_j.
\end{split}
\end{equation}
Gauss quadratures with weight function $(1-\eta)^{1-\alpha}$ will be taken in the numerical process.
\begin{rem}
In the process of computation, we must pay attention to the fact that there is no link between the Lagrange interpolation points and Gauss quadrature points on the triangle.
\end{rem}
A piece of pseudocode illustrates the formation of the left global fractional spatial stiffness matrix.
\begin{algorithm}
\caption{Construction of the fractional global spatial stiffness matrix $_l\widetilde{S}_x$
\begin{algorithmic}[1]
\State \% denote $\pmb{\ell}^k = [\ell_1^k, \ell_2^k, \cdots, \ell_{Np}^k]^T$
\State Initialize every block of $_l\widetilde{S}_x$ with zero matri
\For{k = 1:K}
\For{j = 1:$Q_D$}
\State \% $Q_D$ is the total number of Gauss quadrature points
\State Find the set A and lXdir of every Gauss point $(x_j^k,y_j^k)$ on triangle $D^k$
\State \% lXdir is a length(A)$\times$2 matrix to store the intervals across by $PA$ on each triangle
\State $(_l\widetilde{S}_x) _{kk} = (_l\widetilde{S}_x) _{kk} + (_{x_{k-1}}D_x^{\alpha-2}\pmb{\ell}^k(\textbf{\emph{x}}),\pmb{\ell}^k(\textbf{\emph{x}}))_{D^k}$
\For{m=1:length(A)}
\State $(_l\widetilde{S}_x) _{km} = (_l\widetilde{S}_x) _{km} + (\frac{1}{\Gamma(2-\alpha)}\int_{x_{m-1}}^{x_m}(x-\xi)^{1-\alpha}\pmb{\ell}^m(\xi,y)d\xi,\pmb{\ell}^k(\textbf{\emph{x}}))_{D^k}$
\EndFor
\EndFor
\EndFor
\end{algorithmic}\label{alg1}
\end{algorithm}
\begin{rem}
It is unnecessary to store the full global fractional spatial stiffness matrices $_l\widetilde{S}_x$ and
$_l\widetilde{S}_y$ as they are both sparse.
\end{rem}
We recover the global semi-discrete form (\ref{eq3:12})-(\ref{eq3:14}) as
\begin{eqnarray}
M \frac{\partial \bm{u}_h}{\partial t} &=& S_x \textbf{\emph{q}}_h^x + S_y \textbf{\emph{q}}_h^y - \bigcup_{k=1}^K \int_{\partial D^k}\textbf{n} \cdot ((\textbf{\emph{q}}_h^x, \textbf{\emph{q}}_h^y) - \flux{\textbf{\emph{q}}}_h) \ell(\textbf{\emph{x}}) d\bm{s},\\
M \textbf{\emph{q}}_h^x &=& _l\widetilde{S}_x \textbf{\emph{p}}_h^x, \quad M \textbf{\emph{q}}_h^y = {_l}\widetilde{S}_y \textbf{\emph{p}}_h^y, \\
M \textbf{\emph{p}}_h^x &=& S_x \bm{u}_h - \bigcup_{k=1}^K \int_{\partial D^k} n_x(\bm{u}_h - \flux{\bm{u}}_h) \ell(\textbf{\emph{x}}) d\bm{s}, \\
M \textbf{\emph{p}}_h^y &=& S_y \bm{u}_h - \bigcup_{k=1}^K \int_{\partial D^k} n_y(\bm{u}_h - \flux{\bm{u}}_h) \ell(\textbf{\emph{x}}) d\bm{s},
\end{eqnarray}
where $M, S_x, S_y$ are global mass and stiffness matrices and there non-zero diagonal block are constructed by $M^k, S_x^k, S_y^k$ respectively, and $\bigcup_{k=1}^K$
refers to the construction of the equations simultaneously for numerical computation.
\section{Stability analysis and error estimates}
Before initiating the theoretical analysis, we need following result:
\begin{lem}
Assume that $\Omega$ has been triangulated into $K$ elements, $D^k$. Then
\begin{equation}
\sum_{k=1}^{K} (\textbf{n}\cdot\textbf{\emph{u}}, v)_{\partial D^k} = \oint_{\Gamma}\{\textbf{\emph{u}}\}\cdot [v] ds + \oint_{\Gamma_i}\{v\}[\textbf{\emph{u}}]ds.
\end{equation}
\end{lem}
The proof follows directly by rewriting and then summing the averages and jumps of all the terms along one edge.
By using this Lemma and summing all the terms of (\ref{eq3:9})-(\ref{eq3:11}), we obtain the primal formulation:
Find $(u_h,\textbf{\emph{p}}_h,\textbf{\emph{q}}_h) \in H^1(0,T;V_h)\times (L^2(0,T;V_h))^2\times (L^2(0,T;V_h))^2$ such that
for all $(v,\pmb{\phi},\pmb{\pi}) \in H^1(0,T;V_h)\times (L^2(0,T;V_h))^2\times (L^2(0,T;V_h))^2$ the following holds
\begin{equation}
B(u_h,\textbf{\emph{p}}_h,\textbf{\emph{q}}_h;v,\pmb{\phi},\pmb{\pi}) = \mathcal{L}(v,\pmb{\phi},\pmb{\pi}).
\end{equation}
We denote $(\cdot,\cdot) = (\cdot,\cdot)_{\Omega}$ when no misunderstanding is possible. Recall that the numerical flux is single valued and the homogeneous boundary condition is assumed. Then the discrete bilinear form $B$ can be defined as
\begin{equation}
\begin{split}
& B(u_h,\textbf{\emph{p}}_h,\textbf{\emph{q}}_h;v,\pmb{\phi},\pmb{\pi}) \\ & := \int_{0}^{T}\left(\frac{\partial u_h}{\partial t},v\right) dt + \int_{0}^{T}(\textbf{\emph{q}}_h,\nabla v)dt\\
& ~~~ + \int_{0}^{T} (\textbf{\emph{q}}_h,\pmb{\phi})dt - \int_{0}^{T}\left(\left(\frac{\partial^{\alpha-2}p_h^x}{\partial x^{\alpha-2}},\frac{\partial^{\beta-2}p_h^y}{\partial y^{\beta-2}}\right),\pmb{\phi} \right)dt \\
& ~~~ + \int_{0}^{T} (\textbf{\emph{p}}_h,\pmb{\pi})dt + \int_{0}^{T}(u_h,\nabla\cdot\pmb{\pi})dt\\
& ~~~ - \int_{0}^{T} \left( \oint_{\Gamma}\flux{\textbf{\emph{q}}}_h\cdot[v]ds + \oint_{\Gamma_i}\flux{u}_h[\pmb{\pi}] ds \right)dt.
\end{split}
\label{eq4:3}
\end{equation}
The discrete linear form $\mathcal{L}$ is given by
\begin{equation}
\mathcal{L}(u_h,\textbf{\emph{p}}_h,\textbf{\emph{q}}_h;v,\pmb{\phi},\pmb{\pi}) = \int_{0}^{T}(f,v)dt.
\end{equation}
Provided that the numerical fluxes $\flux{u}_h, \flux{\textbf{\emph{q}}}_h$ are consistent, the primal formulation (\ref{eq4:3}) is consistent, ensuring that the exact solution $(u,\textbf{\emph{p}},\textbf{\emph{q}})$ of (\ref{eq3:2}) satisfies
\begin{equation}
B(u,\textbf{\emph{p}},\textbf{\emph{q}};v,\pmb{\phi},\pmb{\pi}) = \mathcal{L}(v,\pmb{\phi},\pmb{\pi}).\label{eq4:5}
\end{equation}
for all $(v,\pmb{\phi},\pmb{\pi}) \in H^1(0,T;V_h)\times (L^2(0,T;V_h))^2\times (L^2(0,T;V_h))^2$.
\subsection{Numerical stability}
Let $(\widetilde{u}_h,\widetilde{\textbf{\emph{p}}}_h,\widetilde{\textbf{\emph{q}}}_h) \in H^1(0,T;V_h)\times (L^2(0,T;V_h))^2\times (L^2(0,T;V_h))^2$ be the approximation of the solution
$(u_h,\textbf{\emph{p}}_h,\textbf{\emph{q}}_h)$. We denote $e_{u_h} := u_h - \widetilde{u}_h$, $e_{\textbf{\emph{p}}_h} := \textbf{\emph{p}}_h - \widetilde{\textbf{\emph{p}}_h}$ and $e_{\textbf{\emph{q}}_h} := \textbf{\emph{q}}_h - \widetilde{\textbf{\emph{q}}}_h$ as the roundoff errors.
\begin{thm}{($L^2$ stability).}
Numerical scheme (\ref{eq4:5}) with central flux is $L^2$ stable, and for all $t\in (0,T)$ its solution satisfies
\begin{equation}
\begin{split}
& \parallel e_{u_h}(\cdot,t)\parallel^2_{L^2(\Omega)} \\
& = \parallel e_{u_h}(\cdot,0)\parallel^2_{L^2(\Omega)} \\
& ~~~ - \ 2cos((\alpha/2-1)\pi)\int_{0}^{t} \int_{c}^{d}\parallel e_{p_h^x}(\cdot,y,t) \parallel^2_{H^{\frac{\alpha}{2}-1}} dydt\\
& ~~~ - \ 2cos((\beta/2-1)\pi) \int_{0}^{t} \int_{a}^{b}\parallel e_{p_h^y}(x,\cdot,t) \parallel^2_{H^{\frac{\beta}{2}-1}} dxdt.
\end{split}
\end{equation}
\label{thm1}
\end{thm}
\begin{proof}
We just prove the case $t = T$. From (\ref{eq4:3}), we recover the perturbation equation
\begin{equation}
B(e_{u_h},e_{\textbf{\emph{p}}_h},e_{\textbf{\emph{q}}_h};v,\pmb{\phi},\pmb{\pi}) = 0
\end{equation}
for all $(v,\pmb{\phi},\pmb{\pi}) \in H^1(0,T;V_h)\times (L^2(0,T;V_h))^2\times (L^2(0,T;V_h))^2$. Take $v = e_{u_h}$, $\pmb{\phi} = - e_{\textbf{\emph{p}}_h}$, and
$\pmb{\pi} = e_{\textbf{\emph{q}}_h}$, to obtain
\begin{equation}
\begin{split}
0 &= B(e_{u_h},e_{\textbf{\emph{p}}_h},e_{\textbf{\emph{q}}_h};e_{u_h},-e_{\textbf{\emph{p}}_h},e_{\textbf{\emph{q}}_h}) \\
&= \frac{1}{2}\int_{0}^{T} \frac{\partial }{\partial t} \parallel e_{u_h}(\cdot,t)\parallel^2_{L^2(\Omega)} dt + \int_{0}^{T} \int_{\Omega}\nabla\cdot (e_{u_h}e_{\textbf{\emph{q}}_h})d\textbf{\emph{x}} dt \\
& ~~~ + \ 2cos((\alpha/2-1)\pi)\int_{0}^{T} \int_{c}^{d}\parallel e_{p_h^x}(\cdot,y,t) \parallel^2_{H^{\frac{\alpha}{2}-1}} dydt\\
& ~~~ + \ 2cos((\beta/2-1)\pi) \int_{0}^{T} \int_{a}^{b}\parallel e_{p_h^y}(x,\cdot,t) \parallel^2_{H^{\frac{\beta}{2}-1}} dxdt\\
& ~~~ - \int_{0}^{T} \left(\oint_{\Gamma}\flux{e}_{\textbf{\emph{q}}_h}\cdot[e_{u_h}]ds + \oint_{\Gamma_i}\flux{e}_{u_h}[e_{\textbf{\emph{q}}_h}]ds\right)dt.
\end{split}
\label{eq4:8}
\end{equation}
In (\ref{eq4:8}), integration by parts yields
\begin{equation}
\begin{split}
\int_{\Omega_h}\nabla\cdot(e_{u_h}e_{\textbf{\emph{q}}_h})d\textbf{\emph{x}} & = \sum_{k=1}^{K}\int_{D^k}\nabla\cdot(e_{u_h}e_{\textbf{\emph{q}}_h})d\textbf{\emph{x}}
= \sum_{k=1}^{K}\oint_{\partial D^k}\textbf{n}\cdot e_{\textbf{\emph{q}}_h}e_{u_h}ds\\%(e_{u_h}^- = 0 on \Gamma_b)
&= \oint_{\Gamma_i}(\textbf{n}^+\cdot e_{\textbf{\emph{q}}_h}^+e_{u_h}^+ + \textbf{n}^-\cdot e_{\textbf{\emph{q}}_h}^- e_{u_h}^-)ds + \oint_{\Gamma_b} \textbf{n} \cdot e_{\textbf{\emph{q}}_h} e_{u_h} ds.
\end{split}\label{eq4:9}
\end{equation}
With the central flux we recover
\begin{equation}
\begin{split}
& \oint_{\Gamma}\flux{e}_{\textbf{\emph{q}}_h}\cdot[e_{u_h}]ds + \oint_{\Gamma_i}\flux{e}_{u_h}[e_{\textbf{\emph{q}}_h}]ds \\
&= \oint_{\Gamma_i}(\textbf{n}^+\cdot e_{\textbf{\emph{q}}_h}^+e_{u_h}^+ + \textbf{n}^-\cdot e_{\textbf{\emph{q}}_h}^- e_{u_h}^-)ds + \oint_{\Gamma_b} \textbf{n} \cdot e_{\textbf{\emph{q}}_h} e_{u_h} ds.
\end{split}\label{eq4:10}
\end{equation}
Combining (\ref{eq4:8})-(\ref{eq4:10}), the desired result follows.
\end{proof}
\subsection{Error estimate}
For the error estimate, we define the orthogonal projection operators, $\mathbb{P}: H^1(\Omega_h)\rightarrow V_h$, $\mathbb{Q}:((L^2(\Omega_h))^2\rightarrow (V_h)^2$ and $\mathbb{S}: (H^1(\Omega_h))^2\rightarrow (V_h)^2$. For all the elements, $D^k,$ $k= 1,2,\cdots,K$,
$\mathbb{P}, \mathbb{Q}, \mathbb{S}$ are defined to satisfy
\begin{eqnarray}
(\mathbb{P}u-u, v)_{D^k} = 0 &\forall v\in P_N^2(D^k), \nonumber\\
(\mathbb{Q}\bm{u}-\bm{u}, \bm{v})_{D^k} = 0 &\forall \bm{v}\in (P_N^2(D^k))^2, \nonumber\\
(\mathbb{S}\bm{u}-\bm{u}, \bm{v})_{D^k} = 0 &\forall \bm{v}\in (P_N^2(D^k))^2. \nonumber
\end{eqnarray}
\begin{thm}
The $L^2$ error of the numerical scheme (\ref{eq4:5}) with a central flux satisfies
\begin{equation}
\parallel u(\textbf{\emph{x}},t) - u_h(\textbf{\emph{x}},t) \parallel_{L^2(\Omega_h)} \leqslant c(\alpha,\beta)h^{N+1}, \quad \alpha, \beta\in(1,2),
\end{equation}
where $c$ is dependent on $\alpha, \beta, \Omega$ and $N$ represents the order of polynomial.
\end{thm}
\begin{proof}
We denote
$$ e_u = u(\textbf{\emph{x}},t) - u_h(\textbf{\emph{x}},t), \ e_{\textbf{\emph{p}}}(\textbf{\emph{x}},t) = \textbf{\emph{p}}(\textbf{\emph{x}},t) - \textbf{\emph{p}}_h(\textbf{\emph{x}},t), \ e_\textbf{\emph{q}} = \textbf{\emph{q}}(\textbf{\emph{x}},t) - \textbf{\emph{q}}_h(\textbf{\emph{x}},t);$$
and recover the error equation
\begin{equation}
B(e_u,e_\textbf{\emph{p}},e_\textbf{\emph{q}};v,\pmb{\phi},\pmb{\pi}) = 0 \label{eq4:14}
\end{equation}
for all $(v,\pmb{\phi},\pmb{\pi}) \in H^1(0,T;V_h)\times (L^2(0,T;V_h))^2\times (L^2(0,T;V_h))^2$.
Take $$ v = \mathbb{P}u - u_h,\ \pmb{\phi} = \textbf{\emph{p}}_h - \mathbb{Q}\textbf{\emph{p}},\ \pmb{\pi} = \mathbb{S}\textbf{\emph{q}} - \textbf{\emph{q}}_h$$
in (\ref{eq4:14}). After rearranging terms, we obtain
\begin{equation}
B(v,-\pmb{\phi},\pmb{\pi};v,\pmb{\phi},\pmb{\pi}) = B(v^e,-\pmb{\phi}^e,\pmb{\pi}^e;v,\pmb{\phi},\pmb{\pi}), \label{eq4:15}
\end{equation}
where $v^e, \bm{w}^e$ and $\bm{z}^e$ are given as
$$v^e = \mathbb{P}u - u, \pmb{\phi}^e = \textbf{\emph{p}} - \mathbb{Q}\textbf{\emph{p}}, \pmb{\pi}^e = \mathbb{S}\textbf{\emph{q}} - \textbf{\emph{q}}.$$
Following the discussion in the proof of Theorem \ref{thm1}, the left side of (\ref{eq4:15}) becomes
\begin{equation}
\begin{split}
& B(v,-\pmb{\phi},\pmb{\pi};v,\pmb{\phi},\pmb{\pi}) \\
& = \frac{1}{2}\int_{0}^T\frac{\partial}{\partial t}\parallel v(\cdot,t)\parallel^2_{L^2(\Omega_h)}dt \\
& ~~~ + cos((\alpha/2-1)\pi)\int_0^T\int_{c}^{d}\parallel \phi^x(\cdot,y,t)\parallel^2_{H^{\frac{\alpha}{2}-1}{(a,b)}}dydt\\
& ~~~ + cos((\beta/2-1)\pi)\int_0^T\int_{a}^{b}\parallel \phi^y(x,\cdot,t)\parallel^2_{H^{\frac{\beta}{2}-1}{(c,d)}}dxdt;
\end{split}
\end{equation}
and the right hand side can be expressed as
\begin{equation}
B(v^e,-\pmb{\phi}^e,\pmb{\pi}^e;v,\pmb{\phi},\pmb{\pi}) = \Rmnum{1} + \Rmnum{2} + \Rmnum{3} + \Rmnum{4},
\end{equation}
where
\begin{equation}
\Rmnum{1} = \int_0^T \left(\frac{\partial v^e(\cdot,t)}{\partial t}, v(\cdot,t)\right)dt,
\end{equation}
\begin{equation}
\Rmnum{2} = \int_0^T\big[(\pmb{\pi}^e,\nabla v) + (v^e, \nabla\cdot\pmb{\pi}) + (\pmb{\pi}^e, \pmb{\phi}) -(\pmb{\phi}^e, \pmb{\pi}) \big]dt,
\end{equation}
\begin{equation}
\Rmnum{3} = -\int_{0}^{T} \left( \int_{\Gamma}\flux{\pmb{\pi}}^e\cdot[v] ds + \int_{\Gamma_i}v^e \ [\pmb{\pi}] ds \right) dt,
\end{equation}
\begin{equation}
\Rmnum{4} = \int_0^T\left[ \left(\frac{\partial^{\alpha-2}{\phi^e}^x }{\partial x^{\alpha-2}},\phi^x\right) + \left(\frac{\partial^{\beta-2}{\phi^e}^y }{\partial y^{\beta-2}},\phi^y\right) \right]dt.
\end{equation}
Using Cauchy-Schwarz inequality and standard approximation theory, we obtain
\begin{equation}
\begin{split}
\Rmnum{1} &= \frac{1}{2}\int_0^T \parallel \frac{\partial v^e(\cdot,t)}{\partial t} \parallel^2_{L^2(\Omega_h)} dt + \frac{1}{2}\int_0^T \parallel v(\cdot,t)\parallel_{L^2(\Omega_h)} dt \\
&\leqslant ch^{2N+2} + \frac{1}{2}\int_0^T\parallel v(\cdot,t) \parallel^2_{L^2(\Omega)}dt,
\end{split}
\end{equation}
where c is a constant.
In $\Rmnum{2}$, all terms vanish because of the Galerkin orthogonality.
To deal with the term $\Rmnum{3}$, let us reexamine the construction of the Lagrange interpolation bases.
Since we require the same number of Lagrange interpolation points on every element, the basis functions of both elements are equal along the internal edge $e \in \Gamma_i$.
Applying the Cauchy-Schwarz inequality and the trace inequality, $\Rmnum{3}$ can be estimated as
\begin{eqnarray}
\Rmnum{3} &=& -\int_{0}^{T} \int_{\Gamma_b}\flux{\pmb{\pi}}^e\cdot[v] ds dt \nonumber \\
&\leqslant& \frac{1}{2} \int_{0}^{T} \int_{\partial \Omega}(\flux{\pmb{\pi}}^e)^2 ds
+ \frac{1}{2} \int_{0}^{T} \int_{\partial \Omega} ([v])^2ds dt \nonumber \\
&\leqslant& c_1h^{2N+2} + c_2 \int_0^T\parallel v(\cdot,t)\parallel^2_{L^2(\Omega)}dt,
\end{eqnarray}
where $\Gamma_b = \partial\Omega$, $c_1$ and $c_2$ are related with $N$, $\Omega$. Due to the adjoint property and the embedding theorem for the fractional integral, by using the Young's inequality, $\Rmnum{4}$ can be expressed as
\begin{equation}
\begin{split}
\Rmnum{4} &\leqslant \int_0^T \bigg( \frac{1}{2\varepsilon_1}\parallel {\phi^e}^x(\cdot,t) \parallel^2_{L^2(\Omega_h)} + \frac{\varepsilon_1}{2}\int_c^d\parallel\phi^x(\cdot,y,t)\parallel^2_{H^{\alpha-2}(a,b)}dy\\
&~~~ +\frac{1}{2\varepsilon_2}\parallel{\phi^e}^y(\cdot,t)\parallel^2_{L^2(\Omega_h)} + \frac{\varepsilon_2}{2} \int_a^b\parallel\phi^y(x,\cdot,t)\parallel^2_{H^{\beta-2}(c,d)}dx\bigg )dt\\
&\leqslant c/\varepsilon h^{2N+2} + c\varepsilon \int_0^T \Big( \int_c^d\parallel\phi^x(\cdot,y,t)\parallel^2_{H^{\frac{\alpha}{2}-1}(a,b)}dy\\
&~~~ + \int_a^b\parallel\phi^y(x,\cdot,t)\parallel^2_{H^{\frac{\beta}{2}}(c,d)}dx \Big)dt,
\end{split}
\end{equation}
provided $\varepsilon$ is sufficiently small such that $cos((\alpha/2-1)\pi)>c\varepsilon$ and $cos((\beta/2-1)\pi)>c\varepsilon$.
Combining all the above estimates, we obtain
\begin{equation}
\begin{split}
& \frac{1}{2}\int_{0}^T\frac{\partial}{\partial t}\parallel v(\cdot,t)\parallel^2_{L^2(\Omega_h)}dt \\
& ~~~+ (cos((\alpha/2-1)\pi)-c\varepsilon)\int_0^T\int_{c}^{d}\parallel \phi^x(\cdot,y,t)\parallel^2_{H^{\frac{\alpha}{2}-1}{(a,b)}}dydt\\
& ~~~+ (cos((\beta/2-1)\pi)-c\varepsilon)\int_0^T\int_{a}^{b}\parallel \phi^y(x,\cdot,t)\parallel^2_{H^{\frac{\beta}{2}-1}{(c,d)}}dxdt \\
& \leqslant(c/\varepsilon)h^{2N+2} + c\int_0^T\parallel v(\cdot,t)\parallel^2_{L^2(\Omega_h)}dt. \\
\end{split}
\end{equation}
By simplifying further, this yields
\begin{equation}
\begin{split}
&\frac{1}{2}\parallel v(\cdot,T)\parallel^2_{L^2(\Omega_h)}\\
& ~~~+ (cos((\alpha/2-1)\pi)-c\varepsilon)\int_0^T \int_c^d\parallel\phi^x(\cdot,y,t)\parallel^2_{H^{\frac{\alpha}{2}-1}(a,b)}dydt \\
& ~~~+ (cos((\beta/2-1)\pi)-c\varepsilon)\int_{0}^T\int_c^d\parallel\phi^y(x,\cdot,t)\parallel^2_{H^{\frac{\beta}{2}-1}(c,d)} dxdt \\
&\leqslant \frac{1}{2}\parallel v(\cdot,0)\parallel^2_{L^2(\Omega_h)} + (c/\varepsilon)h^{2N+2} + c\int_0^T\parallel v(\cdot,t)\parallel^2_{L^2(\Omega_h)}dt \\
&\leqslant (c/\varepsilon)h^{2N+2} + c\int_0^T\parallel v(\cdot,t)\parallel^2_{L^2(\Omega_h)}dt.
\end{split}
\end{equation}
From the Gr\"{o}nwall's lemma and standard approximation theory, the desired result follows.
\end{proof}
\section{Numerical results}
In this section, we offer a few numerical results to validate analysis. To deal with the method-of-line fractional PDE, i.e., the classical ODE system, we utilize a low-storage five stage fourth order explicit Runge-Kutta method \cite{jan1}. To ensure the overall error is dominated by space error, small time steps are used.
\noindent \textbf{Example 1} We consider the problem
\begin{equation}
\frac{\partial u(x,y,t)}{\partial t} = {_{-1}}D_x^{\alpha}u(x,y,t) + {_{-1}}D_y^{\alpha}u(x,y,t)+ f(x,y,t),
\end{equation}
where
\begin{displaymath}
\begin{split}
f(x,y,t)
&= -e^{-t}\big((x^2-1)^3(y^2-1)^3 + (y^2-1)^3_{-1}D_x^{\alpha-2}(6(x^2-1)(5x^2-1))\\
&+ (x^2-1)^3_{-1}D_y^{\beta-2}(6(y^2-1)(5y^2-1))\big)
\end{split}
\end{displaymath}
on the computational domain $\Omega=(-1,1)\times(-1,1)$ and $\alpha, \beta \in (1,2)$.
We consider the initial condition
\begin{equation}
u(x,y,0) = (x^2-1)^3(y^2-1)^3
\end{equation}
and the Dirichlet boundary condition
\begin{equation}
u(x,y,t) = 0, \quad (x,y) \in \partial \Omega.
\end{equation}
The exact solution is $u(x,y,t)=e^{-t}(x^2-1)^3(y^2-1)^3$.
\begin{table}
\centering
\caption{Numerical errors ($L_2$) and order of convergence on unstructured meshes for Example 1. N denotes the order of polynomial in two variables, and K is the total number of triangle elements.}
\begin{tabular}{*{9}{|c}|}
\hline
&K & \multicolumn{1}{c|}{32} & \multicolumn{2}{c|}{137} & \multicolumn{2}{c|}{378} & \multicolumn{2}{c|}{899} \\
\hline
N &$(\alpha,\beta)$ & error & error & order & error & order & error & order \\ \hline
\multirow{4}{*}{1} & (1.01,1.01) &1.55e-1 &3.48e-2 &2.16 &1.27e-2 &1.99 & 5.39e-3 &2.11 \\
\cline{2-9}
& (1.5,1.5) & 1.44e-1 &3.21e-2 &2.17 &1.22e-2 &1.94 &5.09e-3 &2.16 \\
\cline{2-9}
& (1.8,1.8) & 1.34e-1 &3.03e-2 &2.14 &1.13e-2 &1.93 &5.40e-3 &1.82\\
\cline{2-9}
&(1.99,1.99)& 1.27e-1 &2.94e-2 &2.11 &1.07e-2 &1.98 &4.48e-3 &2.15\\
\hline
&K & \multicolumn{1}{c|}{32} & \multicolumn{2}{c|}{137} & \multicolumn{2}{c|}{378} & \multicolumn{2}{c|}{899} \\\hline
\multirow{4}{*}{2} & (1.01,1.01) &1.65e-2 &4.39e-3 &2.62 &5.22e-4 &2.92 &1.44e-4 &2.97 \\
\cline{2-9}
& (1.5,1.5) &1.40e-2 &3.47e-3 &2.76 &3.41e-3 &3.18 &9.19e-5 &3.02 \\
\cline{2-9}
& (1.8,1.8) &1.33e-2 &1.57e-3 &3.08 &3.37e-4 &3.01 &9.20e-5 &3.20 \\
\cline{2-9}
&(1.99,1.99)&1.36e-2 &1.57e-3 &3.11 &3.56e-4 &2.90 &9.75e-5 &3.20\\
\hline
&K & \multicolumn{1}{c|}{88} & \multicolumn{2}{c|}{137} & \multicolumn{2}{c|}{378} & \multicolumn{2}{c|}{562} \\ \hline
\multirow{4}{*}{3} & (1.01,1.01) &5.51e-4 &2.17e-4 &4.20 &3.07e-5 &3.86 &1.53e-5 &3.82\\
\cline{2-9}
& (1.5,1.5) &3.73e-4 &1.38e-4 &4.49 &1.88e-5 &3.93 &8.53e-6 &4.33\\
\cline{2-9}
& (1.8,1.8) &3.30e-4 &1.33e-4 &4.10 &2.00e-5 &3.73 &8.05e-6 &4.99 \\
\cline{2-9}
&(1.99,1.99)&3.83e-4 &1.61e-4 &3.92 &2.55e-5 &3.63 &1.81e-5 &4.23\\
\hline
\end{tabular}
\label{table1}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.35\textwidth]{mesh1.eps}
\includegraphics[width=0.37\textwidth]{mesh2.eps}\\
\includegraphics[width=0.35\textwidth]{mesh3.eps}
\includegraphics[width=0.38\textwidth]{mesh4.eps}
\caption{Partial unconstructed meshes used in numerical examples }
\label{fig5.2}
\end{figure}
\noindent \textbf{Example 2} Now we expand our algorithm to fractional diffusion equations with both the left and right fractional derivatives.
We consider the model
\begin{equation}
\begin{split}
\frac{\partial u}{\partial t}& = d_+ {_{-1}}D_x^{\alpha}u(x,y,t) + d_- {_x}D_1^{\alpha}u(x,y,t) \\
& ~~~ + e_+ {_{-1}}D_y^{\beta}u(x,y,t) + e_- {_y}D_1^{\beta}u(x,y,t)+ f(x,y,t),
\end{split}
\end{equation}
with the initial condition
$$
u(x,y,0) = (x^2-1)^3(y^2-1)^3
$$
and boundary condition
$$
u(x,y,t)|_{\partial \Omega} = 0.
$$
Here, $(x,y)\in (-1,1) \times (-1,1)$, $\alpha, \beta \in(1,2)$ and $d_+=d_1=1$, $e_+=e_1=1$,
\begin{equation}
\begin{split}
f(x,y,t) &= -e^{-t}\bigg( (x^2-1)^3(y^2-1)^3 \\
&+ (_{-1}D_x^{\alpha-2}+{_x}D_1^{\alpha-2})\big(6(x^2-1)(5x^2-1)\big) (y^2-1)^3 \\
&+(_{-1}D_y^{\beta-2}+{_y}D_1^{\beta-2})\big(6(y^2-1)(5y^2-1)\big) (x^2-1)^3\bigg).
\end{split}
\nonumber
\end{equation}
The exact solution is $u(x,y,t) = e^{-t}(x^2-1)^3(y^2-1)^3$.
\begin{table}
\centering
\caption{Numerical errors ($L_2$) and order of convergence on unstructured meshes for Example 2. N denotes the order of polynomial in two variables, and K is the total number of triangle elements.}
\begin{tabular}{*{9}{|c}|}
\hline
&K & \multicolumn{1}{c|}{32} & \multicolumn{2}{c|}{137} & \multicolumn{2}{c|}{378} & \multicolumn{2}{c|}{899} \\
\hline
N &$(\alpha,\beta)$ & error & error & order & error & order & error & order \\ \hline
\multirow{4}{*}{1} & (1.01,1.01) &1.54e-1 &3.45e-2 &2.16 &1.24e-2 &2.00 & 5.19e-3 &2.15 \\
\cline{2-9}
& (1.1,1.8) & 1.32e-1 &3.11e-2 &2.08 &1.18e-2 &1.91 &5.21e-3 &2.01 \\
\cline{2-9}
& (1.5,1.5) & 1.24e-1 &2.97e-2 &2.17 &1.09e-2 &1.96 &5.19e-3 &2.16 \\
\cline{2-9}
&(1.99,1.99)& 1.13e-1 &2.66e-2 &2.09 &9.69e-3 &1.98 &4.05e-3 &2.15\\
\hline
&K & \multicolumn{1}{c|}{32} & \multicolumn{2}{c|}{137} & \multicolumn{2}{c|}{378} & \multicolumn{2}{c|}{899} \\\hline
\multirow{4}{*}{2} & (1.01,1.01) &1.64e-2 &2.25e-3 &2.87 &5.04e-4 &2.93 &1.66e-4 &2.74 \\
\cline{2-9}
& (1.5,1.5) &1.32e-2 &1.62e-3 &3.03 &3.29e-3 &3.12 &9.96e-5 &2.94 \\
\cline{2-9}
& (1.8,1.1) &1.48e-2 &2.18e-3 &2.76 &4.69e-4 &3.00 &1.28e-4 &3.21 \\
\cline{2-9}
&(1.99,1.99)&1.44e-2 &1.64e-3 &3.13 &3.88e-4 &2.82 &1.10e-4 &3.10\\
\hline
&K & \multicolumn{1}{c|}{32} & \multicolumn{2}{c|}{88} & \multicolumn{2}{c|}{137} & \multicolumn{2}{c|}{378} \\ \hline
\multirow{4}{*}{3} & (1.01,1.01) &3.42e-3 &5.23e-4 &3.67 &2.01e-4 &5.25 &2.47e-5 &4.10\\
\cline{2-9}
& (1.5,1.5) &2.36e-3 &3.01e-4 &4.031 &1.10e-4 &5.53 &1.39e-5 &4.05\\
\cline{2-9}
& (1.2,1.6) &2.62e-3 &3.62e-4 &3.88 &1.46e-4 &4.98 &2.31-5 &3.61 \\
\cline{2-9}
&(1.99,1.99)&2.36e-4 &3.01e-4 &4.03 &1.10e-5 &5.52 &1.39e-5 &4.05\\
\hline
\end{tabular}
\label{table2}
\end{table}
In Table \ref{table1} and Table \ref{table2}, we note a convergence rate $\mathcal{O}(h^{N+1})$, which demonstrates that the central flux is optimal in two dimension.
Note that the small deviations from the optimal order are caused by the unstructured nature of the grid and the numerical quadrature for fractional integral operators.
\section{Conclusions}
In this paper, we propose nodal discontinuous Galerkin methods for two dimensional Riemann-Liouville fractional equations on unstructured meshes, and offer a stability analysis and error estimates. Numerical experiments confirm that the optimal order of convergence is recovered. For the convenience of the theoretical analysis we have restricted our models to homogeneous boundary conditions, although this is not an essential restriction.
\section*{Acknowledgements}
This work was supported by NSFC 11271173, NSF DMS-1115416, and OSD/AFOSR FA9550-09-1-0613.
|
\section{Introduction}\label{sec:intro}
The matrix model has been playing a crucial role in quantum field theory
(QFT) for the last few decades, as a zero dimensional QFT model, or
rather a toy model for the infinite dimensional path integral.
Since it is just defined by finite dimensional integral of a matrix
itself, one can obtain an exact solution in various situations, which
provides a significant insight for the understanding of QFT.
This methodology is also refereed to as random matrix theory (RMT), and
is now applied to the extremely wide range of research
fields~\cite{Akemann:2011RMT}.
In QFT, in order to compute correlation functions, it is convenient to
introduce the generating function by adding the extra source term
\begin{equation}
\mathcal{Z}[J] = \int \mathcal{D}\phi \,
e^{-\frac{1}{\hbar} S[\phi] + \int \! d^Dx \, J(x) \phi(x)}
\, .
\end{equation}
The correlation function can be obtained by taking the functional
derivative with the source field
\begin{equation}
\Big< \phi(x_1) \cdots \phi(x_k) \Big>
=
\frac{\delta}{\delta J(x_1)} \cdots \frac{\delta}{\delta J(x_k)}
\log \mathcal{Z}[J] \Big|_{J=0}
\, .
\end{equation}
This generating function is defined in the sense of path integral, and
thus it is quite difficult to compute in general.
On the other hand, the matrix model having an external source, which is
just given by finite dimensional integral, plays a similar role to the
generating function in QFT,
\begin{equation}
\mathcal{Z}_N (A) = \int dX \, e^{-\frac{1}{\hbar} \mathrm{Tr}\, W(X) + \mathrm{Tr}\, XA}
\, ,
\label{MM_ext}
\end{equation}
where the integral is taken over a size $N$ matrix.
We can compute the correlation function by taking the derivative with
respect to the source matrix $A$, as well as the QFT generating
function.
A remarkable feature of the matrix model with the external source is the
duality with a characteristic polynomial, which was found
by~\cite{Brezin:2000CMP} especially for the Gaussian matrix model.
The claim of the duality is as follows: The $M$-point correlation function of
characteristic polynomials in the size $N$ matrix model with an external
source is equivalent to the $N$-point function with a size $M$ matrix,
under exchanging the arguments of the characteristic polynomial and the
external source.
This duality is now extended to generic
$\beta$-ensemble~\cite{Desrosiers:2008tp}, chiral
ensemble~\cite{Forrester:2013JPA}, and beyond the Gaussian
model~\cite{Kimura:2014mua}, and has various interesting interpretations
in terms of conformal field theory, string/M-theory, knot theory,
algebraic geometry, and so on.
Another interesting property of the external source matrix model is the
connection with the integrable system.
The matrix integral with an external source and also the characteristic
polynomial average satisfy the integrable equation, e.g., the Toda
lattice equation, and they can be interpreted as the corresponding
$\tau$-function.
See, for example,~\cite{Morozov:1994hh}.
Such integrability of the matrix model is quite useful to determine the
spectral density and also the correlation function.
In this paper we study a supersymmetric version of the matrix model
involving the external source based on a Hermitian supermatrix.
The supermatrix method is now well-known both in high energy physics and
condensed matter physics.
For example, one can avoid technical difficulty in dealing with the
normalization factor of the correlation function by applying the
supermatrix method, in a similar way to the replica method.
This property helps us to take the disorder average
in a random system~\cite{Efetov199610}.
In Sec.~\ref{sec:source} we will compute the $\mathrm{U}(N|N)$ symmetric
supermatrix partition function involving the external source.
As in the case of the ordinary Hermitian matrix model, we can utilize
the determinantal structure in order to perform the integral.
We will obtain a determinantal formula for the partition function, which
can be expressed as a size $N$ determinant of the $\mathrm{U}(1|1)$ partition
function as a determinantal kernel.
As in the case of the ordinary matrix model, there have been some proposals for
the duality of the supermatrix model.
For example, in particular for the Gaussian supermatrix model, one can
show an explicit duality between the external source and the
characteristic polynomial~\cite{Desrosiers:2009pz}.
In Sec.~\ref{sec:ch_poly} we will exhibit that this duality generally
holds for the supermatrix with an arbitrary potential function, and
is just interpreted as Fourier transformation.
In addition, it will be pointed out that the relation between the
characteristic polynomial ratio with the ordinary matrix
model~\cite{Fyodorov:2002jw} and supermatrix
models~\cite{Zirnbauer:1996zz,Brezin:2003JPA,Kieburg:2009jd} is naturally
explained as some specialization of the characteristic polynomial
expectation value in the supermatrix.
In Sec.~\ref{sec:integrable} we will also show the integrable equations
for the supermatrix integral with the external source and the
characteristic polynomial, based on the determinantal formula derived in
this article.
Since we have more parameters for the partition function rather than the
ordinary matrix model, we can obtain several integrable equations
corresponding to variables parametrizing the external source and the
characteristic polynomial.
\section{Supermatrix model with external source}\label{sec:source}
Let us introduce a Hermitian supermatrix model with a potential
function $W(x)$, involving the external source matrix $C$, as a natural
generalization of the Hermitian matrix model~(\ref{MM_ext}),
\begin{equation}
\mathcal{Z}_{N,M} (\{a_i\}_{i=1}^N, \{b_j\}_{j=1}^M)
=
\int dZ \, e^{-\frac{1}{\hbar} \mathrm{Str}\, W(Z) + \mathrm{Str}\, Z C}
\, ,
\label{SMM01}
\end{equation}
where $\hbar$ is the coupling constant.
The Hermitian supermatrices $Z$ and $C$ are given by
\begin{equation}
Z =
\left(
\begin{array}{cc}
X & \xi \\ \xi^\dag & Y
\end{array}
\right)
\, , \qquad
C =
\left(
\begin{array}{cc}
A & \eta \\ \eta^\dag & B
\end{array}
\right)
\, ,
\end{equation}
where $X$, $A$ are $N \times N$, and $Y$, $B$ are $M \times M$
(bosonic) Hermitian matrices.
$\xi$ and $\eta$ are $N \times M$ fermionic matrices, whose elements are
given by Gra{\ss}mannian variables.
We can assume $C$ is a diagonal matrix without loss of generality, and
in this paper we also assume $N \ge M$.
The matrix measure of (\ref{SMM01}) is invariant under the supergroup
transformation, $dZ = d(U Z U^{-1})$ with $U \in \mathrm{U}(N|M)$, and is
normalized by the corresponding volume factor $\operatorname{Vol.} \mathrm{U}(N|M)$.
As well as the ordinary Hermitian matrix integral, one can decompose
this measure into the diagonal and angular parts
\begin{equation}
dZ = \Delta_{N,M}(x;y)^2
d^N x \, d^N y \, dU
\, ,
\end{equation}
and the corresponding Jacobian $\Delta_{N,M}(x;y)$ is now given by the
generalized Cauchy determinant~\cite{Basor:1994MN}
\begin{align}
\Delta_{N,M}(x;y)
&\equiv
\Delta_N(x) \Delta_M(y)
\prod_{i=1}^N \prod_{j=1}^M (x_i - y_j)^{-1}
\nonumber \\
& =
\de
\left(
\begin{array}{c}
\displaystyle x_i^{k-1} \\
\displaystyle (x_i-y_j)^{-1}
\end{array}
\right)
\quad
\mbox{with}
\quad
\begin{cases}
i = 1, \cdots, N \\
j = 1, \cdots, M \\
k = 1, \cdots, N-M
\end{cases}
\, ,
\label{Cauchy_det}
\end{align}
where $\Delta_N(x)$ is the Vandermonde determinant
\begin{equation}
\Delta_N(x)
= \det_{1 \le i, j \le N} x_i^{j-1}
= \prod_{i<j}^N (x_i - x_j)
\, .
\end{equation}
The fact that the Jacobian factor can be written as a determinant will
play a crucial role in the following discussion.
To perform the angular part integral of (\ref{SMM01}), we now apply
the supergroup version of the Harish-Chandra--Itzykson--Zuber (HCIZ)
formula, which is written in terms of the Cauchy
determinant~\cite{Guhr:1991JMP,Alfaro:1994ca,Guhr:1996CMP},
\begin{align}
\mathcal{I}_{N,M}(Z,C)
& \equiv
\int_{\mathrm{U}(N|M)} dU \, e^{\mathrm{Str}\, Z U C U^{-1}}
\nonumber \\
&
=
\frac{1}{\Delta_{N,M}(x;y) \Delta_{N,M}(a;b)}
\det_{1 \le i, j \le N} e^{x_i a_j}
\det_{1 \le i, j \le M} e^{- y_i b_j}
\, .
\label{HCIZ_U(N|M)}
\end{align}
This is a quite natural extension of the original formula for the $\mathrm{U}(N)$
group, which is given by replacing the Cauchy determinant with the
Vandermonde determinant~\cite{HC:1957,Itzykson:1979fi},
\begin{eqnarray}
\mathcal{I}_N (X,A)
& \equiv &
\int_{\mathrm{U}(N)} dU \, e^{\mathrm{Tr}\, X U A U^{-1}}
\nonumber \\
& = &
\frac{1}{\Delta_N(x) \Delta_N(a)}
\det_{1 \le i, j \le N} e^{x_i a_j} \, .
\label{HCIZ_U(N))}
\end{eqnarray}
Applying the formula (\ref{HCIZ_U(N|M)}), we obtain the expression only
in terms of eigenvalues
\begin{align}
\mathcal{Z}_{N,M}
& =
\frac{1}{\Delta_{N,M}(a;b)}
\int
\prod_{i=1}^N dx_i \,
e^{-\frac{1}{\hbar} W(x_i) + x_i a_i}
\prod_{j=1}^M dy_j \,
e^{\frac{1}{\hbar} W(y_j) - y_j b_j} \,
\Delta_{N,M}(x;y)
\, .
\end{align}
Let us comment on convergency of this integral.
Naively speaking, it is not convergent integral in general.
We can regard this as a formal integral,
but one possible way to avoid this divergence is to modify the matrix
potential by inserting a constant supermatrix $I_{N,M} = \mathrm{diag}(\mathbbm{1}_N,
i \mathbbm{1}_M) \in \mathrm{U}(N|M)$, i.e.,
\mathrm{Str}\, Z^2 = \mathrm{Tr}\, X^2 - \mathrm{Tr}\, Y^2 \to \mathrm{Str}\, I_{N,M} Z^2 = \mathrm{Tr}\, X^2 + \mathrm{Tr}\,
Y^2$~\cite{Desrosiers:2008tp}.
However, this treatment does not work for the generic potential function
$W(x)$.
Another possibility is considering the coupling constant as an imaginary
number, $\hbar \in i \mathbb{R}$.
In this case, although the integrand becomes an oscillating function,
one can provide interpretation as the Fresnel integral.
For example, this interpretation works well for the Chern--Simons matrix
model~\cite{Marino:2002fk} and its supermatrix generalization, which is
called the ABJM matrix
model~\cite{Kapustin:2009kz,Drukker:2009hy,Marino:2009jd}, although the
corresponding coupling constant is pure imaginary, e.g., $\hbar = 2 \pi
i/(k+N)$, or $\hbar = 2 \pi i / k$.
Then, due to the Cauchy determinantal formula (\ref{Cauchy_det}), it can
be written as a size $N$ determinant, consisting of $N \times M$ and $N
\times (N-M)$ blocks,
\begin{equation}
\mathcal{Z}_{N,M} (\{a_i\}_{i=1}^N,\{b_j\}_{j=1}^M)
=
\frac{1}{\Delta_{N,M}(a;b)}
\det
\left(
\begin{array}{c}
Q_{k-1}(a_i) \\ R(a_i;b_j)
\end{array}
\right)
\, ,
\label{SMM02}
\end{equation}
where the indices run as $i = 1, \cdots, N$, $j = 1, \cdots, M$, $k = 1,
\cdots, N-M$, and we have introduced the following functions,
\begin{align}
P_{i-1}(x)
& =
x^{i-1}
\, , \quad
\tilde{R}(x;y) = \frac{1}{x-y}
\, , \label{P-func} \\
Q_{i-1}(a)
& =
\int dx \, P_{i-1}(x) \, e^{-\frac{1}{\hbar} W(x) + xa} \, ,
\label{Q-func}
\\
R(a;b)
& =
\int dx dy \,
\tilde{R}(x;y) \, e^{-\frac{1}{\hbar} W(x) + \frac{1}{\hbar} W(y) + xa - yb}
\, .
\label{R-func}
\end{align}
Note that the function (\ref{Q-func}) obeys
\begin{equation}
Q_{i}(a) = \frac{d}{d a} Q_{i-1}(a) \, ,
\end{equation}
and $Q_{i=0}(a)$ is seen as a generalized Airy function
\begin{equation}
Q_{i=0}(a) = \int dx \, e^{-\frac{1}{\hbar} W(x) + x a} \, .
\end{equation}
As we will see later, these functions are just seen as (double) Fourier
or rather Laplace transform of $\displaystyle P_{i-1}(x)
e^{-\frac{1}{\hbar}W(x)}$ and $\displaystyle \tilde{R}(x;y)
e^{-\frac{1}{\hbar} W(x) + \frac{1}{\hbar} W(y)}$, respectively.%
\footnote{
If we modify the source term in the supermatrix integral
\begin{equation}
\mathcal{Z}_{N,M} (\{a_i\}_{i=1}^N, \{b_j\}_{j=1}^M)
=
\int dZ \, e^{-\frac{1}{\hbar} \mathrm{Str}\, W(Z) + i \, \mathrm{Str}\, Z C}
\, ,
\end{equation}
they just become Fourier transforms:
\begin{equation}
Q_{i-1}(a) = \int dx \, P_{i-1}(x) \,
e^{-\frac{1}{\hbar} W(x) + i x a}
\, , \quad
R(a;b)
=
\int dx dy \,
\tilde{R}(x;y) \,
e^{-\frac{1}{\hbar} W(x) + \frac{1}{\hbar} W(y) + i xa - i yb} \, .
\end{equation}
}
Let us comment on some specialization of the formula (\ref{SMM02}).
If we take the limit $M=0$, the determinantal expression (\ref{SMM02})
is reduced to the the well-known formula for the ordinary Hermitian
matrix model with the external source
\begin{align}
\mathcal{Z}_{N} (\{a_i\}_{i=1}^N)
& =
\int \! dX \, e^{-\frac{1}{\hbar} \mathrm{Tr}\, W(X) + \mathrm{Tr}\, XA}
\nonumber \\
& =
\frac{1}{\Delta_N(a)} \det_{1 \le i, j \le N} Q_{j-1}(a_i)
\, .
\end{align}
For example, see \cite{Morozov:1994hh} for details.
Then, in the case with $M=N$, the partition function becomes
\begin{align}
\mathcal{Z}_{N,N} (\{a_i\}_{i=1}^N,\{b_j\}_{j=1}^N)
& =
\frac{1}{\Delta_{N,N}(a;b)}
\det_{1 \le i, j \le N}
R(a_i;b_j)
\, .
\end{align}
This shows that the formula for $\mathrm{U}(N|N)$ theory is reduced to a size $N$
determinant of the ``dual'' kernel function $R(a;b)$, which is given by
the $\mathrm{U}(1|1)$ partition function
\begin{equation}
R(a;b) = \frac{\mathcal{Z}_{1,1} (a,b)}{a-b} \, .
\end{equation}
This factorization property is called Giambelli
compatibility~\cite{Borodin:2006AAM}, and follows the Fay
identity~\cite{Fay:1973}.
\section{Characteristic polynomial average and duality}\label{sec:ch_poly}
In this section we consider expectation values of the
characteristic polynomial ratio with the supermatrix model in the
presence of an external source
\begin{align}
&
\Psi_{N,M;\,p,q} (\{a_i\}_{i=1}^N, \{b_j\}_{j=1}^M;
\{\lambda_\alpha\}_{\alpha=1}^p, \{\mu_\beta\}_{\beta=1}^q)
\nonumber \\
& \hspace{5em}
=
\int \! dZ \, e^{-\frac{1}{\hbar} \mathrm{Str}\, W(Z) + \mathrm{Str}\, Z C}
\prod_{\alpha=1}^p \mathrm{Sdet} \left( \lambda_\alpha - Z \right)
\prod_{\beta=1}^q \mathrm{Sdet} \left( \mu_\beta - Z \right)^{-1}
\, ,
\label{EV_ratio01}
\end{align}
and then show the duality between the $(p|q)$-point function with
$\mathrm{U}(N|M)$ theory and $(N|M)$-point function with $\mathrm{U}(p|q)$ theory.
The characteristic polynomial average (\ref{EV_ratio01}) includes
several useful situations:
\begin{itemize}
\item $M=q=0$: Characteristic polynomial with the Hermitian matrix model
\begin{equation}
\Psi_{N;\,p} (\{a_i\}_{i=1}^N; \{\lambda_\alpha\}_{\alpha=1}^p)
=
\int dX \,
e^{-\frac{1}{\hbar}\mathrm{Tr}\, W(X) + \mathrm{Tr}\, XA}
\prod_{\alpha=1}^p \det (\lambda_\alpha - X)
\, .
\label{M=q=0}
\end{equation}
\item $M=p=0$: Characteristic polynomial inverse
with the Hermitian matrix model
\begin{equation}
\Psi_{N;\,q} (\{a_i\}_{i=1}^N; \{\mu_\beta\}_{\beta=1}^q)
=
\int dX \,
e^{-\frac{1}{\hbar}\mathrm{Tr}\, W(X) + \mathrm{Tr}\, XA}
\prod_{\beta=1}^q \det (\mu_\beta - X)^{-1}
\, .
\label{M=p=0}
\end{equation}
\item $M=0$: Characteristic polynomial ratio with the Hermitian matrix model
\begin{align}
&
\Psi_{N;\,p, q}
(\{a_i\}_{i=1}^N;
\{\lambda_\alpha\}_{\alpha=1}^p, \{\mu_\beta\}_{\beta=1}^q)
\nonumber \\
& \hspace{5em}
=
\int dX \,
e^{-\frac{1}{\hbar}\mathrm{Tr}\, W(X) + \mathrm{Tr}\, XA}
\prod_{\alpha=1}^p \det (\lambda_\alpha - X)
\prod_{\beta=1}^q \det (\mu_\beta - X)^{-1}
\, .
\label{M=0}
\end{align}
\end{itemize}
Therefore the expression (\ref{EV_ratio01}) provides a {\em master formula}
for the characteristic polynomial average in various matrix models.
For example, the duality in the case with $M=0$ (\ref{M=0}) claims that the
characteristic polynomial ratio with the ordinary Hermitian matrix model
is dual to another supermatrix model, as shown in~\cite{Brezin:2003JPA}.
As discussed in Sec.~\ref{sec:source}, the angular part of the
integral is performed using the HCIZ formula~(\ref{HCIZ_U(N|M)}), and then we
obtain the expression only in terms of eigenvalues,
\begin{align}
\Psi_{N,M;\,p,q}
& =
\frac{1}{{\Delta}_{N,M}(a;b)}
\int
\prod_{i=1}^N dx_i \, e^{-\frac{1}{\hbar} W(x_i) + x_i a_i}
\prod_{j=1}^M dy_j \, e^{\frac{1}{\hbar} W(y_j) - y_j b_j}
\, \Delta_{N,M}(x;y)
\nonumber \\
&
\hspace{10em} \times
\prod_{\alpha=1}^p
\left(
\frac{\prod_{i=1}^N (\lambda_\alpha - x_i)}
{\prod_{j=1}^M (\lambda_\alpha - y_j)}
\right)
\prod_{\beta=1}^q
\left(
\frac{\prod_{j=1}^M (\mu_\beta - y_j)}
{\prod_{i=1}^N (\mu_\beta - x_i)}
\right)
\, .
\end{align}
It is now convenient to apply the following identity to this expression,
\begin{equation}
{\Delta}_{N+p,M+q}(x,\lambda;y,\mu)
=
{\Delta}_{N,M} (x;y)
{\Delta}_{p,q} (\lambda;\mu)
\prod_{\alpha=1}^p
\left(
\frac{\prod_{i=1}^N (\lambda_\alpha - x_i)}
{\prod_{j=1}^M (\lambda_\alpha - y_j)}
\right)
\prod_{\beta=1}^q
\left(
\frac{\prod_{j=1}^M (\mu_\beta - y_j)}
{\prod_{i=1}^N (\mu_\beta - x_i)}
\right)
\, .
\end{equation}
If we assume $N+p > M+q$, the LHS of this equation is a large Cauchy
determinant of the size $N+p$, with several block structures,
\begin{align}
\Delta_{N+q, M+q}(a,\lambda; b,\mu)
& =
\det
\left(
\begin{array}{cc}
x_i^{k-1} & \lambda_\alpha^{k-1} \\
\left(x_i - y_j\right)^{-1} & \left(\lambda_\alpha - y_j\right)^{-1} \\
\left(x_i - \mu_\beta\right)^{-1}
& \left(\lambda_\alpha - \mu_\beta\right)^{-1}
\end{array}
\right)
\\
& = (-1)^{Np+Mq}
\det
\left(
\begin{array}{cc}
\lambda_\alpha^{k-1} & x_i^{k-1} \\
\left(\lambda_\alpha - \mu_\beta\right)^{-1}
& \left(x_i - \mu_\beta\right)^{-1} \\
\left(\lambda_\alpha - y_j\right)^{-1} & \left(x_i - y_j\right)^{-1} \\
\end{array}
\right)
\, ,
\end{align}
where $i = 1, \cdots, N$, $j = 1, \cdots, M$,
$\alpha = 1, \cdots, p$, $\beta = 1, \cdots, q$, and $k = 1, \cdots, N+p-M-q$.
Due to this expression, we obtain the determinantal formula for the
characteristic polynomial ratio expectation value
\begin{align}
&
\Psi_{N,M;\,p,q}
( \{a_i\}_{i=1}^N, \{b_j\}_{j=1}^M;
\{\lambda_\alpha\}_{\alpha=1}^p, \{\mu_\beta\}_{\beta=1}^q )
\nonumber \\
& =
\frac{1}{{\Delta}_{N,M}(a;b) {\Delta}_{p,q}(\lambda;\mu)}
\int
\prod_{i=1}^N dx_i \, e^{-\frac{1}{\hbar} W(x_i) + x_i a_i}
\prod_{j=1}^M dy_j \, e^{\frac{1}{\hbar} W(y_j) - y_j b_j}
\, \Delta_{N+p,M+q}(x,\lambda; y,\mu)
\nonumber \\
& =
\frac{1}{{\Delta}_{N,M}(a;b) {\Delta}_{p,q}(\lambda;\mu)}
\det
\left(
\begin{array}{cc}
Q_{k-1} (a_i) & P_{k-1}(\lambda_\alpha) \\
R(a_i;b_j) & S_{\rm R} (\lambda_\alpha;b_j) \\
S_{\rm L} (a_i;\mu_\beta) & \tilde{R}(\lambda_\alpha;\mu_\beta) \\
\end{array}
\right)
\, ,
\label{SChPoly_det01}
\end{align}
where we have introduced the following auxiliary functions in addition
to (\ref{P-func}), (\ref{Q-func}) and (\ref{R-func}),
\begin{equation}
S_{\rm L} (a;\mu)
=
\int \! dx \,
\frac{1}{x-\mu} \, e^{-\frac{1}{\hbar} W(x) + xa}
\, , \quad
S_{\rm R} (\lambda;b)
=
\int \! dy \,
\frac{1}{\lambda - y} \, e^{\frac{1}{\hbar} W(y) - y b}
\label{R_tilde}\, .
\end{equation}
The formula (\ref{SChPoly_det01}) shows a duality between the external
source and the characteristic polynomials in the supermatrix model with
an arbitrary potential function $W(x)$, since it can be also written as
follows
\begin{align}
&
\Psi_{N,M;\,p,q}
( \{a_i\}_{i=1}^N, \{b_j\}_{j=1}^M;
\{\lambda_\alpha\}_{\alpha=1}^p, \{\mu_\beta\}_{\beta=1}^q )
\nonumber \\
& \hspace{3em}
=
\frac{(-1)^{Np+Mq}}{{\Delta}_{N,M}(a;b) {\Delta}_{p,q}(\lambda;\mu)}
\det
\left(
\begin{array}{cc}
P_{k-1}(\lambda_\alpha) & Q_{k-1} (a_i) \\
\tilde{R}(\lambda_\alpha;\mu_\beta) & S_{\rm L} (a_i;\mu_\beta) \\
S_{\rm R} (\lambda_\alpha;b_j) & R(a_i;b_j) \\
\end{array}
\right)
\, .
\end{align}
This duality was originally shown in the
Gaussian model with the harmonic potential $W(x) = \frac{1}{2}
x^2$~\cite{Desrosiers:2009pz}.
Under this duality, we see the correspondence between the auxiliary
functions:
\begin{equation}
P_k(\lambda)
\ \longleftrightarrow \
Q_k(a)
\, , \quad
S_{\rm L}(a;\mu)
\ \longleftrightarrow \
S_{\rm R}(\lambda;b)
\, , \quad
R(a;b)
\ \longleftrightarrow \
\tilde{R}(\lambda;\mu)
\, ,
\end{equation}
and then the $(p|q)$-point function for $\mathrm{U}(N|M)$
supermatrix model with the potential $W(x)$ is converted into the
$(N|M)$-point function for $\mathrm{U}(p|q)$ supermatrix model with another
potential $\tilde{W}(x)$, which is obtained through Fourier transformation:
\begin{align}
&
\Psi_{N,M;\,p,q}
( \{a_i\}_{i=1}^N, \{b_j\}_{j=1}^M;
\{\lambda_\alpha\}_{\alpha=1}^p, \{\mu_\beta\}_{\beta=1}^q )
\nonumber \\
& \hspace{3em} \stackrel{\mbox{\scriptsize F.T.}}{=}
{\Psi}_{p,q;N,M}
( \{\lambda_\alpha\}_{\alpha=1}^p, \{\mu_\beta\}_{\beta=1}^q;
\{a_i\}_{i=1}^N, \{b_j\}_{j=1}^M )
\, .
\label{BH-duality}
\end{align}
In this sense, it is easier to see this duality especially for the
Gaussian matrix model, because the Gaussian function is self-dual with
respect to Fourier transformation.
We remark that this kind of duality can be also found in the Chern--Simons
matrix model and the ABJM matrix model, which are obtained by replacing
the Vandermonde determinant with the exponentiated one,
\begin{equation}
\Delta(x) =
\prod_{i<j}^N (x_i - x_j)
\ \longrightarrow \
\prod_{i<j}^N \left( 2 \sinh \frac{x_i - x_j}{2} \right)
\, .
\end{equation}
In this case insertion of an external source implies the
expectation value of the Wilson loop operator, which is characterized by
Schur function~\cite{Eynard:2014rba}.
To see this duality more explicitly, we now introduce another
correlation function by multiplying the weight functions
\begin{align}
&
\Phi_{N,M;\,p,q}
( \{a_i\}_{i=1}^N, \{b_j\}_{j=1}^M;
\{\lambda_\alpha\}_{\alpha=1}^p, \{\mu_\beta\}_{\beta=1}^q )
\nonumber \\
& \hspace{3em}
=
\Psi_{N,M;\,p,q}
( \{a_i\}_{i=1}^N, \{b_j\}_{j=1}^M;
\{\lambda_\alpha\}_{\alpha=1}^p, \{\mu_\beta\}_{\beta=1}^q )
\prod_{\alpha=1}^p e^{-\frac{1}{\hbar} W(\lambda_\alpha)}
\prod_{\beta=1}^q e^{\frac{1}{\hbar} W(\mu_\beta)}
\, .
\end{align}
This can be also written in the determinantal form
(\ref{SChPoly_det01}), but by replacing the auxiliary functions
as follows,
\begin{align}
&
P_k(\lambda)
\ \to \
\lambda^k \, e^{\mp \frac{1}{\hbar} W(\lambda)}
\, , \quad
Q_k(a)
\ \to \
\int dx \, x^k \, e^{\mp \frac{1}{\hbar} W(x) \pm xa}
\, , \label{PQ-func} \\
&
R(a;b)
\ \to \
\int dx dy \,
\frac{1}{x-y} \,
e^{-\frac{1}{\hbar} W(x) + \frac{1}{\hbar} W(y) + xa - yb}
\, , \quad
\tilde{R}(\lambda;\mu)
\ \to \
\frac{1}{\lambda - \mu} \,
e^{-\frac{1}{\hbar} W(\lambda) + \frac{1}{\hbar} W(\mu)}
\, , \\
&
S_{\rm L}(a;\mu)
\ \to \
\int dx \,
\frac{1}{x-\mu} \,
e^{-\frac{1}{\hbar} W(x) + \frac{1}{\hbar} W(\mu) + xa}
\, , \quad
S_{\rm R}(\lambda;b)
\ \to \
\int dy \,
\frac{1}{\lambda - y} \,
e^{-\frac{1}{\hbar} W(\lambda) + \frac{1}{\hbar} W(y) - yb}
\, .
\end{align}
The sign factor in (\ref{PQ-func}) depends on whether $M + p > N + q$ or
$M + p < N + q$.
In this case it is easier to see the relation between auxiliary
functions, which is summarized in Fig.~\ref{FT_flow}.
\begin{figure}[t]
\begin{center}
\begin{tikzpicture}
\node (0) at (0,0)
{$\displaystyle
\frac{ e^{ -\frac{1}{\hbar} W(\lambda) + \frac{1}{\hbar} W(\mu)} }
{\lambda - \mu}$};
\node (L) at (-4,-1.7)
{$\displaystyle S_{\rm L}(a;\mu)$};
\node (R) at (4,-1.7)
{$\displaystyle S_{\rm R}(\lambda;b)$};
\node (LR) at (0,-3.2)
{$\displaystyle R(a;b)$};
\node (P) at (7.5,0)
{$\displaystyle P_k(\lambda)$};
\node (Q) at (7.5,-3.2)
{$\displaystyle Q_k(a)$};
\draw [thick] (-1.5,-.7) rectangle (1.5,.7);
\draw [thick] (-4.8,-2.2) rectangle (-3.2,-1.2);
\draw [thick] (4.8,-2.2) rectangle (3.2,-1.2);
\draw [thick] (-1.2,-3.7) rectangle (1.2,-2.7);
\draw [thick] (6.8,-0.5) rectangle (8.2,0.5);
\draw [thick] (6.8,-3.7) rectangle (8.2,-2.7);
\draw [very thick,->] (-1.7,-0.5) -- (-3,-1.3);
\draw [very thick,->] (1.7,-0.5) -- (3,-1.3);
\draw [very thick,->] (-3,-2.1) -- (-1.7,-2.8);
\draw [very thick,->] (3,-2.1) -- (1.7,-2.8);
\draw [very thick,->] (7.5,-0.7) -- (7.5,-2.5);
\node (FT_L1) at (-2.7,-.5) {FT$_\lambda$};
\node (FT_L2) at (-2.7,-2.9) {FT$_\mu$};
\node (FT_R1) at (2.7,-.5) {FT$_\mu$};
\node (FT_R2) at (2.7,-2.9) {FT$_\lambda$};
\node (FT) at (8.2,-1.6) {FT$_\lambda$};
\end{tikzpicture}
\end{center}
\caption{{\em Fourier transform web}: the relation between the
auxiliary functions. FT$_x$ means Fourier transformation with respect to
$x$ variable.
The duality (\ref{BH-duality}) is seen as the double Fourier
transformation with $\lambda$ and $\mu$ variables.
}
\label{FT_flow}
\end{figure}
\section{Integrable equations}\label{sec:integrable}
In the previous sections, we have shown that the partition functions of
the Hermitian supermatrix models have the determinantal formula, and
such a determinant structure plays an important role in the relation to
the integrable systems~\cite{Morozov:1994hh}:
it behaves as a $\tau$-function for the corresponding
integrable equation.
In this section we study the connection between the supermatrix
partition function and the integrable system.
\subsection*{Supermatrix model with external source}
We now show that the supermatrix integral involving the external source
(\ref{SMM02}) can be interpreted as the $\tau$-function for the Toda
lattice equations.
In this paper we especially derive the integrable equation, obtained by
taking all the external source parameters to the same value.
We first split external source parameters into the center of mass and
relative parts
\begin{equation}
a_i = \delta a_i + a \, , \qquad
b_j = \delta b_j + b \, , \qquad
\begin{cases}
i = 1, \cdots, N \\ j = 1, \cdots, M
\end{cases}
\, .
\end{equation}
With this separation the determinant (\ref{Cauchy_det}) is given by
\begin{equation}
\Delta_{N,M}(a;b)
=
\Delta_N (\delta a) \Delta_M (\delta b)
\prod_{i=1}^N \prod_{j=1}^M
(\delta a_i - \delta b_j + a - b)^{-1}
\, .
\end{equation}
Let us then consider the Taylor expansion around the center of mass for
the functions (\ref{Q-func}) and (\ref{R-func}),
\begin{align}
Q_{k-1}(a_i)
& =
\sum_{l=1}^\infty \frac{(\delta a_i)^{l-1}}{(l-1)!} \, Q_{k+l-2}(a)
\, , \\
R(a_i;b_j)
& =
\sum_{l,m=1}^\infty
\frac{(\delta a_i)^{l-1}}{(l-1)!}
\frac{(\delta b_j)^{m-1}}{(m-1)!} \,
R^{(l-1,m-1)}(a;b)
\, ,
\end{align}
with
\begin{equation}
R^{(i,j)}(a;b)
=
\frac{\partial^{i}}{\partial a^i}
\frac{\partial^j}{\partial b^j} R(a;b)
\, .
\end{equation}
By taking the limit $\delta a_i = \delta b_j = 0$ for $\forall i,
j$, the determinant in (\ref{SMM02}) is written
\begin{align}
\det
\left(
\begin{array}{c}
Q_{k-1} (a_i) \\ R(a_i; b_j)
\end{array}
\right)
=
\det_{1 \le i, j \le N} \frac{\left(\delta a_i\right)^{j-1}}{(j-1)!}
\cdot
\det_{1 \le i, j \le M} \frac{\left(\delta b_i\right)^{j-1}}{(j-1)!}
\cdot
\det
\left(
\begin{array}{c}
Q_{i+k-2} (a) \\ R^{(i-1,j-1)} (a; b)
\end{array}
\right)
\, ,
\end{align}
where the range of indices is given by $i=1, \cdots, N$, $j = 1,
\cdots, M$, and $k = 1, \cdots, N - M$ in the determinants of the LHS
and the third one in the RHS.
Therefore, after some cancellation, we obtain the following formula
\begin{align}
&
\mathcal{Z}_{N,M} (\{a_i=a\}_{i=1}^N,\{b_j=b\}_{j=1}^M)
=
c_{N,M} \, \tilde{\mathcal{Z}}_{N,M}(a,b) \, , \\
&
\tilde{\mathcal{Z}}_{N,M}(a,b)
=
\det \left(
\begin{array}{c}
Q_{i+k-2}(a) \\ R^{(i-1,j-1)}(a;b)
\end{array}
\right)
\, ,
\label{SMM03}
\end{align}
with the constant
\begin{equation}
c_{N,M} =
(a-b)^{NM}
G(N+1)^{-1} G(M+1)^{-1}
\, ,
\end{equation}
where $\displaystyle G(n) = \prod_{i=0}^{n-2} i! = \prod_{i=1}^{n-1}
\Gamma(i)$ is the Barnes G-function.
We remark that this factor corresponds to the volume element for $\mathrm{U}(N)$ and
$\mathrm{U}(M)$ groups.
This shows that the supermatrix integral (\ref{SMM01}) is finally written as a
kind of Wronskian in the equal parameter limit, $\delta a_i = \delta b_j
= 0$.
In order to derive the integrable equation for the supermatrix integral, we
now apply the Jacobi identity for a determinant
\begin{equation}
D \cdot D \left(
\begin{array}{cc}
i & j \\
k & l \\
\end{array}
\right)
=
D \left(
\begin{array}{c}
i \\ k
\end{array}
\right)
\cdot
D \left(
\begin{array}{c}
j \\ l
\end{array}
\right)
-
D \left(
\begin{array}{c}
i \\ l
\end{array}
\right)
\cdot
D \left(
\begin{array}{c}
j \\ k
\end{array}
\right)
\, ,
\label{Jacobi_id}
\end{equation}
where $D$ is a size $n$ determinant, and the size $n-1$ minor determinant
$D \left(\begin{array}{c} i \\ j \end{array} \right)$ is obtained by
removing $i$-th row and $j$-th column from the matrix.
Similarly
$D \left(\begin{array}{cc} i & j \\ k & l \\ \end{array} \right)$ is the
size $n-2$ determinant obtained by
getting rid of $i, j$-th row and $k,l$-th column.
Setting $i=k=N$ and $j=l=N-1$, we have
\begin{equation}
\tilde{\mathcal{Z}}_{N,M} \cdot \tilde{\mathcal{Z}}_{N-2,M-2}
=
\tilde{\mathcal{Z}}_{N-1,M-1} \cdot \partial_a \partial_b \tilde{\mathcal{Z}}_{N-1,M-1}
- \partial_a \tilde{\mathcal{Z}}_{N-1,M-1} \cdot \partial_b \tilde{\mathcal{Z}}_{N-1,M-1}
\, ,
\end{equation}
and by putting $i=k=N-M$ and $j=l=N-M-1$, we obtain
\begin{equation}
\tilde{\mathcal{Z}}_{N,M} \cdot \tilde{\mathcal{Z}}_{N-2,M}
=
\tilde{\mathcal{Z}}_{N-1,M} \cdot \partial^2_a \tilde{\mathcal{Z}}_{N-1,M}
- \left( \partial_a \tilde{\mathcal{Z}}_{N-1,M} \right)^2
\, .
\end{equation}
These yield the Toda lattice equations~\cite{Hirota:2004}:
\begin{itemize}
\item 2D Toda equation:
\begin{equation}
\frac{\tilde{\mathcal{Z}}_{N+1,M+1} \cdot \tilde{\mathcal{Z}}_{N-1,M-1}}
{\left( \tilde{\mathcal{Z}}_{N,M} \right)^2}
=
\frac{\partial^2}{\partial a \partial b} \log \tilde{\mathcal{Z}}_{N,M} \, .
\end{equation}
\item 1D Toda equation:
\begin{equation}
\frac{\tilde{\mathcal{Z}}_{N+1,M} \cdot \tilde{\mathcal{Z}}_{N-1,M}}
{\left( \tilde{\mathcal{Z}}_{N,M} \right)^2}
=
\frac{\partial^2}{\partial a^2} \log \tilde{\mathcal{Z}}_{N,M} \, .
\end{equation}
\end{itemize}
This means that the partition function is the $\tau$-function for the
Toda lattice system both in one and two dimensions simultaneously.
This property can be also found in the ordinary Hermitian matrix
model~\cite{Kimura:2014mua}.
\subsection*{Characteristic polynomial with external source}
We then show that the characteristic polynomial average with the
external source (\ref{SChPoly_det01}) similarly obeys the integrable
equations.
In addition to the external source, we split the parameters for
characteristic polynomials,
\begin{equation}
\lambda_\alpha = \delta \lambda_\alpha + \lambda \,, \qquad
\mu_\beta = \delta \mu_\beta + \mu \,, \qquad
\begin{cases}
\alpha = 1, \cdots, p \\
\beta = 1, \cdots, q
\end{cases}
\, .
\end{equation}
As well as the previous case, by considering the Taylor expansion
around the center of mass, we obtain the Wronskian-type determinantal
formula in the equal parameter limit $\delta a_i = \delta b_j = \delta
\lambda_\alpha = \delta \mu_\beta = 0$ for $\forall i, j, \alpha, \beta$,
\begin{align}
\Psi_{N,M;\,p,q} (a,b;\lambda,\mu)
& =
c_{N,M;\,p,q} \,
\tilde \Psi_{N,M;\,p,q} (a,b;\lambda,\mu)
\, , \\
\tilde \Psi_{N,M;\,p,q} (a,b;\lambda,\mu)
& =
\det
\left(
\begin{array}{cc}
Q_{i+k-2}(a) & P_{k-1}^{(\alpha-1)} (\lambda) \\
R^{(i-1,j-1)}(a;b) & {S}_{\rm L}^{(\alpha-1,j-1)} (\lambda;b) \\
S_{\rm R}^{(\beta-1,i-1)} (a;\mu) &
\tilde{R}^{(\alpha-1,\beta-1)}(\lambda;\mu) \\
\end{array}
\right) \, ,
\label{SChPoly_det02}
\end{align}
with the constant
\begin{equation}
c_{N,M;\,p,q} =
\frac{(a-b)^{NM} (\lambda-\mu)^{pq}}
{G(N+1) G(M+1) G(p+1) G(q+1)} \, .
\end{equation}
Here $\displaystyle P_{k}^{(\alpha)}(\lambda) =
\frac{d^{\alpha}}{d\lambda^{\alpha}} P_{k}(\lambda)$,
$\displaystyle S_{\rm L}^{(\alpha,j)}(\lambda;b) =
\frac{\partial^\alpha}{\partial\lambda^\alpha}
\frac{\partial^j}{\partial b^j}S_{\rm L}(\lambda;b)$,
and so on.
In this case, since the determinantal formula (\ref{SChPoly_det02})
consists of six blocks, we correspondingly obtain six equations by
applying the identity (\ref{Jacobi_id}):
\begin{align}
\tilde\Psi_{N,M;\,p,q} \cdot \tilde\Psi_{N,M;\,p-2,q-2}
& =
\tilde\Psi_{N,M;\,p-1,q-1} \cdot
\partial_\lambda \partial_\mu\tilde\Psi_{N,M;\,p-1,q-1}
- \partial_\lambda \tilde\Psi_{N,M;\,p-1,q-1} \cdot
\partial_\mu\tilde\Psi_{N,M;\,p-1,q-1} \, ,
\label{eq1} \\
\tilde\Psi_{N,M;\,p,q} \cdot \tilde\Psi_{N-2,M;\,p,q-2}
& =
\tilde\Psi_{N-1,M;\,p,q-1} \cdot
\partial_a \partial_\mu \tilde\Psi_{N-1,M;\,p,q-1}
- \partial_a \tilde\Psi_{N-1,M;\,p,q-1} \cdot
\partial_\mu \tilde\Psi_{N-1,M;\,p,q-1}
\, , \\
\tilde\Psi_{N,M;\,p,q} \cdot \tilde\Psi_{N,M-2;\,p-2,q}
& =
\tilde\Psi_{N,M-1;\,p-1,q} \cdot
\partial_\lambda \partial_b \tilde\Psi_{N,M-1;\,p-1,q}
- \partial_\lambda \tilde\Psi_{N,M-1;\,p-1,q} \cdot
\partial_b \tilde\Psi_{N,M-1;\,p-1,q}
\, , \\
\tilde\Psi_{N,M;\,p,q} \cdot \tilde\Psi_{N-2,M-2;\,p,q}
& =
\tilde\Psi_{N-1,M-1;\,p,q} \cdot
\partial_a \partial_b \tilde\Psi_{N-1,M-1;\,p,q}
- \partial_a \tilde\Psi_{N-1,M-1;\,p,q} \cdot
\partial_b \tilde\Psi_{N-1,M-1;\,p,q}
\, , \\
\tilde\Psi_{N,M;\,p,q} \cdot \tilde\Psi_{N,M;\,p-2,q}
& =
(p-1)
\left(
\tilde\Psi_{N,M;\,p-1,q} \cdot
\partial_a \partial_\lambda \lambda \tilde\Psi_{N,M;\,p-1,q}
- \partial_\lambda \tilde\Psi_{N,M;\,p-1,q} \cdot
\lambda \partial_a \tilde\Psi_{N,M;\,p-1,q}
\right)
\, , \\
\tilde\Psi_{N,M;\,p,q} \cdot \tilde\Psi_{N-2,M;\,p,q}
& =
p \left(
\tilde\Psi_{N-1,M;\,p,q} \cdot
\lambda \partial^2_a \tilde\Psi_{N,M;\,p-1,q}
- \partial_a \tilde\Psi_{N,M;\,p-1,q} \cdot
\lambda \partial_a \tilde\Psi_{N,M;\,p-1,q}
\right)
\, .
\label{eq6}
\end{align}
We have used the following identity to derive some of these
equations
\begin{eqnarray}
\det \left(
\begin{array}{ccc}
(x^N)^{(0)} & \cdots & (x^{N})^{(M-1)} \\
\vdots & & \vdots \\
(x^{N+M-2})^{(0)} & \cdots & (x^{N+M-2})^{(M-1)} \\
(x^{N+M})^{(0)} & \cdots & (x^{N+M})^{(M-1)} \\
\end{array}
\right)
& = & M \, x \, \det_{1 \le i, j \le M} (x^{N+i-1})^{(j-1)}
\, ,
\nonumber \\
\end{eqnarray}
where we denote $\displaystyle (x^j)^{(k)} = \frac{d^k}{dx^k} x^j$.
From the first equation (\ref{eq1}), for example, we obtain the
two-dimensional Toda lattice equation
\begin{equation}
\frac{\tilde{\Psi}_{N,M;\,p+1,q-1} \cdot \tilde{\Psi}_{N,M;\,p-1,q-1}}
{\tilde{\Psi}_{N,M;\,p,q}}
=
\frac{\partial}{\partial \lambda} \frac{\partial}{\partial \mu}
\log \tilde{\Psi}_{N,M;\,p,q}
\, .
\end{equation}
We can similarly obtain the two-dimensional integrable equations from
others, except for (\ref{eq6}), which leads to the one-dimensional Toda
lattice equation.
Therefore, in this sense, the characteristic polynomial average with the
supermatrix model plays a role of the $\tau$-function for the Toda
lattice system.
Let us comment on higher order integrable equations, which the matrix
partition function would satisfy under generic parametrization of the source.
In order to derive such equations it is convenient to introduce the Miwa
coordinate to the time variables,
\begin{equation}
t_n(A) = \frac{1}{n} \mathrm{Tr}\, A^{-n} \, .
\end{equation}
In this sense, since there are two and four kinds of source parameters
in (\ref{SMM03}) and (\ref{SChPoly_det02}), we can correspondingly apply
two and four series of the time variables, respectively.
Although these variables are decoupled in the equal parameter limit, and
we obtain the well-known Toda lattice equations, they might interact with
each other when we consider more generic parametrization.
If so, it would be interesting to study such a situation providing more
involved integrable equations.
\section{Discussion}
In this paper we have studied the supermatrix model involving the
arbitrary potential function $W(x)$ with the external source term.
We have derived the determinantal formula for the partition function and
the characteristic polynomial expectation value with this supermatrix model.
Based on this formula, we have exhibited the duality between the external
source and the characteristic polynomial, and then pointed out that this
duality is just interpreted as Fourier transformation.
We have also shown that the partition function and the characteristic
polynomial average satisfy the Toda lattice equation both in one and two
dimensions, especially in the equal parameter limit.
This implies that we can obtain the corresponding $\tau$-function as
the supermatrix integral with an external source as well as the ordinary
Hermitian matrix model.
As pointed out in Sec.~\ref{sec:ch_poly}, we can obtain various
situations from the supermatrix model by taking the corresponding limit.
From this point of view, it is interesting to study not only simple
reductions of the characteristic polynomial average (\ref{EV_ratio01}),
e.g., $M = 0$, $q=0$, but also the analytic continuation to negative numbers.
Actually the $\mathrm{U}(N|M)$ supermatrix integral is, at least perturbatively,
equivalent to non-supersymmetric matrix model through the analytic
continuation $M \to -M$, but with the two-cut solution, which breaks the
original symmetry, $\mathrm{U}(N+M) \to \mathrm{U}(N) \times \mathrm{U}(M)$.
This kind of equivalence has been recently gathering a great deal of
attention in the area of string/M-theory~\cite{Marino:2009jd}.
For example, it would be interesting to see how the duality relation is
affected by such an analytic continuation.
It is expected that we can obtain more various kinds of correlation
functions, and all the reductions could
preserve the determinantal
structure~\cite{Strahov:2002zu,Bergere:2009zm}, through the analytic
continuation.
From the random matrix theoretical point of view, it is interesting to
extend the present result to other situations, e.g., $\mathrm{OSp}(N|M)$
supermatrix model, chiral ensemble, complex matrix model, two-matrix
model, and so on.
To investigate the matrix model with external source, the
HCIZ formula plays an important role to
integrate out the angular part of the matrix.
Thus it seems better to start with deriving the corresponding formula,
for example, by considering the differential equation for the matrix integral.
\subsection*{Acknowledgements}
We would like to thank B. Eynard for the collaboration on related subjects.
We are also grateful to M. Berg\`ere and E. Br\'ezin for correspondence.
This work is supported in part by Grant-in-Aid for JSPS
Fellows~(\#25-4302).
\bibliographystyle{ytphys}
|
\section{Intoduction}
The discovery of Seiberg-Witten theory \cite{SW1,SW2} and related integrable systems \cite{GKMMM} in the mid-nineties gave rise to many new insights in the theory of integrable systems. In particular it led to a new understanding \cite{BMMM'2000,FGNR,MM,MM2,GM,BGOR,AMMZ} of the classical $p-q$ duality \cite{Ruj'87,Ruijs} of the Calogero-Ruijsenaars family \cite{GM} via its relation with the Seiberg-Witten construction of the low-energy limit of ${\cal N}=2$ SUSY gauge theories with adjoint matter hypermultiplets: the $4d$ theory is described by the elliptic Calogero-Moser system (the elliptic modulus being associated in physical theory with the bare coupling constant) \cite{DW,IM'1996,IM'1997}, the $5d$ theory (with one compactified Kaluza-Klein dimension) is described by the elliptic Ruijsenaars system \cite{BMMM'1999} and the $6d$ theory (with two compactified Kaluza-Klein dimensions) is described by the double-elliptic integrable system (the second elliptic modulus being associated with the compactification torus or an abelian surface) \cite{BMMM'2000,MM,MM2,Braden2003}. All these integrable systems have an elliptic dependence on particle momenta and the most interesting ones are the double-elliptic integrable systems, where both coordinates and momenta have compact values.
The $p-q$ duality admits various deformations. For instance, with a one parametric deformation one can lift it to quantum integrable systems \cite{Ruj'87,Ruijs,EV,FMTV,TV,GM} where the $p-q$ duality can be studied at the level of wave functions; this provides an additional tool to work with. This deformation corresponds to the Nekrasov-Shatashvili limit, $\epsilon_2\to 0$ \cite{NS2009} of the Nekrasov functions with $\epsilon_1$ playing role of Planck's constant.
One may also consider a two-parametric deformation, where in this case one adds to the quantum integrable system its Whitham deformation controlled by the second deformation parameter. The latter is done within the framework of the AGT correspondence
\cite{AGT,Wyl,MM1'2009} and the most effective tools here are matrix models \cite{DV,IMO,EM'2009,EM'2010,SW'2009,MMSh'2009,MMSh'2010}. Note that at the level of AGT
with one of the deformation parameters set zero (so corresponding to a quantum integrable system \cite{NS,MM2'2009,MM3'2009,MT,MMM}), the $p-q$ duality
has quite an unexpected avatar: it is the spectral duality that describes a duality between the $SL(N)$ spin chain and the Gaudin model
with $N+2$ marked points (these two systems describe the two sides of the AGT correspondence) \cite{MMZZ13,MMRZZ13}. The spectral duality is lifted to $5d$ where it describes a duality of the $SL(N)$ spin chain on $M+2$ sites and the $SL(M)$ spin chain on $N+2$ sites \cite{MMRZZ34,GK2013}. Moreover, one can consider the full two-parametric AGT at this $5d$ level \cite{AHKSh,MMZ} where the spectral duality has a clear brane-picture interpretation.
However, an exact correspondence between $p-q$ duality and spectral duality has yet to be established. Moreover,
in contrast to the above cases, the probable generalizations of spectral duality to $6d$ systems are not known even at the classical level, while the $p-q$ duality corresponding to $6d$ systems is known, as explained above, and is described by the self-duality of the double-elliptic systems.
In this paper we continue the recent study \cite{AMMZ} of the double elliptic systems at the classical level. Our goal is to derive the theta-constant equations for the period matrix of the Seiberg-Witten curve of the double-elliptic integrable system. The main tool is the involutivity condition for the $N$-particle Hamiltonians constructed in \cite{BMMM'2000,MM,Braden2003} with respect to the Poisson bracket introduced in \cite{AMMZ}. Initially the Hamiltonians for the integrable systems under consideration were defined in \cite{BMMM'2000,MM} in the form of the ratios of theta-functions on Abelian varieties associated with the Jacobians of the corresponding Seiberg-Witten curves (see below). The hypothesis of \cite{MM} was that the Hamiltonians are Poisson commuting with respect to the Seiberg-Witten symplectic structure. Following this idea a new approach, which deals with arbitrary Riemann surfaces instead of the Seiberg-Witten curves was proposed in \cite{AMMZ}. Within this approach the concept of the Poisson bracket in terms of the coordinates on the Jacobian and the elements of the period matrix was introduced. An advantage of this approach is that it could lead to integrable systems not necessarily associated with the Seiberg-Witten curves. Indeed,
it was demonstrated in \cite{AMMZ} that the Poisson commutativity of the three-particle Hamiltonians is related just to some new theta-function identities of genus two, not making uses of the Seiberg-Witten curve. In this paper, we extend the approach to the case of $N$-particle systems with $N>3$. When $N=4$ we describe evidence that the four-particle Hamiltonians are still in involution for an arbitrary Riemann surface of genus \eqs{g=3}. However, in the general case when \eqs{N>4} the involutivity condition places restrictions on the period matrix and we find these satisfied for the special period matrices corresponding to the Seiberg-Witten curve of the double-elliptic integrable system, in accordance with the original expectation in \cite{MM}. Thus, the involutivity condition can be used to determine the dependence of the period matrix on the Seiberg-Witten flat moduli, providing an independent method for calculating the Seiberg-Witten prepotentials (including their instanton corrections).
In Section \ref{sec:Setting} we recall the definition of the $N$-particle Hamiltonians \cite{BMMM'2000,MM,Braden2003} and introduce the involutivity condition with respect to the Poisson bracket from \cite{AMMZ}. In Section \ref{sec:linProb} the involutivity condition is reformulated as a linear problem in terms of the vectors in a linear space of weight $3$ theta functions. In Section \ref{sec:example} we prove the theta-function identities that lie behind the Poisson commutativity of the three-particle Hamiltonians. In Section \ref{sec:SWrep} the method for calculating the Seiberg-Witten prepotentials of the double-elliptic integrable systems is presented. To illustrate the method we derive the Seiberg-Witten perturbative prepotential in the case of $N=5$.
\section{The Setting}
\label{sec:Setting}
Motivated by considerations of duality and the known rational, trigonometric and elliptic dependencies of the Calogero-Moser system the work of \cite{BMMM'2000} sought to construct integrable systems with compact momentum dependence, extending the rational and hyperbolic dependence of momentum of the Calogero-Moser and Ruijsenaars-Schneider systems respectively. A 2-particle system that reproduced the Ruijsenaars-Schneider (and so also the Calogero-Moser) system as a limit was constructed in which the
reduced momentum had elliptic dependence. That momentum dependence led to this class of models being called `double-elliptic', though actually the momenta posited in the paper more generally took values in some abelian variety $\mathfrak{A}$ (possibly a Jacobian) and the Hamiltonians were argued to be ratios of
theta functions, $\theta_a({\bf{z}}\,|\,\Omega)/\theta_b({\bf{z}}\,|\,\Omega)$ where $\bf{z}$ is the
momenta and $\Omega$ the period matrix of $\mathfrak{A}$ . The relevant theta functions were introduced by analogy to those of the $N$-particle elliptic Calogero-Moser system. There the genus $N$ spectral curve $\mathcal{C}$ is an $N$-fold covering of an elliptic curve $\mathcal{E}$ and the Jacobian $\mathop{\rm Jac}\nolimits(\mathcal{C})$ is isogenous to an Abelian variety of the form $\mathop{\rm Jac}\nolimits(\mathcal{E})\oplus \mathfrak{A}$. If $T$ is the period matrix of $\mathcal{C}$ the general theory of coverings enables us to write\footnote{Here the Riemann $\theta$-function with
characteristics $\boldsymbol{a},\boldsymbol{b}\in\mathbb{Q}$ and $g\times g$ period matrix $T$ is
$$\theta\left[\begin{matrix}\boldsymbol{a}
\\ \boldsymbol{b}\end{matrix}\right](\boldsymbol{\hat z}\,|\,T)=
\sum_{\boldsymbol{n}\in\mathbb{Z}^g}\mathrm{exp}
\left\{\imath\pi(\boldsymbol{n}+\boldsymbol{a})^T T
(\boldsymbol{n}+\boldsymbol{a})
+2\imath\pi
(\boldsymbol{n}+\boldsymbol{a})^T(\boldsymbol{\hat z}+\boldsymbol{b})
\right\}.
$$}
\begin{equation}
\label{thetadecomp}
\Theta(\mathbf{\hat z}| T)
=\sum_{\vec{\alpha}\in\mathbb{Z}^{N-1}/N\mathbb{Z}^{N-1}}
\theta\bsq{\at{-\frac1N\sum_{j=1}^{N}\alpha_{j}}{0}}\brc{z|\,N\tau}\,
\theta\bsq{\at{\frac{\vec{\alpha}}N}{\textbf{0}}}\brc{{\mathbf{z}}|\,\Omega}.
\end{equation}
where $(z, {\mathbf{z}})\sp{T}=M\mathbf{\hat z}\sp{T}$ for an appropriate $M$ separating out the centre of mass mode $z$ and $MTM\sp{T}=\begin{pmatrix}N\tau&0\\0&\Omega\end{pmatrix}$. Such decompositions are far from unique, for example the identity (here $e(x):=e\sp{2 i\pi x}$)
\begin{equation}
\label{eq:trans}
{\theta\bsq{\at{\textbf{0}}{\textbf{b}/l}}(\textbf{z}|l^{-1}\Omega)= \sum_{0\leq a_i<l} \textrm{e}\brc{\dfrac{\textbf{a}\cdot\textbf{b}}{l}}\theta\bsq{\at{\textbf{a}/l}{\textbf{0}}}(l\textbf{z}|l\Omega),\quad 0\leq b_i<l}
\end{equation}
with inverse
\begin{equation}
\theta\bsq{\at{\textbf{a}/l}{\textbf{0}}}(l\textbf{z}|l\Omega)= \dfrac1{l^g}\sum_{0\leq b_i<l} \textrm{e}\brc{-\dfrac{\textbf{a}\cdot\textbf{b}}{l}} \theta\bsq{\at{\textbf{0}}{\textbf{b}/l}}(\textbf{z}|l^{-1}\Omega),\quad 0\leq a_i<l
\end{equation}
lead to alternative expressions and we write these simply as
\begin{equation}
\Theta(\mathbf{\hat z}| T)
=\sum_a \theta_a\brc{z|\,N\tau}\, \theta_a
\end{equation}
specifying $\theta_a:=\theta_a({\mathbf{z}}|\,\Omega)$ as required. Although \cite{BMMM'2000} suggested the form of the Hamiltonians of the dual system, the exact nature of the $\theta_a$ was left unspecified (for $N>2$) and neither their Poisson commutativity nor their geometric setting was addressed at that time. The Seiberg-Witten picture tells us that $\Omega$ depends on the conjugate coordinates, but the dependence is left unspecified.
Before turning to the issue of Poisson commutativity we recall that Braden and Hollowood have given a geometric setting for such Hamiltonians \cite{Braden2003}. There the spectral curve is of genus $N+1$ and lies in a $(1,N)$-polarized abelian surface and is given by an equation of the form
\begin{equation}
0=\sum_{a=0}\sp{N-1}\Theta\left[\begin{matrix} 0&\frac{a}{N}\\ 0&0\end{matrix}\right]( z_1,z_2\,|\, \Gamma)\, \theta_a, \qquad
\theta_a=\sum_{{\vec{\alpha}\in\mathbb{Z}^{N-1}/N\mathbb{Z}^{N-1}} \atop {a+\sum_j \alpha_j\in N\mathbb{Z} }}
\theta\bsq{\at{\frac{\vec{\alpha}}N}{\textbf{0}}}\brc{{\mathbf{z}}|\,\Omega}.
\end{equation}
Here $\Theta$ is the theta function of the abelian surface and
$\theta_a$ are of the form (\ref{thetadecomp}). Degenerations of this give precisely the Ruijsenaars and elliptic Calogero-Moser systems.
Regarding the Poisson commutativity of the ratios of theta functions (Hamiltonians) associated with the dual models of \cite{BMMM'2000}, significant evidence for this was amassed in \cite{MM} using a perturbative (in instanton number) expansion.
Here it was observed that resulting equations for commutativity gave relations between the various terms of the instanton expansion of the Seiberg-Witten prepotential associated with $\mathcal{C}$.
(A proof of the Poisson commutativity in the case of the elliptic Calogero-Moser system was later provided in \cite{Marshakov:1999bw} .) Following \cite{BM01} we have that (for all $a,b,c,d$)
\begin{equation}
\label{posscom}
0=\left\{ {\theta_a\over\theta_b},
{\theta_{c}\over\theta_{d}}\right\}
\Longleftrightarrow
0=\sum_{i=1}\sp{N}
\left| \begin{array}{ccc}
\theta_{a}&\dfrac{\partial\theta_{a}}{\partial \hat z_{i}}&\dfrac{\partial\theta_{a}}{\partial \hat a\sp{i}}
\\
\theta_{b}&\dfrac{\partial\theta_{b}}{\partial \hat z_{i}} &\dfrac{\partial\theta_{b}}{\partial \hat a\sp{i}}
\\
\theta_{c}&\dfrac{\partial\theta_{c}}{\partial \hat z_{i}}&\dfrac{\partial\theta_{c}}{\partial \hat a\sp{i}}
\end{array}\right|
=
\sum_{r=1}\sp{N-1}
\left| \begin{array}{ccc}
\theta_{a}&\dfrac{\partial\theta_{a}}{\partial z_{r}}&\dfrac{\partial\theta_{a}}{\partial a\sp{r}}
\\
\theta_{b}&\dfrac{\partial\theta_{b}}{\partial z_{r}} &\dfrac{\partial\theta_{b}}{\partial a\sp{r}}
\\
\theta_{c}&\dfrac{\partial\theta_{c}}{\partial z_{r}}&\dfrac{\partial\theta_{c}}{\partial a\sp{r}}
\end{array}\right|
\end{equation}
where the action $\hat a\sp{i}$ is conjugate to the variable $\hat z_i$ and in obtaining the final equality we
express the simplectic form as
\[ \omega=\sum_{i=1}\sp{N}d \hat z_i \wedge d\hat a\sp{i}= dz\wedge d\tau +\sum_{r=1}\sp{N-1}
dz_r\wedge da\sp{r}
\]
noting that $\theta_a$ is independent of the centre of mass. Geometrically the Seiberg-Witten curve is the fibre over the moduli space of actions: this means the dependence of the theta functions on the action $\hat a\sp{i}$ is via the period matrix $T$ of $\mathcal{C}$ and hence $\Omega$. Upon using the heat equation,
that $(\tau, {\mathbf{a}})\sp{T}=M\sp{-1\, T}\mathbf{\hat a}\sp{T}$, and the prepotential $\mathcal{F}$ we have that
\[ \dfrac{\partial\theta_{a}}{\partial a\sp{r}}= \dfrac{\partial\theta_{a}}{\partial \Omega_{st}}
\dfrac{\partial \Omega_{st}}{\partial a\sp{r}}= \frac1{4i\pi} \dfrac{\partial\theta_{a}}{\partial z_{r}z_s}\dfrac{\partial \Omega_{st}}{\partial a\sp{r}}= \frac1{4i\pi} \dfrac{\partial\theta_{a}}{\partial z_{r}z_s}
\, M_{ri}M_{sj}M_{tk}\, \dfrac{\partial T_{ij} }{\partial \hat a\sp{k}}
=
\frac1{4i\pi} \dfrac{\partial\theta_{a}}{\partial z_{r}\partial z_s}
\, M_{ri}M_{sj}M_{tk}\,
\dfrac{\partial\sp3 \mathcal{F} }{\partial \hat a\sp{i}\partial \hat a\sp{j}\partial \hat a\sp{k}};
\]
thus the Poisson-commutativity reduces to showing that
\begin{equation}
\label{pcomm}
0=
\sum_{r,s,t=1}\sp{N-1}P_{rst} H_{rst}\sp{abc},
\end{equation}
where $P_{rst}$ is totally symmetric (for this example $P_{rst}= M_{ri}M_{sj}M_{tk}\,
{\partial\sp3 \mathcal{F} }/{\partial \hat a\sp{i}\partial \hat a\sp{j}\partial \hat a\sp{k}} $) and
\begin{equation}
\label{eq:Hdet}
H_{rst}\sp{abc}:=
\left| \begin{array}{ccc}
\theta_{a}&\dfrac{\partial\theta_{a}}{\partial z_{r}}&\dfrac{\partial\theta_{a}}{\partial z_{s}\partial z_t}\\
\theta_{b}&\dfrac{\partial\theta_{b}}{\partial z_{r}} &\dfrac{\partial\theta_{b}}{\partial z_{s}\partial z_t}\\
\theta_{c}&\dfrac{\partial\theta_{c}}{\partial z_{r}}&\dfrac{\partial\theta_{c}}{\partial z_{s}\partial z_t}\end{array}\right| .
\end{equation}
As noted above, this Poisson commutativity has been established for systems arising as degenerations of the elliptic Calogero-Moser system.
In the works just described we have utilised properties of the underlying spectral curve $\mathcal{C}$ to construct examples of double-elliptic systems. The recent paper \cite{AMMZ} goes beyond this. Let us henceforth assume Hamiltonians of the form ($ a=1,\dots,N-1$)
\begin{equation}
\label{eq:NHam}
{H_a\brc{\textbf{z}\,|\,\Omega}=\frac{\displaystyle\theta\bsq{\at{0\;\dots\;0}{\frac{a}N\dots\frac{a}N}} \brc{\textbf{z}\,|\,\Omega}}{\theta\brc{\textbf{z}\,|\,\Omega}}
:=\frac{\theta_a \brc{\textbf{z}\,|\,\Omega}}{\theta_0\brc{\textbf{z}\,|\,\Omega}} , }
\end{equation}
where $\Omega$ is $(N-1)\times (N-1)$ period matrix. (These are simply related to the Braden-Hollowood Hamiltonians via (\ref{eq:trans})\footnote{The Hamiltonians
$\displaystyle{\wt H_j=\frac{A_{0j}}{A_{00}} }$, $\displaystyle{A_{0j}=\sum_{\bfi{i_k}_j} \theta\bsq{\at{\frac{i_1}N\dots\frac{i_{N-1}}N}{0\,\,\dots\;\;\;\;0}} \brc{{\bf z}\,|\,\Omega}},
$
where \eqs{i_k=0,\dots,N-1} and the elements from \eqs{\bfi{i_k}_j} satisfy
${j+\sum_{k=1}^{N-1}i_k\in N\cdot\mathbb{Z}}$ may be expressed as
$${H_i\brc{{\textbf{z}}\,|\,\Omega}=\frac{\displaystyle\theta\bsq{\at{0\;\dots\;0}{\frac{i}N\dots\frac{i}N}} \brc{{\bf z}/N\,|\,\Omega/N^2}}{\theta\brc{{\bf z}/N\,|\,\Omega/N^2}} =\frac{\displaystyle\sum_{j=0}^{N-1}\textrm{e}\brc{-\frac{i\,j}N}A_{0j}}{\displaystyle\sum_{j=0}^{N-1}A_{0j}},\quad i=1,\dots, N-1.}
$$
}.) Assuming only the Jacobi identity \cite{AMMZ} sought
solutions to (\ref{pcomm}) for genus $2$ theta functions ($N=3$). They discovered that $P_{rst}$
were totally symmetric and their solutions were expressed in terms of theta-function identities.
We reformulate and extend this as follows. Suppose one has a family of abelian varieties for which there is a symplectic structure on the total space with respect to which the abelian varieties are Lagrangian: we have in the above coordinates
\begin{equation}
\label{eq:PS0}
\{z_r,\Omega_{st}\}=P_{rst}(a),\quad
\{z_r,z_s\}=0,\quad \{\Omega_{rs},\Omega_{tu}\}=0,
\end{equation}
where the period matrix $\Omega=\Omega(a)$ are some special functions of the Seiberg-Witten flat moduli $a$. We know from the work of Donagi and Markman \cite{DM96} that (assuming holomorphicity) the differential of the period map at each point is the contraction of a cubic, and so $P_{rst}$ is totally symmetric. (This was a result of \cite{AMMZ} that in the light of \cite{DM96} we assume from the outset.)
What can be said about the solutions to (\ref{pcomm}) for arbitrary $g=N-1$? We will show that for \eqs{g=2} (\ref{pcomm}) holds for an arbitrary symmetric \eqs{ g\times g} period matrix $\Omega$ and are actually
theta-function identities. In the case when \eqs{g\geqslant4} the relations (\ref{pcomm}) define some special constraints on the elements of the period matrix $\Omega$. We will show that these constraints describe the dependence of $\Omega$ on the flat moduli of the corresponding Seiberg-Witten curves. Thus, in the general case, the Poisson commutativity of the Hamiltonians (\ref{eq:NHam}) holds only for some special class of period matrices, as suggested in \cite{BMMM'2000,MM}.
\section{Poisson commutativity as a linear problem}
\label{sec:linProb}
The relations (\ref{pcomm}) are strongly connected with weight $3$ theta functions and their corresponding linear spaces \cite{Mumford}.
Recall that an entire function \eqs{f\brc{\textbf{z}\,|\,\Omega}=f\brc{\textbf{z}}} on \eqs{\mathbb{C}^g} with fixed symmetric period matrix $\Omega$ is called a theta function of weight $\lambda\in \mathbb{N}$ and characteristic $\bsq{\at{\boldsymbol{\delta}}{\boldsymbol{\epsilon}}}$, if
\begin{equation}
f(\mathbf{z}+\mathbf{p}\Omega+\mathbf{q})=e\left( -\frac{\lambda}2 \mathbf{p}\Omega\mathbf{p}-
\lambda \mathbf{p}\cdot\mathbf{z} + \boldsymbol{\delta}\cdot \mathbf{q} -
\boldsymbol{\epsilon}\cdot \mathbf{p})
\right) f(\mathbf{z})
\end{equation}
for all $\mathbf{p},\mathbf{q}\in\mathbb{Z}^g$. Such functions form a linear space
$\Theta^{\Omega}_{\lambda}\bsq{\at{\boldsymbol{\delta}}{\boldsymbol{\epsilon}}}$
of dimension \eqs{\lambda^g} with standard bases \cite{Fay1973}:
\begin{align}
\label{eq:basis1}
(1)&\qquad
\theta
\left[\begin{matrix} \frac{\boldsymbol{\delta}+\boldsymbol{\rho} } {\lambda}\\ \boldsymbol{\epsilon}\end{matrix}\right]
\brc{\lambda\textbf{z}\,|\,\lambda\Omega},\quad 0\leqslant \rho_i<\lambda,\\
\label{eq:basis2}
(2)&\qquad
\theta
\left[\begin{matrix} \boldsymbol{\delta}\\ \frac{\boldsymbol{\epsilon}+\boldsymbol{\rho} }{\lambda} \end{matrix}\right]
\brc{\textbf{z}\,|\,\lambda^{-1}\Omega},\quad 0\leqslant \rho_i<\lambda.
\end{align}
It is convenient to define the general
\begin{equation}
\label{eq:cDet}{H^{\textbf{abc}}_{rst}\brc{\textbf{z}\,|\,\Omega}=\left|\matr{ccc}
{
\theta \bsq{\at{\textbf{a}}{\textbf{a}^{\prime} }} \brc{\textbf{z}\,|\,\Omega} &
\partial_{z_r}\theta \bsq{\at{\textbf{a}}{\textbf{a}^{\prime} }} \brc{\textbf{z}\,|\,\Omega}& \partial_{z_s}\partial_{z_t}\theta \bsq{\at{\textbf{a}}{\textbf{a}^{\prime} }}\brc{\textbf{z}\,|\,\Omega}\\
\\
\theta\bsq{\at{\textbf{b}}{\textbf{b}^{\prime} }}\brc{\textbf{z}\,|\,\Omega} &\partial_{z_r}\theta\bsq{\at{\textbf{b}}{\textbf{b}^{\prime} }}\brc{\textbf{z}\,|\,\Omega}& \partial_{z_s}\partial_{z_t}\theta\bsq{\at{\textbf{b}}{\textbf{b}^{\prime} }}\brc{\textbf{z}\,|\,\Omega}\\
\\
\theta\bsq{\at{\textbf{c}}{\textbf{c}^{\prime} }}\brc{\textbf{z}\,|\,\Omega}&
\partial_{z_r}\theta\bsq{\at{\textbf{c}}{\textbf{c}^{\prime} }}\brc{\textbf{z}\,|\,\Omega}& \partial_{z_s}\partial_{z_t}\theta\bsq{\at{\textbf{c}}{\textbf{c}^{\prime} }}\brc{\textbf{z}\,|\,\Omega}
}\right|,}
\end{equation}
where \eqs{ \textbf{a},\textbf{a}^{\prime} ,\textbf{b},\textbf{b}^{\prime} ,\textbf{c},\textbf{c}^{\prime} \in\mathbb{Q}^g}.
With appropriate choices these will give the determinants (\ref{eq:Hdet}). Now the symmetry of the coefficients $P_{rst}$ means that we
can work with the fully symmetric combinations
\begin{equation}
{H^{\textbf{abc}}_{{\{rst\}}}:=H^{\textbf{abc}}_{rst}+H^{ab}_{str}+H^{\textbf{abc}}_{trs},\quad r\leqslant s\leqslant t.}
\end{equation}
A simple calculation establishes that
\begin{equation}
H^{\textbf{abc}}_{\bfi{rst}}(\mathbf{z}+\mathbf{p}\Omega+\mathbf{q})=e\left( -\frac{3}2\, \mathbf{p}\Omega\mathbf{p}- 3\, \mathbf{p}\cdot\mathbf{z} +
\left[ \textbf{a}+\textbf{b}+\textbf{c}\right]\cdot \mathbf{q} -
\left[ \textbf{a}^{\prime} +\textbf{b}^{\prime} +\textbf{c}^{\prime} \right]\cdot \mathbf{p})
\right)
H^{\textbf{abc}}_{\bfi{rst}}(\mathbf{z})
\end{equation}
and so
\begin{equation}
\label{Hwtheta}
H^{\textbf{abc}}_{\bfi{rst}}(\mathbf{z})\in \Theta^{\Omega}_{3}
\left[\begin{matrix}\textbf{a}+\textbf{b}+\textbf{c} \\ \textbf{a}^{\prime} +\textbf{b}^{\prime} +\textbf{c}^{\prime}.\end{matrix}\right]
\end{equation}
We emphasise that although none of the terms $H^{\textbf{abc}}_{rst}$ individually possess this property the symmetrised sum is a third order theta function. This result has several important consequences. First, let
$\{ f_{\vec{\alpha}}( \textbf{z}) \}$ be any basis for $\Theta^{\Omega}_{3}
\left[\begin{matrix}\textbf{a}+\textbf{b}+\textbf{c} \\ \textbf{a}^{\prime} +\textbf{b}^{\prime} +\textbf{c}^{\prime}.\end{matrix}\right]
$; then we have an expansion
\begin{equation}
\label{Hexpansion}
H^{\textbf{abc}}_{\bfi{rst}}(\mathbf{z})=\sum_{ \vec{\alpha} } C_{\{rst\}}\sp{ \vec{\alpha}}\, f_{\vec{\alpha}}( \textbf{z})
\end{equation}
and (\ref{pcomm}) becomes for each $\vec{\alpha}$
\begin{equation}
\label{eq:g3ThetaId}
0=\sum_{r,s,t=1}\sp{g}P_{rst} \, C_{\{rst\}}\sp{ \vec{\alpha}}.
\end{equation}
This important relation entails several things. First it expresses that the $g\brc{g+1}\brc{g+2}/6$ vectors $C^{\vec{\alpha}}_{\bfi{rst}}$ (each one has \eqs{3^g} coordinates) are linearly dependent. Second, that we have a linear problem to determine the $P_{rst} $'s; and third, that the $P_{rst} $'s will be expressible in terms of the constants $C^{\vec{\alpha}}_{\bfi{rst}}$ if there is a nontrivial solution. Because the $C^{\vec{\alpha}}_{\bfi{rst}}$'s are given in terms of theta-constants, these are the theta-constant identities referred to earlier and generalise those obtained in \cite{AMMZ}. In the next section we shall illustrate this general setting for the Hamiltonians (\ref{eq:NHam}). For ease of description in what follows we will describe $C^{\vec{\alpha}}_{\bfi{rst}}$ as a $3^g \times g\brc{g+1}\brc{g+2}/6$ matrix.
\section{An Example}
\label{sec:example}
We shall now apply the above considerations to the Hamiltonians (\ref{eq:NHam}) and prove a conjecture raised in \cite{AMMZ}. The relevant characteristics are
\[ \bsq{\at{\textbf{a}}{\textbf{a}^{\prime} }}
=\left[ \begin{matrix} 0&\ldots&0\\ a/N&\ldots&a/N \end{matrix} \right], \quad
\bsq{\at{\textbf{b}}{\textbf{b}^{\prime} }}
=\left[ \begin{matrix} 0&\ldots&0\\ b/N&\ldots&b/N \end{matrix} \right], \quad
\bsq{\at{\textbf{c}}{\textbf{c}^{\prime} }}
=\left[ \begin{matrix} 0&\ldots&0\\ 0&\ldots&0 \end{matrix} \right],
\]
and we may obtain the expansion (\ref{Hexpansion}) as follows. The determinants \eqs{H^{\text{ab}}_{rst}} have the following Fourier decomposition:
\eqlb{eq:detFourier}{H^{\text{ab}}_{rst}=(2\pi\imath)^3\sum_{\textbf{n},\textbf{m},\textbf{l}\in\mathbb{Z}^g} \textrm{e}\brc{\frac12\textbf{n}^t\,\Omega\,\textbf{n}+\frac12\textbf{m}^t\,\Omega\,\textbf{m}+ \frac12\textbf{l}^t\,\Omega\,\textbf{l}+\brc{\textbf{n}+\textbf{m}+\textbf{l}}\cdot\textbf{z}}
\textrm{e}\brc{\frac{\textbf{a}\cdot\textbf{m}+\textbf{b}\cdot\textbf{l}}N}\abs{\textbf{n},\textbf{m},\textbf{l}}_{rst},}
where
\eq{\abs{\textbf{n},\textbf{m},\textbf{l}}_{rst}=\left|\matr{ccc}
{1 & n_r & n_s n_t\\
1 & m_r & m_s m_t\\
1 & l_r & l_s l_t
}\right|.}
The symmetrized functions $H^{\text{ab}}_{\bfi{rst}}$ then have an analogous expression in terms of
\eq{\abs{\textbf{n},\textbf{m},\textbf{l}}_{\bfi{rst}}= \abs{\textbf{n},\textbf{m},\textbf{l}}_{rst}+\abs{\textbf{n},\textbf{m},\textbf{l}}_{str} +\abs{\textbf{n},\textbf{m},\textbf{l}}_{trs}.}
That $H^{\textbf{ab}}_{\bfi{rst}}$ is a theta functions of weight $\lambda=3$ is reflected by the
relation:
\eqlb{eq:detrel}{\forall\textbf{p}\in\mathbb{C}^g:\qquad \abs{\textbf{n}+\textbf{p},\,\textbf{m}+\textbf{p},\,\textbf{l}+\textbf{p}}_{\bfi{rst}} =\abs{\textbf{n},\textbf{m},\textbf{l}}_{\bfi{rst}}.}
To obtain the decomposition in the basis (\ref{eq:basis1}) we change the summation variables in (\ref{eq:detFourier}) as follows
\eq{\left\{\matr{l}{
\textbf{n}\;\rightarrow3\textbf{k}+\vec{\alpha}-\textbf{m}-\textbf{l},\\
\textbf{m}\rightarrow\textbf{k}+\textbf{i},\\
\textbf{l}\;\,\,\rightarrow\textbf{k}+\textbf{j},}\right.
\qquad
\textbf{i},\textbf{j},\textbf{k}\in\mathbb{Z}^g,
\quad\textrm{and}\quad
\vec{\alpha}\in\mathbb{Z}^g/3\mathbb{Z}^g.
}
Then, using the relation (\ref{eq:detrel}), we obtain
\eqlb{eq:Hdecomp}{H^{\text{ab}}_{\bfi{rst}}=\sum_{\vec{\alpha}\in\mathbb{Z}^g/3\mathbb{Z}^g}
C^{\vec{\alpha}}_{\bfi{rst}}\,
\,\theta\left[ \begin{matrix} &\vec{\alpha}/3&\\ (a+b)/N&\ldots&(a+b)/N \end{matrix} \right]
\brc{3\textbf{z}\,|\,3\Omega},}
where
\begin{equation}
\begin{split}
{C^{\vec{\alpha}}_{\{rst\}}=(2\pi\imath)^3
\textrm{e}\brc{-\frac{(a+b)\sum_k{\alpha_k}}{3N}}\,
\sum_{\textbf{i},\textbf{j}\in\mathbb{Z}^g}
\textrm{e}\brc{\brc{\textbf{i}+\frac{\textbf{j}-\vec{\alpha}}2}\Omega \brc{\textbf{i}+\frac{\textbf{j}-\vec{\alpha}}2}+
\brc{\frac{\textbf{j}}2-\frac{\vec{\alpha}}6}3\Omega
\brc{\frac{\textbf{j}}2-\frac{\vec{\alpha}}6}}\times
} \\
{\times
\textrm{e}\brc{\frac{\text{a}\sum_k i_k+\text{b}\sum_k j_k}N} \abs{\vec{\alpha}-\textbf{i}-\textbf{j},\,\textbf{i},\,\textbf{j}}_{\{rst\}}}
\end{split}
\end{equation}
or in terms of theta constants (with $\theta'_{r}\brc{0\,|\,\Omega}\equiv \left.\partial_{z_r}
\theta\brc{\textbf{z}\,|\,\Omega}\right|_{\textbf{z}=\textbf{0}}$)
\eqn{C^{\vec{\alpha}}_{\{rst\}}=2\,
\sum_{ \at{\vec{\beta}\in\mathbb{Z}^g/2\mathbb{Z}^g}{r,s,t}}
\left(9\theta'_{r}
\left[ \begin{matrix} &\frac{\vec{\beta}-\vec{\alpha}}2&\\ a/N&\ldots&a/N \end{matrix} \right]
\brc{0\,|\,2\Omega}\,
\theta''_{st}
\bsq{\at{\frac{\vec{\beta}}2-\frac{\vec{\alpha}}6}{\brc{2\text{b}-\text{a}/N \ldots2\text{b}-\text{a}/N }} }
\brc{0\,|\,6\Omega}
\right.}
\eqlb{eq:thetacoord}{\left.-
\theta'''_{rst}
\left[ \begin{matrix} &\frac{\vec{\beta}-\vec{\alpha}}2&\\ a/N&\ldots&a/N \end{matrix} \right]
\brc{0\,|\,2\Omega}\,
\theta
\bsq{\at{\frac{\vec{\beta}}2-\frac{\vec{\alpha}}6}{\brc{2\text{b}-\text{a}/N \ldots2\text{b}-\text{a}/N }} }
\brc{0\,|\,6\Omega}\right).
}
Now when $N=3$ ($g=2$) we have just two Hamiltonians ($a=1$, $b=2$ or vica versa) and the functions \eqs{H^{\text{ab}}_{rst}} are invariant under the transformation
\eq{H^{\text{ab}}_{rst}\brc{\textbf{z}+\frac{\text{k}}3(1,1)\,|\,\Omega}= H^{\text{ab}}_{rst}\brc{\textbf{z}\,|\,\Omega}. }
In terms of the basis elements we have
\eq{\theta\bsq{\at{\vec{\alpha}/3}{\boldsymbol{\epsilon}}}\brc{3\brc{\textbf{z}+\frac{\text{k}}3(1,1) }\,|\,3\Omega}=
\textrm{e}\brc{\frac{ k \sum_l \alpha_l }3}
\theta\bsq{\at{\vec{\alpha}/3}{\boldsymbol{\epsilon}}}\brc{3\textbf{z}\,|\,3\Omega}.}
Hence the only nonzero coefficients in the decomposition (\ref{eq:Hdecomp}) are those \eqs{C^{\vec{\alpha}}_{\bfi{rst}}} for which
\eqlb{eq:symRel}
{\textrm{e}\brc{\frac{ k \sum_l \alpha_l }3}=1
}
so leaving \eqs{3^{g-1}} possible nonzero coordinates in each column of \eqs{C^{\vec{\alpha}}_{\bfi{rst}}}.
Moreover, the symmetry
\eq{H^{\text{ab}}_{rst}\brc{-\textbf{z}\,|\,\Omega}= H^{\text{ab}}_{rst}\brc{\textbf{z}\,|\,\Omega}}
reduces the independent terms further to only
${(3^{g-1}+1)/2}$ different nonzero rows in the matrix \eqs{C^{\vec{\alpha}}_{\bfi{rst}}}.
Now in the present setting we have
$g\brc{g+1}\brc{g+2}/{6}>{(3^{g-1}+1)}/{2}$ and so more variables $P_{\{rst\}}$ than equations, from which we deduce that for arbitrary $\Omega$ equation (\ref{eq:g3ThetaId})
holds for nontrivial $P_{rst}$ and the corresponding $C^{\vec{\alpha}}_{\bfi{rst}}$ are linearly dependent. This establishes the theta-constant relations conjectured
in \cite{AMMZ}.
We remark in passing the above argument actually shows that for each pair of nonzero vectors \eqs{\textbf{a},\textbf{b}\in\mathbb{Z}^g/3\mathbb{Z}^g} with property \eqs{\textbf{a}+\textbf{b}\equiv\textbf{0}\mod 3} there exists a nontrivial set of quantities \eqs{P_{rst}} for which the corresponding Hamiltonians commute for
$g\le4$. This is because $\textbf{z}\rightarrow\textbf{z}+{\textbf{a}}/3$ is still a symmetry of $H^{\textbf{ab}}_{rst}$, now leading to nonzero \eqs{C^{\vec{\alpha}}_{\bfi{rst}}} when $\textrm{e}\brc{{\vec{\alpha}\cdot\textbf{a}}/3}=1$. Together with the restriction coming from parity, we have $\brc{g+1}\brc{g+2}/{6}>{(3^{g-1}+1)}/{2}$
for $g\le4$ and again there are corresponding theta-constant identities. When $\textrm{e}\brc{{\vec{\alpha}\cdot\textbf{a}}/3}\ne 1$ we find the identities $C^{\vec{\alpha}}_{\bfi{rst}}=0$.
\section{Theta-constant representation for the Seiberg-Witten curves}
\label{sec:SWrep}
In this section we shall use the constraint (\ref{pcomm}) to make various deductions about the prepotential.
This work will focus on the perturbative prepotential to establish the method, leaving the instanton corrections to a later work.
It is helpful to isolate the assumptions being made. First we are assuming that there exists an underlying Seiberg-Witten curve $\mathcal{C}$ with period matrix $T$ given by a prepotential $\mathcal{F}$.
Second, we shall assume that the curve $\mathcal{C}$ is of genus $N$ and covers an elliptic curve, and so is related to the elliptic Calogero-Moser family. With a choice of $M$ given by
\eq{\forall i:\quad M_{1i}=M_{i1}=\dfrac{1}{N},\qquad \forall i>1,j>1:\quad M_{ij}=-\dfrac{\delta_{ij}}N.}
we find
\eqlb{eq:OmegaSWN}{\Omega_{ij}=
\delta_{ij}\brc{\frac{g}{N}\tau-\sum_{k\neq i+1}T_{i+1,k}}+
(1-\delta_{ij})\brc{T_{i+1,j+1}-\frac1N\tau}
.}
In passing we note that at this stage one could have chosen to use a different $M$ (for example \cite{BM01})
giving equivalent expansions, or chosen to have the Braden-Hollowood genus $N+1$ curve. Together these
assumptions provide us with an instanton expansion (for \eqs{i\neq j})
\eqlb{eq:SWTdec}{T_{ij}=\frac{\partial^2\mathcal{F}}{\partial\hat a_i\partial\hat a_j}=-\frac{1}{\pi\imath}\ln F^{(0)}_{ij}+\sum_{k\in\mathbb{N}}q^k\frac{\partial^2 F^{(k)}}{\partial\hat a_i\partial\hat a_j},\quad q\equiv \textrm{e}^{2\pi\imath\tau}.}
Here $F^{(0)}$ is the perturbative prepotential and the instanton corrections \eqs{F^{(k)}} to the Seiberg-Witten prepotential $\mathcal{F}$ are only known to low order. We may obtain new information about this expansion as follows. With
$P_{rst}= M_{ri}M_{sj}M_{tk}\, {\partial\sp3 \mathcal{F} }/{\partial \hat a\sp{i}\partial \hat a\sp{j}\partial \hat a\sp{k}} $
the constraints (\ref{eq:g3ThetaId}) with (\ref{eq:thetacoord}) and the instanton expansions of these
constrain the $P_{rst}$'s. Let us write the perturbative expansions (corresponding to the trigonometric limit \eqs{\textrm{Im}\,\tau\rightarrow+\infty}) as \eq{Q_{ij}\equiv\textrm{e}\brc{T_{ij}}= q_{ij}+ \sum_{k\in\mathbb{N}}q^k\,q^{(k)}_{ij},}
\eq{P_{ijk}=p_{ijk}+\sum_{l\in\mathbb{N}}q^{l}\,p^{(l)}_{ijk}.}
The constraints (\ref{eq:g3ThetaId}) become relations between the coefficients of these expansions. Before
illustrating this we note that although the dependence on the coordinates is not a priori known, we know that in the case of the $GL\brc{N}$ systems under consideration (for example, from the Toda limits of \cite{MM})
the quantities $F^{(0)}_{ij}$ from (\ref{eq:SWTdec}) are just functions of $\brc{\hat a_i- \hat a_j}$, the difference of the Seiberg-Witten flat moduli. In general, the variables
${q_{ij}\equiv\brc{F^{(0)}_{ij}}^{-2}}$
may be represented by the following series:
\begin{equation}
\label{eq:qgenDec}
q_{ij}=\sum _{l=0} c_l \left(\hat a_i-\hat a_j\right)^{2l+2}.
\end{equation}
In the previous section we considered the case when $N=3$. When $N=4$ we have $10$ vectors \eqs{C^{\vec{\alpha}}_{\bfi{rst}}} in a $27$ dimensional space. Consider the determinants \eqs{H^{ab}_{rst}} with $a=1$ and $b=3$. Then due to the symmetry
\eq{H^{ab}_{rst}\brc{-z\,|\,\Omega}=H^{ab}_{rst}\brc{z\,|\,\Omega}}
we have only $14$ different rows in the matrix consisting of vectors \eqs{C^{\vec{\alpha}}_{\bfi{rst}}}. The resulting $10\times14$ matrix has rank $9$ due to some theta-constant identities, thus giving the relations (\ref{pcomm}).
We therefore consider the case $N=5$. Here we have $20$ vectors \eqs{C^{\vec{\alpha}}_{\bfi{rst}}} in an $81$ dimensional space.
Again, if we choose the determinants \eqs{H^{ab}_{rst}} with $a=1$ and $b=N-1=4$, we obtain $41$ different rows in the matrix consisting of vectors $C^{\vec{\alpha}}_{\bfi{rst}}$.
Then in order to satisfy the relations (\ref{pcomm}), the rank of this matrix must be at most equal to $19$. The latter gives at least $22$ equations on the elements of the period matrix \eqs{\Omega}. The number $22$ alone does not give useful information, as some of these equations may be equivalent. Now taking
the constraints (\ref{eq:g3ThetaId}) in the case of \eqs{N=5} and the perturbative expansions above, we obtain
the linear system
\eqlb{eq:LinSys}{L\bold{p}=0}
for the first nonzero order of the elements \eqs{p_{ijk}}:
\eq{i\neq j:\quad p_{ijj}=-p_{iij},\qquad
i\neq j\neq k:\quad p_{ijk}=0.}
Here
\begin{equation*}{
\bold{p}^{\textrm{T}}=
\left(\matr{cccccccccc}{
p_{111}& p_{112}& p_{113}& p_{114}& p_{222}& p_{223}& p_{224}& p_{333}& p_{334}& p_{444}}\right)}
\end{equation*}
(the remaining $p$'s appear at higher orders) and
\begin{equation*}
{L=}
{
\left(
\small
\begin{array}{cccccccccc}
q_{13}-q_{23} & 0 & q_{13}-q_{23} & q_{13}-q_{23} & q_{12}-q_{23} & q_{12}-q_{23} & q_{12}-q_{23} & 0 & 0 & 0 \\
q_{24}-q_{14} & q_{24}-q_{14} & 0 & q_{24}-q_{14} & 0 & q_{12}-q_{24} & 0 & q_{24}-q_{12} & q_{24}-q_{12} & 0 \\
q_{25}-q_{15} & q_{25}-q_{15} & q_{25}-q_{15} & 0 & 0 & 0 & q_{12}-q_{25} & 0 & q_{12}-q_{25} & q_{25}-q_{12} \\
0 & 0 & 0 & 0 & 0 & q_{35}-q_{45} & q_{34}-q_{45} & 0 & q_{34}-q_{35} & 0 \\
0 & q_{14}-q_{34} & q_{13}-q_{34} & 0 & q_{34}-q_{14} & 0 & q_{34}-q_{14} & q_{34}-q_{13} & q_{34}-q_{13} & 0 \\
0 & q_{15}-q_{35} & 0 & q_{13}-q_{35} & q_{35}-q_{15} & q_{35}-q_{15} & 0 & 0 & q_{13}-q_{35} & q_{35}-q_{13} \\
0 & 0 & q_{25}-q_{45} & q_{24}-q_{45} & 0 & 0 & 0 & 0 & q_{24}-q_{25} & 0 \\
0 & q_{25}-q_{35} & 0 & q_{23}-q_{35} & 0 & 0 & q_{23}-q_{25} & 0 & 0 & 0 \\
0 & q_{24}-q_{34} & q_{23}-q_{34} & 0 & 0 & q_{23}-q_{24} & 0 & 0 & 0 & 0 \\
0 & 0 & q_{15}-q_{45} & q_{14}-q_{45} & 0 & q_{15}-q_{45} & q_{14}-q_{45} & q_{45}-q_{15} & 0 & q_{45}-q_{14} \\
\end{array}
\right).}
\end{equation*}
Thus, the first constraint is of the form
\eqlb{eq:SWCons}{\det L=0.}
Now the constraint (\ref{eq:SWCons}) fixes the coefficients \eqs{c_l} in (\ref{eq:qgenDec}) for \eqs{l\geqslant4}. In particular, for \eqs{c_4} and \eqs{c_5} the constraint gives
\eq{c_4=\frac23\frac{c_1^4}{c_0^3}-\frac73\frac{c_1^2 c_2}{c_0^2}+ \frac23\frac{c_2^2}{c_0}+2\frac{c_1 c_3}{c_0},}
\eq{c_5=\frac{20}{33}\frac{c_1^5}{c_0^4}-\frac{49}{33}\frac{c_2 c_1^3}{c_0^3}+ \frac{14}{11}\frac{c_3 c_1^2}{c_0^2}-\frac{37}{33}\frac{c_2^2 c_1}{c_0^2}+ \frac{19}{11}\frac{c_2 c_3}{c_0}.}
Moreover, the constraint fixes the lowest term in the series (\ref{eq:qgenDec}) to be $1$ or \eqs{\left(\hat a_i-\hat a_j\right)^2} up to some constant factor. These recurrences are satisfied by
the three sets of functions \eqs{q_{ij}}:
\begin{itemize}
\item{
corresponding to the $GL(N)$ elliptic Calogero system \eq{\label{tr1}q^{I}_{ij}=\brc{1-\frac{m^2}{\left(\hat a_i-\hat a_j\right)^2}}^{-1},}}
\item{
the $GL(N)$ elliptic Ruijsenaars model
\eq{\label{tr2}q^{II}_{ij}=\brc{1-\frac{m^2}{\sinh\left(\hat a_i-\hat a_j\right){}^2}}^{-1},}}
\item{the $GL(N)$ double elliptic system (where \eqs{\bar{\tau}} is the modulus of the second torus)
\eq{\label{tr3}q^{III}_{ij}=\brc{1-\frac{m^2}{\text{sn}\left(\left.\hat a_i-\hat a_j\right|\bar{\tau}\right){}^2}}^{-1}.}
}
\end{itemize}
Let us elaborate on the parameter count for the last of these (the former two being obtained as scaling limits of this). As we remarked above, there is an overall scaling of the functions $q_{ij}$ left undetermined by the constraint (\ref{eq:SWCons}). Further the scale $\sqrt{\mu}$ of the moduli $\hat a_i$, the mass $m$ and the period $\bar\tau$ are also parameters. Setting $q^{III}_{ij}={m^2}/{\mu}\brc{1-{m^2}/{\text{sn}\left(\left.\sqrt{\mu}\brc{\hat a_i-\hat a_j}\right|\bar{\tau}\right){}^2}}^{-1}$ scales $c_0=-1$ and the coefficients $c_{1,2,3}$ encode these
three parameters; the recursions for \eqs{c_l} (\eqs{l\geqslant4}) then express the remaining coefficients
implicitly in terms of $\mu, m, \tau$.
The results just obtained suggest the following hypothesis. The constraints (\ref{eq:g3ThetaId}) are actually theta-constant equations of the period matrix $T$ for the Seiberg-Witten curve associated with the $N$-particle double-elliptic integrable system for $N\geq 5$.
Thus, the equations can be used to determine the dependence of the period matrix on the Seiberg-Witten flat moduli, providing an independent method for calculating the Seiberg-Witten prepotentials (including the instanton corrections) of the $GL\brc{N}$ elliptic Calogero, Ruijsenaars and double-elliptic systems.
\section*{Acknowledgements}
The work is partly supported by grant
NSh-1500.2014.2, by RFBR grants 13-02-00457 (A.Mir.), 13-02-00478 (A.Mor.), 12-02-00594 (A.Z. and G.A.), by joint grants 13-02-91371-ST, 14-01-92691-Ind, by RFBR grant 14-01-31353-mol\_a (G.A.),
by grant RSCF 14-50-00005 (A.Z.), by "Dynasty" fund (A.Z. and G.A.),
by the Brazil National Counsel of Scientific and
Technological Development (A.Mor.).
\bibliographystyle{unsrt}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.